def __resetData_definemergeCostReductions(self):
self.mergeCostReductions = SymmetricTable_Sparse(orderMatters=False)
clusters = self.clusters
numClusters = len(self.clusters)
for i in range(numClusters):
c1 = self.clusters[i]
for j in range(i):
c2 = self.clusters[j]
mergeCost = self.mergeCost_SinglePair(c1, c2)
self.mergeCostReductions.set(c1, c2, mergeCost)
def __resetData_defineClusterCostTable(self):
numClusters = len(self.clusters)
self.clusterCostTable = SymmetricTable_Sparse(orderMatters=False)
for i in range(numClusters):
c1 = self.clusters[i]
cnt_c1 = self.clusterCount_getClusterCount(c1)
for j in range(i, numClusters):
c2 = self.clusters[j]
cost = self.clusterCost_findSingleClusterPairCost(c1, c2, cnt_c1=cnt_c1)
self.clusterCostTable.set(c1, c2, cost)
def __resetData_defineClusterCounts(self):
# clusterSequence = [self.START_CLUSTER_SYMBOL] + clusterSequence
sequenceLength = len(self.clusterSequence)
self.clusterUnigramCounts = {}
self.clusterBigramCounts = SymmetricTable_Sparse(orderMatters=True)
current = self.clusterSequence[0]
self.clusterUnigramCounts[current] = 1.0
for i in range(1, sequenceLength):
previous = current
current = self.clusterSequence[i]
self.clusterUnigramCounts[current] = self.clusterUnigramCounts.get(current, 0.0) + 1.0
self.clusterBigramCounts.set(current, previous, self.clusterBigramCounts.get(current, previous, 0.0) + 1.0)
# CONSIDER: Does the concept of discounted probabilities mean anything here?
self.clusters = self.clusterUnigramCounts.keys()
self.originalClusters = copy(self.clusters)
log.debug("Found cluster sequence for %d words.", self.sequenceLength)
log.debug("%d clusters were found.", len(self.clusterUnigramCounts))
log.debug("%d bigrams were found.", len(self.clusterBigramCounts))
if self.NO_CLUSTER_SYMBOL in self.clusterUnigramCounts:
log.debug("Unclustered words remain in the dataset.")
else:
log.debug("No unclustered words remain in the dataset.")
class BrownClusterModel:
START_CLUSTER_SYMBOL = -1
NO_CLUSTER_SYMBOL = -2
def __init__(self):
self.wordSequence = []
self.sequenceLength = 0
self.TaskStack = TaskStack()
#################### Data Addition Methods
# TODO: Make more efficient. (You don't really need to redo the statistics every time you add data)
def addTrainingData(self, wordSequence):
(wordDataLength, wordDataWidth) = wordSequence.shape # Assume it comes in as numpy array
# Data Validation / Word Data Width
if not hasattr(self, "wordDataWidth"):
self.wordDataWidth = wordDataWidth
if wordDataWidth != self.wordDataWidth:
raise Exception("New words not of same length as old words")
log.debug("Adding new word sequence of length %d...", wordDataLength)
# Add Data. Simple Concatenation
self.wordSequence = self.wordSequence + self.__addTrainingData_tuplefyWordSequence(wordSequence)
# Update Sequence Length
self.sequenceLength += wordDataLength
# Reset Variables and Flags
self.__resetDataStatistics()
log.debug("Added new word sequence. Total word sequence is now %d.", self.sequenceLength)
def __addTrainingData_tuplefyWordSequence(self, wordSequence):
wordDataLength = len(wordSequence)
newWordSequence = [None] * wordDataLength # Pre-allocate. Probably not worthwhile...
