def train(self, samples, maxRuleSize=9999, mineOnlyClass=None):
"""Train with CPAR on the sample set, returning an AssociationRuleSet."""
self.MAX_RULE_SIZE = maxRuleSize
self.objPNData = PNData(samples)
self.lstRules = []
classes = self.objPNData.getClassList()
log("Dataset has %d classes over %d samples." %
(len(classes), len(samples)))
for current_class in classes:
if mineOnlyClass != None:
if current_class != mineOnlyClass:
continue
log("Processing class %s" % (current_class))
self.objPNData.setCurrentClass(current_class)
dblMinTotalWeight = self.dblTotalWeightFactor * self.objPNData.getTotalWeight(
)
lstAntecedent = []
while self.objPNData.getTotalWeight() > dblMinTotalWeight:
self.objPNData.refreshPNAData()
if self.objPNData.noValidGainsinPNarray(
self.dblMinGainThreshold):
#log("NO VALID GAINS....Breaking!");
break
#log('BEGIN DEPTH FIRST SEARCH - total weight %f > %f'%(self.objPNData.getTotalWeight(),dblMinTotalWeight))
self._CPARdfs(self.objPNData.copyPrimes(), lstAntecedent,
[current_class])
trules = len(self.lstRules)
self.removeDuplicateRules()
#log("End of rule search. Found %d rules total, %d after duplicates removed."%(trules,len(self.lstRules)))
arset = AssociationRuleSet()
arset.extend(self.lstRules)
arset.set_target_accuracy("laplace")
return self.remap_index_to_feature(arset, samples)
def train(self,samples,maxRuleSize=9999,mineOnlyClass=None):
"""Train with CPAR on the sample set, returning an AssociationRuleSet."""
self.MAX_RULE_SIZE = maxRuleSize
self.objPNData = PNData(samples)
self.lstRules = []
classes = self.objPNData.getClassList()
log("Dataset has %d classes over %d samples."%(len(classes),len(samples)))
for current_class in classes:
if mineOnlyClass != None:
if current_class != mineOnlyClass:
continue
log("Processing class %s"%(current_class))
self.objPNData.setCurrentClass(current_class)
dblMinTotalWeight = self.dblTotalWeightFactor * self.objPNData.getTotalWeight()
lstAntecedent = []
while self.objPNData.getTotalWeight() > dblMinTotalWeight:
self.objPNData.refreshPNAData()
if self.objPNData.noValidGainsinPNarray(self.dblMinGainThreshold):
#log("NO VALID GAINS....Breaking!");
break
#log('BEGIN DEPTH FIRST SEARCH - total weight %f > %f'%(self.objPNData.getTotalWeight(),dblMinTotalWeight))
self._CPARdfs(self.objPNData.copyPrimes(),lstAntecedent,[current_class])
trules = len(self.lstRules)
self.removeDuplicateRules()
#log("End of rule search. Found %d rules total, %d after duplicates removed."%(trules,len(self.lstRules)))
arset = AssociationRuleSet()
arset.extend(self.lstRules)
arset.set_target_accuracy("laplace")
return self.remap_index_to_feature(arset,samples)
class CPARTrainer(BaseTrainer):
"""Main class for the CPAR implementation"""
def __init__(self, parameters=None):
self.objPNData = None
self.lstRules = []
self.dblGainSimilarityRatio = 0.99
self.dblTotalWeightFactor = 0.05
self.dblDecayFactor = 2.0 / 3.0
self.dblMinGainThreshold = 0.7
self.intKValue = 5
self.MAX_RULE_SIZE = 9999
def remap_index_to_feature(self, association_rule_set, sample_set):
return association_rule_set.remap_index_to_feature(sample_set)
def remap_feature_to_index(self, association_rule_set, sample_set):
return association_rule_set.remap_feature_to_index(sample_set)
def train(self, samples, maxRuleSize=9999, mineOnlyClass=None):
"""Train with CPAR on the sample set, returning an AssociationRuleSet."""
