matrix factorization techniques for recommender systems

computer 30 COVER FE ATURE Published by the IEEE Computer Society 0018-9162/09/$26.00 © 2009 IEEE Such systems are particularly useful for entertainment products such as movies, music, and TV shows. Many cus- tomers will view the same movie, and each customer is likely to view numerous different movies. Customers have proven willing to indicate their level of satisfaction with particular movies, so a huge volume of data is available about which movies appeal to which customers. Com- panies can analyze this data to recommend movies to particular customers. RecommendeR system stRategies Broadly speaking, recommender systems are based on one of two strategies. The content filtering approach creates a profile for each user or product to characterize its nature. For example, a movie profile could include at- tributes regarding its genre, the participating actors, its box office popularity, and so forth. User profiles might include demographic information or answers provided on a suitable questionnaire. The profiles allow programs to associate users with matching products. Of course, content-based strategies require gathering external infor- mation that might not be available or easy to collect. A known successful realization of content filtering is the Music Genome Project, which is used for the Internet radio service Pandora.com. A trained music analyst scores M odern consumers are inundated with choices. Electronic retailers and content providers offer a huge selection of prod- ucts, with unprecedented opportunities to meet a variety of special needs and tastes. Matching consumers with the most appropriate products is key to enhancing user satisfaction and loy- alty. Therefore, more retailers have become interested in recommender systems, which analyze patterns of user interest in products to provide personalized recommenda- tions that suit a user’s taste. Because good personalized recommendations can add another dimension to the user experience, e-commerce leaders like Amazon.com and Netflix have made recommender systems a salient part of their websites. As the Netflix Prize competition has dem- onstrated, matrix factorization models are superior to classic nearest-neighbor techniques for producing product recom- mendations, allowing the incorporation of additional information such as implicit feedback, temporal effects, and confidence levels. Yehuda Koren, Yahoo Research Robert Bell and Chris Volinsky, AT&T Labs—Research MATRIX FACTORIZATION TECHNIQUES FOR RECOMMENDER SYSTEMS 31AuGuSt 2009 well-defined dimensions such as depth of character de- velopment or quirkiness; or completely uninterpretable dimensions. For users, each factor measures how much the user likes movies that score high on the correspond- ing movie factor. Figure 2 illustrates this idea for a simplified example in two dimensions. Consider two hypothetical dimen- sions characterized as female- versus male-oriented and serious versus escapist. The figure shows where several well-known movies and a few fictitious users might fall on these two dimensions. For this model, a user’s predicted rating for a movie, relative to the movie’s average rating, would equal the dot product of the movie’s and user’s lo- cations on the graph. For example, we would expect Gus to love Dumb and Dumber, to hate The Color Purple, and to rate Braveheart about average. Note that some mov- ies—for example, Ocean’s 11—and users—for example, Dave—would be characterized as fairly neutral on these two dimensions. matRix factoRization methods Some of the most successful realizations of latent factor models are based on matrix factorization. In its basic form, matrix factorization characterizes both items and users by vectors of factors inferred from item rating patterns. High correspondence between item and user factors leads to a each song in the Music Genome Project based on hundreds of distinct musical characteristics. These attributes, or genes, capture not only a song’s musical identity but also many significant qualities that are relevant to understanding listeners’ musi- cal preferences. An alternative to content filtering relies only on past user behavior—for example, previous transactions or product ratings— without requiring the creation of explicit profiles. This approach is known as col- laborative filtering, a term coined by the developers of Tapestry, the first recom- mender system.1 Collaborative filtering analyzes relationships between users and interdependencies among products to identify new user-item associations. A major appeal of collaborative fil- tering is that it is domain free, yet it can address data aspects that are often elusive and difficult to profile using content filter- ing. While generally more accurate than content-based techniques, collaborative filtering