for i in range(wordDataLength):
newWordSequence[i] = tuple(wordSequence[i, :])
return newWordSequence
def __resetDataStatistics(self):
# (Superfluous) Delete Existing Data
self.sortedWords = []
self.wordClusterMapping = {}
self.clusterSequence = []
self.clusters = []
self.clusterUnigramCounts = {}
self.clusterBigramCounts = None
self.clusterCostTable = None
self.mergeCostReductions = None
self.mergeHistory = []
self.binaryRepresentation = {}
self.interClusterDistance = {}
# Add New Tasks
self.TaskStack.push(
lambda: self.__resetData_definemergeCostReductions(),
"definemergeCostReductions",
"Finding Merge Cost Reduction Shortcuts",
)
self.TaskStack.push(
lambda: self.__resetData_defineClusterCostTable(),
"defineClusterCostTable",
"Creating Cluster Cost Table (Finding Bigram Graph Edge Costs)",
)
self.TaskStack.push(
lambda: self.__resetData_defineClusterCounts(),
"defineClusterCounts",
"Counting Cluster NGrams from Cluster Sequence",
)
self.TaskStack.push(
lambda: self.__resetData_defineClusterSequence(),
"defineClusterSequence",
"Creating Cluster Sequence from Word-to-Cluster Mapping and Word Sequence",
)
self.TaskStack.push(
lambda: self.__resetData_defineWordClusterMap(), "defineWordClusterMap", "Defining Word-to-Cluster Mapping"
)
# Clear The Tasks
self.TaskStack.clear()
def __resetData_defineWordClusterMap(self, numInitialClusters=None):
self.sortedWords = [x[0] for x in Counter(self.wordSequence).most_common()]
if numInitialClusters is None:
self.wordClusterMapping = dict(zip(self.sortedWords, range(0, len(self.sortedWords))))
else:
raise NotImplementedError("Have not implemented non-used clusters yet.")
self.wordClusterMapping = dict(zip(self.sortedWords[0:numInitialClusters], range(0, numInitialClusters)))
log.debug("Defined Word Cluster Map for %d words.", self.sequenceLength)
log.debug("%d words found.", len(self.sortedWords))
log.debug("%d clusters created.", len(self.wordClusterMapping))
def clusterCount_getBigramCount(self, cluster1, cluster2):
return self.clusterBigramCounts.get(cluster1, cluster2, 0.0)
def clusterCount_getClusterCount(self, cluster1):
return self.clusterUnigramCounts[cluster1] # Throw an error if not found
def __resetData_defineClusterCounts(self):
# clusterSequence = [self.START_CLUSTER_SYMBOL] + clusterSequence
sequenceLength = len(self.clusterSequence)
self.clusterUnigramCounts = {}
self.clusterBigramCounts = SymmetricTable_Sparse(orderMatters=True)
current = self.clusterSequence[0]
self.clusterUnigramCounts[current] = 1.0
for i in range(1, sequenceLength):
previous = current
current = self.clusterSequence[i]
self.clusterUnigramCounts[current] = self.clusterUnigramCounts.get(current, 0.0) + 1.0
self.clusterBigramCounts.set(current, previous, self.clusterBigramCounts.get(current, previous, 0.0) + 1.0)
# CONSIDER: Does the concept of discounted probabilities mean anything here?
self.clusters = self.clusterUnigramCounts.keys()
self.originalClusters = copy(self.clusters)
log.debug("Found cluster sequence for %d words.", self.sequenceLength)
log.debug("%d clusters were found.", len(self.clusterUnigramCounts))
log.debug("%d bigrams were found.", len(self.clusterBigramCounts))
if self.NO_CLUSTER_SYMBOL in self.clusterUnigramCounts:
log.debug("Unclustered words remain in the dataset.")
else:
log.debug("No unclustered words remain in the dataset.")
