SEMAX: Multi-Task Learning for Improving Recommendations

Abstract

Personalization plays an essential role in recommender systems, in which the key task is to predict personalized ratings of users on new items. Recently, a lot of work investigates deep learning-based collaborative filtering techniques to increase the accuracy of rating prediction. However, most exiting works focus on the recommendation task itself. Actually, the multi-task learning exploits an inductive transfer mechanism to enhance the generalization performance of the main task by using the domain information contained in other related tasks. Multi-task learning has shown effectiveness in various real-world problems, including regression and machine translation. To this end, this study proposes a new framework, called SEMAX that extends our previous model SEMA via multi-task learning for improving recommendations, in which the recommendation task gets the domain information from the other task. Specifically, in the recommendation task, SEMAX learns semantic meanings from texts and temporal dynamics from text sequences for both users and items based on our developed hierarchical and symmetrical recurrent neural networks (RNNs) with the long short-term memory. Furthermore, SEMAX exploits the related task that predicts the rating of a text written by a user for an item to reinforce the recommendation task that predicts the rating of the user on the item, because the text can be an important predictor of the rating given by the user to the item. Moreover, SEMAX predicts the rating of a text based on an attention mechanism to choose user-item-specific words so as to generalize the performance of learned word embeddings, user and item representations. Finally, we conduct a comprehensive evaluation for SEMAX using two large-scale real-world review datasets collected from Amazon and Yelp. The experimental results show that the SEMAX achieves significantly superior performance compared to other state-of-the-art recommendation techniques.