Exploration & Exploitation: A Unified Framework for Data Purification & Augmentation in Recommendation Systems

Abstract

In the context of recommendation scenarios, the utilization of data purification and data augmentation methodologies has demonstrated their efficiency in enhancing the quality of representations. Nevertheless, within an open environment, the constrained nature of interaction data and the diverse range of interaction intentions present formidable challenges, giving rise to insufficient generalization ability and generalization bias in these methodologies. To address this issue, in this paper we introduce Exploitation & Exploration: A Unified Framework for Data Purification & Augmentation in Recommendation Systems, which not only ensures the precise purification of current data but also delves into potential noisy data for further exploration. To be specific, based on the traditional collaborative filtering method calculating user-item correlation, we first implement an efficient multi-head SENet block to remove potential noise from the interaction data. After this, we deploy a diffusion module to remove the added adversarial noise based on its ability to denoise all kinds of noise. And finally we use mutual-learning method to coordinate two parts' learning. We conducted experiments on three publicly available datasets, evaluating our model against current state-of-the-art algorithms in recommendation robustness tasks. The experimental results validate the effectiveness of our model.