def testIsDictionary(self): # - - - - - - - - - - - - - - - - - - - - - - -
"""Test 'check_is_dictionary' function."""
assert (auxiliary.check_is_dictionary('TestArgument',{}) == None)
assert (auxiliary.check_is_dictionary('TestArgument', {1:2,6:0}) == None)
assert (auxiliary.check_is_dictionary('TestArgument', \
{'a':4,'t':1,(1,4,6):'tr'}) == None)
def testIsDictionary(self): # - - - - - - - - - - - - - - - - - - - - - - -
"""Test 'check_is_dictionary' function."""
assert (auxiliary.check_is_dictionary('TestArgument',{}) == None)
assert (auxiliary.check_is_dictionary('TestArgument', {1:2,6:0}) == None)
assert (auxiliary.check_is_dictionary('TestArgument', \
{'a':4,'t':1,(1,4,6):'tr'}) == None)
def pairs_quality(weight_vec_dict, match_check_funct):
"""Pairs quality is the ratio of true matches divided by the total number of
matches of the compared record pairs returned after blocking. It is
measured as:
pq = |TP| / all_matches
with TP being the true positives, and all matches being the number of
weight vectors given.
The arguments that have to be set when this method is called are:
weight_vec_dict A dictionary containing weight vectors.
match_check_funct This has to be a function (or method), assumed to
have as arguments the two record identifiers of a
record pair and its weight vector, and returns True
if the record pair is from a true match, or False
otherwise. Thus, 'match_check_funct' is of the form:
match_flag = match_check_funct(rec_id1, rec_id2,
weight_vec)
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_function_or_method('match_check_funct',
match_check_funct)
total_num_rec_pairs = len(weight_vec_dict)
logging.info('')
logging.info('Calculate pairs quality:')
logging.info(' Number of record pairs in weight vector dictionary: %d' % \
(total_num_rec_pairs))
# Get number of true matches in weight vector dictionary - - - - - - - - - -
#
num_true_matches = 0
num_false_matches = 0
for (rec_id_tuple, this_vec) in weight_vec_dict.iteritems():
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1],
this_vec) == True):
num_true_matches += 1
else:
num_false_matches += 1
assert total_num_rec_pairs == (num_true_matches + num_false_matches)
logging.info(' Number of true and false matches in weight vector ' + \
'dictionary: %d / %d' % (num_true_matches, num_false_matches))
pq = float(num_true_matches) / total_num_rec_pairs
logging.info(' Pairs quality: %.4f%%' % (100.0 * pq)) # As percentage
assert pq <= 1.0
return pq
def pairs_quality(weight_vec_dict, match_check_funct):
"""Pairs quality is the ratio of true matches divided by the total number of
matches of the compared record pairs returned after blocking. It is
measured as:
pq = |TP| / all_matches
with TP being the true positives, and all matches being the number of
weight vectors given.
The arguments that have to be set when this method is called are:
weight_vec_dict A dictionary containing weight vectors.
match_check_funct This has to be a function (or method), assumed to
have as arguments the two record identifiers of a
record pair and its weight vector, and returns True
if the record pair is from a true match, or False
otherwise. Thus, 'match_check_funct' is of the form:
match_flag = match_check_funct(rec_id1, rec_id2,
weight_vec)
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_function_or_method('match_check_funct', match_check_funct)
total_num_rec_pairs = len(weight_vec_dict)
logging.info('')
logging.info('Calculate pairs quality:')
logging.info(' Number of record pairs in weight vector dictionary: %d' % \
(total_num_rec_pairs))
# Get number of true matches in weight vector dictionary - - - - - - - - - -
#
num_true_matches = 0
num_false_matches = 0
for (rec_id_tuple, this_vec) in weight_vec_dict.iteritems():
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1], this_vec) == True):
num_true_matches += 1
else:
num_false_matches += 1
assert total_num_rec_pairs == (num_true_matches + num_false_matches)
logging.info(' Number of true and false matches in weight vector ' + \
'dictionary: %d / %d' % (num_true_matches, num_false_matches))
pq = float(num_true_matches) / total_num_rec_pairs
logging.info(' Pairs quality: %.4f%%' % (100.0*pq)) # As percentage
assert pq <= 1.0
return pq
def testIsDictionary(
self): # - - - - - - - - - - - - - - - - - - - - - - -
"""Test 'check_is_dictionary' function."""
assert auxiliary.check_is_dictionary("TestArgument", {})
assert auxiliary.check_is_dictionary("TestArgument", {1: 2, 6: 0})
assert (auxiliary.check_is_dictionary("TestArgument", {
"a": 4,
"t": 1,
(1, 4, 6): "tr"
}))
def SaveMatchStatusFile(w_vec_dict, match_set, file_name):
"""Save the matched record identifiers into a CVS file.
