DSMN: An Improved Recommendation Model for Capturing the Multiplicity and Dynamics of Consumer Interests
Pengtao Lv, Zhenhan Guan, Qinghui Zhang, Jiale Gu, Shuai Zhang, Lei Lv, Mengya Zhang, Lei Li, Chengyang Zhao
Abstract
Electronic shopping has become an important way of shopping in our daily life. A good recommendation system can greatly improve the consumer shopping experience through alleviating consumer and product information overload problem. In the real scenario, each consumer has a wide variety of interests, and consumers interests are constantly changing. Many recent models usually give an overall embedding for consumer behavior sequence or only capture features of customers and characteristics of products from single perspective. However, these methods can hardly model consumers multiple interests and evolving of interests completely. Thus, in this study, we consider both the diversity and dynamics of consumer interests, and model them in a unified model. In light of this, we propose a novel model named Deep Self-Attention Mulit-Interest Networks (DSMN), which captures latent multiple interests by using capsule networks. It can perceive evolving of interests from consumer behavior sequences by using attention mechanism. We evaluate our model on three real-world datasets. In the experiments, our DSMN is 4.934% higher than the average of competitors in terms of AUC. Compared to our original silver medal method on the Kaggle competition, our DSMN is 18.978% higher in terms of MAP@12.