Semantic web based learning styles identification for social learning environments personalization

Abstract

Personal learning is a hot research topic in the field of web-based learning systems as there is no one appropriate learning path for all students. Many researchers are using semantic web technologies to find new approaches to develop personalized learning environments based on describing knowledge using ontologies. In this paper, we investigate the personalization of learning process taking advantage of the Social Semantic Web, Learning Styles and Bayesian Networks to provide students with recommendations of collaborators and relevant resources that best fit their needs. The proposed personalization approach is based on discovering students’ learning styles by means of an analysing of their behaviours. Semantic Web concepts are applied to describe pertinent entities and variables of the proposed Bayesian Network, defined by use of inference mechanisms. Some experiments were conducted and results on students’ learning styles estimation are compared with those obtained by use of MBTI questionnaire.

Keywords

Personalization recommendation Semantic Web learning styles Bayesian Networks

1 Introduction

Technology Enhanced Learning is a research discipline that addresses the use of Information and Communication Technologies within the context of collaborative learning [30]. Many systems have been implemented and many Learning Management Systems (LMS) have been really used offering environments with numerous tools to be used by learners to accomplish their different learning activities. However, what commonly characterizes these LMSs is the fact that most of them do not pay much regard to the social character of learning, which is an inherent characteristic of human being learning [22].

Social interactions are very important in enhancing the learning process [40] and researchers used the social web tools to enhance the learning quality. Recently, Web based learning environments have taken great advantages using the social web and emerging technologies like semantic web; learners and educators currently live in the world of Web 2.0 and social interactions are becoming very essential [11].

The Web was built to be used by human not to be understandable by machines. Machines cannot read and understand resources’ content. A great representation of entities with use of Semantic Web tools is required in order to make machines able to understand the meaning of terms and to process, share, and reuse resources [24]. Semantic Social Web was established after merging Social media and Semantic Web technologies aiming essentially the creation of well and explicitly represented knowledge of social interactions on the Web. The Social Semantic Web combines semantic web, social software and Web 2.0 technologies [21].

But even in a social-based environment, providing learners with the same content is not appropriate to satisfy their needs. They have different learning styles, different knowledge and preferences influencing the learning quality. Hence, the need for Personal Learning Environments (PLEs) became very crucial recently.

Researchers have used Semantic Social Web (SSW) to understand learners’ needs (from a cognitive perspective) and to provide them an accurate content according to their real knowledge; but it is not sufficient to provide an appropriate content which is suitable to their learning tendencies. Therefore, a real need to involve learning styles in the personalization process is very essential [31], accordingly, many researchers opt to incorporate learners’ learning styles usually determined by tests, these tests are a set of questions organized in such a way, which allow detecting the learning style after collecting the entire questionnaire answers [34]. The main and real problem that prevents the success of using these tests is the length of questions where learners usually don’t pay attention when giving answers, they choose arbitrary responses instead of thinking carefully before choosing the right answers [36]. Thus, obtained results can be inaccurate and may not reflect the real learning styles. To overcome this problem, other approaches have been proposed to automatically detect learning styles by observing and analysing students’ behaviours in the system, for example, the number of exercises done, number of collaboration invitations, number comments, tags and published resources can be used to discover each student’s learning style automatically.

In this paper, an enhanced social personal learning environment is addressed, the aim is to semantically represent all necessary system components in order to give meaningful learner’s model, pedagogical model and learning domain model to provide users with accurate content taking advantage of the intelligence mechanisms provided by the semantic web (inferring and reasoning). In a second stage, learners’ initial learning style is used to provide filtered and more suitable content. Final stage, which represents the main contribution in this work, is centred mainly on how automatically define learning styles in a social learning environment enhanced with semantic descriptions in order to provide the best personalization. A new method that automatically estimates learners’ learning styles with respect to the MBTI model is proposed, it is based on analysing their behaviours by semantic inferences. Quantitative and qualitative information about learners’ behaviours are analysed and modelled with use of Bayesian network model, semantic web inference mechanisms and semantic rules are used to define its explicit and implicit variables discovered by analysing learners’ actions and activities in the system. A set of semantic rules to discover these variables is described. Learners will have an accurate, appropriate and asserted content, personalized according to their behaviours and profiles.

The rest of this paper is organized as follows. In next section, we give a brief overview on Social Semantic Web, Personal Learning Environments and Bayesian Networks. In Section 3, some implemented approaches related to our research work are presented. In Section 4, the proposed personalization approach is described in detail. In Section 5, experimental results and the implemented system are presented. Finally, Section 6 provides a conclusion and future work.

2 State of the art

In this section, a brief overview on Personal Learning Environments, Social Semantic Web, Learning Styles and Bayesian Networks is given.

2.1 PLE and Social Semantic Web

A Personal Learning Environment (PLE) allows learners to find and to deal with resources according to their learning needs and preferences [28]. It comprises all learning tools and services used by learners to direct their own learning and to follow their pedagogical goals. Unlike LMSs, PLEs adopt a learner-centric learning approach, allowing learners to improve their learning process. In this context, a PLE based on tools and services inspired from recent technologies (namely Semantic Web, Web 2.0, Social Networks, Learning Styles, etc.) is proposed, by providing relevant recommendations to learners and putting them in charge of their learning process.

PLEs have taken great advantages from the Social Semantic Web [33]. Many researches done in this field shared the use of formal and shared vocabularies or ontologies to provide learning material with semantics [13]. For instance, in [25] an approach of workspace personalization using a service-oriented architecture and RDF/S is presented, in [41] researchers study semantic social interactions in e-learning system. The use of the Semantic Web models and technologies in e-learning systems allows the customization and the personalization of the learning content, the composition of new learning resources according to learners’ needs and adaption of interactions between the system and learners according to their learning objectives. Therefore, and based on some definitions in the literature, some citations approved that a PLE is a self direction platform and only learners are responsible to organize, maintain and increase their learning process. However, this requires some abilities in self awareness and maturity [5]. Effectively, learners are not at a maturity to promote their learning by themselves in an autonomous learning environment. Learners are habitually inexperience and irresponsible and failed in organizing their learning process. Consequently, we believe that a PLE can be: an environment providing for each learner an own learning plan and thus giving best learning paths, resources and collaborators according to her/his preferences, interests and knowledge [22].

However, the content provided may not interest them, due to the incompatibility with their preferences. Therefore, the need to incorporate learners’ learning styles in the personalization process is essential to provide them with best learning resources that best fit their needs.

2.2 What is a learning style

Many researches aim at identifying individual differences [39]. The learning style theory has emerged according to these researches. It is defined as mental processes used by people to collect and treat information in an optimal way. Learning style may be defined as “the attitudes and behaviours which determine an individual’s preferred way of learning” [26]. Till now, a lot of work has been done on the development of learning models. But, many discussions are still open in order to accept different theories [6]. Different tools are used to determine learners’ learning styles [1] and principally with use of questionnaires that categorize each person according to her/his learning styles, namely: Kolb questionnaire [29], Honey and Mumford questionnaire [26]. Felder and Solman proposed the Index of Learning Style Questionnaire ILSQ [16]. Among the different proposals to model learning styles, the Myers Briggs Type Indicator (MBTI) test has been chosen in this work since it has achieved great success.

The MBTI test is based on Carl Jung’s theory of types outlined in 1921 in his work Psychological Types [3]. The theory indicates that human beings are either introverts or extraverts, and their behaviour follows these psychological types. The Myers-Briggs theory classifies personality preferences according to the four Jungian psychological types:

Extraversion (E) or Introversion (I): Does the learner prefer to communicate with the external or with the internal world?

Sensing (S) or Intuition (N): How the learner prefers to take information?

Thinking (T) or Feeling (F): How the learner prefers to take decisions?

Judging (J) or Perceiving (P): How the learner deals with the external world?

Combinations of these preferences resulting in 16 learning style, they are typically designated by four letters to represent a person’s tendency towards four dimensions. For example, ESTP refers to Extroversion, Sensitive, Thinking, and Perceiving. This does not mean that a person has only these four preferences, but that the latter are the dominant ones. This is very interesting to plan pedagogical approach and control the learning process. For example the dominant preference found in ISFP, INFP, ESFJ and ENFJ is sensing [14].

For instance, Sensing learners like so much using the five senses for acquiring information, they prefer concrete facts and organization, they prove themselves when it comes to memorization, they are realistic and they dislike theory. Intuitive learners deal with theory firstly before using facts, they like to be creative, and innovative. Thinking learners use their minds more than they use their hearts, and most of the time they do not pay any attention to feelings of others, they like reasoning and solving logical problems. Feeling learners use their heart more than their minds, they take the decision basing on their feelings towards peoples and their likes and dislikes, they prove themselves in social work and other helping situations, they feel good after helping others [14].

The fact that relying only on using Web-based tests to provide best personalization is not guaranties, since students are not conscious of future uses of questionnaires and they do not give importance when using the test, researchers tend to use other approaches aiming to automatically detect learning styles by observing and analysing learners’ behaviours.

Recently, two methods that automatically detect learning styles are used: the data-driven and the literature-based [36]. In this work, we have only considered the first method – Detection is based on analysing students’ behaviour in the system, the method serves to overcome the drawbacks of the manual method. Researchers associated with the first method the use of: Bayesian Network [18]; Hidden Markov Models and Decision Trees [9]. The data-driven approach uses sample data in order to build a model that simulates the questionnaire for detecting learning styles from the students’ behaviour, its advantage is that the model can be very accurate due to the use of real data. The authors in Cha et al. [9] investigated the use of Decision Trees (DT) and Hidden Markov Models (HMM) for identifying learning styles according to Felder-Silverman learning style model (FSLSM), they conclude that DT and HMM are appropriate for determining certain dimensions of the FSLSM. However, the proposed method gives good results only when students have a moderate or strong preference, due to data using restriction. In another study, Garcia et al. [18] observed the behaviour of learners during an online course and performed two experiments to show the efficiency of using Bayesian networks for identifying learning styles based on students’ behaviour analysing. The result showed that the Bayesian network obtains good results for detecting learning styles.

In this work, a Bayesian network is used to represent and to estimate students’ learning styles. In the following, we present a brief overview of BN, and then we motivate our choice in using it.

2.3 Bayesian Networks

A Bayesian Network (BN) is a probabilistic graphical model that enables discovering new knowledge through the use of expert domain knowledge and statistical data [27]. Nodes represent by random variables of features in a certain domain of interest. In this work, random variables are the different learning styles of MBTI model and factors determining each style. These factors are extracted from interactions between learners, teachers and the environment (learner-teacher, learner-learner & learner-system). Throughout the use of a BN, relationships between learning styles and factors determining them can be easily modelled. Arcs connecting the graph nodes represent a probabilistic relation between variables. A BN also represents a joint probability distribution specified by a set of conditional probability tables (CPT). Each node has a particular CPT that specifies the probability of each node in relation with each possible combination of states of its parents [37].

