Ensure that all linear regression weights are non-negative when specifying monotone_constraints = 1 and linear_tree= True

[hyang57]

Summary

I hope (to confirm) that when specifying monotone_constraints = 1 and linear_tree= True, all regression parameters in the Linear tree are non-negative.

Motivation

I hope to use a linear regression model on the leaf nodes of the LightGBM model while maintaining the strong correlation between the factors and the predictions.

Description

I am grateful to Microsoft for providing such an excellent model, which has deepened my research. However, I encountered some issues while using the model and hope you can assist me.

During the model training process, I used monotone_constraints = 1 for all my 'x' values. I believe there is a strong positive correlation between my factors and my predicted values. Additionally, I am not keen on handling multicollinearity because the factors lose their interpretability after addressing it. Therefore, having a positive correlation can help resolve many of my concerns.

Therefore,I hope (to confirm) that when specifying monotone_constraints = 1 and linear_tree= True, all regression parameters in the Linear tree are non-negative.

References

Maybe apply a bounded regression method or use np.exp() on weights?