Feature Selection and Ranking for Pattern Classification in Wireless Sensor Networks

Feature selection is a classical problem in the discipline of pattern recognition, for which many solutions have been proposed in the literature. In the current paper we consider feature selection in the context of pattern classification in wireless sensor networks. One of the main objectives in the design of wireless sensor networks is to keep the energy consumption of sensors low. This is due to the restricted battery capacity of today’s sensors. Assuming that the features of a pattern recognition systems are acquired by the network’s sensors, the objective of keeping the energy consumption of the sensors low becomes equivalent to minimizing the number of features employed in object classification. In fact, this objective is related with, but not identical to, classical feature selection, where one wants to optimize the recognition performance of a system by detecting and eliminating noisy, redundant, and irrelevant features. This paper introduces a general framework for pattern classification in wireless sensor networks that aims at increasing the lifetime of the underlying system by using a number of features as small as possible in order to reach a certain recognition performance. In experiments with data from the UCI repository, we demonstrate the feasibility of this approach. We also compare a number of classical procedures for feature subset selection in the context of pattern classification in wireless sensor networks.