Lab 4 - Pageviews

Professor Brian Keegan
Department of Information Science, CU Boulder
This notebook is copyright and made available under the Apache License v2.0 license.

This is the third of five lab notebooks that will explore how to analyze the structure of collaborations in Wikipedia data about users' revisions across multiple articles. This lab will extend the methods in the previous two labs about analyzing a single article's revision histories and analyzing the hyperlink networks around a single Wikipedia page. You do not need to be fluent in either to complete the lab, but there are many options for extending the analyses we do here by using more advanced queries and scripting methods.

I'd like to thank the Wikimedia Foundation for the PAWS system and related Wikitech infrastructure that this workbook runs within. Yuvi Panda, Aaron Halfaker, Jonathan Morgan, and Dario Taraborelli have all provided crucial support and feedback.

Confirm that basic Python commands work

a = 3
b = 4
c = 5

Import modules and setup environment

Load up all the libraries we'll need to connect to the database, retreive information for analysis, and visualize results.

# Makes the plots appear within the notebook
%matplotlib inline

# Two fundamental packages for doing data manipulation
import numpy as np                   #
import pandas as pd                  #

# Two related packages for plotting data
import matplotlib.pyplot as plt      #
import seaborn as sb                 #

# Package for requesting data via the web and parsing resulting JSON
import requests                      #
import json                          #
from bs4 import BeautifulSoup        #

# Two packages for accessing the MySQL server
import pymysql                       #
import os                            #

# Setup the code environment to use plots with a white background and DataFrames show more columns and rows
pd.options.display.max_columns = 100
pd.options.display.max_rows = 110

Define an article to examine pageview dynamics.

page_title = 'Star Wars'

Get pageview data for a single article

Details about the Wikimedia REST API for pageviews is available here. Unfortunately, this data end point only provides information going back to July 1, 2015.

This is what the API returns as an example.

# Get today's date and yesterday's date
today =
yesterday = - pd.to_timedelta('1 day')

# Convert to strings
today_s = pd.datetime.strftime(today,'%Y%m%d00')
yesterday_s = pd.datetime.strftime(yesterday,'%Y%m%d00')

# Get the pageviews for today and yesterday
url_string = '{0}/daily/{1}/{2}'
print(url_string.format(page_title.replace(' ','_'),yesterday_s,today_s))

Write a function to get the pageviews from January 1, 2015 (in practice, the start date will be as late as August or as early as May) until yesterday.

def get_daily_pageviews(page_title):
    url_string = '{0}/daily/2015010100/{1}'
    today = pd.datetime.strftime(,'%Y%m%d00')
    req = requests.get(url_string.format(page_title,today))

    json_s = json.loads(req.text)
    if 'items' in json_s.keys():
        _df = pd.DataFrame(json_s['items'])[['timestamp','views','article']]
        _df['timestamp'] = pd.to_datetime(_df['timestamp'],format='%Y%m%d00')
        _df['weekday'] = _df['timestamp'].apply(lambda x:x.weekday())
        return _df

Get the data for your page.

pageview_df = get_daily_pageviews(page_title)

Interpret page view results

What does the pageview activity look like? Are there any bursts of attention? What might these bursts be linked to?

ax = pageview_df.plot.line(x='timestamp',y='views',logy=False,legend=False)

Use a logarithmic scaling for the y-axis to see more of the detail in the lower-traffic days.

ax = pageview_df.plot.line(x='timestamp',y='views',logy=True,legend=False)

What are the dates for the biggest pageview outliers? Here we define an "outlier" to be more than 3 standard deviations above the average number of pageviews over the time window.

std_threshold = 4
threshold_val = pageview_df['views'].mean() + pageview_df['views'].std() * std_threshold
peak_days = pageview_df[pageview_df['views'] > threshold_val]


How much of the total pageview activity occurred on these days compared to the rest of the pageviews?

peak_fraction = pageview_df.loc[peak_days.index,'views'].sum()/pageview_df['views'].sum()

print('{0:.1%} of all pageviews occurred on the {1} peak days.'.format(peak_fraction,len(peak_days)))

How does pageview activity change over the course of a week?

g = sb.factorplot(x='weekday',y='views',data=pageview_df,kind='bar',color='grey',
ax = g.axes[0][0]
ax.set_ylabel('Average pageviews')

Compare pageviews to another page

Lets write a function that takes a list of article names and returns a DataFrame indexed by date, columned by articles, and values being the number of pageviews.

def get_multiple_pageviews(page_list):
    today = str(
    multiple_pv_df = pd.DataFrame(index=pd.date_range('2015-05-01',today))
    for page in page_list:
        pv_df = get_daily_pageviews(page)
            multiple_pv_df[page] = pv_df.set_index('timestamp')['views'] 
            print("Error on: {0}".format(page))
            multiple_pv_df[page] = np.nan
    return multiple_pv_df.dropna(how='all')

Enter two related pages for which you want to compare their pageview behavior.

page_list = ['Star Wars','George Lucas']

Get both of their data.