The Bayes’ theorem (shown in Eq. (1)) links conditional and marginal probabilities. It gives the conditional probability distribution of a random variable A. Equation (1), reads: the probability of A given B is the probability of B given A times the probability of A, divided by the probability of B. $P (A | B) = (P (B | A) P (A)) / P (B)$ (1) An essential characteristic of BN is the Bayesian inference mechanism which aims in general to find the conditional distribution of variables [36].

Bayesian networks allow the acquisition, representation and use of knowledge of a certain system [27], like a learning environment in our case. Such operation is performed according to the usage context of the system, in order to: predict, simulate and control the behaviour of students and to analyse data and make decisions about their preferences. As a Bayesian network may be constructed either from data, by learning, or from explicit modelling of the domain [27], it is sufficient that either form of knowledge is available to consider using this technique in an application like a learning environment.

A Bayesian network is the representation of a probability distribution [32]. If the structure of the distribution is imposed, it would be easy to calculate the impact of each new instance of the parameters of the distribution, in our case, these parameters represent the student’s actions that highlight the decision about his/her learning style.

Compared to systems based on deterministic rules, most often used in expert systems, Bayesian networks allow incorporating uncertainty in the reasoning. They are therefore suitable for problems where uncertainty is present, either in the observations or in the decision rules [10]. This is one of the most important characteristics that pushed us to use BN in order to estimate the student’s learning style, since she/he can perform a lot of actions in the system and because of the immaturity and lack of awareness of the student in expressing his/her needs, the estimation can give inaccurate results. Here comes the importance of BNs in analysing multiple simultaneous actions. Deterministic techniques such as decision trees often lead to only one analysis at a time.

Bayesian networks are particularly suited to analysing student’s behaviour, because they offer the possibility of integrating heterogeneous sources of knowledge (human expertise and statistical data). In our case, human expertise is the knowledge experts of the learning domain (content creators, teachers, etc.); they decide to match the appropriate learning resources to different styles.

Bayesian networks are probably one of the most appropriate technologies to build intelligent systems. They assure [23]: Autonomy, which is presented with the ability to provide decisions under uncertainty, or in the absence of certain information. Motivation, which can be presented by certain types of inferences. Responsiveness, which is the principle of inference in Bayesian networks. Adaptation to the environment, which is made possible by incremental learning capabilities of Bayesian networks.

Bayesian networks are an ideal model for injecting knowledge and intelligence to the learning system, allowing it by the use of a decision-making model to decide in an uncertain environment, and to adapt when new information is entered. For instance, when a student performs contradictory actions or radically changes her/his behaviour.

3 Related work

To enhance the learning personalization, some research works have been carried out on leveraging the use of the Social Semantic Web technologies, Learning Styles, and Bayesian networks. For instance, we can find: the Ensemble project [15] which explores the potential of semantic technologies to support and enhance learning in higher education, the adopted research approach assumes the usage of Semantic Web technologies combining with features of Social Software in order to allow reuse, reconfiguration and adaptation of resources. Didaskon [12] is a platform for automatic composition of learners’ learning paths. The selection of learning objects is based on semantic annotation of users’ profile. It can create a learning path fitting better the need of a particular learner. It employs the User’s profile that can store learner’s needs, skills, and his/her traces, etc. The work in [7] describes how the content model is updated according to the learner model, the learner type being modelled, etc.

Many works addressing learning styles have been carried out, among them: ARTHUR [20] uses three learning styles (visual-interactive, reading-listener, textual), CS388 [8] uses Felder and Silverman styles. The INSPIRE system [35], allows interaction personalization in a web based educational hypermedia system.

Works employing BN models are presented, among them: ANDES [19] using BN techniques to model knowledge about learners in Physics. IDEAL [38] uses BN technique to classify the learners’ level into novice, beginner, intermediate, advanced, or expert. In [2], authors build a Bayesian model to detect a learner’s behaviour towards the e-learning system; the observable behaviour is recorded in a log file. In [17], authors use a BN to detect the learners’ learning style and to evaluate his/her knowledge in an intelligent tutoring system. In [42], a BN is used to model the learners’ behaviour in order to detect their preferences and helping teachers to make decisions.

In our previous work, which addresses the use of semantic social web techniques in order to enhance the learning personalization [22], we showed on the one hand that the use of semantic web techniques is very useful in providing students with an accurate content basing on a semantic representation of learning resources and properties relating them with users and their needs. On the other hand, we argue that the use of social web techniques helps significantly students in creating ties between them in order to increase the communication process and thus to increase personalization results. But, it is found that using SSW techniques only is not enough to fulfil students’ needs. The resources’ content may be suited to the learners’ learning goals, but not surely suitable to her/his preferences, this is what prompted us to incorporate the identification of students’ learning styles through a deep observation and analysing of their behaviour on the system basing on the use of SSW techniques combined with a BN model in order to provide them with the most suitable content.

4 Personalization approach by means of learning styles detection

A hybrid approach of social learning personalization is our proposal, it is based on the use of semantic social web in order to provide learners with an accurate content, the use of learning styles to provide an appropriate content and the decision of providing personalized content will be validated by a Bayesian Network model.

4.1 Objectives

The aim in this proposal is the personalization of learner’s learning process (tasks and content) according to his/her styles (determined by the MBTI test and verified by the Bayesian model). Therefore, all knowledge presented in the system is modelled by use of semantic web techniques, to provide learners with accurate content, proceeding from the intelligence mechanisms provided by the semantic web (inferring and reasoning). In a second stage, learner’s initial learning style is added to the provided content in order to generate an accurate and appropriate content. In the last stage, filtered content is subjected to a Bayesian Network (BN) model in order to generate an accurate, appropriate and asserted (verified) content due to the Bayesian inference mechanism which make inferences about learners’ learning styles (LS) basing on the analysis of their behaviour. This process is called the 3A strategy of content personalization (Accurate, Appropriate and Asserted content), as presented in Fig. 1:

SSW: provides an accurate content, but may not be liked by learners.

SSW+LS: provides an accurate and appropriate content, but because the learner may not express his/her real learning style in the beginning; there is a chance that he/she will not like the content.

SSW+LS+BN: provides an accurate, appropriate and asserted content – verified by the BN model – that should be the real content requested by the learner.

Fig. 1.

The 3A strategy of content personalization.

As shown in Fig. 2, the approach of content personalization is resumed in the following Phases:

Detecting the initial learning style of learners using the MBTI questionnaire.

Knowledge modelling: SSW techniques are used to represent the learner model (Cognitive and behavioural profile – generated according to learners’ interactions within the system and their collaborators); to represent also the learning domain model and the pedagogical model.

Content Filtering: the semantic web inferring mechanism is used in this phase in order to discover additional knowledge on the learner’s needs, thus to prepare a potential list of resources fitting the initial model of the learner and adjusting her/his needs, preferences and skills in order to bring closer with other learners. The proposed pedagogical model is also used to identify pedagogical resources types that must be presented to learner according to his/her profile.

Linking learners’ behaviour with learning styles: combining the characteristics of the learner learning style and the characteristics of the filtered resources to provide him/her the appropriate content.

Decision model: a Bayesian Network is used to model quantitative and qualitative information about a learner’s behaviour with the e-learning system. The model can infer the learner’s learning style given different observed behaviours and knowledge.

In a further Phase, Learning style adjustment: updates and adjusts the learner’s learning style according to the BN decisions. And finally, Classification: after detecting learner’s learning styles. The model classifies the content to appropriate and non appropriate resources.

Recommend the appropriate resources to the learner as an [accurate [and appropriate and [asserted]]] content.

Fig. 2.

Model of content personalization.

The following algorithm expresses the different steps of the proposed approach of content personalization.

Algorithm of content personalization: search for adequate pedagogical materials depending on the learning style, the pedagogical activity and behavioural indicators.
Detect the initial learning style;
Find the most appropriate pedagogical material
Compose the new pedagogical material and recommend it to learners;
Observe learner behaviour and calculate behaviour indicators;
Call the BN and estimate the LS from behaviour;
While evaluation_score does not reach a certain threshold
Make a new estimate of the learning style;
Find the most appropriate pedagogical material;
Compose the new pedagogical material and recommend it to the learner;
Observe learner behaviour and calculate behaviour indicators
Call the BN and estimate the LS from behaviours;
End.

4.2 Phase 1: Learner’ initial learning style detection

When a learner is connected to the system, a session is initiated based on her/his learning style and a pedagogical strategy is recommended to him/her. It also includes a part to test the knowledge acquired in each lesson, which contains several multi-choice questions and other evaluation methods.

Each profile stores personal information provided by the learner, e.g. the name, age, educational level and languages and other meaningful attributes. When learners are registered, the system finds deals with detecting and storing the result that indicates a preference of one of the 16 learning styles. These styles are stored in the learner model used in the first initial learning personalization result and indicate a preference on one of the 16 learning styles.

4.3 Phase 2: Knowledge representation in the learning environment

In order to personalize the learning process, all managed knowledge in the system is described with use of ontologies to formalize the domain knowledge of the Personal Learning Environment. The use of Semantic Web technologies allows us to: enhance the semantic representation with standardized tools, associate formal descriptions to learning resources, make formal reasoning (resources retrieval, resources compositions, etc.), search pedagogical resources tailored to the learner, compose new resources from existing resources and adapt the interaction between the system and the learner according to her/his objectives. As part of our work, the use of ontology allows us to formally define the different actors (learners and teachers), learning resources as well as learning domains and learners’ interests, etc.

4.3.1 The proposed ontology

The ontology describes knowledge about the domain of learning personalization, it allows to formally define different users with their roles (learner, teacher, etc.), resources (courses, tutorials, videos, etc.), learning styles and preferences, pedagogical strategies, tags (keywords that represent the user’s understanding towards a learning domain) and the communication tools (comments and messages, etc.). Specifically, the use of the ontology allows us to:

Annotate and describe learning resources by a formal vocabulary which allows their sharing and reuse.

Define formally the learning domain.

Define formally the users.

Provide accurate learning resources at user’s requests, as long as resources are indexed by the ontology, this allows to compose new learning support from several elements of knowledge (e.g. aggregation of definitions, illustrations and exercises, etc.).

Reduce the ambiguity of the used terminology in the system to improve the efficiency of research and to find the best appropriate resources suitable for a specific learning situation.

