TUM Data Innovation Lab

Graph learning has recently become more popular in the field of recommender systems, as many of the problems that recommender systems try to solve can be modeled as a graph. For fashion recommendations, we typically deal with very large and sparse data sets that make it especially challenging to build high-quality personalized recommender systems. Another challenge in fashion is fairness, as we usually deal with sensitive customer data. This project aims to implement a fashion recommender system using graph-based learning to provide personalized recommendations for customers with fairness evaluation. For this purpose, we explore the public data set from H&M, which contains article data (e. g. name, description), customer data (e. g. age, membership), and their transactions. To capture the complex relationships between customer, article, and their associated properties, we designed a graph database and implemented three graph-based recommender approaches: Random Walk as our baseline, a Graph Embedding, and a Graph Neural Network approach. We evaluated and compared these strategies in terms of how relevant and fair the recommendations are. Our results show that different approaches performed best for differing metrics and demonstrate the potential for graph-based recommender systems in the fashion domain.