# Get the data
multiple_pvs = get_multiple_pageviews(page_list)

# Show the top rows

Plot the data.


What is the correlation coefficient between these two articles' behavior?


How did the ratio between the two articles' pageviews change over time?

ratio_s = multiple_pvs[page_list[0]].div(multiple_pvs[page_list[1]])
ax = ratio_s.plot()

Use the functions for resolving redirects and getting page outlinks from prior labs.

# From
def make_chunks(l, n):
    """Yield successive n-sized chunks from l."""
    for i in range(0, len(l), n):
        yield l[i:i + n]

def resolve_redirects(page_title_list):
    # Chunk the pages into a list of lists of size 50
    chunks = make_chunks(page_title_list,50)
    # Create an empty list to fill with the redirected titles
    redirected_page_titles = []
    # For each chunk try to get the redirects
    for chunk in chunks:
        # Create the query string that separates spaces within page titles by '+' 
        # and separates page titles by '|'
        page_titles_string = '|'.join([page.replace(' ','+') for page in chunk])
        # Put this large string into the URL
        url_string = '{0}+&redirects=1'.format(page_titles_string)
        # Do the query and parse the JSON response into a dictionary
        req = json.loads(requests.get(url_string).text)
        # Convert the returned values containing redirects into a dictionary
        if 'redirects' in req['query'].keys():
            redirected_titles = {d['from']:d['to'] for d in req['query']['redirects']}
            # Add the redirected titles to the list
            for title in chunk:
                    #print(len(redirected_page_titles), title, redirected_titles[title])
                # If they don't have a redirect just add the original title
                except KeyError:
                    #print(len(redirected_page_titles), '\nFrom: ', title, '\nTo: ', title)
            for title in chunk:
    # Make sure the number of page titles remained the same, otherwise raise a warning
    if len(page_title_list) == len(redirected_page_titles):
        return redirected_page_titles
        print("WARNING! The number of page titles in the redirected list ({0}) is not equal to the input list ({1})".format(len(redirected_page_titles),len(page_title_list)))
        return redirected_page_titles

def get_page_outlinks(page_title,redirects=1):
    # Replace spaces with underscores
    #page_title = page_title.replace(' ','_')
    bad_titles = ['Special:','Wikipedia:','Help:','Template:','Category:','International Standard','Portal:','s:']
    # Get the response from the API for a query
    # After passing a page title, the API returns the HTML markup of the current article version within a JSON payload
    req = requests.get('{0}&redirects={1}&prop=text&disableeditsection=1&disabletoc=1'.format(page_title,redirects))
    # Read the response into JSON to parse and extract the HTML
    json_string = json.loads(req.text)
    # Initialize an empty list to store the links
    outlinks_list = [] 
    if 'parse' in json_string.keys():
        page_html = json_string['parse']['text']['*']

        # Parse the HTML into Beautiful Soup
        soup = BeautifulSoup(page_html,'lxml')

        # Delete tags associated with templates
        for tag in soup.find_all('tr'):

        # For each paragraph tag, extract the titles within the links
        for para in soup.find_all('p'):
            for link in para.find_all('a'):
                if link.has_attr('title'):
                    title = link['title']
                    # Ignore links that aren't interesting
                    if all(bad not in title for bad in bad_titles):

        # For each unordered list, extract the titles within the child links
        for unordered_list in soup.find_all('ul'):
            for item in unordered_list.find_all('li'):
                for link in item.find_all('a'):
                    if link.has_attr('title'):
                        title = link['title']
                        # Ignore links that aren't interesting
                        if all(bad not in title for bad in bad_titles):

    return outlinks_list

Get the outlinks.

raw_outlinks = get_page_outlinks(page_title)
redirected_outlinks = resolve_redirects(raw_outlinks)

Get the data.

This stage may take several minutes.

