Research On Dimensionality Reduction Classification Of T-SNE Combined With Support Vector Machine

In the era of big data,better use of data and better extraction of information and knowledge from large amounts of data have become common needs in all walks of life.Traditional classification algorithms usually perform classification based on the distance or density between data.High dimensions are usually sparse spaces,so commonly used classification methods lose their meaning.In addition,certain calculations usually result in huge time costs and computer memory overhead for directly using classification algorithms to classify high-dimensional data,which also leads to the inability to use conventional classification algorithms.The dimensionality reduction algorithm processing of high-dimensional data in advance is not only for more convenient classification and follow-up research,but also a means of data preprocessing.Dimensionality reduction can help us better understand and feel the data.This paper proposes a t-Distributed Stochastic Neighbor Embedding(t-SNE)dimensionality reduction combined with Support Vector Machine(SVM)multi-class combination method based on data augmentation.First,the decoder and encoder are designed to be a three-layer Variational Auto-Encoder(VAE).The encoder extracts the main feature information of the sample data,and then the decoder outputs a sample that meets the requirements and is similar to the original input sample.To achieve the effect of data enhancement,make the original data more regular,make the same category closer,and different categories further away.Secondly,the data enhanced by the data uses t-SNE for dimensionality reduction processing,t-SNE can maintain high-dimensional space It has the same probability distribution as the low-dimensional space,and uses KL divergence to optimize the objective function,builds a VAE-t-SNE data enhancement dimensionality reduction model,and then uses SVM for multi-classification.This article uses two multi-classification methods,one pair One-to-many classification and one-to-many classification are used to classify the data after dimensionality reduction.This paper compares the VAE-t-SNE algorithm with other five mainstream dimensionality reduction algorithms in subsequent experiments.Among them,the accuracy of the VAE-t-SNE algorithm in this paper is better than other algorithms,and it can be read from the figure,It can be clearly observed that the aggregation between the same categories and the outlines between different categories are clearer.the feasibility and effectiveness of the model are therefore verifyied in this paper.