"Tin Kam Ho"

Slide 2

Automatic Pattern Recognition

Automatic Pattern Recognition

Slide 5

Slide 6

Many classifiers are in close rivalry with each other.  Why?

Sources of Difficulty
in Classification

Class Ambiguity

Boundary Complexity

Classification Boundaries As Decided by Different Classfiers

Slide 12

Match between Classifiers and Problems

Measures of Geometrical Complexity
of Classification Problems

Geometry of Datasets and Classifiers

Measures of Geometric Complexity

Measures of Geometrical Complexity

Experiments with Controlled Data Sets

Patterns in Complexity Measure Space
lin.sep¯   lin.nonsep+   random

Slide 20

Loadings of the First 6 Prin. Comp.

Interpretation of the First 4 Prin. Comp.

Slide 23

Slide 24

Slide 25

Domains of Competence of Classifiers

Domain of Competence Experiment

Classifier Domains of Competence

Best Classifier Being
nn,lp,odt vs an ensemble technique

Other Studies on Data Complexity

Extension to Multiple Classes

Global vs. Local Properties

Intrinsic Ambiguity

Tackling Intrinsic Ambiguity

Sampling Density

Real Problems Have a Mixture of Difficulties

"To Conclude:"