# Get the data
hl_pvs_df = get_multiple_pageviews(redirected_outlinks + [page_title])

# Show the top rows

What are the most-viewed articles in the hyperlink network?

most_viewed_articles = hl_pvs_df.cumsum().ix[str(]
most_viewed_articles = most_viewed_articles.sort_values(ascending=False)

Most and least correlated articles

Which articles are most correlated with each other?

# Log the pageview data to reduce skew from bursty outliers abd make the correlation table
hl_corr_df = hl_pvs_df.apply(np.log).corr()

# Correlation table is symmetric, drop one half of them
# From:
hl_corr_df = hl_corr_df.where(np.triu(np.ones(hl_corr_df.shape)).astype(np.bool))

# Unstack the DataFrame into a series and sort
hl_corr_s = hl_corr_df.unstack().sort_values(ascending=False)

# Drop NaNs
hl_corr_s = hl_corr_s.dropna()

# Drop values equal to 1
hl_corr_s = hl_corr_s[hl_corr_s < 1]

List out the 10 most correlated articles.


Inspect this correlation from the raw data.

_df = hl_pvs_df[list(hl_corr_s.index[0])]

ax = _df.plot(logy=True)

Look at the 10 least-correlated articles.


Plot the correlation between the two most anti-correlated articles. These show some kinda wacky properties that are interesting to explore or think more about.

_df = hl_pvs_df[list(hl_corr_s.index[-1])]

ax = _df.plot(logy=True)

Is there a relationship between the position of the link on the page and the correlation between the linked article's pageviews and the seed article's pageviews? For instance, links closer to the top of the page might reflect more important topics while links towards the end of the page might be less relevant.

link_corrs = []

for num,link in enumerate(redirected_outlinks):
    except KeyError:
        print("Error on: {0}".format(link))

Plot the results.

ax = pd.DataFrame(link_corrs).plot.scatter(x='position',y='corr')
ax.set_xlabel('Link position')

Get page revisions

In this section, we'll repurpose and adapt code from the last lab to get data about page revisions. Rather than looking at the number of times a user contributed to a given article, we'll simply count the number of times the article was edited on a given date.

def get_page_edits_by_date(page_title,conn,date_string='2014-12-31'):
    """ Takes a page title and returns the number of revisions made on each date.
      page_title = a string for the page title to get its revisions
      date_string = a string for the date in YYYY-MM-DD format
      conn = a database connection
      A DataFrame with username, page title, edit count, and min/max timestamps
    # In case you pass a page title with spaces in it, replace the spaces with underscores
    page_title = page_title.replace(' ','_').encode('utf8').decode('latin1')
    # The MySQL query string used to retrieve the data. By line, it is
    ## converting the timestamp to a date and 
    ## counting the number of elements
    ## from the "revisions" table
    ## joining the "page" table on it
    ## using the page_id and rev_page columns as keys
    ## limiting the results to entries that have the pagetitle, 
    ## occur in the namespace, and happen after Dec 31, 2014
    ## grouping the results by date
    s = """
                DATE(rev_timestamp) as date,
                COUNT(*) as edits
                page ON page.page_id = revision.rev_page
                page.page_title = "{0}" 
                AND page_namespace = 0
                AND DATE(rev_timestamp) > '{1}'
            GROUP BY

    # Use the connection to run the query and return the results as a DataFrame
    _df = pd.read_sql_query(s,conn)
    _df['page_title'] = _df['page_title'].str.decode('utf8')
    _df['page_title'] = _df['page_title'].str.replace('_',' ')
    # Return the data, with a clean index
    return _df

def get_neighbors_revisions(page_title,conn):
    """ Takes a page title and returns revisions for the page and its neighbors.
      page_title = a string for the page title to get its revisions
      A pandas DataFrame containing all the page revisions.
    # Get the outlinks from the page and include the page itself in the list
    alters = get_page_outlinks(page_title) + [page_title]
    # Resolve the redirects in the list of alters
    alters = list(set(resolve_redirects(alters)))
    # Create an empty container to hold the DataFrames
    df_list = []
    # For each page, get the revision counts and append to the df_list
    for alter in alters:
        _df = get_page_edits_by_date(alter,conn)
    # Concatenate the list of revision count DataFrames into a giant DataFrame
    df = pd.concat(df_list)
    # Return the data
    return df.reset_index(drop=True)

Get the authentication information and connect to the database.

host, user, password = os.environ['MYSQL_HOST'], os.environ['MYSQL_USERNAME'], os.environ['MYSQL_PASSWORD']
conn = pymysql.connect(host=host,user=user,password=password,database='enwiki_p',connect_timeout=3600)
conn.cursor().execute('use enwiki_p');

Get the number of revisions per day for all the articles.

hl_daily_rev_df = get_neighbors_revisions(page_title,conn)

Reindex the edit data so it's starting and ending on the same dates as the pageviews data.

