Compressive neighborhood embedding for classification
Yuan Chen, Zheng Zheng · 2014
Recently, spectral manifold learning algorithms on pattern recognition and machine learning orientation have found wide applications. The common strategy for these algorithms, e.g., Locally Linear Embedding (LLE), facilitates neighborhood relationships which can be constructed by knn or ϵ criterion. This paper presents a simple technique for constructing the nearest neighborhood by combining ℓ2and ℓ1norm. The proposed criterion, called Compressive Neighborhood Embedding (CNE), gives rise to a modified spectral manifold learning technique. The validated discriminating power of sparse representation has illuminated in [1], we additionally formulate the semi-supervised learning variation of CNE, SCNE for short, based on the proposed criterion to utilize both labeled and unlabeled data for inference on a graph. Extensive experiments on semi-supervised classification demonstrate the superiority of the proposed algorithm.