Fuzzy-inspired decision making for dependability recommendation in e-commerce industry

Abstract

Recommender System has become one of the most effective tools for provisioning user-interest based decision-making services. With its capability to generate efficient recommendations, users are directed towards items that are optimal with compliance to their needs, and preferences. Inspired from these aspects, this paper presents a novel recommendation technique based on context-specific information and social network analysis for determining dependable items. Context specific information provides a quantifiable measure of user interest for dependability whereas social network analysis determines the degree of similarity among other users. Both types of information are acquired and analyzed in the form of linguistic terms. This fuzzy-based quantification provides an effective way to evaluate social-ratings and social-similarity. For validation, it is evaluated in the on-line mobile purchase scenario. Based on the numerous simulations performed on different data sets, performance estimators in the form of Temporal Delay, Statistical Analysis and System Stability are estimated. It is concluded that the proposed mechanism of recommendation is effective and efficient in comparison to state-of-the-art recommender systems.

Keywords

Recommender system social network analysis context-specific information item-dependability fuzzy number set

1. Introduction

Recommender Systems (RS) have been a center of attraction among researchers, and innovators from the last three decades [1, 2, 3, 4]. Developments in data analytic have enhanced the applicability domain of RS to provide effective recommendations extending from individual personalization to group generalization [5]. Specifically, contributions in the fields of Semantic Web [6], Artificial Intelligence [7], Connectivity [8], and Ubiquity [6] resulting in Web 3.0 have revolutionized RS to provide effective services in important areas of E-Governance, E-Commerce, E-Tourism, E-Shopping, E-Library, and E-Learning systems [9, 10]. In addition to this, RS is beneficially utilized by an individual user, user-groups and even large business organizations for provisioning recommendations depending upon the service requirements [11]. Before going into detail, it is important to define this innovative conception for aiding in effective decision-making.

Definition 1. Recommender System (RS) are defined as programs or software tools used by users to determine suitable services or items by predicting the degree of user’s interest.

Definition 1 provides a comprehensive overview of the user-specific deployment of RS. Specifying technical perspective, RS tools require complex underneath computations and deep-data analysis for generating time-sensitive effective recommendations. Consequently, numerous recommendation techniques have been put forth by researchers to enhance decision-making accuracy in different user-oriented applications [3, 12]. Some of these techniques include the Collaborative Filtering (CF) mechanism, Content-Based Filtering (CB), and Knowledge-Based approach (KB) [13, 14]. CF provides a recommendation of items on the basis of users with similar interest factors. CB filtering, on the other hand, recommends items on the basis of the historical experience of the user for specific item. KB is based on domain knowledge about user preferences [15]. With the advancement of Web 3.0, numerous mechanisms like Social Network Analysis (SNA), Data Mining, and Neural Network (NN) have begun to present significant enhancements in the field of recommendation [16]. A brief overview of various RS techniques and corresponding domain of application have been provided in Table 1.

Table 1
Review of recommender system and applications

Domain	Recommendation applicability	Technique
Software engineering	Effort Estimation Methods	Knowledge-based Recommendation
Software engineering	Problem Solving Methods	Knowledge-based Recommendation
Software engineering	API Call Completions	Collaborative Filtering
Software engineering	Example Code Fragments	Content-based Recommendation
Software engineering	Contextualized Artefacts	Knowledge-based Recommendation
Software engineering	Software defects Analysis	Machine Learning
Software engineering	Requirements Prioritization	Group Recommendation
Software engineering	Software requirements	Collaborative Filtering
Data and knowledge engineering	Related Constraints	Knowledge-based Recommendation
Data and knowledge Engineering	Explanations (Hitting Sets)	Knowledge-based Recommendation
Data and knowledge engineering	Constraint re-writings	Knowledge-based Recommendation
Data and knowledge engineering	Database queries	Collaborative Filtering
Knowledge-based configuration	Relevant Features	Collaborative Filtering
Knowledge-based configuration	Requirements and Repairs	Collaborative Filtering
Persuasive technologies	Game Task Complexities	Collaborative Filtering
Persuasive technologies	Development Practices	Knowledge-based Recommendation
Smart homes	Equipment Configurations	Knowledge-based Recommendation
Smart home	Control Actions	Collaborative Filtering
People	Criminals	Probability-based Recommendation
People	Reviewers	Collaborative Filtering
People	Physicians	Collaborative Filtering
Points of interest	Tourism Services	Collaborative Filtering
Points of interest	passenger	Probability-based Recommendation
Help services	schools	Content-based Recommendation
Help services	Financial services	Knowledge-based Recommendation
Help services	Lifestyles	Knowledge-based Recommendation
Help services	Receipes	Content-based Recommendation
Help services	Health Information	Content-based Recommendation
Innovation management	Innovation Teams	Knowledge-based Recommendation
Innovation management	Business Plans	Knowledge-based Recommendation
Innovation management	Ideas	Data Mining

1.1 Limitations of recommender system techniques

Despite the numerous beneficial aspects of different RS mechanisms, each of the techniques mentioned above has its respective limitations and deployment challenges. For instance, sparseness, and cold-start are some of the drawbacks of CF [3, 13], while filtering with CB can lead to over-specialized recommendations [3, 17]. In addition to these, there are several other limiting factors of conventional recommender techniques like scalability, delay-latency, and sparsity [18]. Each of these limiting factors is discussed ahead in detail. Moreover, a complete overview of different limitations is itemized in Table 2.

Table 2
Limitations of recommender system techniques

Technique	Cold- start	Synonymy	Shilling attacks	Privacy issue	Over specialization	Grey sheep	Sparsity	Scalability	Latency problem	Context- awareness
Collaborative filtering	Yes	Yes	Possible	Yes	No	Yes	Yes	Yes	No	No
Content based	Yes	Yes	Yes	Yes	Yes	No	No	Yes	No	Yes
Knowledge based	No	No	Possible	Possible	No	Possible	No		No	No
Group recommendation	Yes	Yes	Yes	Yes	Possible	Possible	Yes	Yes	Possible	Possible
Probability based	Yes	No	No	No	No	Possible	Possible	No	No	Possible
Data mining based	Yes	Yes	No	No	Yes	No	Possible	Possible	Possible	Possible
Machine learning	Yes	No	No	Yes	Yes	No	No	Yes	Yes	Yes
Social network analysis	Yes	Possible	Yes	Yes	No	Yes	Yes	No	Possible	No
Context aware based	Yes	Possible	Yes	Yes	Possible	No	Yes	Yes	Yes	No

(a) Cold-Start problem

Cold-Start problem depicts that new users have been registered in the system or new items are added to the database. In such scenarios, neither the behavior or taste of new users is determined nor new items are rated or purchased by the users, resulting in reduced initial accuracy of RS.

(b) Synonymy

Synonymy indicates that a specific item is represented by two or more different entries or names that have a similar meaning. In these cases, RS is unable to identify whether multiple terms represent the same or different items.

(c) Shilling Attack

Shilling Attacks happen when a malicious, unauthorized, and impostor user or competitor enters into a system and begins to provide inaccurate ratings on items either to raise the item popularity or to diminish its generality. This type of attack results in minimizing of trust on RS, and thereby decreasing performance, and quality of RS.

(d) Privacy

Since RS requires feeding of user-specific personal information to systems for better recommendation services. However, it may lead to issues of data privacy, misuse, and security [19].

(e) Over-specialization

Insufficient availability of item-related content results in problems like specification or over-specialization. In this case, items are characterized by their subjective attributes, in which selecting an item is based only on their subjective attributes.

(f) Grey Sheep

Grey Sheep problem occurs when user-opinion does not match with any of the formulated user-groups. Therefore, RS is unable to give sufficient benefits to its users for recommendations.

(g) Sparsity

The availability of large data regarding items in the database and disinclination of users to provide ratings for items make the item-profile matrix very dispersed. This leads to the generation of minimum accuracy of recommendations.

