Predicting Missing Ratings in Recommender Systems: Adapted Factorization Approach

the paper presents a factorization-based approach to make predictions in recommender systems. these systems are widely used in electronic commerce to help customers find products according to their preferences. taking into account the customer’s ratings of some products available in the system, the recommender system tries to predict the ratings the customer would give to other products in the system. the proposed factorization-based approach uses all the information provided to compute the predicted ratings, in the same way as approaches based on singular Value Decomposition (sVD). the main advantage of this technique versus sVD-based approaches is that it can deal with missing data. It also has a smaller computational cost. Experimental results with public data sets are provided to show that the proposed adapted factorization approach gives better predicted ratings than a widely used sVD-based approach. KEy worDs AnD phrAsEs: Factorization technique, recommender systems, singular value decomposition. Since the amount of information available on the World Wide Web increases constantly, sometimes it becomes difficult to focus on interesting information and discard redundant content. For this reason, there is a high demand for methods that select interesting information with respect to users’ preferences. Recommender systems target this demand by helping users to find items, using previous knowledge about the user’s preferences. Users give ratings only to some of the items and therefore the system is able to predict their preferences on the rest of items (this is known as prediction task). The system can also recommend products according to the user’s preferences (recommen‐ dation task). These two powerful tools are widely used on e-commerce sites. Since their introduction in the 1990s, recommender systems have been used to filter information on the Web and to provide recommendations for books, CDs, movies, news, electronics, financial services, travel, and other products. One of the most popular recommender systems is the one at www.amazon. com. The customer rates some books and the system suggests other books, considering information from other customers. A different recommender system is used at www.everyonesacritic.net, where users give their opinion about movies and the system makes recommendations for people who share similar tastes. Another example (www.gnomoradio.org) consists of a music recommender system, where the user rates the music, and the system builds a This work was supported by the government of Spain under projects TRA200762526/AUT and DPI2007-66556-C03-03, and research program Consolider-Ingenio 2010: MIPRCV (CSD2007-00018). The authors thank Nathan Faggian for providing an incremental SVD code. 03 julia.indd 89 10/13/2009 10:33:23 PM 90 JuLIà, SAppA, LumbrErAS, SErrAt, AnD LópEz listening profile based on the user’s ratings. In addition, it recommends music from other users with similar profiles. Thus, in most cases, the main goal of a recommender system is to discover the customer’s preferred products in order to increase sales. This also helps customer, because they will only receive information filtered according to their individual taste. Recommender systems store data in a large table of users (also denoted as customers) and items (or products). Hence, the information is stored into a matrix of data, whose rows and columns correspond to each user and item respectively, and whose entries correspond to the ratings customers give to items. In real problems, the number of customers and items is huge, so it is necessary to deal with large data matrices. Since each user only rates a subset of the items, most entries in the matrix of ratings are empty, which means that the matrix tends to be very sparse.