Feature Selection based on manifold-learning with dynamic constraint handling differential evolution

Feature Selection in high dimensional feature space is the main challenge in statistic learning field. In this paper, a novel feature selection method based on manifold learning is proposed. The distance metric weight vector are optimized to maximize the multi-class margin in the manifold embedded in low dimension space, as well as minimize its L1-norm. This multi objectives optimization problem is solved by a Differential Evolution (DE) with dynamic constraint-handling mechanism. And a criterion to determine the best feature subset based on the optimal weight vector is given. The test result for selecting the optimal feature subset of UCI breast tissue dataset indicates that this real coded feature selection method could find some feature subset which has good classification robustness.