(h) Scalability

Growth-rate of web-based data and social networking depicts a linear relation with the frequency of items and the number of users involved in the e-commerce industry. Consequently, it becomes difficult for a conventional RS to process and analyze data on a large scale in real-time.

(i) Latency problem

RS face latency problem in cases when new items are added in a continuous manner to the data repository. In these case, RS recommends only already-rated items since the newly added items are not rated, which is another limiting factor of conventional RS.

(j) Context-awareness

From functional view-point, context-awareness accumulates all categories of items that represent the environmental settings in which RS is to be deployed, e.g., current location, activity, and time. The inaccuracy of conventional RS in the acquisition of such information makes them inefficient and inaccurate.

In order to cope with these limitations, advanced techniques have been proposed by researchers around the world, namely Fuzzy-inspired RS [20, 21, 22], Social Network based RS [23, 24], Group RS [25], Context-Aware RS [25], and Computational Intelligence based RS [9]. Each of these has been efficiently implemented in their respective application fields.

1.2 Motivation

With the development of Information and Communication Technology (ICT) for Social Network Analysis, and Ubiquitous Connectivity resulting in Web 3.0, Social Network based RS (SNRS) [1] and Context-Aware RS (CARS) [9] have arisen from its infancy to enhance efficacy in the field of RS. Since the current research is inspired by both these techniques, it is important to discuss each of them in detail.

(a) SNRS provides a recommendation on the basis of web-based social engagements of the user, user entanglements, and relationships in the social world. In addition to this, SNRS integrates information with the recommendation procedure resulting in accurate, efficient, and personalized generation of recommendations for decision-making [3, 12]. Confining to the user’s social network, Trust is an important relational parameter in SNA. In the field of RS, it is defined as “Degree of trust that Alice has on Bob for a specific item about item feature or characteristic” [13]. Based on this aspect, many variants of SNRS like Trust-based RS have been developed for enhancement in accuracy and effectiveness of RS [15, 20]. SNRS analyses the propagation mechanism of trust among different users based on social interaction and decision-making. In addition to these, RS models have been developed in different application domains based on cooperative and mutual understanding in users’ relationship with consideration of Physical context [21], Social tags [22], and Co-authorship [1]. These relationship parameters substitute trust to filter out accurate user preferences. Other than social parameters, studies have also been conducted that incorporate user-item ratings for sorting item preferences in SNRS.

(b) CARS are another futuristic RS mechanism based on utilization of contextual information such as user-context (e.g. mood, geographic location, preference etc.), item-context (e.g. bandwidth, screen resolution etc.), and physical context (e.g. time, weight etc.) for generating effective recommendations [25]. Researchers of numerous domains have incorporated ubiquitous informative analysis in RS to provide optimized, effectiveness, and accuracy. In addition to this, with the involvement of contextual information of the item, its semantic meta-data is appropriately assessed. Henceforth, the degree of acceptability by users from different domains has been enhanced significantly [26, 27]. Moreover, since context varies from one application to another, depending upon the deployment domain of RS, numerous context-sensitive systems have been developed [28, 29]. Furthermore, approaches utilizing contextual information with conventional techniques like CF, and KB have also been a significant research area of in recent times.

1.3 Major contribution

Inspired from the aforementioned aspects, this paper presents a novel RS technique for provisioning effective recommendations for user-specified dependable items.

1.
The proposed methodology is a collaborative framework that utilizes user-based social information and item-based contextual data for provisioning accurate item-recommendations for user decision-making.
2.
Social network-based data analysis is incorporated for generating context-specified recommendations.
3.
SNA provides a determination of user-based similarity measure, whereas contextual information is utilized for preference analysis between the specified and evaluated degree of dependability.
4.
Different parameters are considered for enhancing the user experience like ranking, and ratings. In addition to this, the entire user and item-specific information are quantified in the form of a triangular fuzzy number system theory for effective analysis.

1.3.1 Paper organization

The remainder of this paper is organized into various sections. Section 2 provides an overview of related literature work in the field of social network-based RS and Context-aware RS. A preliminary study about dependability and its attributes has been discussed in Section 3. Section 4 presents the proposed framework of fuzzy-inspired Dependability RS. Experimental simulations and corresponding results have been presented in Section 5. Finally, Section 6 concludes the paper with some important discussions about the proposed system.

2. Literature review

This section reviews some of the important contributions that have been performed recently in the field of RS. Specifically, two sub-sections are made depicting numerous contributions in Social Network-based RS and Context-specific RS.

2.1 Social network-based recommender systems

Significant works have been performed by researchers in the field of Social Network-based RS. Consequently, numerous recommendation approaches have been developed by researchers in the aforementioned field for enabling recommendation based efficient decision-making. Sun et al. [1] have presented a regularization approach based on social interaction with friends for enhancing recommendation accuracy. Missing values, corresponding to user and item, are determined using rating-records and user-friendship. In addition to this, authors have utilized a bi-clustering algorithm for determining a suitable group of user-friends for recommending items. He and Chu [30] have utilized the user’s social network information to make a recommendation on the basis of personal preference. A probabilistic model has been proposed in order to eliminate data sparsity and cold-start problems. A similar kind of probability model has been presented by Domingos and Richardson [31] for determining the level of customer potentiality about purchasing a specific item by using the user’s social network information. Liu and Lee [32] have clubbed SNA with collaborative filtering for enhancing the recommendation generating procedure. Ma et al. [33] have presented a matrix-factorization framework by incorporating social regularization for improving the efficacy of prediction in conventional recommendation techniques. Besides these, much work has also been presented with consideration of “Trust” as a social attribute of the user’s relation. As in [34], Golbeck and Ann have proposed an algorithm named TidalTrust for determining Trust-based ratings, which was used in the formation of trust networks in RS. Local-trust matrix and Global-trust matrix have been analyzed by Hwang et al. [24] for enhancing the recommendation domain and its corresponding accuracy. In [35], Palau et al. have provided various statistical measures to determine collaborative relationships using SNA for recommendations. As mentioned earlier, RS are not confined to individual user but have been useful for group of users. With developments of social interaction, and ubiquitous connectivity leading to Web 3.0, social networking has effectively aided in the development of group recommendations. Li et al. [36] have provided social networking-based mechanism of group-coupon system for recommending location sensitive products to the group of users. User willingness-to-purchase has been considered as an evaluating metric, which included user preference, location, and friend influence. The proposed system, thereby determines group similarity using the evaluated score. Similar kind of work has been presented by Kim et al. [37] with the enhancement of individual users’ satisfaction along with group recommendation quality. Chen et al. [38] presented a hybrid model for group recommendation based on conventional collaborative filtering and genetic algorithm for determining group interaction and estimating group rating for an item.

2.2 Context-aware based recommendation systems

Context-Aware based systems for recommendation generation have evolved considerably over the past few years. However, substantial thrust has been provided recently due to incorporation of ubiquitous computing in Web 3.0. Moreover, for deployment of numerous user-oriented applications, different recommendation methodologies have been put forth by researchers around the world. Dao et al. [39] have presented a methodology for location-based advertising using context-aware information and collaborative filtering in mobile systems. Similarity between contexts is determined using genetic algorithm for enhancing location-specific recommendation. Mobasher et al. [40] depicted that similar items may receive extremely different preferences in heterogeneous context. Context-aware information is utilized by Colombo et al. [25] for developing a mobile-based RS. It is aimed at recommending movie, movie-theatre, and show timings to the user based on his/her preference and contextual information (time and crowd). Moreover, authors have developed a comprehensive multi-platform interface for users in order to leverage mobile sensors. A novel approach based on context-sensitive information have been proposed by Alhamid et al. [28]. In the study, authors have incorporated biological signal of user for proposing a collaborative filtering-based recommendation algorithm. Rating of user-information is performed by social tags for determining user preference. Apart from these many applications have been equipped with recommender module that collaborates social and contextual information for generation of recommendations. Development of ICT has extended the factors that influence user-context to include many environmental variables. Such surrounding factors play a crucial role in filtering out the similarity between friends more accurately and thereby, elevating the efficacy and effectiveness of RS.

