Study On The Non-parameter Kernel Learning Method Based On Feature Extraction

In recent years,with machine learning algorithms are lucubrating,the kernel learning is becoming a new hotspot,which focuses on studying how to build a model and obtain an excellent kernel matrix to describe the intrinsic geometric structure of data felicitously.Traditional kernel learning models need to establish on the predefined kernel functions,such as multiple kernel learning and spectral kernel learning model and so on.Due to the influence of the kernel functions themselves,those models cannot adapt to multitudinous data,and algorithms scale poorly.In addition,the processing effect on high-dimensional data is undesirable,and the corresponding kernel matrixes always get lower accuracy and higher consumption when solving practical problems.In consideration of the above shortcomings on the traditional kernel learning methods,combining with the feature extraction methods,this thesis abandons the limitation of using specified kernel function,while establishes optimization model aiming at desirable kernel matrix directly,and mine the cluster structure of data with the help of supervisory information better.After studying the combination of the nonparameter kernel learning model and sparse representation method,sparse auto-encoder method,this thesis solve the kernel learning problem on high-dimensional data effectively.The studies mainly include the following three aspects:(1)In view of the problem about poor scalability and high consumption,this thesis adopt a non-parameter low rank kernel learning method which construct the proportional optimization problem with the supervision information and topology structure of data.And the model is processed with the low-rank decomposition method.The scalability and learning efficiency of kernel learning method is improved.(2)To solve the problem that kernel learning model is weak on processing the highdimensional data,the thesis propose a low rank non-parameter kernel learning method based on sparse representation.The sparse representation theory is applied to the low rank non parametric kernel learning model,which increases the process of feature extraction and improves the accuracy of kernel clustering.(3)In order to solve the problem of high complexity in non-parametric kernel learning model,a non-parametric learning model based on sparse auto-encoder is proposed.Using sparse auto-encoder method to extract the feature of data can reduce the dimensionality fast and effectively,which greatly improves the quality and efficiency of kernel learning.