Fashion Outfit Compatibility
Fashion recommendation by estimating compatibility of items in an outfit
This project was done as part of the course CSE 6240 Web Search and Text Mining at GeorgiaTech.
In this project, we follow two existing approaches that employ graphs to represent outfits and use modified versions of the Graph neural network (GNN) frameworks. Both Node-wise Graph Neural Network (NGNN) and Hypergraph Neural Network (HGNN) aim to score a set of items according to the outfit compatibility of items
We recreated the two existing architectures and compare their performance on two tasks
- Fill-in-the-blank (FITB): finding an item that completes an outfit
- Compatibility prediction: estimating compatibility of different items grouped as an outfit
You can read our full report in the github repository, but to be brief we used two different methods involving GNNs to create embeddings. A GNN learns a target node’s representation by propagating the neighbouring information in an iterative manner thus ensuring similarity of representation of items linked closely:
- The first approach called Node-wise Graph Neural Network (NGNN) improves on GNN by eschewing parameter sharing to capture item characteristics better. It uses the category of the item as a mainstay in message propagation by using weighted edges based on the category of nodes and also using category-specific item representation. Finally, it uses an attention mechanism to score the outfit.
- The second approach, HGNN (Hypergraph-GNN), uses hyperedges to denote each outfit. A hypergraph is a generalization of a graph in which an edge can join any number of nodes. Using this approach, theoretically, captures more complex interactions between each item in an outfit leading to key features being captured accurately in the embeddings.