def build_tree(qs, field, scoref=entropy, ignore_fields=None, include_fields=None):
"""Build a classification decision tree
>>> from pug.nlp.data.tobes_example import tobes_data
>>> print_tree(build_tree(tobes_data)) # doctest: +NORMALIZE_WHITESPACE
0:google?
T-> 3:21?
T-> {'Premium': 3}
F-> 2:yes?
T-> {'Basic': 1}
F-> {'None': 1}
F-> 0:slashdot?
T-> {'None': 3}
F-> 2:yes?
T-> {'Basic': 4}
F-> 3:21?
T-> {'Basic': 1}
F-> {'None': 3}
"""
if ignore_fields is None:
ignore_fields = ('pk', 'id')
N = qs.count()
if not N:
return DecisionNode()
if include_fields is None:
include_fields = qs[0]._meta.get_all_field_names()
current_score=scoref(qs, field)
# Set up some variables to track the best criteria
best_gain=0.0
best_criteria=None
best_sets=None
for col in include_fields:
if col in ignore_fields or col == field:
continue
# Set of unique values in this column
# TODO: should do this once for all columns and cache it somewhere
column_values = count_unique(qs, col)
# Try dividing the table up for each value in this column
for value in column_values:
(set1, set2) = divide(qs, field=col, target=value, ignore_fields=ignore_fields, include_fields=include_fields)
# Information improvement
p = float(set1.count()) / N
gain = current_score - p * scoref(set1, field) - (1 - p) * scoref(set2, field)
if gain > best_gain and set1.count() > 0 and set2.count() > 0:
best_gain = gain
best_criteria = (col, value)
best_sets = (set1, set2)
# Create the sub branches
if best_gain > 0:
trueBranch = build_tree(best_sets[0], field, ignore_fields=ignore_fields, include_fields=include_fields)
falseBranch = build_tree(best_sets[1], field, ignore_fields=ignore_fields, include_fields=include_fields)
return DecisionNode(col=best_criteria[0], value=best_criteria[1],
tb=trueBranch, fb=falseBranch)
else:
return DecisionNode(results=count_unique(qs, field=field))
def build_tree(qs, field, scoref=entropy, ignore_fields=None, include_fields=None):
"""Build a classification decision tree
>>> print_tree(build_tree(tobes_data)) # doctest: +NORMALIZE_WHITESPACE
0:google?
T-> 3:21?
T-> {'Premium': 3}
F-> 2:yes?
T-> {'Basic': 1}
F-> {'None': 1}
F-> 0:slashdot?
T-> {'None': 3}
F-> 2:yes?
T-> {'Basic': 4}
F-> 3:21?
T-> {'Basic': 1}
F-> {'None': 3}
"""
if ignore_fields is None:
ignore_fields = ('pk', 'id')
N = qs.count()
if not N:
return DecisionNode()
if include_fields is None:
include_fields = qs[0]._meta.get_all_field_names()
current_score=scoref(qs, field)
# Set up some variables to track the best criteria
best_gain=0.0
best_criteria=None
best_sets=None
for col in include_fields:
if col in ignore_fields or col == field:
continue
# Set of unique values in this column
# TODO: should do this once for all columns and cache it somewhere
column_values = count_unique(qs, col)
# Try dividing the table up for each value in this column
for value in column_values:
(set1, set2) = divide(qs, field=col, target=value, ignore_fields=ignore_fields, include_fields=include_fields)
# Information improvement
p = float(set1.count()) / N
gain = current_score - p * scoref(set1, field) - (1 - p) * scoref(set2, field)
if gain > best_gain and set1.count() > 0 and set2.count() > 0:
best_gain = gain
best_criteria = (col, value)
best_sets = (set1, set2)
# Create the sub branches
if best_gain > 0:
trueBranch = build_tree(best_sets[0], field, ignore_fields=ignore_fields, include_fields=include_fields)
falseBranch = build_tree(best_sets[1], field, ignore_fields=ignore_fields, include_fields=include_fields)
return DecisionNode(col=best_criteria[0], value=best_criteria[1],
tb=trueBranch, fb=falseBranch)
else:
return DecisionNode(results=count_unique(qs, field=field))
def entropy_and_impurity(qs, field, num_categories=2):
"""Gini impurity evaluation of predicted segmentation/categorization
Returns a tuple of the entropy (in nats, base e bits) and the impurity ( a probability between 0 and 1 inclusive)
Impurity is the probability or frequency with which the *wrong* category or prediction is assigned to an element.
>>> from pug.nlp.data.tobes_example import tobes_data
>>> entropy_and_impurity(tobes_data, -1) # doctest: +ELLIPSIS
(1.50524..., 0.6328125)
"""
from math import log
N = qs.count()
counts = count_unique(qs, field)
impurity = 0.0
entropy = 0.0
for k1 in counts:
p1 = float(counts[k1]) / N
if p1:
entropy -= p1 * log(p1, num_categories)
for k2 in counts:
if not k1 == k2:
p2 = float(counts[k2]) / N
impurity += p1 * p2
return entropy, impurity
def gini_impurity(qs, field):
'''Gini impurity evaluation of set of values
Returns the probability [0, 1], that the wrong category/prediction has been assigned.
'''
N = qs.count()
counts = count_unique(qs, field)
impurity = 0.0
for k1 in counts:
p1 = float(counts[k1]) / N
for k2 in counts:
if not k1 == k2:
p2 = float(counts[k2]) / N
impurity += p1 * p2
return impurity
def gini_impurity(qs, field):
'''Gini impurity evaluation of set of values
Returns the probability [0, 1], that the wrong category/prediction has been assigned.
