FLICM clustering with matrix factorization based course recommendation in an E-learning platform

Abstract

Technology-enabled learning has progressively grown for research areas with wide application of information and communication technologies for numerous standard-compliant Learning and Open Educational Resources. This provides formidable support to users for the selection of courses when they want to develop the course with available learning materials. But selecting a course via searching learning objects is an inherently complex operation having various repositories. In an E-learning Platform, many complexities arise due to various software tools and specification formats that hinder the success of the course. In this paper, many obstacles in the E-learning platform are eradicated by utilizing Fuzzy Local Information C-Means (FLICM) clustering with matrix factorization for the selection of courses. The dataset utilized in this work is E-Khool review data, from which an agglomerative matrix is generated. Here, the agglomerative matrix consists of the learner series matrix and course series matrix along with their binary matrix. After this process, course grouping is carried out by FLICM clustering with matrix factorization. Moreover, group course bilevel matching, followed by relevant learner retrieval and group user is done by Minkowski and Chebyshev distance. From this learner’s preferred course is retrieved and then a recommendation using matrix factorization is carried out. Finally, the course is recommended for the user based on maximum rating. Furthermore, the performance of developed FLICM_matrix factorization is achieved by performance metrics, like precision, recall, and f-measure with values 0.915, 0.850, and 0.882, accordingly.

Keywords

Fuzzy Local Information C-Means Minkowski and Chebyshev distance matrix factorization E-Khool review data bilevel matching

1. Introduction

The development of technology via the internet increased the available amounts of many data types and has resulted in information overload issues [11,12]. The recommender system is probably utilized in many applications of the internet for helping users to find their favorite services or items. This helps in overcoming many options or information related to selecting a course [18]. Nowadays, during era of preparation for post-COVID disease, the global education market accelerated its transition from offline education to online education [17]. As the market of online education [25] proliferates, online course websites, like edX, K-MOOC, and Coursera are turning widespread and many increasing subscribers are noticed. Hence, the range of course-related information that is available to users is rapidly increasing. Many studies show users to face difficult situational issues, because of the huge information available when choosing courses on the online educational website. This process of selection is time-consuming and challenging while choosing the preferred course on those websites. Valuable information provided by the course recommender system includes related resource information like users’ interests and job opportunities. Hence, course recommender systems on online education [26] websites use diverse resources for matching users’ individual goals, knowledge structure, and interests. The condition of user in choosing a personalized course is highly complex and challenging [16]. E-learning is defined by experiences on learning, and instructional content-enabled electronic technology. This is particularly followed in computer networks and standalone computers, which are gradually diffused with community settings. Course review process is needed for finding the quality of individual courses for establishing areas in every course along with its merits [2].

Most general recommendation systems have their base recommendations on ratings of user for particular items in accordance with ratings. For estimating usefulness of not rated items to be known for particular user, many filtering methods are utilized to find ratings, like content-enabled filtering [20], hybrid filtering [4], and collaborative filtering [8]. Content-enabled filtering method to recommend items to user depending on past preferences and past details. Moreover, collaborative filtering methods give base recommendations of course for given user on past preferences of another user. In first stage of recommendation process, both content enabled filtering and collaborative filtering methods suffer from cold start issue. For latter, successful recommendations consider availability of critical mass of user. For leveraging benefits of both methods, hybrid methods, that combine collaborative filtering and content enabled, is suggested by users. Another method that enhances recommendation is data [10] mining enabled recommendation that is useful for users. More than using user ratings, this method learns behavioural models or rules from stored data’s [14,19] history and makes recommendations based on learned rules. To the best of knowledge, only few course recommendation methods are developed. Student Course Recommender (SCR) network learns from information stored about students who have utilized system. The system is unable to predict recommendations for students who have not considered any courses at University. System utilizes the experience of past students and their course histories of them as basis for advising course. For determining similarities among course histories, system utilizes metric commonly utilized in bio-informatics named edit distance [3].

In process of E-learning, course recommendation system recommends optimal courses where students are participating on the development of knowledge [9]. Numerous studies show that users face many difficulties, while choosing courses on online educational website,due to massive or vast quantity of data. Course selection process is highly time consuming and purely challenging task. Important information given by course recommendation method includes relevant resource information, like users’ job opportunities, time schedule and interests on particular course. Therefore, course recommendation methods on online education websites exploit many of resources for matching objectives, interests of individual users, and structure of knowledge in course [28]. Moreover, selecting proper course in studies is very important, as futures of user are more dependent on those decisions. Course recommendation method is necessary for assisting student or academics in the selection of appropriate courses. It gives solution for helping student in receiving appropriate target results. On other hand, the process of selection of personalized course is highly challenging and intricate for user [16]. Nowadays, the recommendation system is becoming more popular in both academia and industry as these lows down information overloading issue. By considering many applications, recommendation systems make effort for evaluating targeted ratings of user on unrated items and hence recommend items with highly predicted ratings to minimize user attempts and accordingly improvise user contentment. Also, data scarcity is most commonly found the problem in recommendation systems where users have ratings on less number of items with huge reviews that makes much difficult for learning efficient recommendation model system [2].

This research work concentrates on course recommendations based on FLICM clustering-enabled matrix factorization. Here, an agglomerative matrix with learner series and course series is generated from E-khool data. From this binary learner series matrix and binary course series matrix is evaluated. Then, course grouping is done using FLICM, followed by bilevel matching by Minkowski and Chebyshev distance. First, bilevel matching is done between group course and binary query, and second bilevel matching is done between binary query and binary best group. After that preferred course is retrieved and the recommendation is done using matrix factorization, At last, final course is recommended based on the maximum ratings of the course.

Important contributions followed in this work are explained as below,

FLICM clustering_matrix factorization: Course recommendation in the E-learning platform is done by FLICM clustering, through which the grouping of courses is done. Here, courses are recommended by the matrix factorization method and the final course is suggested by considering maximum ratings.

