Feature Selection from Incomplete Multi-Sensor Information System Based on Positive Approximation in Rough Set Theory

Information fusion is necessary to conserve energy in multi-sensor information system of wireless sensor networks. A key step in the process of information fusion is feature selection (attribute reduction), which uses to remove the redundancy of conditional attributes and retain the discriminatory power of original features. Rough set theory can select feature effectively from incomplete data. However, feature selection algorithms are often computationally time-consuming. In order to address this issue, positive approximation is proposed. Positive approximation is a kind of multi-granulation sequence which is inconformity with the principle of positive partial relation. Applying positive approximation in heuristic feature selection can diminish the computational universe of each granulation gradually, and will also improve computational efficiency of feature selection algorithm greatly. In a heuristic feature selection algorithm, one key element is core attributes set. This paper proposes the properties of the core attributes among different granulations of positive approximation, and also introduces an efficient method to search core attributes by discernibility vectors, then integrates this method into heuristic feature selection algorithm. Lastly, this paper verifies the advantages of this new algorithm.