• deep graph networks
• deep learning
• demand forecasting
• inductive learaning
• recommendation systems
• transductive learning

The assortments of global retailers are composed of hundreds of thousands of products among which several types of relationships, denoting compatible items, "bought together" products, or "watched together" items can be defined. Fashion experts often manually label compatible items to produce relations such as style compatibility. This process is exhaustive and impractical because the number of pairs of compatible items grows quadratically with the number of articles.

A natural way of representing assortments is in the form of graphs, whose nodes are products and relations among them are edges. Since these relations are produced manually by the fashion experts, they do not cover uniformly the whole graph, which implies that through this work we are going to deal with the sparse graphs, having a large number of zero-degree nodes (nodes that do not have any connections with the other nodes).

This work aims to propose a mechanism for enhancing very sparse graphs. We use inductive learning to enhance a graph encoding style compatibility of a fashion assortment, leveraging rich node information comprising textual descriptions and visual data. We show that the proposed graph enhancement improves substantially the performance on transductive tasks, with a minor impact on graph sparsity.