The remaining process of this paper is allotted by following details, like literature feedback regarding course recommendations is enumerated in Section 2 along with techniques challenges. Section 3 presents details of FLICM_matrix factorization for course grouping and recommendation. Section 4 illustrates results and discussions of FLICM_matrix factorization method, and finally, Section 5 completes the paper with a conclusion.

2. Motivation

The E-learning platform is highly developing for course recommendation as it has many tools, information, and resources regarding courses and their full details. They also enhance education with high management ability and skills, but they develop some challenges regarding lack of user activity, slow process of introducing new items, and scalability. These challenges are eradicated by utilizing a proper clustering algorithm namely FLICM and recommending appropriate courses by matrix factorization, as followed in this paper.

2.1. Literature review

Santosh Kumar Banbhrani et al. [2] used Taylor Chimp Optimization Algorithm (Taylor ChOA)-based Random Multimodal Deep Learning (RMDL) for course recommendation. This method resolved issues of information overload in the online educational field improved performance, however processing time of this method was too high. Gao, M., et al. [7] applied Convolutional Neural Network (CNN) with negative sequence mining for online course recommendation. This method efficiently predicted that, which courses are most commonly to be misselected by users for improving overall performance. But this scheme was not suitable for real-time applications. Ren, X.,et al. [22] established Long- and Short-Term Memory networks (LSTM) and Attention mechanisms for multimodal course recommendation. Although, this method effectively reduced the time for users in choosing courses and realized Personalized Course recommendation service, this method failed in performing sentiment analysis for evaluation of the relationship between learners’ ratings and course preferences. Ali, S., et al. [1] presented E-learning Recommendation Architecture (ELRA) in a virtualized environment for enabling recommendation of course. This method utilized two distinct features that reduced contact and confusion among consumers and the high-quality education team. However, memory consumption in this method was high.

Qinglong Li, and Jaekyeong Kim [16] evaluated the Deep learning-based Course Recommender system (DECOR) for robust recommendation performance. Here, personalized course recommendations and many features integrations were included so as to provide suitable recommendations for users’ individual needs. However, there are not enough public datasets to train domain recommender systems. Muhammad Sajid Rafiq et al. [21] found Intelligent query optimization and course recommendations for enhancing online lectures. Although, this scheme improved the survivability and accessibility of intelligent educational knowledge, but failed in using genetic algorithms and swarm intelligence to improvise course accuracy of online query optimization and course recommendation. Chao Wang et al. [27] introduced Demand-aware Collaborative Bayesian Variational Network (DCBVN) for personalized employee training course recommendation. This method proved to be robust against cold-start and sparse scenarios. But, failed in recommending the most appropriate training courses for employees. Shahbazi, Z. and Byun, Y.C., [24] designed Knowledge Discovery and Machine Learning Approaches for an agent-based recommendation. Here, user requests without any interruptions or inconvenience are preferred but failed to improvise its performance with other algorithms in focusing on gaming. Sara Lazarevic et al. [15] proposes a system for recommending courses that is powered by machine learning and based on commonalities between courses. The approach suggested in this work is based on rankings according to study areas. Using these methods, we were able to develop an algorithm that, given an input, returns courses that meet certain criteria. The outcomes meet the needs of students who will utilize it to further their education in discovering comparable courses offered through cross-platform applications. Mursalin Islam Emon et al. [6] provide a well-organized framework for a system of intellectual advice for self-learners. The Recommendation System (RS) employs hybrid methods that combine user-based collaborative filtering and association rule mining. We chose to use hybrid methods since traditional recommendation algorithms can only meet students’ needs to a certain extent. Hybrid approaches are better able to pinpoint each student’s individual learning preferences, resulting in more tailored recommendations for the students. Table 1 displays the literature review of the existing methods.

Table 1
Literature review of the existing methods

Authors Methods Advantages Disadvantages

Santosh Kumar Banbhrani et al. [2] Taylor Chimp Optimization Algorithm (Taylor ChOA)-based Random Multimodal Deep Learning (RMDL) for course recommendation. This method resolved issues of information overload in the online educational field improved performance. Processing time of this method was too high.

Gao, M., et al. [7] Convolutional Neural Network (CNN) with negative sequence mining for online course recommendation. This method efficiently predicted that, which courses are most commonly to be misselected by users for improving overall performance. This scheme was not suitable for real-time applications.

Ren, X., et al. [22] Long- and Short-Term Memory networks (LSTM) and Attention mechanisms for multimodal course recommendation. This method effectively reduced the time for users in choosing courses and realized Personalized Course recommendation service. This method failed in performing sentiment analysis for evaluation of the relationship between learners’ ratings and course preferences.

Ali, S., et al. [1] E-learning Recommendation Architecture (ELRA) in a virtualized environment for enabling recommendation of course. This method utilized two distinct features that reduced contact and confusion among consumers and the high-quality education team. Memory consumption in this method was high.

Qinglong Li, and Jaekyeong Kim [16] Deep learning-based Course Recommender system (DECOR) for robust recommendation performance. Here, personalized course recommendations and many features integrations were included so as to provide suitable recommendations for users’ individual needs. There are not enough public datasets to train domain recommender systems.

Muhammad Sajid Rafiq et al. [21] Intelligent query optimization and course recommendations for enhancing online lectures. This scheme improved the survivability and accessibility of intelligent educational knowledge. It failed in using genetic algorithms and swarm intelligence to improvise course accuracy of online query optimization and course recommendation.

Wang et al. [27] Demand-aware Collaborative Bayesian Variational Network (DCBVN) for personalized employee training course recommendation. This method proved to be robust against cold-start and sparse scenarios. It failed in recommending the most appropriate training courses for employees.

Shahbazi, Z. and Byun, Y.C., [24] Knowledge Discovery and Machine Learning Approaches for an agent-based recommendation. Here, user requests without any interruptions or inconvenience are preferred. It failed to improvise its performance with other algorithms in focusing on gaming.