def __resetData_defineClusterSequence(self):
self.clusterSequence = self.__resetData_defineClusterSequence_generic(self.wordSequence)
def __resetData_defineClusterSequence_generic(self, wordSequence):
sequenceLength = len(wordSequence)
clusterSequence = [None] * sequenceLength
for i in range(sequenceLength):
clusterSequence[i] = self.wordClusterMapping.get(wordSequence[i], self.NO_CLUSTER_SYMBOL)
return clusterSequence
def __resetData_defineClusterCostTable(self):
numClusters = len(self.clusters)
self.clusterCostTable = SymmetricTable_Sparse(orderMatters=False)
for i in range(numClusters):
c1 = self.clusters[i]
cnt_c1 = self.clusterCount_getClusterCount(c1)
for j in range(i, numClusters):
c2 = self.clusters[j]
cost = self.clusterCost_findSingleClusterPairCost(c1, c2, cnt_c1=cnt_c1)
self.clusterCostTable.set(c1, c2, cost)
def clusterCost_findSingleClusterPairCost(self, c1, c2, cnt_c1=None, cnt_c2=None):
if cnt_c1 is None:
cnt_c1 = self.clusterCount_getClusterCount(c1)
if c1 == c2:
bigramCount = self.clusterCount_getBigramCount(c1, c1)
return self.mutualInfo_Equation(bigramCount, cnt_c1, cnt_c1)
else:
if cnt_c2 is None:
cnt_c2 = self.clusterCount_getClusterCount(c2)
bigramCount = self.clusterCount_getBigramCount(c1, c2)
cost = self.mutualInfo_Equation(bigramCount, cnt_c1, cnt_c2)
bigramCount = self.clusterCount_getBigramCount(c2, c1)
cost += self.mutualInfo_Equation(bigramCount, cnt_c1, cnt_c2)
return cost
def mutualInfo_Equation(self, bigramCount, unigram1Count, unigram2Count):
if bigramCount > 0:
# I've been uncomfortable with logs being above/below 0, but what I'm coming to realize is that the zero
# intercept of mutual information is where information goes from useful to not..?
# Sample probability? or Population probability? Population.
# n = sequenceLength - 1 # Sample probability
n = self.sequenceLength # Population probability
return bigramCount / n * lgrthm(n * bigramCount / unigram1Count / unigram2Count, 2)
else:
if unigram1Count == 0 or unigram2Count == 0:
raise Exception("Erroneous clusters")
return 0.0
# This is the mutual information from treating two separate clusters as one.
# If you want to treat 1 and 2 as A, vis-a-vis 3, then:
# you need to account for 1->3, 2->3, 3->1, and 3->2 connections, and treat them all as A->3 and 3->A;
# you also need to account for P(A) = P(1)+P(2)
def mutualInfo_PairVersusAnother(self, mc1, mc2, c3, cnt_mc1=None, cnt_mc2=None, cnt_c3=None):
if cnt_mc1 is None:
self.clusterCount_getClusterCount(mc1)
if cnt_mc2 is None:
cnt_mc2 = self.clusterCount_getClusterCount(mc2)
if cnt_c3 is None:
cnt_c3 = self.clusterCount_getClusterCount(c3)
mutualInformation = 0.0
bigram1Count = self.clusterCount_getBigramCount(c3, mc1) # 3->1
bigram2Count = self.clusterCount_getBigramCount(c3, mc2) # 3->2
mutualInformation += self.mutualInfo_Equation(
(bigram1Count + bigram2Count), cnt_c3, (cnt_mc1 + cnt_mc2)
) # 3->A
bigram1Count = self.clusterCount_getBigramCount(mc1, c3) # 1->3
bigram2Count = self.clusterCount_getBigramCount(mc2, c3) # 2->3
mutualInformation += self.mutualInfo_Equation(
(bigram1Count + bigram2Count), cnt_c3, (cnt_mc1 + cnt_mc2)
) # A->3
return mutualInformation
# This is the mutual information from treating two separate clusters as one.
# If you want to treat 1 and 2 as A, then:
# you need to account for 1->2, 2->1, 1->1, and 2->2 connections, and treat them all as A->A;
# you need to account for P(A) = P(1)+P(2)
def mutualInfo_PairIntoOne(self, c1, c2, cnt_c1=None, cnt_c2=None):
if cnt_c1 is None:
cnt_c1 = self.clusterCount_getClusterCount(c1)
if cnt_c2 is None:
cnt_c2 = self.clusterCount_getClusterCount(c2)
bigram1Count = self.clusterCount_getBigramCount(c1, c2) # 1->2
bigram2Count = self.clusterCount_getBigramCount(c2, c1) # 2->1
bigram3Count = self.clusterCount_getBigramCount(c1, c1) # 1<->1
bigram4Count = self.clusterCount_getBigramCount(c2, c2) # 2<->2
totalBigramCount = bigram1Count + bigram2Count + bigram3Count + bigram4Count
return self.mutualInfo_Equation(totalBigramCount, (cnt_c1 + cnt_c2), (cnt_c1 + cnt_c2))
