Welcome to DKRegression

DKRegression is a Python package for distributional kernel regression that is built using PyTorch as backend. Distributional kernel regression represents an extension to the classical kernel regression and therefore belongs to the class of non-parametric regression methods. The key difference to classical kernel regression is that distributional kernel regression—as indicated by the name—estimates the parameters for a distribution rather than just a point estimate as it is the case for the classic kernel regression.

The idea of distributional kernel regression is not new and has been explored in the 1980s, but mostly has been forgotten due to the popularity of Gaussian processes. In many ways, distributional kernel regression is similar to Gaussian processes, but more flexible in terms of the observation likelihood and computationally more efficient (\(\mathcal{O}(n^2)\) compared to \(\mathcal{O}(n^3)\) for GP). You can find more information about the mathematical probem forumulation of distributional kernel regression in the section "What is DKR?".

If you use our software package for your research, please consider citing:

@article{schlichting2099dkr,
  title={DKRegression: A framework for distributional kernel regression},
  author={Schlichting, Marc R and Kochenderfer, Mykel J},
  journal={TBD},
  volume={999},
  number={999},
  pages={1--999},
  year={9999}
}