Speeding up the testing and training time for the support vector machines with minimal effect on the performance

Support vector machine faces some problems associated with training time in the presence of large data sets due to the need for high memory and high computational cost. The main problem with Support vector machine occurs during the training phase, which is computationally expensive and depends on the size of the input data set. The high boundary complexity among the classes is another major problem with the most data sets, which will reduce the generalizability. Therefore, this study was conducted to present a useful method for reducing the number of data and boundary complexity, in which the training set is divided into the boundary, non-boundary, and harmful patterns. It has four phases: In the first phase, the information is determined about the neighborhood of each sample with other samples. In the second phase, harmful patterns are removed to reduce the data complexity. In the third phase, the remaining training data samples are divided into the boundary and non-boundary patterns. In phase 4, representatives of non-boundary data are determined and a reduced set is formed through combination of these representatives and boundary patterns. The proposed method is tested with 33 data sets and comparatively evaluated against five of the most successful instance-based condensation algorithms. Experiments showed that the proposed method is better than other methods presented in the research literature.