# This is the mutual information from treating 4 separate clusters as two.
# If you want to treat 1 and 2 as A, and 3 and 4 as B, then:
# you need to account for 1->3, 1->4, 2->3, and 2->4 connections, and treat them all as A->B;
# you need to account for 3->1, 4->1, 3->2, and 4->2 connections, and treat them all as B->A;
# you need to account for P(A) = P(1)+P(2)
# you need to account for P(B) = P(3)+P(4)
def mutualInfo_PairVersusPair(self, c1, c2, c3, c4, cnt_c1=None, cnt_c2=None, cnt_c3=None, cnt_c4=None):
if cnt_c1 is None:
cnt_c1 = self.clusterCount_getClusterCount(c1)
if cnt_c2 is None:
cnt_c2 = self.clusterCount_getClusterCount(c2)
if cnt_c3 is None:
cnt_c3 = self.clusterCount_getClusterCount(c3)
if cnt_c4 is None:
cnt_c4 = self.clusterCount_getClusterCount(c4)
mutualInformation = 0.0
# Group 1 to Group 2
bigram1Count = self.clusterCount_getBigramCount(c1, c3)
bigram2Count = self.clusterCount_getBigramCount(c1, c4)
bigram3Count = self.clusterCount_getBigramCount(c2, c3)
bigram4Count = self.clusterCount_getBigramCount(c2, c4)
totalBigramCount = bigram1Count + bigram2Count + bigram3Count + bigram4Count
mutualInformation += self.mutualInfo_Equation(totalBigramCount, (cnt_c1 + cnt_c2), (cnt_c3 + cnt_c4))
# Group 2 to Group 1
bigram1Count = self.clusterCount_getBigramCount(c3, c1)
bigram2Count = self.clusterCount_getBigramCount(c3, c2)
bigram3Count = self.clusterCount_getBigramCount(c4, c1)
bigram4Count = self.clusterCount_getBigramCount(c4, c2)
totalBigramCount = bigram1Count + bigram2Count + bigram3Count + bigram4Count
mutualInformation += self.mutualInfo_Equation(totalBigramCount, (cnt_c1 + cnt_c2), (cnt_c3 + cnt_c4))
return mutualInformation
# This function gives the merge reduction cost for a single cluster pair using the naive algorithm.
# This algorithm has been verified.
def mergeCost_SinglePair(self, c1, c2):
cnt_c1 = self.clusterCount_getClusterCount(c1)
cnt_c2 = self.clusterCount_getClusterCount(c2)
clusterCostReduction = 0.0
clusterCostAddition = 0.0
for c3 in self.clusters:
if c3 == c1 or c3 == c2:
continue # deal with these separately.
clusterCostReduction += self.clusterCostTable.get(c1, c3) # 1<->3 (Encompasses 1->3 and 3->1)
clusterCostReduction += self.clusterCostTable.get(c2, c3) # 2<->3 (Encompasses 2->3 and 3->2)
# This is the procedure you get if you try to combine two nodes into one:
# P(c,c')*log(P(c,c')/P(c)/P(c'))
cnt_c3 = self.clusterCount_getClusterCount(c3)
clusterCostAddition += self.mutualInfo_PairVersusAnother(c1, c2, c3, cnt_c1, cnt_c2, cnt_c3)
# Deal with connections among the pair
clusterCostReduction += self.clusterCostTable.get(c1, c2) # 1<->2 (Encompasses 1->2 and 2->1)
clusterCostReduction += self.clusterCostTable.get(c1, c1) # 1<->1
clusterCostReduction += self.clusterCostTable.get(c2, c2) # 2<->2
clusterCostAddition += self.mutualInfo_PairIntoOne(c1, c2, cnt_c1, cnt_c2)
return clusterCostAddition - clusterCostReduction
def __resetData_definemergeCostReductions(self):
self.mergeCostReductions = SymmetricTable_Sparse(orderMatters=False)
clusters = self.clusters
numClusters = len(self.clusters)
for i in range(numClusters):
c1 = self.clusters[i]
for j in range(i):
c2 = self.clusters[j]
mergeCost = self.mergeCost_SinglePair(c1, c2)
self.mergeCostReductions.set(c1, c2, mergeCost)
def mergeClusters_findMergeClusters(self):
if len(self.mergeCostReductions) == 0: # Necessary?
