A team of students from Politecnico di Milano, coordinated by Maurizio Ferrari Dacrema, researcher at the Department of Electronics, Information and Bioengineering, and Andrea Pisani, PhD student, won the best academic team prize at the RecSys Challenge 2025 for the third consecutive year. The team is composed by Simone Colecchia, Mauro Orazio Drago, Jihad Founoun, Paolo Gennaro, Ernesto Natuzzi, Luca Pagano, Sajjad Shaffaf, Giuseppe Vitello.
The RecSys Challenge 2025 lasted three months, with the final phase in Prague from September 22 to 26, 2025 as part of the 19th ACM Conference on Recommender Systems, and was organized by Synerise, a Polish company developing an AI-driven behavioral data platform, in collaboration with Maastricht University, the Indian Institute of Management Visakhapatnam, the Technical University of Denmark, the University of Bari Aldo Moro, and Politecnico di Bari.
This year’s challenge focused on creating universal behavioral profiles, that is user representations able to generalize across multiple tasks. Teams generated a per-user embedding from a large e-commerce event database that included purchases, add and remove from cart, page visits, and searches. The representations were evaluated on tasks disclosed to participants, such as churn prediction and purchase propensity, as well as on hidden tasks. The key aspect was the ability of these profiles to solve multiple heterogeneous tasks of industrial relevance.
Teams had access to a large real-world dataset provided by Synerise that spans six months of activity from a major online retailer and contains about 170 million events across page visits, searches, and transactions. Profiles were generated for a subset of one million users.
Approximately 400 teams participated, with affiliates from leading institutions and companies such as Université Paris-Saclay, Johannes Kepler University, East China Normal University, NTT DOCOMO, Mitsubishi Research Institute, and Match Group. The team from Politecnico di Milano presented their solution in the paper “From Sequences to Profiles: Generating Universal Behavioral Profiles exploiting Recurrent Neural Networks”