Implementation of locally constant (Nadaraya-Watson) and locally linear kernel regression with automatic bandwidth selection and adaptive kernel, compatible with sklearn.
Expanded from Jan Hendrik Metzen's kernel_regression package. Major code redundancies between routines that I've yet to clean up.
- Improved numerical convergence of fit values far from training points
- Allows validation on "parallel sample" rather than cross-validation, which is useful when some points would be dominated by own kernel
Note: the code is fairly rough, untested, and poorly documented, but is made public since I couldn't find any equivalent functionality elsewhere. Feel free to contact me with improvements.
Example of locally linear regression advantages in 1 and 2-d:
