Tang Mao Pseudo code

tangmao.py

INPUT: 
  U # underlying set of features
  EXPECTED_FEATURE_SET_SIZE # how many features should be selected 

OUTPUT: 
  S # selected set of features

while(not is_good_enough(S)):
  Ux, Uz = split_into_continous_nominal_features(U)
  
  # select the best continous feature
  X´ = None
  # if the measure needs to be maximized or minimized depends on the evaluation measure. 
  # lets asume a lower value indicates a better feature set
  best_J = +Inf
  
  measure = if(S contains nominal feature): Jm else: Jc 
  
  for ux in Ux:
    j = calculate_subset_evaluation_measure(ux, measure)
    if j < best_J:
      best_J = j
      X´ = ux
  
  
  # select the best nominal feature
  Z´ = None
  # if the measure needs to be maximized or minimized depends on the evaluation measure. 
  # lets asume a lower value indicates a better feature set
  best_J = +Inf
  
  measure = if(S contains continous feature): Jm else: Jn 
  
  for uz in Uz:
    j = calculate_subset_evaluation_measure(uz, measure)
    if j < best_J:
      best_J = j
      Z´ = uz
      
  classifier_performance_ux = train_classifier_on(S + X´).error
  
  classifier_performance_uz = train_classifier_on(S + Z´).error
  
  if classifier_performance_ux > classifier_performance_uz:
    U = U \ Z´
    S = S + Z´
  else:
    U = U \ X´
    S = S + X´
    
return S


# example for an abort condition
def is_good_enough(S):
  return S.size >= EXPECTED_FEATURE_SET_SIZE