Zhiwei Qin

Off-policy evaluation (OPE) is widely applied in various sectors such as pharmaceuticalsand e-commerce to evaluate the efficacy of new products or policiesfrom offline datasets. This paper introduces a causal framework thatrelaxes the mean-field structural assumption commonly imposed in existingOPE approaches for handling spatio-temporal interferences. The mean-fieldassumption often falls short in real-world scenarios, limiting the capabilityof existing OPE methods to accurately capture complex interference effects.In response, we advocate for a more generic permutation invariance (PI) assumptionthat enables the data-driven, adaptive learning of the mean-fieldfunction, offering a more flexible estimation of the interference effect. Furthermore,we develop novel permutation-invariant estimators (PIEs) basedon neural networks that incorporate the PI assumption and thoroughly examinetheir theoretical properties. Our numerical analyses demonstrate that ourapproach yields significantly more accurate estimations than existing baselinemethods, substantially improving the practical applicability of OPE in thepresence of spatio-temporal interference. A Python implementation of our proposed method is available at https://github.com/BIG-S2/Causal-PIE.