self.MAX_RULE_SIZE = maxRuleSize
self.objPNData = PNData(samples)
self.lstRules = []
classes = self.objPNData.getClassList()
log("Dataset has %d classes over %d samples." %
(len(classes), len(samples)))
for current_class in classes:
if mineOnlyClass != None:
if current_class != mineOnlyClass:
continue
log("Processing class %s" % (current_class))
self.objPNData.setCurrentClass(current_class)
dblMinTotalWeight = self.dblTotalWeightFactor * self.objPNData.getTotalWeight(
)
lstAntecedent = []
while self.objPNData.getTotalWeight() > dblMinTotalWeight:
self.objPNData.refreshPNAData()
if self.objPNData.noValidGainsinPNarray(
self.dblMinGainThreshold):
#log("NO VALID GAINS....Breaking!");
break
#log('BEGIN DEPTH FIRST SEARCH - total weight %f > %f'%(self.objPNData.getTotalWeight(),dblMinTotalWeight))
self._CPARdfs(self.objPNData.copyPrimes(), lstAntecedent,
[current_class])
trules = len(self.lstRules)
self.removeDuplicateRules()
#log("End of rule search. Found %d rules total, %d after duplicates removed."%(trules,len(self.lstRules)))
arset = AssociationRuleSet()
arset.extend(self.lstRules)
arset.set_target_accuracy("laplace")
return self.remap_index_to_feature(arset, samples)
def _CPARdfs(self, objPNDatacopy, lstAntecedent, lstConsequent):
"""Depth first search for constructing a new rule."""
blnMaxExceeded = False
if (len(lstAntecedent) == self.MAX_RULE_SIZE):
blnMaxExceeded = True
'GET ALL ATTRIBUTES WITHIN 1.0 - dblGainSimilarityRatio of max'
objPNDatacopy.recalculateGains()
[dblGain, lstAttributes
] = objPNDatacopy.getBestGainAttributes(self.dblGainSimilarityRatio)
if dblGain > self.dblMinGainThreshold and not blnMaxExceeded:
for intAttribute in lstAttributes:
"""
This is a check to see if this attribute still within 1%...it may not
after the depth first search
"""
lstAntecedentCopy = deepcopy(lstAntecedent)
lstAntecedentCopy.append(intAttribute)
objPNDatacopyTemp = objPNDatacopy.copyPrimes()
objPNDatacopyTemp.lstAprime[intAttribute] = 0
objPNDatacopyTemp.removeExamplesNotSatisfying(
lstAntecedentCopy)
self._CPARdfs(objPNDatacopyTemp, lstAntecedentCopy,
lstConsequent)
else:
if len(lstAntecedent) > 0:
lstAntecedent.sort()
dblLaPlaceAccuracy = self.objPNData.getLaPlaceAccuracy(
lstAntecedent, lstConsequent)
objRule = AssociationRule(lstAntecedent, lstConsequent,
{"laplace": dblLaPlaceAccuracy})
objPNDatacopy.updateWeights(objRule, self.dblDecayFactor)
self.addRule(objRule)
else:
log("Empty antecedent.")
pass
def addRule(self, objRule):
"""Add the rule to the AssociationRuleSet."""
self.lstRules.append(objRule)
#log("FOUND RULE: %s\n"%(str(objRule)))
def removeDuplicateRules(self, objRule=None, startindex=0):
"""Ensure that there are no duplicate rules."""
if objRule == None:
i = 0
while i < len(self.lstRules):
r = self.lstRules[i]
self.removeDuplicateRules(r, i + 1)
i += 1
else:
i = startindex
while i < len(self.lstRules):
r = self.lstRules[i]
if objRule.ls == r.ls and objRule.rs == r.rs:
del self.lstRules[i]
i -= 1
i += 1
class CPARTrainer(BaseTrainer):
"""Main class for the CPAR implementation"""
def __init__(self,parameters=None):
self.objPNData = None
self.lstRules = []
self.dblGainSimilarityRatio = 0.99
self.dblTotalWeightFactor = 0.05
self.dblDecayFactor = 2.0/3.0
self.dblMinGainThreshold = 0.7
self.intKValue = 5
self.MAX_RULE_SIZE = 9999
def remap_index_to_feature(self,association_rule_set,sample_set):
return association_rule_set.remap_index_to_feature(sample_set)
def remap_feature_to_index(self,association_rule_set,sample_set):
return association_rule_set.remap_feature_to_index(sample_set)
def train(self,samples,maxRuleSize=9999,mineOnlyClass=None):
"""Train with CPAR on the sample set, returning an AssociationRuleSet."""