Sara Lazarevic et al. [15] A well-organized framework for a system of intellectual advice for self-learners. The outcomes meet the needs of students who will utilize it to further their education in discovering comparable courses offered through cross-platform applications. -

Mursalin Islam Emon et al. [6] Awell-organized framework for a system of intellectual advice for self-learners. Hybrid approaches are better able to pinpoint each student’s individual learning preferences, resulting in more tailored recommendations for the students.

Authors	Methods	Advantages	Disadvantages
Santosh Kumar Banbhrani et al. [2]	Taylor Chimp Optimization Algorithm (Taylor ChOA)-based Random Multimodal Deep Learning (RMDL) for course recommendation.	This method resolved issues of information overload in the online educational field improved performance.	Processing time of this method was too high.
Gao, M., et al. [7]	Convolutional Neural Network (CNN) with negative sequence mining for online course recommendation.	This method efficiently predicted that, which courses are most commonly to be misselected by users for improving overall performance.	This scheme was not suitable for real-time applications.
Ren, X., et al. [22]	Long- and Short-Term Memory networks (LSTM) and Attention mechanisms for multimodal course recommendation.	This method effectively reduced the time for users in choosing courses and realized Personalized Course recommendation service.	This method failed in performing sentiment analysis for evaluation of the relationship between learners’ ratings and course preferences.
Ali, S., et al. [1]	E-learning Recommendation Architecture (ELRA) in a virtualized environment for enabling recommendation of course.	This method utilized two distinct features that reduced contact and confusion among consumers and the high-quality education team.	Memory consumption in this method was high.
Qinglong Li, and Jaekyeong Kim [16]	Deep learning-based Course Recommender system (DECOR) for robust recommendation performance.	Here, personalized course recommendations and many features integrations were included so as to provide suitable recommendations for users’ individual needs.	There are not enough public datasets to train domain recommender systems.
Muhammad Sajid Rafiq et al. [21]	Intelligent query optimization and course recommendations for enhancing online lectures.	This scheme improved the survivability and accessibility of intelligent educational knowledge.	It failed in using genetic algorithms and swarm intelligence to improvise course accuracy of online query optimization and course recommendation.
Wang et al. [27]	Demand-aware Collaborative Bayesian Variational Network (DCBVN) for personalized employee training course recommendation.	This method proved to be robust against cold-start and sparse scenarios.	It failed in recommending the most appropriate training courses for employees.
Shahbazi, Z. and Byun, Y.C., [24]	Knowledge Discovery and Machine Learning Approaches for an agent-based recommendation.	Here, user requests without any interruptions or inconvenience are preferred.	It failed to improvise its performance with other algorithms in focusing on gaming.
Sara Lazarevic et al. [15]	A well-organized framework for a system of intellectual advice for self-learners.	The outcomes meet the needs of students who will utilize it to further their education in discovering comparable courses offered through cross-platform applications.	-
Mursalin Islam Emon et al. [6]	Awell-organized framework for a system of intellectual advice for self-learners.	Hybrid approaches are better able to pinpoint each student’s individual learning preferences, resulting in more tailored recommendations for the students.

2.2. Challenges

The major challenges of many traditional course recommendation techniques are given as below.

The developed technique in [2] attained better performance using precision, recall, and f1-score with higher values. However, the performance of the devised approach is not much evaluated utilizing more evaluation metrics. Thus, the challenge lies in extending the proposed method by developing deep learning classifiers and evaluating performance using more evaluation metrics.

In [22], the method considers the integration of multiple modal features, so that recommended personalized course content meets the needs of users. Because the proposed method utilizes the operation of users browsing courses for obtaining implicit feedback data, most of the implicit feedback data of users in daily life are difficult to capture which is considered as a challenging task.

In [1], ELRA techniques greatly increase accomplishments and skills, along with learning success. However, the proposed method failed in updating the learning materials and syllabus. Hence, the challenges of the scheme focus on creating ELRA to continuously update learning materials and syllabus, as well as in verifying the progress of both teachers and learners in improving the quality of courses in the online learning environment.

In [16], the method solved many challenges in terms of sparsity and high dimensionality issues. Hence, the challenge lies in addressing high-dimensionality and sparsity problems in practical online education platforms.

E-learning sites are useful in improving awareness and skills of academic backbones, such as students, instructors, administrative staff, as well as those who are searching for present information about various educational institutes. Despite all benefits of the online learning platform, users face challenges and complexities, like selecting appropriate learning material and courses based on preferences and needs.

3. Course recommendation using FLICM clustering with matrix factorization

The selection of optional courses based on students’ knowledge, interests, and skills is termed as course recommendation. The primary intention of this research is to design and develop a new method for recommending maximum-rated items to the user. At first, E-khool data [5] is taken as a dataset, and then the learner or course agglomerative matrix is computed. After that, course grouping is done by FLICM Clustering [13]. Furthermore, Minkowski and Chebyshev Distance-based bi-level matching is performed between learner query and group so that retrieval of learner preferred courses is done. At last, courses having maximum ratings are predicted using matrix factorization [23] based on user ID and course ID for recommending courses to the learner. Figure 1 represents the block diagram of FLICM clustering with matrix factorization to predict user ratings based on course recommendations.

Fig. 1.

Block diagram of FLICM clustering with matrix factorization to predict user ratings based on course recommendation.

3.1. Data acquisition

The initial stage in this research work is the acquisition of data from the dataset [5], which contains the course and learner. Each learner is interested to learn a particular course and based on this learner and course is represented. Here, the learner and course are represented as below, $\begin{array}{c} (1) & L = C_{a} \\ (2) & C = B_{d} \end{array}$

Here, L represents learner, and C indicates course.

3.2. Computation of agglomerative matrix

After the learner and course is acquired from E-khool dataset, the agglomerative matrix is computed. This learner or course agglomerative matrix has a learner series matrix, learner series binary matrix, course series matrix, and course rating binary matrix. These are explained in detail as below.

