Research On Weighted Multi-View Clustering Algorithm Based On Internal Evaluation Of Clustering Results

As an important data analysis approach,clustering is widely used in various fields,including data mining,pattern recognition and image analysis.In traditional single-view clustering,one data object is represented by a single set of features,and this places restrictions on the clustering accuracy to a certain extent.Compared with single-view clustering,multiview clustering integrates multiple sets of features from different views,making full use of the consistency and complementarity among different views to obtain more accurate clustering results.In multi-view clustering,weighted multi-view clustering assigns weights to individual views to obtain a combined view,which is then used to generate the final clustering result.This paper proposes a weighted multi-view clustering algorithm based on the internal evaluation of clustering results.Firstly,it is noticed that in many cases the real data obey the Gaussian distribution approximately,and the internal evaluation of clustering results is rather accurate for data of this distribution in general.Motivated by this observation,this paper proposes to estimate the clustering quality in the combined view with an internal evaluation criterion,and then determine the weights of views by maximizing the estimated clustering quality in the combined view.Secondly,this paper uses the normalized cut algorithm as the single-view clustering approach,adopts the Dunn index as the internal evaluation criterion of clustering results,introduces a heuristic algorithm to determine the scale parameter,designs an adaptive weight initialization and updating method based on internal evaluation of clustering results.As a result,a weighted multi-view clustering algorithm based on internal evaluation clustering results is presented.Finally,experiments are conducted with several publicly available image and text datasets.Each main component of the proposed algorithm is tested separately,and its effectiveness is verified by experimental results.In comparison with some other multi-view clustering algorithms,the proposed algorithm generates better or comparable results,demonstrating the effectiveness of the proposed algorithm.