This function saves the record identifiers of all record pairs that are in
the given match set into a CSV file with four columns:
- First record identifier
- Second record identifier
- Summed matching weight from the corresponding weight vector
- A unique match identifier (generated in the same way as the ones in the
function SaveMatchDataSet below).
"""
auxiliary.check_is_dictionary('w_vec_dict', w_vec_dict)
auxiliary.check_is_set('match_set', match_set)
auxiliary.check_is_string('file_name', file_name)
match_rec_id_list = list(match_set) # Make a list so it can be sorted
match_rec_id_list.sort()
if (len(match_set) > 0):
num_digit = max(1, int(math.ceil(math.log(len(match_set), 10))))
else:
num_digit = 1
mid_count = 1 # Counter for match identifiers
# Try to open the file for writing
#
try:
f = open(file_name, 'w')
except:
logging.exception('Cannot open file "%s" for writing' %
(str(file_name)))
raise IOError
for rec_id_tuple in match_rec_id_list:
w_vec = w_vec_dict[rec_id_tuple]
w_sum = sum(w_vec)
mid_count_str = '%s' % (mid_count)
this_mid = 'mid%s' % (mid_count_str.zfill(num_digit))
rec_id1 = rec_id_tuple[0]
rec_id2 = rec_id_tuple[1]
f.write('%s,%s,%f,%s' % (rec_id1, rec_id2, w_sum, this_mid) +
os.linesep)
mid_count += 1
f.close()
def SaveMatchStatusFile(w_vec_dict, match_set, file_name):
"""Save the matched record identifiers into a CVS file.
This function saves the record identifiers of all record pairs that are in
the given match set into a CSV file with four columns:
- First record identifier
- Second record identifier
- Summed matching weight from the corresponding weight vector
- A unique match identifier (generated in the same way as the ones in the
function SaveMatchDataSet below).
"""
auxiliary.check_is_dictionary('w_vec_dict', w_vec_dict)
auxiliary.check_is_set('match_set', match_set)
auxiliary.check_is_string('file_name', file_name)
match_rec_id_list = list(match_set) # Make a list so it can be sorted
match_rec_id_list.sort()
if (len(match_set) > 0):
num_digit = max(1,int(math.ceil(math.log(len(match_set), 10))))
else:
num_digit = 1
mid_count = 1 # Counter for match identifiers
# Try to open the file for writing
#
try:
f = open(file_name, 'w')
except:
logging.exception('Cannot open file "%s" for writing' % (str(file_name)))
raise IOError
for rec_id_tuple in match_rec_id_list:
w_vec = w_vec_dict[rec_id_tuple]
w_sum = sum(w_vec)
mid_count_str = '%s' % (mid_count)
this_mid = 'mid%s' % (mid_count_str.zfill(num_digit))
rec_id1 = rec_id_tuple[0]
rec_id2 = rec_id_tuple[1]
f.write('%s,%s,%f,%s' % (rec_id1, rec_id2, w_sum, this_mid) + os.linesep)
mid_count += 1
f.close()
def GenerateHistogram(w_vec_dict, bin_width, file_name=None, match_sets=None):
"""Print and/or save a histogram of the weight vectors stored in the given
dictionary, and according to the match sets (if given).
The histogram is rotated 90 degrees clockwise, i.e. up to down instead of
left to right.
This function sums up the number of weight vectors with a matching weight
in a given bin (according to the given bin width).
If given, the match sets must be a tuple containing three sets, the first
being a set with matches, the second with non-matches, and the third with
possible matches, as generated by classifiers in the classification.py
Febrl module.
For each bin, the number of weight vectors in this bin is printed as well,
and if the match sets are given the number of matches, non-matches and
possible matches in this bin.
If a file name is given, the output will be written into this text file.
This function returns a list of containing the histogram as text strings.
"""
MAX_HISTO_WIDTH = 80 # maximum width in characters
auxiliary.check_is_dictionary('w_vec_dict', w_vec_dict)
auxiliary.check_is_number('bin_width', bin_width)
auxiliary.check_is_positive('bin_width', bin_width)
if (file_name != None):
auxiliary.check_is_string('file_name', file_name)
if (match_sets != None):
auxiliary.check_is_tuple('match_sets', match_sets)
if (len(match_sets) != 3):
logging.exception('Match sets must be a tuple containing three sets.')
raise Exception
auxiliary.check_is_set('match_sets[0]', match_sets[0])
auxiliary.check_is_set('match_sets[1]', match_sets[1])
auxiliary.check_is_set('match_sets[2]', match_sets[2])
if (len(w_vec_dict) != (len(match_sets[0]) + len(match_sets[1]) + \
len(match_sets[2]))):
logging.exception('Lengths of weight vector dictionary differs from' + \
'summed lengths of match sets.')
