Icml2022 | establishing a continuous time model of counterfactual results using neural control differential equations

Thesis link ：https://arxiv.org/abs/2206.08311

as time goes on , The estimation of counterfactual results may help decision makers answer “ If ” problem , To unlock personalized healthcare . Existing causal inference methods usually consider the rules between observation and processing decisions 、 Discrete time interval , Therefore, irregular sampling data cannot be modeled naturally , This is a common setting in practice . To handle arbitrary observation patterns , We interpret the data as a sample of a continuous time process , The explicit modeling of its potential trajectory using control differential equation mathematics is proposed . This led to a new approach , That is, the treatment effect of neural control differential equation (TE-CDE), It allows you to evaluate potential outcomes at any point in time . Besides , Antagonistic training is used to adjust for time-dependent confusion , This is crucial in the vertical setup , Is an additional challenge not encountered in traditional time series . To evaluate the solution to this problem , We propose a method based on tumor Controllable simulation environment for growth model , For a series of scenarios , Irregular sampling reflects various clinical scenarios .TE-CDE It is superior to the existing methods in all simulation scenarios with irregular sampling .