Evolving regression trees robust to missing data

Data quality is a major concern in several fields of knowledge that rely on data analysis. Missing data, in particular, have a strong negative impact in machine learning, potentially harming the knowledge extraction process by skewing results and affecting the predictive performance of the induced models. For dealing with the problem of missing data, the literature in machine learning offers a variety of strategies which can be either in the form of a preprocessing step or of an embedded solution within a predictive method. In this paper, we propose a novel evolutionary algorithm for regression tree induction, which has embedded in its evolutionary cycle a robust framework for dealing with missing data. For comparison purposes, we evaluate six traditional regression algorithms over 10 public regression datasets that were artificially modified to present different levels of missing data. Results from the experimental analysis show that the proposed approach is the one that is less affected by the increasing levels of missing data, presenting an interesting trade-off between model interpretability and predictive performance especially for datasets with more than 40% of missing data.