Top-N-Targets-Balanced Recommendation Based on Attentional Sequence-to-Sequence Learning

Abstract

User's behaviors and preferences alter with the temporal evolution dynamically, which leads to low performance, such as the Hit Rate and Normalized Discounted Cumulative Gain (NDCG). Understanding the dynamics of users' behaviors and preferences can improve the performance of recommendation system. In this paper, we propose a Top-N-targets-balanced recommendation based on attentional sequence-to-sequence (Seq2Seq) learning to capture the users' transient interests. The attentional Seq2Seq learning is introduced to largely exploit the coherence of users' sequential behaviors and preferences in the out module of our sequential recommendation. We use two methods to get the Top-N outputs as inputs of the attentional Seq2Seq learning. The direct selection method is choosing the k items with high probability. The simulated generation method is an improvement of the network by simulating the output. We balance the loss between Top-N outputs and the sequence targets to train the neural networks, which include Long Short-Term Memory and attentional Seq2Seq learning. Besides, we modify the recommendation list generation method to further improve the performance. Experimental results demonstrate that our methods outperform existing algorithms including state-of-the-art NCF, ItemPop, ItemKNN, BPR, eAls, and SVD++ on the performance of HR and NDCG. In the best case, the NDCG of our proposal is 14.72% higher than that of NCF.