Research On Fairness Of Recommendation System Based On Knowledge Graph

Knowledge-aware recommendation system has attracted considerable interest in academia and industry,which comes in handy to solve the cold-start problem and offer a reliable solution for the business to grow.This kind of recommendation system can not only solve the problem of cold start well but also provide reliable information for the growth of enterprise business.Solutions,our social life is inseparable from it.Therefore,when designing and using these systems,we need to consider the issue of fairness,to avoid these systems from discriminating against certain individuals or groups,or even deepening existing discrimination.However,we found that even without explicitly introducing the protected sensitive attribute information into the knowledge graph,the recommender system can still capture these sensitive attribute information,resulting in unfair recommendation results.Most of the existing debiasing methods require complex model design or can only be applied to a specific base model,and the methods can only deal with the fairness constraint of a single sensitive attribute,which is difficult to be used in debiasing scenarios with multiple sensitive attributes.However,some methods that can be used for fairness constraints on multiple sensitive attributes simply combine a single sensitive filter linearly to implement multiple sensitive attribute debiasing work,ignoring the potential correlation between sensitive attributes,and cannot guarantee that any multiple sensitive attribute groups can Satisfy fairness constraints.In addition,some methods have special requirements on the fairness of the input data,which are difficult to apply to real-life recommendation scenarios.To address the above issues,we propose a fair representation learning model suitable for arbitrary knowledge graph recommender systems.We learn latent relationships between sensitive attributes by introducing a sensitivity graph to flexibly adapt to fairnessconstrained scenarios with multiple sensitive attributes.Specifically,given a knowledge graph-based recommendation model,we first use the protected sensitive attributes as nodes to dynamically learn the hidden relationships between sensitive attribute nodes to construct the sensitivity graph? then we combine the representation of the sensitivity graph with the original The representation of the knowledge graph is fused,and without changing the original recommendation algorithm,an adversarial learning framework is introduced to remove the sensitive attribute information in the user representation generated by using the sensitivity graph and the original knowledge graph? Extensive experiments have been conducted on the recommendation dataset,and the results show that the proposed model can improve the fairness performance of any knowledge graph-based recommendation system without losing the recommendation performance of the system.