| With the development of emerging technologies,artificial intelligence has gradually stepped into people’s daily lives,including face recognition,intelligent robots,and intelligent recommendations.As one of important methods of machine learning and image recognition,clustering analysis has been widely studied and applied.However,traditional clustering methods such as k-means and spectral clustering still have some challenges when facing high-dimensional data: the performance of clustering is very sensitive to outliers data and highly dependent on learned similarity graphs.In recent years,subspace clustering has become a research hotspot for its good ability to process high-dimensional data.Sparse Subspace Clustering(SSC)integrates sparse representation and spectral clustering algorithms,it has been used in issues for its efficient calculation and robust.Clustering with Adaptive Neighbors(CAN)obtains a wide range of applications because its unique similarity graph learning style and stable clustering performance.However,unlabeled data in the SSC are high dimensionality and cluster distribution are scattered,the clustering accuracy depends on the similarity graph.CAN cannot capture the local structure of the data well,and there is room for improvement in the clustering performance.Therefore,this article studies and discusses some problems in the existing subspace clustering,aiming at the maximum of data separability and the processing ability of nonlinear data,and the improvement of clustering performance.The main work of this paper is as follows:1.Through the study on the data preprocessing process of SSC and model,this paper proposed a novel Sparse Subspace Clustering model integrated with Unsupervised Metric Learning(UML-SSC),which combines distance metric and subspace clustering to a joint framework,design a two-step subspace clustering process.This method reconstructs the sparse subspace clustering model,which data preprocessing is no longer a separate step,and maximizes the similarity between the unlabeled data,improves the subspace clustering performance and strengthens the generalization ability of model.2.Through the study on CAN and its expansion on autoencoder,this paper proposed Deep Subspace Clustering fused with Auto-Weight Learning and Structure Information(DSC-AWSI),This method use autoencoder as a feature extraction method to iteratively enhances the correlation between features in latent space,and uses autoweight learning to assign smaller weights for noise and unimportant features,effectively distinguish important features and noise features.This model taking into account the global and local structure information of the data,has better stability and clustering performance than original method. |
