Abstract:

A semi-supervised nearest neighbour classification algorithm was proposed, in which both labeled points and pair-wise constraints were employed to determinate the label of data points. To solve the problem that some data points may not be assigned to any class label, the ratio sorting was designed to reduce the number of conflict points. An active learning strategy based on CitationkNN score was developed to search valuable supervision information and improve the quality of clustering by querying the label of a point incompatible with its neighbours. Experiments show that the learning strategy can improve the clustering performance, and the comparison with COP-kmeans and CCL illustrates the efficiency of the active SNN from the view of CRI.

Key words: semi-supervised clustering; active learning; supervision information; nearest neighbour