Abstract:
Supervised classification is a sub-task of machine learning where the goal is to infer a classification function using labeled data. A vast amount of research has been conducted in this area addressing various classification problems such as binary, multi-class and multi-label classification. However, we often encounter classification problems in real world in groups of tasks with complex interactions among them. Methods that are able to take advantage of the additional information regarding task relationships and interactions are able to perform better in terms of classification accuracy. Furthermore, with the vast amount of data that is being accumulated in the recent years the real world problems that have any practical utility have exploded in terms of problem size; with respect to number of data elements, feature size and number of class labels. Therefore, there is an urgent need for scalable methods that are able to gracefully scale to web-scale problems. In my thesis, I have tried to address these issues by developing novel classification methods for large scale hierarchical classification and multi-task learning.
