Information-Theoretic Ensemble Feature Selection With Multi-Stage Aggregation for Sensor Array Optimization

In an electronic nose (e-nose) system, the gas sensor array is the key component for detecting the volatile profile of a sample. An inappropriate sensor combination can lead to several issues, including overlapping selectivities, high computational overhead, performance degradation, etc. To deal with this problem, typically, a feature selection algorithm (FSA) is utilized to optimize the gas sensor combination in the sensor array. However, the instability of the FSA output is a serious problem in sensor array optimization. An unstable FSA output makes it difficult to conclude the general sensor combination. Hence, in this study, the Information-Theoretic Ensemble Feature Selection (ITEFS) method is proposed to deal with FSA instability. In the experiment, twelve homogeneous e-nose data sets were used that corresponded to twelve types of beef samples. Moreover, the outputs from eleven information-theoretic FSAs were aggregated to determine the general sensor combination. The performance of the selected sensors was validated based on the number of overlapping selectivities, F-measure, and execution time. The results indicated that ITEFS can more effectively reduce overlapping selectivities than other FSAs. The selected sensors also displayed comparable performance in classification tasks to other FSAs even when using fewer sensors, with an average F-measure of more than 99% using k-NN and SVM on all data sets. This indicates that the combination of selected sensors had sufficiently good generalization to detect the various types of beef samples. In addition, the utilization of a lower number of sensors also reduced the execution time in the training and testing processes.