Python: Generating Network Graph of Twitter Follower

Generating a network graph of Twitter followers using Python and NetworkX

In this article I show you how by starting at a single twitter account we can build up a network graph of twitter followers and then visualize that network using the NetworkX library.

The steps are:

• From initial seed account collect followers using the Snowball Sampling technique.
• Process the collected twitter data to generate an output file of relationships between twitter accounts.
• Visualize network data in a network graph using the NetworkX library.

Step 1. Collect follower data from the Twitter API

You will need to have API keys to be able to query the Twitter API. I have written in previous articles how to do this, e.g. Collecting tweets using Python.

When you interact with the Twitter API you will learn quickly that you need to cache data as you go along. This is because the API is rate limited and you will find any script you write will halt frequently when hitting a rate limit if you don’t cache responses. The solution is to check for cached data before making an API call, if you get a cache miss then query the API and write the returned data to disk.

I use two directories for cached data. The directory ‘following’ contains a CSV file for each twitter account queried. The name of each file is the screen name of the twitter account and the content is a tab delimited list, each row contains the twitter id, screen name and account name of a follower, up to a maximum of 200 followers.

$ ls following/
-rw-r--r-- 1 mark mark 7.1K Aug 14 21:04 TEDxMtHood.csv
-rw-r--r-- 1 mark mark 7.0K Aug 14 21:21 TEDxYYC.csv
-rw-r--r-- 1 mark mark 5.7K Aug 15 07:29 TEDxCibeles.csv
-rw-r--r-- 1 mark mark 2.8K Aug 15 07:30 TEDxProvidence.csv
-rw-r--r-- 1 mark mark 6.9K Aug 15 07:46 TEDxUHasselt.csv
-rw-r--r-- 1 mark mark  625 Aug 15 07:46 TEDxWestVillage.csv
-rw-r--r-- 1 mark mark  196 Aug 15 07:46 TEDxESPRIT.csv
-rw-r--r-- 1 mark mark 2.9K Aug 15 08:02 TEDxUU.csv

cat following/TEDxESPRIT
XXXXXXXXX       dediil  hedil jabou
XXXXXXXXX       MehdiBJemia     Mehdi Ben Jemia
XXXXXXXXX       _willywall      _william
XXXXXXXX        MirakHikimori   Hello Hikimori
XXXXXXXX        maroo_king      Marou

The second directory is called ‘twitter-users’, it is a cache of twitter user details, each file contains cached data for a twitter user including friend and follower counts and a list of follower IDs (up to a maximum of 5000 follower IDs can be queried from the API).

$ ls twitter-users/
-rw-r--r-- 1 mark mark  252 Jul 24 16:45 XXXXXXXXX.json
-rw-r--r-- 1 mark mark  57K Jul 24 16:46 XXXXXXXX.json
-rw-r--r-- 1 mark mark 6.3K Jul 24 17:01 XXXXXXXXXX.json

... Lots more ...

$ cat twitter-users/XXXXXXXX
{
 "name": "TEDxSingapore",
 "friends_count": 147,
 "followers_count": 12814,
 "followers_ids": [
  XXXXXXXXXX,
  XXXXXXXXXX,
  XXXXXXXXX,
  ...
  XXXXXXXXXX,
  XXXXXXXXXX
 ],
 "id": XXXXXXXX,
 "screen_name": "TEDxSingapore"
}

Here is the script to collect this data:

import tweepy
import time
import os
import sys
import json
import argparse 

FOLLOWING_DIR = 'following'
MAX_FRIENDS = 200
FRIENDS_OF_FRIENDS_LIMIT = 200

if not os.path.exists(FOLLOWING_DIR):
    os.makedir(FOLLOWING_DIR)

enc = lambda x: x.encode('ascii', errors='ignore')

# The consumer keys can be found on your application's Details
# page located at https://dev.twitter.com/apps (under "OAuth settings")
CONSUMER_KEY = 'XXXXXXXXXXXXXXXXXXXXXXXXX'
CONSUMER_SECRET = 'XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX'

# The access tokens can be found on your applications's Details
# page located at https://dev.twitter.com/apps (located
# under "Your access token")
ACCESS_TOKEN = 'XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX'
ACCESS_TOKEN_SECRET = 'XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX'

# == OAuth Authentication ==
#
# This mode of authentication is the new preferred way
# of authenticating with Twitter.
auth = tweepy.OAuthHandler(CONSUMER_KEY, CONSUMER_SECRET)
auth.set_access_token(ACCESS_TOKEN, ACCESS_TOKEN_SECRET) 

api = tweepy.API(auth)

def get_follower_ids(centre, max_depth=1, current_depth=0, taboo_list=[]):

