Reproduce “Learning Diverse Fashion Collocation by Neural Graph Filtering” on OpenResearch

About this paper

Xin Liu, Yongbin Sun, Ziwei Liu, Dahua Lin

Researchers from Duke University, MIT, and The Chinese University of Hong Kong developed Neural Graph Filtering, a graph neural network framework designed to learn and generate diverse aesthetic fashion styles while supporting flexible item counts in outfits. The system achieved an 8% increase in compatibility AUC on Polyvore and a 17.4% increase in fill-in-the-blank accuracy on a reorganized Amazon Fashion dataset, with 82.5% of users preferring its diverse recommendations.

Key insights

• Modeling fashion collocation as an edge-centric graph filtering problem allows for permutation-invariant processing, accommodating varying outfit sizes and item types more effectively than sequential or fixed-structure models.
• Explicitly defining and training for diverse aesthetic styles, rather than solely relying on visual similarity, leads to more creative and user-preferred fashion recommendations.
• A combination of graph-based relational modeling and focal loss is effective in addressing the inherent data imbalance prevalent in fashion datasets, improving the recognition of minority styles.

Summary by alphaXiv.