A dynamic and optimal construction method for cross-domain user profiles under multi-intelligent body reinforcement learning

For cross-domain users, characterizing user profiles solely based on basic attributes is ineffective and of low quality. Therefore, a method for dynamically optimizing cross-domain user profiles using multi-agent reinforcement learning is proposed. Using a multi-agent reinforcement learning model, multi-source text topic mining is performed. After data preprocessing, model evaluation, and label quality scoring, cross-domain user profile labels are obtained. Dynamically optimized attributes are defined across multiple dimensions, including basic attributes, cross-domain consumption preferences, social interaction preferences, and interest migration, encompassing multiple characteristics. To characterize users, behavioral labels are deeply integrated with dynamically optimized attributes. User similarity is calculated using the cosine similarity algorithm. Multi-agent reinforcement learning is employed to capture the deeper characteristics of user behavior, ultimately completing the dynamic optimization construction of cross-domain user profiles. Experimental results show that a social domain weight of 0.88 for user U01 improves recommendation accuracy by 17.3%. After 150 rounds of training, the average user reward value increases by 214.6%. Compared with the literature [1]-[4], the design method has the highest task execution efficiency, with a hit rate of 92.1% for 100 users and 73.6% for 800 users, and the standard deviation is controlled within 2.1%, proving that it can enhance the quality of user profile construction.