Multi-Objective Hybrid Algorithm For The Classification Of Imbalanced Datasets
Classification of imbalanced datasets remained a significant issue in data mining and machine learning (ML) fields. This research work proposed a new idea based on the optimization for handling the imbalanced datasets. A new self-adaptive hybrid algorithm (CSCMAES) is introduced for optimization....
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/48598/1/Sana%20Saeed%20thesis%20Ph.D%20cut.pdf http://eprints.usm.my/48598/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Classification of imbalanced datasets remained a significant issue in data mining
and machine learning (ML) fields. This research work proposed a new idea based
on the optimization for handling the imbalanced datasets. A new self-adaptive hybrid
algorithm (CSCMAES) is introduced for optimization. The proposed algorithm is
grounded on the two famous metaheuristic algorithms: cuckoo search (CS) and covariance
matrix adaptation evolution strategy (CMA-es). For its fast convergence and for
its efficient search procedure, the self-adaptation is proposed in the parameters of the
proposed hybrid algorithm. The effectiveness of this algorithm is verified by applying
it on the unconstrained and constrained test functions through a simulation study.
From the simulation study, it is shown that CSCMAES performed very well on each
test function and produced the best values with minimum standard deviation and with
faster convergence. Thereafter, a multi-objective hybrid algorithm (MOHA), an extension
of the self-adaptive hybrid algorithm is proposed and tested on the established
multi-objective (MO) test functions. The proposed MOHA performed very well on
these test functions. A new methodology is presented for the classification of the imbalanced
datasets. The key idea of this methodology is to estimate the probabilities
for each case in both classes separately. For this purpose, the normal distributions are
applied to each class. The parameters of this distribution are optimized by applying the
proposed MOHA. An efficient performance of this proposed methodology is observed
with the help of an experimental study in which three types of datasets; simulated
datasets, noisy borderline datasets and real-life imbalanced datasets are engaged. |
|---|