Figure 1.

(a) Proposed recommender system mechanism (b) Proposed work-flow of recommendation generation.

3. Conceptualizing dependability

Before discussing the proposed framework of dependability RS, it is indispensable to understand the conceptual aspects of item-dependability from the perspective of the real-world user. In practical scenarios, dependability is a complex qualitative measure from a user viewpoint. In many cases, dependability represents an optimization measure that forms a balance between item-utilization and item-risks. Many authors have defined dependability in different terminology. In this paper, however, the following definition of dependability has been adopted.

Definition 2. Dependability is defined as the capability of an item to deliver a set of services in a justifiable and efficient manner for item availability, maintainability, safety, security, and reliability.

Though this is a precise definition in a real-world scenario, it can be extended to include many other practical aspects. Due to the high variation in applicability preference of the user, dependability specification embraces a high degree of user sensitiveness. Moreover, dependability itself being a complex measure consisting of numerous parameters further elevates the complexity domain of RS for data analysis. Many authors have presented works depicting dependability as a set of various attributes. However, in this paper, five important attributes are considered, namely: Availability (Av), Maintainability (Ma), Reliability (Re), Safety (Sa) and Security (Se). These are defined as follows.

(a)
Availability of an item can be defined as the readiness of the item for correct service to an authorized user.
(b)
Maintainability is the ability to revive the item or preventing it from failure.
(c)
Reliability of an item is the ability to deliver continuous service in an accurate manner.
(d)
Safety means ensuring a certain degree of measures to item making it less vulnerable to damage.
(e)
Security means the requirement of secure procedures during an operational state of an item for prevention from unauthorized attacks.

As mentioned above, various attributes constitute to provide a quantifiable form of dependability. Mathematically, it can be represented in the form of ordered set attributes as shown by Eq. (1).

$\displaystyle\vspace{-2cm}\textbf{Dependability}\!=<\!\![\text{Av}],[\text{Ma}% ],[\text{Re}],[\text{Sa}],[\text{Se}]\!\!>$ (1)
4. Proposed framework

In this study, a fuzzy-inspired methodology is presented to facilitate users with the recommendation in terms of a specified degree of dependability from the available list of items. The proposed methodology is mainly composed of four modules namely: Interface Module, Similarity Measurement and Filtering, Dependability Evaluation and finally Recommendation Generation. Figure 1a shows the modular framework of the proposed RS approach. Each module performs its intended functions and provides necessary informative services to adjacent modules. Moreover, a work-flow of the proposed model is presented in Fig. 1b to depict underneath the data analysis technique of the RS mechanism. Moreover, Table 3 provides a symbolic notations that are used in the presented approach

Table 3
Symbolic notation and its description

Symbolic notation	Description
$U(i)$	$i^{\rm th}$ Active user
$I(i)$	$i^{\rm th}$ Item
TAG( $i, j$ )	Fuzzy number for $i^{\rm th}$ user and $j^{\rm th}$ Item
$a$	Real number
$\bar{a}$	Fuzzy number
$\bar{F}_{a}^{X}$	User (a) feedback for attribute $X$
$\bar{F}_{x}$	Feedback for Item $x$
$\phi(i)$	Profile set for User( $i$ )
$\bar{R}^{X(1)}$	Requirement vector for attribute $X(1)$
$N_{1}$	Item-set

(a) Interface module: It is the user-accessible module of the presented system. Various requirements regarding the degree of item-dependability are provided to the system through this interface. Moreover, it contains a feedback section for users to provide user-specific feedbacks, depicting the level of satisfaction about item-dependability. Both kinds of user involvements with the system are performed using linguistic terms and corresponding triangular fuzzy number set.

(b) Similarity measurement and filtering: Determination of user-rating for dependability attributes about items that are new or unknown to the user are performed by this module. It utilizes user-friend feedbacks about items and thereby, aiding to the similarity determination procedure. Moreover, this module filters out the group of friends using an efficient classification algorithm for the determination of similar users.

(c) Dependability evaluation: After evaluating the necessary information about items, user similarity and their corresponding feedbacks, items with missing information are worked upon using it. Appropriate values are determined corresponding to the intended items about dependability.

(d) Recommendation module: Final list of dependability attributes for items are compared with the user-specified requirement. An algorithm is proposed for sorting out the recommended items on a priority basis, which is then presented to the user.

The detailed procedure of each module is explained in the section ahead.

Table 4

User-Item Matrix

User	I(1)	I(2)	I(3)	I(4)	I(5)
U(1)	TAG 11	TAG 12	$-$	TAG 14	TAG 15
U(2)	TAG 21	$-$	TAG 23	TAG 24	TAG 25
U(3)	$-$	TAG 32	$-$	TAG 34	$-$
U(4)	TAG 41	TAG 42	TAG 43	$-$	TAG 45
U(5)	TAG 51	$-$	TAG 53	TAG 54	$-$

4.1 User interface

There are two categories of users involved in the system, active-user, and passive-users (friends of the active user). Active-user is a user to whom the recommendation is to be provided. Passive-users, on the other hand, are the friends of the active user whose feedback influences the final rating during evaluation. For instance, Fig. 2 depicts a social network with 5 users (U (1), U (2), U (3), U (4), U (5)) in a system. Bidirectional arrows between pair of users denote mutual friendship among different users. U (1) is considered as an active user for requirement specification while other users act as the passive ones for feedback consideration. Analogous to this, items are also categorized into two sets: Known and Unknown Set. Known set includes those items that are already rated or complete information about dependability attributes is available with the active user. However, items whose ratings are unknown or not rated by the active user comprises the unknown set. Table 4 gives a synthetic instance of various users and items in the form of the user-item matrix. With respect to User U (1), Known item set includes Item I(1), I(2), I(4), I(5) while Unknown set includes only Item I(3).

Figure 2.

Instance of social network.

Basically, the user interface is composed of two sub-modules namely Requirement Acquisition Module (RAM) and Feedback Acquisition Module (FAM). RAM performs the task of information acquisition (context specification) from active users regarding various dependability attributes. This information is stored in the form of a vector known as a requirement vector comprising a triangular fuzzy number set. The triangular fuzzy number is a specific case of fuzzy number theory with a triangular membership function representing the probabilistic distribution. For instance, let ( $\bar{a}$ ) be a fuzzy number with triangular membership function $\delta(x)$ , it is depicted as $\bar{a}=(a_{1},a_{2},a_{3})$ where $a_{1}$ , $a_{2}$ and $a_{3}$ are real numbers such that $a_{1}\leqslant a_{2}\leqslant a_{3}$ . FAM, on the other hand, provides an interface for acquiring feedback from passive users regarding unknown items. Such decisions allow the presented model to evaluate user-similarity and generate a recommendation for the active user with high accuracy. This procedure of information acquisition is performed during the recommendation process from passive users. In heterogeneous scenarios, such decisions can also be utilized by item manufactures/developers for enhancing their product from qualitative perspectives. Each feedback rating is also stored in the form of triangular fuzzy numbers representing a form of a social-tag vector. Three-dimensional feedback matrix is used as a representation tool for feed-backs, in which one dimension represents dependability parameters, while the second indicates different users and third depicts various items under consideration. An instance of such kind of matrix is represented in Fig. 3. In the view of providing an active-user decision for the requirement and effective feedback generation from passive users, triangular fuzzy numbers are associated with linguistic terms. The same can be depicted in Table 5. Such a rating allows efficient quantification of the user’s response to the parameters. Specifically, seven sets are formed for allowing users ratings: Very Low (1), Low (2), Low-Medium (3), Medium (4), High-Medium (5), High (6), Very High (7). However, it is important to understand the representation of such rating values for further evaluation. Referring to Fig. 2 and Table 4, with respect to User U(1), the social-tag rating in User-Item matrix for Item I(1) i.e. $<$ TAG 11 $>$ is fuzzy number set comprises of ordered fuzzy numbers i.e. $<$ T11, T12, T13, T14, T15 $>$ as shown in feedback matrix. T11, T12 $\ldots$ so on are the triangular fuzzy numbers as depicted in Table 4.