suffers from what is called the cold start problem, due to its inability to ad- dress the system’s new products and users. In this aspect, content filtering is superior. The two primary areas of collaborative filtering are the neighborhood methods and latent factor models. Neighbor- hood methods are centered on computing the relationships between items or, alternatively, between users. The item- oriented approach evaluates a user’s preference for an item based on ratings of “neighboring” items by the same user. A product’s neighbors are other products that tend to get similar ratings when rated by the same user. For example, consider the movie Saving Private Ryan. Its neighbors might include war movies, Spielberg movies, and Tom Hanks movies, among others. To predict a par- ticular user’s rating for Saving Private Ryan, we would look for the movie’s nearest neighbors that this user actually rated. As Figure 1 illustrates, the user-oriented approach identifies like-minded users who can complement each other’s ratings. Latent factor models are an alternative approach that tries to explain the ratings by characterizing both items and users on, say, 20 to 100 factors inferred from the ratings patterns. In a sense, such factors comprise a computerized alternative to the aforementioned human- created song genes. For movies, the discovered factors might measure obvious dimensions such as comedy versus drama, amount of action, or orientation to children; less Joe #2 #3 #1 #4 figure 1. The user-oriented neighborhood method. Joe likes the three movies on the left. To make a prediction for him, the system finds similar users who also liked those movies, and then determines which other movies they liked. In this case, all three liked Saving Private Ryan, so that is the first recommendation. Two of them liked Dune, so that is next, and so on. COVER FE ATURE computer 32 vector q i ∈ R f, and each user u is associ- ated with a vector p u ∈ R f. For a given item i, the elements of q i measure the extent to which the item possesses those factors, positive or negative. For a given user u, the elements of p u measure the extent of interest the user has in items that are high on the corresponding factors, again, posi- tive or negative. The resulting dot product, q i T p u , captures the interaction between user u and item i—the user’s overall interest in the item’s characteristics. This approximates user u’s rating of item i, which is denoted by r ui , leading to the estimate rˆui = q i T p u . (1) The major challenge is computing the map- ping of each item and user to factor vectors q i , p u ∈ R f. After the recommender system completes this mapping, it can easily esti- mate the rating a user will give to any item by using Equation 1. Such a model is closely related to singular value decom- position (SVD), a well-established technique for identifying latent semantic factors in information retrieval. Applying SVD in the collaborative filtering domain requires factoring the user-item rating matrix. This often raises difficulties due to the high portion of missing values caused by sparse- ness in the user-item ratings matrix. Conventional SVD is undefined when knowledge about the matrix is incom- plete. Moreover, carelessly addressing only the relatively few known entries is highly prone to overfitting. Earlier systems relied on imputation to fill in missing ratings and make the rating matrix dense.2 However, im- putation can be very expensive as it significantly increases the amount of data. In addition, inaccurate imputation might distort the data considerably. Hence, more recent works3-6 suggested modeling directly the observed rat- ings only, while avoiding overfitting through a regularized model. To learn the factor vectors (p u and q i ), the system minimizes the regularized squared error on the set of known ratings: min * *,q p ( , )u i ∈ ∑ κ (r ui - q i Tp u )2 + λ(|| q i ||2 + || p u ||2) (2) Here, κ is the set of the (u,i) pairs for which r ui is known (the training set). The system learns the model by fitting the previously observed ratings. However, the goal is to generalize those previous ratings in a way that predicts future, unknown ratings. Thus, the system should avoid overfitting the observed data by regularizing the learned parameters, whose magnitudes are penalized. The constant λ controls recommendation. These methods have become popular in recent years by combining good scalability with predictive accuracy. In addition, they offer much flexibility for model- ing various real-life situations. Recommender systems rely on different types of input data, which are often placed in a matrix with one dimension representing users and the other dimension representing items of interest. The most convenient data is high-quality explicit feedback, which includes explicit input by users regarding