Size And Binary Constrained Spectral Clustering:Methods And Application

Spectral clustering algorithm is a basic task in machine learning,pattern recognition and data mining.The existing research work shows that adding a small amount of size or binary prior constraints to the clustering process can effectively improve the clustering effect.Size constraints refer to the number of known clusters as prior constraints;binary constraints refer to the must-link and cannot-link constraints,must-link indicate two samples must belong to the same class,cannot-link indicate two samples must belong to different classes.After introducing the constraints,the algorithm can optimize the clustering results in the case of embedding adaptive constraints.The existing constrained spectral clustering methods are not significant in clustering results improvement,and time loss is higher than that without constraints.It is necessary to optimize the way to embed constraints to improve clustering accuracy.This paper proposes a spectral clustering algorithm based on size constraints and binary constraints.These constraints are processed as regular terms in the objective function,our approach is based on Normalized Cut framework,so the solution is solved by the method of eigenvalue decomposition.During the experiment,we select UCI data set and video face clustering data sets for the verification of reliability and robustness of the model,and compare the experimental results with the existing optimal constrained spectral clustering methods,our approach have a very big advantage on accuracy improvement and less cost of experiment time.The main contributions are as follows.(1)We use radial basis function method to construct similarity matrix to reduce the time cost,it is better than random forest,and don't need to do feature selection and pruning.And we also don't need to do the SDP procedure.(2)The constraint embedded in the objective function reflects the unity and self-adaptation ability of the spectral clustering in the way of regularization terms.(3)Try to apply the size constrained spectral clustering to the real face clustering experiment,which provides a new way for the video face recognition.