Table 5

Linguistic terms and fuzzy values

Linguistic terms	Triangular fuzzy values
Very low	(0,1,2)
Low	(1,2,3)
Low-medium	(2,3,4)
Medium	(3,4,5)
High-medium	(4,5,6)
High	(5,6,7)
Very high	(6,7,8)

Figure 3.

Instance of feedback matrix.

4.2 Social similarity evaluation

Users are categorized as similar or different based upon the similarity in taste, favor, choices and other parameters. In practical-world scenarios, these factors act as important influencing parameters for determining user’s mutual relationships. In the current research, social similarity aims at identifying the degree of similarity between the active user and passive users. For this purpose, two types of similarity are computed in this study, namely: Social-Tag based similarity and Profile-based similarity. Both types of similarities are then balanced by a probabilistic factor $\beta$ , which is pre-adjusted by the system administrator depending upon the user’s social information. This similarity measure is used to form a cluster of user-friends involving active users and similar passive users.

(a) Social-Tag based similarity evaluation

Social-Tag based similarity measure is evaluated on the basis of feedback values that users have assigned to items belonging to Known item set. In order to determine this kind of similarity, Euclidean Fuzzy Similarity is utilized due to its efficiency and feasibility in computation. For instance, let X be a dependability attribute of known item. $\bar{F}_{a}^{X}$ and $\bar{F}_{p}^{X}$ are the feedback triangular fuzzy number values of active user and any passive user respectively for attribute $X$ . Henceforth, $\bar{F}_{a}^{X}=(f_{a}^{1},f_{a}^{2},f_{a}^{3}$ ) and $\bar{F}_{p}^{X}=(f_{p}^{1},f_{p}^{2},f_{p}^{3}$ ). Euclidean fuzzy similarity for $X$ is defined by Eq. (2).

$\displaystyle\textit{Sim}{}_{E}{}^{X}[\bar{F}_{a}^{X},\bar{F}_{p}^{X}]=\frac{1% }{\sqrt{3}}\left(\sum_{i=1}^{3}\left|f_{a}^{i}-f_{p}^{i}\right|\right)^{1/2}$ (2)

Moreover, given an item with two sets of feedback fuzzy numbers for different attributes of dependability, $\bar{F}_{a}=[\bar{F}_{a}^{X(1)}\bar{F}_{a}^{X(2)}\bar{F}_{a}^{X(3)}\bar{F}_{a}% ^{X(4)}\bar{F}_{a}^{X(5)}]$ and $\bar{F}_{p}=[\bar{F}_{p}^{X(1)}\bar{F}_{p}^{X(2)}\bar{F}_{p}^{X(3)}\bar{F}_{p}% ^{X(4)}\bar{F}_{p}^{X(5)}]$ , each attribute feedback of active user $\bar{F}_{a}^{X(i)}$ is compared with the corresponding passive user feedback $\bar{F}_{p}^{X(i)}$ . Furthermore, friend similarity for specific item i is evaluated based on Eq. (2) and considering every attribute for dependability. This similarity measure is represented by Eq. (4.2).

$\displaystyle\textit{Friend-Sim}_{E}{}^{i}[a,p]=\frac{1}{5}$ (3) $\displaystyle\qquad\left(\sum_{i=1}^{5}\textit{Sim}_{E}{}^{X}[\bar{F}_{a}^{X(i% )},\bar{F}_{p}^{X(i)}]\right)$

Above equation denotes one kind of similarity between active user and passive user by evaluating Euclidean fuzzy similarity for various feedbacks. As the Euclidean similarity gets larger, the similarity between two users increases. Let there are $N_{1}$ items that form the Known item set between users, Normalized overall similarity between two users is estimated by using Eq. (4.2).

$\displaystyle\textit{N-Friend-Sim}_{E}[a,p]=\frac{1}{N_{i}}$ (4) $\displaystyle\qquad\left(\sum_{i=1}^{N‘}\textit{Friend-Sim}_{E}{}^{i}[a,p]\right)$

(b) Profile-based similarity evaluation

Information about different user profiles is acquired by social network analytics to represent a form of social influence. As stated earlier, different users have heterogeneous preferences, taste, and favours depending upon their respective social world. These social parameters provide certain type of information about characteristics of the users regarding items. For instance, it is more likely that two computer researchers prefer anti-virus in their laptops rather than graphic card. Therefore, certain kind of similarity is evaluated by such statistics through social network-based information. The similarity metrics used in this paper for determining this type of measure is Jaccard similarity. The same is depicted by Eq. (5). In this, the function, $\phi(i)$ , depicts the profile set of user $U(i)$ about specific item.

$\displaystyle\textit{P-Friend-Sim}_{J}[a,p]=\frac{\left|\varphi\left(a\right)% \cap\varphi\left(b\right)\right|}{\left|\varphi\left(a\right)\cup\varphi\left(% b\right)\right|}$ (5)

(c) Net similarity

Based on the two types of similarities evaluated, a balanced similarity measure between two different users is established. In the evaluating Eq. (6), the balancing factor $\beta(0<\beta<1)$ is predetermined by the system administrator. Moreover, such feasibility of balancing eliminates the conventional sparsity problem to a great extent.

$\displaystyle\textit{Similarity}[a,p]=(\beta)\textit{N-Friend-Sim}_{E}[a,p](1-% \beta)\textit{P-Friend-Sim}_{J}[a,p]$ (6)

4.3 Similar friend classification

In practical scenarios, the similarity between the real-world user and his friends represents similarity in taste, flavor, choice, and effective decision-making. Therefore, it is important to evaluate passive users on the similarity scale for predicting the attributes of an unknown item-set. The similarity values determined in the previous section between users are analyzed to form a similar group of friends. In other words, group formation results in determining those users that are more similar to the active user. This classification of passive users as similar friends aids in evaluating the missing information about items that belong to Unknown Set with respect to the active user. Though there are many similarity algorithms available for the formation of similarity groups, in this study however, it is performed using the k-Nearest Neighbour (k-NN) classification algorithm due to its simplicity and accuracy. It is a classical algorithm for generating similar groups or items that are utilized by various researchers for determining efficient similarity measures in numerous applications. Basically, it is based on comparing similarity values between the active user and other passive users. k-NN algorithm filters out $k$ the number of similar friends that are further considered for dependability evaluation. The detailed algorithm (Pseudo-code 1) mentioned generates a neighbor set in terms of similar friend group formation from an available friend list of active users. Referring to Fig. 2, User U(2), User U(4) and User U(5) are the available friends for active User U(1), out of which the nearest neighbor set is formed and similar friends are generated.

Pseudo-code 1: k-Nearest Neighbour Algorithm for Neighbour Set Generation
Input: Given the similarity value set $S=(S_{1},S_{2},S_{3}\ldots)$ , Active user a, Passive user set $P=(U_{1},U_{2}\ldots U_{p})$ .
$k$ denotes Number of Neighbours, Neighbour set $N$ .
Step 1: Initialize $N=\phi$
Step 2: For each user $i$ belonging to passive user set $P$ ,
Do
Step 2.1: If number of users in neighbour set $N$ is less then $k$ , then insert $i$ user to $N$
Step 2.2: If there exist any user $u\in N$ such that Similarity [ $a, u$ ] $\geqslant$ Similarity [ $a, i$ ]
Then Update $N$ by removing user $u$ from N and inserting user [ $i$ ] in $N$
Output: Neighbour Set $N=(U_{1},U_{2}\ldots U_{k})$ for active user

4.4 Dependability evaluation

After the determination of various passive users that are similar to an active user, the proposed model then uses their feedback values for estimating ratings related to unknown item-set of active users. In other words, the system evaluates a proxy rating of similar passive users of unknown items for the active user. The influence of passive users on the rating of unknown items is dependent on the similarity measure that they have with the active user. The feedback provided by the passive similar users, which have been stored in the database, aids in computing the final item rating over the dependability scale. Moreover, since dependability is composed of five attributes, each attribute is estimated in parallel by the system and final dependability, in the form of a triangular fuzzy number, is estimated for consideration by the active user. In scenarios where numbers of items are very large, parallelism can be achieved using Map-Reduce functions in a cloud computing environment. Furthermore, in the evaluation of dependability, fractional values are computed to a near integer. For instance, the value of 3.123 is considered as 3 and therefore, the corresponding triangular fuzzy number is (3, 4, 5). Suppose $\bar{F}^{X}$ be the feedback rating in the form of fuzzy number by user p to attribute X for specific unknown item $i$ . The estimated feedback ratings (Tag value) for attribute $X$ of item I(i) according to the active user U(a), is computed according to Eq. (4.4).