their interest in products. For example, Netflix collects star ratings for movies, and TiVo users indicate their preferences for TV shows by pressing thumbs-up and thumbs-down buttons. We refer to explicit user feedback as ratings. Usually, explicit feedback com- prises a sparse matrix, since any single user is likely to have rated only a small percentage of possible items. One strength of matrix factorization is that it allows incorporation of additional information. When explicit feedback is not available, recommender systems can infer user preferences using implicit feedback, which indirectly reflects opinion by observing user behavior, including pur- chase history, browsing history, search patterns, or even mouse movements. Implicit feedback usually denotes the presence or absence of an event, so it is typically repre- sented by a densely filled matrix. a Basic matRix factoRization modeL Matrix factorization models map both users and items to a joint latent factor space of dimensionality f, such that user-item interactions are modeled as inner products in that space. Accordingly, each item i is associated with a Geared toward males Serious Escapist The Princess Diaries Braveheart Lethal Weapon Independence Day Ocean’s 11 Sense and Sensibility Gus Dave Geared toward females Amadeus The Lion King Dumb and Dumber The Color Purple figure 2. A simplified illustration of the latent factor approach, which characterizes both users and movies using two axes—male versus female and serious versus escapist. 33AuGuSt 2009 data aspects and other application-specific requirements. This requires accommodations to Equation 1 while staying within the same learning framework. Equation 1 tries to cap- ture the interactions between users and items that produce the different rating values. However, much of the observed variation in rating values is due to effects associated with either users or items, known as biases or intercepts, indepen- dent of any interactions. For example, typical collaborative filtering data exhibits large systematic tendencies for some users to give higher ratings than others, and for some items to receive higher ratings than others. After all, some products are widely perceived as better (or worse) than others. Thus, it would be unwise to explain the full rating value by an interaction of the form q i Tp u . Instead, the system tries to identify the portion of these values that individual user or item biases can explain, subjecting only the true interaction portion of the data to factor modeling. A first-order approxi- mation of the bias involved in rating r ui is as follows: b ui = µ + b i + b u (3) The bias involved in rating r ui is denoted by b ui and ac- counts for the user and item effects. The overall average rating is denoted by µ; the parameters b u and b i indicate the observed deviations of user u and item i, respectively, from the average. For example, suppose that you want a first-order estimate for user Joe’s rating of the movie Titanic. Now, say that the average rating over all movies, µ, is 3.7 stars. Furthermore, Titanic is better than an average movie, so it tends to be rated 0.5 stars above the average. On the other hand, Joe is a critical user, who tends to rate 0.3 stars lower than the average. Thus, the estimate for Titanic’s rating by Joe would be 3.9 stars (3.7 + 0.5 - 0.3). Biases extend Equation 1 as follows: rˆui = µ+ b i + b u + q i Tp u (4) Here, the observed rating is broken down into its four components: global average, item bias, user bias, and user- item interaction. This allows each component to explain only the part of a signal relevant to it. The system learns by minimizing the squared error function:4,5 min * * *, ,p q b ( , )u i ∈ ∑ κ (r ui - µ - b u - b i - p u Tq i )2 + λ (|| p u ||2 + || q i ||2 + b u 2 + b i 2) (5) Since biases tend to capture much of the observed signal, their accurate modeling is vital. Hence, other works offer more elaborate bias models.11 additionaL inPUt soURces Often a system must deal with the cold start problem, wherein many users supply very few ratings, making it the extent of regularization and is usually determined by cross-validation. Ruslan Salakhutdinov and Andriy Mnih’s “Probabilistic Matrix Factorization”7 offers a probabilistic foundation for regularization. LeaRning aLgoRithms Two approaches to minimizing Equation 2 are stochastic gradient descent and alternating least squares (ALS). stochastic gradient descent Simon Funk popularized a stochastic gradient descent optimization of Equation 2 (http://sifter.org/~simon/ journal/20061211.html) wherein the algorithm loops through all ratings in the training set. For each given training case, the system predicts r ui and computes the associated prediction error e ui = def r ui - q i T p u . Then it modifies the parameters by a magnitude pro- portional to g in