Key Feature Mining Algorithm Based on Self-organizing Fuzzy Rough Analysis for Target Uncertainty Sensing Information

Feature selection is a complex problem of knowledge discovery, especially when both perceptual features and target labels are ambiguous. In this paper, we propose an algorithm based on the theory of “self-organization analysis” and “fuzzy rough approximation” for mining key features from raw low-density features captured by various sensors such as radar, sonar, optical devices, etc. First, the spatial distributed clusters (SDCs) of the label-missing samples are obtained by iterative fuzzy self-organizing analysis. Second, a subset of candidate features is selected, which is then used to create a fuzzy correlation function and a fuzzy equivalence partition for the target sample. This fuzzy equivalence partition helps us to approximate the original SDCs and obtain an approximate degree value. Finally, all possible combinations of feature elements are traversed by the above method, and the feature subset with the largest approximation degree is selected as the key feature subset. Experiments based on target recording data show that the proposed algorithm can mine the key feature set efficiently without sacrificing the classification rate for target uncertainty sensing information.