A Survey on Trust Prediction in Online Social Networks

Level of Trust can determine which source of information is reliable and with whom we should share or from whom we should accept information. There are several applications for measuring trust in Online Social Networks (OSNs), including social spammer detection, fake news detection, retweet behaviour detection and recommender systems. Trust prediction is the process of predicting a new trust relation between two users who are not currently connected. In applications of trust, trust relations among users need to be predicted. This process faces many challenges, such as the sparsity of user-specified trust relations, the context-awareness of trust and changes in trust values over time. In this paper, we analyse the state-of-the-art in pair-wise trust prediction models in OSNs, classify them based on different factors, and propose some future directions for researchers interested in this field.


I. INTRODUCTION
In early human societies, (hunter-gatherer) people realised that to fulfil their needs, they had to interact with each other. Quickly, they found that not all interactions were beneficial for them. For instance, their experiences in trading with other people (traders) were not always satisfactory, and sometimes they were deceived by the traders. At that point, they learned to interact with trustworthy people. Trust can be defined as the 'willingness of a party to be vulnerable to actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party' [1]. 'Trust is necessary in order to face the unknown, whether that unknown is another human being, or simply the future and its contingent events'. 1 Sociologically speaking, 'a complete absence of trust would prevent [one] even getting up in the morning' [2].
There are several applications for measuring trust levels in Online Social Networks (OSNs), including social spammer The associate editor coordinating the review of this manuscript and approving it for publication was Renato Ferrero . 1 https://reviews.history.ac.uk/review/287a detection [3], fake news detection [4], retweet behaviour detection [5], [6], recommender systems [7], [8] and influence spread problem [9], [10]. Trust prediction can be defined as the process of predicting a new trust relation between a pair of users that may not be connected in a social network. In applications of trust, trust relations among users need to be predicted. This process faces many challenges, such as the sparsity of user-specified trust relations, the contextawareness of trust and changes in trust values over time.
Although, there were some minor attempts in providing overviews on trust prediction approaches, they either mainly focus on one particular type of trust prediction approaches (Liu et al. [11] mainly focus on supervised trust prediction approaches) or they have been published several years ago and they may fail to overview the trust prediction approaches that were proposed in recent years [12]- [14]. In this paper, we aim to find the research gaps in the literature of trust, classify the state-of-the-art pair-wise trust prediction models based on how they address the research gaps, and finally suggest some future direction for researchers in this field.
The rest of the paper is organised as follows: Section II-A provides various definitions for the concept of trust from different aspects. Sections II and IV discuss the required preliminaries and challenges of trust prediction process. We present the properties of trust, mechanisms for collecting trust information and the ways to present trust in Sections II-B, II-C and III. We discuss the trust prediction process and the current state-of-the-art approaches in Sections II-E and V. Finally, we present the related studies on the impact of users' personality on trust and suggest some future directions for researchers in Sections V-E and IX, before concluding the paper in Section X.

II. PRELIMINARIES
This section briefly introduces the main concepts of this paper.

A. DEFINITION OF TRUST
With the development of human societies, trust has played an important role in people's lives, including in their relationships, families and their businesses and in social management systems. With the development of science and scientific knowledge, different branches of science that focused on human behavioural analysis and human interaction analysis started to study the concept of trust. Trust has different definitions in different scientific fields. Here, a brief overview is provided on the definition of trust in psychology, sociology, economics and, of particular relevance to the subject of this paper, computer science.

1) TRUST IN PSYCHOLOGY
Schlenker et al. [15] provided a definition for trust: being confident about received information from another party in an uncertain environmental state. Psychologists also define trust as 'the subjective probability by which an individual expects that another performs a given action on which its welfare depends' [14]. Psychologically speaking, an inclination towards trusting others can be considered a personality trait [13]. Moreover, 'trusting behaviour takes place when an individual confronts an ambiguous path leading to a perceived either beneficial or harmful result contingent on the action of another person' [16].

2) TRUST IN SOCIOLOGY
Although in sociology studies, the main focus is on the trust in the society or social relations, some research has also focused on trust at the individual level. At this level, the definition of trust is similar to that in psychology [16]; for example, Sztompka stated that 'trust is a bet about the future contingent actions of others' [17]. At the society or social relations level, sociologists consider trust as a properties of social groups [16] and define it as 'a set of expectations shared by all those involved in an exchange' [18]. Another sociologist defined trust as 'a means for reducing the complexity of society' [2]. A different definition of trust was provided by Seligman [19]: 'trust enters into social interaction in the interstices of systems, when for one reason or another systematically defined role expectations are no longer viable'.
Hence, according to Seligman, if people play their expected roles, we can safely have our own transactions [19].

3) TRUST IN ECONOMICS
In economics, trust is defined as 'the property of a business relationship, such that reliance can be placed on the business partners and the business transactions developed with them' [20]. Economists also conceptualise trust as 'existing when one party has confidence in an exchange partner's reliability and integrity' [21]. Moreover, in online trading environments, where there is a lack of direct interaction with customer and products, 'trust can reduce transaction risks, mitigate information asymmetry and generate price premiums for reputable vendors' [16], [22], [23].

4) TRUST IN COMPUTER SCIENCE
The concept of trust is widely used in computer science. Artz and Gil [24] classified trust related research domains in computer science into four major categories: i) policy-based trust, which covers studies in topics related to network security credentials, security policies and trust languages; ii) reputation-based trust, which includes research on trust in peer-to-peer networks, and grids and trust metrics in a web of trust; iii) general models of trust, encompassing research addressing general considerations and properties of trust and software engineering; and iv) trust in information resources, which focus on trust concerns on the Web, the semantic Web and information filtering based on trust.
Trust also plays a significant role in the online activities of users of platforms such as Online Social Networks (OSNs). Tang et al. [25] provided a popular definition for trust in OSNs: 'Trust provides information about with whom we should share information, from whom we should accept information and what considerations to give to information from people when aggregating or filtering data'. There are many applications for trust in OSNs, including: social spammer detection [3], fake news detection [4], retweet behaviour detection [5], [6] and recommender systems [7], [8]. All these applications require predicting the trust relations among users.

B. PROPERTIES OF TRUST
The properties of trust have been listed as context-specific, dynamic, propagative, subjective, asymmetric and event sensitive [12]. We should keep in mind that trust is a concept that is not closed to the specific OSN studied, because it depends on ethical, social, cultural, historical aspects outside the network as well. This is beyond the property of Context Specific, because it only refers to topics that can be developed within the same social network (e.g. science, arts, politics, sports, etc.). However, our intention in this paper is to analyze features of trust relations that can be found within the OSNs.

