for writing_type, source in (('email', urls1), ('forum', urls2)):
for i,url in enumerate(source,1):
print ' Downloading %s %i/%i' % (writing_type, i, len(source))
io.mg_insert_one( MG_COLL_NAME,
{ 'author': '' }, # dummy query
{ 'author': author,
'type': writing_type,
'content': requests.get(ROOT_URL + url).content },
verbose=True)
# sanity check:
print 'No. of e-mails \t', io.mg_find(MG_COLL_NAME,
{ 'author': 'satoshi',
'type' : { '$eq': 'email' } } ).count()
print 'No. of posts \t', io.mg_find(MG_COLL_NAME,
{ 'author': 'satoshi',
'type' : { '$eq': 'forum' } } ).count()
print '\n'
##--- Parse HTML for content ---
MG_SRC_COLL = MG_COLL_NAME
MG_DEST_COLL = 'raw-docs'
html_parser = HTMLParser.HTMLParser()
class MySVCModel(object):
def __init__(self, num_class=2):
"""
:type num_classes: int
:rtype: None
"""
self.__ctrl__ = None
self.__case__ = None
with open('../../.dbname', 'r') as f:
self.__DB_NAME__ = json.load(f)['dbname']
self.__MG_DOCS_COLL__ = 'raw-docs' # raw docs
self.__MG_SENTS_COLL__ = 'bag-of-sents' # raw sentences
self.__MG_TOKENS_COLL__ = 'sample-tokens' # clean tokens (words)
self.__PG_STATS_TBL__ = 'stats' # stylometric features
self.__PG_RESULTS_TBL__ = 'results_' + \
str(num_class) + \
'class' # cross val results
self.__PG_PROBAS_TBL__ = 'probabilities' # cross val probabilities
self.__model__ = Pipeline([ \
# ('scaler2', StandardScaler()),
# ('scaler', MinMaxScaler()),
# ('scaler3', Normalizer()),
('classifier', SVC(probability=True,
kernel='poly',
degree=2,
class_weight='balanced') \
if num_class-1 \
else OneClassSVM(kernel='rbf',
nu=0.7,
gamma=1./250))
])
print 'Instantiated classifier %s.' % \
self.__model__.named_steps['classifier'].__class__.__name__
self.__io__ = DBIO(MG_DB_NAME=self.__DB_NAME__,
PG_DB_NAME=self.__DB_NAME__)
self.__tagger__ = None # initialise if re-creating samples
self.__bootstrap__ = None # initialise in fit
def fit(self, author1, author2, wts1=None, wts2=None,
bootstrap=False, verbose=False):
"""
:type author1: str
:type author2: str
:type wts1: str/List[str]
:type wts2: str/List[str]
:type verbose:bool
:rtype: bool
:
: Prepares databases and tables/collections.
:
"""
self.__bootstrap__ = bootstrap
cases = []
for i, (author, wts) in enumerate([(author1, wts1), (author2, wts2)]):
if not wts:
wts = [wt.encode('ascii') \
for wt in self.__io__.mg_distinct(self.__MG_DOCS_COLL__,
'type',
{ 'author':author } )]
if not isinstance(wts, list):
wts = [wts]
cases += (author, wts, (1,-1)[i]), # use 1, -1 to match output
# from sklearn's OneClassSVM
self.__ctrl__ = cases[0] # assign label 1 in y vector
self.__case__ = cases[1] # assign be label 0 in y vector
self.__MG_TOKENS_COLL__ += '-' + cases[0][0] + \
'-' + cases[1][0] + \
'-' + \
''.join(wt[:3] for wt in cases[0][1]) + \
'-' + \
''.join(wt[:3] for wt in cases[1][1]) + \
'-' + \
('nobs','bs')[bootstrap]
self.__PG_STATS_TBL__ += '_' + cases[0][0] + \
'_' + cases[1][0] + \
'_' + \
''.join(wt[:3] for wt in cases[0][1]) + \
'_' + \
''.join(wt[:3] for wt in cases[1][1]) + \
'_' + \
('nobs','bs')[bootstrap]
if verbose:
print 'Control:', self.__ctrl__
print 'Case: ', self.__case__
print 'Saving tokens to', self.__MG_TOKENS_COLL__
print 'Saving stats to', self.__PG_STATS_TBL__
return self.__prep_sents__(verbose=verbose) # err in preparing sentences
def __prep_sents__(self, verbose=False):
"""
:type verbose: bool
:rtype: bool
:
: Prepares bags of raw sentences if not already done so.
