Popularity bias and Explainability in Recommendation Systems

Abstract

Consumer behaviour has changed as a result of how digital technology has changed con sumer behaviours. As businesses have grown more aware of the influence of technology on consumer behaviour, they have been keen to improve their services with a focus on creating positive customer experiences. There is a need to filter, prioritise, and efficiently communicate critical information on the Internet, where the variety of alternatives is overwhelming, in order to alleviate the problem of information overload, which has emerged as a potential issue for many Internet users. To solve this problem, recommender systems sift through massive volumes of constantly created data and offer users with relevant content and services. One of the important is sues in the recommendation systems is popularity bias, which has a negative effect on both consumers and item producers. The study offers a summary of the recommendation sys tem in addition to an empirical analysis of the popularity bias. It also emphasizes the fact that not all bias is harmful. Unbiased learning pursued blindly could eliminate patterns in the data, lowering accuracy and user satisfaction. A statistical approach, Popularity bias Deconfounding and Adjusting (PDA) is used to alter recommendation scores via causal intervention and reduce confounding popular bias from model training. Finally, to add fair ness to the recommendation system, a linear collaborative filtering recommendation model that makes use of user profiles within the context of item metadata is used. The idea be hind the study is to learn the feature embedding from the tags to better understand the use behavior; for that, a well-known TEASER model is used, and that can be fused with any user-item matching function.