I am currently investigating the M5 tree algorithm by Quinlan(1992) link here: https://sci2s.ugr.es/keel/pdf/algorithm/congreso/1992-Quinlan-AI.pdf
An example of a linear regression model of the algorithm can be seen below:
An implementation of the model similar to Scikitlearn can be found here: https://github.com/ankonzoid/LearningX/tree/master/advanced_ML/model_tree
The M5 model is a more advanced implementation of the standard decision trees such as the IDE3 or C4.5. Instead of simple binary splits of the training features the data is split to the Standard Deviation Reduction calculated as follows against linear functions at leaf nodes:
At the end the model produces composite linear regression sections for the different split portions of the model at the leaf nodes.
Would the same robustness principles hold for outliers, missing values and noise as for regression trees?
