Abstract of the Paper: A vast amount of user behavior data is constantly accumulating ontoday’s large recommendation platforms, recording users’ variousinterests and tastes. Preserving knowledge from the old data whilenew data continually arrives is a vital problem for recommendersystems. Existing approaches generally seek to save the knowledgeimplicitly in the model parameters. However, such a parametercentric approach lacks scalability and flexibility—the capacity ishard to scale, and the knowledge is inflexible to utilize. Hence, inthis work, we propose a framework that turns massive user behavior data to retrievable knowledge (D2K). It is a data-centric approachthat is model-agnostic and easy to scale up. Different from onlystoring unary knowledge such as the user-side or item-side information, D2K proposes to store ternary knowledge for recommendation,which is determined by the complete recommendation factors—user, item, and context. The knowledge retrieved by target samplescan be directly used to enhance the performance of any recommendation algorithms. Specifically, we introduce a Transformer-basedknowledge encoder to transform the old data into knowledge withthe user-item-context cross features. A personalized knowledgeadaptation unit is devised to effectively exploit the informationfrom the knowledge base by adapting the retrieved knowledge tothe target samples. Extensive experiments on two public datasetsshow that D2K significantly outperforms existing baselines and iscompatible with a major collection of recommendation algorithms.Original Paper: https://arxiv.org/pdf/2401.11478.pdf