The Good, the Bad, and the Ugly of Markov Boundary for Tabular Prediction

Researchers tested whether Markov boundaries can improve tabular prediction across thousands of SCM3K synthetic tasks and several regressors.

Using the true boundary can boost accuracy, especially in larger and sparser feature spaces.

But current causal discovery methods rarely recover the boundary well enough to beat training on all features.

The study argues that structural recovery and predictive accuracy are not the same objective, so exact boundary discovery may not be necessary for good prediction.