Improve the result of users and learning resources recommendations through the use of the inference mechanisms.

The ontology describes two types of knowledge: concepts and properties organized hierarchically. For each property, a subject and object concept is defined (the domain and co-domain). Knowledge of the learning domain is instantiated to describe the concepts of the domain (e.g. the domain of object-oriented programming), learner’s characteristics (e.g. cognitive level, preferences, etc.) and annotations on learning resources (e.g. resource type, author, etc.).

The approach aims at fostering interchange, cooperation and increase interactivity between users, so that each one can achieve his/her learning objectives with an easy and accurate way, and in accordance with his/her learning style and needs. Indeed, interactions and collaboration are at the heart of the social learning, this is why everything that connects the user to any other entity is presented. Through this study, it is found that the most important elements that create relationships between users and have a direct influence on their feedback are: learning content (resources, tag), communication tools (comments, messages and chat), learning styles and preferences, etc. These elements represent the key concepts of our personal learning ontology.

Fig. 3.

Personal Learning Ontology.

Figure 3 shows the personal learning ontology “PL”. It is composed of different classes and sub-classes devoted to formalization of different knowledge about users, knowledge about learning units, styles and preferences, and knowledge about the pedagogical strategies and activities. The “ User ” class, for example, represents the users of the system, it includes learners and teachers, users are instances of the concept Learner or Teacher. The “ Learning Domain ” class represents domains of learning, it is divided into two main sub-classes, the “Theoretical” class to represent the theoretical learning domains (mathematical logic, data analysis, etc.) and “Practical” class for the representation of practical learning domains (programming language, development environment, etc.). The “ Learning Unit ” class represents learning resources used in the system. In our work, a learning unit contains a set of modules, each module contains several chapters and each chapter contains a series of: courses, tutorials and practical works. Each course is associated with a set of educational content (Definition, Abstract, Example, etc.) that have different format (for example: the definition is a pdf file and illustration is a video file). Semantic relations between these entities are found, for example, the definition is the first pedagogical component to be presented to the learner, it means that the definition is a prerequisite for all other elements. The “ Learning Styles ” class represents the learning styles of learners, they learn easily when they have the opportunity to use their preferred learning style. Each preference has some predictable effects on learning styles. User’s intuition type, for example, prefers to use metaphors and symbolic language and they tend to associate ideas.

Properties connecting objects

Using classes alone is not enough to describe the knowledge manipulated by the system. Significant properties that describe different objects are also identified. With the perspective of formalization in OWL, the properties of object values to those of literal values are distinguished. Object properties are presented in Fig. 4. For example, the property hasTeacher, connects the concept “Learner” to the concept “Teacher”, it represents the fact that a learner follows the teacher’s courses hasTeacher(Learner, Teacher), the property WantToLearn, connects the concept “Learner” to the concept “LearningDomain”, describes that a learner is interested in learning a certain domain. The property mayTeacherOf , connects the concept “Teacher” to the Concept “Learner”, represents the fact of recommending a teacher to a learner. The hasRequirement property, is used to connect the concept “LearningDomain” to its post requisites “LearningDomain”, it defines a prerequisite relation between concepts, i.e. relations where concepts are taken sequentially.

Fig. 4.

Properties connecting the ontology objects.

Fig. 5.

Links between the ontology and different models.

Fig. 6.

Excerpt from the hierarchy of learning domains.

4.3.2 System’s knowledge modelling

Different models of the personalized learning system (the domain model, the pedagogical model and learning model) are presented in RDF. Figure 5 presents an overview of links that connect the ontology to different models. RDF annotations are knowledge instantiation described in OWL. For example, the triplet <Halimi, hasTeacher, Seridi> is instantiating the following knowledge described in OWL: hasTeacher (Halimi, Seridi) and natural language “Seridi is the teacher of Halimi.” This RDF annotation belongs to the learner model, it is used to model the characteristics of the learner Halimi, knowing that hasTeacher is a property in the ontology. The “pl” prefix is used to denote the URI of the ontology.

4.3.3 Learning domains knowledge modelling

The ontology allows us to describe different domains of learning, as shown in Fig. 3, learning objects in the system are broken down into two main classes, the “Theory” class to represent the theoretical learning domains (mathematical logic, data analysis, operations research, etc.) and “Practical” class for the representation of practical domains of learning (programming languages, development environments, design tools, etc.). In Fig. 6, an excerpt from the hierarchy of learning domains is presented.

4.3.4 Learner’s model

The learner’s model represents the knowledge that the system has on the learner. In this work, the knowledge of the learner is modelled, concentrating on his/her level of topic/course mastery, his/her preferences (collaborators, preferred format of resources, etc.), his/her learning goals, learning history and learning style. The learner’s model is built by instantiating concepts belonging to the ontology of personal learning PL.

Fig. 7.

Learner’s model.

For example, in Fig. 7, the learner “Halimi” with the id “learner37” is an instance of the “Learner” concept defined in PL and is connected with the property learned to the Concept/instance “java”, means that the learner “Halimi” has learned the java language. The property hasStyle describes the fact that the learner “Halimi” has a learning style “ISTP” and the property isCollaboratorOf describes that the learner “Halimi” and “Kirati” are collaborators. This knowledge of the learner is modelled as a set of attribute-value pairs, the knowledge level of the learner in relation to a specific learning domain is presented, this knowledge is obtained by applying the system’s evaluation methods, namely: MCQs, exercises, etc., by the end of each assessment the learner profile will be updated by one of the following values: “low, medium and good”. Figure 8, represents a serialization of the learner model in N3 representation.

Fig. 8.

Serialization of the learner’s model.

4.3.5 Pedagogical approach modelling

To achieve the goal of providing learner with the best learning content personalized according to his/her needs, knowledge, preferences and learning styles, a pedagogical approach is provided.

The approach consists of dividing the learning resources to a set of units and each unit is divided into a set of modules and each module is also divided into a set of chapters and each chapter contains a series of lectures, tutorials and directed works. The pedagogical model is used to describe how is the sequence of the various pedagogical activities (courses, explanation, definition, exercise, etc.) is carried out. This model allows us to decide what type of learning resources must be presented to the learner taking into account his/her profile. For example, some learners prefer to understand a particular concept, from a concrete example of the concept and then go to a conceptualization of the notion.

The model describes different pedagogical strategies that enable a learner to take a concept of a certain learning domain. It is constructed by defining the different pedagogical entities (classes, exercise, definition, example, etc.) and their sequencing order, for example, an introduction of a concept must be presented to the learner firstly, then he/she must have explanations or illustrations, the latter do not have a specific learning order, i.e. the learner can have the illustration before the explanation or vice versa (they may be presented to the learner in parallel), finally, he/she must have an assessment exercise. This sequencing order is represented by semantic relationships that connect concepts, e.g. some pedagogical entities cannot be presented to the learner because they are Prerequisites of other entities (x hasRequirement y) that have not yet been acquired by the learner (the learner cannot have an exercise on a specific concept before he/she must assimilate definitions, illustrations and examples).

Fig. 9.

Pedagogical model.

Another advantage of our pedagogical model is the possibility to take into account the learning styles and preferences during the personalization process, the fact that learners are different, and one course structure cannot serve everyone, for example a learner who has the ISTP profile does not like a purely theoretical content, he/she likes to go to the essentials and spend many time for experimentation; another learner who has the ISTJ profile, likes to start the learning by theoretical concepts, definitions, theorems, etc., hence, this decomposition approach is very useful to avoid learners to have a content that does not comply with their attitudes.

When creating the learning content, the teacher must indicate the type, level, target profile, pedagogical strategy and assessment form, of each pedagogical entity, for example, he/she creates a resource as follows: Document1 (Type: Illustration, Level: Easy, Strategy: byThinking, Evaluation: QMC). The system then automatically assigns pedagogical entities to appropriate profiles.

The pedagogical model is presented in Fig. 9, where the “LearningUnit” class describes knowledge about pedagogical objectives components used in the system. Each course includes a set of pedagogical entities (definition, abstract, example, etc.), entities may have different format (for example: the definition is a Pdf file and illustration is a video file). Semantic relationships that connect those entities are presented also, for example, the introduction is the first pedagogical component to be presented to the learner, means that the definition is a prerequisite for all other elements, represented by isPrerequisiteOf property in the triplet <Introduction, isPrerequisiteOf, definition>. The Illustration and the Example entity may be presented to the learner in parallel because taking illustration can be helped by the assimilation of the example and vice versa. The concept Document represents the physical documents published by different users, namely: pdf files, images, videos, texts, etc. Each document represents a unique pedagogical entity such as Document1 is an Introduction, it has also semantic relations, some of them are presented in the following:

hasPedagogicalStrategy (Doc1, ByReading), means that the Doc1 is an introduction and has a strategy by reading therefore destined to learners with visual preference.

hasForm (Doc1, TextFile), means that the teacher has created the Doc1 and he/she has published it as a text file.

hasEavaluationType (Doc1, GapText) means that the evaluation of learners on Doc1 is using text gaps.

definitionIn (Java, Doc1): means that a definition of java is presented in Doc1.

4.4 Phase 3: Content filtering

The semantic web inferring mechanism is used in this phase in order to discover additional knowledge on the learner’s needs, thus to prepare a potential list of resources fitting the initial model of the learner and adjusting her/his needs, preferences and skills in order to bring closer with other learners. The proposed pedagogical model is also used to identify pedagogical resources types that must be presented to learner according to his/her profile. In the following, some inferring rules used to infer new knowledge from learners’ interactions are presented.

4.4.1 Inference about content and users

In general, inference on the Semantic Web can be characterized by the discovery of new relationships among instances. It is a process of reasoning that is based on acquired knowledge, which revolves around the fundamental rules to allow obtaining new information. This additional information can be defined using a set of rules.

