Improving Sensor Subset Selection of Machine Olfaction using Multi-class SVM

An approach of sensor subset selection is considered one of significant issues in machine olfaction. Basically, each sensor should provide different selectivity profiles over the range of target odor application so that a unique odor pattern is produced from each sensor in the array. However, some or most of the features obtained from an array of sensors in practice are redundant and irrelevant due to cross-sensitivity and odor characteristics. The goal in this study is to optimize the number of sensors and also propose a fast searching strategy to the optimal solution. In this study, a state-of-the-art classification algorithm, Support Vector Machine (SVM), is employed by selecting the first few seed sensors based on maximum margin criterion among different odor classes. These identified sensors are subsequently used as an initial candidate in the search algorithm. From the experimental results on the soda data set, the number of selected sensors is not only significantly reduced but the classification performance is also increased.