A performance comparison of feature selection techniques with SVM for network anomaly detection

Feature selection techniques are important to select features for good classification results. In this work, Correlation based Feature Selection, Motif Discovery using Random Projection, Hybrid Feature Selection and Convex Hull Feature Selection techniques with Support Vector Machine are compared using the same dataset for network anomaly detection. The performance metrics are number of features, an accuracy rate and training time. The results showed that HFS-SVM gave the minimum number of features and CH-SVM gave the good accuracy rate for all range of training data records. However, with the training data more than 300,000 records, the HFS-SVM gave similar accuracy rate as CH-SVM. For the training time, HFS-SVM gave minimum training time since it used minimum number of features.