To extract more additional knowledge about the content and users in order to increase the process of personalization, a set of rules defined with the SWRL language (Semantic Web Reasoning Language) are added to the ontology, the necessary rules in the approach of personalization are cited in the following:

collaboratorOf ( $? z, ? y$ ) ∧ hasStudent( $? x, ? y$ ) → mayStudentOf ( $? z, ? x$ )

If z is a collaborator of y and x is a teacher of y therefore z shall probably be a learner of teacher x.

hasTeacher( $? x, ? y$ ) ∧ collaboratorOf ( $? z, ? y$ ) → mayTeacherOf ( $? x, ? z$ )

If x is a teacher of the learner y and y is a collaborator of the learner z, therefore x is likely a teacher of learner z.

learned( $? x, ? d$ ) ∧ wantToLearn( $? y, ? d$ ) → canHelp( $? x, ? y$ )

If the learner x mastering a learning object d (PHP for example) and if another learner is trying to learn the same object, then the learner x can give help to the learner y.

putTag( $? x, ? t$ ) ∧ putTag( $? y, ? t$ ) → mayCollaboratorOf ( $? x, ? y$ )

If two learners have put the same tag on the same learning object, therefore, these two learners can be collaborators, the fact that the action of tagging reflects the understanding of a user towards a learning domain.

post( $? x, ? c$ ) ∧ postedOn( $? c, ? d$ ) ∧ User( $? x$ ), (publish max 0 Document)( $? x$ ) ∧ talksAbout( $? d, ? l o$ ) → wantToLearn( $? x, ? l o$ )

If a learner x posted a comment on a document d that addresses a learning object and x is not the creator of this document, therefore, the system can discover that the learner wants to learn the object $l o$ .

like( $? x, ? d$ ) ∧ talksAbout( $? d, ? l o$ ) → wantToLearn( $? x, ? l o$ )

If a user x has liked a document d that addresses a learning object $l o$ , therefore, the system can infer that the user x has an interest towards the learning object $l o$ .

download( $? x, ? d$ ) ∧ talksAbout $(? d, ? l o))$ ∧ putTag( $? x, ? t$ ) ∧ putOn( $? t, ? d$ ) → learned( $? x, ? l o$ )

If a user x has downloaded a document d that deals with a learning object $l o$ and he/she has added a tag on the document d, then, the system can infer that the user x has learned the object $l o$ .

download( $? x, ? d$ ) ∧ publishedBy( $? d, ? y$ ) → hasInterest( $? x, ? y$ )

If a user x has downloaded a document d and the latter was published by the user y, therefore, the system can infer that the user x interested in user y.

putTag( $? u, ? t$ ) ∧ putOn( $? t, ? d$ ) ∧ talksAbout( $? d, ? l o$ ) → learned( $? u, ? l o$ )

If a user u put a tag t on a document d that deals with a learning object $l o$ , the system can infer that the user u has learned the object $l o$ .

publish( $? u, ? d$ ) ∧ wantToLearn( $? z, ? d$ ) ∧ collaboratorOf ( $? u, ? z$ ) → canHelp( $? u, ? z$ )

If a user u has published a document d and the user z wants to learn the same document d, therefore, the system can infer that the user u can give help to the user z.

4.4.2 Recommendation of users and content

According to our definition of personalization (presented in the Section 2.1) the system should provide learners with best users and learning resources that meet with their needs, according to his/her interests, preferences and knowledge. This means that enhancing recommendations is a very important operation in order to increase the personalization; therefore, through this study it is found that merging the Social Web and the Semantic Web can give very significant improvements for the result of recommendations, the fact that the system will be able to understand the real needs of its users and to discover new hidden knowledge.

The enhanced recommendation approach based on the inference mechanisms of the semantic web is presented in the following.

Recommendation of users

Properties used for users’ recommendation are: mayCollaboratorOf, mayStudentOf, mayTeacherOf, and canHelp, i.e. if one of these properties exists in the RDF model of a user; this means that he/she is able to get recommendations. The rules of users’ recommendation are presented in the following.

collaboratorOf ( $x, y$ ) & collaboratorOf ( $y, z$ ) → mayCollaboratorOf ( $x, z$ )

y and z are collaborators means that there’s a strong possibility that mutual interest exists between y and $z, x$ and y are collaborators and they have a common interest, so z is recommended to x because there will be a high probability that x and z have the same interests.

putTag( $x, t$ ) & putTag( $y, t$ ) → mayCollaboratorOf ( $x, y$ )

If two different users put the same tag for the same learning object: each one is recommended to the other because both have expressed the same understanding towards the learning object.

putTag( $x, t$ ): the user x put a tag “t”

putTag( $y, t$ ): the user y also has put the same tag “t”.

So, y is recommended to x, the fact that both are expected to have the same needs and the same interests.

putTag( $x, t$ ) & putTag( $y, s$ ) & sameAs( $t, s$ ) → mayCollaboratorOf ( $x, y$ )

If a user put a tag on a learning object and another user put a different tag on the same object, but the second tag has a relation like sameAs or hasRequirement with the first tag, each one is recommended to the other

putTag( $x, t$ ): the user x put a tag “t”

putTag( $y, s$ ) the user y put another tag “s” different of “t” but, “t” and “s” are the same concept. So, y is recommended to x, the fact that both are expected to have the same needs and interests, despite each of them was expressed with different way.

collaboratorOf ( $x, y$ ) & studentOf ( $y, z$ ) → mayStudentOf ( $x, z$ )

If the property mayStudentOf connects a learner to a teacher, then, learner x is recommended directly to the teacher z.

collaboratorOf ( $x, y$ ): learner x is a collaborator of y

studentOf ( $y, z$ ): the learner y is a learner of the teacher z. So, learner x is recommended to teacher z, the fact that the learner x could be a potential learner of teacher z.

collaboratorOf ( $x, y$ ) & teacherOf ( $y, z$ ) → mayTeacherOf ( $y, x$ )

If the property mayTeacherOf connects a teacher y to a learner x, then, the teacher y is recommended directly to the learner x.

collaboratorOf ( $x, y$ ): learner x is a collaborator of learner y

teacherOf ( $y, z$ ): y is a teacher of learner z. So, the teacher y is recommended to the learner xthat the teacher y could be a potential teacher to learner x.

canHelp( $x, y$ )

If the property canHelp connects a user x to a user y, then, the user x is directly recommended to user y.

learned( $x, d$ ) & wantToLearn( $y, d$ ) → mayCollaboratorOf ( $x, y$ )

If a user x masters a learning domain and a user y wants to learn the same domain, then user x is directly recommended to user y.

learned( $x, d$ ): User x masters the learning domain d

wantToLearn( $y, d$ ): the user y wants to learn the same domain d. So, user x is recommended to user y that user x can help the user y.

learned( $x, d$ ) & wantToLearn( $y, f$ ) & required ( $f, d$ ) → canHelp( $x, y$ )

If a user x mastering a domain d and a user y wants to learn a domain f and that f has a relationship with the domain d (e.g. f required d), then user x is recommended to user y.

learned( $x, d$ ): User x masters the domain d

wantToLearn( $y, f$ ): User y wants to learn the domain f, required( $f, d$ ): but in the ontology, it is defined that in order to learn the domain f , the user must first learn the field d. So the user x is recommended to the user y, the fact that x has already learned the domain d which is a requirement for the domain f, so the user x can help the user y.

Learning resources recommendation

A very important rule in our approach of content recommendation is to recommend automatically to learners the content of any user that may be interesting for their learning, i.e. providing for learners the content related to users recommended by the rules defined in the previous section (users discovered by the properties: mayCollaboratorOf, mayStudentOf, mayTeacherOf, and canHelp). The property wantToLearn is used to start the recommendation of learning resources, i.e. if this property exists in the RDF model of a user (wantToLearn (user, resource)), this means that the user is looking to learn the learning domain represented by this resource. The recommendation rules of learning resources are presented in the following.

collaboratorOf ( $x, y$ ) & publishedBy( $d, x$ ) → wantToLearn( $y, d$ )

Learning resources of a user are recommended directly to his/her collaborators.

collaboratorOf ( $x, y$ ): learner x is a collaborator of learner y;

PublishedBy( $d, x$ ): document d published by the user x. So, the document d published by the user x is recommended to the learner y, the fact that two users can have the same interests and they are present in the same community.

hasStudent( $x, y$ ) & publishedBy( $d, x$ ) → wantToLearn( $y, d$ )

Resources of a teacher are directly recommended to his/her learners.

hasStudent( $x, y$ ): y is a learner of teacher x;

publishedBy( $d, x$ ): document d is published by the teacher x. So, the document published by the teacher x is recommended to learner y, the fact that the learner takes courses of Teacher x.

(putTag( $x, t$ ) & putOn( $t, d$ )) & ((putTag( $y, s$ ) & putOn( $s, e$ )) & sameAs( $t, s$ ) → wantToLearn( $x, e$ ) or wantToLearn( $y, d$ )

Different content tagged by the same tag is recommended to users.

putTag( $x, t$ ) & putOn( $t, d$ ): user x put a tag t on the document d.

putTag( $y, s$ ) & putOn( $s, e$ ): the user y put a tag s on the document e.

sameAs( $t, s$ ): tag t and tag s are the same.

So, documents published by the user y are recommended to user x, and vice versa, the fact that the documents are tagged with the same tag.

4.5 Phase 4: Linking learner’s behaviour with learning styles

Each learner has a natural preference. Since the beginning, individuals exhibit differences between ways of learning: some prefer to receive complete and accurate instructions before starting a new task, some prefer to immediately go to action and learn globally and others need to complete the current task before moving on to the next, etc. [14].

There are specialized questionnaires to determine the learner’s learning style, e.g. ISTP, ENFJ, but in this work, the aim is to identify it according to analysing his/her behaviour in the system, the relationships between the learning style and the factors determining it (learner’s actions) have been determined. The information used to detect the style is obtained by analysing learner’s model, which is an RDF file. This file contains records of the tasks carried out by the learner in the system and his/her participation in activities such as adding resources, tagging, commenting, chatting, etc. Table 1, resumes relations between actions carried out by the learner in the system and the preferred learning characteristics that determine the learning style.

Table 1
Learner’s behaviour vs. learning styles

Style Preferred characteristics Learner’s behaviour (Actions observed on the system)

Sensory ISTJ ISFJ ESTP ESFP Uses traditional curriculum and Phase by Phase. When a learner looks for a learning object xand the system answer him/her that in order to take xhe/she must take ybefore. If he/she accepts and search again for the recommended object y, the system records that he/she prefers Phase-by-Phase method.

Likes using past experiences and standard ways to solve problems. If he/she is looking in the forum for an old exercises on his/her domain of interest, Search_Old_Exercise.

Enjoys applying what is already known by giving examples and details. If he/she publishes objects that he/she masters, or makes comment about objects that he/she masters (Makes Comment_on_Mastered_Object) or answering message about objects that he/she masters or tag objects that he/she masters.

Respect of the deadlines and rules. If he/she returns the solution in the right time and exactly as requested, Solution_Delivery_Time.

Likes suggestions that are straightforward and feasible. If he/she likes the recommendations provided by the system and finds it appropriate with his/her needs, Likes_Recommendations.

Is inclined to follow an agenda. If he/she respects the agenda proposed by the teacher to take different objects, for example: read theoretical concept, then read examples, finally, test knowledge with a MCQ, Respect_Agenda.

Likes to do practical things and prefers realistic applications. He/She prefers solving exercises, quizzes and exams, creates simulations, etc.

