Credibility score based multi-criteria recommender system

https://doi.org/10.1016/j.knosys.2020.105756Get rights and content

Highlights

  • A multi-criteria recommender system based on credibility of an item is proposed.

  • Personal and public views are computed to achieve more accurate prediction.

  • GA is used to discover the relationship between each criterion.

  • Aggregate the credibility score on various criteria to find overall credibility of an item.

Abstract

Recommender system has been emerged as a personalization tool to solve the issue of information overload in an e-commerce environment. Traditional collaborative filtering (CF) based recommender systems (RSs) suggest items to users based on their overall ratings which are used to find out similar users. Multi-criteria ratings are used to capture user preferences efficiently in multi-criteria recommender systems (MCRS), and incorporation of various criteria ratings can lead to higher performance in MCRS. Usually, user relies on the credibility of an item provided through his/her social circle or similar users, which is called a personal view on items from their close ones. However, it is not generally sufficient to depend exclusively on the personal view of the user. Therefore, public view that includes whole community can play a key role in the credibility of an item. In this paper, we propose a MCRS based on the credibility score of an item, which is an aggregated value of credibility scores on various criteria of an item. These credibility scores are computed based on personal and public views. However, different users have different priorities to various criteria of an item. Therefore, we use genetic algorithm (GA) to learn appropriate weights in the aggregation task of credibility score. The experiment results on Yahoo! Movies and modified MovieLens dataset demonstrate the effectiveness of proposed credibility score based MCRS in terms of coverage, recall, precision, and f-measure.

Introduction

The rapid expansion of Web has brought huge convenience for users as well as causing a problem of information overload, which makes it problematic for users to find useful information according to their requirements [1], [2]. RSs are intelligent tools that help to solve this overload problem by suggesting the relevant information to users [1], [3]. With the intense growth of the Internet, RS has been deployed to a variety of online systems [4], such as online videos (YouTube), movies (Netflix), songs (Last.FM), books (Amazon), online social networks (Facebook), news articles (Globo.com) [5], hotels (Goibibo) [6] etc.

Generally, RS can be built based on various filtering techniques such as content-based filtering (CBF), collaborative filtering (CF) and hybrid filtering [1], [2]. CBF technique [7], recommends items to users based on their past preferences and the contents of preferred items [8], while collaborative filtering (CF) generates recommendations to users based on similar users. CF, the most prevalent technique in the area of RS, has successfully explored in various domains such as movies, music, news, etc. The CF technique has been categorized into two categories- model-based filtering and memory-based filtering. In these two categories, memory-based CF provides more accurate recommendations to users [7]. These memory-based techniques are further classified into two categories such as user-based CF and item-based CF. In user-based CF, items are recommended based on similar users directly, while item-based CF generates recommendations based on similar items, however, these similarities are computed based on common users [4], [9]. Among those two techniques, CF is used more frequently to build RSs. However, each filtering technique has its own pros and cons. Therefore, hybrid filtering [10], [11] came into existence in building efficient RS by overcoming their weaknesses.

There are two major extensions of RS i.e. group RS and multi-criteria RS. Standard recommendation approaches used in various domains focus mostly on a single user. But there are plenty of situations where recommendations are needed for a group of users. For such scenarios, the group recommendation is the optimal solution [12]. In RSs, users provide their ratings to experienced items for describing their preferences. Majority of existing RSs use these overall ratings to generate recommendations to users. These overall ratings are not sufficient to capture users preferences efficiently because users may like/dislike any item based on some specific attributes [13]. For example, user u1 has given an overall rating 5 to a particular movie based on its story criteria and other user u2 has given same rating to that movie based on visual effects. However, these two users are not similar in real life. Therefore, it would be better to incorporate criteria ratings for generating effective recommendations to users. Hence, multi-criteria recommender systems (MCRSs) [1], [13], [14], [15] came into existence to provide quality items to users by considering their various criteria ratings. MCRS can be viewed as multi-criteria decision-making (MCDM) problem in decision support systems (DSS) where decision-makers can be visualized as users. Under this perception, users in MCRS will get recommendations from a set of objects estimated on several points of view, which are stated as criteria or attributes.

In traditional CF, input data does not capture user preferences efficiently because the data has users’ overall ratings only. Therefore, incorporation of multi-criteria ratings can be a good alternative for RSs. Usually, heuristic approaches and model-based approaches are discussed in [1], [13] to build MCRS framework. However, in these approaches, aggregation is a major concern to build an efficient MCRS. Further, they have suggested that these ratings can be engaged in both prediction and recommendation task. In recommendation task, defining utility of an item can be considered as a challenge in MCRS. To quickly establish a relationship between users and items, the important prerequisite is to evaluate the credibility of an item as a possible choice of the utility of an item for the target user. Credibility of an item is not explored very much in the area of RS. However, it can provide a good basis for generating efficient recommendations to users. More is the credibility of an item, higher are the chances to be recommended to users. Therefore, computation of credibility of an item is an important challenge in MCRS. In [16], authors proposed an approach that scores the credibility of user-reviews by seeing the following aspects such as reputation of the reviewer and value of the content of reviews in traditional RSs. The reputation of the reviewer can be considered as a public view in their approach. Different disciplines handle credibility in several ways according to their observations and needs. From a different perspective, credibility can be considered as a user’s trust among various users [17], [18], [19], [20]. Further, selecting an appropriate way to provide final recommendations to users based on credibility score of each criterion is another challenge in MCRS.