return (False, None) # Necessary?
return (True, sorted(self.mergeCostReductions.items(), key=itemgetter(1), reverse=True)[0][0])
# Change NGram Counts from merging 1 into 2
# We're deleting 2
def mergeClusters_changeNGramCounts(self, mc1, mc2):
# Change Unigram Counts / Delete Cluster mc2
self.clusterUnigramCounts[mc1] += self.clusterCount_getClusterCount(mc2)
del self.clusterUnigramCounts[mc2]
# Change Unigram Counts / Reset Saved Cluster Keys
self.clusters = self.clusterUnigramCounts.keys()
# Change Bigram Counts / Change Non-Merging Clusters
for c3 in self.clusters:
if c3 == mc1 or c3 == mc2:
continue # deal with these separately.
else:
bigramCount = self.clusterBigramCounts.delete(c3, mc1, 0) # 3->1
bigramCount += self.clusterBigramCounts.delete(c3, mc2, 0) # 3->2
self.clusterBigramCounts.set(c3, mc1, bigramCount) # (3->2 + 3->1) => 3->1
bigramCount = self.clusterBigramCounts.get(mc1, c3, 0) # 1->3
bigramCount += self.clusterBigramCounts.delete(mc2, c3, 0) # 2->3
self.clusterBigramCounts.set(mc1, c3, bigramCount) # (1->3 + 2->3) => 1->3
# Change Bigram Counts / Change Merging Clusters
bigramCount = self.clusterBigramCounts.delete(mc1, mc1, 0) # 1->1
bigramCount += self.clusterBigramCounts.delete(mc2, mc1, 0) # 2->1
bigramCount += self.clusterBigramCounts.delete(mc1, mc2, 0) # 1->2
bigramCount += self.clusterBigramCounts.delete(mc2, mc2, 0) # 2->2
self.clusterBigramCounts.set(mc1, mc1, bigramCount) # all => 1->1
# Change Cluster Costs from merging 1 into 2
# We're deleting 2
# Depends on NEW Cluster Counts (i.e. NGram Counts)
def mergeClusters_changeClusterCosts(self, mc1, mc2):
new_cnt_mc1 = self.clusterCount_getClusterCount(mc1) # Depends on NEW Cluster Counts
# Change ClusterCost Table / Change Non-Merging Clusters
for c3 in self.clusters:
if c3 == mc1 or c3 == mc2:
continue # deal with these separately.
else:
cost = self.clusterCost_findSingleClusterPairCost(
mc1, c3, cnt_c1=new_cnt_mc1
) # Depends on NEW Cluster Counts
self.clusterCostTable.set(mc1, c3, cost)
self.clusterCostTable.delete(mc2, c3, silent=True)
# Change ClusterCost Table / Change Merging Clusters
cost = self.clusterCost_findSingleClusterPairCost(mc1, mc1, cnt_c1=new_cnt_mc1)
self.clusterCostTable.set(mc1, mc1, cost)
self.clusterCostTable.delete(mc2, mc2, silent=True)
def mergeClusters_mergeTop(self, updateClusterSequence=False, verbose=False):
# 1) Find Merge Clusters
(success, mergeClusters) = self.mergeClusters_findMergeClusters()
if not success:
return False
(c1, c2) = mergeClusters
if verbose: # TODO: Remove
cost = self.mergeCostReductions.get(c1, c2)
print "Merging Cluster {1} into Cluster {0}, with clusterCostReduction={2}".format(c1, c2, cost)
# 2) Change Merge Cost Reduction using OLD Ngram Counts
self.mergeClusters_changeMergeCostReductions_UnmergedClusters(c1, c2)
self.mergeClusters_changeMergeCostReductions_RemoveDeletedCluster(c2)
# 3) Change Unigram and Bigram Counts
self.mergeClusters_changeNGramCounts(c1, c2)
# 4) Change Cluster Cost Table
self.mergeClusters_changeClusterCosts(c1, c2)
# 5) Change Merge Cost Reduction using NEW Ngram Counts
self.mergeClusters_changeMergeCostReductions_MergedClusters(c1)
# 6) Record Change in MergeHistory
self.mergeHistory.append((c1, c2))