raise Exception
# Check if weight vector dictionary is empty, if so return empty list
#
if (w_vec_dict == {}):
logging.warn('Empty weight vector dictionary given for histogram ' + \
'generation')
return []
# Get a random vector dictionary element to get dimensionality of vectors
#
(rec_id_tuple, w_vec) = w_vec_dict.popitem()
v_dim = len(w_vec)
w_vec_dict[rec_id_tuple] = w_vec # Put back in
histo_dict = {} # A combined histogram dictionary
if (match_sets != None): # Also matches, non-matches and possible matches
match_histo_dict = {}
non_match_histo_dict = {}
poss_match_histo_dict = {}
max_bin_w_count = -1 # Maximal count for one binned weight entry
# Loop over weight vectors - - - - - - - - - - - - - - - - - - - - - - - - -
#
for (rec_id_tuple, w_vec) in w_vec_dict.iteritems():
w_sum = sum(w_vec) # Sum all weight vector elements
binned_w = w_sum - (w_sum % bin_width)
binned_w_count = histo_dict.get(binned_w,0) + 1 # Increase count by one
histo_dict[binned_w] = binned_w_count
if (binned_w_count > max_bin_w_count): # Check if this is new maximum count
max_bin_w_count = binned_w_count
if (match_sets != None):
if (rec_id_tuple in match_sets[0]):
binned_w_count = match_histo_dict.get(binned_w,0) + 1
match_histo_dict[binned_w] = binned_w_count
elif (rec_id_tuple in match_sets[1]):
binned_w_count = non_match_histo_dict.get(binned_w,0) + 1
non_match_histo_dict[binned_w] = binned_w_count
else: # A possible match
binned_w_count = poss_match_histo_dict.get(binned_w,0) + 1
poss_match_histo_dict[binned_w] = binned_w_count
# Sort histogram according to X axis values - - - - - - - - - - - - - - - - -
#
x_vals = histo_dict.keys()
x_vals.sort()
assert sum(histo_dict.values()) == len(w_vec_dict)
if (match_sets == None): # Can use 68 characters for histogram
scale_factor_y = float(MAX_HISTO_WIDTH-19) / max_bin_w_count
elif (len(poss_match_histo_dict) == 0): # No possible matches
scale_factor_y = float(MAX_HISTO_WIDTH-30) / max_bin_w_count
else: # All three set non-empty
scale_factor_y = float(MAX_HISTO_WIDTH-41) / max_bin_w_count
# Generate the histogram as a list of strings - - - - - - - - - - - - - - - -
#
histo_list = []
histo_list.append('Weight histogram:')
histo_list.append('-----------------')
if (match_sets == None):
histo_list.append(' Counts | w_sum |')
histo_list.append('-------------------')
elif (len(poss_match_histo_dict) == 0): # No possible matches
histo_list.append(' Counts |')
histo_list.append(' Match | Non-Match| w_sum |')
histo_list.append('------------------------------')
else:
histo_list.append(' Counts |')
histo_list.append(' Match | Non-Match|Poss-Match| w_sum |')
histo_list.append('-----------------------------------------')
for x_val in x_vals:
this_count = histo_dict[x_val]
if (match_sets == None):
line_str = '%9d | %5.2f |' % (this_count, x_val)
elif (len(poss_match_histo_dict) == 0): # No possible matches
this_match_count = match_histo_dict.get(x_val, 0)
this_non_match_count = non_match_histo_dict.get(x_val, 0)
line_str = '%9d |%9d | %5.2f |' % (this_match_count,
this_non_match_count, x_val)
else:
this_match_count = match_histo_dict.get(x_val, 0)
this_non_match_count = non_match_histo_dict.get(x_val, 0)
this_poss_match_count = poss_match_histo_dict.get(x_val, 0)
line_str = '%9d |%9d |%9d | %5.2f |' % (this_match_count,
this_non_match_count,
this_poss_match_count, x_val)
line_str += '*'*int(this_count*scale_factor_y)
histo_list.append(line_str)
histo_list.append('')
# If a file name is given open it for writing - - - - - - - - - - - - - - - -
#
if (file_name != None):
try:
f = open(file_name, 'w')
except:
logging.exception('Cannot open file "%s" for writing' % (str(file_name)))
raise IOError
for line in histo_list:
f.write(line + os.linesep)
f.close()
logging.info('Histogram written to file: %s' % (file_name))
if (match_sets != None):
print match_histo_dict.items()
print non_match_histo_dict.items()
return histo_list
def GenerateHistogram(w_vec_dict, bin_width, file_name=None, match_sets=None):
"""Print and/or save a histogram of the weight vectors stored in the given
dictionary, and according to the match sets (if given).
The histogram is rotated 90 degrees clockwise, i.e. up to down instead of
left to right.
This function sums up the number of weight vectors with a matching weight
in a given bin (according to the given bin width).
If given, the match sets must be a tuple containing three sets, the first
being a set with matches, the second with non-matches, and the third with
possible matches, as generated by classifiers in the classification.py
Febrl module.
For each bin, the number of weight vectors in this bin is printed as well,
and if the match sets are given the number of matches, non-matches and
possible matches in this bin.
