Yehuda Koren

Matrix Factorization Techniques for Recommender Systems

As the Netflix Prize competition has demonstrated, matrix factorization models are superior to classic nearest neighbor techniques for producing product recommendations, allowing the incorporation of additional information such as implicit feedback, temporal effects, and confidence levels.

Factorization meets the neighborhood: a multifaceted collaborative filtering model

Recommender systems provide users with personalized suggestions for products or services. These systems often rely on Collaborating Filtering (CF), where past transactions are analyzed in order to establish connections between users and products. The two more successful approaches to CF are latent factor models, which directly profile both users and products, and neighborhood models, which analyze similarities between products or users. In this work we introduce some innovations to both approaches. The factor and neighborhood models can now be smoothly merged, thereby building a more accurate combined model. Further accuracy improvements are achieved by extending the models to exploit both explicit and implicit feedback by the users. The methods are tested on the Netflix data. Results are better than those previously published on that dataset. In addition, we suggest a new evaluation metric, which highlights the differences among methods, based on their performance at a top-K recommendation task.

Collaborative Filtering for Implicit Feedback Datasets

A common task of recommender systems is to improve customer experience through personalized recommendations based on prior implicit feedback. These systems passively track different sorts of user behavior, such as purchase history, watching habits and browsing activity, in order to model user preferences. Unlike the much more extensively researched explicit feedback, we do not have any direct input from the users regarding their preferences. In particular, we lack substantial evidence on which products consumer dislike. In this work we identify unique properties of implicit feedback datasets. We propose treating the data as indication of positive and negative preference associated with vastly varying confidence levels. This leads to a factor model which is especially tailored for implicit feedback recommenders. We also suggest a scalable optimization procedure, which scales linearly with the data size. The algorithm is used successfully within a recommender system for television shows. It compares favorably with well tuned implementations of other known methods. In addition, we offer a novel way to give explanations to recommendations given by this factor model.

Performance of recommender algorithms on top-n recommendation tasks

In many commercial systems, the best bet recommendations are shown, but the predicted rating values are not. This is usually referred to as a top-N recommendation task, where the goal of the recommender system is to find a few specific items which are supposed to be most appealing to the user. Common methodologies based on error metrics (such as RMSE) are not a natural fit for evaluating the top-N recommendation task. Rather, top-N performance can be directly measured by alternative methodologies based on accuracy metrics (such as precision/recall).n An extensive evaluation of several state-of-the art recommender algorithms suggests that algorithms optimized for minimizing RMSE do not necessarily perform as expected in terms of top-N recommendation task. Results show that improvements in RMSE often do not translate into accuracy improvements. In particular, a naive non-personalized algorithm can outperform some common recommendation approaches and almost match the accuracy of sophisticated algorithms. Another finding is that the very few top popular items can skew the top-N performance. The analysis points out that when evaluating a recommender algorithm on the top-N recommendation task, the test set should be chosen carefully in order to not bias accuracy metrics towards non-personalized solutions. Finally, we offer practitioners new variants of two collaborative filtering algorithms that, regardless of their RMSE, significantly outperform other recommender algorithms in pursuing the top-N recommendation task, with offering additional practical advantages. This comes at surprise given the simplicity of these two methods.

Collaborative filtering with temporal dynamics

Customer preferences for products are drifting over time. Product perception and popularity are constantly changing as new selection emerges. Similarly, customer inclinations are evolving, leading them to ever redefine their taste. Thus, modeling temporal dynamics should be a key when designing recommender systems or general customer preference models. However, this raises unique challenges. Within the eco-system intersecting multiple products and customers, many different characteristics are shifting simultaneously, while many of them influence each other and often those shifts are delicate and associated with a few data instances. This distinguishes the problem from concept drift explorations, where mostly a single concept is tracked. Classical time-window or instance-decay approaches cannot work, as they lose too much signal when discarding data instances. A more sensitive approach is required, which can make better distinctions between transient effects and long term patterns. The paradigm we offer is creating a model tracking the time changing behavior throughout the life span of the data. This allows us to exploit the relevant components of all data instances, while discarding only what is modeled as being irrelevant. Accordingly, we revamp two leading collaborative filtering recommendation approaches. Evaluation is made on a large movie rating dataset by Netflix. Results are encouraging and better than those previously reported on this dataset.