: (lump docs by the same auther and of the same type into the same bag)
:
"""
for author, wts, _ in (self.__ctrl__, self.__case__):
for wt in wts: # process type by type so sents can be reused
if verbose:
print "Checking for %s's %s sentences..." % (author, wt)
query_dic = { 'author' : author,
'type' : wt }
if self.__io__.mg_find(self.__MG_SENTS_COLL__,
query_dic).count() == 0:
print " No sentences found. Preparing from raw docs..."
query_results = self.__io__.mg_find(self.__MG_DOCS_COLL__,
query_dic)
# gather raw lines
contents = []
for result in query_results:
if isinstance(result['content'], list):
contents += result['content'] # list of lines
else:
contents += result['content'], # entire doc = a str
if not contents:
print ' Error: No docs found.'
return False
# create a bag of sentences from contents
sentences = bag_of_sentences(contents)
if not self.__io__.mg_insert_one(self.__MG_SENTS_COLL__,
query_dic,
dict({'sentences':sentences},
**query_dic),
verbose=True):
print 'Error in saving sentences.'
return False
return True
def cross_validate(self, chunk_size=500, k=3,
verbose=False,
recreate_samples=False, plot_samples=False):
"""
:type chunk_size: int
:type k: int
:type verbose: bool
:type recreate_samples: bool
:type plot_samples: bool
:rtype: bool
:
: Perform k-fold cross validation.
:
"""
# create results table if not exist
pg_header_dic = {'author1': 'varchar(20)',
'wts1' : 'text[]',
'author2': 'varchar(20)',
'wts2' : 'text[]',
'bs' : 'boolean',
'fold' : 'int' }
# create table to store stylometric results if not already exist
if isinstance(self.__io__.pg_find("""
SELECT COUNT(*) FROM %s
""" % self.__PG_RESULTS_TBL__),
type(None)):
self.__io__.recreate_tables(PG_TBL_NAME = self.__PG_RESULTS_TBL__,
pg_header_dic = pg_header_dic,
verbose=True)
# create table to store prediction probabilities if not already exist
if self.__model__.named_steps['classifier'].__class__.__name__ not in \
set(['OneClassSVM', 'LinearSVC']) and \
isinstance(self.__io__.pg_find("""
SELECT COUNT(*) FROM %s
""" % self.__PG_PROBAS_TBL__),
type(None)):
self.__io__.recreate_tables(PG_TBL_NAME = self.__PG_PROBAS_TBL__,
pg_header_dic = dict({'uid': 'int'},
**pg_header_dic),
verbose=True)
# create samples if not already exist or recreate samples when forced to
if self.__io__.mg_find(self.__MG_TOKENS_COLL__, {}).count() == 0 or \
recreate_samples:
print 'Initialising Spacy POS tagger...'
self.__tagger__ = English() # spacy tagger
# recreate tables (role = "trian", "test")
self.__io__.recreate_tables(MG_COLL_NAME = self.__MG_TOKENS_COLL__,
PG_TBL_NAME = self.__PG_STATS_TBL__,
pg_header_dic={'fold' : 'int',
'role' : 'varchar(5)',
'uid' : 'int',
'author': 'varchar(20)',
'type' : 'text[]'},
verbose=True)
# split raw sentences into folds
print 'Splitting folds...'
folds_dic = self.__create_folds__(k, verbose=verbose)
if verbose:
for key, item in folds_dic.iteritems():
print ' %s: %i folds (%s sentences)' % \
(key, len(item), [len(lst) for lst in item])
# create chunks then clean, tokenize and count words in each chunk
print 'Creating and cleaning chunks...'
if not self.__create_samples__(folds_dic, k, chunk_size,
bootstrap=self.__bootstrap__,
verbose=verbose):
print 'Error in creating samples.'
return False
for fold in xrange(k):
# create feature matrix
print 'Preparing feature matrix.'
dfs = self.__prepare_df__(fold, verbose=verbose)
if plot_samples:
print 'Hello Samples :)'
self.__plot_samples__(dfs, fold)
X_train = dfs[0].drop('label', axis=1)
y_train = dfs[0]['label']
X_test = dfs[1].drop('label', axis=1)
y_true = dfs[1]['label']
# train model
print 'Fit and predict...'
self.__model__.fit(X_train, y_train)
y_pred = self.__model__.predict(X_test)
error = False
with warnings.catch_warnings(record=True) as w:
# compute metrics
CM = confusion_matrix(y_true, y_pred)
f1 = f1_score(y_true, y_pred)
accuracy = accuracy_score(y_true, y_pred)
recall = recall_score(y_true, y_pred)
precision = average_precision_score(y_true, y_pred)
if len(w): # caught f1 UndefinedMetricsWarning from sklearn
print w[-1].message
error = True
print ' f1 score :', f1
print ' accuracy score :', accuracy
print ' recall score :', recall
print ' area under the precision-recall curve:', precision
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
probas = self.__model__.predict_proba(X_test)
print ' probabilities:\n', probas
# save results to postgres
header_dic = {'author1' : self.__ctrl__[0],
'wts1' : self.__ctrl__[1],
'author2' : self.__case__[0],
'wts2' : self.__case__[1],
'bs' : self.__bootstrap__,
'fold' : fold }
# CM: row = true, col = pred, ind 0 = case, ind 1 = ctrl
detail_dic = { 'tp': CM[1][1], # ctrl classified as ctrl
'fp': CM[0][1],
'fn': CM[1][0], # ctrl classified as case
'tn': CM[0][0],
'f1' : f1,
'accuracy': accuracy,
'recall' : recall,
'avg_prec': precision,
'und_err' : error }
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
detail_dic.update({ 'pdiff_mean': np.mean(abs(probas[:,0] - \
probas[:,1])),
'pdiff_std' : np.std(abs(probas[:,0] - \
probas[:,1])) })
if not self.__io__.pg_insert_one(self.__PG_RESULTS_TBL__,
header_dic,
detail_dic,
verbose=True):
print 'Error in saving results.'
