Is RuleFit monotonic?

[bdatko]

Section 5.6.4 Disadvantages describes the benefit of using a monotonicity constraint for RuleFit, implying that RuleFit isn't a monotonic function. Below is the excerpt:

Sometimes RuleFit creates many rules that get a non-zero weight in the Lasso model. The interpretability degrades with increasing number of features in the model. A promising solution is to force feature effects to be monotonic, meaning that an increase of a feature has to lead to an increase of the prediction.

I apologize if this is obvious, but can you provide your reasoning/source for the emphasized text in the passage above?

In the passage, are you implying:

• A general use of monotonicity constraint to improve the fit
  AND/OR
• Because RuleFit doesn't enforce any monotonicity constraint, you can improve the fit by imposing the constraint yourself

I have seen two conflicting references citing your book where the authors say RuleFit is and isn't monotonic. [1,2]

Reference 1

The author from reference 1 shows a table, where citation [12] in the screenshot below is

[12] C. Molnar, Interpretable Machine Learning, 2019. ISBN 978-0-244-76852-2 https://christophm.github.io/interpretable-ml-book/.

Reference 2

The authors from reference 2 shows a table, where citation [70] in the screenshot below is

[70] Molnar, C. Interpretable Machine Learning. 2019. Available online: https://christophm.github.io/interpretable-ml-book/ (accessed on 22 January 2019).

References

1. On the impact of geospatial features in real estate appraisal with interpretable algorithms
2. Machine Learning Interpretability: A Survey on Methods and Metrics