Synthetic Nonlinear Regression for Domain Generalization

A synthetic regression dataset simulating shortcut learning in nonlinear settings. Inputs are composed of invariant features and spurious features, with domain shifts created by varying the correlation between spurious features and the output. The data is generated from passing random inputs through a random 4-layer MLP and includes two benchmark variants (regular and harder) with different training environment correlations. Used to evaluate the ability of models to ignore spurious features during domain generalization.