Feature selection is the task of finding a subset of original features which is as small as possible yet still sufficiently describes the target concepts. Feature selection has been approached through both heuristic and meta-heuristic approaches. Hyper-heuristics are search methods for choosing or generating heuristics or components of heuristics, to solve a range of optimisation problems. This paper proposes a genetic-programming-based hyper-heuristic approach to feature selection. The proposed method evolves new heuristics using some basic components (building blocks). The evolved heuristics act as new search algorithms that can search the space of subsets of features. The classification performance (accuracy) of classifiers are improved by using small subsets of features found by evolved heuristics.
展开▼
