|Title||:||Learning with Complex Performance Measures: Algorithms, Theory and Applications|
|Speaker||:||Harikrishna Narasimhan (IISc Bangalore)|
|Details||:||Fri, 15 May, 2015 3:00 PM @ BSB 361|
|Abstract:||:||In many real-world applications of machine learning, ranging from information retrieval to medicine to bioinformatics, one is required to learn accurate prediction models from data sets that have severe imbalanced characteristics. Such applications often warrant the use of complex/non-additive performance measures to evaluate the learned prediction model. These include several evaluation measures used in text retrieval (e.g. F-measure), in rare event prediction problems like fraud detection (e.g. G-mean), and in ranking/prioritization tasks such as drug discovery (e.g. Precision@K). How does one design efficient learning algorithms for such complex performance measures? Can the algorithms developed be shown to be consistent, i.e., shown to converge in the limit of infinite data to the optimal prediction model? What are interesting applications that benefit from such algorithms? These are some questions that I have been exploring for the past few years.
In this talk, I shall focus on a recent work of ours in this space on multiclass learning with complex performance measures, a setting where efficient (let alone consistent) algorithms have largely remained elusive. I shall present the first efficient and consistent learning algorithms for a large family of multiclass performance measures that are complex functions of the confusion matrix, including both concave evaluation measures such as the multiclass G-mean and ratio-of-linear measures such as the micro F-measure. Our algorithms outperform the state-of-the-art SVMPerf method in experiments and are often orders of magnitude faster. I shall conclude by briefly discussing some of our other related directions of research, including an applied work with life scientists on a problem in personalized cancer treatment.
Speaker Bio: Harikrishna Narasimhan is a PhD candidate in the Department of Computer Science and Automation at the Indian Institute of Science (IISc), Bangalore, working with Shivani Agarwal, and is supported by a Google India PhD Fellowship in Machine Learning. Prior to joining PhD, he completed his masters from the same department and bachelors from College of Engineering, Guindy, Anna University, Chennai. He has been a short-term research fellow at Harvard University, working with David Parkes and Yaron Singer, and a research intern at Microsoft Research India with Prateek Jain. His research interests lie in the areas of Machine learning, Optimization, and Learning Theory, with Applications to the Life Sciences, and more recently in problems at the interface of Machine Learning and Social Science. Details of his publications can be found at: http://clweb.csa.iisc.ernet.in/harikrishna/.