Practical research on optimizing bodybuilding training plan and execution efficiency with transformer

Abstract

This study aims to address the issues of difficult modeling of users’ dynamic interests and low behavioral execution efficiency in personalized bodybuilding training recommendations. This study proposes a Transformer-xDeepFM (T-xDeepFM) model that integrates the Transformer structure with eXtreme Deep Factorization Machine (xDeepFM), constructing a training plan recommendation system centered on user behavior sequences. The T-xDeepFM model integrates the expressive ability of the Compressed Interaction Network (CIN) module in xDeepFM for explicit feature interaction with the advantages of Transformer in temporal modeling, achieving joint learning of high-dimensional sparse features and user interest evolution. On the MovieLens-1M and gym exercise datasets, T-xDeepFM outperforms traditional integrated learning models and existing deep sequence recommendation methods in recommendation performance, with AUC reaching 0.824 and 0.848 respectively, and Logloss decreasing to 0.424 and 0.427. In the field test of user behavior, 60 users are recruited for training tests. The results show that the completion rate of the recommended plans reaches 87.4%, the interruption rate is controlled at 8.2%, the average adaptation time is 3.7 days, and the satisfaction score is 8.6 points. Experiments verify the practical application potential of the proposed model in promoting fitness behavior execution and personalized health management. The study provides a feasible path and theoretical basis for the design of intelligent recommendation systems and user behavior intervention in the health field.