If a file name is given, the output will be written into this text file.
This function returns a list of containing the histogram as text strings.
"""
MAX_HISTO_WIDTH = 80 # maximum width in characters
auxiliary.check_is_dictionary('w_vec_dict', w_vec_dict)
auxiliary.check_is_number('bin_width', bin_width)
auxiliary.check_is_positive('bin_width', bin_width)
if (file_name != None):
auxiliary.check_is_string('file_name', file_name)
if (match_sets != None):
auxiliary.check_is_tuple('match_sets', match_sets)
if (len(match_sets) != 3):
logging.exception(
'Match sets must be a tuple containing three sets.')
raise Exception
auxiliary.check_is_set('match_sets[0]', match_sets[0])
auxiliary.check_is_set('match_sets[1]', match_sets[1])
auxiliary.check_is_set('match_sets[2]', match_sets[2])
if (len(w_vec_dict) != (len(match_sets[0]) + len(match_sets[1]) + \
len(match_sets[2]))):
logging.exception('Lengths of weight vector dictionary differs from' + \
'summed lengths of match sets.')
raise Exception
# Check if weight vector dictionary is empty, if so return empty list
#
if (w_vec_dict == {}):
logging.warn('Empty weight vector dictionary given for histogram ' + \
'generation')
return []
# Get a random vector dictionary element to get dimensionality of vectors
#
(rec_id_tuple, w_vec) = w_vec_dict.popitem()
v_dim = len(w_vec)
w_vec_dict[rec_id_tuple] = w_vec # Put back in
histo_dict = {} # A combined histogram dictionary
if (match_sets != None): # Also matches, non-matches and possible matches
match_histo_dict = {}
non_match_histo_dict = {}
poss_match_histo_dict = {}
max_bin_w_count = -1 # Maximal count for one binned weight entry
# Loop over weight vectors - - - - - - - - - - - - - - - - - - - - - - - - -
#
for (rec_id_tuple, w_vec) in w_vec_dict.iteritems():
w_sum = sum(w_vec) # Sum all weight vector elements
binned_w = w_sum - (w_sum % bin_width)
binned_w_count = histo_dict.get(binned_w,
0) + 1 # Increase count by one
histo_dict[binned_w] = binned_w_count
if (binned_w_count >
max_bin_w_count): # Check if this is new maximum count
max_bin_w_count = binned_w_count
if (match_sets != None):
if (rec_id_tuple in match_sets[0]):
binned_w_count = match_histo_dict.get(binned_w, 0) + 1
match_histo_dict[binned_w] = binned_w_count
elif (rec_id_tuple in match_sets[1]):
binned_w_count = non_match_histo_dict.get(binned_w, 0) + 1
non_match_histo_dict[binned_w] = binned_w_count
else: # A possible match
binned_w_count = poss_match_histo_dict.get(binned_w, 0) + 1
poss_match_histo_dict[binned_w] = binned_w_count
# Sort histogram according to X axis values - - - - - - - - - - - - - - - - -
#
x_vals = histo_dict.keys()
x_vals.sort()
assert sum(histo_dict.values()) == len(w_vec_dict)
if (match_sets == None): # Can use 68 characters for histogram
scale_factor_y = float(MAX_HISTO_WIDTH - 19) / max_bin_w_count
elif (len(poss_match_histo_dict) == 0): # No possible matches
scale_factor_y = float(MAX_HISTO_WIDTH - 30) / max_bin_w_count
else: # All three set non-empty
scale_factor_y = float(MAX_HISTO_WIDTH - 41) / max_bin_w_count
# Generate the histogram as a list of strings - - - - - - - - - - - - - - - -
#
histo_list = []
histo_list.append('Weight histogram:')
histo_list.append('-----------------')
if (match_sets == None):
histo_list.append(' Counts | w_sum |')
histo_list.append('-------------------')
elif (len(poss_match_histo_dict) == 0): # No possible matches
histo_list.append(' Counts |')
histo_list.append(' Match | Non-Match| w_sum |')
histo_list.append('------------------------------')
else:
histo_list.append(' Counts |')
histo_list.append(' Match | Non-Match|Poss-Match| w_sum |')
histo_list.append('-----------------------------------------')
for x_val in x_vals:
this_count = histo_dict[x_val]
if (match_sets == None):
line_str = '%9d | %5.2f |' % (this_count, x_val)
elif (len(poss_match_histo_dict) == 0): # No possible matches
this_match_count = match_histo_dict.get(x_val, 0)
this_non_match_count = non_match_histo_dict.get(x_val, 0)
line_str = '%9d |%9d | %5.2f |' % (this_match_count,
this_non_match_count, x_val)
else:
this_match_count = match_histo_dict.get(x_val, 0)
this_non_match_count = non_match_histo_dict.get(x_val, 0)
this_poss_match_count = poss_match_histo_dict.get(x_val, 0)
line_str = '%9d |%9d |%9d | %5.2f |' % (
this_match_count, this_non_match_count, this_poss_match_count,
x_val)
line_str += '*' * int(this_count * scale_factor_y)
histo_list.append(line_str)
histo_list.append('')
# If a file name is given open it for writing - - - - - - - - - - - - - - - -
#
if (file_name != None):
try:
f = open(file_name, 'w')
except:
logging.exception('Cannot open file "%s" for writing' %
(str(file_name)))
raise IOError
for line in histo_list:
f.write(line + os.linesep)
f.close()
logging.info('Histogram written to file: %s' % (file_name))
if (match_sets != None):
print match_histo_dict.items()
print non_match_histo_dict.items()
return histo_list
def pairs_completeness(weight_vec_dict, dataset1, dataset2, get_id_funct,
match_check_funct):
"""Pairs completeness is measured as
pc = Nm / M
with Nm (<= M) being the number of correctly classified truly matched
record pairs in the blocked comparison space, and M the total number of
true matches.
