Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Estimation and Inference under Recommender Interference
Authors: Ruohan Zhan, Shichao Han, Yuchen Hu, Zhenling Jiang,
Presenting Author: Ruohan Zhan*
Recommendation algorithms are crucial in digital platforms for tailoring content to viewer preferences. These platforms rely on content creators to maintain a dynamic viewer community. This study evaluates interventions targeting creators. Using canonical creator-side randomization A/B experiments, we find that the standard difference-in-mean estimator is biased in estimating treatment effects. This bias stems from the interference among creators competing for visibility through recommendation algorithms. To address this, we propose an innovative method to eliminate interference bias in measuring treatment effects. We introduce a semi-parametric model for recommender choice and develop influence functions for treatment effects that satisfy Neyman orthogonality. This approach enables us to create a consistent and asymptotically normal estimator for treatment effects, supporting inference and hypothesis testing. We demonstrate the efficacy of our method through simulations and practical applications on a leading short video platform.