3.2.1. Learner series matrix

Here, input data is taken from the dataset and then presented to the learner series matrix $D_{a}$ . This matrix includes many courses visited by learners and is represented in the formula as below, $\begin{matrix} (3) & D_{a} = {B_{1}^{q}, B_{2}^{a}, \dots, B_{b}^{a}, \dots, B_{c}^{a}} \end{matrix}$ where, $D_{a}$ is course visited by the learner a, $B_{c}^{a}$ is $c^{th}$ course visited by $a^{th}$ learner, and $B_{b}^{a}$ is $b^{th}$ course visited by $a^{th}$ learner.

3.2.2. Learner series binary matrix

This binary matrix is achieved by courses searched by the learner, depending on the learner series matrix and hence this matrix is indicated as 0 and 1. Here, each learner series corresponds to binary values of courses given in binary sequence and is formulated in the matrix as follows, $\begin{matrix} (4) & {ND}_{a}^{b} = \{\begin{array}{ll} 1; & B_{b}^{a} \in B_{d} \\ 0; & otherwise \end{array} \end{matrix}$ where, 1 is course searched by the learner a, 0 is a course not searched by the learner a, and ${ND}_{a}^{b}$ is learner series binary matrix.

3.2.3. Course series matrix

Course series matrix is formed depending on learners visiting courses and indicates learners along courses in matrix form that is denoted as below, $\begin{matrix} (5) & C_{d} = {N_{1}^{d}, N_{2}^{d}, \dots, N_{i}^{d}, \dots, N_{n}^{d}} \end{matrix}$ where, $N_{n}^{d}$ is $n^{th}$ learner visited $d^{th}$ course, $N_{i}^{d}$ is $i^{th}$ learner visited $d^{th}$ course, and $C_{d}$ is learner visited $d^{th}$ course.

3.2.4. Course rating binary matrix

Learners searching courses is varied as 1 or 0, depending on the courses provided is termed as course rating binary matrix. This course rating binary matrix is represented in binary sequence as 1 or 0 and based on the course selected by learners. This considers the course series matrix and is formulated as, $\begin{matrix} (6) & {NC}_{d}^{i} = \{\begin{array}{ll} 1; & N_{i}^{d} \in C_{a} \\ 0; & otherwise \end{array} \end{matrix}$ where, 1 is learner searched course d, 0 is learner not searched course d, and ${NC}_{d}^{i}$ is the course series binary matrix.

3.3. Course grouping by FLICM

As the agglomerative matrix is calculated, course grouping is carried out by FLICM, from the course rating binary matrix ${NC}_{d}^{i}$ . This course grouping is needed for selecting the best groups of courses. Course grouping is a set of the total count of groups generated by FLICM [13]. Course grouping is formulated as below, $\begin{matrix} (7) & G = {G_{1}, G_{2}, \dots, G_{l}} \end{matrix}$ where, G is grouped courses, and l is total number of groups. FLICM used for course grouping is clearly explained in below section,

3.3.1. FLICM clustering algorithm

This algorithm is used for image or data-based clustering mechanism that incorporates gray-level information and spatial information by enhancing clustering performance. This preserves details of the image as well as data, and also guarantees insensitiveness in noise, because of the use of fuzzy local similarity measure [13].

a ) Introduction of Fuzzy factor

Many methods are utilized for clustering mechanisms, but all those methods lack insensitiveness to noise and also lack robustness. Furthermore, those methods assign a parameter in objective functions for balancing robustness in noise and effective preservation of details in data, but this selection parameter must be assigned only by trial and error experiments. Also, all clustering methods are only applied to static data rather than using original data. Such drawbacks are overcome by FLICM by introducing a fuzzy factor in FCM objective function. This factor has many characteristics, such as preserving robustness, avoiding noise insensitiveness, controlling neighborhood data based on central data distance, avoiding pre-processing steps by using original data and avoiding missing details in data, and avoiding parameter selection. Here, the fuzzy factor is formulated as, $\begin{matrix} (8) & F_{x y} = \sum_{\begin{array}{c} z \in T_{y} \\ y \neq z \end{array}} \frac{1}{e_{y z} + 1} {(1 - f_{x z})}^{g} ‖ h_{z} - j_{x} ‖^{2} \end{matrix}$ where, $y^{th}$ data is the center of the local window, x is the reference cluster, $z^{th}$ data belong in set of neighbors that fall into the window around $y^{th}$ data. $e_{y z}$ is Euclidean distance among data y and z, $f_{x z}$ is membership degree of $z^{th}$ data in $x^{th}$ cluster, g is weighting exponent on every fuzzy membership, $h_{z}$ is gray level value of neighbors of window center, and $j_{x}$ is the prototype of the center of the cluster x. This factor controls noise and details of data.

b ) Framework of FLICM

This framework incorporates both gray-level information and local spatial into an objective function that is formulated in term $M_{g}$ as below, $\begin{matrix} (9) & M_{g} = \sum_{y = 1}^{T} \sum_{x = 1}^{k} [f_{x y}^{g} ‖ h_{y} - j_{x} ‖^{2} + F_{x y}] \end{matrix}$

Necessary conditions for $M_{g}$ in accordance with $f_{x y}$ and $j_{x}$ for being local minimal extreme is, $\begin{array}{c} (10) & f_{x y} = \frac{1}{\sum_{z = 1}^{k} {(\frac{‖ h_{y} - j_{x} ‖^{2} + F_{x y}}{‖ h_{y} - j_{z} ‖^{2} + F_{z y}})}^{1 / g - 1}} \\ (11) & j_{x} = \frac{\sum_{y = 1}^{T} f_{x y}^{g} h_{y}}{\sum_{y = 1}^{T} f_{x y}^{g}} \end{array}$ where, k is cluster prototype number, g is fuzzification parameter, T is set of neighbors, and $h_{y}$ is window center. When the algorithm gets converged, the process of defuzzification happens to change the fuzzy partition matrix to a crisp partition. This is attained by the maximum membership procedure and this assigns data y to the class E with the largest membership. $\begin{matrix} (12) & E_{y} = {arg}_{x} {max {f_{x y}}}, x = 1, 2, . ., k . \end{matrix}$ where, $f_{x y}$ is the condition of $M_{g}$ to be local minimal extreme, and k is a number of clusters. Thus, FLICM adopts fuzzy factors for preserving data details by providing noise immunity, lacking in selecting a parameter, preserving details of data, and application is done with original data. Therefore, by FLICM, grouped courses is generated and indicated by the term G.

