Comparison of Efficiency, Stability and Interpretability of Feature Selection Methods for Multiclassification Task on Medical Tabular Data

摘要

Feature selection is an important step of machine learning pipeline. Certain models may select features intrinsically without human interactions or additional algorithms applied. Such algorithms usually belong to neural networks class. Others require help of a researcher or feature selection algorithms. However, it is hard to know beforehand which variables contain the most relevant information and which may cause difficulties for a model to learn the correct relations. In that respect, researchers have been developing feature selection algorithms. To understand what methods perform better on tabular medical data, we have conducted a set of experiments to measure accuracy, stability and compare interpretation capacities of different feature selection approaches. Moreover, we propose an application of Bayesian Inference to the task of feature selection that may provide more interpretable and robust solution. We believe that high stability and interpretability are as important as classification accuracy especially in predictive tasks in medicine.