Graph-Based Relation-Aware Representation Learning for Clothing Matching

Yang Li, Yadan Luo, Zi Huang

Abstract: Learning mix-and-match relationships between fashion items is a promising yet challenging task for modern fashion recommender systems, which requires to infer complex fashion compatibility patterns from a large number of fashion items. Previous work mainly utilises metric learning techniques to model the compatibility relationships, such that compatible items are closer to each other than incompatible ones in the latent space. However, they ignore the contextual information of the fashion items for compatibility prediction. In this paper, we propose a Graph-based Type-Relational Neural Network (GTR-NN) framework, which first generates item representations through multi-layer ChebNet considering k-hop neighbour information, and then outputs compatibility score by predicting the binary label of an edge between two nodes under a specific type relation. Extensive experiments for two fashion-related tasks demonstrate the effectiveness and superior performance of our model.