Personalized Neural Usefulness Network for Rating Prediction

Abstract

Reviews are crucial to rating prediction on e-commerce websites such as Amazon and Yelp. They reflect user preferences and item properties. Considering not all parts of the reviews are equally important for rating prediction, some works select the more informative words of reviews, others focus on finding the more useful reviews for each user and item. However, on the one hand, the meaning of words may be incomplete, further they may distort the original meaning of sentences, on the other hand, for the more useful reviews, they still contain more useful parts and less useful parts for rating prediction. To address these problems, we propose to model reviews in a view of sentence usefulness, since different sentences have varied importance for reviews modeling. In addition, the preferences of a user may vary with different item properties, it leads to the inconstant usefulness of a user sentence for different items, and vice verse. Hence, in this paper, we show a Personalized Neural Usefulness Network (PUN) to capture the varied sentence usefulness of reviews for rating prediction. Expensive experiments on benchmark datasets demonstrate that PUN significantly outperforms the state-of-the-art baseline models. In the meantime, according to personalized sentence usefulness, PUN also attains highly-useful sentences to better understand the varied user preferences and item properties.