Model Definition

1
The Chinese University of Hong Kong
Learning Classifiers from Imbalanced Data Based
on Biased Minimax Probability Machine
Kaizhu Huang, Haiqin Yang, Irwin King, and
Michael R. Lyu Department of Computer Science and
Engineering The Chinese University of Hong Kong,
Shatin, N.T., Hong Kong SAR kzhuang,hqyang,
king, lyu_at_cse.cuhk.edu.hk
1
2
Introduction
Contributions

• The problem of the binary classification on
  imbalanced data, where
• nearly all the instances are labeled as one class
  x, while far fewer instances are labeled as the
  other class y, usually the more important class.
• Traditional imbalanced learning methods
• Common learning methods seeking an accurate
  performance over a full range of instances
  tending to classify all the data into the
  majority, usually the less important class.
• current methods utilizing some intermediate
  factors, 1) the distribution of the training set,
  2) the decision thresholds 3) the cost matrices
  being not rigorous and not systematic

• Proposed a novel model named Biased Minimax
  Probability Machine for imbalanced
  classification.
• A rigorous model for imposing a bias on the
  important data.
• Directly controlling the worst-case real accuracy
  of classification of the future data to build up
  biased classifiers
• The comparison with the Naive Bayesian
  classifier, the k-Nearest Neighbor method, and
  the decision tree method C4.5, demonstrates the
  superiority of our novel model

.
3
4
Model Definition
Optimization Methods
Transforming a Concave-Convex Fractional
Programming problem
the lower bound of the accuracy for the
classification of future x data the lower
bound of the accuracy for the classification of
future y data. the acceptable level of the
less important class y If , z is the x
class otherwise it belongs to the y class.
We should maximize the accuracy of the
important class while maintaining the accuracy of
the less important class acceptable.
Every local optimum is the global optimum
5
Model Illustration
6
Experimental Results
IEEE Computer Vision and Pattern Recognition 2004
Dept. of C.S. E., The Chinese University of
Hong Kong