[intermediate/advanced] Learning from Class Imbalances
The class imbalance problem was first coined in the mid-1990s, when machine learning algorithms became robust enough to be applied in real-world settings. At that point, a myriad of new problems came up including the ubiquitous class imbalance problem. Since then, many remedies have been proposed, but the problem persists. This course will introduce the problem and relate it to other common problems in machine learning including cost-sensitive learning, long tailed distributions, data scarcity, fairness, anomaly detection, and evaluation. It will then show how the seriousness of the problem increases in the presence of different kinds of data characteristics and what the effect of increasing neural network depth has on it. It will then present the different kinds of solutions that have been proposed to deal with the problem, including cost-sensitive approaches, data resampling methods, and one-class learning. The discussion will span both classification and other learning paradigms.
Lecture 1: Introduction to the class imbalance problem and its relation to other common problems
Lecture 2: Understanding the causes of the class imbalance problem and the effect of network depth on it
Lecture 3: Proposed solutions for dealing with class imbalances
- Japkowicz, N. and Shaju Stephen. “The class imbalance problem: A systematic study.” Intell. Data Anal. 6 (2002): 429-449.
- He, Haibo and Edwardo A. Garcia. “Learning from Imbalanced Data.” IEEE Transactions on Knowledge and Data Engineering 21 (2009): 1263-1284.
- Branco, Paula et al. “A Survey of Predictive Modeling on Imbalanced Domains.” ACM Computing Surveys (CSUR) 49 (2016): 1 – 50.
- Krawczyk, B.. “Learning from imbalanced data: open challenges and future directions.” Progress in Artificial Intelligence 5 (2016): 221-232.
- Johnson, Justin M. and T. Khoshgoftaar. “Survey on deep learning with class imbalance.” Journal of Big Data 6 (2019): 1-54.
- Ghosh, Kushankur et al. “On the combined effect of class imbalance and concept complexity in deep learning.” ArXiv abs/2107.14194 (2021): n. pag.
Introductory course on Machine Learning or Data Mining.
Nathalie Japkowicz is a Professor and Chair of the Computer Science Department at American University, Washington DC. She was previously with the School of Electrical Engineering and Computer Science at the University of Ottawa where she lead the Laboratory for Research on Machine Learning for Defense and Security. Her work has spanned different areas of Machine learning, but focused primarily on the class imbalance problem, anomaly detection using one-class learning, and machine learning evaluation. She worked in a number of domains in the areas of radiation protection, cyber security, medicine and molecular biology among others. She has supervised over thirty graduate students, received funding from Canadian and American institutions, worked with governmental agencies as well as private companies, and published over 180 peer-reviewed journal articles and conference papers, together with special issues and books including Evaluating Learning Algorithms: A Classification Perspective, with Mohak Shah (Cambridge University Press, 2011). She is a past president of the Canadian Artificial Intelligence Association and she received a number of best paper awards as well as the Canadian Artificial Intelligence Association’s Distinguished Service Award.