Title :
Structure relation between classes for supervised learning using pretopology
Author :
Lebourgeois, F. ; Bouayad, M. ; Emptoz, H.
Author_Institution :
Lab. de Reconnaissance de Formes et Vision, Inst. Nat. des Sci. Appliquees, Villeurbanne, France
Abstract :
The article presents a pretopological approach for class proximity evaluation according to cluster boundaries. This is a generic method independent of the classifier used, resulting in a neighborhood structure between objects for each class. This proximity measure of classes also gives an objective quality measure of a given training set and an evaluation of class separability. The theoretical context of the method is presented, followed by an application featuring evaluation for printed character recognition and an example of prototype selection for an improved training set selection
Keywords :
character recognition; learning (artificial intelligence); pattern classification; topology; class proximity evaluation; class separability; classifier; cluster boundaries; generic method; neighborhood structure; objective quality measure; pretopological approach; printed character recognition; prototype selection; proximity measure; structure relation; supervised learning; training set; training set selection; Data mining; Density measurement; Electrical capacitance tomography; Pattern recognition; Probability; Prototypes; Reconnaissance; Statistical analysis; Supervised learning; Testing;
Conference_Titel :
Document Analysis and Recognition, 1999. ICDAR '99. Proceedings of the Fifth International Conference on
Conference_Location :
Bangalore
Print_ISBN :
0-7695-0318-7
DOI :
10.1109/ICDAR.1999.791718
