BEHAV-E! You are Not Just a Number to Us, but an R^2048 Embedding

Abstract

BEHAV-E is a self-supervised architecture for learning transferable 2048-dimensional user representations from multi-event behavioural data. It combines temporal activity modelling, product and category embeddings, and LSTM-based representations of search semantics. Evaluated in the RecSys Universal Behavioural Modelling Challenge, the approach captures complex multimodal user behaviour in a compact embedding intended to support multiple downstream personalization tasks.

Publication
_RecSysChallenge ‘25: Proceedings of the Recommender Systems Challenge 2025, pp. 16-20, https://doi.org/10.1145/3758126.3758130_