return False
# store probablilities if using two-class SVC
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
for i, row in enumerate(probas):
if not self.__io__.pg_insert_one(self.__PG_PROBAS_TBL__,
dict({ 'uid' : i }, **header_dic),
detail_dic = { 'y_pred': y_pred[i],
'case_0': row[0],
'ctrl_1': row[1] },
verbose=True):
print 'Error in saving results.'
return False
return True
def __create_folds__(self, k=3, verbose=False):
"""
:type k: int
:type verbose: bool
:rtype: dict
:
: Divide sentences into k folds.
:
"""
folds_dic, target_count = {}, 0
for author, wts, _ in (self.__ctrl__, self.__case__):
query_dic = { 'author' : author,
'type' : { '$in': wts } }
query_results = self.__io__.mg_find(self.__MG_SENTS_COLL__,
query_dic)
sentences = [result['sentences'] for result in query_results]
sentences = list(chain(*sentences))
if verbose:
print ' (%s, %s): %i sentences' % (author, wts, len(sentences))
# randomly shuffle sentences
shuffle(sentences) # in place shuffle
folds_dic[(author, str(wts))] = [sentences[i::k] for i in xrange(k)]
return folds_dic
def __create_samples__(self, folds_dic, k=3, chunk_size=500,
bootstrap=False, verbose=False):
"""
:type folds_dic: dict
:type k: int
:type chunk_size: int
:type bootstrap: bool
:type verbose: bool
:rtype: bool
:
: Create a sample out of each chunk in each fold.
:
"""
# process fold by fold
for fold in xrange(k):
print ' Starting fold %i...' % fold
# create training chunks
chunks_train_dic, chunks_test_dic = {}, {}
for author, wts, _ in (self.__ctrl__, self.__case__):
chunks_train_dic[(author, str(wts))] = \
create_chunks(list(chain(*folds_dic[(author, str(wts))][1:])),
chunk_size,
0,
verbose=False)
# compute amount to bootstrap
target_count = max(len(chunks) \
for chunks in chunks_train_dic.values())
for author, wts, _ in (self.__ctrl__, self.__case__):
# bootstrap minority class to 50/50
if bootstrap and \
len(chunks_train_dic[(author, str(wts))]) < target_count:
if verbose:
print ' Bootstrap minority to %i chunks.' % target_count
chunks_train_dic[(author, str(wts))] = \
create_chunks(list(chain(*folds_dic[(author,
str(wts))][1:])),
chunk_size,
target_count,
verbose=False)
# do not bootstrap test
chunks_test_dic[(author, str(wts))] = \
create_chunks(folds_dic[(author, str(wts))][0],
chunk_size,
0,
verbose=False)
# prepare folds_dic for next fold-iteration
folds_dic[(author, str(wts))] = \
folds_dic[(author, str(wts))][1:] + \
folds_dic[(author, str(wts))][:1]
if verbose:
for author, wts, _ in (self.__ctrl__, self.__case__):
print ' (%s, %s): %i train, %i test' % \
(author,
wts,
len(chunks_train_dic[(author, str(wts))]),
len(chunks_test_dic[(author, str(wts))]))
# Clean, tokenize and compute stats for each chunk
for author, wts, _ in (self.__ctrl__, self.__case__):
for role, chunks \
in [('train', chunks_train_dic[(author, str(wts))]),
('test', chunks_test_dic[(author, str(wts))])]:
for i, chunk in enumerate(chunks,1):
print ' Processing %s %s %s chunk %i/%i...' % \
(author, wts, role, i, len(chunks))
# clean chunks, tokenize and compute stats
stats_ctr, \
words, \
words_lemma, \
misspellings, \
gb_spellings, \
us_spellings = crunch_statistics(chunk, self.__tagger__)
query_dic = { 'fold' : fold,
'role' : role,
'uid' : i,
'author' : author,
'type' : wts }
entry_dic = { 'words': words,
'words_lemma': words_lemma,
'misspellings': misspellings,
'us_spellings': us_spellings,
'gb_spellings': gb_spellings }
# save words in mongodb
if not self.__io__.mg_insert_one(self.__MG_TOKENS_COLL__,
query_dic,
dict(query_dic,
**entry_dic),
verbose=True):
print 'Error in saving words.'
return False
# save stats in postgres
if not self.__io__.pg_insert_one(self.__PG_STATS_TBL__,
query_dic,
stats_ctr,
verbose=True):
print 'Error in saving stats.'
return False
# break # break chunks
return True
def __prepare_df__(self, fold, verbose=False):
"""
:type fold: int
:type verbose: bool
:rtype: pandas.DataFrame
:
: Prepare feature matrix.
