Abstract
Current recommendation systems are significantly affected by a serious issueof temporal data shift, which is the inconsistency between the distribution ofhistorical data and that of online data. Most existing models focus onutilizing updated data, overlooking the transferable, temporal data shift-freeinformation that can be learned from shifting data. We propose the TemporalInvariance of Association theorem, which suggests that given a fixed searchspace, the relationship between the data and the data in the search space keepsinvariant over time. Leveraging this principle, we designed a retrieval-basedrecommendation system framework that can train a data shift-free relevancenetwork using shifting data, significantly enhancing the predictive performanceof the original model in the recommendation system. However, retrieval-basedrecommendation models face substantial inference time costs when deployedonline. To address this, we further designed a distill framework that candistill information from the relevance network into a parameterized moduleusing shifting data. The distilled model can be deployed online alongside theoriginal model, with only a minimal increase in inference time. Extensiveexperiments on multiple real datasets demonstrate that our frameworksignificantly improves the performance of the original model by utilizingshifting data.