def dataset():
"""Returns partial dataset with pagecounts."""
df = features.pageids_canonical().set_index('pageid')
np.testing.assert_(df.index.is_unique)
n_records = len(df)
# merge titles
titles = features.titles()
df = pd.merge(df, titles, left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
# merge redirects
redirects = _redirects_of_interest(df.title)
df = pd.merge(df, redirects, how='left', left_on='title', right_index=True)
df['redirect_set'] = df['redirect_set'].fillna(set())
df['n_redirect'] = df['redirect_set'].apply(len)
np.testing.assert_equal(n_records, len(df))
del titles
del redirects
# gather pagecounts for all titles associated with a page
# this code is awkward because we need to avoid making many lookups in a
# very large HDF5 file. We make just one query and then rearrange results.
titles_and_redirect_titles_of_interest = set(df.title.tolist())
for s in df.redirect_set:
titles_and_redirect_titles_of_interest.update(s)
# to avoid memory errors we do this in stages
def chunks(iterable, size):
it = iter(iterable)
item = list(itertools.islice(it, size))
while item:
yield item
item = list(itertools.islice(it, size))
chunk_size = 30
titles_chunks = chunks(titles_and_redirect_titles_of_interest, chunk_size)
pagecounts = dict()
logger.info("Retrieving pagecounts. This may take several hours.")
for titles in titles_chunks:
pagecounts_partial = pagecounts_for_titles(titles)
pagecounts.update(pagecounts_partial)
if len(pagecounts) % 10000 <= chunk_size:
logging.info("Retrieving pagecount {} of {}".format(len(pagecounts), len(titles_and_redirect_titles_of_interest)))
pagecounts_ordered = []
for i, (index, row) in enumerate(df.iterrows()):
count = pagecounts.get(row['title'], 0)
for redirect_title in row['redirect_set']:
count += pagecounts.get(redirect_title, 0)
pagecounts_ordered.append(count)
df['pagecount'] = pagecounts_ordered
return df[['title', 'pagecount']]
def _redirects_of_interest(titles_of_interest):
titles = features.titles()
redirects = features.redirects()
# restrict redirects to universe and add pageid
redirects = redirects[redirects.redirect_to.isin(titles_of_interest)].reset_index(drop=True)
redirects = pd.merge(redirects, titles, how='left', left_on='redirect_from', right_index=True)
del redirects['redirect_from'] # old redirect_from, was pageid
redirects.rename(columns={'title': 'redirect_from'}, inplace=True)
redirects = redirects.set_index('redirect_to')
# remove NaN; these correspond to pages that have no title in the dump (likely deleted?)
redirects = redirects[pd.notnull(redirects.redirect_from)]
# now the redirects DataFrame is in a usable form
redirects = redirects.groupby(redirects.index)['redirect_from'].apply(set)
redirects.name = 'redirect_set'
return pd.DataFrame(redirects)
def dataset(langlink_max_features, death_decade_start, death_decade_stop):
"""Returns the full dataset
Limit language link indicators to `langlink_max_features`. Create death
decade indicators for decades between `death_decade_start` and
`death_decade_stop`.
"""
df = features.pageids_canonical().set_index('pageid')
np.testing.assert_(df.index.is_unique)
np.testing.assert_(df.notnull().values.all())
n_records = len(df)
# merge titles
titles = features.titles()
df = pd.merge(df, titles, left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge redirects
redirects = features.redirects_of_interest()
df = pd.merge(df, redirects, how='left', left_on='title', right_index=True)
df['redirect_set'] = df['redirect_set'].fillna(set())
df['n_redirect'] = df['redirect_set'].apply(len)
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
del titles
del redirects
# merge topics
df = pd.merge(df, features.topics(), left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge pagecounts
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.pagecounts(), left_index=True, right_index=True, how='left')
# because the only column with NaN values is pagecounts, we may fill with 0 safely
# Note that a pagecount of NaN means a page got 0 or 1 hits during the period studied
df = df.fillna(0)
np.testing.assert_((df.pagecount >= 0).all())
np.testing.assert_equal(n_records, len(df))
# merge langlinks counts
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.langlink_counts(), left_index=True, right_index=True, how='left')
# because the only columns with NaN values are dummies, we may fill with 0 safely
df = df.fillna(0)
np.testing.assert_(df.notnull().values.all())
np.testing.assert_equal(n_records, len(df))
# merge imagelinks
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.imagelink_counts(), left_index=True, right_index=True, how='left')
df = df.fillna(0)
np.testing.assert_(df.notnull().values.all())
np.testing.assert_equal(n_records, len(df))
# merge article lengths
df = pd.merge(df, features.article_lengths(df.index), left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge last revision timestamp
df = pd.merge(df, features.last_revision(df.index), left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
# merge article age
df = pd.merge(df, features.article_age(df.index), left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
# merge revisions per day rates
df = pd.merge(df, features.revisions_per_day(df.index), left_index=True, right_index=True)
np.testing.assert_equal(n_records, len(df))
# merge categories (comma-separated list of categories)
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.categories_flat(), left_index=True, right_index=True, how='left')
# because the only columns with NaN values are dummies, we may fillna safely
df = df.fillna(set())
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge death decade dummies
df = pd.merge(df, features.death_decade_dummies(death_decade_start, death_decade_stop), left_index=True, right_index=True, how='left')
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge category dummies of interest
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.category_dummies_of_interest(), left_index=True, right_index=True, how='left')
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge langlink dummies
np.testing.assert_(df.notnull().values.all())
df = pd.merge(df, features.langlink_dummies(max_features=langlink_max_features), left_index=True, right_index=True, how='left')
# because the only columns with NaN values are dummies, we may fillna safely
df = df.fillna(0)
np.testing.assert_equal(n_records, len(df))
np.testing.assert_(df.notnull().values.all())
# merge online books page data
df = pd.merge(df, features.digital_editions(), left_index=True, right_index=True, how='left')
df.obp_digital_editions = df.obp_digital_editions.fillna(0)
np.testing.assert_equal(n_records, len(df))
# merge death year (there will be many NaNs)
df = pd.merge(df, features.death_year(), left_index=True, right_index=True, how='left')
np.testing.assert_equal(n_records, len(df))
return df