Evolutionary Approach for Interpretable Feature Selection Algorithm in Manufacturing Industry

Feature selection techniques in prediction play a role in manufacturing industries of late. However, it is very challenging to achieve an optimal subset of features as well as interpretable relationship among features due to computation complexity and variable diversity. In order to address those difficulties, this paper presents a novel evolutionary approach for feature selection algorithm to improve the effectiveness of existing meta-heuristic approaches. In other words, their optimal combinations with minimal difference between prediction and actual values can be achieved by applying an estimation of distribution algorithms (i.e., extended compact genetic algorithm) on the collected candidate feature sets. The approach discovers a less complicated and more closely related probabilistic-model structure on population space in each generation, thereby encouraging the comprehension power of feature selection results. We tested our method on six real-world data sets from manufacturing industries (open to the public). It demonstrated that higher interpretability on features selection results is achieved in comparison with well-known methods.