# Convert into a format like the hl_pageviews DataFrame
# Index are dates between Jan 1, 2015 and today; columns are article titles; values are number of edits
hl_edits_df = hl_daily_rev_df.set_index(['date','page_title'])['edits'].unstack(1)

# Reindex so dates are continuous
pv_start_ix = str(hl_pvs_df.index.min().date())
pv_end_ix = str(( - pd.Timedelta('1 d')).date())
_date_range = pd.date_range(pv_start_ix,pv_end_ix)
hl_edits_df = hl_edits_df.reindex(index=_date_range)

# Fill in empty observations with 0s
hl_edits_df = hl_edits_df.fillna(0)


Are pageviews and edits correlated with each other?

_s1 = hl_pvs_df[page_title]
_s2 = hl_edits_df[page_title]

single_pv_edits_df = pd.DataFrame({'pageviews':_s1,'edits':_s2})
ax = single_pv_edits_df.plot(secondary_y='edits',logy=True)

Can Wikipedia supply information to keep up with demand?

The ratio between the cumulative pageviews and cumulative edits.

ax = (_s1.cumsum()/_s2.cumsum()).plot()

ax.set_ylabel('Cumulative pageviews per edit')
def zscore(series):
    return np.abs((series - series.mean())/series.std())

Look at the normalized (z-score) excitation and relaxation in edits and pageviews by day. Each point is a single day in the article's history and they're connected if they come one day after each other. Values along the diagonal in red suggest that increases in attention to the article are matched by similar increases in editing activity on the article. Alternatively, data points in the upper-left triangle suggest increases in pageviews are not matched by increases in edits while data points in the lower-right triangle suggest increases in edits are not matched by increases in pageviews.

f,ax = plt.subplots(1,1)

ax.set_xlabel('Edits (z-score)')
ax.set_ylabel('Pageviews (z-score)')

plt.text(1e-1,1e1,'More views than edits',ha='center',weight='bold')
plt.text(1e1,1e-1,'More edits than views',ha='center',weight='bold')


_s1 = zscore(hl_edits_df[page_title])
_s2 = zscore(hl_pvs_df[page_title])
f,ax = plt.subplots(1,1)

ax.set_xlabel('Edits (z-score)')
ax.set_ylabel('Pageviews (z-score)')

for page in hl_edits_df.columns:
    _s1 = zscore(hl_edits_df[page])
    _s2 = zscore(hl_pvs_df[page])
# Diagonal red line showing perfect matching

Did a burst of pageviews diffuse to adjacent pages?

On what day did the maximum number of pageviews occur?

page_max_pv_date = hl_pvs_df[page_title].idxmax()
# For each article, what was the date when the maximum number of 
max_pv_dates_df = hl_pvs_df.idxmax()

# Check whether these dates were also the page_max_pv_date
_fraction = max_pv_dates_df.isin([page_max_pv_date,page_max_pv_date+1]).sum()/len(max_pv_dates_df)

print("{0:.1%} of articles in the hyperlink network also had the max pageview on the same date as the article.".format(_fraction))

What are these "co-bursting" pages?

cobursting_pv_articles = max_pv_dates_df[max_pv_dates_df.isin([page_max_pv_date,page_max_pv_date+1])].sort_index()

Did this burst of pageviews translate into edits on these pages as well?

Get the articles with the peak amount of activity by date and then subset to only include those articles whose peak dates are the same as the peak pageview dates.

max_edits_dates_df = hl_edits_df.idxmax()
cobursting_edit_articles = max_edits_dates_df[max_edits_dates_df.isin([page_max_pv_date,page_max_pv_date+1])].sort_index()

How many edits beyond the target article did this translate into? Subset the dataframe to the relevant dates and articles and inspect.

dates = [page_max_pv_date,page_max_pv_date+1]
cobursting_edit_articles_list = list(set(cobursting_edit_articles.index) - set(page_title))
cobursting_hl_edits_df = hl_edits_df.loc[dates,cobursting_edit_articles.index]

Now sum up all these edits.