Research Article
A Novel Feature-Selection Approach Based on Particle Swarm Optimization Algorithm for Intrusion Detection Systems (Workshop Paper)
@INPROCEEDINGS{10.1007/978-3-030-30146-0_32, author={Jianzhen Wang and Yan Jin}, title={A Novel Feature-Selection Approach Based on Particle Swarm Optimization Algorithm for Intrusion Detection Systems (Workshop Paper)}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 15th EAI International Conference, CollaborateCom 2019, London, UK, August 19-22, 2019, Proceedings}, proceedings_a={COLLABORATECOM}, year={2019}, month={8}, keywords={Feature selection Discrete particle swarm algorithm Correlation analysis Correct classification rate Modeling efficiency}, doi={10.1007/978-3-030-30146-0_32} }- Jianzhen Wang
Yan Jin
Year: 2019
A Novel Feature-Selection Approach Based on Particle Swarm Optimization Algorithm for Intrusion Detection Systems (Workshop Paper)
COLLABORATECOM
Springer
DOI: 10.1007/978-3-030-30146-0_32
Abstract
This paper proposes a feature selection approach, based on improved Discrete Particle Swarm Optimization (DPSO), to solve the “dimension disaster” problem in data classification; it is named Progressive Binary Particle Swarm Optimization (PBPSO). This feature selection approach is highly problem-dependent and influenced by the locations of particles. It adopts the principle of “partial retention - change - reduction of duplication - update” in the process of selection, and defines a new fitness function describing the correlation between the features and class labels. Experimentation was conducted using of the KDDCup99 data set to evaluate our proposed PBPSO. The experimental results show that 14 features were selected from the original data space with 41 features. Three classic classifiers, namely J48, Naive Bayes and ID3, were then used to further evaluate the performance of the selected features. The classification accuracy rates on the different classifiers achieved using the selected feature subset are similar to those achieved using the original feature set. The training time is, however, significantly reduced. In comparison with other similar algorithms, including Genetic Algorithm GA and Greedy Algorithm FGA. The results show that the PBPSO extracts fewer features, achieves slightly higher classification accuracy, and less time consuming in terms of model training. It has been demonstrated that the PBPSO enhances the practicability of certain classification algorithms in handling high-dimensional data.
