Adaptive Edge Weights for Supervised Graph Embedding

Abstract

Subspace learning is crucial for feature extraction and dimensionality reduction which play important role for pattern recognition and machine learning. It is generally believed that many subspace learning algorithms can be considered as linear cases of graph-based manifold learning with special edge weights. We develop a robust subspace learning method by designing reasonable edge weights which give rise to good generalization. The value of the edge weights can reflect the distribution of the data of each class and thus the consequent subspace may have good generalization property. Experiments results on face recognition show the effectiveness of the proposed method.