If both data sets are the same a deduplication is assumed, otherwise a
linkage.
The arguments that have to be set when this method is called are:
weight_vec_dict A dictionary containing weight vectors.
dataset1 The initialised first data set object.
dataset2 The initialised second data set object.
get_id_funct This has to be a function (or method), assumed to
have argument a record (assumed to be a list fo field
values), and returns the record identifier from that
record.
match_check_funct This has to be a function (or method), assumed to
have as arguments the two record identifiers of a
record pair and its weight vector, and returns True
if the record pair is from a true match, or False
otherwise. Thus, 'match_check_funct' is of the form:
match_flag = match_check_funct(rec_id1, rec_id2,
weight_vec)
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_not_none('dataset1', dataset1)
auxiliary.check_is_not_none('dataset2', dataset2)
auxiliary.check_is_function_or_method('get_id_funct', get_id_funct)
auxiliary.check_is_function_or_method('match_check_funct',
match_check_funct)
# Check if a deduplication will be done or a linkage - - - - - - - - - - - -
#
if (dataset1 == dataset2):
do_dedup = True
else:
do_dedup = False
logging.info('')
logging.info('Calculate pairs completeness:')
logging.info(' Data set 1: %s (containing %d records)' % \
(dataset1.description, dataset1.num_records))
if (do_dedup == True):
logging.info(' Data sets are the same: Deduplication')
else:
logging.info(' Data set 2: %s (containing %d records)' % \
(dataset2.description, dataset2.num_records))
logging.info(' Data sets differ: Linkage')
logging.info(' Number of record pairs in weight vector dictionary: %d' % \
(len(weight_vec_dict)))
num_all_true_matches = 0 # Count the total number of all true matches
# For a deduplication only process data set 1 - - - - - - - - - - - - - - - -
#
if (do_dedup == True):
# Build a dictionary with entity identifiers as keys and a list of their
# record identifier (rec_ident) as values
#
entity_ident_dict = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 1: %d' % \
(len(entity_ident_dict)))
for (ent_id, rec_list) in entity_ident_dict.iteritems():
num_this_rec = len(rec_list)
if (num_this_rec > 1):
num_all_true_matches += num_this_rec * (num_this_rec - 1) / 2
# More efficent version: Only count number of matches ber record don't
# store them
#
entity_ident_dict2 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict2.get(ent_id, 0) + 1
entity_ident_dict2[ent_id] = ent_id_count
assert sum(entity_ident_dict2.values()) == dataset1.num_records
tm = 0 # Total number of true matches (without indexing)
for (ent_id, ent_count) in entity_ident_dict2.iteritems():
tm += ent_count * (ent_count - 1) / 2
assert num_all_true_matches == tm
else: # For a linkage - - - - - - - - - - - - - - - - - - - - - - - - - - -
# Build two dictionaries with entity identifiers as keys and a list of
# their record identifier (rec_ident) as values
#
entity_ident_dict1 = {}
entity_ident_dict2 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict1.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict1[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 1: %d' % \
(len(entity_ident_dict1)))
for (rec_ident, rec) in dataset2.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict2.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict2[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 2: %d' % \
(len(entity_ident_dict2)))
# Now calculate total true match number (loop over smaller dict)
#
if (len(entity_ident_dict1) < len(entity_ident_dict2)):
for (ent_id1, rec_list1) in entity_ident_dict1.iteritems():
if (ent_id1 in entity_ident_dict2):
rec_list2 = entity_ident_dict2[ent_id1]
num_all_true_matches += len(rec_list1) * len(rec_list2)
else:
for (ent_id2, rec_list2) in entity_ident_dict2.iteritems():
if (ent_id2 in entity_ident_dict1):
rec_list1 = entity_ident_dict1[ent_id2]
num_all_true_matches += len(rec_list1) * len(rec_list2)
# More efficent version: Only count number of matches ber record don't
# store them
#
entity_ident_dict3 = {}
entity_ident_dict4 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict3.get(ent_id, 0) + 1
entity_ident_dict3[ent_id] = ent_id_count
for (rec_ident, rec) in dataset2.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict4.get(ent_id, 0) + 1
entity_ident_dict4[ent_id] = ent_id_count
assert sum(entity_ident_dict3.values()) == dataset1.num_records
