Abstract
Statistical machine learning provides a statistical framework to improve the performance of classifiers or predictors learned from large amounts of training data collected from past experiences. In this thesis, we propose a novel AdaBoost-based algorithm based on paired feature learning to improve the performance of classifiers for several different applications. The proposed learning system contains four major improvements with the goal to achieve higher discrimination and more information preservation in the paired feature learning for the AdaBoost classifiers. These improvements include feature space warping, joint feature representation, ID3-like plane quantization and weak probabilistic classifiers. The proposed algorithm updates the sample weights via the proposed processes and achieves a more efficient and accurate AdaBoost classifier. The proposed classification method has been applied to four different applications to demonstrate its discrimination power. First, the proposed algorithm is applied to content-based image retrieval (CBIR) with relevance feedback, which can be formulated as a classification problem with a small number of training samples. Our experiments show superior performance of the proposed system compared to some previous methods. Secondly, we develop a new face detection algorithm that works directly on the wavelet compressed domain based on the same paired feature learning framework. The third application is the vertebra detection and segmentation from a spinal magnetic resonance (MR) image. The modified AdaBoost classifier is used in conjunction with a robust spinal curve fitting technique and iterative normalized cut segmentation, and the proposed system can achieve nearly 98% vertebra detection rate and high segmentation accuracy on a variety of testing spinal MR images. Finally, we apply the same paired feature learning technique to improve the contrarian trading strategy in computational finance. A dual classifiers model is adopted to discover the time-series patterns constructed by the integrated knowledge of finance and machine learning techniques. Our experiments on S&P 100 index component stocks show dramatic improvement by using the proposed algorithm. In summary, all these experiments demonstrate the improvement in accuracy by using the proposed paired feature learning in an AdaBoost classification framework.