LAUSR

The existing review-based recommendation methods usually employ the same model to learn the review representation of users and items. However, for different user-item pairs, the same words or similar reviews may deliver different information, and hence have different degrees of importance. Besides, many reviews are often noisy and even misleading. Hence, this paper proposes a multi-task Dual Attention Recommendation Model using Reviews and their Helpfulness (DARMH). The model separately constructs the local and interactive attentions to extract the personalized preference of specific users for specific items. By defining the helpfulness of a review, the attention weight of the review is acquired to better extract the features of the items review. However, many reviews do not have helpfulness ratings. Therefore, DARMH simultaneously performs multi-task learning of review helpfulness and rating prediction. The results of the experiments performed on four real datasets of Amazon show that the proposed DARMH model has a lower mean square error (MSE), and it achieves a performance improvement of 3.9%-5.4% compared to other review-based rating prediction algorithms.