1) CONTEXT-SPECIFIC
Trust is a context-dependent notion. A trust relation in one context does not guarantee its existence in another context. VOLUME 8, 2020 2) DYNAMIC Trust is a time-dependent concept. Trusting someone at one point in time does not mean the trust relation will exist at another point in time. Trust relations can change because of new experiences, new behaviours on the part of target users or a shift in interests of either or both users.

3) PROPAGATIVE
'Because of its propagative nature, trust information can be passed from one member to another in a social network, creating trust chains' [12]. As an example, if David trusts Sarah, and Sarah trusts Mathew, there is a trust relation between David and Mathew, whereby David may derive some amount of trust towards Mathew from the strength of the trust relations between David and Sarah, and Sarah and Mathew [12].

4) SUBJECTIVE
Trust is a subjective concept. Being trustworthy in one's mind does not imply a person is considered trustworthy by all others. For instance, suppose David and Sarah are two PhD students in the computer science department, and Mathew is a PhD supervisor and a lecturer in this department. David may believe that Mathew is trustworthy, while Sarah does not. Such differences in opinion arise from people's diverse expectations, biases and interests.

5) ASYMMETRIC
'Trust is typically asymmetric' [12]. In other words, if David trusts Sarah, he may not necessary be trusted by her.

6) EVENT SENSITIVE
Establishing a trust relation may take a great deal of effort and time, but a high-impact event can destroy it [12], [26].

C. COLLECTING TRUST INFORMATION
There are three different sources from which to collect trust information [12]; that is, attitude, experience and behaviour.

1) ATTITUDE
Our attitude is the way we think or feel (positively/negatively) about something. Information about a person's attitude can be captured by their online interactions using a measure such as a Likert scale [12].

2) EXPERIENCE
Experience can refer to the 'knowledge or skill that you get from doing, seeing, or feeling things, or the process of getting this'. 2 In OSNs, users can gain experience information about other users by interacting with them. This experience can be captured by the feedback among users, and better feedback may result in more interactions in future [12]. 2 https://dictionary.cambridge.org/dictionary/english/experience

3) BEHAVIOUR
Human behaviour refers to 'the range of behaviours exhibited by humans . . . [which are] typically influenced by culture, attitudes, emotions, values, ethics, authority, persuasion, coercion and/or genetics' [12], [27]. In OSNs, we may notice a sudden change in the frequency of interaction between two users, or the amount of activity of a user. The first case may indicate that the trust level between those users has decreased. While the second may represent a decline in the user's trust towards the community in which he or she participate [12].

D. ONLINE SOCIAL NETWORKS
Garton et al. [28] widely accepted definition concerning OSNs holds that 'when a computer network connects people or organisations, it is a social network. Just as a computer network is a set of machines connected by a set of cables, a social network is a set of people (or organisations or other social entities) connected by a set of social relationships, such as friendship, co-working or information exchange'. OSNs are relatively new and evolving phenomena on the Web. Users of these online platforms can communicate with others and present themselves through their profiles [29], [30].
Social network analysis is an area of study focusing on OSNs that looks for patterns of relations among people [28]. In OSNs, relations can be characterised by their content (i.e., resources exchanged, such as information), direction and strength [28]. One of the relations on which social network analysis focuses is the trust among people in OSNs. These studies aim to understand why people trust each other and establish trust relations in OSNs, with a view to predicting trust relations among people in OSNs.

E. TRUST PREDICTION IN ONLINE SOCIAL NETWORKS
Trust networks in OSNs are usually sparse [25]. They follow the power law distribution whereby a small number of users account for the majority of the trust relations [25]. As a result, the explicit trust relations among many users in OSNs are unknown [13]. Therefore, to employ trust information in different applications in OSNs (e.g., recommender systems and retweet behaviour prediction), we need to predict unknown trust relations among users. Figure 4 illustrates a simple example of the trust prediction procedure; on the left side, we have some users and their explicit trust relations, as shown by the green arrow and the label '1'. We want to know if there is a trust relation between Sarah and John. A trust prediction approach can be used to predict that the existence of this trust relation.

III. TRUST REPRESENTATION
A pair-wise trust relation ( Figure 1) is a relationship between a source user (trustor) and a target user (trustee) that indicates that the trustor trusts the trustee. With the help of trust, the trustor may seek information from the trustee, to avoid being confused by the huge amount of available data (i.e., mitigated information overload) and to be confident about the credibility of the received information (i.e., increased information credibility) [13].
To denote the naivest notion of trust (e.g., singledimensional trust), one can use a representation similar to On the left side, there is a trust network representation among these users, where the green arrow with the label '1', indicates the existence of a trust relation between two users, and its absence means that there is not any trust relation between them. To the right of this figure, there is a corresponding adjacency matrix, showing the trust network between any two users. In this matrix, '0' represents the lack of trust and '1' illustrates the existence of trust between two users.
However, trust may have multiple dimensions. For instance, trust is a context-dependent concept. Context is the information about the condition of an entity [31]. As an illustration of a single context (focusing on the domain of the trust), consider Sarah, a football player, who trusts her coach in football. This does not necessarily mean that she also trusts her coach regarding music. Hence, to represent trust relations among users in different contexts, we need a representation with more dimensions. As another example, if Mathew trusts Jack (as two users in an OSN) at time T 1 , this does not necessary mean that Mathew will also trust Jack at time T 2 (where T 2 = T 1 + h, and h is a fraction of time). Hence, matrices cannot appropriately represent a multi dimensional trust network. Instead, tensors are one of the most favoured representations for trust relations as they can store data in several dimensions. Figure 3 illustrates an example of representing trust relations in different contexts of trust. In this figure, which demonstrates a single dimension of context (e.g., domain of expertise), there are three contexts of trust (football player, computer scientist and plumber). There are also three users (A, B and C). As shown, there are three trust relations between users in the first context (football player). According to these trust relations, A trusts B, B trusts C and C trusts A as a football player. For representing these trusts relations and other trusts relations between these users in other contexts, we can use a tensor. For instance, Figure 3 shows a three dimensional tensor with two dimensions for representing users' relations and a third dimension denoting the contexts of trust. Since all the mentioned trust relations are related to the football player (context 1), they are stored in the matrix of context 1 of this tensor.

IV. CHALLENGES OF TRUST PREDICTION IN ONLINE SOCIAL NETWORKS
This paper will focus on the following three significant problems in OSNs: sparsity of user-specified trust relations, context-aware pair-wise trust relations and time-aware pairwise trust relations.

