SelfTraining

1
Self-Training Co-Training

• Huang Shujian

2
Self-Training

• Situation
• You are the only one in the world
• What can we do
• Just trust yourself!
• Using its own predictions to teach itself.
• Most confident ones are regarded as truths.

3
Self-Training (cont.)

• Procedure
• Training the classifier with labeled data
• Using the classifier to classify unlabeled data
• Adding the most confident unlabeled instances to
  the training set, together with their labels
• If no unlabeled data, exit Else, go to step1,
  re-training the classifier.
• Problems
• Mistakes may reinforce themselves ? Threshold

4
Co-Training Situation

• Situation changed
• Youve got a partner
• Identify you partner
• Same work
• Good worker
• Not cheating (working independently)
• Keep away from bias.
• Able to get along.
• Both get improved.

5
Co-Training Partner

• Find a partner
• Divide the attributes into two sets.
• Random division still helps.
• Use two different classifiers.
• Work with your partner
• Trust him completely
• Only use common opinions.
• The distribution of data the parameters

6
Co-Training Procedure

• Procedure
• Training the two classifiers independently with
  labeled data
• Using the two classifiers the classify unlabeled
  data
• Adding most confident unlabeled instances of one
  classifier into the others training set,
  together with their labels
• If no unlabeled data or some threshold is
  reached, exit Else, go to step1, re-training
  the two classifiers.

7
Co-Training Result

• Multiplying the two results to give the final
  result.
• P(c, x)P(c, x1) P(c, x2)
• The weighted vote of two results can generate a
  final result, with confidence as the weight.
• We hope partners helped each other. However, we
  never know what the truth is.
• Not strictly independent

8

• Thank you

Next