Reproduce “Abacus: Self-Supervised Event Counting-Aligned Distributional Pretraining for Sequential User Modeling” 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

Sullivan Castro, Artem Betlei, Thomas Di Martino, Nadir El Manouzi

Abacus trains sequential models to predict the proportions of event types in a user timeline, combining this aggregate target with tasks that preserve temporal information. On a private advertising dataset, hybrid pretraining raised GRU test AUC from 0.7094 ± 0.0149 to 0.7525 ± 0.0036, a 6.07% relative improvement over no pretraining; gains vary across datasets and encoders.

Key insights

  • Counting-aligned objectives like Abacus consistently outperform traditional reconstruction (MSM) and contrastive (BT) pretraining methods in enhancing downstream purchase prediction.
  • Hybrid multi-task learning, which blends diverse pretext tasks, provides the most robust and stable performance gains, demonstrating effective regularization through complementary inductive biases.
  • Distributional pretraining (Abacus), targeting global event statistics, is more reliable and less prone to negative transfer compared to tasks requiring exact sequence fidelity in sparse, noisy advertising data.
  • GRUs can perform comparably or better than Transformers (BERT) in this domain due to relatively short sequence lengths and small vocabularies, suggesting that simpler models with stronger inductive biases are sometimes more suitable.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.