Dual Attention Modeling between Users and Items Using Review for Recommendation

Recommendation systems were developed to predict the users’ preferences by tracking the interactions between users and items. While the interaction history and the basic information have been the prevailing data of the predictive models, other available types of information, like user feedback, could also provide abundant and subjective information. However, there is no efficient way to extract emotional information, and hence user comments are rarely fully utilized. With the recent development of Natural Language Processing (NLP) in sentiment extraction, user and item reviews could be processed to sequential word embeddings via Transformer encoders and then further extracted as features, which enables itself as a piece of auxiliary information added to the recommendation systems. Thus, this paper aims to propose a dual attention model using review for recommendation called DAM4R, which employs deep bidirectional self-attention to model user preferences and item characters respectively from review text and predict the score via an MLP layer. Through evaluation on Yelp and Amazon datasets, we show that our model can match the predictive performance of prior work.