:
"""
# compute bag-of-word
query_results = self.__io__.mg_find( self.__MG_TOKENS_COLL__,
{'fold' : fold,
'role' : 'train',
'author': self.__ctrl__[0],
'type' : self.__ctrl__[1] } )
words_lists = [result['words_lemma'] for result in query_results]
words = list(chain(*words_lists))
bag_of_words = get_topN(words, 250).keys()
if verbose:
print ' No. of words lists:', len(words_lists)
print ' No. of lemmatized words:', len(words)
print ' No. of unique lemmatized words', len(set(words))
# prepare features dataframe for model
dfs = []
for role in ('train', 'test'):
df_stylo, df_freq = pd.DataFrame(), pd.DataFrame()
# get stylometric features
for author, wts, label in (self.__ctrl__, self.__case__):
# train one-class SVM with only literature from ctrl
if \
self.__model__.named_steps['classifier'].__class__.__name__ == \
'OneClassSVM' \
and role == 'train' \
and label == self.__case__[2]:
continue
# get term frequency features
q = '''
SELECT * FROM %s
WHERE fold = %i
AND role = '%s'
AND author = '%s'
AND type = ARRAY%s;
''' % (self.__PG_STATS_TBL__, fold, role, author, wts)
df = self.__io__.pg_find(q)
if isinstance(df, type(None)):
print 'Error: no stats found.'
return False
df_stylo = df_stylo.append(compute_stylo_features(df, label))
# get term frequency features
results = [results for results in \
self.__io__.mg_find( self.__MG_TOKENS_COLL__,
{ 'fold' : fold,
'role' : role,
'author': author,
'type' : wts } ) ]
df_freq = df_freq.append(compute_freqs_features(results,
bag_of_words,
label))
# horizontally stack DataFrames
dfs += df_stylo.join(df_freq.drop('label', axis=1)),
if verbose:
print ' df_train shape:', dfs[0].shape
print ' df_test shape:', dfs[1].shape
return dfs
def __plot_samples__(self, dfs, fold):
"""
:type dfs: List[pandas DataFrame] # [training df, testing df]
:type fold: int
:rtype: None
"""
mds = MDS(n_components=2, max_iter=3000, eps=1e-9, dissimilarity='euclidean', n_jobs=-1)
tsne = TSNE(n_components=2)
# change label to color index
# author 1 train (0 = light blue), author 1 test (1 = dark blue)
# author 2 train (2 = light green), author 2 test (3 = dark green)
df_all = pd.DataFrame(columns = dfs[0].columns)
df0_copy = dfs[0].copy()
df0_copy.loc[(df0_copy.label == 1).values, 'label'] = 0
df0_copy.loc[(df0_copy.label == -1).values, 'label'] = 2
df_all = df_all.append(df0_copy)
df1_copy = dfs[1].copy()
df1_copy.loc[(df1_copy.label == 1).values, 'label'] = 1
df1_copy.loc[(df1_copy.label == -1).values, 'label'] = 3
df_all = df_all.append(df1_copy)
legend = {0: 'Author 1 Training Sample',
1: 'Author 1 Test Sample',
2: 'Author 2 Training Sample' ,
3: 'Author 2 Test Sample' }
# fit on training data
pos_lst = [('Multi-Dimensional Scaling (MDS)',
mds.fit(df_all.drop('label', axis=1)).embedding_),
('t-Distributed Stochastic Neighbor Embedding (TSNE)',
tsne.fit(df_all.drop('label', axis=1)).embedding_)]
# plot
colors = sns.color_palette('Paired', 4)
fig = plt.figure(figsize=(16,7))
plt.hold(True)
for k, (title, pos) in enumerate(pos_lst, 1):
