KernelPopper: A Scikit-Learn Compatible Package That Calculates Feature Weights From Trained Kernel Ridge Regression Models

Kernel Ridge Regression (KRR) models extend linear least squares regression using the kernel trick to learn non-linear relationships between input and target variables. During ordinary least squares regression training, the model learns by determining a weight for each feature that captures its contribution to predicting the target value. KRR models instead focus on dual weights, which evaluate the importance of each training sample rather than the features. To calculate feature weights from the learned dual weights, we have developed a package for the KernelRidge model class from the widely used scikit-learn Python package. This enables interpretation of the weights learned by trained KRR models. The package is currently restricted to polynomial kernels of degrees two and three.