A direct ensemble classifier for learning imbalanced multiclass data
A traditional direct single classifier can be easily applied to solve a multiclass classification problem. However, the performance of a single classifier is decreased with the existence of imbalanced data in multiclass classification tasks. Thus, an ensemble of classifiers is one of the methods use...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/38557/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/38557/2/FULLTEXT.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A traditional direct single classifier can be easily applied to solve a multiclass classification problem. However, the performance of a single classifier is decreased with the existence of imbalanced data in multiclass classification tasks. Thus, an ensemble of classifiers is one of the methods used to solve multiclass classification tasks. In this thesis, the problem of learning from imbalanced multiclass data classification is studied. In the multiclass classification problem, decision can be estimated not only by the final single class label, but also by other appropriate class. Many real-world multiclass classification problems can be represented into a setting where non-crisp label need to be observed. An in-depth review and method to solve this special learning task is explained in this thesis. An alternative ensemble learning framework called Direct Ensemble Classifier for Imbalance Learning (DECIML) is proposed combining the advantages of existing single classifiers and ensemble methods and strategies. The learning framework consists of ensemble learning and decision combiner model with general supervised learning algorithms as base learner. Feature selection is also applied in DECIML in order to increase the performance of the ensemble learning. In order to facilitate the experiments and future research on the imbalanced multiclass problem, a standard pool of benchmark data is created, which consists of 16 datasets with different degrees of imbalanced ratio and 4 datasets for imbalanced multiclass with feature selection purposes. The benchmark data is used to evaluate and compare the proposed frameworks with several ensemble methods, such as bagging and adaboost. The DECIML with feature selection is also evaluated and compared with methods named CFsSubsetEval and Filteredsubseteval. The results obtained show that the proposed learning frameworks are comparable to other methods. In addition, the selected benchmark data, experiments and the results are useful for future research on the imbalanced multiclass classification problem. Furthermore, the DECIML framework was applied to the real world leaf classification problem based on the shape features. Extensive experiments and results show that the DECIML method does provide a promising performance in imbalanced multiclass with highly noisy data. |
|---|