Hybrid Incremental Learning Algorithm Based On Support Vector Machine And Its Application

SVM (Support Vector Machine, SVM) is based on statistical theory of a Machine learning methods, it has a solid theory foundation.In dealing with small samples, high altitude, it have more advantages in the nonlinear pattern recognition learning problems.It is essentially a quadratic programming problem.The research emphasis of this paper is based on incremental learning algorithm of support vector machine research and its application in network intrusion.On the analysis of the support vector machine on the incremental learning behavior, and the lack of some existing incremental learning algorithm, this paper proposes a new incremental learning algorithm based on support vector machine, which is based on the cosine similarity and distance function incremental SVM learning algorithm, referred to as "CSFD-ISVM".The main core idea of the algorithm is to use the training sample set distribution characteristics and historical results as far as possible to reduce the number of samples. This article puts forward several ways to filter sample and accelerate the process of sample selection.This article puts forward several innovative ways to filter sample and accelerate the process of sample selection.The main work and innovation summarized below:1) Based on the idea of SVM, the final decision surface with only a few support vectors, so you can advance in support vector, a sampler can greatly reduce participate in the training sample, so to speed up the training.A by category centroid and Angle cosine samples primary strategy.By the law of large Numbers, a large number of samples randomly selected a certain number of center of mass of the sample is close to the overall center of mass of the sample, therefore put forward using a certain proportion randomly sample instead of the whole sample cosine method to calculate the sample.On this basis, this paper based on the cosine similarity and distance function incremental learning algorithm. The experimental results show that:based on the cosine similarity and distance function incremental learning algorithm effectively reduces the training time, and also can improve the training accuracy.2) Based on incremental learning before and after the change of the sample distribution analysis found that the useful information in the historical samples in addition to the support vector, there is also very important boundary support vector, vector for edge extraction, this paper puts forward a kind of based on cosine similarity and function of the distance prefetch strategy, and analyzes the algorithm through the experiment of the function of the distance tolerance factor influence on incremental learning, show that the method of incremental learning accuracy has been increased obviously.By UCI standard data sets and other data sets of simulation experimental results show that the proposed based on the cosine similarity and function from the prefetch strategy of incremental learning algorithm is feasible and effective.Finally it is applied to network intrusion detection, by KDD CUP99 data simulation experiments, the proposed hybrid incremental learning algorithm in training speed and has good performance in detection rate, and maintain a high detection rate and low false alarm rate, so it's suitable for training the classification model of intrusion detection.