A. SPARSITY OF USER-SPECIFIED TRUST RELATIONS
User-specified trust relations are extremely rare [32]. For instance, 'the density of a typical trust network in social media is less than 0.01' [13], [33]. As another example, 'the sparsity of Advogato, Ciao, and Epinions, FriendFeed, and Flixster [frequently used datasets in trust prediction related research], i.e., the ratio of the observed trust relations to all the possible relations, is 0.0011%, 0.0028%, 0.0042%, 0.0041% and 0.0035%, respectively [4], [13], [34], [35]. It is challenging to predict the trust relations well with so limited observed links' [32]. Moreover, trust relations follow the rules of the power law distribution: many trust relations can be accounted for a small number of users and a large number of users participate in only a few trust relations [25]. For any trust prediction approach in OSNs, the number of known user-specified trust relations compared to all possible relations among users is low. This makes the pair-wise trust   prediction problem in OSNs a challenging task; any trust prediction approach should be able to deal with this data sparsity problem.

B. CONTEXT-AWARE PAIR-WISE TRUST RELATIONS
The notion of trust is context-dependent [4], [13]: Trusting someone in one context does not guarantee trusting them in another context [13]. As an example, the context dependency of trust has been investigated by [36] in the collected data from a real-world product review website. 3 In this website there is an option for users to explicitly indicate which users are trustworthy. Tang et al. [25] used this information as the ground truth of their analysis. They considered items' categories (e.g., electronics, sports and entertainment) as the context of trust and reported that: 'less than 1% of users, trust their friends in all categories' and 'on average, people trust only 35.4% of their trust networks for a specific category'.
Hence, people trust each other in certain contexts. Context is the information about the condition of an entity [31]. As an illustration of a single context (focusing on the domain of the trust), consider David who is a PhD student at the Computing Department. He trusts his supervisor in the computer science field; however, he does not necessarily trust him in sports. As a result, predicting pair-wise trust relations with respect to the different context of trust can be a daunting task.

C. TIME-AWARE PAIR-WISE TRUST RELATIONS
Trust values can also change over time. Users can establish new trust relations or eliminate their existing trust relations after a period of time. For instance, if Jack trusts David (as two users in an OSN) at time T 1 , this does not necessary mean that he trusts him at time T 2 (where T 2 = T 1 + h and h is a fraction of time). As another example, David may not trust Sarah at time T 1 , but he could trust her at time T 2 .
Hence, predicting the pair-wise trust relations statically may not be a realistic approach for OSNs. Trust is timesensitive: if John trusts David at time T 1 , this trust relation may change at time T 2 . This can be affected by many factors, such as some new behaviour on David's part or a change in John's interests. Hence, predicting pair-wise trust relations in OSNs dynamically can be a challenging task.

V. TRUST PREDICTION APPROACHES
In this section, we describe related work in four areas: representation of the network, type of prediction algorithms, context-awareness and time-awareness. Finally, we classify the existing pair-wise trust prediction approaches (Table 3).

A. REPRESENTATION OF TRUST NETWORKS
We broadly categorise trust prediction approaches into three categories: graph-based trust models, interaction-based trust models and hybrid trust models [12].

1) GRAPH-BASED TRUST PREDICTION MODELS
Approaches in the category of graph-based trust models are mostly based on the concept of web-of-trust or Friend-of-a-Friend (FOAF). Each user is assumed to have a trust network that contains friends (i.e., social network actors/users) as nodes, with the relationships (i.e., value of their trust relations) among them as the edges [12]. This assumption can be invalid or too strong because, in many online communities, there is either no way to identify a web-of-trust or the connectivity is sparse [37]. Moreover, in some cases, this kind of approach may fail to capture the actual interactions among members [12]. Trust propagation-based [38] and inferencebased [39] methods belong to this category.
Golbeck et al. [38] proposed another trust inference approach based on the FOAF concept that can determine which pairs of users trust each other and on which topic. Similarly, Zhang et al. [40] presented an approach by which the source user accepts the recommendation from similar neighbour nodes (i.e., other users directly connected to the target user). Kim et al. [41] proposed an approach to build a web-of-trust based on the implicit feedback of users in a certain context. Golbeck [42] proposed another trust prediction approach, TidalTrust, also based on the FOAF concept. In TidalTrust, if two neighbours have a high trust rating, it is more likely they would agree on other users' trustworthy levels [12]. Ziegler and Lausen [43] developed another network-based trust prediction model, Appleseed, for use in the semantic Web. They focused on local group trust metrics to improve the efficiency of the trust prediction procedure. Hang and Singh [44] introduced a new trust prediction approach based on the similarity of users' trust networks; they treated the recommendation problem as a graph similarity problem [44]. Zuo et al. [45] proposed a trust prediction framework based on trust chains and a trust graph. This framework can 'calculate trust along a trust chain and evaluate a trust based on a trust certificate graph' [45]. Caverlee et al. [46] developed another trust prediction framework, SocialTrust, focusing on social relationships and users' feedback. SocialTrust also allocates a weight value to feedback according to the PageRank algorithm.
Zhang and Yu [47] designed a semantic-based trust reasoning mechanism for trust prediction in OSNs. They noted that trust is a category-dependent concept and traditional trust prediction approaches required much human effort to predict pair-wise trust relations. They also inferred trust relations by designing a domain ontology and exploiting role-based and behaviour-based reasoning functions [47]. Liu et al. [48] proposed a heuristic approach, called the Heuristic Social Context-Aware Trust Network Discovery algorithm, adopting the K-best-first search for addressing the trust network extraction problem by developing a contextual social network structure and proposing the concept of Quality of Trust Network [48]. Azadjalal et al. [49] proposed a trust-aware recommendation system and use these trust values to improve the accuracy of their model in present of the sparsity of user-item ratings matrix. They propose a model for identifying implicit trust relations. Moreover, they detect the most prominent users based on the Pareto dominance and confidence concepts to use their opinions in their recommendation model. Guo et al. [50] developed a trust-aware recommender system based on a matrix factorization model. Their proposed approach focus on item recommendation rather than rating prediction. They also use the trust values provided by explicit users' feedback. Ghavipour and Meybodi [51] proposed a trust inference method based on aggregation strategy and learning automata. Parvin et al. [52] proposed a collaborative filtering recommender system based on users' trust network and ant colony optimization (ACO) algorithm. They ranked users based on social trust relationships and then, using ACO, they assigned proper weight values to users to identify the similarity levels among users. Jiang et al. [53] presented a slope one algorithm using trusted data and user similarity for designing a collaborating filtering-based recommender systems. Ruan et al. [54] developed a trust inference approach for OSNs using trust's transitivity property. They also proposed a metric for measuring the trust level and its certainty.

