DATA-DRIVEN OPERATION


There are four important operations in constructing a decision tree:

• Question selection: choosing a set of questions to categorize your data (some algorithms can derive questions automatically).
  
  
• Splitting: partitioning data assigned to a node into N groups
  (N=2 for binary trees).
  
  
• Growing: expanding the tree to better represent the training data.
  
  
• Pruning: removing nodes to improve generalization.

In speech recognition, we operate on continuous-valued feature vectors, and use likelihood computations derived directly from HMM training. This is a major reason why decision trees are so popular in speech recognition systems - the implementation is very elegant.