I am Daniel Puente Viejo.

This challenge has posed the problem of carrying out an analysis of the garments of the Lookiero ecommerce, as well as relating these based on their characteristics. However, taking into account that Lookiero's business is to send looks to its customers, graph theory will be used to facilitate the selection of products for shoppers.
In this way, the aim is to relate the garments in a correct way, so that the customers are satisfied with the personalised selection that has been created for them and so that the amount of products that are returned can be reduced. Consequently, highlighting which combinations of garments are the most successful for each client.

The idea focuses on the possibility that a garment with specific characteristics that is not found in the database enters the scene. The model is able to relate this garment to the most appropriate ones once it has learned the established rules, and even predict the weight of these links, the latter is optional, but gives added value to the personal shopper who can see which of the links she is forming are robust and which are not.

To carry out this link prediction, GNNs using the pytorch-geometric library have been used. Once the links are predicted, their weights can be calculated using a Catboost model, the input data being the embbeddings generated by the network.