2) INTERACTION-BASED TRUST PREDICTION MODELS
Approaches in the previous category may fail to 'capture actual interactions among members. The volume, frequency and even the nature of interaction are important indicators of trust in social networks' [12]. By contrast, interaction-based trust prediction models mainly focus on the interactions among users. Liu et al. [37] proposed a classification approach for trust prediction in OSNs based on the action and interactions of users. A similar approach presented by Nepal et al. [26] proposed a trust prediction model that considers two types of trust: the trust of other users towards a target user and the trust value that a user has towards a community. Adali et al. [55] developed a trust prediction approach focusing on users' communication behaviour and more specifically on conversational trust (i.e., duration and frequency of communication between two users) and propagation trust. Sacco and Breslin [56] proposed a trust prediction approach centering on the subjective trust values of connected users, based on their social interactions [56]. They stated that most of the existing trust prediction approaches are 'propagating known trust values among peers in a trusted network and do not provide measures for asserting a trust value from user interactions between peers' [56]. These approaches only focus on users' interactions and do not consider the social network structure, which may contain important information about users and the type of relations among them.

3) HYBRID TRUST PREDICTION MODELS
Hybrid trust models combine the network-based and interaction-based models. In particular, they simultaneously consider users' previous interactions and the social network's structure [57]. We proposed a trust prediction model called TDTrust [4]. They proposed a set of context factors for capturing contexts of trust relations among users in OSNs. We also mathematically modeled our trust prediction approach, based on three-dimensional tensor decomposition to consider the context of trust directly in their model and to predict trust relations in different contexts of trust. In another study [58], we proposed a new unsupervised approach, SETTrust, which incorporates the social exchange theory. We proposed that a trust relation can be established if the costs of that relation is less than its benefits. Zhang et al. [59] developed a trust link detection scheme. Their approach tens to find subjective trust, reputation, and indirect link between users. They calculate the subjective trust according to the users' previous interactions and assess the users' reputation based on collective objective trust.

B. TYPE OF PREDICTION ALGORITHMS
We now discuss past work from the perspective of the type of algorithms used. We can roughly categorise trust prediction approaches into supervised and unsupervised approaches.

1) SUPERVISED APPROACHES
Liu et al. [37] developed a supervised trust prediction model and a classifier that works with a set of users' features and interactions. Ma et al. [60] proposed a personalised and cluster-based classification trust prediction model that creates user clusters and then trains a classifier for them. Matsuo and Yamamoto [61] focused on a Japanese e-commerce website called @cosme, and became the first to explain the concept of community gravity: a two-way effect of trust and rating. They followed this with a model to formulate the trust prediction and rating prediction problems. Grana et al. [62] introduced a supervised trust prediction approach: a binary classification that focuses on users' reputation. Wang et al. [63] proposed a trust-distrust prediction approach that simultaneously employed Dempster-Shafer theory and neural networks. They also analysed the effects of homophily theory, emotion tendency and status theory in trust relations [63]. Zhao and Pan [64] developed another supervised trust prediction approach: a classifier with a feature set that included several trust-related factors. These features could be demographic features (e.g., age and Gender), profile features (e.g., number of follwoers and followees), numeric representation of textual contents provided by users and etc. [65]. They used the existing trust labels for training their classifier. However, the main shortcoming of these approaches is the fact that because of the sparsity of trust relations in OSNs, they have not enough label data available for their training process. Bachi et al. [66] developed a new trust inference framework to infer trust-distrust relationships. Their approach was based on frequent subgraph mining, signed networks, social balance theory, edge classification and rule-based link prediction [66]. It decomposed a 'trust network into its ego 4 network components and mining on this ego network set the trust relationships' [66].
Korovaiko and Thomo [68] designed a classifier that works with users' provided ratings on product review websites. They analysed the effects of similarities in users' ratings on their trust relations. Borzymek and Sydow [69] focused on analysing graph-based and users' rating-based attributes and employed a C4.5 decision tree-based algorithm to predict users' trust-distrust relations in OSNs. Lopez and Maag [70] proposed a generic trust prediction framework as a multi-class classifier, employing the RESTful web-service architecture and support vector machines technique [71]. We developed a deep classifier for pair-wise trust prediction in OSNs, called DCAT [65]. They proposed some demographic factors and textual contents-based factors for our classifier. To improve the accuracy of DCAT, they also used the word embeddings of users' textual contents.
Zolfaghar and Aghaie [72] developed a supervised time-aware trust prediction approach. They considered the trust prediction problem as a temporal link prediction problem. Their main focus was analysing historical information on the trust relations (or links). Raj and Babu [73] presented a probabilistic reputation feature model as a supervised trust prediction approach. They proposed a framework using reputation features to solve the cold start problem in trust prediction. They also employed the SMOTE-Boost algorithm to establish balanced classes in their datasets [73]. Zhao et al. [74] introduced a trust prediction approach to evaluate the trustworthiness of users and tweets on Twitter, focusing on Twitter data from Latin America. Their approach 'jointly consider users' social and contextual relationships in a Twitter social graph' [74]. Their approach used a novel topic-focused trustworthiness estimator model based on a similarity metric. For instance, if a tweet is similar to trustworthy tweets, it can also be considered trustworthy.
Zhang et al. [75] with the aim of addressing the 'all good reputation' problem, proposed a multidimensional trust prediction approach called CommTrust, which evaluated trust by mining users' feedback comments [75].
Chakraverty et al. [76] introduced a logistic regression-based model that focused on the ratings similarity of users to predict their pair-wise trust relations. Their experimental results is somewhat contradict those of Tang et al. [25]. Chakraverty et al.'s study focused on the implicit similarity and co-rated item-count thresholds, finding low precision, recall and coverage for the similarity threshold and better precision, recall and coverage for the co-rated item-count threshold [76]. Nunez-Gonzalez et al. [77] considered the trust prediction problem as a classification problem. They focused on the reputation features of users, because they believed that their reputation information could be used to evaluate the trustworthiness of a user [77]. Raj and Babu [73] proposed a probabilistic reputation feature model to compute the level of trustworthiness in OSNs by identifying the features that determine a user's trustworthiness level. some researchers focused on employing Bayesian network model in their trust prediction approaches. Denko et al. [78] proposed a trust management approach to assess relationships among devices in pervasive computing environments. This approach enables the devices to evaluate the trustworthiness of other devices even if they did not have enough interactions before. Fung et al. [79] proposed a Bayesian trust management model for host-based detection system (HIDS) and for tracking the uncertainty in evaluating the HIDS's trustworthiness. Sharma et al. [80] presented a trust management framework for pervasive online social networks (POSNs) using concepts of lock door policy and intermediate state management procedure to identify trustworthy and untrustworthy users.

