Dynamic Edge-Based Graph Attention Network and Visual-Semantic Embedding for Outfits Recommendation

Currently, outfit compatibility matching used in fashion e-commerce has been based on visual and textual information to learn global features from items to choose from, which has ignored local feature relationships between spatially adjacent items. We propose a new method that combines a dynamic graph neural network and visual semantic embedding to effectively discover the latent fashion compatibility. It takes visual and semantic information of items as input and extracts global and local information of items through the network to learn the compatibility of the outfit. This method enhances the local connections of items in an outfit by capturing the relationship between items, thus revealing potential information about the outfit collocation at multiple dimensions. Experimental results conducted on a real-world Polyvore dataset show that our proposed approach outperforms all the state-of-the-art methods in both accuracy and efficiency.