$\displaystyle\textit{Tag}(a,i^{X})=$

(7) $\displaystyle\quad\textit{Avg}.(\bar{F}^{X}_{p},i^{\prime})+\frac{\makecell{% \sum_{1\subseteq i\subseteq N_{2}}}{\sum_{P=1}^{k}\textit{Similarity}[a,p]}}$

where Avg. ( $\bar{F}^{X}_{p},i^{\prime})$ is the average rating by the user for attribute $X$ . Therefore, when an active user requires a recommendation for certain unknown item $i$ , the system evaluates the rating for each attribute for that item. In other words, Tag ( $a,i^{X}$ ) $=$ Null for all $X$ is evaluated. In this way, the system is able to determine the dependability vector for all items that belong to the Unknown Set.
4.5 Recommendation module

This module aims at providing an optimal list of items based on specifications provided by the user. User-specification for various levels of dependability attributes are acquired in the initial phase of the system. This information is quantified in the form of a triangular fuzzy number set represented by the requirement vector as mentioned earlier. The ratings for various available items are also stored in the form of a fuzzy vector known as a form of social-tags (feedbacks). The recommendation is generated based on the optimal similarity between requirement-vector and tag-vector. In other words, let suppose $\bar{R}=$ $<\bar{R}^{X(1)},\bar{R}^{X(2)},\bar{R}^{X(3)},\bar{R}^{X(4)},\bar{R}^{X(5)}>$ be the requirement vector where $X(i)$ be the $i^{\rm th}$ attribute for dependability. This vector is compared with the vectors $\bar{F}(i)=<\bar{F}^{X(1)},\bar{F}^{X(2)},\bar{F}^{X(3)},\bar{F}^{X(4)},\bar{F% }^{X(5)}>$ (where $1\leqslant i\leqslant N$ ) for similarity analysis, thereby generating a sorted list of item recommendations. This problem of decision-making regarding a recommendation is reduced to an optimization problem where the system generates an optimized list of items to the intended user. In this study, the system utilizes the proposed algorithm to predict top-k similar items from an available list of items.

Definition 3. Let F be the set of available feedback fuzzy vectors [ $F_{1},F_{2}\ldots F_{N}$ ] corresponding to various items such that $F_{k}\in F$ , and $R$ be the requirement vector, then top-k similar vectors is defined as $\forall F_{m}\in F$ , top-k ( $F_{m},R$ ) $=$ [ $F_{R1},F_{R2}\ldots F_{Rk}$ ], where $F_{Rk}$ is $k^{\rm th}$ similar vector to $R$ .

In order to generate ranking in various items for generating recommendation, the system incorporates priority list data structure. That is, more similar items are stored with high priority and least similar items are tagged with low priority. Moreover, vectors are compared on optimal similarity function for fuzzy number based on Euclidean distance. Optimal similarity function is described in Eq. (4.5).

$\displaystyle\textit{O-Sim}[\bar{F}_{m},\bar{F}]=$

(8) $\displaystyle\quad\frac{1}{5}\sum_{k=1}^{5}[\textit{Sim}_{E}{}^{X}(\bar{F}^{X(% k)},\bar{R}^{X(k)})*w^{X(k)}]$

where $w^{X(k)}(0<w^{(X(k))}<1$ ) is the weight associated with each attribute of the dependability, depicting importance of the specific attribute. The various steps for generation of top-k list of recommended items are mentioned in the form of Pseudo-code 2.

Pseudo-code 2: Recommendation List Generation

Input: Given requirement vector from active user and Feedback vectors for $N$ items.

Priority List $L$ for storing item list. $X\in$ [Av, Ma, Re, Sa, Se]

Step 1: Initialize $L=\phi$

Step 2: For items 1 to N Repeat Steps 2.1 to 2.4

Step 2.1: For Passive users 1 to p, Determine O-Sim [ $\bar{F}_{m}$ , $\bar{R}$ ]

Step 2.2: For Min (O-Sim [ $\bar{F_{m}}$ , $\bar{R}$ ]), Add $i$ to $L$ with highest available priority.

Step 2.3: $L=L+1$

Step 2.4: $N=N+[i]$

Step 2.5: Repeat until $N=$ Null or $L=K$

Output: Priority List $L$ which includes user-desired items with highest priority to most recommended item.

5. Experimental evaluation

Pseudo-code 2: Recommendation List Generation
Input: Given requirement vector from active user and Feedback vectors for $N$ items.
Priority List $L$ for storing item list. $X\in$ [Av, Ma, Re, Sa, Se]
Step 1: Initialize $L=\phi$
Step 2: For items 1 to N Repeat Steps 2.1 to 2.4
Step 2.1: For Passive users 1 to p, Determine O-Sim [ $\bar{F}_{m}$ , $\bar{R}$ ]
Step 2.2: For Min (O-Sim [ $\bar{F_{m}}$ , $\bar{R}$ ]), Add $i$ to $L$ with highest available priority.
Step 2.3: $L=L+1$
Step 2.4: $N=N+[i]$
Step 2.5: Repeat until $N=$ Null or $L=K$
Output: Priority List $L$ which includes user-desired items with highest priority to most recommended item.

This section provides an experimental implementation of the proposed fuzzy-inspired dependability recommendation framework. The presented system is deployed in a real-world scenario of On-line Mobile-Phone purchase for performance evaluation, thereby validating the practical applicability of the system. Moreover, since dependability is a universal qualitative aspect that every user desire from a real-world entity, the scenario selected for system deployment is highly relevant in practical situations. In addition to this, the proposed system is experimentally analyzed to determine three important objectives i.e.

1)
To identify the temporal aspects of similar group generation by k-NN Algorithm
2)
To verify recommendation generated by the proposed framework in terms of different statistical parameters.
3)
To determine overall system stability when the number of items is altered.

The first objective determines temporal effectiveness in the estimation of group similarity depending on the number of selected as passive users. The second objective, however, evaluates the efficiency of the system based on various statistical parameters such as Sensitivity, Precision, Accuracy and F-measure. The third objective provides an estimation of stability that the proposed framework persists. Since the system is enveloped around dependability, it is vital to define mobile-phones over its various attributes. Table 6 gives an overview of different dependability attributes defined for a mobile-phone.
5.1 Simulation data set

As a premier requirement of the system, two different data sets are acquired from the real world namely: User Data set and Item Data set.

(a) User data set

For acquiring information about users and his friends, a graduation class with nearly 120 students having an account on social-networking website Facebook.com is selected and the corresponding friend list is obtained. The friend list with respect to the user acts as passive users. However, the selection of the user (active) is constrained by two parameters.

1)
The student must be activated for at least 1 year.
2)
The student must have at least 50 friends in his profile.

After implying these constraints, nearly 93 students are finally short-listed.

(b) Item dataset

The mobile-phone dataset includes a list of various available mobile-phones in the market. This type of data is acquired from an online shopping portal Flipkart.com. Affordable range of Rs. 10000-Rs. 18000 is randomly selected and a list of 398 mobile phones is obtained when last accessed, out of which popular 53 were finally selected for dependability generation. This dataset includes mobile-phones of different brands including Sony, Samsung, HTC, and LG.

