Retrieval and Distill: A Temporal Data Shift-Free Paradigm for Online Recommendation System

TL;DR

This paper introduces a novel recommendation framework that leverages temporal invariance principles to mitigate data shift issues, improving online recommendation accuracy while maintaining low inference costs through a distillation approach.

Contribution

The paper proposes a temporal invariance theorem and a retrieval-based framework that learns shift-free relevance, along with a distillation method to deploy efficient online models.

Findings

Significant performance improvements on multiple datasets.

Effective mitigation of temporal data shift effects.

Low inference overhead with the distilled model.

Abstract

Current recommendation systems are significantly affected by a serious issue of temporal data shift, which is the inconsistency between the distribution of historical data and that of online data. Most existing models focus on utilizing updated data, overlooking the transferable, temporal data shift-free information that can be learned from shifting data. We propose the Temporal Invariance of Association theorem, which suggests that given a fixed search space, the relationship between the data and the data in the search space keeps invariant over time. Leveraging this principle, we designed a retrieval-based recommendation system framework that can train a data shift-free relevance network using shifting data, significantly enhancing the predictive performance of the original model in the recommendation system. However, retrieval-based recommendation models face substantial inference time costs when deployed online. To address this, we further designed a distill framework that can distill information from the relevance network into a parameterized module using shifting data. The distilled model can be deployed online alongside the original model, with only a minimal increase in inference time. Extensive experiments on multiple real datasets demonstrate that our framework significantly improves the performance of the original model by utilizing shifting data.