## fig.add_subplot() works in ipython notebook but creates a
## mysterious 3rd axes in python...
# ax = fig.add_subplot(1,2,k)
ax = plt.subplot(1,2,k)
ax.set_title(title)
for i in xrange(len(colors)):
samples = pos[(df_all.label == i).values, :]
ax.scatter(samples[:,0], samples[:,1],
c=colors[i], edgecolor='none',
label=legend[i])
ax.legend()
plt.hold(False)
plt.savefig('../figs/' + \
self.__PG_STATS_TBL__[self.__PG_STATS_TBL__.find("_")+1:] + \
'fold' + str(fold) + '.png',
dpi=300, transparent=True)
plt.close(fig)
print ' Downloading %s %i/%i' % (writing_type, i, len(source))
io.mg_insert_one(
MG_COLL_NAME,
{'author': ''}, # dummy query
{
'author': author,
'type': writing_type,
'content': requests.get(ROOT_URL + url).content
},
verbose=True)
# sanity check:
print 'No. of e-mails \t', io.mg_find(MG_COLL_NAME, {
'author': 'satoshi',
'type': {
'$eq': 'email'
}
}).count()
print 'No. of posts \t', io.mg_find(MG_COLL_NAME, {
'author': 'satoshi',
'type': {
'$eq': 'forum'
}
}).count()
print '\n'
##--- Parse HTML for content ---
MG_SRC_COLL = MG_COLL_NAME
MG_DEST_COLL = 'raw-docs'
class MySVCModel(object):
def __init__(self, num_class=2):
"""
:type num_classes: int
:rtype: None
"""
self.__ctrl__ = None
self.__case__ = None
with open('../../.dbname', 'r') as f:
self.__DB_NAME__ = json.load(f)['dbname']
self.__MG_DOCS_COLL__ = 'raw-docs' # raw docs
self.__MG_SENTS_COLL__ = 'bag-of-sents' # raw sentences
self.__MG_TOKENS_COLL__ = 'sample-tokens' # clean tokens (words)
self.__PG_STATS_TBL__ = 'stats' # stylometric features
self.__PG_RESULTS_TBL__ = 'results_' + \
str(num_class) + \
'class' # cross val results
self.__PG_PROBAS_TBL__ = 'probabilities' # cross val probabilities
self.__model__ = Pipeline([ \
# ('scaler2', StandardScaler()),
# ('scaler', MinMaxScaler()),
# ('scaler3', Normalizer()),
('classifier', SVC(probability=True,
kernel='poly',
degree=2,
class_weight='balanced') \
if num_class-1 \
else OneClassSVM(kernel='rbf',
nu=0.7,
gamma=1./250))
])
print 'Instantiated classifier %s.' % \
self.__model__.named_steps['classifier'].__class__.__name__
self.__io__ = DBIO(MG_DB_NAME=self.__DB_NAME__,
PG_DB_NAME=self.__DB_NAME__)
self.__tagger__ = None # initialise if re-creating samples
self.__bootstrap__ = None # initialise in fit
def fit(self,
author1,
author2,
wts1=None,
wts2=None,
bootstrap=False,
verbose=False):
"""
:type author1: str
:type author2: str
:type wts1: str/List[str]
:type wts2: str/List[str]
:type verbose:bool
:rtype: bool
:
: Prepares databases and tables/collections.
:
"""
self.__bootstrap__ = bootstrap
cases = []
for i, (author, wts) in enumerate([(author1, wts1), (author2, wts2)]):
if not wts:
wts = [wt.encode('ascii') \
for wt in self.__io__.mg_distinct(self.__MG_DOCS_COLL__,
'type',
{ 'author':author } )]
if not isinstance(wts, list):
wts = [wts]
cases += (author, wts, (1, -1)[i]), # use 1, -1 to match output
# from sklearn's OneClassSVM
self.__ctrl__ = cases[0] # assign label 1 in y vector
self.__case__ = cases[1] # assign be label 0 in y vector
self.__MG_TOKENS_COLL__ += '-' + cases[0][0] + \
'-' + cases[1][0] + \
'-' + \
''.join(wt[:3] for wt in cases[0][1]) + \
'-' + \
''.join(wt[:3] for wt in cases[1][1]) + \
'-' + \
('nobs','bs')[bootstrap]
self.__PG_STATS_TBL__ += '_' + cases[0][0] + \
'_' + cases[1][0] + \
'_' + \
''.join(wt[:3] for wt in cases[0][1]) + \
'_' + \
''.join(wt[:3] for wt in cases[1][1]) + \
'_' + \
('nobs','bs')[bootstrap]
if verbose:
print 'Control:', self.__ctrl__
print 'Case: ', self.__case__
print 'Saving tokens to', self.__MG_TOKENS_COLL__
print 'Saving stats to', self.__PG_STATS_TBL__
return self.__prep_sents__(
verbose=verbose) # err in preparing sentences
def __prep_sents__(self, verbose=False):
"""
:type verbose: bool
:rtype: bool
:
: Prepares bags of raw sentences if not already done so.
