Research On Clustering Analysis Based On Graph Representation

Clustering aims to analyze the potential community structure of unknown labeled data according to their inherent similarity,and is one of the important methods of data mining.Traditional clustering algorithms mainly rely on the attribute characteristics of samples to calculate the similarity between samples,and implicitly assume that samples are independent of each other,thus ignoring the potential correlation between samples.Graph neural network can effectively extract and mine topological structure features and connection patterns between samples,and has been widely used in real complex relational networks,such as social relationship networks,chemical molecular networks,traffic prediction networks,etc.In this paper,on the basis of using the graph neural network to learn the characteristics of the relationship between data samples,the spatial location distribution of data samples is used,and the relationship between samples in different views is integrated,which effectively promotes the clustering of data clusters.Structural predictions.The main research results and innovations of this paper are as follows:(1)Adaptive edge samples recognition for depth clustering algorithm.The existing deep clustering algorithm has insufficient utilization of topological information between samples,does not consider the spatial distribution of samples and the relationship between samples,poor separability of edge samples,and high category uncertainty.Adaptive Deep Clustering Algorithm(Auto-CB)for Edge Sample Identification.While learning sample feature representation and mining structural information between samples,data samples are adaptively divided into cluster center samples and edge samples according to the spatial distribution of samples,and the correlation between samples is used to reduce the sensitivity of the algorithm to edge samples.It makes the model focus on improving the core characteristics of the category,reduces the ambiguity of the edge sample category,and improves the ability of the clustering algorithm to describe the data.This paper compares the performance of 8 deep clustering and graph neural network-based algorithms on 4 public data sets.After comparative analysis of experiments,the Auto-CB algorithm has improved the accuracy by more than 2% compared with most algorithms.It shows that the fusion of feature information and spatial location information of samples can effectively use the correlation relationship between samples,and then improve the final clustering effect.(2)Multi-view clustering algorithm based on graph enhancement.Because the existing clustering algorithm contains too much noise information in the topological structure of a single sample association relationship,and cannot deeply utilize its topological structure characteristics,the clustering accuracy of the algorithm is not high.A multi-view clustering algorithm based on graph enhancement is proposed(Multi-CB).In order to improve the data quality,improve the robustness and generalization ability of the model,and reduce the uncertainty of the sample,this paper enhances the characteristics of the data sample;and utilizes the complementarity and consistency between the multi-views of the sample correlation structure The property promotes the spatial topology information extraction of the potential relationship between the data samples by the graph neural network,and extracts the topological structure features of different information angles.This paper verifies the effectiveness of this algorithm on 4 datasets.Experimental analysis shows that Multi-CB is superior to Auto-CB algorithm on most datasets,indicating that constructing multiple views can effectively compensate for the existence of single-view clustering samples.Insufficient topological information to avoid high sample misclassification rate.