    # print 'current depth: %d, max depth: %d' % (current_depth, max_depth)
    # print 'taboo list: ', ','.join([ str(i) for i in taboo_list ])

    if current_depth == max_depth:
        print 'out of depth'
        return taboo_list

    if centre in taboo_list:
        # we've been here before
        print 'Already been here.'
        return taboo_list
    else:
        taboo_list.append(centre) 

    try:
        userfname = os.path.join('twitter-users', str(centre) + '.json')
        if not os.path.exists(userfname):
            print 'Retrieving user details for twitter id %s' % str(centre)
            while True:
                try:
                    user = api.get_user(centre) 

                    d = {'name': user.name,
                         'screen_name': user.screen_name,
                         'id': user.id,
                         'friends_count': user.friends_count,
                         'followers_count': user.followers_count,
                         'followers_ids': user.followers_ids()}

                    with open(userfname, 'w') as outf:
                        outf.write(json.dumps(d, indent=1))

                    user = d
                    break
                except tweepy.TweepError, error:
                    print type(error)

                    if str(error) == 'Not authorized.':
                        print 'Cant access user data - not authorized.'
                        return taboo_list

                    if str(error) == 'User has been suspended.':
                        print 'User suspended.'
                        return taboo_list

                    errorObj = error[0][0]

                    print errorObj

                    if errorObj['message'] == 'Rate limit exceeded':
                        print 'Rate limited. Sleeping for 15 minutes.'
                        time.sleep(15 * 60 + 15)
                        continue

                    return taboo_list
        else:
            user = json.loads(file(userfname).read())

        screen_name = enc(user['screen_name'])
        fname = os.path.join(FOLLOWING_DIR, screen_name + '.csv')
        friendids = []

        # only retrieve friends of TED... screen names
        if screen_name.startswith('TED'):
            if not os.path.exists(fname):
                print 'No cached data for screen name "%s"' % screen_name
                with open(fname, 'w') as outf:
                    params = (enc(user['name']), screen_name)
                    print 'Retrieving friends for user "%s" (%s)' % params 

                    # page over friends
                    c = tweepy.Cursor(api.friends, id=user['id']).items()

                    friend_count = 0
                    while True:
                        try:
                            friend = c.next()
                            friendids.append(friend.id)
                            params = (friend.id, enc(friend.screen_name), enc(friend.name))
                            outf.write('%s\t%s\t%s\n' % params)
                            friend_count += 1
                            if friend_count >= MAX_FRIENDS:
                                print 'Reached max no. of friends for "%s".' % friend.screen_name
                                break
                        except tweepy.TweepError:
                            # hit rate limit, sleep for 15 minutes
                            print 'Rate limited. Sleeping for 15 minutes.'
                            time.sleep(15 * 60 + 15)
                            continue
                        except StopIteration:
                            break
            else:
                friendids = [int(line.strip().split('\t')[0]) for line in file(fname)] 

            print 'Found %d friends for %s' % (len(friendids), screen_name) 

            # get friends of friends
            cd = current_depth
            if cd+1 < max_depth:
                for fid in friendids[:FRIENDS_OF_FRIENDS_LIMIT]:
                    taboo_list = get_follower_ids(fid, max_depth=max_depth,
                        current_depth=cd+1, taboo_list=taboo_list) 

            if cd+1 < max_depth and len(friendids) > FRIENDS_OF_FRIENDS_LIMIT:
                print 'Not all friends retrieved for %s.' % screen_name 

    except Exception, error:
        print 'Error retrieving followers for user id: ', centre
        print error

        if os.path.exists(fname):
            os.remove(fname)
            print 'Removed file "%s".' % fname 

        sys.exit(1) 

    return taboo_list 

if __name__ == '__main__':
    ap = argparse.ArgumentParser()
    ap.add_argument("-s", "--screen-name", required=True, help="Screen name of twitter user")
    ap.add_argument("-d", "--depth", required=True, type=int, help="How far to follow user network")
    args = vars(ap.parse_args())

    twitter_screenname = args['screen_name']
    depth = int(args['depth']) 

    if depth < 1 or depth > 3:
        print 'Depth value %d is not valid. Valid range is 1-3.' % depth
        sys.exit('Invalid depth argument.')

    print 'Max Depth: %d' % depth
    matches = api.lookup_users(screen_names=[twitter_screenname])

    if len(matches) == 1:
        print get_follower_ids(matches[0].id, max_depth=depth)
    else:
        print 'Sorry, could not find twitter user with screen name: %s' % twitter_screenname
view raw
get_followers.py hosted with ❤ by GitHub

Python file: get_followers.py

I ran this script twice first without a filter on the screen name but limiting the maximum number of following accounts to 20 then again but this time filtering for accounts starting with ‘TED’ (line 102) and allowing up to 200 following accounts to be queried. This will give a mix of TED and non-TED twitter accounts. Running the script:

$ python get_followers.py -s TEDxSingapore -d 3

Max Depth: 3
Found 147 friends for TEDxSingapore
Found 200 friends for TEDWomen
Already been here.
Found 72 friends for TEDxDanteSchool
Found 33 friends for TEDHelp
Retrieving user details for twitter id XXXXXXXX from API... 

... Lots more output ...