self.MAX_RULE_SIZE = maxRuleSize
self.objPNData = PNData(samples)
self.lstRules = []
classes = self.objPNData.getClassList()
log("Dataset has %d classes over %d samples."%(len(classes),len(samples)))
for current_class in classes:
if mineOnlyClass != None:
if current_class != mineOnlyClass:
continue
log("Processing class %s"%(current_class))
self.objPNData.setCurrentClass(current_class)
dblMinTotalWeight = self.dblTotalWeightFactor * self.objPNData.getTotalWeight()
lstAntecedent = []
while self.objPNData.getTotalWeight() > dblMinTotalWeight:
self.objPNData.refreshPNAData()
if self.objPNData.noValidGainsinPNarray(self.dblMinGainThreshold):
#log("NO VALID GAINS....Breaking!");
break
#log('BEGIN DEPTH FIRST SEARCH - total weight %f > %f'%(self.objPNData.getTotalWeight(),dblMinTotalWeight))
self._CPARdfs(self.objPNData.copyPrimes(),lstAntecedent,[current_class])
trules = len(self.lstRules)
self.removeDuplicateRules()
#log("End of rule search. Found %d rules total, %d after duplicates removed."%(trules,len(self.lstRules)))
arset = AssociationRuleSet()
arset.extend(self.lstRules)
arset.set_target_accuracy("laplace")
return self.remap_index_to_feature(arset,samples)
def _CPARdfs(self,objPNDatacopy,lstAntecedent,lstConsequent):
"""Depth first search for constructing a new rule."""
blnMaxExceeded = False
if (len(lstAntecedent) == self.MAX_RULE_SIZE):
blnMaxExceeded = True
'GET ALL ATTRIBUTES WITHIN 1.0 - dblGainSimilarityRatio of max'
objPNDatacopy.recalculateGains()
[dblGain,lstAttributes] = objPNDatacopy.getBestGainAttributes(self.dblGainSimilarityRatio)
if dblGain > self.dblMinGainThreshold and not blnMaxExceeded:
for intAttribute in lstAttributes:
"""
This is a check to see if this attribute still within 1%...it may not
after the depth first search
"""
lstAntecedentCopy = deepcopy(lstAntecedent)
lstAntecedentCopy.append(intAttribute)
objPNDatacopyTemp = objPNDatacopy.copyPrimes()
objPNDatacopyTemp.lstAprime[intAttribute] = 0
objPNDatacopyTemp.removeExamplesNotSatisfying(lstAntecedentCopy)
self._CPARdfs(objPNDatacopyTemp,lstAntecedentCopy,lstConsequent)
else:
if len(lstAntecedent)>0:
lstAntecedent.sort()
dblLaPlaceAccuracy = self.objPNData.getLaPlaceAccuracy(lstAntecedent,lstConsequent)
objRule = AssociationRule(lstAntecedent, lstConsequent, {"laplace":dblLaPlaceAccuracy})
objPNDatacopy.updateWeights(objRule,self.dblDecayFactor)
self.addRule(objRule)
else:
log("Empty antecedent.")
pass
def addRule(self,objRule):
"""Add the rule to the AssociationRuleSet."""
self.lstRules.append(objRule)
#log("FOUND RULE: %s\n"%(str(objRule)))
def removeDuplicateRules(self,objRule=None,startindex=0):
"""Ensure that there are no duplicate rules."""
if objRule==None:
i = 0
while i < len(self.lstRules):
r = self.lstRules[i]
self.removeDuplicateRules(r,i+1)
i+=1
else:
i = startindex
while i < len(self.lstRules):
r = self.lstRules[i]
if objRule.ls == r.ls and objRule.rs == r.rs:
del self.lstRules[i]
i-=1
i+=1
