Weighted Linear Ridge Regression¶
Purpose of the Module¶
This module solves the weighted linear ridge regression problem calculating the linear parameters of a model selected by the user that minimize the deviations of the predictions from the references of the data set. Therefore, it is a supervised learning tool that optimizes the linear parameters of an analytical expression in order to fit a data set. Each element of the data set can be weighted according to the relative importance or reliability attributed by the user. The regularization provides a protection from the over-fitting, this inconvenient can occur if the flexibility of the model is too high in relation to the available data. Moreover, the module calculates the leave-one-out cross-validation error for the employed data set. The WLRR module is a component of the LRR-DE software tool, developed to parametrize force fields of metal ions. In the LRR-DE software tool, the WLRR module is combined with the metaheuristic optimization algorithm differential evolution in order to tune the hyper-parameters of the model (the regularization parameter and the non-linear parameters of the model).
The LRR-DE module has been developed to parametrize force fields of metal ions, however the method can be applied to optimize the parameters of a general functional form with respect to reference data.
Background Information¶
The theoretical background of the LRR-DE procedure is illustrated in the paper [FF2018]. The LRR-DE procedure is a supervised learning methodology that combines the weighted linear ridge regression algorithm, to obtain the linear patameters of the model, with the differential evolution optimizer, to obtain the non-linear parameters of the model, using the leave-one-out cross-validation error as objective function. This module uses the GNU Scientific Library.
Building and Testing¶
To compile the code execute the Makefile (including the demo.c file provided in the ./test directory). In ./test directory a multi-objective data set is provided. The demo.c file includes an example for the definition of a model. The example is the parametrization of a force field with three components (Coulomb, Lennard-Jones 12-6) of the zinc ion in water with respect the solvatation energy and the forces on the ion for a set of clusters. The linear parameters calculated by the module should be 2.40203305, 0.00001364, and -0.10986800. They appear in the third column of the output. The values of the first and second columns are the scaled parameters and the scaling factors respectively.
Source Code¶
The source code of the algorithm is available from the Weighted Linear Ridge Regression repository. The ./source directory includes two files: i) wlrr.c contains the functions that perform the scaling of the data, the operation of fitting and the calculation of the leave-one-out cross-validation error; ii) wlrr.h define the data types employed by wlrr.c.
| [FF2018] | Fracchia F., Del Frate G., Mancini G., Rocchia W., Barone V., Force Field Parametrization of Metal Ions from Statistical Learning Techniques. J. Chem. Theory Comput., 2018, 14(1), pp 255-273 |