Graph Fusion And Embedding For Clustering

The lack of most data class labels in machine learning and data mining makes clustering tasks a key task in the learning.The ultimate purpose of clustering is to reveal the hidden characteristics of the data,so it is very important for the clustering task to get a reasonable representation of the relationship among samples.For clustering of higher dimensional data,it is difficult to obtain useful information due to its low data density.At present,the commonly used solutions are unsupervised feature selection and subspace clustering.In unsupervised learning tasks,because the data has no class labels,the relationship among samples,which can be obtained through graph model learning,that is,mining data information to build a learning model,cannot be obtained intuitively.Unsupervised feature selection requires similarity or discriminative information between samples to replace class labels,and spectral clustering method is a commonly used method to generate sample pseudo-label matrices and itself is a graphbased clustering method.At present,there are many methods to measure the similarity between samples,but most unsupervised feature selection algorithms choose a method to build a graph model based on experience.There is no clear reason to believe that this method can truly reflect the sample distribution.How to fuse graphs constructed by different methods for unsupervised feature selection to improve clustering performance is a problem to be solved in this paper.In addition,most subspace clustering methods first generate an affinity matrix and apply it to spectral clustering to complete the clustering process.There is a kind of symbolic social network data,which is characterized by user positive-negative link matrix data in addition to the sample feature matrix.In this paper,whether graph embedding can be used to obtain an affinity matrix which is closer to the true relationship of the samples,and then improve the clustering performance is another problem to be solved.This paper discusses and analyzes the problem of adaptive graph fusion unsupervised feature selection and graph embedded clustering for clustering tasks to improve clustering performance.The main works include:(1)An unsupervised feature selection method for adaptive graph fusion for clustering tasks is proposed.Using parameter-free auto-weighted multiple graph learning method,graph models constructed by different methods are adaptively fused for the spectral clustering based methods aim to calculate the pseudo-label matrix for unsupervised feature selection and improve the clustering performance.(2)A clustering method based on graph embedding is proposed.The positive and negative link graphs in the symbol network data are embedded into the clustering process to help learning the similarity between samples,and then improve the clustering performance.