Using Markov Boundary Approach for Interpretable and Generalizable Feature Selection
share
Predictive power and generalizability of models depend on the quality of features selected in the model. Machine learning (ML) models in banks consider a large number of features which are often correlated or dependent. Incorporation of these features may hinder model stability and prior feature screening can improve long term performance of the models. A Markov boundary (MB) of features is the minimum set of features that guarantee that other potential predictors do not affect the target given the boundary while ensuring maximal predictive accuracy. Identifying the Markov boundary is straightforward under assumptions of Gaussianity on the features and linear relationships between them. This paper outlines common problems associated with identifying the Markov boundary in structured data when relationships are non-linear, and predictors are of mixed data type. We have proposed a multi-group forward-backward selection strategy that not only handles the continuous features but addresses some of the issues with MB identification in a mixed data setup and demonstrated its capabilities on simulated and real datasets.
READ FULL TEXT