Towards mining the density and discriminative information of data in dimensionality reduction
Deqin Yan, Hongzhe Jia, Shenglan Liu · 2014
Recently these has been a lot of interest in geometrically motivated approaches to data analysis in high dimensional space. Locally linear embedding (LLE) is one of the nonlinear dimensionality reduction algorithms. LLE is an unsupervised learn ing algorithm. It computes low-dimensional by neighborhood-preserving embeddings of high-dimensional inputs. However, different structure of neighbors will produce different reconstruction errors, which will make the results hurt. In this paper, we propose a novel algorithm called Density and Discriminate-based Weight ed locally linear Embedding (DDWLLE). Different from WLLE[1], DDWLLE aims at preserving the local neighborhood structure and mining the density and discriminative information of the neighborhoods on the data manifold. Several experiments on gene expression profile database and image retrieval database demonstrate the effectiveness of our algorithm.