Design methods for improving robustness of ensemble learning classifiers


Student thesis: Doctoral Thesis

View graph of relations


  • Jingjing CAO

Related Research Unit(s)


Awarding Institution
Award date15 Feb 2013


Ensemble learning plays a significant role in supervised learning field for improving the pattern recognition capability of multiple classifier systems. The inner spirit of ensemble methodology is to train several base classifiers and combine them by assigning different weights for a better performance, which usually outperforms each individual classifier. However, the current ensemble methods still suffer from different sorts of problems, such as the unstable performance caused by the presence of noisy data, the ineffective fusion strategies for combining individual classifier outputs, and the lack of multi-objective evolutionary algorithm based methods to break multi-class problem into dichotomies. Further, with regard to the fuzzy genetic system with multi-objectives, it is easy to apply the ensemble pruning and diversity techniques to select less fuzzy rule classifiers and increase the diversity between each pair of individuals in evolutionary process. Inspired by the above problems, this thesis investigates several methodologies for improving robustness of ensemble learning classifiers which involves a noisy-detection based approach for Adaptive Boosting (AdaBoost), a class-specific weighted fusion method for Extreme Learning Machine (ELM), an indicator selection based multi-objective evolutionary algorithm with preference for multi-class classification systems, and an algorithm by applying ensemble pruning and diversity techniques for multi-objective hierarchical evolutionary algorithm. Specifically, the main contributions of this thesis are outlined as follows: 1. A new boosting approach, named noise-detection based AdaBoost (ND-AdaBoost), is exploited to combine classifiers by emphasizing on training misclassified noisy instances and correctly classified non-noisy instances. This algorithm is proposed based on the fact that AdaBoost is prone to overfitting in dealing with the noisy data sets due to its consistent high weights assignment on hard-to-learn instances (mislabeled instances or outliers). Concretely, the algorithm is designed by integrating a noise-detection based loss function into AdaBoost to adjust the weight distribution in each iteration. A k-nearestneighbor (k-NN) and an expectation maximization (EM) based evaluation criteria are both constructed to detect noisy instances. Further, a regeneration condition is presented and analyzed to control the ensemble training error bound of the proposed algorithm which provides theoretical support to the algorithm. 2. A class-specific weight based soft voting method is presented for the design of ELM ensembles (CSSV-ELM). The class-specific soft voting method is a common approach to deal with the base learner which produces outputs with class probabilities. Further, the new algorithm is designed based on the incorporation of two important characteristics for improving the reliability of ELM. Firstly, the individual ELM classifiers have unequal performances since the initialization of the hidden node learning parameters are randomly generated. Secondly, as a linear equation system based algorithm, ELM may suffer the illconditioned problem. According to these two factors, the classic weighted voting scheme and the condition number of matrix are integrated into CSSV-ELM algorithm. Additionally, compared with other neural networks, there is no empirical research in weighted voting based ELM ensembles. In this work, we also compare and analyze seven weighted voting methods with the proposed method. 3. One of the most difficult components for multi-class classification system is to find an appropriate Error-Correcting Output Codes (ECOC) matrix, which is used to decompose the multi-class problem into several binary class problems. In this thesis, an indicator based multi-objective evolutionary algorithm with preference involved is designed to search the high-quality ECOC matrix. Specifically, the Harrington's one-sided desirability function is integrated into an indicator-based evolutionary algorithm (IBEA), which aims to approximate the relevant regions of pareto front (PF) according to the preference of the decision maker. 4. The contributions of the proposed evolutionary algorithm are two-fold: firstly, it employs a multi-objective evolutionary hierarchical algorithm (MOHEA) to obtain a non-dominated fuzzy rule classifier set with interpretability and diversity preservation. Secondly, a reduce-error based ensemble pruning method is utilized to decrease the size and enhance the accuracy of the combined fuzzy rule classifiers. In this algorithm, each chromosome represents a fuzzy rule classifier and consists of three different types of genes: control, parameter and rule genes. In each evolution iteration, each pair of classifiers in non-dominated solution set with the same multi-objective qualities is examined in terms of Q statistic diversity values. Then, similar classifiers are removed to preserve the diversity of the fuzzy system.

    Research areas

  • Machine learning