Research On Deep Ensemble-Driven Semi-supervised Cluster Ensemble

Clustering ensemble can often integrate multiple base clusterings into higher accuracy and more robust one.At present,some work has focused on weighting the division of each base clusterings and integrating them to get better results.However,the current research still has some shortcomings.First,these methods tend to regard each base clustering or clusters within the base clustering as the basic unit of weighting,while ignoring the differences between samples.Secondly,most of the existing methods evaluate the weights solely based on the quality of a single base clustering or cluster,without considering the complementarity between the base clustering and its clusters,and the weights obtained are local.How to use the integration effect to give each base clustering a global weighting and improve consistency is still an open problem.In this article,we refer to the global weighting method as Ensemble-driven.Finally,in many cases the data has a small amount of supervised information available for use.How to use this semi-supervised information to guide the clustering to ensemble efficiently is a hot research topic.To this end,this paper proposes two algorithms,which are ensemble-driven weighted semi-supervised clustering ensemble methods.This two algorithms uses semi-supervised information to train ensemble weightings which considers the complementarity between the divisions of base clusterings fully.And take the performance of each base clustering on each sample as weighting uint,which consider the difference between different samples fully.What’s more,the graph network-based method proposed in this article can integrate the consensus graph information and the original feature space at the same time,learn a low-dimensional graph representation that is easy to cluster.The specific work is as follows:First,this article proposes an ensemble-driven semi-supervised clustering ensemble algorithm based on soft attention network.The model is composed of a shallow attention network and a multi-layer classification neural network.The model can learn an explicit weighted co-coordination matrix.During the training process,the constraint information can be expanded based on the weighted co-coordination matrix to further improve the ensemble effect.Secondly,we propose an ensemble-driven semi-supervised clustering ensemble algorithm based on graph convolutional network.The model is based on graph neural network and the differentiable graph pooling method of graph neural network.The division and original feature space were constructed into an integrated graph structure.Then,a hierarchical representation of the integrated graph structure was generated.In addition to fully considering the complementarity between the base clusterings,the model can simultaneously extract original features and base division information to generate integrated hierarchical representations to guide cluster ensemble.Finally,rich verification experiments were designed.We compare the performance of the algorithm in this paper and other 11 ensemble algorithms and 16 semi-supervised classification methods based on neural networks.We conduct a comprehensive analysis and evaluation of the algorithm in this paper.