New interpretable algorithm

[mathias-von-ottenbreit]

Hi.

I would like to contribute with a section in Chapter 5 about Automatic Piecewise Linear Regression (APLR). It can be used for regression or classification. APLR is often able to compete with tree-based methods on predictiveness and it is inherently interpretable. The algorithm automatically handles interactions and performs variable selection.

Below are some of the possibilities in APLR regarding interpretability:

• Get the shapes of main effects and interactions (for each effect, relevant predictor values and contributions to the linear predictor). For main effects and two-way interactions this is easy to plot in charts.
• See the regression coefficients and estimated importance of each term in the model.
• Compute the local feature contribution to the linear predictor on arbitrary data (training data, test data, new data etc.).

Is that ok?

Best regards,
Mathias