Abstrato

Large-scale data classification based on clustering feature tree decomposition

Yanfeng Li

When the scale of training dataset is large, the demand for computing resource of traditional classifiers will increase fast. So we need to expand SVM algorithms to largescale dataset. With the analysis on the development and direction of semi-supervised algorithms at home and abroad, this paper introduces clustering feature tree to organize large-scale data using local learning strategy. First, based on the idea of local learning, we use CF tree to organize and separate the samples into a series of local sub-set, to divide original problem into limited small-scale sub-problems; Next, we propose the computing method to improve the Euclidean distance of CF tree, to measure the distance between test samples and multiple local classifiers, and to select the closest classifier for testing; Finally, SVM is used to construct multiple local classifiers for the local labeled clusters. Then these local classifiers are combined to a global classifier to acquire an integrated classification model. Several groups of large-scale data experiments show that the improved algorithm increases the training speed and test speed, with higher test accuracy.