Enhanced grey wolf optimisation algorithm for feature selection in anomaly detection
Anomaly detection deals with identification of items that do not conform to an expected pattern or items present in a dataset. The performance of different mechanisms utilized to perform the anomaly detection depends heavily on the group of features used. Thus, not all features in the dataset can be...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | eng eng |
| Published: |
2022
|
| Subjects: | |
| Online Access: | https://etd.uum.edu.my/10192/1/s903692_01.pdf https://etd.uum.edu.my/10192/2/s903692_02.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Anomaly detection deals with identification of items that do not conform to an expected pattern or items present in a dataset. The performance of different mechanisms utilized to perform the anomaly detection depends heavily on the group of features used. Thus, not all features in the dataset can be used in the classification process since some features may lead to low performance of classifier. Feature selection (FS) is a good mechanism that minimises the dimension of high-dimensional datasets by deleting the irrelevant features. Modified Binary Grey Wolf Optimiser (MBGWO) is a modern metaheuristic algorithm that has successfully been used for FS for anomaly detection. However, the MBGWO has several issues in finding a good quality solution. Thus, this study proposes an enhanced binary grey wolf optimiser (EBGWO) algorithm for FS in anomaly detection to overcome the algorithm issues. The first modification enhances the initial population of the MBGWO using a heuristic based Ant Colony Optimisation algorithm. The second modification develops a new position update mechanism using the Bat Algorithm movement. The third modification improves the controlled parameter of the MBGWO algorithm using indicators from the search process to refine the solution. The EBGWO algorithm was evaluated on NSL-KDD and six (6) benchmark datasets from the University California Irvine (UCI) repository against ten (10) benchmark metaheuristic algorithms. Experimental results of the EBGWO algorithm on the NSL-KDD dataset in terms of number of selected features and classification accuracy are superior to other benchmark optimisation algorithms. Moreover, experiments on the six (6) UCI datasets showed that the EBGWO algorithm is superior to the benchmark algorithms in terms of classification accuracy and second best for the number of selected features. The proposed EBGWO algorithm can be used for FS in anomaly detection tasks that involve any dataset size from various application domains. |
|---|