Federated Learning for an Autoencoder Based Recommender System for Look-Alike Modeling

This work provides the design of an Autoencoder Based Recommender System, created in the context of digital advertising, where it aims at suggesting new categories to users based on their previous URLs browsing history. We propose a particular training procedure that considers recency, frequency and time of the click views. We tuned the architecture considering different features regarding the browsed web pages: we consider also textual information, such as the page title and keywords. This information turns out to be relevant for our purpose. Furthermore, this project deals with the privacy issues related to the end of third-party cookies and the GDPR: the Federated Learning provides a solution for the segmentation of the users that takes places completely on the user device. We compared multiple algorithms (FedSGD, FedAdam and FedAVG) noticing an advantage of FedAdam. We also aim at refining a pre-trained model with federated learning, with the purpose to adapt the model on new data: we propose a solution for a proper training of the model in this setting.