Research On Improved Support Vector Machine Classification Algorithm Based On Tensorflow

With the enhancement of computer processing ability,informatization produces a large number of data.Machine learning is the way to solve problems by learning large amounts of data.Tensorflow has different optimization algorithms that provide computational space for a variety of machine learning tasks.One of the main tasks in machine learning is to deal with classification problems.Support Vector Classification algorithm is a classical classification algorithm in machine learning.In the classification problem of Support Vector Machine,samples close to the separation plane decide the final segmentation plane,and these samples are called "support vector".The expression of the final prediction model is related to "support vector".As the large sample size leads to high solution complexity,saving possible "support vector" reduced sample set before training can shorten the solution time.Support vector machine and neural network both have nonlinear approximation characteristics,but neural network is prone to fall into local optimality.Based on the above background,the research contents of this paper are as follows:(1)To solve the problem of high complexity due to the large sample size of SVM classification algorithm,this paper proposes to use the peripheral probability of k-nearest neighbor samples to reduce the sample set,retain the possible "support vector" samples,and eliminate the non-support vector sample reduced data set.In support vector machine(SVM)problem solving,there is a significant difference between the original problem and dual problem in computing rate.When the number of sample sets is large,the dual problem solving complexity is high and the rate is low,which is suitable for the original problem solving.This paper USES Tensorflow framework to solve the original problem and dual problem of support vector classification algorithm,and USES grid search to find the optimal hyperparameter of support vector machine.Firstly,the effectiveness of using the K nearest neighbor sample peripheral probability to reduce the data set was verified.Finally,under the UCI data set,by comparing Tensorflow to solve SVM and SVM in python library,good accuracy and F1 were achieved on some data sets.(2)Based on the study of support vector classification algorithm,in order to solve the neural network easy to fall into local problems such as excellent performance,network is not stable,based on neural network and support vector machine(SVM)distinction,analysis of the causes of poor neural network effect,the priori knowledge,using the gaussian support vector clustering features as prior to optimize network,because the neural network is influenced by the initialization easily trapped in local optimum,this paper use noise reduction from the encoder to advance training for initial weights of the network,combined with neural network and support vector machine(SVM)is used,giving full play to the advantages of both.First contrast noise reduction from the encoder to initialize the network weights and random initialization of the network weights of network accuracy and the effect of F1,secondly compared to the gaussian kernel support vector clustering characteristics as a priori characteristics affect the performance of network,finally under the UCI data sets,the algorithm in this paper with commonly used classification algorithms(decision tree,logistic regression,neural network and support vector machine),most of the data set has better accuracy and F1.