Primary Submission Category: Causal Discovery
Causal Discovery and Prediction with Interventional Data
Authors: James Long, Yumeng Yang, Kim-Anh Do,
Presenting Author: James Long*
The field of causal discovery has historically focused on parameter estimation, rather than the prediction performance of models. Recently, systems biology experiments have begun generating large-scale interventional data sets in which certain variables are manipulated (intervened on) and the resulting system state observed. These data sets offer the possibility to assess the prediction performance of causal discovery algorithms: parameters of a causal model are learned on a subset of the intervention data and then the model is used to predict the effects of test interventions. Here we derive some of the first analytic results connecting causal discovery models with standard regression approaches for intervention response prediction. Causal discovery models require estimation of many more parameters than standard regression approaches but can extrapolate to predict the effect of untested interventions. We study the performance of causal discovery models and regression approaches in simulations and an application to predicting the effect of drug interventions in a Melanoma cancer cell line. In the Melanoma data set, we obtain state of the art performance with regression modeling, outperforming a substantially more complex causal discovery model proposed in the computational biology literature.