3.4. Bilevel product matching by Minkowski and Chebyshev distance

This process is done after course grouping based on the binary query and grouped courses for retrieving the best group for further learner retrieval. This is done by Minkowski and Chebyshev’s distance between the grouped course and binary sequence query. Chebyshev distance is a metric defined on vector space where the distance among dual vectors is the maximum of their differences along any coordinate dimension, wherein Minkowski distance is the distance measured among dual points in N-dimensional space. Minkowski is a generalization of Euclidean distance and Manhattan distance. This bilevel product matching is done based on the below formula, $\begin{matrix} (13) & {Mat}_{1} (G, {NQ}_{u}) = α d_{md} (G, {NQ}_{u}) + (1 - α) d_{cd} (G, {NQ}_{u}) \end{matrix}$ where, ${Mat}_{1} (G, {NQ}_{u})$ is bilevel product matching, $d_{md}$ is Minkowski distance, $d_{cd}$ is Chebyshev distance, α is constant, G is grouped courses, and ${NQ}_{u}$ is binary sequence query. Here, query and binary query is explained as followed for bilevel product matching.

3.4.1. Query

The query is matched with the course group and represented in a set format with many numbers of courses as formulated below, $\begin{matrix} (14) & Q_{u} = {q_{u 1}, q_{u 2}, \dots, q_{u r}, \dots, q_{u s}} \end{matrix}$ where, $q_{u r}$ is $r^{th}$ course in query, $q_{u s}$ is $s^{th}$ course in query, and s is number of courses.

3.4.2. Binary query

The binary series of the query is computed from the query and this tends to product matching with the grouped course and hence formulated by, $\begin{matrix} (15) & {NQ}_{u} = \{\begin{array}{ll} 1; & q_{u r} \in B_{d} \\ 0; & otherwise \end{array} \end{matrix}$ where, $B_{d}$ is course data, and $q_{u r}$ is $r^{th}$ course in query.

Thus, the binary query and the grouped course is matched using Minkowski and Chebyshev distance and is denoted as ${Mat}_{1} (G, {NQ}_{u})$ . This is then fed to the relevant customer retrieval stage.

3.5. Relevant learner retrieval

After Bilevel product matching by Minkowski and Chebyshev distance, relevant learner retrieval is carried out for finding the best group of learner series. This best group learner series is depicted in the formula as given below, $\begin{matrix} (16) & P^{m} = {m_{1}^{w}, m_{2}^{w}, \dots, m_{u}^{w}, \dots, m_{v}^{w}} \end{matrix}$ where, $P^{m}$ is best group learner series, v is total best group learner, and $m_{u}^{w}$ is best group learner in $u^{th}$ position in series. Its binary best group is indicated by the below formula, $\begin{matrix} (17) & {NP}_{u}^{m} = \{\begin{array}{ll} 1; & m_{u}^{w} \in B_{d} \\ 0; & otherwise \end{array} \end{matrix}$ where, ${NP}_{u}^{m}$ is the binary best group, $B_{d}$ is course data, and $m_{u}^{w}$ is best group learner in $u^{th}$ position in series.

3.6. Bilevel matching series for learners’ preferred course by Minkowski and Chebyshev distance

After attaining the binary best group and binary sequence query, group learner matching is done by Minkowski and Chebyshev distance among both binary best group and binary sequence query. This is formulated as below, $\begin{matrix} (18) & {Mat}_{2} ({NQ}_{u}, {NP}_{u}^{m}) = β d_{md} ({NQ}_{u}, {NP}_{u}^{m}) + (1 - β) d_{cd} ({NQ}_{u}, {NP}_{u}^{m}) \end{matrix}$ where, β is constant, $d_{md}$ is Minkowski distance, $d_{cd}$ is Chebyshev distance, ${NP}_{u}^{m}$ is the binary best group, and ${NQ}_{u}$ is binary sequence query. The output of the bilevel matching result is learner preferred course that is formulated as, $\begin{matrix} (19) & A_{c} = {A_{1}, A_{2}, \dots, A_{n}} \end{matrix}$ where, $A_{n}$ is $n^{th}$ course preferred by the learner.

3.7. Recommendation using matrix factorization

Implicit feedbacks are contained in matrix factorization where information is derived by finding user behavior towards item rather than direct involvement of user towards item [23]. This user behavior towards item is derived on basis of user-item interaction matrix. This method is thus helpful for predicting or estimating user ratings towards specific item. Moreover, rating matrix T is approximation of product among two matrices, such as U and V, where $U \in T^{p \times t}$ is user latent matrix and $V \in T^{o \times t}$ is item latent matrix with t as total count of latent factors and $t ≪ p, o$ . Here, each item o is related with vector $v_{x_{o}} \in T^{o \times t}$ and every user p is related with vector $u_{x_{p}} \in T^{p \times t}$ . Also, for item o, element of $v_{x_{o}}$ measures how much item possesses positive or negative factors, and for user p, element $u_{x_{p}}$ measures how much user p shows interest in a specific item. Dot product $v_{x_{o}} . u_{x_{p}}$ predict interaction between item and user as followed, $\begin{matrix} (20) & {prediction}_{interaction} = v_{x_{o}}^{Z} u_{x_{p}} \end{matrix}$ where, ${prediction}_{po}$ is predicted interaction’s approximation.

3.8. Final recommendation

The final product or course is recommended based on the maximum rating of the particular course, which is analyzed by learner-preferred course prediction. The final product is indicated by the term ${Fin}_{pro}$ . Figure 2 shows final recommendation result enhancing preferred course.

Fig. 2.

Final recommendation result.

The Pseudocode of above mentioned process is depicted in Algorithm 1.