Rarely makes errors of facts. She/He shows a high score in quizzes and exams, High_Score_Evaluation.

He/She is inclined to express by critics. If he/she uses in his/her comments or in his/her messages critic words, Critic_comments.

Dislike the purely Theoretical courses. Most of times, he/she looks for examples, exercises, podcast tutorials, diagrams, etc. he/she will be very enthusiastic when he/she receives recommendations of practical objects, Search_for_ practical_objects, likes_ practical_objects.

Communicative and sociable. He/she Sends and accepts a lot of collaboration requests. He/She has many collaborators, Lot_of_collaborators.

Intuitive INFJ INTJ ENFP ENTP Focuses on conceptual understanding and the use of self-instructional methods for learning. He/she likes to deal (publish, tag, comment, like, etc.) with theoretical objects, Likes_ theoretical_objects, he/she prefers finding alone his/her objects; he/she does not pay attention to recommended objects.

Likes solving new and complex problems. Looks for solving new problems initiated by teachers, and tries to be the first one to find the solution, he/she does not like old problems being processed by other users, Like_new_problems.

Enjoys learning new skills more than using them. She/He prefers learning knew knowledge, but, he/she does not participate in sharing it with other users, Learning_knew_objects.

Likes novel and unusual suggestions. She/He likes new recommendations and the last one liked by others, Like_ new_recommended_objects

Likes to do innovative things. She/He seeks to give meaning to his/her existence by providing new types and forms of resources and ideas, etc. for example he/she adds a new method of problem resolution.

Works quickly. Likes to finish his/her work as fast as possible, Works_quickly.

May make errors of facts. The fact that he/she prefers to work lonely and quickly on new resources, he/she will commit a lot of errors when using the MCQ or adding tags, Commit_Errors.

Feeling INFP ISFP ESFJ ENFJ Uses values to reach conclusions. Takes with strong conviction decisions based on personal values and the impact they may have on others. On the system, the learner clinging to his/her opinion when sharing his/her personnel views, for example in comments or in tagging, etc.

Works best in harmony. When the learner works with his/her collaborators he/she chosen by him/her self in his/her network, he/she achieves good results in different evaluations.

Tends to be sympathetic and has difficulty providing criticism. The learner does not like receiving criticism or make it.

Feels rewarded when people’s needs are met. Learners expresses a positive emotion when he/she helps someone in solving an exercise for example, he/she expresses a big ability to make a lot of actions, like: send or accept invitations, add learning resources, comments, add tags, etc.

Seeks involvement with people. He/she works in a participative mode, he/she is characterized by adding a lot of collaborators to his/her learning network.

Presents points of agreement first. Before taking action to carry out a learning task, he/she tries to put his/her own touches in advance, like making many comments to explain his/her point of view on the subject.

Is sociable and collaborative. He/she adds a lot of collaborators to his/her learning network and accepts a lot of collaboration invitations.

Uses emotions. He/she needs to be alone, he/she can then engage in a new project. He/She can also write a diary and share it with his/her collaborators.

Logic and a good observer. The learner has high score in logical problem solving, case studies, planned interactive activities and tests to progress.

Thinking ISTP INTP ESTJ ENTJ Uses teacher-directed instructional approaches and peer tutoring. She/He respects teachers rules and instructions, also he/she pays big attention to respect times and agenda.

Can work without harmony. Dislikes expression by others of a strong emotion, he/she likes works lonely.

Is firm-minded and has little trouble giving criticism. He/she does not matter with criticisms, because most of times he/she prefers work lonely. He/She dislikes the invasion of his/her personal space.

Feels rewarded when task is done. He/she makes positive reactions, when he/she finished his/her task, for example solving a problem in the right time.

Presents goals and objectives first. She/He does not like having to peruse many theoretical resources.

Tends to be brief and concise. His/her learning must go to the basics and spend plenty of space to experimentation.

Style	Preferred characteristics	Learner’s behaviour (Actions observed on the system)
Sensory ISTJ ISFJ ESTP ESFP	Uses traditional curriculum and Phase by Phase.	When a learner looks for a learning object xand the system answer him/her that in order to take xhe/she must take ybefore. If he/she accepts and search again for the recommended object y, the system records that he/she prefers Phase-by-Phase method.
Likes using past experiences and standard ways to solve problems.	If he/she is looking in the forum for an old exercises on his/her domain of interest, Search_Old_Exercise.
Enjoys applying what is already known by giving examples and details.	If he/she publishes objects that he/she masters, or makes comment about objects that he/she masters (Makes Comment_on_Mastered_Object) or answering message about objects that he/she masters or tag objects that he/she masters.
Respect of the deadlines and rules.	If he/she returns the solution in the right time and exactly as requested, Solution_Delivery_Time.
Likes suggestions that are straightforward and feasible.	If he/she likes the recommendations provided by the system and finds it appropriate with his/her needs, Likes_Recommendations.
Is inclined to follow an agenda.	If he/she respects the agenda proposed by the teacher to take different objects, for example: read theoretical concept, then read examples, finally, test knowledge with a MCQ, Respect_Agenda.
Likes to do practical things and prefers realistic applications.	He/She prefers solving exercises, quizzes and exams, creates simulations, etc.
Rarely makes errors of facts.	She/He shows a high score in quizzes and exams, High_Score_Evaluation.
He/She is inclined to express by critics.	If he/she uses in his/her comments or in his/her messages critic words, Critic_comments.
Dislike the purely Theoretical courses.	Most of times, he/she looks for examples, exercises, podcast tutorials, diagrams, etc. he/she will be very enthusiastic when he/she receives recommendations of practical objects, Search_for_ practical_objects, likes_ practical_objects.
Communicative and sociable.	He/she Sends and accepts a lot of collaboration requests. He/She has many collaborators, Lot_of_collaborators.
Intuitive INFJ INTJ ENFP ENTP	Focuses on conceptual understanding and the use of self-instructional methods for learning.	He/she likes to deal (publish, tag, comment, like, etc.) with theoretical objects, Likes_ theoretical_objects, he/she prefers finding alone his/her objects; he/she does not pay attention to recommended objects.
Likes solving new and complex problems.	Looks for solving new problems initiated by teachers, and tries to be the first one to find the solution, he/she does not like old problems being processed by other users, Like_new_problems.
Enjoys learning new skills more than using them.	She/He prefers learning knew knowledge, but, he/she does not participate in sharing it with other users, Learning_knew_objects.
Likes novel and unusual suggestions.	She/He likes new recommendations and the last one liked by others, Like_ new_recommended_objects
Likes to do innovative things.	She/He seeks to give meaning to his/her existence by providing new types and forms of resources and ideas, etc. for example he/she adds a new method of problem resolution.
Works quickly.	Likes to finish his/her work as fast as possible, Works_quickly.
May make errors of facts.	The fact that he/she prefers to work lonely and quickly on new resources, he/she will commit a lot of errors when using the MCQ or adding tags, Commit_Errors.
Feeling INFP ISFP ESFJ ENFJ	Uses values to reach conclusions.	Takes with strong conviction decisions based on personal values and the impact they may have on others. On the system, the learner clinging to his/her opinion when sharing his/her personnel views, for example in comments or in tagging, etc.
Works best in harmony.	When the learner works with his/her collaborators he/she chosen by him/her self in his/her network, he/she achieves good results in different evaluations.
Tends to be sympathetic and has difficulty providing criticism.	The learner does not like receiving criticism or make it.
Feels rewarded when people’s needs are met.	Learners expresses a positive emotion when he/she helps someone in solving an exercise for example, he/she expresses a big ability to make a lot of actions, like: send or accept invitations, add learning resources, comments, add tags, etc.
Seeks involvement with people.	He/she works in a participative mode, he/she is characterized by adding a lot of collaborators to his/her learning network.
Presents points of agreement first.	Before taking action to carry out a learning task, he/she tries to put his/her own touches in advance, like making many comments to explain his/her point of view on the subject.
Is sociable and collaborative.	He/she adds a lot of collaborators to his/her learning network and accepts a lot of collaboration invitations.
Uses emotions.	He/she needs to be alone, he/she can then engage in a new project. He/She can also write a diary and share it with his/her collaborators.
Logic and a good observer.	The learner has high score in logical problem solving, case studies, planned interactive activities and tests to progress.
Thinking ISTP INTP ESTJ ENTJ	Uses teacher-directed instructional approaches and peer tutoring.	She/He respects teachers rules and instructions, also he/she pays big attention to respect times and agenda.
Can work without harmony.	Dislikes expression by others of a strong emotion, he/she likes works lonely.
Is firm-minded and has little trouble giving criticism.	He/she does not matter with criticisms, because most of times he/she prefers work lonely. He/She dislikes the invasion of his/her personal space.
Feels rewarded when task is done.	He/she makes positive reactions, when he/she finished his/her task, for example solving a problem in the right time.
Presents goals and objectives first.	She/He does not like having to peruse many theoretical resources.
Tends to be brief and concise.	His/her learning must go to the basics and spend plenty of space to experimentation.

4.6 Phase 5: A BN to estimate learner’s learning style and to validate the personalization accuracy

A Bayesian Network model is used, which allows discovering user’s learning style automatically using the inference mechanisms based on the user’s behaviour analysing, parameters of the structure are obtained from the learner model (in RDF format), which records all his/her actions: his/her primary learning styles, his/her cognitive level towards a learning domain, etc.

To provide the personalization, observing the learner’s behaviour during his/her learning is needed. The system memorizes the learner’s actions and then it uses these data to build the learner’s model. As said before, the learner model comprises the learner’s learning style, her/his cognitive level of a given course, number of published, downloaded, tagged, liked and commented resources, number of sending, accepting and cancelling collaboration invitations, exams or exercises evaluation score, and type of exercises done, etc. Through the observation of the learner’s behaviour, the system can estimate the learner’s learning style using the Bayesian network model. For example, if a learner frequently likes recommendations provided by the system and prefers solving exercises, the system can infer that the learner belongs to the sensory category and has the ISTJ profile.

After receiving recommendations, the learner can provide feed-backs to the system. These feedbacks can be explicit, like: adding comments and tags, or implicit, when the system observes the learner’s behaviour after assisting him/her, like detection his/her pedagogical objectives. The system uses the learner’s feed-back to adjust her/his model for future use.

In order to complete the structure of the Bayesian model we need to set the conditional probability tables for the independent and the dependent nodes. Values of the probabilities are obtained by analysing the learner’s model. As an example, Table 2 shows the probability values obtained for a certain learner.

