Attention Neural Network for User Behavior Modeling

Abstract

The recommendation system can effectively and quickly provide valuable information for users by filtering out massive useless data. User behavior modeling can extract all kinds of aggregated features over the heterogeneous behaviors to help recommendation. However, the existing user behavior modeling method cannot solve the cold-start problem caused by data sparse. Recent recommender systems which exploit reviews for learning representation can alleviate the above problem to a certain extent. Therefore, a user behavior modeling is proposed for recommendation task using attention neural network based on user reviews (AT-UBM). Firstly vanilla attention was used to sample reviews, and then CNN+Pooling method was applied to extract user behavior features. Finally the long-term behavior was combined with short-term behavior in feature spaces. Experimental results on real datasets show that the review-based user behavior model has better prediction accuracy and generalization capability.