Fashion Outfit Recommendation using Graph Neural Network

Abstract

Recently, as the demand for online shopping has increased, the outfit recommendation service has played an important role in driving customer purchases in the online fashion industry and outfit recommendation is actively being researched. However, most of the relevant studies focus on compatibility learning, and most studies related to outfit generation are based on Bi-LSTM. Bi-LSTM based model rely heavily on the fixed order of the outfit items, resulting in incompatible outfit generation. In this study, we proposed a method to generate fashion outfits using graph neural network to overcome the limitations of Bi-LSTMs. We propose two embedding method to encode a new item feature into a node feature. The first method is to learn a multi-layer perceptron that transforms an image feature into a node feature. The MLP model is trained using the image feature as input and the node features as label extracted from the existing GNN-based outfit learning model. The second method is to find the most similar item in the outfit graph based on the cosine similarity and use the surrounding information associated with the item. Finally, we propose a method of generating multiple outfits by receiving one or more input items using the embedded node feature. To verify the proposed method, we conducted experiments with real data. As a result, this method has quantitative performance similar to the base model, but it recommends a more compatible outfit than the base model.