Dual Attention-based Interest Network for Personalized Recommendation System

With the rapid development of information technology, a large amount of resources can be obtained quickly anytime and anywhere, but these resources bring the difficulties for users because of the information explosion. Recommendation systems are often designed to help users filter the resources. At present, it is difficult to capture the potential user preferences from rich information of implicit feedbacks. To solve this problem, we propose Dual Attention-based interest Network for Recommendation system(DANR). Firstly, we use the dual attention mechanism to connect the clicked item representation and implicit feedback representation to obtain the user overall preference. Specially, implicit feedback information is extended and it helps express the user's interest in the candidate. Next, we mine the relationships between users and items from item-level and behavior-level, to get multiple representations of the user-item relevance. Finally, the click-through rate is calculated through multiple layer perceptron(MLP) after concatenating. Extensive comparative experiments are conducted with the main-stream recommendation models on the real dataset. The results show that, compared with the existing models, our model has achieved further improvement.

I. Introduction

In the Internet era, the sharing of massive resources has brought convenience to people's life [1], such as online shopping, online movie-going and e-book reading. But excessive information makes it difficult for people to select and filter resources because of information overload [2]. Therefore, recommendation systems are used to filter the resources for users and adopt the traditional recommendation algorithm below. content-based recommendation algorithm mines similar items for recommendation from the items selected or rated by users [3]. collaborative filtering-based recommendation algorithm assumes that similar users have similar preferences and mines users' potential preferences for recommendationp [4]. Hybrid recommendation algorithm combines two or more recommendation algorithms through the methods of weighting, switching, feature combination and so on [4]. However, traditional recommendation algorithm usually has shortcomings such as data heterogeneity, data sparsity, and cold start problems [5].