Large-scale Data Analysis via Online Optimization and Distance Learning

In the big data era, many existing machine learning algorithms are not applicable due to various performance constraints. In this thesis, approaches using online optimization and distance learning have been proposed under the large-scale setting for some typical machine learning topics, such as: 1) streaming data problem 2) rich data with limited label problem and 3) multimodal distribution and imbalanced data problem. These machine learning topics are inspired from real world applications. In addition, a unified framework has been proposed for a general large-scale classification problem. This framework involves four major components: 1) feature extraction 2) feature selection 3) distance measure and 4) classification. Finally, some real world Large-scale classification problems solved by this framework have been included in this thesis, such as: Mars Crater detection, Boston Crime prediction and activity recognition from accelerometer data