Attributes with Costs

Consider

• medical diagnosis, BloodTest has cost $150
• robotics, $Width\_from\_1ft$ has cost 23 sec.

How to learn a consistent tree with low expected cost?

One approach: replace gain by

• Tan and Schlimmer (1990)
  
  $$\frac{Gain^{2}(S,A)}{Cost(A)}.$$
  
  

• Nunez (1988)
  
  $$\frac{2^{Gain(S,A)} - 1}{(Cost(A) + 1)^{w}}$$
  
  
  where $w \in [0,1]$ determines importance of cost