Traditional Collaborative Filtering algorithms for recommendation are designed for stationary data. Likewise, conventional evaluation methodologies are only applicable in offline experiments, where data and models are static. However, in real world systems, user feedback is continuously being generated, at unpredictable rates. One way to deal with this data stream is to perform online model updates as new data points become available. This requires algorithms able to process data at least as fast as it is generated. One other issue is how to evaluate algorithms in such a streaming data environment. In this paper we introduce a simple but fast incremental Matrix Factorization algorithm for positive-only feedback. We also contribute with a prequential evaluation protocol for recommender systems, suitable for streaming data environments. Using this evaluation methodology, we compare our algorithm with other stateof-the-art proposals. Our experiments reveal that despite its simplicity, our algorithm has competitive accuracy, while being significantly faster.

Numerous stream mining algorithms are equipped with forgetting mechanisms, such as sliding windows or fading factors, to make them adaptive to changes. In recommender systems those techniques have not been investigated thoroughly despite the very volatile nature of users' preferences that they deal with. We developed five new forgetting techniques for incremental matrix factorization in recommender systems. We show on eight datasets that our techniques improve the predictive power of recommender systems. Experiments with both explicit rating feedback and positive-only feedback confirm our findings showing that forgetting information is beneficial despite the extreme data sparsity that recommender systems struggle with. Improvement through forgetting also proves that users' preferences are subject to concept drift.

Classic Collaborative Filtering (CF) algorithms rely on the assumption that data are static and we usually disregard the temporal effects in natural user-generated data. These temporal effects include user preference drifts and shifts, seasonal effects, inclusion of new users, and items entering the system-and old ones leaving-user and item activity rate fluctuations and other similar time-related phenomena. These phenomena continuously change the underlying relations between users and items that recommendation algorithms essentially try to capture. In the past few years, a new generation of CF algorithms has emerged, using the time dimension as a key factor to improve recommendation models. In this overview, we present a comprehensive analysis of these algorithms and identify important challenges to be faced in the near future.