Learning Diverse Fashion Collocation by Neural Graph Filtering

Fashion recommendation systems are highly desired by customers to find visually-collocated fashion items, such as clothes, shoes, bags, etc. While existing methods demonstrate promising results, they remain lacking in flexibility and diversity, e.g. assuming a fixed number of items or favoring safe but boring recommendations. In this paper, we propose novel collocation framework, Neural Graph Filtering , that models flexible set via graph neural network. Specifically, consider the visual embeddings each garment node graph, describe inter-garment relationship edge between nodes. By applying symmetric operations on vectors, framework allows varying numbers inputs/outputs is invariant their ordering. We further include style classifier augmented with focal loss enable significantly diverse styles, which inherently imbalanced training set. To facilitate comprehensive study collocation, reorganize Amazon dataset carefully designed evaluation protocols. evaluate proposed approach three popular benchmarks, Polyvore dataset, Polyvore-D our reorganized dataset. Extensive experimental results show outperforms state-of-the-art over 10% improvements standard AUC metric. More importantly, 82.5% users prefer diverse-style recommendations other alternatives real-world perception study.