: (lump docs by the same auther and of the same type into the same bag)
:
"""
for author, wts, _ in (self.__ctrl__, self.__case__):
for wt in wts: # process type by type so sents can be reused
if verbose:
print "Checking for %s's %s sentences..." % (author, wt)
query_dic = {'author': author, 'type': wt}
if self.__io__.mg_find(self.__MG_SENTS_COLL__,
query_dic).count() == 0:
print " No sentences found. Preparing from raw docs..."
query_results = self.__io__.mg_find(
self.__MG_DOCS_COLL__, query_dic)
# gather raw lines
contents = []
for result in query_results:
if isinstance(result['content'], list):
contents += result['content'] # list of lines
else:
contents += result[
'content'], # entire doc = a str
if not contents:
print ' Error: No docs found.'
return False
# create a bag of sentences from contents
sentences = bag_of_sentences(contents)
if not self.__io__.mg_insert_one(
self.__MG_SENTS_COLL__,
query_dic,
dict({'sentences': sentences}, **query_dic),
verbose=True):
print 'Error in saving sentences.'
return False
return True
def cross_validate(self,
chunk_size=500,
k=3,
verbose=False,
recreate_samples=False,
plot_samples=False):
"""
:type chunk_size: int
:type k: int
:type verbose: bool
:type recreate_samples: bool
:type plot_samples: bool
:rtype: bool
:
: Perform k-fold cross validation.
:
"""
# create results table if not exist
pg_header_dic = {
'author1': 'varchar(20)',
'wts1': 'text[]',
'author2': 'varchar(20)',
'wts2': 'text[]',
'bs': 'boolean',
'fold': 'int'
}
# create table to store stylometric results if not already exist
if isinstance(
self.__io__.pg_find("""
SELECT COUNT(*) FROM %s
""" % self.__PG_RESULTS_TBL__),
type(None)):
self.__io__.recreate_tables(PG_TBL_NAME=self.__PG_RESULTS_TBL__,
pg_header_dic=pg_header_dic,
verbose=True)
# create table to store prediction probabilities if not already exist
if self.__model__.named_steps['classifier'].__class__.__name__ not in \
set(['OneClassSVM', 'LinearSVC']) and \
isinstance(self.__io__.pg_find("""
SELECT COUNT(*) FROM %s
""" % self.__PG_PROBAS_TBL__),
type(None)):
self.__io__.recreate_tables(PG_TBL_NAME=self.__PG_PROBAS_TBL__,
pg_header_dic=dict({'uid': 'int'},
**pg_header_dic),
verbose=True)
# create samples if not already exist or recreate samples when forced to
if self.__io__.mg_find(self.__MG_TOKENS_COLL__, {}).count() == 0 or \
recreate_samples:
print 'Initialising Spacy POS tagger...'
self.__tagger__ = English() # spacy tagger
# recreate tables (role = "trian", "test")
self.__io__.recreate_tables(MG_COLL_NAME=self.__MG_TOKENS_COLL__,
PG_TBL_NAME=self.__PG_STATS_TBL__,
pg_header_dic={
'fold': 'int',
'role': 'varchar(5)',
'uid': 'int',
'author': 'varchar(20)',
'type': 'text[]'
},
verbose=True)
# split raw sentences into folds
print 'Splitting folds...'
folds_dic = self.__create_folds__(k, verbose=verbose)
if verbose:
for key, item in folds_dic.iteritems():
print ' %s: %i folds (%s sentences)' % \
(key, len(item), [len(lst) for lst in item])
# create chunks then clean, tokenize and count words in each chunk
print 'Creating and cleaning chunks...'
if not self.__create_samples__(folds_dic,
k,
chunk_size,
bootstrap=self.__bootstrap__,
verbose=verbose):
print 'Error in creating samples.'
return False
for fold in xrange(k):
# create feature matrix
print 'Preparing feature matrix.'
dfs = self.__prepare_df__(fold, verbose=verbose)
if plot_samples:
print 'Hello Samples :)'
self.__plot_samples__(dfs, fold)
X_train = dfs[0].drop('label', axis=1)
y_train = dfs[0]['label']
X_test = dfs[1].drop('label', axis=1)
y_true = dfs[1]['label']
# train model
print 'Fit and predict...'
self.__model__.fit(X_train, y_train)
y_pred = self.__model__.predict(X_test)
error = False
with warnings.catch_warnings(record=True) as w:
# compute metrics
CM = confusion_matrix(y_true, y_pred)
f1 = f1_score(y_true, y_pred)
accuracy = accuracy_score(y_true, y_pred)
recall = recall_score(y_true, y_pred)
precision = average_precision_score(y_true, y_pred)
if len(w): # caught f1 UndefinedMetricsWarning from sklearn
print w[-1].message
error = True
print ' f1 score :', f1
print ' accuracy score :', accuracy
print ' recall score :', recall
print ' area under the precision-recall curve:', precision
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
probas = self.__model__.predict_proba(X_test)
print ' probabilities:\n', probas
# save results to postgres
header_dic = {
'author1': self.__ctrl__[0],
'wts1': self.__ctrl__[1],
'author2': self.__case__[0],
'wts2': self.__case__[1],
'bs': self.__bootstrap__,
'fold': fold
}
# CM: row = true, col = pred, ind 0 = case, ind 1 = ctrl
detail_dic = {
'tp': CM[1][1], # ctrl classified as ctrl
'fp': CM[0][1],
'fn': CM[1][0], # ctrl classified as case
'tn': CM[0][0],
'f1': f1,
'accuracy': accuracy,
'recall': recall,
'avg_prec': precision,
'und_err': error
}
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
detail_dic.update({ 'pdiff_mean': np.mean(abs(probas[:,0] - \
probas[:,1])),
'pdiff_std' : np.std(abs(probas[:,0] - \
probas[:,1])) })
if not self.__io__.pg_insert_one(
self.__PG_RESULTS_TBL__, header_dic, detail_dic,
verbose=True):
print 'Error in saving results.'
