In this research, we address the problem of learning a single causal network structure from multiple dataset generated from different experiments. The experiments can be observational or interventional. We assume that each dataset is generated by an unknown causal network altered under different experimental conditions (interventions, manipulation or perturbation). As a result, we get a collection of heterogeneous datasets having different distributions. Manipulated distribution implies manipulated graphs over the variables. Combining all the data to learn a network might increase statistical power but only if it assumes a single encapsulating network that is true for all the datasets, which is not always the case under uncertain interventions. Pooling under uncertainty leads to spurious changes in correlations among variables. While learning causal network by pooling data from different experiments is common, we found by experimenting that this paves the way for false causal discoveries, if the effects of interventions are uncertain.
We address these issues and present a Bayesian approach of combining data from multiple experiments with observations to learn a single and accurate causal network. Our approach, called ‘Learn and Vote’ learns causal links using data from each experiment and combines them by weighted averaging. We show through studies on synthetic and natural datasets that our method
outperforms many state of the art approaches and is more robust with respect to modelling assumptions about the nature of the interventions.