인기도 편향 문제 해결을 위한 임베딩 전이 기법

Abstract

Datasets used in the recommendation system are collected based on the user's behavioral history, and various biases are included in this process in addition to the user's preference. Recommendation models learned through these datasets are also recommended by learning the bias contained in the dataset, and popularity bias mean that some popular items are highly recommended regardless of the user's preference, and most of the remaining items are ignored. Various methods have been attempted to solve this problem, mostly suggesting that the overall accuracy is improved by increasing the accuracy of the remaining items at the expense of the accuracy of some of the popular items. In this paper, we propose a method to improve accuracy by reducing popularity bias without sacrificing the accuracy of some popular items based on the user's preference related to popularity.