2) UNSUPERVISED APPROACHES
Tang et al. [25] proposed an unsupervised trust prediction model called hTrust. It exploits the homophily effect on the trust prediction procedure by focusing on similar users. In this way, Tang et al. identified similar users based on the users' ratings similarity. They considered three factors for rating similarities: users who rated similar items, users who gave similar ratings for similar items and users who had similar ratings patterns. Wang et al. [81] developed an unsupervised model, sTrust, using social status theory and the PageRank algorithm [82], based on MF. In this approach, if a user has a higher social status in an OSN, he or she is more likely to be trusted by other users.
Guha et al. [83] developed a trust prediction model that propagate trust based on users' trust or distrust relations with others. Golbeck [84] put forward a website called FilmTrust which used trust to produce movie recommendations. Wang et al. [32] proposed a trust prediction approach that, in addition to learning low-rank representations of users, also learned these sparse components of the trust network [32]. Zheng et al. [31] suggested an unsupervised trust prediction model based on the concept of trust transference, to transfer trust between different contexts [31]. Wang et al. [39] introduced an unsupervised trust prediction model to infer trust among users with an indirect connection. Liu et al. [85] proposed a trust inference model, incorporating factors such as residential location and outdegree. Wang et al. [86] proposed a novel trust prediction model, CATrust, for auction websites, using Bayesian inference based on Markov Chain Monte Carlo. More importantly, their model considered the contexts of trust.
Moradi and Ahmadian [87] proposed a trust-aware recommender system, Reliability-based Trust-aware Collaborative Filtering, to address the problem of the accuracy of ratings predictions in recommender systems. This system dynamically extracts trust networks among users based on similarity values and trust statements. Sanadhya and Singh [88] designed a trust prediction approach based on ant colony optimization (ACO), called Trust-ACO, to calculate trust path and trust cycle and identify the most trustworthy path to find trustworthy services [88]. Their approach is based on probabilistic trust rule, social intimacy pheromone. Fazeli et al. [89] proposed a trust prediction approach based on social trust, using MF. They first studied the effect of existing trust metrics in predicting pair-wise trust relations, employing those they deemed most effective in their prediction approach.
Massa and Avesani [90] stated that 'predicting a distrust statement is harder than predicting a trust statement'; however, Tang et al. [91] have proposed an approach to predict distrust in OSNs. Specifically, their approach facilitates computational understanding of distrust. Zhang et al. [92] proposed a context-aware trust prediction approach focusing on 'the ratings of past transactions, the nature of both past transactions and the new transaction' [92]. This approach used transaction context similarities to 'identify and prevent potentially malicious transactions with the value imbalance problem' [92]. Matsutani et al. [93] assumed that the trust prediction problem could be solved in the same way as a link prediction problem. They proposed an approach based on non-negative MF (NMF) methods. This approach 'incorporates people's evaluation of users' activities as well as trust-links and users' activities themselves' [93].
Tang et al. [94] delved into the evolution of trust as a result of interpersonal interactions. They proposed a dynamic MF-based trust prediction approach, called eTrust, which focused on the dynamic preferences of users on product review websites [94]. Huang et al. [95] believed that 'people who are in the same social circle often exhibit similar behaviour and tastes'. They treated the trust prediction problem as a link prediction problem and proposed a joint manifold factorisation method that aggregated heterogeneous social networks to explore 'the user group level similarity between correlated graphs and simultaneously [learn] the individual graph structure' [95]. Moturu and Liu [96] proposed an unsupervised approach for evaluating the trustworthiness of shared content, particularly shared health content. They proposed an approach based on feature identification, for determining the features most relevant to trust and quantification. Yao et al. [97] proposed a trust inference approach based on MF. They addressed the trust prediction problem as a recommendation problem. Their model 'characterizes multiple latent factors for each trustor and trustee from the locally-generated trust relationships'. To improve the accuracy of their approach, they also employed prior knowledge (e.g., trust bias and trust propagation). Huang et al. [98] stated that, since trust matrices are of low-rank, they could consider the trust prediction problem as a recommendation problem. Specifically, they proposed a rank-k matrix completion approach that was robust to noise. Liao et al. [99] developed a ranking system for evaluating users' reputation which they used it in evaluating the trust relationships and social acquaintances of users. Su et al. [100] developed a trust-aware approach for a reliable personalized Quality of Service (QoS) assessment. They employed a beta reputation system to calculate the reputation of users. Next, they identify similar trustworthy users and finally using user-contributed QoS data of these users, they predict the QoS. Ruan et al. [101] proposed a trust-aware approach for increasing the correlation between social media and financial data in the stock market. They collected stock-related data (tweets) from Twitter and they proposed a reputation-based mechanism to identify a firm's Twitter sentiment valence and its stock abnormal returns.

C. CONTEXT-AWARENESS OF TRUST
Existing trust prediction approaches can be classified into two groups based on their consideration of the context of trust: approaches that consider context and those that do not. Before discussing the approaches that fall into these categories, we first discuss the notion of the context of trust as it relates to OSNs.

1) DEFINITION OF CONTEXT
Context, which influences the building of a trust relationship between the trustor and the trustee [102], is multifaceted [31]. In a society, the interactions between two participants can form a context that can provide information such as the time or location of that interaction. Uddin et al. [102] provided a definition for context of trust in OSNs: 'a context is a situation, which influences in the building of a trust relationship between the trustor and the trustee'.

2) CONTEXT-LESS APPROACHES
The context-less approaches do not consider context to predict a trust relation in OSNs. The majority of existing trust prediction approaches can be considered context-less (see Tang et al. [25], Wang et al. [81], Golbeck [84] and Wang et al. [32]). These approaches assume that if John trusts Jack, this means John trusts Jack in all fields of expertise (e.g., electronics, sports, music, movies and science), for a lifetime and in any location. This assumption is too simplistic for real-word scenarios, because people only trust each other in certain contexts [4], [13], [36].

