Reproduce “Encode Me If You Can: Learning Universal User Representations via Event Sequence Autoencoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anton Klenitskiy, Artem Fatkulin, Daria Denisova, Anton Pembek, Alexey Vasilev
The Sber AI Lab developed an ensemble approach combining sequence autoencoders, collaborative filtering, and language model-based embeddings to create universal user representations. This method secured second place overall and first place on the academic leaderboard in the RecSys Challenge 2025, effectively modeling diverse user behaviors for robust performance across various predictive tasks.
Key insights
- The ensemble strategy significantly outperforms any single model, demonstrating that combining diverse representation learning techniques captures a broader spectrum of user behavior effectively.
- Sequence autoencoders, particularly GRU-based architectures, are highly effective for learning general-purpose user embeddings by compelling the latent vector to retain maximal behavioral information through reconstruction.
- Proper normalization of individual embeddings (e.g., unit-length for model-generated, quantile transformation for handcrafted) before concatenation is critical for maximizing the ensemble's performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.