Geodesic distance based semi-supervised locality dimensinality reduction

Abstract

Semi-supervised dimensionality reduction is becoming one of the most popular fields nowadays. But the existing algorithms can not fully utilize the information in dimensionality reduction as the side information is treated equally. A new semi-supervised dimensionality reduction algorithm called Geodesic distance based semi-supervised locality dimensionality reduction (GSLDR) is proposed for the handwriting data to overcome the shortcomings. Since Euclidean distance cannot really reflect the structure of data, we adopt geodesic distance as the measurement. Then the algorithm expands the pairwise constraints to strengthen the guiding role of the constraints in dimensionality reduction, and add the constraints to the nearest neighbor graph to make the graph reflect realistic manifold structure of the data. At last, the proposed method is applied to the writer identification. The experimental results on datasets show the effectiveness of the algorithm.