3) CONTEXT-AWARE APPROACHES
Liu et al. [103] and Zhang and Wang [104] highlighted the importance of the context of trust as an essential factor for trust prediction approaches. However, little effort has been made to consider the context of trust for a first class citizen. One exception is Zheng et al. [31], who proposed a context-aware approach that considers both user's properties and the features of contexts. Social trust proposed as a novel probabilistic social context-aware trust inference approach, exploits textual information to deliver better results [39]. In Zheng et al.'s approach, trust is inferred along the paths connecting two users. Thus, if two users are not connected by any path, no trust among them can be predicted. Similarly, Liu et al. [85] developed a context-aware trust prediction approach based on the web-of-trust concept, which considered social context factors, such as users' location, previous interactions, social intimacy degree with other users, existing trust relations and so on. Zolfaghar and Aghaie [105] proposed a supervised context-aware trust prediction approach. They investigated the effects on trust relations of certain social trust factors, such as contextual similarity, users' reputation and relationship-based trust factors.
Zhang et al. [106] proposed a novel context-aware trust prediction approach based on contextual transaction factors, categorised into those relating to service and those relating to transaction [106]. This approach considered the context of past transactions and forthcoming transactions to evaluate the reputation profile of the seller [106]. In another study, Zhang et al. [107] aimed to develop a context-aware trust prediction approach. They designed a data structure to support the Contextual Transaction Trust (CTT) computation in e-commerce environments [107]. They also proposed 'an approach for promptly responding to a buyer's CTT query' [107]. Liu et al. in [108], [109] and [110] noted that 'predicting the trust between two unknown participants based on the whole large-scale social network can lead to very high computation costs' [108]. Hence, they proposed an approach to extract a sub-network of the trust network that contained the most important nodes and trust relations. Since this sub-network extraction problem is a NP-complete problem, they proposed a strong social component-aware trust sub-network extraction model, So-BiNet, to address this [108]. Zheng et al. [111] proposed another solution to 'extract a small-scale contextual network that contains most of the important participants as well as trust and contextual information' [111]. They developed a context-aware trust sub-network extraction model. They also used ant colony algorithm sub-network extraction.
Liu and Datta [112] introduced a new context-aware trust prediction approach based on the Hidden Markov Model (HMM). This approach can dynamically model a user's interactions in OSNs. Rettinger et al. [113] proposed a context-aware trust prediction approach, called the Infinite Hidden Relational Trust Model. They expressed that 'from the truster's point of view trust is best expressed as one of several relations that exist between the agent to be trusted (trustee) and the state of the environment'. Xiong and Liu [114] developed a novel context-aware trust prediction model, PeerTrust, for e-commerce platforms, based on a transaction-based feedback system. They also introduced the factors of transaction context and community context for capturing the contexts of trust relations. Rehak et al. [115] designed a situational (context-dependent) trust prediction approach. They proposed a mechanism that 'describes the similarity among the situations using their distance in a metric space and defines a set of reference contexts in this space to which it associates the trustfulness data'.
Uddin et al. [102] proposed an interaction-based contextaware trust prediction approach, called CAT. They also suggested the concept of context similarity, which can be used for decision making in similar situations [102]. Kim et al. [41] believed that existing trust prediction approaches mostly relied on the web-of-trust concept, which may fail to accurately predict trust relations among users because of the data sparsity problem. They developed a context-aware trust prediction approach focusing on users' expertise and affinity in a particular context (topic). Li and Wang [116] developed a fuzzy comprehensive evaluation based method to evaluate the trustworthiness of a service provider in an upcoming transaction based on the trust ratings in its transaction history. This approach is grounded in context-based trust normalisation, which focuses on 'the familiarity between each rater and the service client of the upcoming transaction' [116]. Wu et al. [117] proposed a linguistic trust model for direct trust relations of group experts in social network group decision making (SN-GDM) using trust/distrust values. Then, they combined the social network trust with the collaborative filtering to propose a comprehensive estimation method for incomplete information. Burt et al. [118] analyzed the well-being, business differences, political views and demographic features of users in the strong ties known in China as guanxi. Their finding illustrate that there is a strong relationship between trust and social network and '' Trust variance is 60% network context, and 10% individual differences''. Li et al. [119] designed a context-aware and trust-aware recommendation based on Gaussian mixture model (GMM). Their assumed that decisions and preferences of users may be affected by their trusted friends.

D. TIME-DEPENDENCY OF TRUST
Although time can be considered one of the elements of context, because of its importance we investigate it more deeply. The literature on time-aware trust prediction in OSNs can be divided into two categories based on the approach taken: static approaches and dynamic approaches.

1) STATIC TRUST PREDICTION APPROACHES
Static trust prediction approaches assume that trust relations among users do not change over time. However, in real-world scenarios, trust relations among people may be terminated at any time for various reasons (e.g., changes in interests, expectations or opinions). The majority of existing trust prediction approaches belong to this category (see Liu et al. [37], Ma et al. [60], Matsuo and Yamamoto [61], Tang et al. [25], Wang et al. [81] Ghafari et al. [4], [58] and Wang et al. [32]).

2) DYNAMIC TRUST PREDICTION APPROACHES
Dynamic trust-prediction approaches can be classified into three main categories: Beta models, HMM-based models and others.
In the Beta models, Beta probability density functions consider reputation and feedback simultaneously (see Ismail and Josang [120]). In another work [121], a decay factor was used to give more weight to recent events based on Recency bias (i.e., a person will remember the most recent events more easily compared to older events). Zhang and Cohen [122] introduced an approach that monitors the dynamic behaviour of an agent based on the concept of time windows. In each time window, the number of successful and unsuccessful transactions is considered.
HMM-based models use HMM to propose dynamic trust prediction models. These approaches are of two main types. The first type focuses on the outcomes of past transactions and observations of HMM [16], [123], [124]. Although these may have better performance compared to the Beta models, they fail to consider contextual information about each transaction [16]. In the second type, researchers seek to consider contextual information about the transactions (see Liu and Datta [112]). Zheng et al. [125] developed a dynamic trust prediction approach based on HMM, which focused on the hidden characteristics of the HMM model as well as the outcomes. They used a service provider's historical transactions to predict its trust level. They considered 'static features, such as the provider's reputation and item price and the dynamic features, such as the latest profile changes of a service provider and price changes' [125]. Malik et al. [126] presented a means of assessing reputation in a service oriented approach for service oriented environments based on HMM. This approach can predict trust-based interactions among Web services.
Falling under the third category of dynamic trust prediction approaches, Cai et al. [127] proposed a MF-based trust prediction model. They incorporated temporal dynamics to model the dynamics of users' preferences. Laifa et al. [128] tested a research model using structural equation modeling and delivered the outputs to an artificial neural network and fuzzy logic model developing their dynamic prediction approach. Liu and Datta [129] designed another dynamic trust prediction approach. They believed that modelling the behaviour of people is challenging as people may change their behaviour strategically to increase their profits [129]. By measuring similarity among the contexts of transactions, they estimated the trustworthiness of a transaction based on previous cases of similar transactions. Although these approaches give outstanding performance in some situations, they may fail when a user's 'behaviour is highly dynamic or is changing strategically' [16].