assert sum(entity_ident_dict4.values()) == dataset2.num_records
tm = 0 # Total number of true matches (without indexing)
if (len(entity_ident_dict3) < len(entity_ident_dict4)):
for (ent_id, ent_count) in entity_ident_dict3.iteritems():
if ent_id in entity_ident_dict4:
tm += ent_count * entity_ident_dict4[ent_id]
else:
for (ent_id, ent_count) in entity_ident_dict4.iteritems():
if ent_id in entity_ident_dict3:
tm += ent_count * entity_ident_dict3[ent_id]
assert num_all_true_matches == tm
logging.info(' Number of all true matches: %d' % (num_all_true_matches))
# Get number of true matches in weight vector dictionary - - - - - - - - - -
#
num_true_matches = 0
num_false_matches = 0
for (rec_id_tuple, this_vec) in weight_vec_dict.iteritems():
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1],
this_vec) == True):
num_true_matches += 1
else:
num_false_matches += 1
assert len(weight_vec_dict) == num_true_matches + num_false_matches
logging.info(' Number of true and false matches in weight vector ' + \
'dictionary: %d / %d' % (num_true_matches,num_false_matches))
if (num_all_true_matches > 0):
pc = float(num_true_matches) / float(num_all_true_matches)
logging.info(' Pairs completeness: %.4f%%' %
(100.0 * pc)) # As percentage
else:
pc = 0.0
logging.info(' No true matches - cannot calculate pairs completeness')
assert pc <= 1.0, pc
return pc
def quality_measures(weight_vec_dict, match_set, non_match_set,
match_check_funct):
"""Calculate several quality measures based on the number of true positives,
true negatives, false positives and false negatives in the given match
and non-match sets and weight vector dictionary using the given match
check function.
The function calculates and returns:
- Accuracy: (|TP|+|TN|)
---------------------
(|TP|+|TN|+|FP|+|FN|)
- Precision: |TP|
-----------
(|TP|+|FP|)
- Recall: |TP|
-----------
(|TP|+|FN|)
- F-Measure: 2 * (Precision * Recall)
--------------------------
(Precision + Recall)
With TP the True Positives, TN the True negatives, FP the False Positives
and FN the False Negatives.
For a discussion about measuring data linkage and deduplication quality
please refer to:
Quality and Complexity Measures for Data Linkage and Deduplication
Peter Christen and Karl Goiser
Book chapter in "Quality Measures in Data Mining"
Studies in Computational Intelligence, Vol. 43
F. Guillet and H. Hamilton (eds), Springer
March 2007.
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_set('match set', match_set)
auxiliary.check_is_set('non match set', non_match_set)
auxiliary.check_is_function_or_method('match_check_funct',
match_check_funct)
if ((len(match_set) + len(non_match_set)) != len(weight_vec_dict)):
logging.exception('Match and non-match set are not of same length as ' + \
'weight vector dictionary: %d, %d / %d' % \
(len(match_set),len(non_match_set),len(weight_vec_dict)))
raise Exception
tp = 0.0
fp = 0.0
tn = 0.0
fn = 0.0
for rec_id_tuple in match_set:
w_vec = weight_vec_dict[rec_id_tuple]
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1],
w_vec) == True):
tp += 1
else:
fp += 1
for rec_id_tuple in non_match_set:
w_vec = weight_vec_dict[rec_id_tuple]
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1],
w_vec) == False):
tn += 1
else:
fn += 1
logging.info('')
logging.info('Classification results: TP=%d, FP=%d / TN=%d, FN=%d' % \
(tp, fp, tn, fn))
if ((tp != 0) or (fp != 0) or (tn != 0) or (fn != 0)):
acc = (tp + tn) / (tp + fp + tn + fn)
else:
acc = 0.0
if ((tp != 0) or (fp != 0)):
prec = tp / (tp + fp)
else:
prec = 0.0
if ((tp != 0) or (fn != 0)):
reca = tp / (tp + fn)
else:
reca = 0.0
if ((prec != 0.0) or (reca != 0.0)):
fmeas = 2 * (prec * reca) / (prec + reca)
else:
fmeas = 0.0
logging.info('Quality measures:')
logging.info(' Accuracy: %.6f Precision:%.4f Recall: %.4f ' % \
(acc, prec, reca)+'F-measure: %.4f' % (fmeas))
return acc, prec, reca, fmeas
def testIsDictionary(self): # - - - - - - - - - - - - - - - - - - - - - - -
"""Test 'check_is_dictionary' function."""
assert auxiliary.check_is_dictionary("TestArgument", {}) == None
assert auxiliary.check_is_dictionary("TestArgument", {1: 2, 6: 0}) == None
assert auxiliary.check_is_dictionary("TestArgument", {"a": 4, "t": 1, (1, 4, 6): "tr"}) == None
def pairs_completeness(weight_vec_dict, dataset1, dataset2, get_id_funct,
match_check_funct):
"""Pairs completeness is measured as
pc = Nm / M
with Nm (<= M) being the number of correctly classified truly matched
record pairs in the blocked comparison space, and M the total number of
true matches.
