Trust-aware Recommender Systems Chapter 1 Trust-aware Recommender Systems Trust-aware Recommender Systems 1.1 Recommender Systems Trust-aware Recommender Systems

Recommender systems are an effective solution to the information overload problem, specially in the online world where we are constantly faced with inordinately many choices. These systems try to find the items such as books or movies that match best with users’ preferences. Based on the different approaches to finding the items of interests to users, we can classify the recommender systems into three major groups. First, content based recommender systems use content information to make a recommendation. For example, such systems might recommend a romantic movie to a user that showed interest in romantic movies in her profile. Second, collaborative filtering recommender systems rely only on the past behavior of the users such as their previous transactions or ratings. By comparing this information, a collaborative filtering recommender system finds new items or users to users. In order to address the cold-start problem and fend off various types of attacks, the third class of recommender systems, namely trust-aware recommender systems, is proposed. These systems use social media and trust information to make a recommendation, which is shown to be promising in improving the accuracy of the recommendations. In this chapter, we give an overview of state-of-theart recommender systems with a focus on trust-aware recommender systems. In particular, we describe the ways that trust information can help to improve the quality of the recommendations. In the rest of the chapter, we introduce recommender systems, then trust in social media, and next trust-aware recommender systems. Trust-Aware Recommender Systems 3 1.1 Recommender Systems With the development of Web 2.0, information has increased at an unprecedented rate which aggravates the severity of the information overload problem for online users. For example, a search for “smartphone” returns 1,664,253 results in Amazon products or a search for “best movies to watch” in Google videos returns about 219,000,000 results. Due to the information overload problem, the decision-making process becomes perplexing when one is exposed to excessive information [58, 55, 19, 1]. Therefore, with the rapidly growing amount of available information and digital items on the web, it is necessary to use tools to filter this information in order to find items that are more likely to be of interest to the users. One can use search engines to overcome the information overload problem. In this case, the user has to refine the search terms or has to pick more specific query terms to narrow down the results. Another solution to overcome the information overload problem is to use top-k recommendations. In this approach, the system keeps a list of the most popular items and utilizes the list to recommend items to the user. For example, Ted is a website that uses this technique to recommend items to users. It can be seen in Figure 1.1, users can sort items bases on the different approaches such as overall popularity (most viewed), popularity in the past week (most emailed this week), or popularity in the past month (most popular this month) among others. Similar to search engines, top-k items are not usually customized based on users’ preferences and interest. In particulate, a top-k-item system returns the same list of items to people with different preferences. Therefore, customization is the major problem associated with these two approaches. Recommender systems are introduced to tackle the information overload, and the customization problem. Recommender systems are a subclass of information filtering systems that consider users’ preferences and recommended items that match with users’ preferences and interests [23]. These systems have become extremely common in recent years and are applied in a variety of applications including recommending products, social links, and digital items. The most popular ones are probably movies, music, news, books, and products in general [58, 70, 19, 26, 60]. Further, recommender systems are frequently used on recommending social links such as recommending people to follow on Twitter, befriend on social networks or dating sites [67, 37]. Furthermore, these systems are also used to accurately estimate the degree to which a particular user (from now on termed the target user) will like a particular item (the target item) [73]. Based on the type of data that recommender systems use, we can classify 1http://www.amazon.com 2https://www.google.com/#q=best+movies+to+watch&safe=active&tbm=vid 3http://www.ted.com/ 4 Trust-Aware Recommender Systems FIGURE 1.1: Ted.com uses a top-k item recommendation approach to rank items them into two major classes: content-based and collaborative filtering based recommender systems [76, 60]. Content-based recommendation systems use items’ features and characteristics to rank the items based on the user’s preferences. Collaborative filtering recommendation systems rely on the user’s past behavior e.g., purchases or ratings, to find similar users or items and utilize this information in order to find the items of interests to the user. In general, recommender systems are utility functions that predict the rating of item i from the item set I for user u from the user set U in the form of U × I → R, where rui is the rating of the item i for the given user u. The task of recommender systems is to predict user u’s rating for the given item i for which rui is unknown and use r̂ui to represent the predicted rating. The ratings, ru,i, can be any real number but often ratings are integers in the range [1, 5]. We use R to show all of the ratings. In real-world recommender systems, only a few users rate the items of interests (this number for many recommender system is less than 1%). Matrix 1.1 shows an example of a rating matrix with missing values. The goal of recommender systems is to predict these missing values. R =  5 2 3 4 3 4 2 2 5 3 5 5 3  (1.1) Trust-Aware Recommender Systems 5 Algorithm 1 Content-based recommendation 1: Describe the items that may be recommended. 