Algorithm 1

Pseudocode of course recommended framework

Thus, course recommendation based on the E-khool dataset by FLICM enabled matrix factorization is designated by ${Fin}_{pro}$ .

4. Results with discussion

Enhancement of performance with various experimental results on FLICM-based matrix factorization by considering evaluation measures, like precision, recall, and f-measure is elaborated in this section.

4.1. Experimental setup

Implementation of FLICM-based matrix factorization is carried out in the PYTHON tool with PC having Windows 10 OS, 4 GB RAM, and an Intel I3 processor using E-khool dataset [5]. The batch size is 20 and the epoch varied as 20, 40, 60, 80, 100.

4.2. Dataset description

The dataset used in this paper is the E-khool learning platform dataset [5]. This dataset consists of one data file with 1 lakh rows, i.e. over 25 courses and 1000 learners. Here, Course ID, Date of subscription, Learner ID, and Ratings ranging between 1 to 5, along the date of ratings and review are clearly given.

4.3. Performance metrics

The performance of developed FLICM-based matrix factorization is analyzed by considering evaluation measures, like precision, recall, and F-measure.

Precision : It is the value obtained by the ratio of true positives to total positives. Here, total positives include the sum of true positives along with false positives. The precision measure is expressed as below, $\begin{matrix} (21) & χ = \frac{{Pos}_{true}}{{Pos}_{true} + {Pos}_{false}} \end{matrix}$ where, χ is precision, ${Pos}_{true}$ is true positives, and ${Pos}_{false}$ signifies false positives.

Recall : These measures the ratio of true positives to a total of false negatives and true positives and the formula is given by, $\begin{matrix} (22) & δ = \frac{{Pos}_{true}}{{Pos}_{true} + {Neg}_{false}} \end{matrix}$ where, recall is signified as δ, and ${Neg}_{false}$ symbolizes false negatives.

f-measure : This indicates the harmonic average value between precision and recall. This also indicates a weighted measure of both recall and precision that is given as, $\begin{matrix} (23) & f_{m} = 2 \times (\frac{χ \times δ}{χ + δ}) \end{matrix}$ where, $f_{m}$ denotes f-measure.

4.4. Performance analysis

FLICM_matrix factorization is analyzed for its performance by altering iteration values. Here, the iterations included are 20, 40, 60, 80, and 100. This is assessed with three performance metrics, such as f-measure, precision, and accuracy.

4.4.1. Performance assessment with cluster size 4

Figure 3 shows the performance assessment ofFLICM_matrix factorization with a cluster size of 4 by varying its iterations. Figure 3 (a) depicts an f-measure-based performance assessment with a cluster size of 4. Here, when the query is considered as 3, then the f-measure value for FLICM_matrix factorization with iteration 100 is 0.738, wherein f-measure reduces to 0.718, 0.719, 0.736, and 0.726 for FLICM_matrix factorization with varying iterations of 20, 40, 60, and 80. Precision-based performance analysis with cluster size 4 is depicted in Fig. 3 (b). For query 1, the precision values of FLICM_matrix factorization are 0.726 for iteration 20, 0.746 for iteration 40, 0.755 for iteration 60, 0.786 for iteration 80, and 0.806 for iteration 100. Figure 3 (c) enhances recall-based performance analysis with a cluster size of 4. For query 2, recall is 0.751 for iteration 100, and it becomes lesser for other iterations of 20, 40, 60, and 80 with values of 0.703, 0.715, 0.728, and 0.740.

Fig. 3.

Performance analysis of FLICM_matrix factorization with cluster size-4, (a) f-measure, (b) precision, and (c) recall.

4.4.2. Performance assessment with cluster size 5

Performance analysis of FLICM_matrixfactorization with cluster size 5 is indicated in Fig. 4. Figure 4 (a) indicates an f-measure-based performance analysis with cluster size = 5. When the query is 5, then the f-measure value for FLICM_matrix factorization having iteration 100 is 0.753, whereas f-measure decrease to 0.732, 0.746, 0.738, and 0.725 for FLICM_matrixfactorization with varying iterations of 80, 60, 40, and 20. Figure 4 (b) indicates precision-based performance analysis for cluster size = 5. When query = 4, precision values are 0.745 for iteration 20, 0.778 for iteration 40, 0.783 for iteration 60, 0.813 for iteration 80, and 0.845 for iteration 100. Figure 4 (c) shows recall indicated performance analysis with a cluster size of 5. For query = 3, recall is 0.839 for iteration 100, and it becomes less for iterations 20, 40, 60, and 80 with values of 0.742, 0.762, 0.779, and 0.794.

Fig. 4.

Performance analysis of FLICM_matrix factorization with cluster size-5, (a) f-measure, (b) precision, and (c) recall.

4.5. Comparative methods

This proposed method is compared with many methods, like TaylorChOA_RMDL [2], CNN [7], LSTM [22], ELRA [1], DECOR [16], and DCBVN [27] evaluated with three performance metrics.

4.6. Comparative analysis

Comparative analysis is done for FLICM_matrix factorization by altering cluster sizes of 4 and 5 and analyzed with three metrics by the varying query.

4.6.1. Comparative assessment with cluster size 4

Figure 5 demonstrates a comparative assessment of FLICM_matrix factorization with cluster size 4. Figure 5 (a) shows an f-measure-based comparative assessment with cluster size 4. When the query is 1, then the f-measure value is 0.852 for FLICM_matrix factorization, whereas it reduces to 0.824 for TaylorChOA_RMDL, 0.806 for CNN, 0.741 for LSTM, 0.814 for ELRA, 0.830 for DECOR, and 0.837 for DCBVN. This shows improvement in performance with values of 3.31%, 5.36%, 13.07%, 4.48%, 3.03%, and 1.76%. Figure 5 (b) inhibits precision-based comparative assessment with cluster size 4. When the query is 2, then precision is 0.865 for the proposed model, wherein precision decreases to 0.839 for TaylorChOA_RMDL, 0.816 for CNN, 0.779 for LSTM, 0.824 for ELRA, 0.848 for DECOR and 0.858 for DCBVN. Figure 5 (c) shows a recall-based comparative assessment with a cluster size of 4. If the query is considered as 5, then recall is 0.850 for FLICM_matrix factorization, whereas it is lesser with values of 0.844, 0.830, 0.768, 0.843, 0.846, and 0.848, for methods, such as TaylorChOA_RMDL, CNN, LSTM, ELRA, DECOR, and DCBVN. Here, performance improvement of recall with other methods is 0.68%, 2.35%, 9.64%, 0.66%, 0.55%, and 0.23%.

