Research on the prospects of optimizing recommendation systems by restricting the retrospective of training data

This article presents an empirical study investigating the feasibility of optimizing recommendation systems by reducing the volume of training data through limiting the temporal retrospective of user interaction records. The research addresses key challenges arising from continuously growing datasets in operational services, such as increased computational complexity, model training costs, and conceptual drift in user preferences. Utilizing the open yambda dataset from Yandex.Music streaming platform, which contains anonymized timestamped user interactions over a ten-month period, the study experiments with two baseline models: DecayPop and itemKNN. Two primary strategies for restricting data retrospect were implemented and compared: a global strategy, which truncates data based on the latest timestamp across all users, and a personalized strategy, which sets individual cut-offs per user based on their last interaction. Model accuracy was evaluated using standard metrics, including NDCG@N, DCG@N, and Coverage@N for N values of 10, 50, and 100. The results demonstrate that strategically limiting the training data timeline can indeed improve recommendation accuracy for certain models, with an optimal retrospective depth identified at approximately 7–8 weeks for the studied dataset. This work confirms the potential of this optimization method and suggests its prospective inclusion in the recommendation system design pipeline, while noting the need for further validation on more complex models.