Locally Linear Discriminant Embedding for Feature Gene Extraction Based on Dynamical Neighborhood

Extracting feature genes is important for disease identification and therapy. Today the state-of-the-art methods are manifold learning methods. In this paper, we use a dynamical neighborhood parameter, instead of a stable neighborhood parameter, to construct the tangent subspace of each point. Dynamical neighborhood can select the neighbors based on the sampling density and the manifold curvation, which overcomes the drawbacks in modified locally linear discriminant embedding. To obtain reliable experimental results, we have not only used the original division of the data set for training and testing, but also reshuffled the data set randomly in the experiments. Then we select the average accuracy of KNN as final prediction classification. The result of our experiments shows that our method is more accurate and stable than other methods. It further explores the feasibility of manifold learning methods in the bioinformatics.