Balancing Exploitation And Exploration Search Behavior On Nature-Inspired Clustering Algorithms
Nature-inspired optimization-based clustering techniques are powerful, robust and more sophisticated than the conventional clustering methods due to their stochastic and heuristic characteristics. Unfortunately, these algorithms suffer with several drawbacks such as the tendency to be trapped or sta...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/47436/1/Balancing%20Exploitation%20And%20Exploration%20Search%20Behavior%20On%20Nature-Inspired%20Clustering%20Algorithms.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Nature-inspired optimization-based clustering techniques are powerful, robust and more sophisticated than the conventional clustering methods due to their stochastic and heuristic characteristics. Unfortunately, these algorithms suffer with several drawbacks such as the tendency to be trapped or stagnate into local optima and slow convergence rates. The latter drawbacks are consequences of the difficulty in balancing the exploration and exploitation processes which directly affect the final quality of the clustering solutions. Hence, this research has proposed three enhanced frameworks, namely, Optimized Gravitational-based (OGC), Density-Based Particle Swarm Optimization (DPSO), and Variance-based Differential Evolution with an Optional Crossover (VDEO) frameworks for data clustering. In the OGC framework, the exhibited explorative search behavior of the Gravitational Clustering (GC) algorithm has been addressed by (i) eliminating the agent velocity accumulation, and (ii) integrating an initialization method of agents using variance and median to subrogate the exploration process. Moreover, the balance between the exploration and exploitation processes in the DPSO framework is considered using a combination of (i) a kernel density estimation technique associated with new bandwidth estimation method and (ii) estimated multi-dimensional gravitational learning coefficients. Lastly, (i) a single-based solution representation, (ii) a switchable mutation scheme, (iii) a vector-based estimation of the mutation factor, and (iv) an optional crossover strategy are proposed in the VDEO framework. The overall performances of the three proposed frameworks have been compared with several current state-of-the-art clustering algorithms on 15 benchmark datasets from the UCI repository. The experimental results are also thoroughly evaluated and verified via non-parametric statistical analysis. Based on the obtained experimental results, the OGC, DPSO, and VDEO frameworks achieved an average enhancement up to 24.36%, 9.38%, and 11.98% of classification accuracy, respectively. All the frameworks also achieved the first rank by the Friedman aligned-ranks (FA) test in all evaluation metrics. Moreover, the three frameworks provided convergent performances in terms of the repeatability. Meanwhile, the OGC framework obtained a significant performance in terms of the classification accuracy, where the VDEO framework presented a significant performance in terms of cluster compactness. On the other hand, the DPSO framework favored the balanced state by producing very competitive results compared to the OGC and DPSO in both evaluation metrics. As a conclusion, balancing the search behavior notably enhanced the overall performance of the three proposed frameworks and made each of them an excellent tool for data clustering. |
|---|