Using community preference for overcoming sparsity and cold-start problems in collaborative filtering system offering soft ratings

Highlights

• This paper proposes a new collaborative filtering recommender system offering soft ratings.

• In the system, community preferences are used for overcoming sparsity and cold-start problems.

• The new system is experimentally tested on Flixster data set.

Abstract

This paper introduces a new collaborative filtering recommender system that is capable of offering soft ratings as well as integrating with a social network containing all users. Offering soft ratings is known as a new methodology for modeling subjective, qualitative, and imperfect information about user preferences, as well as a more realistic and flexible means for users to express their preferences on products and services. Additionally, in the system, community preferences that are extracted from the social network are employed for overcoming sparsity and cold-start problems. In the experiment, the new system is tested using a data set culled from Flixster, a social network focused on movies. The experiment’s results show that this system is more effective than the selected baseline in terms of recommendation accuracy.

Introduction

Recommender systems (RSs) (Adomavicius and Tuzhilin, 2005, Hwang et al., 2010, Kim et al., 2002) have been developing rapidly since they were introduced in 1990s; in practice, RSs have been applied in a variety of e-commerce applications (Linden et al., 2003, Shambour and Lu, 2011, Kim et al., 2011). In general, RSs collect information about user preferences from multiple sources, estimate user preferences on unseen items, and then generate suitable recommendations based on the estimated data. Logically, the quality of the recommendations in an RS mainly depends on the representation of user preferences and the accuracy of estimations.

Conventional RSs described in the literature usually provide a rating domain defined as a totally ordered finite set of rating scores, and represent a user’s preference for an item as a rating score (i.e., a hard rating). However, user preferences are naturally subjective and qualitative; therefore, representing user preferences as hard ratings is not appropriate in some cases (Nguyen and Huynh, 2015). For example, let us consider an RS offering hard ratings with a rating domain $\mathrm{\Theta }=\{1\text{,}2\text{,}3\text{,}4\text{,}5\}$. Suppose that a user has rated two items, $I_1$ and $I_3$, with rating scores of 4 and 5, respectively; further, this user wants to rate item $I_2$ with a rating score indicating that it is better than $I_1$ but worse than $I_3$. By using a rating score, the user can evaluate item $I_2$ only as 4 or 5; as a consequence, this user may hesitate to express his/her preference for the item. In this scenario, it will be more comfortable for the user to use a combination of rating scores, such as $\{4\text{,}5\}$ (i.e., a soft rating). Moreover, even though a user has evaluated an item by using a hard rating, this rating might encode imperfect information inside. For instance, when a user has rated an item with a rating score, such as 3, we cannot know exactly what this user thought about the item because the user probably wanted to evaluate the item as at least 3 or 3 for 90%. To addess such situations, the use of soft ratings in RSs has recently been studied and developed for the purpose of capturing subjective, qualitative, and imperfect information about user preferences (Wickramarathne et al., 2011, Nguyen and Huynh, 2014, Nguyen and Huynh, 2015). With a rating domain $\mathrm{\Theta }=\{1\text{,}2\text{,}3\text{,}4\text{,}5\}$, a user can express his/her preference for an item by using a soft rating as follows:

• 3 for sure ($\{3\}$ with a probability of 1.0)

• At least 3 ($\{3\text{,}4\text{,}5\}$ with a probability of 1.0)

• 3 for 90% ($\{3\}$ and $\mathrm{\Theta }$ with probabilities of 0.9 and 0.1, respectively)

• Less than 3 ($\{1\text{,}2\}$ with a probability of 1.0)

According to the aforementioned studies, RSs offering soft ratings are developed based on the Dempster-Shafer theory (DST) (Dempster, 1967, Shafer, 1976), which is considered as one of the most general theories for modeling imperfect information (Wickramarathne et al., 2011).

Furthermore, in the RS research area, recommendation techniques can be divided into three main categories: collaborative filtering, content-based, and hybrid (Adomavicius and Tuzhilin, 2005). Among these, collaborative filtering is regarded as the most well-known technique (Ricci et al., 2011, Jannach et al., 2012). To provide personalized recommendations to an active user, collaborative filtering RSs commonly try to find other users who are expected to have preferences similar to those of the user, and employ those users’ ratings to estimate the original user’s preferences for unseen items. However, collaborative filtering RSs also suffer from two fundamental limitations known as sparsity and cold-start problems (Adomavicius and Tuzhilin, 2005). The first problem is caused when each user only rates a very small number of items; as a result, the number of ratings is insufficient and recommendation performance is significantly affected (Huang et al., 2004). The second problem arises because of missing information about new items and new users.

Social networks are growing very fast and play a significant role on the Internet. Additionally, communication and collaboration in social networks have become more and more convenient and frequent. In a social network, users naturally form into various communities whose members interact frequently with one another (Tang and Liu, 2010). These social relations might influence individual behaviors and decisions, including those related to buying items. Commonly, when consulting for advice before buying new items, most people tend to believe in recommendations from friends in the same community rather than in recommendations from outside users. In a community, each member has his/her own preference for an item, and the overall preference of all members is called the community preference for that item. In practice, community preferences can be used for better understanding of user behaviors and ratings; thus, these preferences can potentially be used for predicting the missing information about user preferences. In other words, community preferences might be useful for overcoming the limitations in collaborative filtering RSs.

On the basis of this observation, this paper proposes a new collaborative filtering RS that is capable of (1) representing user preferences as soft ratings and (2) exploiting community preferences derived from a social network containing all users, which helps address the aforementioned sparsity and cold-start problems. In the system, community preferences are employed for predicting all unprovided user ratings on items (including new users and new items); subsequently, active users receive personalized recommendations based mainly on the provided and predicted ratings.

The remainder of this paper is organized as follows. In Section 2, background information about DST is introduced. In Section 3, related work is discussed. The proposed system is described in Section 4. In Section 5, experiments are presented, and their results are discussed. Finally, conclusions and suggestions for future research are presented in Section 6.