return False
# store probablilities if using two-class SVC
if self.__model__.named_steps['classifier'].__class__.__name__ \
not in set(['OneClassSVM', 'LinearSVC']):
for i, row in enumerate(probas):
if not self.__io__.pg_insert_one(self.__PG_PROBAS_TBL__,
dict({'uid': i}, **
header_dic),
detail_dic={
'y_pred': y_pred[i],
'case_0': row[0],
'ctrl_1': row[1]
},
verbose=True):
print 'Error in saving results.'
return False
return True
def __create_folds__(self, k=3, verbose=False):
"""
:type k: int
:type verbose: bool
:rtype: dict
:
: Divide sentences into k folds.
:
"""
folds_dic, target_count = {}, 0
for author, wts, _ in (self.__ctrl__, self.__case__):
query_dic = {'author': author, 'type': {'$in': wts}}
query_results = self.__io__.mg_find(self.__MG_SENTS_COLL__,
query_dic)
sentences = [result['sentences'] for result in query_results]
sentences = list(chain(*sentences))
if verbose:
print ' (%s, %s): %i sentences' % (author, wts,
len(sentences))
# randomly shuffle sentences
shuffle(sentences) # in place shuffle
folds_dic[(author,
str(wts))] = [sentences[i::k] for i in xrange(k)]
return folds_dic
def __create_samples__(self,
folds_dic,
k=3,
chunk_size=500,
bootstrap=False,
verbose=False):
"""
:type folds_dic: dict
:type k: int
:type chunk_size: int
:type bootstrap: bool
:type verbose: bool
:rtype: bool
:
: Create a sample out of each chunk in each fold.
:
"""
# process fold by fold
for fold in xrange(k):
print ' Starting fold %i...' % fold
# create training chunks
chunks_train_dic, chunks_test_dic = {}, {}
for author, wts, _ in (self.__ctrl__, self.__case__):
chunks_train_dic[(author, str(wts))] = \
create_chunks(list(chain(*folds_dic[(author, str(wts))][1:])),
chunk_size,
0,
verbose=False)
# compute amount to bootstrap
target_count = max(len(chunks) \
for chunks in chunks_train_dic.values())
for author, wts, _ in (self.__ctrl__, self.__case__):
# bootstrap minority class to 50/50
if bootstrap and \
len(chunks_train_dic[(author, str(wts))]) < target_count:
if verbose:
print ' Bootstrap minority to %i chunks.' % target_count
chunks_train_dic[(author, str(wts))] = \
create_chunks(list(chain(*folds_dic[(author,
str(wts))][1:])),
chunk_size,
target_count,
verbose=False)
# do not bootstrap test
chunks_test_dic[(author, str(wts))] = \
create_chunks(folds_dic[(author, str(wts))][0],
chunk_size,
0,
verbose=False)
# prepare folds_dic for next fold-iteration
folds_dic[(author, str(wts))] = \
folds_dic[(author, str(wts))][1:] + \
folds_dic[(author, str(wts))][:1]
if verbose:
for author, wts, _ in (self.__ctrl__, self.__case__):
print ' (%s, %s): %i train, %i test' % \
(author,
wts,
len(chunks_train_dic[(author, str(wts))]),
len(chunks_test_dic[(author, str(wts))]))
# Clean, tokenize and compute stats for each chunk
for author, wts, _ in (self.__ctrl__, self.__case__):
for role, chunks \
in [('train', chunks_train_dic[(author, str(wts))]),
('test', chunks_test_dic[(author, str(wts))])]:
for i, chunk in enumerate(chunks, 1):
print ' Processing %s %s %s chunk %i/%i...' % \
(author, wts, role, i, len(chunks))
# clean chunks, tokenize and compute stats
stats_ctr, \
words, \
words_lemma, \
misspellings, \
gb_spellings, \
us_spellings = crunch_statistics(chunk, self.__tagger__)
query_dic = {
'fold': fold,
'role': role,
'uid': i,
'author': author,
'type': wts
}
entry_dic = {
'words': words,
'words_lemma': words_lemma,
'misspellings': misspellings,
'us_spellings': us_spellings,
'gb_spellings': gb_spellings
}
# save words in mongodb
if not self.__io__.mg_insert_one(
self.__MG_TOKENS_COLL__,
query_dic,
dict(query_dic, **entry_dic),
verbose=True):
print 'Error in saving words.'
return False
# save stats in postgres
if not self.__io__.pg_insert_one(self.__PG_STATS_TBL__,
query_dic,
stats_ctr,
verbose=True):
print 'Error in saving stats.'
return False
# break # break chunks
return True
def __prepare_df__(self, fold, verbose=False):
"""
:type fold: int
:type verbose: bool
:rtype: pandas.DataFrame
:
: Prepare feature matrix.