Fig. 5.

Comparative analysis of FLICM_matrix factorization with cluster size-4, (a) f-measure, (b) precision, and (c) recall.

4.6.2. Comparative assessment with cluster size 5

Figure 6 depicts a comparative assessment of the developed model with a cluster size of 5. Figure 6 (a) shows the f-measure-based comparative analysis of FLICM_matrix factorization with cluster size = 5. When the query is 5, then f-measure values are 0.882, 0.849, 0.790, 0.843, 0.859, 0.866, and 0.872, for FLICM_matrix factorization, ELRA, LSTM, CNN, TaylorChOA_RMDL, DECOR, and DCBVN. Figure 6 (b) depicts the precision-related comparative analysis of FLICM_matrix factorization with cluster size 5. If query = 2, then precision is 0.840 for TaylorChOA_RMDL, 0.821 for CNN, 0.791 for LSTM, 0.828 for ELRA, 0.848 for DECOR, 0.857 for DCBVN and 0.880 for FLICM_matrix factorization. Figure 6 (c) enumerates a recall-based comparative analysis of FLICM_matrix factorization with cluster size 5. If query = 4, then recall values are 0.839, 0.829, 0.760, 0.834, 0.843, 0.845, and 0.844 for methods, such as TaylorChOA_RMDL, CNN, LSTM, ELRA, DECOR, DCBVN, and FLICM_matrix factorization.

Fig. 6.

Comparative analysis of FLICM_matrix factorization with cluster size-5, (a) f-measure, (b) precision, and (c) recall.

4.7. Comparative discussion

Table 2 shows a comparative discussion of FLICM_matrix factorization with other methods. Here, the precision value is 0.915 for FLICM_matrix factorization with cluster size 5, whereas it is lesser for other methods, such as TaylorChOA_RMDL, CNN, LSTM, ELRA, DECOR, and DCBVN with values of 0.860, 0.836, 0.826, 0.850, 0.865, and 0.880. This improvement in the precision of FLICM_matrix factorization is due to the development of an agglomerative matrix, where the binary matrix enhances the performance rating. Moreover, the recall value is more with 0.850 for the proposed model with cluster size = 5, as it is lesser for other methods and this is because of using the FLICM method for the clustering process. Furthermore, the f-measure value is 0.882 for FLICM_matrix factorization with a cluster size of 5, and for other techniques such as TaylorChOA_RMDL, CNN, LSTM, ELRA, DECOR, and DCBVN, f-measure values are lesser. Here, f-measure enhancement in the proposed model is due to matrix factorization utilized for a recommendation of a course from a preferred course.

Table 2
Comparative discussion of FLICM_matrix factorization

Methods/Metrics TaylorChOA_RMDL CNN LSTM ELRA DECOR DCBVN Proposed FLICM_matrix factorization

Cluster size = 4 Precision 0.854 0.836 0.816 0.845 0.863 0.875 0.882

Recall 0.844 0.830 0.768 0.844 0.846 0.848 0.850

f-measure 0.854 0.837 0.782 0.845 0.864 0.872 0.882

Cluster size = 5 Precision 0.860 0.836 0.826 0.850 0.865 0.880 0.915

Recall 0.844 0.830 0.768 0.845 0.850 0.850 0.850

f-measure 0.859 0.843 0.790 0.849 0.866 0.872 0.882

Methods/Metrics	TaylorChOA_RMDL	CNN	LSTM	ELRA	DECOR	DCBVN	Proposed FLICM_matrix factorization
Cluster size = 4	Precision	0.854	0.836	0.816	0.845	0.863	0.875	0.882
Recall	0.844	0.830	0.768	0.844	0.846	0.848	0.850
f-measure	0.854	0.837	0.782	0.845	0.864	0.872	0.882
Cluster size = 5	Precision	0.860	0.836	0.826	0.850	0.865	0.880	0.915
Recall	0.844	0.830	0.768	0.845	0.850	0.850	0.850
f-measure	0.859	0.843	0.790	0.849	0.866	0.872	0.882

5. Conclusion

The field of education is improvised nowadays due to new development in technologies, eventually via the Internet. E-learning platform mainly helps students to enroll with decisions and provides a recommendation of optional and selective courses based on student’s skills, interests in subjects, free slots of time in the timetable, and knowledge of students. In this work, course recommendation is done based on FLICM clustering with matrix factorization. The dataset used in this work is E-khool review data, from which an agglomerative matrix is generated. This matrix has a learner series matrix, learner series binary matrix, course series matrix, and course series binary matrix. After this matrix generation, course grouping is done using FLICM, and further bilevel matching is carried out using Minkowski and Chebyshev distance. Then, the relevant learner and group user are retrieved by Minkowski and Chebyshev distance, from which the learner-preferred course is retrieved. Finally, the course is recommended using matrix factorization that has a maximum rating. The performance of FLICM clustering with matrix factorization is enhanced by three performance metrics like f-measure, recall, and precision with values of 0.882, 0.850, and 0.915, accordingly. The proposed method’s precision is limited to 0.915, which is a drawback. We’ll keep updating the models in subsequent work to improve prediction precision. For instance, we can train the models with restrictions like prerequisite courses or nonnegative contents in the matrices, or we can use a multi-relational strategy to employ more relational information. In addition, we’ll strive to use instances from actual life to illustrate our explanation. One technique for reducing dimensionality is matrix factorization, which has been used in numerous applications. Matrix factorization techniques in feature extraction break down a matrix into its component pieces, which simplifies the procedure and improves speed while requiring less computational work.

