PRACTITIONER TRACK | Oct. 11 • 4:00 pm • Room 14–15
Learning and Reasoning with Imperfect Data
Kamal Premaratne, PhD, Victor P. Clarke Professor of Electrical and Computer Engineering, University of Miami
The practical utility and effectiveness of machine learning algorithms for reasoning and inference depend on how well one may extract the relevant parameters from training data. However, adequate representative statistical training data are often unavailable and, when available, real-world training data are often mired in incomplete/missing data. This is particularly true in the detection of cybersecurity threats. Imputation of such data must be guided by how different variables are related to each other and/or by the underlying distribution which dictates data ‘missingness’. Interval-valued probabilities are better suited to deal with situations when such information is unknown or indeterminate and when one is called on to harness the more qualitative subjective human-based information; they are also what arises naturally from incomplete or partial elicitation. In this presentation, we illustrate a framework that allows for parameter learning and reasoning with interval-valued probabilities in much the same manner as one would reason with probabilities (as, for example, in a Bayesian network). For datasets where attribute values could be unknown/missing or are known to lie within a set of values, we show that an intuitive frequency counting method can be employed to learn interval-valued parameters. Importantly, the probabilities associated with an arbitrary imputation strategy, including the underlying ‘true’ probabilities, are guaranteed to be contained within these intervals. Experimental results demonstrate the utility of the proposed framework.
About Dr. Kamal Premaratne
Kamal Premaratne received a BSc in electronics and telecommunication engineering (1982) with First-Class Honors from University of Moratuwa, Sri Lanka. He obtained both his MS (1984) and PhD (1988) degrees in electrical and computer engineering under the supervision of Professor Eliahu I. Jury at the University of Miami, Coral Gables, Florida, where he is presently the Victor P. Clarke Professor. He has received the Mather Premium (1992/93) and Heaviside Premium (1999/00) of the Institution of Electrical Engineers (IEE), London, UK, and the Eliahu I. Jury Excellence in Research Award (1991, 1994, 2001) and the Johnson A. Edosomwan Researcher of the Year Award (2014) of the College of Engineering, University of Miami. He has served as an associate editor of IEEE Transactions on Signal Processing and the Journal of the Franklin Institute. He is a Fellow of IET (formerly IEE) and a Senior Member of IEEE. His current research interests include imprecise probability formalisms, especially Dempster-Shafer (DS) belief theory, evidence fusion, machine learning and knowledge discovery from imperfect data, and opinion and consensus dynamics in social networks. His work has been funded by the Office of Naval Research (ONR) and the National Science Foundation (NSF).