Fair Recommendation Method Based on Meta-Learning and Feature Enhancement

Existing recommender systems have difficulty in solving the cold-start problem, which can be mitigated by improvements in the data and model layers. At the data layer, the model needs to extract more significant user representations; at the model layer, the whole model framework needs to be optimized. However, cold-start recommendation can also be discriminatory, leading to unfair recommendation of user preferences. Existing recommendation models are generally unable to consider users and items comprehensively, and are difficult to cope with user privacy leakage and group fairness issues. In this paper, we design a recommendation model that can reasonably utilize user feature information and improve fairness, and name it MFCN. Unlike existing recommendation models, MFCN focuses on modeling user interaction history by integrating the items given by users in the interaction process and the user-embedded representations. On the other hand, MFCN utilizes generative adversarial learning to construct the model as a gaming problem, where the model uses a discriminator to obtain the user's sensitive information and recommendation preferences, and the entire recommendation model as a generator needs to keep the generated information from being captured by the discriminator, which in turn improves the fairness of the recommendation system. The MAML algorithm is introduced in the process of training to improve the efficiency of the recommendation model in generating generalization parameters, thus improving the accuracy and generalization of the cold-start model. In this paper, extensive experiments are conducted on relevant datasets, and the experiments show that in terms of mitigating user cold-start, user preferences can be more accurately captured compared to the traditional MELU model, and in terms of recommender fairness, the generated relevant embedded representations are more protective and private, while satisfying the group preferences.