Conflict-based Feature Selection for Information Fusion Systems

Applying information fusion systems aims at gaining information of higher quality and simultaneously decreasing computational and communicational efforts. An increased availability of sensors in industrial machines, but also in everyday life, results in large amounts of potential features. Each feature entails computational and communicational costs. An information fusion system may not require all features, supported by the available sensors, to fulfil its purpose. Feature selection methods reduce the amount of features with the aim to maintain or even increase performance. This contribution proposes a feature selection approach exploiting the inherent conflict between features and utilising a state-ofthe-art information fusion operator. The performance of the proposed method is evaluated in the scope of a publicly available data set and benchmarked against an established feature selection method. It is shown that the proposed approach is faster and produces more accurate feature subsets containing very few features, although the established method produces slightly better performing subsets for large feature subsets.