If both data sets are the same a deduplication is assumed, otherwise a
linkage.
The arguments that have to be set when this method is called are:
weight_vec_dict A dictionary containing weight vectors.
dataset1 The initialised first data set object.
dataset2 The initialised second data set object.
get_id_funct This has to be a function (or method), assumed to
have argument a record (assumed to be a list fo field
values), and returns the record identifier from that
record.
match_check_funct This has to be a function (or method), assumed to
have as arguments the two record identifiers of a
record pair and its weight vector, and returns True
if the record pair is from a true match, or False
otherwise. Thus, 'match_check_funct' is of the form:
match_flag = match_check_funct(rec_id1, rec_id2,
weight_vec)
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_not_none('dataset1', dataset1)
auxiliary.check_is_not_none('dataset2', dataset2)
auxiliary.check_is_function_or_method('get_id_funct', get_id_funct)
auxiliary.check_is_function_or_method('match_check_funct', match_check_funct)
# Check if a deduplication will be done or a linkage - - - - - - - - - - - -
#
if (dataset1 == dataset2):
do_dedup = True
else:
do_dedup = False
logging.info('')
logging.info('Calculate pairs completeness:')
logging.info(' Data set 1: %s (containing %d records)' % \
(dataset1.description, dataset1.num_records))
if (do_dedup == True):
logging.info(' Data sets are the same: Deduplication')
else:
logging.info(' Data set 2: %s (containing %d records)' % \
(dataset2.description, dataset2.num_records))
logging.info(' Data sets differ: Linkage')
logging.info(' Number of record pairs in weight vector dictionary: %d' % \
(len(weight_vec_dict)))
num_all_true_matches = 0 # Count the total number of all true matches
# For a deduplication only process data set 1 - - - - - - - - - - - - - - - -
#
if (do_dedup == True):
# Build a dictionary with entity identifiers as keys and a list of their
# record identifier (rec_ident) as values
#
entity_ident_dict = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 1: %d' % \
(len(entity_ident_dict)))
for (ent_id, rec_list) in entity_ident_dict.iteritems():
num_this_rec = len(rec_list)
if (num_this_rec > 1):
num_all_true_matches += num_this_rec*(num_this_rec-1)/2
# More efficent version: Only count number of matches ber record don't
# store them
#
entity_ident_dict2 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict2.get(ent_id, 0) + 1
entity_ident_dict2[ent_id] = ent_id_count
assert sum(entity_ident_dict2.values()) == dataset1.num_records
tm = 0 # Total number of true matches (without indexing)
for (ent_id, ent_count) in entity_ident_dict2.iteritems():
tm += ent_count*(ent_count-1)/2
assert num_all_true_matches == tm
else: # For a linkage - - - - - - - - - - - - - - - - - - - - - - - - - - -
# Build two dictionaries with entity identifiers as keys and a list of
# their record identifier (rec_ident) as values
#
entity_ident_dict1 = {}
entity_ident_dict2 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict1.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict1[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 1: %d' % \
(len(entity_ident_dict1)))
for (rec_ident, rec) in dataset2.readall():
ent_id = get_id_funct(rec)
this_rec_list = entity_ident_dict2.get(ent_id, [])
this_rec_list.append(rec_ident)
entity_ident_dict2[ent_id] = this_rec_list
logging.info(' Number of unique entity identifiers in data set 2: %d' % \
(len(entity_ident_dict2)))
# Now calculate total true match number (loop over smaller dict)
#
if (len(entity_ident_dict1) < len(entity_ident_dict2)):
for (ent_id1, rec_list1) in entity_ident_dict1.iteritems():
if (ent_id1 in entity_ident_dict2):
rec_list2 = entity_ident_dict2[ent_id1]
num_all_true_matches += len(rec_list1) * len(rec_list2)
else:
for (ent_id2, rec_list2) in entity_ident_dict2.iteritems():
if (ent_id2 in entity_ident_dict1):
rec_list1 = entity_ident_dict1[ent_id2]
num_all_true_matches += len(rec_list1) * len(rec_list2)
# More efficent version: Only count number of matches ber record don't
# store them
#
entity_ident_dict3 = {}
entity_ident_dict4 = {}