Table 2
Probability table for some nodes

Resource	%	Collaboration Invitations	%	Messaging	%
Add	60	Send	30	Send	20
Download	20	Accept	50	Receive	40
Like	20	Cancel	20	Answer	40

Once the BN is built, the learner learning style is determined via Bayesian inference. The values of the nodes corresponding to the learning style are inferred through the analysing learner’s behaviour in the system. The learning style of the learner corresponds to the style that has the biggest probability value. The simple example in Table 2, given evidence of the utilization of the collaboration, messaging and actions on resources, there is a possibility to infer whether the learner prefers to focus his/her attention: Extraversion (with his/her collaborators in his/her network) or introversion (by himself/herself).

An important step in the construction of a BN is determining the variables and their states. In our work, variables represent: the learning styles and the different factors analyzed in learners’ behaviour, variables are: pedagogical medias, scores, learning objectives, activities, etc. Arcs between these variables can define different types of relationships: is part of , is a prerequisite of , etc. The variables of the proposed Bayesian network model are defined with two types: Explicit Variables and Implicit Variables, some variables may play the double role, i.e. we can observe a variable directly, at the same time, we can determine it by semantic inference.

4.6.1 Explicit variables

Define variables directly observed by learner’s reaction on the system.

Learner’s behaviour

The fact that there are many parameters that determine the learner’s behaviour and for simplify the learner’s behaviour is represented by three continuous random variables only, $B = {K, S, R}$ .

K: measures the score obtained by learner’s evaluation at the end of each learning sequence. For example, after using a QMC.

S: measures the learner’s sociability degree, i.e. measure the level of participation and exchange of learners, the number of collaborators, sending acceptation and cancellation collaboration requests.

R: measures the degree of compliance with rules and instructions provided by instructors/teachers, for example, the number of given reviews and critics is counted.

(See Section 5.3.3 for how to calculate variables S & R).

4.6.2 Implicit variables

Define variables discovered by analysing users’ actions and activities in the system:

Identifying pedagogical materials

Pedagogical materials describing a pedagogical sequence consist of a set of four variables $M = {M_{1}, M_{2}, M_{3,} M_{4}}$ , where:

$M_{1}$ : represents the learning domain treated with pedagogical materials (e.g. programming or mathematics, etc.).

$M_{2}$ : represents the media used with the pedagogical material. It can take 4 values $M_{1} \in {T, I, S, V}$ (T) Text, (I) Image, (S) Sound, (V) Video.

$M_{3}$ : describes learning entities, which takes 8 values $M_{3} \in {I, S, D, C, U, E, X, T}$ , (T) Introduction, (S) Summary, (D) Definition (C) Clarification, (U) illUstration, (E) Example, (X) eXercise, (T) Tutorial.

$M_{4}$ : represents the type of evaluation associated with the pedagogical material, it can take 7 values, $M_{4} \in {Q, T, G, S, E, Z, X}$ , (Q) MCQ, (T) byTag, (G) GapText, (S) Summary, (E) ExerciseSolving, (Z) quiZ, (X) eXam.

Identifying pedagogical activity

Pedagogical activity is described by a set of two variables $A = {A_{1}, A_{2}}$ , where:

$A_{1}$ : is the learning objective. Based on the Bloom taxonomy [4], this variable can take 6 values, $A_{1} \in {K, U, Ap, An, S, E}$ . This variable describes a pedagogical activity that can target 6 levels of learning objectives: (K) Knowledge, (U) Understanding, (Ap) Application, (An) Analysis, (S) Synthesis and (E) Evaluation.

$A_{2}$ : indicates the pedagogical strategy. It may take 7 values $A_{2} \in {R, W, H, L, Do Ds, T}$ , (R) ByReading, (W) ByWriting, (H) ByHearing, (L) ByLooking, (Do) ByDoing, (Ds) ByDiscussing and (T) ByThinking.

Variables are defined by using the inference mechanisms of the semantic web and they are used also to define other explicit variables. For this end, a set of semantic rules that allow discovering these variables is proposed. In the following, we present some semantic rules.

Discovering Pedagogical Entities

wantToLearn(U, Definition) & hasRequirement(Definition, Introduction) → wantToLearn (U, Introduction)

If a student U wants to take a definition and the latter as described in the ontology has requirement an Introduction, it means that, the introduction should be presented first to the student. Therefore, the variable Pedagogical Entity will be highlighted with the new belief “Introduction”.

Fig. 10.

Overall view of the BN model.

Discovering Learning Objectives

learned(U, Definition) & wantToLearn(U, Exercise) → hasObjective (U, Understanding)

If a student U has taken a definition on a certain learning domain and just after he/she looks for an exercise, it means that the student wants to understand that domain. Therefore, the variable Pedagogical Activity will be highlighted with the new belief “Understanding”.

wantToLearn(U, Example) | wantToLearn(U, Exercise) → hasObjective (U, Application)

If a student U has taken directly an Example on a certain learning domain or he/she has taken directly an Exercise, it means that the student looks to apply his/her knowledge. Therefore, the variable Pedagogical Activity will be highlighted with the new belief “Application”.

Discovering Learning Strategies

wantToLearn(U, Domain) & subClassOf (Domain, Practical) → hasStrategy (U, ByDoing)

If a student U wants to learn a certain domain (e.g. a programming language) and the latter is defined in the ontology as a sub class of the practical domains, it means that, the student wants to do programs. Therefore, the variable Pedagogical Strategy will be highlighted with the new belief “ByDoing”.

Discovering Rules Respecting

hasRequirement(Definition, Introduction) & learned(U, Introduction) & learned(U, Definition) → respctRules(U, “ $Y e s$ ”)

If a student U accepts the recommendation of taking the Introduction before taking the Definition the fact that the latter has a requirement an Introduction, it means that the student U is respecting rules. Therefore, the explicit variable Rules Compliance will be highlighted with the new belief “Yes”.

In this work, the factors analyzed to determine the perceptions of a learner are: whether the learner likes recommendations, he/she follows such pedagogical sequencing; if he/she adds resources and how many resources he/she publishes; the type of learning material the learner prefers (theoretical or practical); the number of examples of a given topic the learner reads; the number of exercises a learner does on a given topic; the number of errors he/she commits; the number of collaborators he/she has, number of sending, accepting or cancelling collaboration invitations; number of comments he/she lets, number of tags, etc. Hence, for example, a learner, who does not commit a lot of errors and follows instructions to accumulate the knowledge and likes explications, has an ISTJ profile. But, if he/she prefers learning globally without respecting the pedagogical sequencing and he/she likes practical materials, then he/she has an ESTP profile.

Fig. 11.

BN structure after student’s selections.

To completely define the structure of Bayesian network, dependencies (or arcs) between nodes must be established. As shown in Fig. 10, dependencies allow expressing that:

The learning domain influences the choice of pedagogical entity.

The pedagogical entity influences the choice of the learning objective and the evaluation method.

The learning objective influences the choice of pedagogical strategy.

The pedagogical strategy influences the choice of pedagogical media.

The combination of pedagogical entity and the evaluation method influence the obtained score.

The combination of the obtained score, the degree of sociability, the compliance to rules degree and the pedagogical media influences the learning style.

Firstly, equal values for the probabilities of independent nodes are given. Then, the values changed and updated as the system collects more information about the learner’s behaviour and knowledge. Therefore, the Bayesian model still updated the fact that new information about learners’ behaviour is obtained, until the probability values show a very small modification, values obtained at this level represent the estimation of the learner’s style.

Illustrative example

Assuming a student logs into the system seeking to learn “Algorithms”, as defined in the ontology, “Algorithms” is a theoretical learning domain. As shown in Fig. 11, the variable LearningDomain highlighted with the belief “Theoretical”, the system using inferring rules will recommend to the student a set of pedagogical entities that correspond to theoretical resources ordered according to the pedagogical model. For instance, it proposes to student respectively: “introduction, definition, clarification, illustration and summary” about algorithms, the student then selects one of those entities, the system records if he/she respects the sequencing order or not, in order to calculate later her/his degree of compliance to rules. Suppose, that the student selects a “definition”, the system automatically highlights the variable Pedagogical_Entity with the belief “definition” due to the explicit action of the learner; and highlights the variable Learning_Objective with the belief “Understanding” and the variable Pedagogical_Strategy by the belief “byReading, ByHearing or byLooking” following the implicit variables discovered using the inferring rules presented above. Thereafter, the system provides student with pedagogical Medias “text, image and video”. Suppose that the student selects the “definition” with the form “text”, so the variable Pedagogical_Media highlighted with the belief “text”. After a period of time, the students will be asked to perform a test to detect her/his score about algorithms; the system automatically highlights the variable Evaluation Method with “QMC or GapText”. Let’s suppose now that the student obtained a low score, he/she respects rules and he/she has a low Sociability Degree. As a result, the system estimates that the student has an “INFJ” style.

4.7 Phase 6: Recommendation of content and learning strategies

After obtaining the final decision about the learner learning style provided by the BN model, the system adjusts the recommendation process, by providing the learner the appropriate content with the most preferred media according to his/her style, best assistance according to his personality type and best collaborators. In Table 3, some recommendations according to learning styles are presented.

Table 3
Examples of recommendations according to learning styles

Style	Recommendations	Preferred Medias
Sensory	It is better to take document Xbefore the document Y.	Chat Messages Animation Communities Pictures Videos Simulation
	It is better to look for introduction about this document first.
	It will be better to solve more exercises.
	Try to read examples as much as you can.
	Check these critics about your participation.
	It is better to add this user to your network.
Intuitive	It is better to take theoretical topics first.	Lectures Recorded events Simulations Tutorial systems Written text
	It is better to take all your time in order to complete your tasks.
	We look forward to receiving your participation about this topic.



Feeling	It is better to cooperate with your collaborators.	Communities Pictures Recorded events Simulations Forums
	It is better to add users of your interest to your network.
	We appreciate so much helping this learner in solving his/her exercises.
	Comment your collaborators’ participation.
Thinking	It is better to start reading chapter’s theory or idea before related examples.	Internet- research Simulation Blogs Wikis
	It is better to look for logical problem solving.
	Read summaries first.

5 Experiment & implementation

An experimental study is in progress within a set of learners and teachers of the science department of Guelma University (Algeria). Actually, the experiment has been divided into two stages. The first stage has been done using the social and the semantic web aspects to enhance the learning personalization, i.e. to see if they will positively affect or not the personalization process quality. A prototype of our system called SoLearn (in alpha testing) is accessible from this URL (for a real demonstration): http://solearn.labstic.com.

