Rethinking Purity and Diversity in Multi-Behavior Sequential Recommendation from the Frequency Perspective

TL;DR

This paper challenges the traditional view that high-frequency information in multi-behavior recommendation is noise, showing it also captures user interest diversity, and proposes a model to leverage this for better recommendations.

Contribution

The paper introduces PDB4Rec, a model that extracts and balances information across frequency bands, emphasizing the importance of high-frequency data in user interest modeling.

Findings

High-frequency information correlates with user interest diversity.

Low-frequency information relates to user interest purity.

PDB4Rec outperforms existing models in real-world datasets.

Abstract

In recommendation systems, users often exhibit multiple behaviors, such as browsing, clicking, and purchasing. Multi-behavior sequential recommendation (MBSR) aims to consider these different behaviors in an integrated manner to improve the recommendation performance of the target behavior. However, some behavior data will also bring inevitable noise to the modeling of user interests. Some research efforts focus on data denoising from the frequency domain perspective to improve the accuracy of user preference prediction. These studies indicate that low-frequency information tends to be valuable and reliable, while high-frequency information is often associated with noise. In this paper, we argue that high-frequency information is by no means insignificant. Further experimental results highlight that low frequency corresponds to the purity of user interests, while high frequency corresponds to the diversity of user interests. Building upon this finding, we proposed our model PDB4Rec, which efficiently extracts information across various frequency bands and their relationships, and introduces Boostrapping Balancer mechanism to balance their contributions for improved recommendation performance. Sufficient experiments on real-world datasets demonstrate the effectiveness and efficiency of our model.