the opposite direction of the gradient, yielding: • q q e p qi i ui u i← + ⋅ ⋅ − ⋅γ λ( ) • p p e q pu u ui i u← + ⋅ ⋅ − ⋅γ λ( ) This popular approach4-6 combines implementation ease with a relatively fast running time. Yet, in some cases, it is beneficial to use ALS optimization. alternating least squares Because both q i and p u are unknowns, Equation 2 is not convex. However, if we fix one of the unknowns, the op- timization problem becomes quadratic and can be solved optimally. Thus, ALS techniques rotate between fixing the q i ’s and fixing the p u ’s. When all p u ’s are fixed, the system recomputes the q i ’s by solving a least-squares problem, and vice versa. This ensures that each step decreases Equation 2 until convergence.8 While in general stochastic gradient descent is easier and faster than ALS, ALS is favorable in at least two cases. The first is when the system can use parallelization. In ALS, the system computes each q i independently of the other item factors and computes each p u independently of the other user factors. This gives rise to potentially massive parallelization of the algorithm.9 The second case is for systems centered on implicit data. Because the training set cannot be considered sparse, looping over each single training case—as gradient descent does—would not be practical. ALS can efficiently handle such cases.10 adding Biases One benefit of the matrix factorization approach to col- laborative filtering is its flexibility in dealing with various COVER FE ATURE computer 34 prove accuracy. Decomposing ratings into distinct terms allows the system to treat different temporal aspects sepa- rately. Specifically, the following terms vary over time: item biases, b i (t); user biases, b u (t); and user preferences, p u (t). The first temporal effect addresses the fact that an item’s popularity might change over time. For example, movies can go in and out of popularity as triggered by external events such as an actor’s appearance in a new movie. Therefore, these models treat the item bias b i as a function of time. The second temporal effect allows users to change their baseline ratings over time. For example, a user who tended to rate an average movie “4 stars” might now rate such a movie “3 stars.” This might reflect several factors including a natural drift in a user’s rating scale, the fact that users assign ratings relative to other recent ratings, and the fact that the rater’s identity within a house- hold can change over time. Hence, in these models, the parameter b u is a function of time. Temporal dynamics go beyond this; they also affect user preferences and therefore the interaction between users and items. Users change their preferences over time. For example, a fan of the psychological thrillers genre might become a fan of crime dramas a year later. Simi- larly, humans change their perception of certain actors and directors. The model accounts for this effect by taking the user factors (the vector p u ) as a function of time. On the other hand, it specifies static item characteristics, q i , because, unlike humans, items are static in nature. Exact parameterizations of time-varying parameters11 lead to replacing Equation 4 with the dynamic prediction rule for a rating at time t: rˆui (t) = µ + b i (t) + b u (t) + q i T p u (t) (7) inPUts With VaRying confidence LeVeLs In several setups, not all observed ratings deserve the same weight or confidence. For example, massive adver- tising might influence votes for certain items, which do not aptly reflect longer-term characteristics. Similarly, a system might face adversarial users that try to tilt the rat- ings of certain items. Another example is systems built around implicit feedback. In such systems, which interpret ongoing user behavior, a user’s exact preference level is hard to quantify. Thus, the system works with a cruder binary representation, stating either “probably likes the product” or “probably not interested in the product.” In such cases, it is valuable to attach confidence scores with the esti- mated preferences. Confidence can stem from available numerical values that describe the frequency of actions, for example, how much time the user watched a certain show or how frequently a user bought a certain item. These numerical values indicate the confidence in each obser- vation. Various factors that have nothing to do with user difficult to reach general conclusions on their taste. A way to relieve this problem is to incorporate additional sources of information about the users. Recommender systems can use implicit feedback to gain insight into user preferences. Indeed, they can gather behavioral information regardless of the user’s willingness to provide explicit ratings. A re- tailer can use its customers’ purchases or browsing history to learn their tendencies, in addition to the ra