Behavior-based safety (BBS), which contains definition, observation, intervention, and test, is proven to be useful in safety management on-site. Safety training is regarded as an effective method for the BBS intervention. However, existing studies provide little attention to workers' personal behavior patterns that limit the persistent effectiveness of interventions. This paper proposes a hybrid recommendation approach that can enhance the effectiveness of safety training in the Chinese construction industry. The proposed approach integrates content-based personalized recommendation and MapReduce-based collaborative filtering for personalized recommendation. The personal behavior patterns of workers are automatically analyzed through data that are generated from training. Then, reasonable training materials are recommended for behavior modification. The workers' behavior modification (WoBeMo) system is designed to implement safety training. A pilot study on metro construction sites showed that the unsafe behavior rate (S) of 20 laborers who were engaged in scaffolding work decreased over 70% after intervention, and were reduced over 60% compared with another group of laborers not using the WoBeMo. Application results indicate the feasibility and practicability of the proposed approach on modifying workers' unsafe behaviors and improving safety performance on construction sites.