A SVM Learning Method Based On Local Linear Embedding

With the rapid development of network technology in recent years,the data scale is also growing rapidly.SVM takes a long time for high-dimensional data classification.PCA-SVM has higher classification rate than SVM but its accuracy is relatively low.An improved LLE method is proposed based on the study of the characteristics of nonlinear dimensionality reduction methods to improve the SVM.The SVM method is used for classification after the LLE dimension reduction of high-dimensional data to reduce the number of new samples.The incremental SVM learning based on local linear embedding(LLE-ISVM)is implemented.The algorithm is applied to the classification process of MNIST database and surface defects of tiles.For the data sets with a small number of unlabeled data and a large number of labeled data,Due to the lack of prior knowledge,there may be labeling errors due to human-labeled labels.The supervised learning is difficult to study the datasets.Experimental results show that the LLESVM incremental algorithm has improved the speed and accuracy of high-dimensional data.LLE-TSVM can make semi-supervised learning better.Compared with existing semi-supervised algorithms,LLE-TSVM has certain improvements in accuracy and speed.The main content of this article is as follows:(1)For the difficulties of traditional SVM methods for high-dimensional data processing,some improvements are proposed.Introduce the background of machine learning,manifold learning methods,SVM,incremental learning,manifold learning and semi-supervised methods.(2)Propose the classical SVM method and an improved method using nonlinear manifold method.After comparing ISOMAP and LLE,local linear embedding(LLE)was chosen as the nonlinear dimensionality reduction method.Analyze the deficiencies of PCA-SVM in highdimensional nonlinear data processing.The LLE method is used to improve the SVM.The SVM classifier hyperplane is only determined by the SV set large data sets.During the increasing learning the data sets have only a few samples with classification information and do not need to train all samples.When the training data is large,the non-incremental SVM training calculation time is often very long,the idea of incremental learning is used to improve the SVM.LLE-SVM algorithm can effectively handle the classification process of a large number of high-dimensionalnonlinear data sets.(3)Since the LLE-SVM Incremental Learning Method is a supervised learning algorithm that applies only to the dataset training process with all data tagged.It is also difficult to obtain the correct labels for all datasets in the actual process.Propose the TSVM classification learning algorithm based on LLE.Add the new data sets to the original SV set,then use the TSVM classification learning algorithm to obtain a new classification interface.LLE-TSVM has a high accuracy for the classification of handwritten data sets,and the tile classification process can be fully realized.LLE-TSVM incremental learning method can be applied to semi-supervised classification.(4)Finally,this paper summarizes and prospects LLE-SVM and LLE-TSVM,and identifies some more details to be improved in the future.