Context-Aware Graph Convolutional Network with Multi-Clusters Mini-Batch for Link Prediction

Abstract: Predicting which fashion items can compose an outfit is not a trivial task since every person has different preferences based on their experiences, location, etc., determining each personal style. The items that can formulate an outfit have to be compatible, but the compatibility of two products should vary based on the context of the products, where context is defined as the products that we already know that are compatible with them. The authors in [1] introduced a model named Context-Aware-Graph-Convolutional-Network (CA-GCN), where they predict the compatibility among fashion items, by taking into consideration their context, after being trained with full-batch Gradient Descent (GD), making its application more difficult on denser and larger graphs. The authors in [2] have proposed the Cluster-GCN method to train GCN [3] models with mini-batch GD for node classification. We propose a novel model named CA-Cluster-GCN that merges a modified version of the Cluster-GCN with the CA-GCN model. We modify the Cluster-GCN in such a way that we can apply it for link-prediction since the CA-GCN predicts the compatibility among the fashion items by predicting if an edge exists among them. Our CA-Cluster-GCN achieves State-Of-The-Art performance on the public Maryland Polyvore [4] dataset and it outperforms the CA-GCN model on the public Amazon [5, 6] dataset. Additional tests have been performed on the industrial datasets provided by H&M.