for (rec_ident, rec) in dataset1.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict3.get(ent_id, 0) + 1
entity_ident_dict3[ent_id] = ent_id_count
for (rec_ident, rec) in dataset2.readall():
ent_id = get_id_funct(rec)
ent_id_count = entity_ident_dict4.get(ent_id, 0) + 1
entity_ident_dict4[ent_id] = ent_id_count
assert sum(entity_ident_dict3.values()) == dataset1.num_records
assert sum(entity_ident_dict4.values()) == dataset2.num_records
tm = 0 # Total number of true matches (without indexing)
if (len(entity_ident_dict3) < len(entity_ident_dict4)):
for (ent_id, ent_count) in entity_ident_dict3.iteritems():
if ent_id in entity_ident_dict4:
tm += ent_count*entity_ident_dict4[ent_id]
else:
for (ent_id, ent_count) in entity_ident_dict4.iteritems():
if ent_id in entity_ident_dict3:
tm += ent_count*entity_ident_dict3[ent_id]
assert num_all_true_matches == tm
logging.info(' Number of all true matches: %d' % (num_all_true_matches))
# Get number of true matches in weight vector dictionary - - - - - - - - - -
#
num_true_matches = 0
num_false_matches = 0
for (rec_id_tuple, this_vec) in weight_vec_dict.iteritems():
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1], this_vec) == True):
num_true_matches += 1
else:
num_false_matches += 1
assert len(weight_vec_dict) == num_true_matches+num_false_matches
logging.info(' Number of true and false matches in weight vector ' + \
'dictionary: %d / %d' % (num_true_matches,num_false_matches))
if (num_all_true_matches > 0):
pc = float(num_true_matches) / float(num_all_true_matches)
logging.info(' Pairs completeness: %.4f%%' % (100.0*pc)) # As percentage
else:
pc = 0.0
logging.info(' No true matches - cannot calculate pairs completeness')
assert pc <= 1.0, pc
return pc
def quality_measures(weight_vec_dict, match_set, non_match_set,
match_check_funct):
"""Calculate several quality measures based on the number of true positives,
true negatives, false positives and false negatives in the given match
and non-match sets and weight vector dictionary using the given match
check function.
The function calculates and returns:
- Accuracy: (|TP|+|TN|)
---------------------
(|TP|+|TN|+|FP|+|FN|)
- Precision: |TP|
-----------
(|TP|+|FP|)
- Recall: |TP|
-----------
(|TP|+|FN|)
- F-Measure: 2 * (Precision * Recall)
--------------------------
(Precision + Recall)
With TP the True Positives, TN the True negatives, FP the False Positives
and FN the False Negatives.
For a discussion about measuring data linkage and deduplication quality
please refer to:
Quality and Complexity Measures for Data Linkage and Deduplication
Peter Christen and Karl Goiser
Book chapter in "Quality Measures in Data Mining"
Studies in Computational Intelligence, Vol. 43
F. Guillet and H. Hamilton (eds), Springer
March 2007.
"""
auxiliary.check_is_dictionary('weight_vec_dict', weight_vec_dict)
auxiliary.check_is_set('match set', match_set)
auxiliary.check_is_set('non match set', non_match_set)
auxiliary.check_is_function_or_method('match_check_funct', match_check_funct)
if ((len(match_set) + len(non_match_set)) != len(weight_vec_dict)):
logging.exception('Match and non-match set are not of same length as ' + \
'weight vector dictionary: %d, %d / %d' % \
(len(match_set),len(non_match_set),len(weight_vec_dict)))
raise Exception
tp = 0.0
fp = 0.0
tn = 0.0
fn = 0.0
for rec_id_tuple in match_set:
w_vec = weight_vec_dict[rec_id_tuple]
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1], w_vec) == True):
tp += 1
else:
fp += 1
for rec_id_tuple in non_match_set:
w_vec = weight_vec_dict[rec_id_tuple]
if (match_check_funct(rec_id_tuple[0], rec_id_tuple[1], w_vec) == False):
tn += 1
else:
fn += 1
logging.info('')
logging.info('Classification results: TP=%d, FP=%d / TN=%d, FN=%d' % \
(tp, fp, tn, fn))
if ((tp != 0) or (fp != 0) or (tn != 0) or (fn != 0)):
acc = (tp + tn) / (tp + fp + tn + fn)
else:
acc = 0.0
if ((tp != 0) or (fp != 0)):
prec = tp / (tp + fp)
else:
prec = 0.0
if ((tp != 0) or (fn != 0)):
reca = tp / (tp + fn)
else:
reca = 0.0
if ((prec != 0.0) or (reca != 0.0)):
fmeas = 2*(prec*reca) / (prec+reca)
else:
fmeas = 0.0
logging.info('Quality measures:')
logging.info(' Accuracy: %.6f Precision:%.4f Recall: %.4f ' % \
(acc, prec, reca)+'F-measure: %.4f' % (fmeas))
return acc, prec, reca, fmeas
