def construct_new_rule(self, rule, idx, ddist, bdist, attr, used, negate):
if attr.var_type == Orange.feature.Type.Discrete:
keys = [k for k,v in ddist.items() if v > 0]
fanout = len(keys) / 5 if len(keys) / self.fanout > 5 else self.fanout
for keyblock in block_iter(keys, fanout):
new_rule = rule.cloneAndAddCondition(attr, keyblock, used=used, negate=negate)
new_rule.parent_rule = rule
if new_rule not in self.bad_rules:
yield new_rule
else:
avgv = Orange.data.Value(bdist.variable, bdist.avg)
minv = Orange.data.Value(bdist.variable, bdist.min) - 0.5
maxv = Orange.data.Value(bdist.variable, bdist.max) + 0.5
if used:
minv, maxv = None, None
for cond in rule.filter.conditions:
if cond.position == idx:
minv = max(cond.min, minv) if minv else cond.min
maxv = min(cond.max, maxv) if maxv else cond.max
block = (maxv - minv) / self.fanout
ranges = [(minv + i*block, minv + (i+1)*block) for i in xrange(self.fanout)]
for minv, maxv in ranges:
if minv == maxv: # edge case
continue
new_rule = rule.cloneAndAddContCondition(attr,
minv,
maxv,
orange.ValueFilter.Between,
used=used,
negate=negate)
new_rule.parent_rule = rule
if new_rule not in self.bad_rules:
yield new_rule
def construct_new_rule(self, rule, idx, ddist, bdist, attr, used, negate):
if attr.var_type == Orange.feature.Type.Discrete:
matches = filter(lambda c: idx == c.position,
rule.filter.conditions)
if matches:
keys = [attr.values[int(v)] for v in matches[0].values]
else:
keys = [k for k,v in ddist.items() if v > 0]
if len(keys) <= 1:
return
fanout = len(keys) / 5 if len(keys) / self.fanout > 5 else self.fanout
fanout = self.fanout
#fanout = 60
fanout = min(fanout, len(keys))#fanout)
#fanout = len(keys)
for keyblock in block_iter(keys, fanout):
new_rule = rule.cloneAndAddCondition(attr, keyblock, used=used, negate=negate)
new_rule.parent_rule = rule
if new_rule not in self.bad_rules:
yield new_rule
else:
minv = bdist.percentile(0)
maxv = bdist.percentile(100)
avgv = bdist.percentile(50)
#avgv = Orange.data.Value(bdist.variable, bdist.avg)
#minv = Orange.data.Value(bdist.variable, bdist.min)# - 0.5
#maxv = Orange.data.Value(bdist.variable, bdist.max)# + 0.5
if minv == maxv:
return
if used:
# This shouldn't be an issues because rule.examples
# should filetr the distribution correctly
minv, maxv = None, None
for cond in rule.filter.conditions:
if cond.position == idx:
minv = max(cond.min, minv) if minv else cond.min
maxv = min(cond.max, maxv) if maxv else cond.max
#block = (maxv - minv) / self.fanout
#ranges = [(minv + i*block, minv + (i+1)*block) for i in xrange(self.fanout)]
ranges = [[minv, avgv], [avgv, maxv]]
for minv, maxv in ranges:
if minv == maxv: # edge case
continue
new_rule = rule.cloneAndAddContCondition(attr,
minv,
maxv,
orange.ValueFilter.Between,
used=used,
negate=negate)
new_rule.parent_rule = rule
if new_rule in self.bad_rules:
continue
#if len(new_rule.examples) == len(rule.examples):
# continue
yield new_rule