Overview / Usage

We consider the problem of full model learning of causal models from data specified in the context of predicting post-partum depression (PPD) from data. A common argument is that when learning only from data, learning causal models is only as informative as learning a correlation model (an opaque one such as deep belief network). In this work-in-progress, we put this hypothesis to test. We aim to learn a causal model using an ensemble of models and methods. This is particularly important as scaling causal learning to large problems without interventions or bias is a significantly challenging task.

Methodology / Approach

To scale causal model learning, we first learn a DN. Then we identify and remove cycles from this DN. We consider
several different metrics employed in causal models to score and remove the edges. Finally, we construct a model based
on the edges that are commonly present across all the metrics (i.e., the intersection of the edges from the different methods). Contrary to popular intuition, we employ two levels of ensemble learning to uncover a causal model - first is on learning a DN using boosting and the second is on learning a causal model from several different metrics. Our evaluations on the real survey data for predicting PPD demonstrates the utility of such an approach. While we present quantitative metrics, qualitatively, the edges that are learned in this model uncover interesting findings.

Technologies Used

Python, bnlearn, R