Generalized Locally Linear Embedding Based on Local Reconstruction Similarity

Manifold learning has currently become a hot issue in the field of machine learning, pattern recognition and data mining. Locally linear embedding (LLE) is one of several promising manifold learning methods. But ordinary LLE can not distinguish effectively the low-dimensional embeddings of noise data. By introducing the reconstruction similarity into LLE, this paper proposes a generalized locally linear embedding algorithm based on local reconstruction similarity. Experimental results show on Columbia object image data that the new generalized version is superior to LLE in revealing the visualization of high-dimensional image dataset containing noise images.