Fast Matrix Factorization for Online Recommendationrnwith Implicit Feedback

摘要

This paper contributes improvements on both the eu000bectivenessrnand eu000eciency of Matrix Factorization (MF) methodsrnfor implicit feedback. We highlight two critical issues of existingrnworks. First, due to the large space of unobservedrnfeedback, most existing works resort to assign a uniformrnweight to the missing data to reduce computational complexity.rnHowever, such a uniform assumption is invalid inrnreal-world settings. Second, most methods are also designedrnin an ou000fine setting and fail to keep up with the dynamicrnnature of online data.rnWe address the above two issues in learning MF modelsrnfrom implicit feedback. We frst propose to weight the missingrndata based on item popularity, which is more eu000bectivernand rexible than the uniform-weight assumption. However,rnsuch a non-uniform weighting poses eu000eciency challenge inrnlearning the model. To address this, we specifcally designrna new learning algorithm based on the element-wisernAlternating Least Squares (eALS) technique, for eu000ecientlyrnoptimizing aMF model with variably-weighted missing data.rnWe exploit this eu000eciency to then seamlessly devise an incrementalrnupdate strategy that instantly refreshes a MF modelrngiven new feedback. Through comprehensive experimentsrnon two public datasets in both ou000fine and online protocols,rnwe show that our eALS method consistently outperformsrnstate-of-the-art implicit MF methods.