Fast Incremental Matrix Factorization for Recommendation with Positive-Only Feedback

Abstract

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.

Personalised Dynamic Viewer Profiling for Streamed Data

“…The results show that our algorithm outperforms the off-line matrix factorization algorithm proposed by Barragáns et al (2010). Regarding the on-line incremental matrix factorization algorithm of Vinagre et al (2014), our algorithm shows improved results in the positive feedback scenario and in the rating feedback scenario with the ML 100k; whereas in the case of the rating feedback scenario with the ML 1M, our individual model updating displays a 3 % increase in prediction errors and a 5 % decrease in T Recall@10.…”

“…The matrix factorization and the learning techniques are SVD and SGD, respectively. The prequential protocol verifies, every time a new rating event occurs, if the rated item would have been recommended to that viewer and, if yes, counts as a hit [11]. In 2015, these authors included a rating-and-recency-based scheme to perform negative preference imputation [12].…”

“…Instead of the standard recall metric used by [3,11], which considers a subset of the top rated items, we defined yet another recall based metric -Target Recall (TRecall) -which contemplates a subset of items centered around the target rating. TRecall evaluates the accuracy of the predictions when compared with the actual viewer ratings.…”

From Think Parallel to Think Sequential

“…GRAPE parallelizes CF by adopting SGD [23] as PEval, and the incremental algorithm ISGD of [34] as IncEval, using master P0 to synchronize the shared factor vectors u.f and p.f .…”

“…(2) IncEval is the incremental algorithm ISGD of [34]. Upon receiving message Mi at worker Pi, it computes Fi ⊕ Mi by treating Mj as updates to factor vectors of nodes in Fi.I, and only modifies affected factor vectors as in PEval based solely on new observations.…”