•   When: Tuesday, February 21, 2017 from 10:00 AM to 12:00 PM
•   Speakers: Azad Naik
•   Location: ENGR 4801
•   Export to iCal

Hierarchical Classification (HC) is an important supervised learning problem where the goal is to classify unlabeled instances into a hierarchy of classes, efficiently and accurately. A vast amount of research has been conducted to address HC challenges including: (i) Class imbalance with large number of classes having very few positive examples for training (rare categories problem), (ii) Multi-label classification were instances can belong to more than one class, (iii) Inconsistencies in the hierarchy due to manual design by domain experts, (iv) Feature selection with large number of irrelevant features and (v) Need for scalability arising due to large number of examples, features and classes. In my thesis, I have developed novel and innovative approaches that deal with the rare categories and inconsistency related problems within the hierarchy definition. For dealing with rare categories a rank-based approach has been developed that exploits the hierarchical structure for generalized model learning. For dealing with hierarchical inconsistencies - node flattening and rewiring approaches were developed. Further, an approach for embedding feature selection into the HC framework has been developed which helps to improve the classification performance while reducing the memory requirements and computational runtime during the learning and prediction phases. Finally, a multi-stage integrated pipeline has been developed to solve the large-scale HC problem that can also detect orphan classes (i.e., classes with no examples in the training set).