We have suggested following solutions in our proposed approach for handling the issues discussed above:

  • Defining the utility of an item: Credibility of an item can be considered as a utility of an item. We have computed the credibility of an item based on the aggregation of credibility scores (CS) on various criteria of an item. These CS on various criteria are evaluated by incorporating the public and personal views of users.

  • Aggregation task: We have suggested to use genetic algorithm (GA) for computing appropriate weights in the aggregation task for combining CS on different criteria.

  • Recommendation based on credibility of items: After computation of credibility of an item, we propose a method to suggest items to users based on the credibility of items. Items with higher credibility scores can be recommended to users.

The organization of this paper is summarized as follows: Section 2 provides background related to collaborative filtering and multi-criteria RS. Section 3 provides literature review relevant to our proposed approach. In Section 4 we introduce proposed approach for generating recommendations. Section 5 describes the computational experiments and results. In Section 6, conclusion work is presented with some future research directions.

Section snippets

Background

This section discusses collaborative filtering, followed bymulti-criteria recommender systems.

Literature review

Recommender systems with multiple criteria can be considered as MCDM problems in DSS (Decision support systems). Both MCRS and DSS based on MCDM, assist users and decision-makers respectively by providing a set of objects with various features. We can summarize the analogy between MCDM and MCRS in Table 2.

The following points should be considered while analyzing MCRS problem based on DSS [1].

Proposed approach

In this section, we propose a credibility score-based approach that provides a more accurate subset of the items to users. We have computed the credibility score of an item based on each criterion by considering users’ personal views (direct score) and public view (indirect score). Direct score of an item on each criteria is computed by using more similar users of a target user (active user) and the indirect score is computed by considering the non-similar users of a target user in our proposed

Experiments and results analysis

The experiments were performed to analyze the effectiveness of the proposed MCRS-CS (multi-criteria recommender system based on credibility score) using Yahoo! movies and modified MovieLens (100k ratings) dataset. Yahoo! movies dataset consists of 6078 users and 976 movies. Each movie is expressed in terms of its 4 criteria i.e. story, acting, direction, and visuals in addition to an overall rating. In this dataset, original ratings are provided on a 13-point rating scale (A+ to F), which is

Conclusion and future work

In this paper, we proposed MCRS based on credibility score (MCRS-CS) of an item to enhance recommendation quality of the system. In this paper, our work has three-fold. First, we compute direct and indirect scores of an item on various criteria based on personal and public view and then these scores are utilized to get the credibility score CSi,c on various criteria of an item. Second, we have computed credibility score of an item by aggregating these CSi,c through GA. Finally, the

CRediT authorship contribution statement

Shweta Gupta: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing - original draft. Vibhor Kant: Resources, Visualization, Supervision, Writing - review & editing.

References (42)

  • AdomaviciusG. et al.

    Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions

    IEEE Trans. Knowl. Data Eng.

    (2005)
  • RicciF. et al.

    Recommender systems: introduction and challenges

  • MoreiraG.d.S.P. et al.

    News session-based recommendations using deep neural networks

    (2018)
  • SagaR. et al.

    Hotel recommender system based on user’s preference transition

  • HaririW. et al.

    A multimedia content recommender system using table of contents and content-based filtering

    Adv. Sci. Lett.

    (2018)
  • AggarwalC.C.

    An introduction to recommender systems

  • BurkeR.

    Hybrid recommender systems: Survey and experiments

    User Model. User-Adapt. Interact.

    (2002)
  • TranT. et al.

    Hybrid recommender systems for electronic commerce

    Proc. Knowledge-Based Electronic MarketsPapers from the AAAI Workshop, Technical Report WS-00-04

    (2000)
  • AdomaviciusG. et al.

    Multi-criteria recommender systems

  • AdomaviciusG. et al.

    New recommendation techniques for multicriteria rating systems

    IEEE Intell. Syst.

    (2007)
  • JannachD. et al.

    Accuracy improvements for multi-criteria recommender systems

  • Cited by (25)

    • A novel top-n recommendation method for multi-criteria collaborative filtering

      2022, Expert Systems with Applications
      Citation Excerpt :

      Traditional CF-based methods typically use an overall rating for recommending items to users that do not meet their preferences efficiently. However, users may be more interested in a specific product-related criterion (Gupta & Kant, 2020). Determination of products or users that exhibit similar behavior is critical for the CF process.

    • Personalized Knowledge Distillation for Recommender System

      2022, Knowledge-Based Systems
      Citation Excerpt :

      In the era of information explosion, Recommender System (RS) has played a key role in helping users’ decisions and improving the cooperate profits [1–6].

    • “How to get consensus with neighbors?”: Rating standardization for accurate collaborative filtering

      2021, Knowledge-Based Systems
      Citation Excerpt :

      It indicates that the process of rating standardization in our NARS framework is scalable with the number of users and also the number of ratings. Potential avenues of future research include a new design of multi-criteria recommender systems [51–53] in the sense of alleviating the absence of consensus that may occur not only among users but also among various evaluation aspects (e.g., ‘cleanliness’, ‘service’, and ‘location’ of hotel in tourism domain) for a user. Hong-Kyun Bae: Conceptualization, Methodology, Software, Writing – original draft.

    • An efficient approach for improving the predictive accuracy of multi-criteria recommender system

      2024, International Journal of Information Technology (Singapore)
    View all citing articles on Scopus

    No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.knosys.2020.105756.

    View full text