Online Active Learning with Imbalanced Classes

Title: Online Active Learning with Imbalanced Classes

Speaker: Zahra Ferdowsi

Abstract:

Real-world machine learning systems are often embedded in large interactive systems involving experts in the loop. Once the classifier is trained on a pool of known examples, such systems classify a large number of new examples and present the experts with a ranked list of examples to review and verify. The experts often have limited time, are expensive, and are concerned primarily with finding positive instances. The interactive nature of such systems, together with limited labeling resources, high labeling costs, and the large number of unlabeled examples lend this problem well to the use of active learning techniques. Selecting the most informative examples to query from the experts could be very challenging since there is no instance selection strategy that consistently works better than others. In this talk, I will discuss the challenges of using these techniques to select the best examples and present a new online algorithm that switches between different candidate instance selection strategies for classification in imbalanced data sets.

About the Speaker:

Zahra is a PhD student at DePaul Univesity working on practical machine learning and active learning applications. She has experience in analytics across healthcare, real estate, and financial services industries. As a data scientist at Groupon, she has been working on risk assessment of the merchants and demand forecast for local businesses.