A deep hybrid model for optimization of user preferences in marriage matching platforms

In the context of the rapid development of marriage and love matching platforms, how to accurately understand user preferences and achieve efficient matching recommendations has become an important challenge for intelligent recommendation systems. Existing methods generally face problems such as difficulty in modeling the dynamic changes of user interests, insufficient utilization of social relationship information, and inflexible strategy updates, which limit the further improvement of recommendation quality. To this end, we propose VGRM-Net, which integrates variational autoencoders (VAE), graph neural networks (GNNs), and policy gradient reinforcement learning (Policy Gradient RL) to construct a user’s potential preference representation, integrate social network structure, and achieve dynamic strategy optimization driven by interactive feedback. The model can effectively capture the latent variable characteristics of user behavior, mine social influence patterns, and optimize matching strategies under long-term goals. Experimental results show that VGRM-Net significantly outperforms multiple benchmark methods on Douban and OkCupid public datasets, and achieves leading performance in multiple indicators such as F1-score and NDCG. The research results provide a more intelligent, personalized and sustainable optimization solution for marriage and love matching systems, which has positive significance for individualized recommendations and interactive intelligent system design.