# TODO: Skip mergeHistory; go straight to dictionaryrepresentation.
# 7) Update Cluster Sequence (Optional)
if updateClusterSequence:
if not hasattr(self, "originalClusterSequence"): # sloppy
self.originalClusterSequence = self.clusterSequence # sloppy
self.mergeClusters_updateClusterSequence(c1, c2)
return True
# This updates the mergeCostReductions after mergeCluster1 and mergeCluster2 are merged.
# This relies on NEW (i.e. post-merge) NGram Counts
# That means mc2 shouldn't be in any tables anymore
def mergeClusters_changeMergeCostReductions_MergedClusters(self, mc1):
for c3 in self.clusters:
if c3 == mc1:
continue
else:
newMergeCostReduction = self.mergeCost_SinglePair(mc1, c3)
self.mergeCostReductions.set(mc1, c3, newMergeCostReduction)
def mergeClusters_changeMergeCostReductions_RemoveDeletedCluster(self, mc2):
for c3 in self.clusters:
if c3 == mc2:
continue
else:
self.mergeCostReductions.delete(mc2, c3, silent=False) # Test. I think I want silent
# This updates the mergeCostReductions after mergeCluster1 and mergeCluster2 are merged.
# This relies on OLD (i.e. pre-merge) NGram Counts
def mergeClusters_changeMergeCostReductions_UnmergedClusters(self, mc1, mc2):
cnt_mc1 = self.clusterCount_getClusterCount(mc1)
cnt_mc2 = self.clusterCount_getClusterCount(mc2)
numClusters = len(self.clusters)
for i in range(numClusters):
c3 = self.clusters[i]
if c3 == mc1 or c3 == mc2:
continue
cnt_c3 = self.clusterCount_getClusterCount(c3)
for j in range(i):
c4 = self.clusters[j]
if c4 == mc1 or c4 == mc2:
continue
cnt_c4 = self.clusterCount_getClusterCount(c4)
mergeCostAddition = 0
mergeCostAddition += self.clusterCostTable.get(c3, mc1) # c3<->mc1
mergeCostAddition += self.clusterCostTable.get(c3, mc2) # c3<->mc2
mergeCostAddition += self.clusterCostTable.get(c4, mc1) # c4<->mc1
mergeCostAddition += self.clusterCostTable.get(c4, mc2) # c4<->mc2
mergeCostAddition += self.mutualInfo_PairVersusPair(
c3, c4, mc1, mc2, cnt_c3, cnt_c4, cnt_mc1, cnt_mc2
) # (c3+c4)<->(mc1+mc2)
mergeCostReduction = 0
mergeCostReduction += self.mutualInfo_PairVersusAnother(
mc1, mc2, c3, cnt_mc1, cnt_mc2, cnt_c3
) # c3<->(mc1+mc2)
mergeCostReduction += self.mutualInfo_PairVersusAnother(
mc1, mc2, c4, cnt_mc1, cnt_mc2, cnt_c4
) # c4<->(mc1+mc2)
mergeCostReduction += self.mutualInfo_PairVersusAnother(
c3, c4, mc1, cnt_c3, cnt_c4, cnt_mc1
) # mc1<->(c3+c4)
mergeCostReduction += self.mutualInfo_PairVersusAnother(
c3, c4, mc2, cnt_c3, cnt_c4, cnt_mc2
) # mc2<->(c3+c4)
mergeCostChange = mergeCostAddition - mergeCostReduction
self.mergeCostReductions.set(
c3, c4, self.mergeCostReductions.get(c3, c4) + mergeCostChange
) # Do we want the 'get' to return 0?
