Diversify or Not: Dynamic Diversification for Personalized Recommendation

Diversity is believed to be an essential factor in improving user satisfaction in recommender systems, while how to take advantage of it has long been a problem worth exploring. Existing work either ignores the influence of diversity or overlooks users’ different diversity demands in recommendations. In this study, we analyze users’ behaviors on a real-world dataset collected from an e-commerce website and find that the demand for diversity changes among different users, even the same user’s demand varies among different shopping scenarios. There is also evidence that users’ behaviors are affected by the diversity of impressions, which has been often ignored by traditional session-based recommendation models. Then, we propose a Dynamic Diversification Recommendation Model (DDRM) with the integration of both click and impression diversities to better make use of diversity for recommendations. Extensive experimental results demonstrate that DDRM outperforms all baseline methods significantly. Further studies show our model can provide more precise and reasonable recommendations.