2: Create a profile of the user that describes the types of items the user likes 3: Compare items to the user profile to determine what to recommend. The profile is often created and updated automatically in response to feedback on the desirability of items that have been presented to the user. 1.1.1 Content-based Recommendation Content-based recommender systems uses items’ and users’ features to create a profile for each item or user. For example, movie profile might include attributes such as gender, participating actors, director, and office box popularity. User profile includes demographic information and users’s interests [28]. These systems use supervised machine learning to induce a classifier that can discriminate between items likely to be of interest to the user and those likely to be uninteresting [52, 5, 49]. The recommender recommends an item to a user based on a description of the item and a profile of the users’ interests. Algorithm 1 shows the main steps of a content-based recommendation. We usually use vector space model to represent users’ and items’ features. In this model, every item or user is represented as a vector. i = (t1, t2, ..., tn) (1.2) where tj is the frequency of term j in item i. To model users or items more accurately, instead of frequency we can use tf-idf which can be calculated as follows: tft,i = ft,i max{fz,i : z ∈ i} idft = log N nt (1.3) wt,i = tft,i × idft (1.4) where ft,i is the frequency of term t in item i, max{fz,i : z ∈ i} is the maximum term frequency in item i, N is the total number of items, nt is the number of items where term t appears. tft,i denotes the frequency of term t in item i, and idft denotes the inverse document frequency of term t, which inversely correlates with the number of items, that term t is appeared in their descriptions. The similarity between user u and item i can be calculated using Equation 1.5. sim(u, i) = ∑ t∈T wt,uwt,i √∑ t∈T w 2 t,u √∑ t∈T w 2 t,i (1.5) where T indicates the set of terms that appeared in item and user description. 6 Trust-Aware Recommender Systems 1.1.2 Collaborative Filtering (CF) Collaborative filtering is the process of filtering the information or patterns using techniques involving collaboration among multiple agents, viewpoints, data sources, etc [69]. Collaborative filtering systems use the user’s past behavior, and recommend items that match their taste. Collaborative filtering recommender systems can be classified into memory-based and model-based collaborative filtering. In memory-based approach we predict the missing ratings based on similarity between users or items. In model-based approach, we use given user-item ratings to construct a model and use the model to predict missing ratings. We’ll give a detailed description of these two approaches in the following sections. The main advantage of this method is that the recommender system does not need to have any information about the users and content of the items to recommend. User-item ratings are the only information the system needs to operate. The following are assumptions for collaborative filtering systems [76]: • Users with similar ratings on some items are more likely to have similar ratings on future items, and • Items with similar ratings in the past are more likely to have similar ratings in the future. Figure 1.2 illustrates this approach for a small set of users and movies. The goal is recommending a new movie to Jack. In the first step, the system finds three other users that have similar movie taste as Jack’s. The next step it looks for other movies that these users liked. All three of them liked “Once Upon a Time in the West”, and two of them liked “Spider man”. Therefore, the top recommendation would be “Once Upon a Time in the West”. 1.1.2.1 Memory-based Collaborative Filtering In a memory-based approach, the recommender system aims to predict the missing ratings based on either similarity between users or similarity between items [6, 53, 30]. The former is built upon the hypothesis that similar users have similar tastes. Hence, to make a reasonable recommendation, it finds similar users, then uses these users’ taste to make a recommendation for the target user. The second approach is built upon the consistency of a user’s taste. If a user liked a product, she will like similar products as well. In both approaches, the recommender system takes two steps to recommend an item to the user. First, it calculates the similarity between users or similarity between items. Then, it uses the most similar users or items to make its recommendation. Collaborative filtering uses the rating information to compute user-user or item-item similarity. Cosine similarity and the Pearson correlation coefficients are two of the most frequently used similarity measures in collaborative filtering. Given two users u, and v, and their ratings, the cosine similarity is Trust-Aware Recommender Systems 7