Imbalanced classification problems are those in which populations and (mis)classification costs are very different. They are pervasive and important in many real-world application fields, from Health to Security, including Economy, Industry, Business, and Social Media, to cite just a few. Unfortunately, the powerful discriminative machines seriously fail when applied to these problems –just as humans do! Of course, there are many re-balancing techniques to reduce this difficulty… But most of them are purely empirical (or even “brute force”), and, consequently, their application faces uncertainties and risks.
In this lecture, a principled re-balancing approach, which is based in Bayes theory, in introduced for the binary classification problems. Several experiments show its advantages with respect to alternative re-balancing algorithms. A two-step principled modification to improve its performance is also presented. The many research avenues that this formulation permits to explore close the lecture.
Aníbal R. Figueiras-Vidal (Tel. Eng., UPM, 1973, Dr. Tel. Eng., UPC, 1976) has served at UPC, UPM (as a Full Prof., since 1978), USC, and at UC3M during the last 20 years. His teaching and research is focused on DSP and ML, together with their applications. He has been PI in almost 100 national and international research projects and contracts, published 80+ SCI journal papers, and supervised 30+ Doctoral dissertations. He is a member of the Spain Real Academia de Ingeniería, serving as its President from 2007 to 2011. He is a Life Fellow of the IEEE, and he has been distinguished with “Honoris Causa” Doctor degrees by Universidad de Vigo (1999) and Universidad San Pablo, Arequipa, Peru (2011).CTTC Auditorium / 13:00hProf. Aníbal R. Figueiras-Vidal, UC3M