Comparative study on feature extraction of mass traffic data using multiple methods

This paper aims at extracting the typical and significant features of the traffic network by using variant feature extraction methods. Combined with the intrinsic tempo-spatial characteristics of traffic flow data, data mining technique is introduced to extract the main features of the temporal and spatial relationship and the typical patterns of the traffic network. We introduce three methods in feature extraction: principal component analysis (PCA), robust PCA and kernel PCA. By selecting the eigenvalues according to decreasing magnitude of eigenvalues, we design a transform matrix to reduce the dimensionality of the original matrix, as well as obtain the features of the traffic network. By comparing the results of feature extraction of different methods, we find a better way to extract the typical features in urban traffic data and attempt to explain some the features.