def setUp(self):
self.dimension, self.signatureLength = 50, 23
self.phraseTextAndDimensionMap = TwoWayMap()
for i in range(self.dimension):
self.phraseTextAndDimensionMap.set(TwoWayMap.MAP_FORWARD, i, i)
self.unitRandomVectors = [
VectorGenerator.getRandomGaussianUnitVector(
dimension=self.dimension, mu=0, sigma=1)
for i in range(self.signatureLength)
]
self.doc1 = Document(
1,
VectorGenerator.getRandomGaussianUnitVector(
dimension=self.dimension, mu=0, sigma=1))
self.doc2 = Document(
2,
VectorGenerator.getRandomGaussianUnitVector(
dimension=self.dimension, mu=0, sigma=1))
self.doc1.setSignatureUsingVectors(self.unitRandomVectors,
self.phraseTextAndDimensionMap)
self.doc2.setSignatureUsingVectors(self.unitRandomVectors,
self.phraseTextAndDimensionMap)
self.pm = SignaturePermutationWithTrie(
signatureLength=self.signatureLength)
self.pm.addDocument(self.doc1)
self.pm.addDocument(self.doc2)
def __init__(self, **settings):
self.settings = settings
self.nearestNeighborThreshold = settings['nearest_neighbor_threshold']
self.unitVector = RandomGaussianUnitVector(
dimensions=settings['dimensions'], mu=0, sigma=1)
self.vectorPermutations = VectorPermutation.getPermutations(
settings['signature_length'], settings['dimensions'],
self.unitVector)
# self.signaturePermutations = [SignaturePermutationWithTrie(settings['signature_length']) for i in range(settings['number_of_permutations'])]
signatureType = settings.get('signature_type', 'signature_type_trie')
if signatureType == 'signature_type_trie':
self.signaturePermutations = [
SignaturePermutationWithTrie(settings['signature_length'])
for i in range(settings['number_of_permutations'])
]
else:
self.signaturePermutations = [
SignaturePermutationWithSortedList(
settings['signature_length'])
for i in range(settings['number_of_permutations'])
]
self.phraseTextAndDimensionMap = TwoWayMap()
self.documentIdToDocumentMap = {}
def test_addDocument_newKey(self):
doc1 = Document(
1,
VectorGenerator.getRandomGaussianUnitVector(
dimension=self.dimension, mu=0, sigma=1))
doc1.setSignatureUsingVectors(self.unitRandomVectors,
self.phraseTextAndDimensionMap)
pm = SignaturePermutationWithTrie(signatureLength=self.signatureLength)
pm.addDocument(doc1)
self.assertEqual(pm.signatureTrie[doc1.signature.permutate(pm).to01()],
set([1]))
def __init__(self, **clustering_settings):
self.thresholdForDocumentToBeInACluster = clustering_settings[
'threshold_for_document_to_be_in_cluster']
self.unitVector = RandomGaussianUnitVector(
dimensions=clustering_settings['dimensions'], mu=0, sigma=1)
self.vectorPermutations = VectorPermutation.getPermutations(
clustering_settings['signature_length'],
clustering_settings['dimensions'], self.unitVector)
signatureType = clustering_settings.get('signature_type',
'signature_type_trie')
if signatureType == 'signature_type_trie':
self.signaturePermutations = [
SignaturePermutationWithTrie(
clustering_settings['signature_length'])
for i in range(clustering_settings['number_of_permutations'])
]
else:
self.signaturePermutations = [
SignaturePermutationWithSortedList(
clustering_settings['signature_length'])
for i in range(clustering_settings['number_of_permutations'])
]
self.phraseTextAndDimensionMap, self.clusters = TwoWayMap(), {}
self.clustering_settings = clustering_settings
def test_permutate(self):
sgnt = Signature('1001011')
self.assertTrue(sgnt.count() == sgnt.permutate(
SignaturePermutationWithTrie(7)).count())
def test_getNearestDocument_emptyTrie(self):
permutationWithEmptyTrie = SignaturePermutationWithTrie(
signatureLength=self.signatureLength)
self.assertEqual(
permutationWithEmptyTrie.getNearestDocuments(self.doc1), set())
def offlineLSHClusteringDemo():
wordToDimensionMap = {}
def createDocumentFromLine(docId, line):
vector = Vector()
words = line.split()
for word in words[1:]:
if word not in wordToDimensionMap:
wordToDimensionMap[word] = len(wordToDimensionMap)
wordDimension = wordToDimensionMap[word]
if wordDimension not in vector: vector[wordDimension] = 1
else: vector[wordDimension] += 1
return Document(docId, vector, clusterId=words[0])
dimensions = 53
signatureLength = 13
numberOfPermutations = 5
unitVector = RandomGaussianUnitVector(dimensions=dimensions, mu=0, sigma=1)
vectorPermutations = VectorPermutation.getPermutations(
signatureLength, dimensions, unitVector)
signaturePermutations = [
SignaturePermutationWithTrie(signatureLength)
for i in range(numberOfPermutations)
]
permutatedUnitVectors = [
unitVector.getPermutedVector(r) for r in vectorPermutations
]
# Build LSH Model.
# Read training documents.
traningDocumentsMap = {}
for docId, l in enumerate(
FileIO.iterateLinesFromFile('../data/train_offline.dat')):
traningDocumentsMap[docId] = createDocumentFromLine(docId, l)
# Construct cluster vectors.
clusterToDocumentsMap = defaultdict(list)
for document in traningDocumentsMap.values():
clusterToDocumentsMap[document.clusterId].append(document)
clusterMap = {}
for k, v in clusterToDocumentsMap.iteritems():
clusterMap[k] = Document(docId=k,
vector=Vector.getMeanVector(v),
clusterId=k)
# Create signatures and signaturePermutations for all the clusters.
map(
lambda document: document.setSignatureUsingVectors(
permutatedUnitVectors), clusterMap.values())
for permutation in signaturePermutations:
for document in clusterMap.values():
permutation.addDocument(document)
# Testing the model.
# Read testing documents.
testDocumentsMap = {}
for docId, l in enumerate(
FileIO.iterateLinesFromFile('../data/test_offline.dat')):
testDocumentsMap[docId] = createDocumentFromLine(docId, l)
# Create signatures for test documents
map(
lambda document: document.setSignatureUsingVectors(
permutatedUnitVectors), testDocumentsMap.values())
predicted, labels = [], []
for t in testDocumentsMap.values():
possibleNearestClusters = reduce(
lambda x, y: x.union(y),
(permutation.getNearestDocuments(t)
for permutation in signaturePermutations), set())
predictedClass = max(
((clusterId, clusterMap[clusterId].cosineSimilarity(t))
for clusterId in possibleNearestClusters),
key=itemgetter(1))
predicted.append(predictedClass[0])
labels.append(t.clusterId)
return EvaluationMetrics.purity(predicted, labels)
