Nowadays,massive data presents the characteristics of multi-source,heterogeneity and low value density.Multi-view learning can make good use of the consistency and complementarity between data from different sources to mine the rich information hidden behind it.However,in real life,complete multi-view relationships are difficult to obtain,and considering the large labor-intensive labeling of large-scale instances,many incomplete multi-view clustering algorithms have been proposed one after another.In view of the shortcomings of existing methods in this field for capturing complex data structures,this paper proposes two incomplete multi-view clustering algorithms based on the idea of self-expressive subspace learning:To solve the interference of potentially redundant data on feature learning for missing views,an incomplete multi-view clustering algorithm IMGLWSLR based on weighted low-rank sparse representation graph learning is proposed.The algorithm utilizes low-rank and sparse constraints to capture the global and local subspace structure of multi-view data through affinity graph learning,so as to select important features for mutual self-representation between data.Meanwhile,a weighting mechanism is designed to suppress the effect of missing instances.Furthermore,a kernel alignment method is integrated,aiming to obtain common feature representations across incomplete views.To solve the influence of the disparity distribution between heterogeneous views on the effective fusion of incomplete multi-views,a dual-aligned self-supervised incomplete multi-view subspace clustering network DASIMC is proposed.The model first designs a deep autoencoder based on inter-view consistent alignment and original geometric manifold alignment,combined with specific weight layers,to achieve reliable fusion of incomplete multi-views.Afterwards,through the bidirectional learning of the self-expressive layer and the spectral clustering module,the easily segmentable subspace structure with intra-class compactness and inter-class repulsion is further obtained.Based on self-expressive technology,algorithms IMGLWSLR and DASIMC respectively carry out learning for two different problems: feature confusion within views and inconsistent distribution across views when data is missing.In this paper,clustering experiments of the two proposed algorithms under different missing rates are performed on several common multi-view datasets.The experimental results show that the IMGLWSLR and DASIMC methods successfully learn high-precision incomplete multi-view clustering features.Its clustering performance significantly outperforms all contrasting state-of-the-art methods,and they are applicable to arbitrary multi-view missing situations.In particular,the IMGLWSLR algorithm performs well in traditional models,while the DASIMC model breaks through the bottleneck of previous shallow models.It can capture the deep abstract features of data instances,and has more excellent clustering ability on large-scale,high-missing multi-view datasets. |