Based on the proposed methodology, the experimentation simulations were performed in different steps. The initial task of the presented system is to acquire linguistic data about dependability attributes from both types of users (active and passive). Specifically, nearly 12527 datasets were accumulated and stored in Amazon EC2 cloud storage for further analyzation.
5.2 Evaluating parameters

The proposed mechanism of RS is analyzed at two stages during its implementation for realizing the aforementioned objectives. In the initial stage, group similarity is evaluated for different passive users. The group generation is analyzed over a temporal basis in comparison with other state-of-the-art classification techniques. In addition to this, for the purpose of realizing the second objective, numerous statistical measures are incorporated. These statistical parameters, namely Precision ( $p$ ), Sensitivity ( $s$ ) and Accuracy ( $\rho$ ) are the main evaluating parameters for analyzing the recommendation performance of the system. In addition to these, F-measure ( $f$ ) is also incorporated for determining the composite effect of the precision and sensitivity statistics. Table 7 represents the mathematical formulation for evaluation of the statistical parameters and corresponding symbolic terminology. Table 8 shows the description of various symbols that are utilized in evaluating precision, sensitivity, accuracy and f-measure.

Table 6
Dependability in terms of mobiles phones

Attributes	Mobile phone descriptions
Availability	High throughput in multitasking
Maintainability	Cost of maintenance for repair
Reliability	Hardware and software capability
Safety	Hardware durability
Security	Software crash resistance

Table 7

Dependability in terms of mobiles phones

Parameter	Mathematical formulae
Precision (p)	$\frac{a}{a+c}$
Sensitivity (s)	$\frac{a}{a+b}$
Accuracy ( $\rho$ )	$\frac{a+d}{a+b+c+d}$
F-measure (f)	$\frac{2ps}{p+s}$

Table 8

Symbol terminology

Symbol	Description
a	Active user agrees with the generated recommendation
b	Active user disagrees with the generated recommendation
c	System did not generate but active user accepts
d	System did not generate and active user did not accept

Table 9

Details of datasets for statistical evaluation

Number of items	a	b	c	d	Total
3	504	45	40	7	596
5	837	74	66	13	990
7	1172	103	92	19	1386
9	1506	133	119	24	1782
11	1841	163	146	28	2178
13	2176	193	172	33	2574
15	2511	222	199	38	2970

5.3 Linguistic data acquisition

User-specification related data and feedback data about different mobile-phones are obtained from various users in the form of linguistic terms. Details of different data sets for a heterogeneous number of mobile-phones have been provided in Table 9. For acquiring such data, an on-line questionnaire is conducted in real-time using a data acquisition tool. Feedbacks obtained from passive users are stored in cloud storage and further analyzed by IBM SPSS Statistics. The implementation procedure is performed in continuous successions with the variable number of passive users forming active user-specific similarity groups.

5.4 Similarity group generation results

The similarity in user-friends and group formation is identified in real-time using the k-NN algorithm. In order to analyze the results generated by the proposed system on the basis of temporal efficiency, two cases are considered.

Case 1: Group formulation is performed using Bayesian Classifier instead of k-NN for a variable number of users in real-time for the current scenario.

Case 2: In this case, group formulation is performed by using Neural Network (NN) for different users in real-time.

Table 10
Temporal delay

Number of items	$K$ -NN	Bayesian	Neural network
3	2.63s	4.56s	6.69s
5	4.56s	8.23s	9.89s
7	6.63s	12.01s	19s
9	8.23s	16.66s	23.36s
11	9.98s	18.25s	28.36s
13	10.23s	20.23s	32.69s
15	12.25s	25.36s	40.02

Table 11

Statistical parameters

Stats	Precision (%)			Sensitivity (%)			Accuracy (%)			F-Measure (%)
items	Proposed	CBF	CF	Proposed	CBF	CF	Proposed	CBF	CF	Proposed	CBF	CF
3	92.07	91.98	89.09	91.80	90.69	90.02	93.69	92.60	92.36	91.23	90.36	89.23
5	94.03	93.02	91.59	95.25	94.36	93.78	96.36	95.25	94.36	95.69	92.36	92.65
7	96.78	95.69	94.56	95.98	95.00	94.68	92.36	91.98	90.01	96.35	92.50	91.07
9	98.25	97.25	92.60	95.60	94.56	93.05	94.56	92.69	89.36	90.26	87.50	88.36
11	92.36	89.96	88.88	91.02	90.23	89.23	91.01	90.22	90.88	89.03	88.59	87.99
13	90.25	86.05	83.05	92.36	90.36	88.89	90.25	88.87	85.69	92.99	90.25	92.88
15	93.60	92.25	92.02	95.69	92.36	91.05	96.05	95.69	94.025	90.99	89.15	88.56

Figure 4.

Classification results.

It is important to mention that in the above two scenarios only the classification technique adopted for group formation is changed while the rest of the system is identical to the proposed framework. As mentioned earlier, the experimental simulations are performed in succession with the different number of users selected for formation of the similar groups. Results depicted in Fig. 4 shows that, in the current scenario, k-NN is temporally efficient in similar group generation as compared to Bayes Classifier and Neural Network Classifier. Specifically, k-NN registers average time of 12.32 seconds for generation of similar groups in comparison to 14.56 seconds by Bayes Classifier and 15.27 seconds by Neural Network. An increase in the number of users increases the time information from similar groups. The time consumed information of groups using k-NN is the analyzation time for the similarity between the active user and passive users. Moreover, the simplicity of applying the k-NN algorithm in the current perspective makes it feasibly efficient in the generation of similar groups. However, in other comparative classification techniques, due to complex underneath structures of evaluation, the running time of the classification is impacted and thereby, overall system execution time is increased. The detailed analysis of the results are provided in Table 10. Henceforth, based on these results, it is concluded that the presented approach with the k-NN classifier is more efficient temporally in comparison to other state-of-the-art classifiers.

Figure 5.

Overall system stability.

5.5 Statistical results

Recommendation evaluation is performed in terms of various statistical parameters, namely Precision, Sensitivity, Accuracy, and F-measure. These parameters are evaluated depending upon user-choice in the purchase of a corresponding Mobile-phone. For comparative analysis, state-of-the-art techniques are used namely Collaborative Filtering (CF) and Content-based Filtering (CBF). Results presented in Table 11 show the parametric values obtained under different scenarios with a variable number of users. Specifically, higher precision results (averaging to 94.03%) are obtained in comparison to other recommendation techniques of CBF (averaging to 93.87%) and CF (averaging to 92.78%) when the numbers of passive users are more, thereby depicting high system performance. Moreover, results also depict that high values of sensitivity (averaging to 94.77%) and accuracy (averaging to 95.71%) were obtained by the proposed system in comparison to CBF (averaging to 94.12%) and CF (averaging to 93.18%). Furthermore, based on the presented results, it is depicted that the evaluated values of f-measure (averaging to 94.87%) which is better in comparison to CBF (averaging to 91.69%) and CF (averaging to 92.97%). As far as experiment implementation is concerned, the accuracy achieved by the system is highly acceptable in practical scenarios. Moreover, observing the results obtained after implementation with the variable number of users, it presents analogousness to the practical situations. In other words, a greater number of users involved in recommendation generation procedure (more feedbacks), results in high efficacy in recommending the most suitable item corresponding to user-specification. Therefore, compiling the aforementioned statistical results, the proposed framework is validated over accuracy, precision and sensitivity scale. Moreover, in the process of analyzing the objectives from various procedural aspects, the proposed methodology depicts its high relevancy for the applicability domain.

