Outfit Recommendation using Graph Neural Networks via Visual Similarity

Abstract

Computer vision plays an important role in the development of the fashion industry. There has been a lot of research done on various fashion recommendations, and determining the compatibility of clothing is a key factor in most of them. Solving this problem can help users buy items that go well with their current wardrobe, and help stores sell multiple clothing items at once. Previous research has mainly focused on learning compatibility between two clothing elements. There are several approaches that take into account the outfit as a whole but they require rich textual data. In this work, we only use images of clothing from the Polyvore dataset to extract visual features. By representing outfits in the form of a graph, we train node embeddings based on graph structure and node features. We then train multi-layer perceptron to classify the set of embeddings representing the outfit. We compare our method with the relevant works in two tasks: outfit compatibility prediction and fill-in-the-blank. Our approach showed the best result among approaches that use only images on the first task and showed state-of-the-art result on the second task.