Feature Selection-Based Approach for Generalized Physical Contradiction Recognition

Abstract

The objective of this paper is to improve a machine learning based methodology for recognizing the features of a Generalized Physical Contradictions (GPC) before knowing the contradiction itself when the system to be improved can be described by a simulated model based on design parameters and performance parameters. The paper starts with the background about identifying contradictions from data. It focuses on physical contradiction parameters identification with quantitative data and machine learning techniques. Although previous approaches are promising, they still have several drawbacks that require to be fixed. For instance, they do not propose any metric to inform the user about the quality of the result, which depends, among others, on the sample size. These drawbacks mainly appear in case of imbalanced data or complex relation between variables. To address these issues, we first tested different feature importance variable provided by decision tree methods (with the XGBOOST library) and retain the total gain. Second, we compared the XGBOOST methods with the previous proposed SVM based approach to see which one better describes the feature importance of variables involved in a GPC. As result XGBOOST was more robust to the noise from non-important variables. Third, we defined a set of measures for helping the user to know which is the sample size required to get good results with the tested methods.