Peter Bühlmann

Spectrally Deconfounded Gradient Boosting

Andrea Nava, Peter Bühlmann, Fabio Sigrist

Flexible machine-learning methods can be sensitive to hidden confounding: they may learn associations induced by unobserved confounders rather than stable signals. Spectral deconfounding mitigates this problem by shrinking high-variance directions of the covariate matrix that, un…Read more