Dissimilarity-based semi-supervised subset selection

Abstract

Extracting useful information from large-scale data is a major challenge in the era of big data. As an effective means of information filtering and data summarization, the subset selection method selects the most informative subset from large-scale data to represent the entire data set to reduce the size of the data that needs to be processed. In this thesis, a kind of dissimilarity-based semi-supervised subset selection method is proposed. To begin with, the subset selection problem is treated as an convex optimization process with regularization. Thus the wanted subset is modeled as an unknown sparse matrix, which non-zero rows represent the target set by the source set. Then alternating optimization method is used to solve the Lagrangian form of the objective function. To utilize the information implicated in the labels of samples, semi-supervised algorithm is proposed to do unsupervised clustering and supervised representatives judgement. Afterwards, the iterative process will update the distribution of representatives based on the overall correlation coefficients of each category of target set. In the end, the optimal matrix and representatives will be output.