Initial results of the first stage have been very encouraging. On the one hand, learners have appreciated the prototype. In addition, the majority of them have appreciated the idea of implementing a learning system as a social network, especially: the possibility to gather in social networks, the accurate recommendations of collaborators and learning objects, the semantic search, etc. On the other hand, teachers have also appreciated so much the system, they agreed that the fact to infer new knowledge and recommend learning objects related to these new facts and related to similar users will certainly improve learning personalization. See [22] for more details about the implementation of SoLearn and the experiment’s first stage results.

5.1 Second phase purpose and objectives

The purpose of the second phase of the study is to determine if there is relationship between learner’s behaviour (analysing by SSW and BN techniques) on the system, namely: score obtained, helping requests, making errors, working quickly, adding resources, commenting, tagging, etc. and the learner’s learning style. Thereafter, to evaluate the precision of the approach comparing the learning style detected by the model against the learning style obtained with the MBTI test.

5.2 Methodology

We are conducting the second phase of the experiment study at the computer science department of Guelma University. The participants are fifty four ( $n = 54$ ) learners (we used a Simple Random Sampling where we have chosen randomly 54 learners from 100 learners (License and Master degree promotion). We consider that the sample is representative because a group of 54 students chosen out of a hat which represents 100 students in global, where each student has an equal chance of being chosen, with a confidence level of 95% and a confidence interval of 9. Topic taken was general concepts on computing. We conducting the experiment as follows: we asked learners to use the system and through analysing their behaviour and applying the proposed BN model, the system tries to find the real learning style of each one. Thereafter, we invite learners to finish the MBTI test (carefully) in order to compare both results.

Table 4
Learner’s actions on the system

Total learner = 54

Action/Action on Number of times Percentage of Learner using the action Action Detail

Comment 123 27[3] 50% of learners comment 3 times. First_Comment [50]

17[2] 32% of learners comment twice.

8[1] 13% of learners one time. On_Comment [73]

2[0] 5% never used comments.

Message 70 42[⩾1] 77% of learners used messages more than one time. Send [47]

12[0] 23% never used messages. Receive [47]

Answer [23]

Collaboration Invitation 57 38[1] 70% of learners used collaboration invitation. Send [38]

16[0] 30% of learners never used collaboration invitation. Accept [11]

Cancel [8]

Resource 147 27[4] 50% of learners treat with resources 4 times. Add [40]

Download [30]

16[2] 30% of learners treat with resources twice. Comment_On [33]

7[1] 12% of learners treat with resources one time. Search [29]

4[0] 8% of learners never treat with resources. Tag [15]

Working Time 54 19[Yes] 35% of learners work quickly. Upload_Solution_Quickly [19]

35[No] 65% of learners work slowly. Ask_for_help [25]

Search_old_Res [10]

Commit Errors 54 36[No] 66% of learners have not committed errors. Correct_Tag [20]

Correct_QMC_Answer [16]

18[Yes] 33% of learners have committed errors. Wrong_Tag [12]

Wrong_QMC_Answer [6]

Total learner = 54
Comment	123	27[3]	50% of learners comment 3 times.	First_Comment [50]
	17[2]	32% of learners comment twice.

	8[1]	13% of learners one time.	On_Comment [73]
	2[0]	5% never used comments.
Message	70	42[⩾1]	77% of learners used messages more than one time.	Send [47]
	12[0]	23% never used messages.	Receive [47]
			Answer [23]
Collaboration Invitation	57	38[1]	70% of learners used collaboration invitation.	Send [38]
	16[0]	30% of learners never used collaboration invitation.	Accept [11]
			Cancel [8]
Resource	147	27[4]	50% of learners treat with resources 4 times.	Add [40]
			Download [30]

	16[2]	30% of learners treat with resources twice.	Comment_On [33]
	7[1]	12% of learners treat with resources one time.	Search [29]

	4[0]	8% of learners never treat with resources.	Tag [15]
Working Time	54	19[Yes]	35% of learners work quickly.	Upload_Solution_Quickly [19]

	35[No]	65% of learners work slowly.	Ask_for_help [25]
			Search_old_Res [10]
Commit Errors	54	36[No]	66% of learners have not committed errors.	Correct_Tag [20]
			Correct_QMC_Answer [16]

	18[Yes]	33% of learners have committed errors.	Wrong_Tag [12]
			Wrong_QMC_Answer [6]

5.3 Procedures

In order to realize the experiment’s objective, we are working as follows (after 6 months of system using):

Collecting users’ preferences (preferred objective, strategy, media, and learning entities).

Classifying users’ actions (Commenting, messaging, collaboration invitation, making errors, etc.).

Calculating the “Sociability Degree” and “Compliance to Rules Degree” for each learner.

Evaluation the learner’s knowledge and determining his/her score towards a learning domain.

Applying results to the proposed BN model and identifying the learning style of each learner.

Comparing the results of the proposed approach and the results obtained by the MBTI test.

5.3.1 Collecting users’ actions

As showed before in Section 4.3.4, all learner’s actions are stored in his/her model (RDF file), where we can find his/her level of topic/course mastery, his/her preferences (collaborators, preferred format of resource, etc.), his/her learning goals, learning history and learning style.

5.3.2 Collecting and classifying users’ actions

In order to realize this step, we need to classify users’ actions and to calculate the rate of each action, namely when the learner uses the commenting tool, messaging tool, collaboration tool, resources manager, if he/she works quickly and if he/she commits errors, etc. Table 4, shows some learners’ actions when using the system (until now), for instance, it describes how many times learners have used the commenting tool, where 50% of them commented 3 times, 32% used comments twice, 13% commented one time and 5% never used the commenting tool, it indicates also that almost 40% of learners added comments at first and 60% of them commented on Users’ comments, etc. Figures 12, 13, 14, 15 and 16 show ratios of learners’ actions on the SoLearn system.

5.3.3 Sociability & compliance to rules degrees calculating

To determine the sociability degree of user $U_{i}$ , we need to calculate his/her social interactions rate against other users’ actions, i.e. how many times the user performs social actions against the total number of actions performed by all other users. We use a SPARQL query to count the student’s number of actions, for instance, we calculate number of comments C of such a user $U_{i}$ as follows:

SELECT DISTINCT ?comment

WHERE {?x ns:putComment ?comment}

For this end, we defined four variables: Collaboration_Invitation_Sending, Collaboration_Invitation_ Accepting, Likes and Commenting. As shown in Eq. (2). We defined four parameters “action importance coefficient”, which determine the importance of the social action, for instance, a user who sends collaboration invitations to other users should supposed to be more social person against another one who accepts invitations only. Therefore, we have given:

The coefficient 0.4 to the $CIS$ (Collaboration Invitation Sending action).

The coefficient 0.3 to the $CIA$ (Collaboration Invitation Accepting action).

The coefficient 0.2 to the L (Likes action).

The coefficient 0.1 to the C (Commenting action).

\begin{array}{l} DS (U_{i}) \\ = {{CIS}_{i} * 0.4 + {CIA}_{i} * 0.3 + L_{i} * 0.2 \\ + C_{i} * 0.1} \times {\sum_{i = 1}^{n} ({CIS}_{i} * 0.4 + {CIA}_{i} * 0.3 \\ {+ L_{i} * 0.2 + C_{i} * 0.1)}}^{- 1} & (2) \end{array}

The Degree_of_sociability $DS$ of a user $U_{i}$ , will have a value of: $DS (U_{i}) \in [0, .., 1]$ , where:

$0 ⩽ DS (U_{i}) < 0.2 \to$ Low degree

$0.2 ⩽ DS (U_{i}) < 0.6 \to$ Medium degree

$0.6 ⩽ DS (U_{i}) ⩽ 1 \to$ High degree

Fig. 12.

Overall users’ actions.

Fig. 13.

Ratios of using comments.

Fig. 14.

Ratios of using messages.

Fig. 15.

Ratios of using collaboration invitations.

Fig. 16.

Ratios of actions on resources.

To determine the compliance to rules degree of user $U_{i}$ , we need to calculate his/her rate of acceptance and respect of different suggestions and recommendation, for this end, we have defined three variables: Accept_Learning_Strategy (e.g. Global or sequencial), Respect_of_Time (e.g. Time to answer evaluation) and Like_Recommendations. As shown in Eq. (3). We defined three parameters “compliance importance coefficient”, which determine the importance of the action. Therefore, we have given:

The coefficient 0.5 to the $AL$ (Accept Learning Strategy action).

The coefficient 0.3 to the $RT$ (Respect of Time action).

The coefficient 0.2 to the $LR$ (Like System Recommendations).

\begin{array}{l} DRC (U_{i}) \\ = \frac{{AL}_{i} * 0.5 + {RT}_{i} * 0.3 + {LR}_{i} * 0.2}{\sum_{i = 1}^{n} ({AL}_{i} * 0.5 + {RT}_{i} * 0.3 + {LR}_{i} * 0.2)} & (3) \end{array}

The Degree_of_Rules_Compliance $DRC$ of a user $U_{i}$ , will have a value of: $DRC (U_{i}) \in [0, .., 1]$ , where:

$0 ⩽ DS (U_{i}) < 0.5 \to$ NO

$0.5 ⩽ DS (U_{i}) ⩽ 1 \to$ Yes

Fig. 17.

The proposed BN model.

5.3.4 Evaluation the learner’s knowledge

To determine the learner’s knowledge towards a learning object, he/she will be subjected to one of the evaluation method provided by the system according to his primary behaviour (a QMC for example), the score of the test will be recorded in his profile (learner model) with this possible values (High, Medium or Low) as presented in the following example:

5.3.5 Applying results to the proposed BN model

In order to estimate the learner’s learning style through analysing his/her behaviour, we need to subject all results of previous steps to the proposed Bayesian Network, the model then, will automatically detect the learning style. As shown in Fig. 17, the score, the pedagogical media, the rules compliance degree and the sociability degree influence the learning style. In case a belief in the probabilities on those parameters is highlighted the learning style will be changed.

Firstly, as shown in Figs 17 and 18, equal values for the probabilities of independent nodes are given. Then, the values changed and updated as the system collects more information about the learner’s behaviour and knowledge. Therefore, the Bayesian model still updated the fact that new information about the learner’s behaviour is obtained, until the probability values show a very small modification. The values obtained at this level represent the learner’s style.

5.3.6 Comparing the results

After applying the proposed approach and after inviting learners to perform the MBTI test on line, we have obtained the results summarized in Table 5. As presented, our approach of learning style identification by analysing learners’ behaviour combined with use of semantic social web techniques allowed us on one hand, to detect four correct categories of learning styles in accordance with the MBTI test, (with some errors in such categories). On the other hand, our approach gives two wrong categories with respect to the MBTI test results.

Fig. 18.

CPT of the learning style node.

