Abstract
Systems and methods are described for exposure-response behavioral curve estimation in recommendation systems with detection of aggregation artifacts consistent with Simpson’s paradox. Interaction logs are used to fit a behavioral curve model at multiple granularities, including individual and aggregate levels, and to compute peak locations. A distortion factor and severity score quantify divergence between aggregate and individual peak parameters and can trigger a paradox determination. A per-user peaked-behavior classifier is null-calibrated by generating synthetic sequences from a fitted monotonic model to estimate a null false-positive rate and adjust decision thresholds. A granularity-adaptive selector chooses among individual, cohort, and aggregate models based on data sufficiency and paradox severity, and outputs an exposure control parameter, such as an exploration budget, derived from the selected peak location.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Simpson's Paradox Detection and Granularity-Adaptive Behavioral Curve Estimation for Recommendation Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10633