The Application Of Heuristic Ideas In Semi-supervised Cluster Ensemble

The development of Internet brings various data to human society.Clustering is an effective tool for minning hirstorical data to make it valuable.Through clustering we can dig the inherent distribution pattern of the collected data and obtain the effective information and support the realistic decision.In the area of big data,the data we face are often high-dimension,multi-faceted.So far,many clustering methods have been proposed to deal with high-dimensional clustering problems,but there are obvious limitations in the existing methods:(1)Most clustering ensemble are unsupervised and do not fully exploit the domain knowledge that is used by the dataset.(2)Some methods are semi-supervised,but only the similarity or the feature contrast method is used to remove the redundant clustering member,which is lack of system optimization.(3)the current semi-supervised clustering ensemble field lacks the selection of priori knowledge.In this paper,two semi-supervised clustering integration frameworks are proposed:First,the semi-supervised clustering ensemble framework based on ensemble member selection(ISSCE),ISSCE's innovation points are:1)The use of random subspace technology to generate ensemble members and the domain knowledge-assisted clustering process 2)The use of semi-supervised clustering algorithm E2CP to maximize the use of the data set of a priori information.3)Using heuristic ensemble member selection algorithm to find the optimal ensemble member set.Second,the double weighting semi-supervised clustering ensemble framework based on constraints projection(DCECP),DCECP's innovation points are:1)weight the pairwised constraints to achieve differential treatment based on Boosting method 2)make constraints projection with weighted constraints to transform the original space 3)proposed the adaptive ensemble member weighting process based on heuristic evolution.Experiments on 18 real datasets show that the frameworks presented are more accurate and robust than those of existing clustering algorithms in high-dimensional clustering scenarios.