Source code

Revision control

Copy as Markdown

Other Tools

# mypy: allow-untyped-defs
from math import log
from collections import defaultdict
class Node:
def __init__(self, prop, value):
self.prop = prop
self.value = value
self.parent = None
self.children = set()
# Populated for leaf nodes
self.run_info = set()
self.result_values = defaultdict(int)
def add(self, node):
node.parent = self
def __iter__(self):
yield self
for node in self.children:
yield from node
def __len__(self):
return 1 + sum(len(item) for item in self.children)
def entropy(results):
"""This is basically a measure of the uniformity of the values in results
based on the shannon entropy"""
result_counts = defaultdict(int)
total = float(len(results))
for values in results.values():
# Not sure this is right, possibly want to treat multiple values as
# distinct from multiple of the same value?
for value in values:
result_counts[value] += 1
entropy_sum = 0
for count in result_counts.values():
prop = float(count) / total
entropy_sum -= prop * log(prop, 2)
return entropy_sum
def split_results(prop, results):
"""Split a dictionary of results into a dictionary of dictionaries where
each sub-dictionary has a specific value of the given property"""
by_prop = defaultdict(dict)
for run_info, value in results.items():
by_prop[run_info[prop]][run_info] = value
return by_prop
def build_tree(properties, dependent_props, results, tree=None):
"""Build a decision tree mapping properties to results
:param properties: - A list of run_info properties to consider
in the tree
:param dependent_props: - A dictionary mapping property name
to properties that should only be considered
after the properties in the key. For example
{"os": ["version"]} means that "version" won't
be used until after os.
:param results: Dictionary mapping run_info to set of results
:tree: A Node object to use as the root of the (sub)tree"""
if tree is None:
tree = Node(None, None)
prop_index = {prop: i for i, prop in enumerate(properties)}
all_results = defaultdict(int)
for result_values in results.values():
for result_value, count in result_values.items():
all_results[result_value] += count
# If there is only one result we are done
if not properties or len(all_results) == 1:
for value, count in all_results.items():
tree.result_values[value] += count
tree.run_info |= set(results.keys())
return tree
results_partitions = []
remove_properties = set()
for prop in properties:
result_sets = split_results(prop, results)
if len(result_sets) == 1:
# If this property doesn't partition the space then just remove it
# from the set to consider
new_entropy = 0.
results_sets_entropy = []
for prop_value, result_set in result_sets.items():
results_sets_entropy.append((entropy(result_set), prop_value, result_set))
new_entropy += (float(len(result_set)) / len(results)) * results_sets_entropy[-1][0]
# In the case that no properties partition the space
if not results_partitions:
for value, count in all_results.items():
tree.result_values[value] += count
tree.run_info |= set(results.keys())
return tree
# split by the property with the highest entropy
results_partitions.sort(key=lambda x: (x[0], prop_index[x[1]]))
_, best_prop, sub_results = results_partitions[0]
# Create a new set of properties that can be used
new_props = properties[:prop_index[best_prop]] + properties[prop_index[best_prop] + 1:]
new_props.extend(dependent_props.get(best_prop, []))
if remove_properties:
new_props = [item for item in new_props if item not in remove_properties]
for _, prop_value, results_sets in sub_results:
node = Node(best_prop, prop_value)
build_tree(new_props, dependent_props, results_sets, node)
return tree