Reproduce “System-2 Recommenders: Disentangling Utility and Engagement in Recommendation Systems via Temporal Point-Processes” 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

Arpit Agarwal, Nicolas Usunier, Alessandro Lazaric, Maximilian Nickel

Researchers from FAIR at Meta developed a generative model leveraging a Hawkes process to disentangle user utility from engagement in recommender systems. By modeling user return probability with distinct System-1 (impulsive) and System-2 (utility-driven) components, the framework enables recommendations that maximize genuine user utility, achieving superior performance over engagement-focused strategies, particularly when preferences are misaligned.

Key insights

  • User return probability to a platform serves as a more reliable indicator of received long-term utility than instantaneous, within-session engagement metrics.
  • Explicitly distinguishing between fast-decaying impulsive effects and slow-decaying utility-driven effects within a temporal point process framework allows for disentangling these influences.
  • A sufficient temporal difference in the decay rates of System-1 and System-2 effects is crucial for the identifiability and accurate disentanglement of user preferences from observational data.

Summary by alphaXiv.

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