Enhancing Rating Prediction by Discovering and Incorporating Hidden User Associations and Behaviors

Abstract

Collaborative filtering (CF)-based rating prediction would greatly benefit by incorporating additional user associations and behavioral similarity. This article focuses on infusing such additional side information in three common techniques used for building CF-based systems. First, multi-view clustering is used over neighborhood-based rating predictions. Secondly, additional user behavior knowledge discovered by mining user reviews are infused into non-negative matrix factorization (NMF) techniques. Finally, the article explores how to infuse such additional behavioral knowledge into a Deep Neural Network (DNN) based DF architecture. The article also explores using term frequency-inverse document frequency (TF-IDF) vectors as the input to DNN. Since TF-IDF does not directly capture the conceptual contents of the text or the behavioral aspects of the writer, the article also proposes a novel scheme called topic proportions-inverse entity frequency (TP-IEF) that uses topics discovered from reviews instead of words to better capture semantic associations between users and items.