5.6 Overall system stability

In addition to the above-mentioned results, the proposed system is also evaluated over stability measures. This measure depicts the variation in system performance over a variable number of items. In general, it is measured in terms of Mean Absolute Shift (MAS). MAS provides a statistical metric to measure the degree of divergence from the optimal value when invalid input is provided to the system. Consequently, a lower value (0.00) of shift implies better system performance, while higher value (0.10) of shift depicts poor performance. During the implementation of the proposed system over the different numbers of mobile phones, intentional invalid user-feedback was fed into the system. In addition to this, the number of mobile phone data was increased to 178 for the determination of OSS. Based on the results, it was observed that when the number of mobile phones was 50 and invalid inputs for certain datasets are fed to the system, MAS is registered as 0.05 value. Similarly, when the number of mobile phones is increased to 60, the value of 0.06 is obtained for MAS. For large number of mobile phones numerating to 80, 105, 125, 152, 175, and 178, MAS acquires the value of 0.07, 0.04, 0.075, 0.04 and 0.05 respectively. Based on these results obtained by deploying the proposed system with a variable number of items, it shows that the presented framework is highly stable even if the numbers of items are increased or decreased. Figure 5 shows the result in implementing the proposed system in the scenario of mobile-phone purchase averaging to 0.044. Henceforth, it is concluded that the system is highly efficient and accurate in generating recommendations based on user-specified dependability. Moreover, the fuzzy nature of the data acquisition, representation, and computation provides significant enhancements and flexibility to the proposed system.

5.7 Comparative analysis

This section performs a comparative analysis between the proposed technique of the recommender system and other state-of-the-art recommendation mechanisms. As mentioned earlier, the conceptualization of dependability is based on several parameters. However, this paper presents certain novel aspects in comparison to other related studies for depicting the overall utility of RS. Specifically, a comparison is made with five state-of-the-art studies on the basis of various parameters. These studies include Qian et al. (2019) [5], Mohamed et al.[41], Alhabashneh et al. [42], Katarya and Verma [43] and Almohammadi et al. [44]. A summary of the comparative analysis is provided in Table 12.

5.7.1 Comparison parameters

For the purpose of comparison, ten parameters have been identified in the current study. These include Major Contribution, Recommendation Technique Used, Domain of Application, Incorporation of Social Network Analysis, Dependability Parameters, Similarity Metric Used, Classification Technique Used, Fuzzy Set Values, Generation of Recommender List, and Mathematical Evaluation. Each of These has been discussed ahead in detail.

(a) Major contribution

This parameter is used to determine the work performed by authors in the presented study. In other words, it provides the information about major contributions of the authors, proposed models, algorithms, and technical outcomes that authors have obtained for the generation of accurate recommendations using fuzzy number set theory.

(b) Recommendation technique used

Recommendation Technique provides information about the specific recommendation mechanisms incorporated by authors for proposing RS. In addition to this, the underneath recommendation technique provides a pivotal novelty to the recommendation generation procedure.

(c) Domain of application

RS and its variants have been utilized in different domains for provisioning effective services to users as well as e-commerce industries. In fact, numerous organizations have incorporated RS for the generation of efficient recommendations for business-oriented decision-making purposes. As a result, it becomes important to determine the underneath domain of application for which the research has been carried out.

(d) Incorporation of social network analysis (SNA)

With the widespread utilization of social networking by users around the world, analyzing social behavior through SNA presents numerous beneficial features to any applications. Consequently, the utilization of SNA for generating user-centric recommendations improves the accuracy and efficiency of the overall system. This parameter provides information about the utilization of SNA in the presented research.

(e) Dependability parameters

The determination of dependable items has been a core area of the presented research. This parameter performs an analysis of the proposed domain as compared to other state-of-the-art researches on the basis of dependability aspect.

(f) Similarity metric used

Analysis of user-behavior and similarity generation are important to determine the efficiency of proposed

Table 12
Comparative analysis

Parameter	Qian et al. (2019)	Mohamed et al. (2017)	Alhabashneh et al. (2017)	Katarya and Verma (2017)	Almohammadi et al. (2017)	Proposed model
Reference	[5]	[41]	[42]	[43]	[44]	$-$
Major contribution	Emotion Aware Recommendation System	Fuzzy-logic-based personalized recommendation system that recommends tasks to the engineers based on historical data	New approach for enterprise information seeking and retrieval systems using fuzzy logic to develop a task, user and document profiles	Novel web-based RS which is based on sequential information of the users’ navigation on web pages	Type-2 fuzzy logic-based method with visual RGB-D features to measure student engagement in remote and on-site education	Generating recommendations of dependable items based on SNA based similarity and context-aware based similarity using linguistic fuzzy number set theory
Recommendation technique used	Hybrid technique	Knowledge based RS	Relevance feedback and Fuzzy logic-based RS	Fuzzy C-Means Clustering	Inference based RS	Social Network-based Analysis and Collaborative Filtering
Domain of application	Emotion RS	Field Services RS	Intranet-based information and document seeking and retrieval RS	Web Page RS	E-Learning RS	E-Commerce RS
Incorporation of social network analysis	Yes	No	No	No	No	Yes
Dependability parameters	No	No	No	No	No	Availability, Maintainability, Safety, Reliability, and Security
Similarity metric used	Jaccard Similarity	Probabilistic Parameter	Unified term-visit instance index	Sequential Approach	Engagement Degree	Social-Tag based Similarity and Profile-based Similarity
Classification technique used	Not Available	Not Available	Fuzzy based Technique	Fuzzy C-Means	Fuzzy based Technique	K-Nearest Neighbour Algorithm
Fuzzy set values	No	Fuzzy Logic	If-Then Fuzzy Rules	Fuzzy Logic	Triangular Logic	Triangular Linguistic Values
Generation of Recommendation List	No	No	Yes	Yes	No	Yes
Mathematical evaluation	Yes	Yes	Yes	No	Yes	Yes

RS. Henceforth, this parameter provides information for the type of similarity metric incorporated in the different researches in comparison to the proposed model.

(g) Classification technique used

For the purpose determination of similar items by RS for analyzing dependability, classification techniques reduce the solution space of items and enable early decision-making. Henceforth, it is indispensable to compare the underneath classification technique used by authors in different state-of-the-art researches.

(h) Fuzzy set values

Since the proposed studies incorporate a triangular fuzzy number set theory for analyzing the item-dependability, the comparison of fuzzy-logic presents another crucial factor for determining the novelty of presented research.

(i) Generation of recommender list

Recommendation list generation provides information about the determination of multiple dependable items simultaneously and ranking them according to utility and user-specific similarity. Henceforth, this parameter is important for analyzing the effectiveness of the proposed framework.

(j) Mathematical evaluation

The mathematical evaluation provides an in-depth analysis technique of the proposed research in comparison to other studies. Moreover, since the proposed study incorporates fuzzy-number theory, it becomes necessary to assess the proposed model mathematically and perform comparative analysis.

6. Conclusion

Recommender systems (RS) have been progressively developed over the last three decades. Moreover, with developments in ICT and Web 3.0, generation of recommendation has been effectively enhanced with the utilization of social network information, and contextual information. The work presented in this paper focusses on developing an efficient RS by incorporating hybrid recommendation techniques and suggesting dependable items to the user. The proposed framework comprises of four modules, namely User Interface, Similarity Measurement and Filtering, Dependability Evaluation, and Recommendation Module. SNA is used for evaluating item-specific dependability and determination of social similarity. Moreover, data analysis of contextual information and generation of recommendations is performed in terms of triangular-fuzzy number sets. The utilization of fuzzy logic enhances the feasibility of parameter quantification, which are then utilized in recommendation generation procedures. Furthermore, the performance of the proposed framework is analyzed in the On-line Mobile-phone purchase scenario. Results obtained upon simulation show that the presented dependability RS is highly efficient in the generation of accurate recommendations to the users.

Footnotes

Conflict of interest

The authors declare that they have no conflict of interest with anyone.

References

Sun

Han

Huang

Wang

Zeng

Wang

, et al. Recommender systems based on social networks. Journal of Systems and Software. 2015; 99: 109-119.

Grandison

Sloman

. A survey of trust in internet applications. IEEE Communications Surveys & Tutorials. 2000; 3(4): 2-16.

Adomavicius

Tuzhilin

. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering. 2005; 17(6): 734-749.

Resnick

Varian

. Recommender systems. Communications of the ACM. 1997; 40(3): 56-58.

