The Technology And Applications Of Subspace Learning

With the advancement of China's smart city strategy,big data has became an important basic resource for our state management.For big data,one important feature is the high dimensionality.For the intelligent auxiliary decision system,it is unscientific to directly use big data as a system input,and the input data dimension is crucial to the correctness of the final output decision result.The dimensionality reduction of input data to obtain accurate and effective feature dimensions,which becomes a hot topic of technology.The main research of this paper is the subspace learning technology and its applications,which mainly includes theoretical study of subspace learning,Krylov subspace theory and applied research and invariant subspace theory and its application research.The goal of the study consists of reducing the redundant dimensions in high-dimensional data and extracting effective features.In terms of learning technology,this paper is based on data clustering to undertake the study on the practical problems that including image content decomposition and video scene clustering.The main research results of this paper are as follows:1.Put forward a subspace learning model which is based on polynomial smooth ranks:With sufficient research on the truncated kernel norm and compressive perception model,we first use the polynomial function to approximate the non-convex and non-smooth low-rank decomposition objective function,and then use the generalized Lagrangian multiplier method to solve the subspace learning model.Finally,using the method of spectral clustering,we conduct an experiment was on the face recognition,and the results of the experiment shows that the proposed model is effective.2.Offer a subspace learning method which is based on Krylov decomposition:The main idea of us is that using the Krylov decomposition to structure solution method,and then complete the low rank model to solve problem.The experimental results in image decomposition show that this method has speed advantage for solving the high-dimensional low-rank decomposition problem.3.Present a clustering technique which is based on invariant subspace:First of all,for the input data,we take minimized error as the criterion to reformulate the data features.And then design clustering method on the new features.Experiments in video scene clustering show that the method is very effective,and the effectiveness of the scene clustering reach 90%.Generally speaking,based on the subspace learning technology,this thesis makes a deep research on the corresponding theory and obtains some meaningful results.The results can be directly used to the algorithm design in the field of the intelligent decision-making system and serve the corresponding intellectualized system.