Factor in the neighbors: Scalable and accurate collaborative filtering

Recommender systems provide users with personalized suggestions for products or services. These systems often rely on collaborating filtering (CF), where past transactions are analyzed in order to establish connections between users and products. The most common approach to CF is based on neighborhood models, which originate from similarities between products or users. In this work we introduce a new neighborhood model with an improved prediction accuracy. Unlike previous approaches that are based on heuristic similarities, we model neighborhood relations by minimizing a global cost function. Further accuracy improvements are achieved by extending the model to exploit both explicit and implicit feedback by the users. Past models were limited by the need to compute all pairwise similarities between items or users, which grow quadratically with input size. In particular, this limitation vastly complicates adopting user similarity models, due to the typical large number of users. Our new model solves these limitations by factoring the neighborhood model, thus making both item-item and user-user implementations scale linearly with the size of the data. The methods are tested on the Netflix data, with encouraging results.

Modeling relationships at multiple scales to improve accuracy of large recommender systems

The collaborative filtering approach to recommender systems predicts user preferences for products or services by learning past user-item relationships. In this work, we propose novel algorithms for predicting user ratings of items by integrating complementary models that focus on patterns at different scales. At a local scale, we use a neighborhood-based technique that infers ratings from observed ratings by similar users or of similar items. Unlike previous local approaches, our method is based on a formal model that accounts for interactions within the neighborhood, leading to improved estimation quality. At a higher, regional, scale, we use SVD-like matrix factorization for recovering the major structural patterns in the user-item rating matrix. Unlike previous approaches that require imputations in order to fill in the unknown matrix entries, our new iterative algorithm avoids imputation. Because the models involve estimation of millions, or even billions, of parameters, shrinkage of estimated values to account for sampling variability proves crucial to prevent overfitting. Both the local and the regional approaches, and in particular their combination through a unifying model, compare favorably with other approaches and deliver substantially better results than the commercial Netflix Cinematch recommender system on a large publicly available data set.

Yahoo! music recommendations: modeling music ratings with temporal dynamics and item taxonomy

In the past decade large scale recommendation datasets were published and extensively studied. In this work we describe a detailed analysis of a sparse, large scale dataset, specifically designed to push the envelope of recommender system models. The Yahoo! Music dataset consists of more than a million users, 600 thousand musical items and more than 250 million ratings, collected over a decade. It is characterized by three unique features: First, rated items are multi-typed, including tracks, albums, artists and genres; Second, items are arranged within a four level taxonomy, proving itself effective in coping with a severe sparsity problem that originates from the unusually large number of items (compared to, e.g., movie ratings datasets). Finally, fine resolution timestamps associated with the ratings enable a comprehensive temporal and session analysis. We further present a matrix factorization model exploiting the special characteristics of this dataset. In particular, the model incorporates a rich bias model with terms that capture information from the taxonomy of items and different temporal dynamics of music ratings. To gain additional insights of its properties, we organized the KddCup-2011 competition about this dataset. As the competition drew thousands of participants, we expect the dataset to attract considerable research activity in the future.

Second workshop on information heterogeneity and fusion in recommender systems (HetRec2011)

1. MOTIVATION AND GOALS In recent years, increasing attention has been given to finding ways for combining, integrating and mediating heterogeneous sources of information for the purpose of providing better personalized services in many information seeking and ecommerce applications. Information heterogeneity can indeed be identified in any of the pillars of a recommender system: the modeling of user preferences, the description of resource contents, the modeling and exploitation of the context in which recommendations are made, and the characteristics of the suggested resource lists.

OrdRec: an ordinal model for predicting personalized item rating distributions

We propose a collaborative filtering (CF) recommendation framework, which is based on viewing user feedback on products as ordinal, rather than the more common numerical view. This way, we do not need to interpret each user feedback value as a number, but only rely on the more relaxed assumption of having an order among the different feedback ratings. Such an ordinal view frequently provides a more natural reflection of the user intention when providing qualitative ratings, allowing users to have different internal scoring scales. Moreover, we can address scenarios where assigning numerical scores to different types of user feedback would not be easy. Our approach is based on a pointwise ordinal model, which allows it to linearly scale with data size. The framework can wrap most collaborative filtering algorithms, upgrading those algorithms designed to handle numerical values into being able to handle ordinal values. In particular, we demonstrate our framework with wrapping a leading matrix factorization CF method. A cornerstone of our method is its ability to predict a full probability distribution of the expected item ratings, rather than only a single score for an item. One of the advantages this brings is a novel approach to estimating the confidence level in each individual prediction. Compared to previous approaches to confidence estimation, ours is more principled and empirically superior in its accuracy. We demonstrate the efficacy of the approach on some of the largest publicly available datasets, the Netflix data, and the Yahoo! Music data.