Ensemble predictions : empirical studies on learners' performance and sample distributions

Publication Type:
Issue Date:
Full metadata record
Imbalanced data problems are among the most challenging in Data Mining and Machine Learning research. This dissertation investigates the performance of ensemble learning systems on different types of data environments, and proposes novel ensemble learning approaches for solving imbalanced data problems. Bagging is one of the most effective ensemble methods for classification tasks. Despite the popularity of bagging in many real-world applications, there is a major drawback on extremely imbalanced data. Much research has addressed the problems of imbalanced data by using over-sampling and/or under-sampling methods to generate an equally balanced training set to improve the performance of the prediction models. However, it is unclear which is the best ratio for training, and under which conditions bagging is outperformed by other sampling schemes on extremely imbalanced data. Previous research has mainly been concerned with studying unstable learners as the key to ensuring the performance gain of a bagging predictor, with many key factors remaining unclear. Some questions have not been well answered: (1) What are the key factors for bagging predictors to achieve the best predictive performance for applications? and (2) What is the impact of varying the levels of class distribution on bagging predictors on different data environments. There is a lack of empirical investigation of these issues in the literature. The main contributions of this dissertation are as follows: 1. This dissertation proposes novel approaches, uneven balanced bagging to boost the performance of the prediction model for solving imbalanced problems, and hybrid-sampling to enhance bagging for solving highly imbalanced time series classification problems. 2. This dissertation asserts that robustness and stability are two key factors for building a high performance bagging predictor. This dissertation also derives a new method, utilizing two-dimensional robustness and stability decomposition to rank the base learners into different categories for the purpose of comparing the performance of bagging predictors with respect to different learning algorithms. The experimental results demonstrate that bagging is influenced by the combination of robustness and instability, and indicate that robustness is important for bagging to achieve a highly accurate prediction model. 3. This dissertation investigates the sensitivity of bagging predictors. We demonstrate that bagging MLP and NB are insensitive to different levels of imbalanced class distribution. 4. This dissertation investigates the impact of varying levels of class distribution on bagging predictors with different learning algorithms on a range of data environments, to allow data mining practitioners to choose the best learners and understand what to expect when using bagging predictors.
Please use this identifier to cite or link to this item: