23 Apr 10:15 Juan Diego Rodríguez: Learning Bayesian network classifiers for multi-dimensional supervised classification problems by means of a multi-objective approach

HIIT seminar, Friday Apr 23, 10:15 a.m. (coffee from 10), Exactum B222

Juan Diego Rodríguez
Intelligent Systems Group, University of the Basque Country, Spain
Visiting Researcher at HIIT (Complex Systems Computation Group)

Learning Bayesian network classifiers for multi-dimensional supervised classification problems by means of a multi-objective approach

Abstract:
A classical supervised classification task tries to predict a single class variable based on a data set composed of a set of labelled examples. However, in many real domains more than one variable could be considered as a class variable, so a generalization of the single-class classification problem to the simultaneous prediction of a set of class variables should be developed. This problem is called multi-dimensional supervised classification.

In this talk, I will present the recently developed Multi-dimensional Bayesian network classifiers, as a generalization of the classical single-class Bayesian network classifier to the prediction of several class variables. I will also present a learning approach following a multi-objective strategy wich considers the accuracy of each class variable separately as the functions to optimize. The solution of the learning approach is a Pareto set of non-dominated multi-dimensional Bayesian network classifiers and their accuracies for the different class variables, so a decision maker can easily choose by hand the classifier that best suits the particular problem and domain.