|
主讲人 |
朱雪宁 |
简介 |
<p>We study offline policy evaluation for assessing promotion effects in online livestreaming sessions with competing network interactions. To model such dependence, we adopt a partial linear spatial autoregression (PL-SAR) model, where promotional actions follow a behavior policy, network interference propagates through a spatial autoregressive term, and covariate effects are captured nonparametrically using machine learning methods. To estimate the average value of a target policy, we propose a Network-Directional Targeted Debiasing Estimation (NTDE) procedure that corrects the plug-in bias of an initial estimator by accounting for the propagation of perturbations through the network structure. We establish valid statistical inference under suitable initial estimation rates. For random covariates, we address a key theoretical challenge involving a network-dependent V-statistic of diverging order and reduce the asymptotic variance through exploiting the empirical distribution of the covariates. We apply the proposed method to a multimodal live-streaming dataset, estimate competing effects among sessions featuring automobiles from the same brand, and evaluate a holiday promotion policy. The results show stronger promotion effects near the holiday, followed by a cost-benefit analysis based on the estimated policy value.</p> |
