A novel method for SVM samples reduction

The support vector machine is one of the best classification algorithms that has been successfully used in many machine learning problems. Despite the many capabilities of SVMs, they have certain weaknesses. Perhaps their main drawback is the high computational cost in the large dataset. Because kernel training has grown in size with datasets, this makes it impossible to use SVM even for a moderate problem. In addition, in some cases, a large number of support vectors are generated, which in practice means an increase in test and execution time. In this paper, a method to reduce the training samples required for SVM training is proposed, which does the training only by selecting a very small subset of training data. This will reduce the time required for training and memory. Also, in this method, the test and execution time will be reduced by reducing the number of support vectors. In addition, in the dataset where there is an imbalance in the number of class samples and there is a tendency for the majority class, the proposed method creates a balance in the two class samples, which makes it possible to use this method, used in unbalanced datasets. The experimental results show that the proposed method, despite a significant reduction in training time and memory requirements, has an acceptable accuracy compared to conventional methods.

کلیدواژه ها:

Support vector machine ، Linear independency ، Large-scale learning ، Classification ، Sampling