Correct. It’s a linear model (using ridge regression) applied to nonlinearly transformed variables. While the model itself is linear, the transformations allow for capturing nonlinear relationships and interactions.
I am not going to make a feature request of P123 to generate new variables with a neural-net to use with P123 classic and the optimizer, myself. Nor am I going to try to duplicate what the paper is doing with P123 downloads at home. I think I will stick with Extra Trees Regressor and LightGBM in order to capture nonlinear relationships and interactions.
Specific potential problems include the potential of overfitting without careful cross-validation, or other methods to prevent overfilling, and a lack of transparency with neural-net generated features. These practical considerations are summed up here I think:
Maybe everyone (or almost everyone) has some limits as to which parts of a kitchen sink they expect to be beneficial to a model. Maybe cross-validation can help provide an objective answer as to what factors to use. In addition to any domain knowledge to guide research. P123 provides some excellent cross-validation methods for those looking for practical answers about their particular models and features.
The paper did motivate me to consider investigating the use of interaction terms with linear models. I think that could be one practical application for the findings of this paper. Most members could start using interactive terms in their models today if they find this interesting and potentially useful. And they could do it base on their domain knowledge about how they think features might interact. With transparency and control of which features are allowed to interact in the model.
For those who are interested, I believe Feldy had a helpful post about interactive terms and how to implement them (including an excellent suggestion for standardizing the interactive terms within a P123 rank node using Frank()): LightGBM and interactions - #9 by feldy