:
"""
# compute bag-of-word
query_results = self.__io__.mg_find(
self.__MG_TOKENS_COLL__, {
'fold': fold,
'role': 'train',
'author': self.__ctrl__[0],
'type': self.__ctrl__[1]
})
words_lists = [result['words_lemma'] for result in query_results]
words = list(chain(*words_lists))
bag_of_words = get_topN(words, 250).keys()
if verbose:
print ' No. of words lists:', len(words_lists)
print ' No. of lemmatized words:', len(words)
print ' No. of unique lemmatized words', len(set(words))
# prepare features dataframe for model
dfs = []
for role in ('train', 'test'):
df_stylo, df_freq = pd.DataFrame(), pd.DataFrame()
# get stylometric features
for author, wts, label in (self.__ctrl__, self.__case__):
# train one-class SVM with only literature from ctrl
if \
self.__model__.named_steps['classifier'].__class__.__name__ == \
'OneClassSVM' \
and role == 'train' \
and label == self.__case__[2]:
continue
# get term frequency features
q = '''
SELECT * FROM %s
WHERE fold = %i
AND role = '%s'
AND author = '%s'
AND type = ARRAY%s;
''' % (self.__PG_STATS_TBL__, fold, role, author, wts)
df = self.__io__.pg_find(q)
if isinstance(df, type(None)):
print 'Error: no stats found.'
return False
df_stylo = df_stylo.append(compute_stylo_features(df, label))
# get term frequency features
results = [results for results in \
self.__io__.mg_find( self.__MG_TOKENS_COLL__,
{ 'fold' : fold,
'role' : role,
'author': author,
'type' : wts } ) ]
df_freq = df_freq.append(
compute_freqs_features(results, bag_of_words, label))
# horizontally stack DataFrames
dfs += df_stylo.join(df_freq.drop('label', axis=1)),
if verbose:
print ' df_train shape:', dfs[0].shape
print ' df_test shape:', dfs[1].shape
return dfs
def __plot_samples__(self, dfs, fold):
"""
:type dfs: List[pandas DataFrame] # [training df, testing df]
:type fold: int
:rtype: None
"""
mds = MDS(n_components=2,
max_iter=3000,
eps=1e-9,
dissimilarity='euclidean',
n_jobs=-1)
tsne = TSNE(n_components=2)
# change label to color index
# author 1 train (0 = light blue), author 1 test (1 = dark blue)
# author 2 train (2 = light green), author 2 test (3 = dark green)
df_all = pd.DataFrame(columns=dfs[0].columns)
df0_copy = dfs[0].copy()
df0_copy.loc[(df0_copy.label == 1).values, 'label'] = 0
df0_copy.loc[(df0_copy.label == -1).values, 'label'] = 2
df_all = df_all.append(df0_copy)
df1_copy = dfs[1].copy()
df1_copy.loc[(df1_copy.label == 1).values, 'label'] = 1
df1_copy.loc[(df1_copy.label == -1).values, 'label'] = 3
df_all = df_all.append(df1_copy)
legend = {
0: 'Author 1 Training Sample',
1: 'Author 1 Test Sample',
2: 'Author 2 Training Sample',
3: 'Author 2 Test Sample'
}
# fit on training data
pos_lst = [('Multi-Dimensional Scaling (MDS)',
mds.fit(df_all.drop('label', axis=1)).embedding_),
('t-Distributed Stochastic Neighbor Embedding (TSNE)',
tsne.fit(df_all.drop('label', axis=1)).embedding_)]
# plot
colors = sns.color_palette('Paired', 4)
fig = plt.figure(figsize=(16, 7))
plt.hold(True)
for k, (title, pos) in enumerate(pos_lst, 1):
## fig.add_subplot() works in ipython notebook but creates a
## mysterious 3rd axes in python...
# ax = fig.add_subplot(1,2,k)
ax = plt.subplot(1, 2, k)
ax.set_title(title)
for i in xrange(len(colors)):
samples = pos[(df_all.label == i).values, :]
ax.scatter(samples[:, 0],
samples[:, 1],
c=colors[i],
edgecolor='none',
label=legend[i])
ax.legend()
plt.hold(False)
plt.savefig('../figs/' + \
self.__PG_STATS_TBL__[self.__PG_STATS_TBL__.find("_")+1:] + \
'fold' + str(fold) + '.png',
dpi=300, transparent=True)
plt.close(fig)