Distance-based feature selection for low-level data fusion of sensor data / Maz Jamilah Masnan ... [et al.]

Low level data fusion offers a mechanism for raw data from different sensor devices to be fused in an attempt to improve classification performance. This scenario creates a challenge for engineers to deal with large number of features over the smaller number of observations, or also known as high dimensional problem. Traditionally, engineers prefer to apply feature extraction in choosing important features for classification task. Unfortunately, the traditional method is bound to certain limitations. Thus, the objective of this study is to propose a feature selection based on unbounded Mahalanobis distance [0, ∞) to replace the feature extraction phase in the low-level data fusion. The average pair-wise distances for the fused features were calculated and filtered from largest to smallest values, and features with the larger distance value are considered important. Classification results using the ranked features in a feature subset selection manner have shown that the proposed distance provides an effective and easy approach in choosing the important features that lead to good classification performance.