'''
N = qs.count()
counts = count_unique(qs, field)
impurity = 0.0
for k1 in counts:
p1 = float(counts[k1]) / N
for k2 in counts:
if not k1 == k2:
p2 = float(counts[k2]) / N
impurity += p1 * p2
return impurity
def entropy(qs, field, num_categories=2):
"""Total entropy (in nats, base e bits) for all the categorizations assigned
sum(p(x) * log(p(x)) for x in count_unique(qs, field)
Which measures how different each categorization (segmentation) is from the others
"""
from math import log
counts = count_unique(qs, field)
ans = 0.0
N = qs.count()
for k in counts:
p = float(counts[k]) / N
if p:
ans -= p * log(p, num_categories)
return ans
def entropy(qs, field, num_categories=2):
"""Total entropy for all the categorizations assigned
sum(p(x) * log(p(x)) for x in count_unique(qs, field)
Which measures how different each categorization is from the others
"""
from math import log
counts = count_unique(qs, field)
ans = 0.0
N = qs.count()
for k in counts:
p = float(counts[k]) / N
if p:
ans -= p * log(p, num_categories)
return ans
def entropy_and_impurity(qs, field, num_categories=2):
"""Gini impurity evaluation of predictions
Returns the probability [0, 1], that the wrong category/prediction has been assigned.
>>> entropy_and_impurity(tobes_data, -1) # doctest: +ELLIPSIS
(1.50524..., 0.6328125)
"""
from math import log
N = qs.count()
counts = count_unique(qs, field)
impurity = 0.0
entropy = 0.0
for k1 in counts:
p1 = float(counts[k1]) / N
if p1:
entropy -= p1 * log(p1, num_categories)
for k2 in counts:
if not k1 == k2:
p2 = float(counts[k2]) / N
impurity += p1 * p2
return entropy, impurity
def prune(tree, mingain):
# If the branches aren't leaves, then prune them
if tree.tb.results == None:
prune(tree.tb, mingain)
if tree.fb.results == None:
prune(tree.fb, mingain)
# If both the subbranches are now leaves, see if they
# should merged
if tree.tb.results != None and tree.fb.results != None:
# Build a combined dataset
tb, fb = [],[]
for v, c in tree.tb.results.items():
tb += [[v]] * c
for v, c in tree.fb.results.items():
fb += [[v]] * c
# Test the reduction in entropy
delta = dt.entropy(tb + fb) - (dt.entropy(tb) + dt.entropy(fb) / 2)
if delta < mingain:
# Merge the branches
tree.tb, tree.fb = None, None
tree.results = count_unique(tb + fb)
def prune(tree, mingain):
# If the branches aren't leaves, then prune them
if tree.tb.results == None:
prune(tree.tb, mingain)
if tree.fb.results == None:
prune(tree.fb, mingain)
# If both the subbranches are now leaves, see if they
# should merged
if tree.tb.results != None and tree.fb.results != None:
# Build a combined dataset
tb, fb = [], []
for v, c in tree.tb.results.items():
tb += [[v]] * c
for v, c in tree.fb.results.items():
fb += [[v]] * c
# Test the reduction in entropy
delta = dt.entropy(tb + fb) - (dt.entropy(tb) + dt.entropy(fb) / 2)
if delta < mingain:
# Merge the branches
tree.tb, tree.fb = None, None
tree.results = count_unique(tb + fb)
from call_center.models import CaseExchange, CaseHDTVHeader, CaseMaster
from pug.db.explore import count_unique
from pug.nlp.db_decision_tree import build_tree, print_tree
N = CaseMaster.objects.count()
UN = CaseMaster.objects.values('case_number').distinct().count()
N_ce = CaseExchange.objects.count()
UN_ce = CaseExchange.objects.values('case_number').distinct().count()
N_hdtv = CaseHDTVHeader.objects.count()
UN_hdtv = CaseHDTVHeader.objects.values('case_number').distinct().count()
un = count_unique(CaseExchange.objects.values('case_number'), 'case_number')
assert(len(un.values()) == UN_ce)
assert(sum(un.values()) == N_ce)
qs = CaseHDTVHeader.objects.filter(case_number__lt=4000000)
ex = qs.all()[0]
ex.service_calls
print_tree(build_tree(qs, field='service_calls', ignore_fields=('id', 'case_number')))
# dispatch_status:Completed ?
# T-> date_time:2008-09-15 12:25:34.270000?
# T-> {1: 1}
# F-> date_time:2008-07-09 08:49:36.437000?
# T-> {0: 1}
# F-> {None: 0}
# F-> {None: 0}
#test_decider.py
from call_center.models import CaseExchange, CaseHDTVHeader, CaseMaster
from pug.db.explore import count_unique
from pug.nlp.db_decision_tree import build_tree, print_tree
N = CaseMaster.objects.count()
UN = CaseMaster.objects.values('case_number').distinct().count()
N_ce = CaseExchange.objects.count()
UN_ce = CaseExchange.objects.values('case_number').distinct().count()
N_hdtv = CaseHDTVHeader.objects.count()
UN_hdtv = CaseHDTVHeader.objects.values('case_number').distinct().count()
un = count_unique(CaseExchange.objects.values('case_number'), 'case_number')
assert (len(un.values()) == UN_ce)
assert (sum(un.values()) == N_ce)
qs = CaseHDTVHeader.objects.filter(case_number__lt=4000000)
ex = qs.all()[0]
ex.service_calls
print_tree(
build_tree(qs, field='service_calls', ignore_fields=('id', 'case_number')))
# dispatch_status:Completed ?
# T-> date_time:2008-09-15 12:25:34.270000?
# T-> {1: 1}
# F-> date_time:2008-07-09 08:49:36.437000?
# T-> {0: 1}
# F-> {None: 0}
# F-> {None: 0}
qs = CaseHDTVHeader.objects.filter(case_number__lt=2000000)
