def learn_tree(self, input_features, data_subset):
"""returns a decision tree
for input_features is a set of possible conditions
data_subset is a subset of the data used to build this (sub)tree
where a decision tree is a function that takes an example and
makes a prediction on the target feature
"""
if (input_features and len(data_subset) >= self.min_number_examples):
first_target_val = self.target(data_subset[0])
allagree = all(self.target(inst)==first_target_val for inst in data_subset)
if not allagree:
split, partn = self.select_split(input_features, data_subset)
if split: # the split succeeded in splitting the data
false_examples, true_examples = partn
rem_features = [fe for fe in input_features if fe != split]
self.display(2,"Splitting on",split.__doc__,"with examples split",
len(true_examples),":",len(false_examples))
true_tree = self.learn_tree(rem_features,true_examples)
false_tree = self.learn_tree(rem_features,false_examples)
def fun(e):
if split(e):
return true_tree(e)
else:
return false_tree(e)
#fun = lambda e: true_tree(e) if split(e) else false_tree(e)
fun.__doc__ = ("if "+split.__doc__+" then ("+true_tree.__doc__+
") else ("+false_tree.__doc__+")")
return fun
# don't expand the trees but return a point prediction
return point_prediction(self.target, data_subset, selection=self.leaf_selection)
def training_error(dataset, data_subset, to_optimize):
"""returns training error for dataset on to_optimize.
This assumes that we choose the best value for the optimization
criteria for dataset according to point_prediction
"""
select_dict = {"sum-of-squares":"mean", "sum_absolute":"median",
"logloss":"Laplace"} # arbitrary mapping. Perhaps wrong.
selection = select_dict[to_optimize]
predictor = point_prediction(dataset.target, data_subset, selection=selection)
error = sum(error_example(predictor(example),
dataset.target(example),
to_optimize)
for example in data_subset)
return error