References

Ali,

Hafeez,

Humayun,

N.S.M.

Jamail,

Aqib and

Nawaz, Enabling recommendation system architecture in virtualized environment for e-learning, Egyptian Informatics Journal 23(1) (2022), 33–45. doi:10.1016/j.eij.2021.05.003.

S.K.

Banbhrani,

Xu,

Lin and

D.K.

Sajnani, Taylor-ChOA: Taylor-chimp optimized random multimodal deep learning-based sentiment classification model for course recommendation, Mathematics 10(9) (2022), 1354. doi:10.3390/math10091354.

Bendakir and

Aïmeur, Using association rules for course recommendation, in: Proceedings of the AAAI Workshop on Educational Data Mining, Vol. 3, 2006, pp. 1–10.

Burke, Hybrid recommender systems: Survey and experiments, User modeling and user-adapted interaction 12(4) (2002), 331–370. doi:10.1023/A:1021240730564.

E-khool learning platforms dataset, https://ekhool.com/, accessed on November 2022.

M.I.

Emon,

Shahiduzzaman,

M.R.H.

Rakib,

M.S.A.

Shathee,

Saha,

M.N.

Kamran and

J.H.

Fahim, Profile based course recommendation system using association rule mining and collaborative filtering, in: The Proceedings of the International Conference on Science & Contemporary Technologies (ICSCT), 2021.

Gao,

Luo and

Hu, Online course recommendation using deep convolutional neural network with negative sequence mining, Wireless Communications and Mobile Computing (2022).

J.L.

Herlocker,

J.A.

Konstan and

Riedl, Explaining collaborative filtering recommendations, in: Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work, pp. 241–250, 2000.

Hua,

Wang,

Xu,

Zhang and

Liang, Predicting corporate financial distress based on integration of support vector machine and logistic regression, Expert Systems with Applications 33(2) (2007), 434–440. doi:10.1016/j.eswa.2006.05.006.

10.

A.N.

Jadhav and

Gomathi, DIGWO: Hybridization of dragonfly algorithm with improved grey wolf optimization algorithm for data clustering, Multimedia Research 2(3) (2019), 1–11.

11.

H.K.

Kim,

Y.U.

Ryu,

Cho and

J.K.

Kim, Customer-driven content recommendation over a network of customers, IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans 42(1) (2011), 48–56. doi:10.1109/TSMCA.2011.2147306.

12.

J.K.

Kim,

H.K.

Kim,

H.Y.

Oh and

Y.U.

Ryu, A group recommendation system for online communities, International journal of information management 30(3) (2010), 212–219. doi:10.1016/j.ijinfomgt.2009.09.006.

13.

Krinidis and

Chatzis, A robust fuzzy local information C-means clustering algorithm, IEEE transactions on image processing 19(5) (2010), 1328–1337. doi:10.1109/TIP.2010.2040763.

14.

Kulkarni and

T.S.

Murugan, Hybrid weed-particle swarm optimization algorithm and C-mixture for data publishing, Multimedia Research 2(3) (2019), 33–42.

15.

Lazarevic,

Zuvela,

Djordjevic,

Sladojevic and

Arsenovic, Machine learning driven course recommendation system, in: The Proceedings of the 21st International Symposium Infoteh-Jahorina (Infoteh), 2022.

16.

Li and

Kim, A deep learning-based course recommender system for sustainable development in education, Applied Sciences 11(19) (2021), 8993. doi:10.3390/app11198993.

17.

Li,

Nishimura,

Yagami and

H.S.

Park, An empirical study on online learners’ continuance intentions in China, Sustainability 13(2) (2021), 889. doi:10.3390/su13020889.

18.

Lu,

Wu,

Mao,

Wang and

Zhang, Recommender system application developments: A survey, Decision Support Systems 74 (2015), 12–32. doi:10.1016/j.dss.2015.03.008.

19.

A.A.

Ntwali, Similarity learning on big data: A case study, Multimedia Research 5(1) (2022).

20.

Perugini,

M.A.

Gonçalves and

E.A.

Fox, Recommender systems research: A connection-centric survey, Journal of Intelligent Information Systems 23(2) (2004), 107–143. doi:10.1023/B:JIIS.0000039532.05533.99.

21.

M.S.

Rafiq,

Jianshe,

Arif and

Barra, Intelligent query optimization and course recommendation during online lectures in E-learning system, Journal of Ambient Intelligence and Humanized Computing (2021), 1–20.

22.

Ren,

Yang,

Jiang,

Jin and

Yu, A deep learning framework for multimodal course recommendation based on LSTM+ attention, Sustainability 14(5) (2022), 2907. doi:10.3390/su14052907.

23.

Sanwal and

Çalişkan, A hybrid movie recommender system and rating prediction model, International Journal of Information Technology and Applied Sciences (IJITAS) 3(3) (2021), 161–168. doi:10.52502/ijitas.v3i3.128.

24.

Shahbazi and

Y.C.

Byun, Agent-based recommendation in E-learning environment using knowledge discovery and machine learning approaches, Mathematics 10(7) (2022), 1192. doi:10.3390/math10071192.

25.

Shirsat, Developing deep neural network for learner performance prediction in EKhool online learning platform, Multimedia Research 3(4) (2020).

26.

Srinivas, Prediction of e-learning efficiency by deep learning in E-khool online portal networks, Multimedia Research 3(4) (2020).

27.

Wang,

Zhu,

Zhang,

Chen and

Xiong, Personalized employee training course recommendation with career development awareness, in: Proceedings of the Web Conference 2020, 2020, pp. 1648–1659. doi:10.1145/3366423.3380236.

28.

Zhang,

Huang,

Lv,

Liu and

Zhou, MCRS: A course recommendation system for MOOCs, Multimedia Tools and Applications 77(6) (2018), 7051–7069. doi:10.1007/s11042-017-4620-2.