Improving user-oriented fairness in recommendation via data augmentation: Don’t worry about inactive users

Abstract

A recommendation system is considered unfair when it does not perform equally well for different user groups according to users’ specific attributes. In recent research, the user groups are divided into active user group and inactive user group according to the number of interaction records in a recommendation system. Intuitively, increasing the number of inactive users’ interaction records would improve the fairness of the recommendation system. Existing data augmentation techniques can increase interaction records, however they usually fail to deeply mine user interaction patterns and fail to generate context-related feedback, which cannot effectively improve the quality of recommendations for inactive users. To resolve the problem, we use the Large Language Models (LLMs) to mine user historical interaction records to achieve data augmentation, which improve the quality of recommendations for inactive user groups. Experimental results on four classic baseline recommendation algorithms show that our data augmentation method for the inactive user group can effectively alleviate the poor recommendation quality caused by the low interaction with the recommendation system, reduce the recommendation quality gap with active user group, and further improve the user group fairness of the recommendation system.