def _expand_rule(self, rule):
"""
If the rule has enough statistics, possible expanding candidates are checked. If the best
candidate verifies the Hoeffding bound, a new predicate is add to the rule.
The rule statistics are update to fit the new description.
"""
if len(rule.observed_class_distribution) >= 2:
class_idx = None
if self.expand_criterion == _HELLINGER:
split_criterion = HellingerDistanceCriterion()
elif self.expand_criterion == _INFOGAIN:
split_criterion = InfoGainExpandCriterion()
else:
split_criterion = FoilGainExpandCriterion()
class_idx = rule.class_idx
should_expand = False
best_expand_suggestions = rule.get_best_expand_suggestion(split_criterion, class_idx)
best_expand_suggestions.sort(key=attrgetter('merit'))
if len(best_expand_suggestions) < 2:
should_expand = len(best_expand_suggestions) > 0
else:
hoeffding_bound = self.compute_hoeffding_bound(split_criterion.get_range_of_merit(
rule.observed_class_distribution), self.expand_confidence, rule.get_weight_seen())
best_suggestion = best_expand_suggestions[-1]
second_best_suggestion = best_expand_suggestions[-2]
if ((best_suggestion.merit - second_best_suggestion.merit) > hoeffding_bound) or \
(hoeffding_bound < self.tie_threshold):
should_expand = True
if self.remove_poor_atts is not None and self.remove_poor_atts:
poor_atts = set()
# Scan 1 - add any poor attribute to set
for i in range(len(best_expand_suggestions)):
if best_expand_suggestions[i] is not None:
split_atts = [best_expand_suggestions[i].att_idx]
if len(split_atts) == 1:
if best_suggestion.merit - best_expand_suggestions[i].merit > hoeffding_bound:
poor_atts.add(int(split_atts[0]))
# Scan 2 - remove good attributes from set
for i in range(len(best_expand_suggestions)):
if best_expand_suggestions[i] is not None:
split_atts = [best_expand_suggestions[i].att_idx]
if len(split_atts) == 1:
if best_suggestion.merit - best_expand_suggestions[i].merit < hoeffding_bound:
try:
poor_atts.remove(int(split_atts[0]))
except KeyError:
pass
for poor_att in poor_atts:
rule.disable_attribute(poor_att)
if should_expand:
best_suggestion = best_expand_suggestions[-1]
new_pred = Predicate(best_suggestion.att_idx, best_suggestion.operator, best_suggestion.att_val)
add_pred = True
for pred in rule.predicate_set:
if (pred.operator == new_pred.operator) and (pred.att_idx == new_pred.att_idx):
if pred.operator == "<=":
pred.value = min(pred.value, new_pred.value)
rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
elif pred.operator == ">":
pred.value = max(pred.value, new_pred.value)
rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
rule._attribute_observers = {}
add_pred = False
break
if add_pred:
rule.predicate_set.append(new_pred)
rule._attribute_observers = {}
rule.observed_class_distribution = {}
if new_pred.operator in ["=", "<="]:
rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
else:
rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
if self.expand_criterion == _FOILGAIN:
if not self.ordered_rules:
for c in rule.observed_class_distribution.keys():
if c != rule.class_idx:
new_rule = copy.deepcopy(rule)
new_rule.class_idx = c
split_criterion = FoilGainExpandCriterion()
should_expand = False
best_expand_suggestions = new_rule.get_best_expand_suggestion(split_criterion, c)
best_expand_suggestions.sort(key=attrgetter('merit'))
if len(best_expand_suggestions) < 2:
should_expand = len(best_expand_suggestions) > 0
else:
hoeffding_bound = self.compute_hoeffding_bound(
split_criterion.get_range_of_merit(
new_rule.observed_class_distribution), self.expand_confidence,
new_rule.get_weight_seen())
best_suggestion = best_expand_suggestions[-1]
second_best_suggestion = best_expand_suggestions[-2]
if ((best_suggestion.merit - second_best_suggestion.merit) > hoeffding_bound) \
or (hoeffding_bound < self.tie_threshold):
should_expand = True
if self.remove_poor_atts is not None and self.remove_poor_atts:
poor_atts = set()
# Scan 1 - add any poor attribute to set
for i in range(len(best_expand_suggestions)):
if best_expand_suggestions[i] is not None:
split_atts = [best_expand_suggestions[i].att_idx]
if len(split_atts) == 1:
if best_suggestion.merit - best_expand_suggestions[i].merit > \
hoeffding_bound:
poor_atts.add(int(split_atts[0]))
# Scan 2 - remove good attributes from set
for i in range(len(best_expand_suggestions)):
if best_expand_suggestions[i] is not None:
split_atts = [best_expand_suggestions[i].att_idx]
if len(split_atts) == 1:
if best_suggestion.merit - best_expand_suggestions[i].merit < \
hoeffding_bound:
try:
poor_atts.remove(int(split_atts[0]))
except KeyError:
pass
for poor_att in poor_atts:
new_rule.disable_attribute(poor_att)
if should_expand:
best_suggestion = best_expand_suggestions[-1]
new_pred = Predicate(best_suggestion.att_idx, best_suggestion.operator,
best_suggestion.att_val)
add_pred = True
for pred in new_rule.predicate_set:
if (pred.operator == new_pred.operator) and (
pred.att_idx == new_pred.att_idx):
if pred.operator == "<=":
pred.value = min(pred.value, new_pred.value)
new_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
elif pred.operator == ">":
pred.value = max(pred.value, new_pred.value)
new_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
new_rule._attribute_observers = {}
add_pred = False
break
if add_pred:
new_rule.predicate_set.append(new_pred)
new_rule._attribute_observers = {}
new_rule.observed_class_distribution = {}
if new_pred.operator in ["=", "<="]:
new_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
else:
new_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
self.rule_set.append(copy.deepcopy(new_rule))
def _create_rule(self):
""" Create a new rule from the default rule.