def performClustering_convertMergeHistoryToBinaryWords(self):
self.binaryRepresentation = dict.fromkeys(self.originalClusters, "")
for (mc1, mc2) in reversed(self.mergeHistory):
self.binaryRepresentation[mc2] = self.binaryRepresentation[mc1] + "1"
self.binaryRepresentation[mc1] = self.binaryRepresentation[mc1] + "0"
maxDepth = 0
for (cluster, binRep) in self.binaryRepresentation.iteritems():
maxDepth = max(maxDepth, len(binRep))
self.binaryRepresentationMaxLen = maxDepth
def performClustering_convertMergeHistoryToDictionary(self):
clusterWordMapping = dict(zip(self.wordClusterMapping.values(), self.wordClusterMapping.keys()))
d = {}
for (mc1, mc2) in self.mergeHistory:
leftNode = d.get(mc1, (mc1, clusterWordMapping[mc1]))
rightNode = d.get(mc2, (mc2, clusterWordMapping[mc2]))
d[mc1] = {-2: leftNode, -1: rightNode}
self.dictionaryRepresentation = {"root": d[mc1]}
def performClustering_findInterClusterDistances(self):
if len(self.binaryRepresentation) == 0: # sloppy.
self.performClustering_convertMergeHistoryToBinaryWords() # sloppy. Use task stack instead.
numClusters = len(self.originalClusters)
self.interClusterDistance = SymmetricTable_FixedSize(numClusters, includeDiagonal=False)
for i in range(numClusters):
binRep_i = self.binaryRepresentation[i]
for j in range(i + 1, numClusters):
matchLength = self.performClustering_findInterClusterDistances_stringMatchLength(
binRep_i, self.binaryRepresentation[j]
)
self.interClusterDistance.set(i, j, self.binaryRepresentationMaxLen - matchLength)
# Consider: could be done more efficiently? (i.e. don't use strings)
def performClustering_findInterClusterDistances_stringMatchLength(self, string1, string2):
length = 0
for i in range(min(len(string1), len(string2))):
if string1[i] != string2[i]:
break
length += 1
return length
# TODO: Can this be done more efficiently?
def performClustering_establishDistanceMeasure(self):
if hasattr(self, "originalClusterSequence"): # sloppy
seq = self.originalClusterSequence # sloppy
else: # sloppy
seq = self.clusterSequence # sloppy
sequenceLength = len(seq)
self.distanceTable = SymmetricTable_FixedSize(sequenceLength, includeDiagonal=False)
for i in range(sequenceLength):
c1 = seq[i]
for j in range(i + 1, sequenceLength):
c2 = seq[j]
if c1 == c2:
self.distanceTable.set(i, j, 0)
else:
self.distanceTable.set(i, j, self.interClusterDistance.get(c1, c2))
def performBrownianClustering(self):
leftToMerge = True
while leftToMerge:
leftToMerge = self.mergeClusters_mergeTop()
self.performClustering_convertMergeHistoryToBinaryWords()
self.performClustering_findInterClusterDistances()
self.performClustering_establishDistanceMeasure()
def mergeClusters_updateClusterSequence(self, mc1, mc2):
for i in range(self.sequenceLength):
if self.clusterSequence[i] == mc2:
self.clusterSequence[i] = mc1
def findTotalClusteringCost(self):
qualityCost = 0
for c1 in self.clusters:
cnt_c1 = self.clusterCount_getClusterCount(c1)
for c2 in self.clusters:
cnt_c2 = self.clusterCount_getClusterCount(c2)
bigramCount = self.clusterCount_getBigramCount(c1, c2)
qualityCost += self.mutualInfo_Equation(bigramCount, cnt_c1, cnt_c2)
return qualityCost