Research On Incomplete Multi-View Clustering Incorporating Prior Information

In recent years,the application of multi-view data in daily life is becoming more and more extensive,the research on multi-view clustering has attracted the attention of a large number of scholars.In real life,due to factors such as human operation errors and machine failures,there are a large number of missing multi-view data.Although it is common for multi-view data to be missing,it does not mean that we can do nothing for multi-view data but abandon it.As long as the prior information about missing positions,weak labels and other information,which is easily obtained in the process of data collection can be reasonably used,the missing multi-view data can also be used efficiently.（1）Incomplete Multi-view Learning via Half-quadratic MinimizationTo make better use of the prior information of the missing positions,a method called Incomplete Multi-view Clustering Method via Half-quadratic Minimization（IMLHM）is proposed.IMLHM is based on the matrix decomposition technology and introduces a ro-bust estimator based on semi-quadratic minimization.The robust estimator leads a more robust loss function,which overcomes the commonly problem thatl₂norm is too sensi-tive.So that the absent samples and noise points are assigned smaller weights,and high quality instances are assigned a greater weight.Additionally,a nuclear norm is introduced to exploit the low-rank structure of the learned representation matrix,further improving the robustness of the proposed method against to noise.Comprehensive experimental re-sults on seven data sets verify the effectiveness of the proposed method.（2）Incomplete Multi-view Clustering Incorporating Compound PriorIn the process of collecting incomplete multi-view data,in addition to the common prior information of missing locations,there are many other types of prior information that are widely available and easily obtained.Pairwise constraints are the most typical weak label priors information.To make full use of both types of prior information simultane-ously,we propose the Incomplete Multi-view Clustering Incorporating Compound Prior（IMC-ICP）method.Concretely,we explicitly express the pairwise constraint as a graph regular term,then with the help of self-expression technology,we organically integrate the matrix factorization-based incomplete multi-view clustering method and the incom-plete multi-view clustering method based on graph.So that the model can combine the advantages of the two types of methods,and make reasonable use of the prior information of the paired constraints simultaneously.The algorithm is optimized and solved by alter-nate iteration,verified on 7 actual data sets and applied to the scene of face clustering in the video.