Controlling Diversity at Inference

Abstract

Due to the feedback loop, conventional recommendation systems fail to respond to users’ intentions on the fly. Furthermore, solely prioritizing accuracy inevitably results in the filter bubble owing to the imbalanced category preference distribution observed in user interactions. One approach to handle these issues is controllable di- versified recommendation system, which aim to adjust the diversity of recommenda- tion lists. There have been several efforts to control the diversity in the system. How- ever, most existing methods do not remove biased category preference data within users’ past interactions. As a result, these methods can not rapidly adapt to the de- sired diversity level and usually improve diversity at the expense of accuracy. To tackle the above problems, we propose D3Rec (Disentangled Diffusion model for Diversified Recommendation) to control the diversity at inference phase, while miti- gating accuracy-diversity dilemma. Functionally D3Rec is composed as follows: 1) Corrupting user interactions by diffusion forward process to eliminate biased cate- gory preference information. 2) Generating users’ future interactions guided by their targeted category preferences to control the diversity and resolve the dilemma. 3) Disentanglement to capture a fine-grained diversity-aware feature from users’ inter- action and category preferences. Experiments conducted on three real-world datasets demonstrate the effectiveness of D3Rec. At last, we expect to improve user satisfac- tion by enhancing the controllability of the recommendation system and mitigating the accuracy-diversity dilemma.