E. PERSONALITY AND TRUST
Alarcon et al. [130], in investigating the relation between personality and trust, focused on the relations between VOLUME 8, 2020 propensity to trust, the five-factor model [131], trust beliefs and behaviours. Thielmann and Hilbig [132] researched the impact of HEXACON, another trait-based personality mechanism, on trustworthiness by designing three trust games. Their work demonstrated the relation between honesty/humility and trustworthiness, independent of the prior level of trust. Another study by Evans and Revelle [133] considered the trust inventory and personality traits and validated this inventory through an economic task. They discovered that trust can be related to the Extraversion personality trait. Sicora [134] focused on trust among co-workers and workplace leaders and its relationship with two personality models. Their aim was to create greater trusting relationships in organisations [134]. Gerris et al. [135] studied the influence of the Big Five personality traits of couples on their marriages. Solomon et al. [136] studied Twitter users based on the Big Five personality model [131] and the Schwartz sociological behaviour model [137] to understand the psycho-sociological homophilic nature of personal networks. We proposed a pattern-based word embedding technique, personality2vec [138] as a novel data analytics pipeline that enables analysis of users' personality patterns and behavioural disorders, based on their activities in OSNs. We also proposed to use domain knowledge to design cognitive services to automatically contextualise raw social data and prepare them for behavioural analytics.
Although there are a rich body of knowledge in the literature of trust related studies in psychology of social science, unfortunately researchers do not pay attention to focusing on the personality traits of users to evaluate their trust relations. Considering users' personality traits could be a good direction for future researches in this domain.

VI. EVALUATION METRICS FOR TRUST PREDICTION PROCESS
In this section, we discuss the evaluation metrics that are frequently used in trust prediction approaches.

A. RANKING-BASED EVALUATION
One of the most well-known trust prediction evaluation metrics is ranking-based evaluation [13], [25], [81]. For this evaluation metric, we divide each of our datasets into two parts. The first part includes users who do not have any trust relations (N). The second part includes users who have trust relations with other users (T). We sort these trust relations based on their time of establishment. At that point, we select the first A% trust relations as old trust relations and denote 1 − A% of them as the New trust relations to predict. We consider four percentage values for A={60,70,80,90}. Further, we employ a trust prediction metric from Liben-Nowell and Kleinberg [139] to evaluate the performance of our approaches. Based on this, we first merge all New (new trust relations) and N (non-trust relations) such that N ∪ New and call them M . Then, we predict the trust relations in M and extract the |New| number of trust relations and call this Predict. Based on these sets, the performance of any trust prediction approach is determined by the following formula: where TPA is the trust prediction quality. The value of TPA is usually small and 'to demonstrate the significance of performance, [a] randomly guessing predictor is usually used as a baseline method' [13]. As we increase the size of A, the size of New decreases. This makes it difficult to accurately predict trust relations in M ; thus, the TPA is expected to decrease.

B. THE MEAN ABSOLUTE ERROR AND ROOT MEAN SQUARED ERROR METRICS
Two widely used prediction accuracy metrics for trust prediction approaches are mean absolute error (MAE) and root mean squared error (RMSE) [4], [13], [65]. Similar to the settings for the ranking based metric, we create M , New and N . Then, the trust values for the pairs of users in N are computed. MAE and RMSE can be defined as follows: where T AC ij is the actual trust relations between u i and u j , and T Pre ij is the predicted trust relations. A lower MAE and RMSE indicate a better performance. A small improvement in terms of RMSE or MAE has a significant effect on the quality of the top-few recommendations [13].

C. PRECISION, RECALL AND F1 SCORE
In the field of machine learning and classification problems, there is a matrix called ''Confusion matrix'' or ''error matrix'' is used to evaluate the performance of algorithms (Table 1). Obviously, results with the true positive and true negative labels are the desired results. Based on this matrix, Precision and Recall metrics can be defined as follows: where Precision is the percentage of relevant results compared to the retrieved results, while Recall is the percentages of relevant instances that were successfully retrieved. Finally, F1 measure is combining both Precision and Recall as follows:

VII. DATASETS FOR TRUST PREDICTION APPROACHES
In this section, we introduce the datasets that researchers are frequently used for evaluating their proposed trust prediction approaches. In most of these datasets, each pair of users has a Boolean label associated with its trust relation (which acts as the ground truth for our experiments). For instance, in the Epinions and Ciao Datasets [25], [81], the trust labels are generated by explicitly asking users to give a '0' or '1' value as the trust value to other users. Epinions and Ciao Datasets also contain the attributes of the users and their reviews, and use a 5-star (one-to-five) rating system to rate reviews. The reviews in these datasets are categorised by topic, such as travel, books, food and drink, house and garden and family.
In addition to the Epinions and Ciao datasets, Table 4 summarize the characteristics of other datasets that researchers can use in their evaluations. In this table, trust network density represents the ratio of the number of known trust relations and the possible trust relations (number of users × (number of users-1)). Advogato dataset is crawled from Advogato.org website which is for development of open-source software. Similar to Epinion and Ciao, the users in Advogato can explicitly certify other users as ''observer'', ''apperentice'',''journeyer'' and ''master'' which can demostrate the level of trustworthy of a user. Flixster and FriendFeed also are similar to the Epinions and Ciao datasets and users can provide a rating for other users' feed using a 5-star (one-to-five) rating system. Moreover, Wiki-RfA and Wiki-Elec datasets are based on the the votes that Wikipedia members are providing for an editor in Wikipedia to become an administrator. 56

VIII. ANALYSIS
In this Section, we first have an analysis over the existing trust prediction approaches from the point of view of their structures. Next, we compare the current approaches with respect to their time complexity.

