Research On Direction-based Clustering Algorithm And Its Application In Community Detection

Unsupervised learning provides a solution to the problem that it is difficult to obtain sample labels in reality,and the cost of manual labeling is expensive.By learning unlabeled samples,its intrinsic properties and regularities are revealed.As an important algorithm in unsupervised learning,clustering is widely used in pattern analysis,text mining and other fields.With the introduction of semi-supervised learning methods,semisupervised clustering can effectively use the known information of samples to improve clustering accuracy under the premise of obtaining a small number of labeled samples,which has become a research hotspot.The direction-based clustering algorithm considers the orientation relationship between samples,so as to better mine the information contained in the data.Based on the research of direction-based unsupervised and semisupervised clustering algorithms,this thesis improves and optimizes them from two branches of Euclidean space data and graph data,and applies the improved algorithm to community detection,thus improving the partitioning effect of complex community networks.The main work is summarized as follows:(1)The Direction-based Clustering Algorithm needs to set the number of neighbors and the angle,which are highly sensitive.This thesis proposes an Improved DBC Algorithm Based on Adaptive Angle and Label Redistribution(ALR-DBC).It determines the receiving range by dynamically adjusting the deviation angle through the concept of high and low density region.After the initial allocation of all samples has been completed,labels are reassigned to samples that do not meet the expectations of the main receiving direction.Experiments show that the proposed algorithm improves the clustering effect and reduces the parameter sensitivity compared with the DBC algorithm.(2)Traditional clustering algorithms can not extract complex structural information of non-Euclidean spatial data.This thesis proposes a Semi-supervised Clustering Algorithm Based on Direction and Graph Convolutional Networks(SC-DGCA).It uses the advantages of graph convolution model to process graph data,embeds the attributes and structures of samples into multidimensional vector space,and makes full use of labeled data to train the model.We add the idea of metric learning,obtain high-scoring similar nodes through directional clustering algorithm and dynamically increase the set of labeled samples to achieve joint optimization.Experiments show that the proposed algorithm can still learn excellent results when only a small number of samples are labeled.(3)Community detection can better obtain the attribute information between community members and understand the internal patterns and functions of complex networks.Based on the basic idea of SC-DGCA algorithm for processing graph data,this thesis proposes an Optimized Algorithm Based on Contrastive Learning and the DGCA Model(CL-DGCA),and applies it to community detection.We introduce a contrastive learning method to make full use of unlabeled data to mine the supervision information of the sample itself.The contrastive learning loss is constructed to control the tolerance of the model to difficult samples.Experiments show that the CL-DGCA algorithm can still maintain good division results on more complex community networks,which provides reference value for further analysis and application of community structure.