We attribute these errors to the fact that either the values entered in the CPT tables of variables were not appropriate or learners’ answers when they have performed the MBTI test are just chosen arbitrary. As many actions are carried out with the continuous use of the system, and as much as possible adjusting the CPT tables, we believe that the results will be more accurate.

Table 5

Comparing results

Figure 19, presents some screens of the SoLearn system focusing on its various features.

Fig. 19.

Main pages of the system.

6 Conclusion

This work addresses the learning personalization. We have gone beyond classical methods of personalization that adjust the content level of knowledge to the learner and we have proposed a new approach that aims to satisfy real needs of learners to provide them the best pedagogical materials, activities and assistances according to their knowledge, preferences and learning styles. We presented a new method based on an automatic estimation of student’s learning styles with respect to the MBTI model. Therefore, we answered the following questions: (1) How to represent resources and users to incorporate richness knowledge to personalization? For this purpose, we have modelled users, learning domains and the pedagogical approach with an ontology that combines the users, domains, pedagogical information, and properties of resources. (2) How to select learning activities to fit the style of the learner? To answer this question, a personalization approach is developed based on the use of a Bayesian Network model that allows deciding whether the provided content is appropriate or not. First experimentation has shown good results. In the future work, we will carry on extensive tests to firmly validate the proposed approach and the efficiency of the method. We project also to apply the approach on other learning style’s models instead of the MBTI model.

References

[1]

Anderson (ed.), The Theory and Practice of Online Learning, Athabasca University Press, 2008.

[2]

Arroyo and

Woolf, Inferring learning and attitudes from a Bayesian network of log file data, in: AIED 05, Twelfth International Conference on Artificial Intelligence in Education, 2005, pp. 33–40.

[3]

Bishop-Clark and

Wheeler, The Myers-Briggs personality type and its relationship to computer programming, Journal of Research on Computing in Education26(3) (1994), 358–370.

[4]

B.S.

Bloom, Taxonomy of Educational Objectives: The Classification of Educational Goals,

McKay , ed., 1956.

[5]

Boroch,

Fillpot,

Hope,

Johnstone,

Mery,

Smith and

Serban, Basic skills as a foundation for student success, Journal of Applied Research in the Community College15(1) (2007), 81–85.

[6]

Bruen, A development framework for re-useable learning resources for different learning styles and requirements, in: World Conference on E-Learning in Corporate, Government, Healthcare, and Higher Education, 2002, pp. 1238–1241.

[7]

Brusilovsky and

Peylo, Adaptive and intelligent web-based educational systems, International Journal of Artificial Intelligence in Education13(2) (2003), 159–172.

[8]

Carver,

Howard and

Lavelle, Enhancing student learning by incorporating learning styles into adaptive hypermedia, in: ED-MEDIA96, World Conference on Educational Multimedia Hypermedia, 1996, pp. 118–123.

[9]

H.J.

Cha,

Y.S.

Kim,

S.H.

Park,

T.B.

Yoon,

Y.M.

Jung and

J.-H.

Lee, Learning style diagnosis based on user interface behaviour for the customization of learning interfaces in an intelligent tutoring system, in: Proc. of the Eighth International Conference on Intelligent Tutoring Systems, Lecture Notes inComputer Science, Springer-Verlag, Berlin, Heidelberg, 2006, pp. 513–524.

10.

[10]

Conati,

Gertner and

VanLehn, Using Bayesian networks to manage uncertainty in student modelling, Journal of User-Modelling and User-Adapted Interaction12(4) (2002), 371–417.

11.

[11]

M.P.

Cuéllar,

Delgado and

M.D.C.P.

Jiménez, Improving learning management through semantic web and social networks in e-learning environments, Expert Systems with Applications38(4) (2011), 4181–4189.

12.

[12]Didaskon, http://didaskon.sourceforge.net/.

13.

[13]

Dzbor,

Stutt,

Motta and

Collins, Representations for semantic learning webs: semantic web technology in learning support, Journal of Computer Assisted Learning23(1) (2007), 69–82.

14.

[14]

ElBachari,

Abdelwahed and

El Adnani, Design of an adaptive e-learning model based on learner’s personality, Ubiquitous Computing and Communication Journal5(3) (2010), 1–8.

15.

[15]Ensemble project: semantic technologies for the enhancement of case based learning, http://www.ensemble.ac.uk/wp/resources.

16.

[16]

R.M.

Felder and

L.K.

Silverman, Learning and teaching styles in engineering education, Engineering Education78(7) (1988), 674–681.

17.

[17]

Gamboa and

Fred, Designing intelligent tutoringsystems: a Bayesian approach, in: Enterprise Information Systems III, Springer Verlag, New York, 2002, pp. 146–152.

18.

[18]

García,

Amandi,

Schiaffino and

Campo, Evaluating Bayesian networks’ precision for detecting students’ learning styles, Computers & Education49(3) (2007), 794–808, Elsevier.

19.

[19]

Gertner and

VanLehn, Andes: a coached problem solving environment for physics, in: Intelligent Tutoring Systems: Fifth International Conference, ITS, 2000, pp. 131–142.

20.

[20]

Gilbert and

Ham, Adapting instruction in search of a significant difference, Journal of Network and Computer Applications22(3) (1999), 149–160.

21.

[21]

Gruber, Collective knowledge systems: where the Social Web meets the Semantic Web, Journal of Web Semantics6(1) (2008), 4–13.

22.

[22]

Halimi,

Seridi-Bouchelaghem and

Faron-Zucker, An enhanced personal learning environment using social semantic web technologies, Interactive Learning Environments22(2) (2014), 165–187.

23.

[23]

Heckerman,

Mamdani and

M.P.

Wellman, Real-world applications of Bayesian networks, Communications of the ACM38(3) (1995), 24–26.

24.

[24]

Heflin and

J.A.

Hendler, Portrait of the Semantic Web in action, IEEE Intelligent Systems16(2) (2001), 54–59.

25.

[25]

Henze, Personalization services for e-learning in the semantic web, in: Proc. of Second International Workshop on Adaptive Systems for Web-Based Education: Tools and Reusability, University of Amsterdam, The Netherlands, 2005.

26.

[26]

Honey and

Mumford, The Manual of Learning Styles,

Honey and

Mumford, eds, 1992.

27.

[27]

Jensen, An Introduction to Bayesian Networks, UCL Press, London, 1996, p. 210.

28.

[28]

Jeremic,

Jovanovic and

Gasevic, Personal learning environments on the social semantic web, 2011, Available at: http://www.semantic-web-journal.net/content/personal-learning-environments-social-semantic-web.

29.

[29]

A.Y.

Kolb and

D.A.

Kolb, Learning styles and learning spaces: enhancing experiential learning in higher education, Academy of Management Learning & Education4(2) (2005), 193–212.

30.

[30]

Koshmann, CSCL: Theory and Practice of an Emerging Paradigm, Routledge, 1996.

31.

[31]

Logan and

Thomas, Learning style in distance educationstudents learning to program, in: Fourteenth Workshop of the Psychology of Programming Interest Group, Brunel University, 2002, pp. 29–44.

32.

[32]

Naïm,

P.H.

Wuillemin,

Leray,

Pourret and

Becker, Réseaux Bayésiens, Eds Eyrolles, 2011.

33.

[33]

Nilsson,

Palmr and

Naeve, Semantic web metadatafor e-learning – some architectural guidelines, in: Eleventh Conference on the World Wide Web (WWW’02), Honolulu, 2002, pp. 7–11.

34.

[34]

Paredes and

Rodriguez, Considering learning styles inadaptive web-based education, in: Sixth World Multiconference on Systemics, Cybernetics and Informatics, 2002, pp. 481–485.

35.

[35]

Papanikolaou,

Grigoriadou,

Knornilakis and

Magoulas, Personalizing the interaction in a webbased educational hypermedia system: the case of INSPIRE, User Modelling and User-Adapted Interaction13 (2003), 213– 267.

36.

[36]

Q.D.

Pham and

A.M.

Florea, A method for detection of learning styles in learning management systems, UPB Scientific Bulletin, Series C75(4) (2013), 3–12.

37.

[37]

Schiaffino and

Amandi, Intelligent user profiling, in: Artificial Intelligence, An International Perspective, Springer, Berlin, Heidelberg, 2009, pp. 193–216.

38.

[38]

Shang,

Shi and

Chen, An intelligent distributed environment for active learning, Journal on Educational Resources in Computing (JERIC)1(2) (2001), 4.

39.

[39]

R.E.

Snow, Aptitude-treatment interaction as a framework for research on individual differences in learning, Journal of Consulting and Clinical Psychology59(2) (1991), 205.

40.

[40]

Soller, Supporting social interaction in an intelligent collaborative learning system, International Journal of Artificial Intelligence in Education12 (2001), 40–62.

41.

[41]

Torniai,

Jovanovic,

Gasevic,

Bateman and

Hatala, E-learning meets the social semantic web, in: Proc. of the Eighth IEEE International Conference on Advanced Learning Technologies, 2008, pp. 389–393.

42.

[42]

Xenos, Prediction and assessment of student behaviour in open and distance education in computers using Bayesian networks, Computers and Education43 (2004), 345–359.

Total learner = 54

Action/Action on	Number of times		Percentage of Learner using the action	Action Detail
Comment	123	27[3]	50% of learners comment 3 times.	First_Comment [50]
		17[2]	32% of learners comment twice.	First_Comment [50]

		8[1]	13% of learners one time.	On_Comment [73]
		2[0]	5% never used comments.	On_Comment [73]
Message	70	42[⩾1]	77% of learners used messages more than one time.	Send [47]
		12[0]	23% never used messages.	Receive [47]
				Answer [23]
Collaboration Invitation	57	38[1]	70% of learners used collaboration invitation.	Send [38]
		16[0]	30% of learners never used collaboration invitation.	Accept [11]
				Cancel [8]
Resource	147	27[4]	50% of learners treat with resources 4 times.	Add [40]
				Download [30]

		16[2]	30% of learners treat with resources twice.	Comment_On [33]
		7[1]	12% of learners treat with resources one time.	Search [29]

		4[0]	8% of learners never treat with resources.	Tag [15]
Working Time	54	19[Yes]	35% of learners work quickly.	Upload_Solution_Quickly [19]

		35[No]	65% of learners work slowly.	Ask_for_help [25]
				Search_old_Res [10]
Commit Errors	54	36[No]	66% of learners have not committed errors.	Correct_Tag [20]
				Correct_QMC_Answer [16]

		18[Yes]	33% of learners have committed errors.	Wrong_Tag [12]
				Wrong_QMC_Answer [6]