If the default rule has enough statistics, possible expanding candidates are checked.
If the best candidate verifies the Hoeffding bound, a new rule is created if a one predicate.
The rule statistics are passed down to the new rule and the default rule is reset.
"""
if len(self.default_rule.observed_class_distribution) >= 2:
if self.expand_criterion in [_INFOGAIN, _HELLINGER]:
if self.expand_criterion == _HELLINGER:
expand_criterion = HellingerDistanceCriterion()
else:
expand_criterion = InfoGainExpandCriterion()
should_expand = False
best_expand_suggestions = self.default_rule.get_best_expand_suggestion(expand_criterion, None)
best_expand_suggestions.sort(key=attrgetter('merit'))
if len(best_expand_suggestions) < 2:
should_expand = len(best_expand_suggestions) > 0
else:
hoeffding_bound = self.compute_hoeffding_bound(expand_criterion.get_range_of_merit(
self.default_rule.observed_class_distribution), self.expand_confidence,
self.default_rule.get_weight_seen())
best_suggestion = best_expand_suggestions[-1]
second_best_suggestion = best_expand_suggestions[-2]
if ((best_suggestion.merit - second_best_suggestion.merit) > hoeffding_bound) or \
(hoeffding_bound < self.tie_threshold):
should_expand = True
if should_expand:
best_suggestion = best_expand_suggestions[-1]
new_pred = Predicate(best_suggestion.att_idx, best_suggestion.operator, best_suggestion.att_val)
self.rule_set.append(self.new_rule(None, copy.deepcopy(self.drift_detector), None))
self.rule_set[-1].predicate_set.append(new_pred)
self.default_rule.restart()
if new_pred.operator in ["=", "<="]:
self.rule_set[-1].observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
self.default_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
else:
self.rule_set[-1].observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
self.default_rule.observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
else:
self.default_rule.weight_seen_at_last_expand = self.default_rule.get_weight_seen()
elif self.expand_criterion == _FOILGAIN:
expand_criterion = FoilGainExpandCriterion()
should_expand = False
for class_idx in self.default_rule.observed_class_distribution.keys():
best_expand_suggestions = self.default_rule.get_best_expand_suggestion(expand_criterion, class_idx)
best_expand_suggestions.sort(key=attrgetter('merit'))
if len(best_expand_suggestions) < 2:
should_expand = len(best_expand_suggestions) > 0
else:
hoeffding_bound = self.compute_hoeffding_bound(expand_criterion.get_range_of_merit(
self.default_rule.observed_class_distribution), self.expand_confidence,
self.default_rule.get_weight_seen())
best_suggestion = best_expand_suggestions[-1]
second_best_suggestion = best_expand_suggestions[-2]
if ((best_suggestion.merit - second_best_suggestion.merit) > hoeffding_bound) or (
hoeffding_bound < self.tie_threshold):
should_expand = True
if should_expand:
best_suggestion = best_expand_suggestions[-1]
new_pred = Predicate(best_suggestion.att_idx, best_suggestion.operator, best_suggestion.att_val)
self.rule_set.append(self.new_rule(None, copy.deepcopy(self.drift_detector), class_idx))
self.rule_set[-1].predicate_set.append(new_pred)
if new_pred.operator in ["=", "<="]:
self.rule_set[-1].observed_class_distribution = best_suggestion. \
resulting_stats_from_split(0).copy()
else:
self.rule_set[-1].observed_class_distribution = best_suggestion. \
resulting_stats_from_split(1).copy()
if should_expand:
self.default_rule.restart()
else:
self.default_rule.weight_seen_at_last_expand = self.default_rule.get_weight_seen()
else:
self.default_rule.weight_seen_at_last_expand = self.default_rule.get_weight_seen()