Back

Session TC5 - Classification / Clustering

Day Tuesday, May 05, 2009
Room Ordre des CGA
President Sylvain Perron

Presentations

03h30 PM-
03h55 PM
Parallel Hyperplanes Separation Method for the Two-Groups Discrimination Problem
  Anthony Guillou, GERAD, HEC Montréal, Montréal, Québec, Canada
Pierre Hansen, HEC Montréal, GERAD et Méthodes quantitatives de gestion, 3000 Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7
Sylvain Perron, GERAD, HEC Montréal, Méthodes quantitatives de gestion, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7

We consider the problem of separating two non linearly separable sets of points A and B in the euclidean space, with two parallel hyperplanes that are respectively the boundaries of half-spaces containing A and B. Both heuristic and exact method are proposed to minimise the euclidean distance between such two hyperplanes.


03h55 PM-
04h20 PM
Optimal Clusterwise Multiple Linear Regression
  Réal Carbonneau, GERAD et HEC Montréal, Méthodes quantitatives de gestion
Gilles Caporossi, GERAD, HEC Montréal, Méthodes quantitatives de gestion, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7
Pierre Hansen, HEC Montréal, GERAD et Méthodes quantitatives de gestion, 3000 Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7

Clusterwise regression has been used as a data mining tool, however, there is no published research on identifying optimal solutions. In this research, both a quadratic programming formulation and a branch and bound algorithm are compared for identifying Optimal Clusterwise Multiple Linear Regression (OCMLR) solutions. The processing time increases with the number of clusters, observations, dimensions and amount of noise.


04h20 PM-
04h45 PM
A Column Generation Algorithm for 2 Groups Discrimination
  Gilles Caporossi, GERAD, HEC Montréal, Méthodes quantitatives de gestion, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7
Sylvain Perron, GERAD, HEC Montréal, Méthodes quantitatives de gestion, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Québec, Canada, H3T 2A7

A method of discrimination among two groups of data is proposed. This method involves a combination of linear classifiers, each of which aims at minimizing the number of (weighted) misclassified observations. The combination of these classifiers is achieved by weighted vote. The learning algorithm uses column generation and could be viewed as a mathematical programming approach for boosting.


Back