A. ANALYZING THE EXISTING TRUST PREDICTION APPROACHES
To gain a better understanding of the existing trust prediction approaches, in Table 3 we classify the existing approaches according to whether they are supervised (S) or unsupervised (U ), where S denotes supervised and U represents unsupervised; whether the context of trust is considered, 5 http://snap.stanford.edu/data/wiki-RfA.html 6 http://snap.stanford.edu/data/wiki-Vote.html where Y represents that the property is satisfied and N denotes that the method cannot satisfy the property; and whether the dynamic, time-dependent nature of trust is considered, where Y likewise denotes that the property is satisfied, while N means it is not. Based on this analysis, we find that around 54% of existing trust-prediction approaches do not consider the context of trust. This means, they assume all trust relations are the same and that if a user trusts another user in one context, he or she will trust that user across all contexts. Surprisingly, only 27% of existing trust-prediction approaches are time-aware. Accordingly, the majority of existing approaches assume that trust relations last a lifetime.

B. TIME COMPLEXITY
In this subsection, we compare different trust prediction approaches from the point of view of their time complexity. Here, we mainly focused on the trust prediction approaches that discussed the time complexity of their approaches explicitly, or at least they provided the algorithms of their models. Table 4 compare these approaches.

IX. FUTURE DIRECTIONS
In this paper, we have investigated the problem of predicting trust between two unknown users in OSNs. We believe this is an important research area that has important applications for business and government in understanding trends in the diffusion of misinformation on OSNs. We believe this important research area, will attract a great deal of attention from the research community over the coming years. Below, we summarise some significant research directions in this area.

A. CONTEXT AND DATA CURATION
Although several context-aware trust prediction approaches have been proposed in the literature, there remains room to study the factors that can accurately capture the context of trust relations. It would be useful to investigate the use of textual contents in trust prediction approaches. As an almost unexplored research topic in trust prediction area, researchers need to use natural language processing techniques [142] to analyse textual contents as a rich source of information about users' activities and behaviour. Such analysis would enrich our available data about users and their relations, potentially helping to alleviate the data sparsity problem.
Accordingly, understanding the content and context of social data can help in understanding the trust relations among users in OSNs. For example, if a user retweets a  tweet on Twitter, it would be helpful to understand the text of the tweet, whether it contains an image or URL, and the keywords or entities (e.g., people, organisations, locations and products) and topics mentioned. In this context, data curation [143]- [145] (i.e., the task of preparing the raw data for analytics) can help in turning raw data into contextualised data and knowledge. For example, curating a raw tweet from Twitter can tell us if the tweet contains a mention of a person named Barak Obama (using entity extraction and coreference reolution techniques [146]) who was the 44th president of the United States (using linking techniques [147] to link this entity to external knowledge sources such as Wikidata 7 ). We can also understand if the topic of the tweet is related to politics (using topic extraction [148]) and if the tweet is discussing a social issue (using rule-based techniques [149]). A future direction would be to use data curation in OSNs to improve the accuracy of predicting the trust relation between two users.

B. TIME AND BUSINESS PROCESSES
Although there were a few attempts to introduce novel time-aware trust prediction approaches that can dynamically predict trust relations, the time complexity of these approaches in real-world scenarios must be critically examined. In other words, the next trust prediction approaches should focus on decreasing the execution times. Many of the existing trust prediction models are based on a computationally complex model with a high execution time. By decreasing the execution time of trust prediction approaches, we make them more feasible for real-world applications.
An important application in this category is to understand customer's personality, behaviour and attitude in business processes [150], [151] and to predict how their trust in companies and products may change over time. Business processes are a set of tasks and activities performed to accomplish a specific organisational goal [152], [153]. For example, consider a bank customer who has decided to change their bank or a specific product offered by a bank. Analysing the time-aware activities of bank customers may allow the loss of a trust relation for an existing product to be predicted. Another interesting avenue for future work in this domain would be to use data provenance [154], [155] to model and understand the evolution of social items over time. For example, to help predict customers' personality, behaviour and attitude in business processes, their retweets, likes and views could be analysed over time [138].

C. BENCHMARKING DATASETS
Surprisingly, even after several years of research in the trust prediction area, researchers still suffer from an absence of test datasets that provide sufficient contextual information about users and the dynamic timestamp of their trust relations. As an urgent need in this domain, providing such a dataset for trust prediction related research could help to attract many more researchers to this research area. Future work in this domain would be to use crowdsourcing techniques [156]- [158] to facilitate the labelling of such datasets.

D. CONTINUOUS TRUST METRICS
Although most of the existing trust prediction approaches assume that trust is a binary concept (''0'' for lack of a trust relation and and ''1'' for a trust relation between a pair of users), in real-word scenarios the trust relations can be Continuous variables and assign any real numbers as trust values [13]. In the future, researcher could more focus on this area of research. However, they employ the right evaluation metrics and datasets (relevant for Continuous trust values) in their evaluations.

X. CONCLUDING REMARKS
OSNs enable users to connect with others, expand their social networks, share multimedia content and write reviews on specific items. Users in OSNs are bombarded with information and trust can play an important role in their decision making. Due to the lack of interactions between the majority of participants on OSNs, predicting pair-wise trust relations in this context is a daunting task. In this paper, we extensively analysed the concept of trust and presented three main research challenges related to the trust prediction process. Next, we classified the state-of-the-art trust prediction approaches based on addressing those challenges. Finally, we suggested some potential research directions for researchers in this field. ADNAN MAHMOOD is currently associated with the Department of Computing, Macquarie University, Sydney, NSW, Australia. Before moving to Macquarie University, he spent a considerable number of years in the diverse academic and industrial settings of the Ireland (where he carried research on a project funded and co-funded by Science Foundation Ireland and European Regional Development Fund -Grant Number: 13/RC/2077, respectively), South Korea, Malaysia, Pakistan, and People's Republic of China. His research interests include softwaredefined networks, network functions virtualization, intelligent transportation systems, the Internet-of-Vehicles, trust management, and next generation heterogeneous wireless networks. He also serves on the technical program committees and editorial boards of several reputed international conferences and journals, respectively. SHAHPAR YAKHCHI received the bachelor's and master's degrees (Hons.) in computer software engineering. She is currently pursuing the Ph.D. degree with Macquarie University, Sydney, NSW, Australia. Her research interests include data analysis and recommender systems.
MEHMET A. ORGUN (Senior Member, IEEE) is currently a Professor with the Division of Information and Communication Sciences, Department of Computing, Macquarie University, Sydney, NSW, Australia. His current research interests include computational intelligence, multiagent systems, trust and security, temporal reasoning, and formal methods. VOLUME 8, 2020