Suppose that out of a large set, S, of attributes that describe the target variable M, there exists a subset Z of S that optimize the performance for say a decision tree model, does that mean that Z will optimize performance for all other techniques, such as K-NN , or Bayes, or SVM?