Forgetting methods for incremental matrix factorization in recommender systems

Abstract

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.

Personalised fading for stream data

“…We not only experiment with time and rating-and-position based forgetting techniques, using individual fixed-size rating queues, but, additionally, apply fading both to the viewer model and to the viewer rating queue. Our results and those of Matuszyk (2015) [4] show that fading outperforms the sliding window results.…”

“…Live viewer profiling is typically performed using sliding windows ( [3,5]) or factor fading ( [4]) together with incremental matrix factorization. In particular, sliding windows are FIFO queues of dynamic or fixed size representing a number of viewer events [3] or a time span [5].…”

An overview on the exploitation of time in collaborative filtering

“…The distribution of the metrics naturally follows the type of data: precision, recall, F1, DCG, and MAP for positive‐only data, and RMSE and MAE for ratings. Two exceptions to this are Refs and that use Precision/Recall to evaluate recommmendations obtained with ratings data. One contribution uses an ad hoc, nonstandard accuracy metric.…”

“…In Ref , Matuszyk and Spiliopoulou proposed and evaluated several selective forgetting strategies for incremental MF algorithms with ratings data. In Ref , Matuszyk et al extended the study with several other forgetting methods and tested them with positive‐only data in addition to ratings data. In both publications, the authors show that selectively forgetting some of the past users' feedback is beneficial to the system.…”

Personalised Dynamic Viewer Profiling for Streamed Data

“…We implemented three main approaches: (i ) the static approach [2,8,1], which creates the SVD model off-line; (ii ) the global adaptive approach [13,12,7], which builds the initial SVD model off-line and applies SGD for incremental on-line update; and (iii) our individual adaptive approach, which generates the initial SVD model off-line, and uses personalised learning and over-fitting parameters with the on-line SGD incremental update. The three approaches were applied to the rating and positive feedback scenarios.…”