Qian

Zhang

Peng

. EARS: Emotion-aware recommender system based on hybrid information fusion. Information Fusion. 2019; 46: 141-146.

Lassila

Hendler

. Embracing “Web 30”. IEEE Internet Computing. 2007; 11(3).

Hendler

Berners-Lee

. From the Semantic Web to social machines: A research challenge for AI on the World Wide Web. Artificial Intelligence. 2010; 174(2): 156-161.

Harris

. Web 2.0 Evolution into The Intelligent Web 3.0: 100 Most Asked Questions on Transformation, Ubiquitous Connectivity, Network Computing, Open Technologies,

\ldots

Databases and Intelligent Applications. Emereo Pty Ltd; 2008.

Mao

Wang

Zhang

. Recommender system application developments: a survey. Decision Support Systems. 2015; 74: 12-32.

10.

Tang

Yuan

Tian

Chua

. Adversarial training towards robust multimedia recommender system. IEEE Transactions on Knowledge and Data Engineering. 2019.

11.

Yao

Zhou

Lev

. Synergies Between Association Rules and Collaborative Filtering in Recommender System: An Application to Auto Industry. In: Data Science and Digital Business. Springer, 2019; pp. 65-80.

12.

Celdrán

Pérez

Clemente

FJG

Pérez

. Design of a recommender system based on usersâ€™ behavior and collaborative location and tracking. Journal of Computational Science. 2016; 12: 83-94.

13.

Ben-Shimon

Tsikinovsky

Rokach

Meisles

Shani

Naamani

. Recommender system from personal social networks. In: Advances in Intelligent Web Mastering. Springer, 2007; pp. 47-55.

14.

Ali

Thanh

Van Minh

, et al. A neutrosophic recommender system for medical diagnosis based on algebraic neutrosophic measures. Applied Soft Computing. 2018; 71: 1054-1071.

15.

Bedi

Kaur

Marwaha

. Trust Based Recommender System for Semantic Web. In: IJCAI. vol. 7; 2007, pp. 2677-2682.

16.

Bhatia

Sood

Kaur

. Quantum-based predictive fog scheduler for IoT applications. Computers in Industry. 2019; 111: 51-67.

17.

Nilashi

Ibrahim

Bagherifard

. A recommender system based on collaborative filtering using ontology and dimensionality reduction techniques. Expert Systems with Applications. 2018; 92: 507-520.

18.

Zhang

Yao

Sun

Tay

. Deep learning based recommender system: A survey and new perspectives. ACM Computing Surveys (CSUR). 2019; 52(1): 5.

19.

Sood

Sarje

Singh

. Dynamic identity-based single password anti-phishing protocol. Security and Communication Networks. 2011; 4(4): 418-427.

20.

Yang

Lyu

King

. Sorec: social recommendation using probabilistic matrix factorization. In: Proceedings of the 17th ACM conference on Information and Knowledge Management. ACM; 2008, pp. 931-940.

21.

Woerndl

Groh

. Utilizing physical and social context to improve recommender systems. In: Web Intelligence and Intelligent Agent Technology Workshops, 2007 IEEE/WIC/ACM International Conferences on. IEEE; 2007, pp. 123-128.

22.

Zhou

Lyu

King

. Improving recommender systems by incorporating social contextual information. ACM Transactions on Information Systems (TOIS). 2011; 29(2): 9.

23.

Sood

. SNA based QoS and reliability in fog and cloud framework. World Wide Web. 2018, pp. 1-16.

24.

Hwang

Wei

Liao

. Coauthorship networks and academic literature recommendation. Electronic Commerce Research and Applications. 2010; 9(4): 323-334.

25.

Colombo-Mendoza

Valencia-García

Rodríguez-González

Alor-Hernández

Samper-Zapater

. RecomMetz: A context-aware knowledge-based mobile recommender system for movie showtimes. Expert Systems with Applications. 2015; 42(3): 1202-1222.

26.

Chen

Wang

. Recommender systems based on user reviews: the state of the art. User Modeling and User-Adapted Interaction. 2015; 25(2): 99-154.

27.

Abbar

Bouzeghoub

Lopez

. Context-aware recommender systems: A service-oriented approach. In: VLDB PersDB workshop, 2009; pp. 1-6.

28.

Alhamid

Rawashdeh

Al Osman

Hossain

El Saddik

. Towards context-sensitive collaborative media recommender system. Multimedia Tools and Applications. 2015; 74(24): 11399-11428.

29.

Ricci

Rokach

Shapira

. Introduction to recommender systems handbook. Recommender systems handbook. Springer, 2011; pp. 1-35.

30.

Chu

. A social network-based recommender system (SNRS). In: Data mining for social network data. Springer, 2010; pp. 47-74.

31.

Domingos

Richardson

. Mining the network value of customers. In: Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining. ACM. 2001; pp. 57-66.

32.

Liu

Lee

. Use of social network information to enhance collaborative filtering performance. Expert Systems with Applications. 2010; 37(7): 4772-4778.

33.

Zhou

Liu

Lyu

King

. Recommender systems with social regularization. Proceedings of the fourth ACM international conference on Web search and data mining. ACM; 2011; pp. 287-296.

34.

Golbeck

. Computing and applying trust in web-based social networks, 2005.

35.

Palau

Montaner

López

De La Rosa

. Collaboration analysis in recommender systems using social networks. In: International Workshop on Cooperative Information Agents. Springer; 2004; pp. 137-151.

36.

Chou

Lin

. A social recommender mechanism for location-based group commerce. Information Sciences. 2014; 274: 125-142.

37.

Kim

Ryu

. A group recommendation system for online communities. International Journal of Information Management. 2010; 30(3): 212-219.

38.

Chen

Cheng

Chuang

. A group recommendation system with consideration of interactions among group members. Expert Systems with Applications. 2008; 34(3): 2082-2090.

39.

Dao

Jeong

Ahn

. A novel recommendation model of location-based advertising: Context-Aware Collaborative Filtering using GA approach. Expert Systems with Applications. 2012; 39(3): 3731-3739.

40.

Mobasher

Burke

Bhaumik

Williams

. Toward trustworthy recommender systems: An analysis of attack models and algorithm robustness. ACM Transactions on Internet Technology (TOIT). 2007; 7(4): 23.

41.

Mohamed

Bilgin

Liret

Owusu

. Fuzzy Logic Based Personalized Task Recommendation System for Field Services. In: International Conference on Innovative Techniques and Applications of Artificial Intelligence. Springer; 2017; pp. 300-312.

42.

Alhabashneh

Iqbal

Doctor

James

. Fuzzy rule based profiling approach for enterprise information seeking and retrieval. Information Sciences. 2017; 394: 18-37.

43.

Katarya

Verma

. An effective web page recommender system with fuzzy c-mean clustering. Multimedia Tools and Applications. 2017; 76(20): 21481-21496.

44.

Almohammadi

Hagras

Yao

Alzahrani

Alghazzawi

Aldabbagh

. A type-2 fuzzy logic recommendation system for adaptive teaching. Soft Computing. 2017; 21(4): 965-979.

Fuzzy-inspired decision making for dependability recommendation in e-commerce industry

Abstract

Keywords

1. Introduction

Table 1 Review of recommender system and applications

Table 2 Limitations of recommender system techniques

1.3 Major contribution

2. Literature review

2.1 Social network-based recommender systems

2.2 Context-aware based recommendation systems

Table 3 Symbolic notation and its description

4.4 Dependability evaluation

Table 6 Dependability in terms of mobiles phones

5.4 Similarity group generation results

Table 10 Temporal delay

5.6 Overall system stability

5.7 Comparative analysis

5.7.1 Comparison parameters

Table 12 Comparative analysis

Footnotes

Conflict of interest

References

Table 1
Review of recommender system and applications

Table 2
Limitations of recommender system techniques

Table 3
Symbolic notation and its description

Table 6
Dependability in terms of mobiles phones

Table 10
Temporal delay

Table 12
Comparative analysis