Interpretable modeling by linearly independent descriptor generation method

For interpretable modeling, linear regression combining descriptor generation and selection operator, such as approach of Ghiringhelli[1] and Ouyan[2], is very effective. However, multicollinearity (MC) (and near multicollinearity) between descriptors that are problematic in linear regression analysis can reduce the quality of the descriptor selection. Therefore, we proposed the linearly independent descriptor generation (LIDG) method[3] that generates descriptors while removing MC. This approach can improve their approach.
In this presentation we will present some application examples.


References:
[1] L. M. Ghiringhelli, J. Vybiral, S. V. Levchenko, C. Draxl, and M. Scheffler, Phys. Rev. Letter {\bf 114}, 105503 (2015).
[2] R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, arXiv:1710.03319v2 [cond-mat.mtrl-sci] (2017).
[3] https://github.com/Hitoshi-FUJII/LIDG.