Step 2. Process twitter data to generate an output file of relationships between twitter accounts

The script below will process the data collected from the twitter API and generate an edge list. That is a list of relationships between twitter accounts. A weight value is included, this value is the total number of followers for the first twitter account, this value is retrieved from the API. The weight value can be used later to prune the network graph.

import glob
import os
import json
import sys
from collections import defaultdict

users = defaultdict(lambda: { 'followers': 0 })

for f in glob.glob('twitter-users/*.json'):
    data = json.load(file(f))
    screen_name = data['screen_name']
    users[screen_name] = { 'followers': data['followers_count'] }

SEED = 'TEDxSingapore'

def process_follower_list(screen_name, edges=[], depth=0, max_depth=2):
    f = os.path.join('following', screen_name + '.csv') 

    if not os.path.exists(f):
        return edges

    followers = [line.strip().split('\t') for line in file(f)]

    for follower_data in followers:
        if len(follower_data) < 2:
            continue

        screen_name_2 = follower_data[1]

        # use the number of followers for screen_name as the weight
        weight = users[screen_name]['followers']

        edges.append([screen_name, screen_name_2, weight])

        if depth+1 < max_depth:
            process_follower_list(screen_name_2, edges, depth+1, max_depth)

    return edges

edges = process_follower_list(SEED, max_depth=3)

with open('twitter_network.csv', 'w') as outf:
    edge_exists = {}
    for edge in edges:
        key = ','.join([str(x) for x in edge])
        if not(key in edge_exists):
            outf.write('%s\t%s\t%d\n' % (edge[0], edge[1], edge[2]))
            edge_exists[key] = True
view raw

twitter_network.py hosted with ❤ by GitHub

Python file: twitter_network.py

The output generated from this script:

...

TEDxSingapore   trendwatchingAP 12814
adaptev TEDxSingapore   321
IS_magazine     TEDxSingapore   9955
trendwatchingAP TEDxSingapore   678
TEDxSingapore   GuyKawasaki     12814
TEDxSingapore   InnovateAP      12814
TEDxSingapore   InnosightTeam   12814
TEDxSingapore   ScottDAnthony   12814
TEDxSingapore   WorldAndScience 12814
TEDxSingapore   EntMagazine     12814
...  

Step 3. Visualizing the Network using the NetworkX library

We now have all the data we need to generate a network graph. Here are the steps used to visualize the network graph:

• Create a directed graph (net.DiGraph) containing all the edge data including metadata.
• Remove nodes based on how connected they are to other nodes in the network (i.e. remove poorly connected nodes)
• Remove edges that have less than a minimum number of followers
• Split nodes into two separate categories, ‘TED’ and ‘non-TED’ sets.
• Render each nodeset
• Render edges between nodes
• Render node labels

Here is the code to generate the twitter network image. I wrote this code in IPython Notebook (this is the reason Line 3 has a magic command that causes matplotlib output to be rendered in the browser):

import networkx as net
import matplotlib.pyplot as plt

from collections import defaultdict
import math

twitter_network = [ line.strip().split('\t') for line in file('twitter_network.csv') ]

o = net.DiGraph()
hfollowers = defaultdict(lambda: 0)
for (twitter_user, followed_by, followers) in twitter_network:
    o.add_edge(twitter_user, followed_by, followers=int(followers))
    hfollowers[twitter_user] = int(followers)

SEED = 'TEDxSingapore'

# centre around the SEED node and set radius of graph
g = net.DiGraph(net.ego_graph(o, SEED, radius=4))

def trim_degrees_ted(g, degree=1, ted_degree=1):
    g2 = g.copy()
    d = net.degree(g2)
    for n in g2.nodes():
        if n == SEED: continue # don't prune the SEED node
        if d[n] <= degree and not n.lower().startswith('ted'):
            g2.remove_node(n)
        elif n.lower().startswith('ted') and d[n] <= ted_degree:
            g2.remove_node(n)
    return g2

def trim_edges_ted(g, weight=1, ted_weight=10):
    g2 = net.DiGraph()
    for f, to, edata in g.edges_iter(data=True):
        if f == SEED or to == SEED: # keep edges that link to the SEED node
            g2.add_edge(f, to, edata)
        elif f.lower().startswith('ted') or to.lower().startswith('ted'):
            if edata['followers'] >= ted_weight:
                g2.add_edge(f, to, edata)
        elif edata['followers'] >= weight:
            g2.add_edge(f, to, edata)
    return g2

print 'g: ', len(g)
core = trim_degrees_ted(g, degree=235, ted_degree=1)
print 'core after node pruning: ', len(core)
core = trim_edges_ted(core, weight=250000, ted_weight=35000)
print 'core after edge pruning: ', len(core)

nodeset_types = { 'TED': lambda s: s.lower().startswith('ted'), 'Not TED': lambda s: not s.lower().startswith('ted') }

nodesets = defaultdict(list)

for nodeset_typename, nodeset_test in nodeset_types.iteritems():
    nodesets[nodeset_typename] = [ n for n in core.nodes_iter() if nodeset_test(n) ]

pos = net.spring_layout(core) # compute layout 

colours = ['red','green']
colourmap = {}

plt.figure(figsize=(18,18))
plt.axis('off')

# draw nodes
i = 0
alphas = {'TED': 0.6, 'Not TED': 0.4}
for k in nodesets.keys():
    ns = [ math.log10(hfollowers[n]+1) * 80 for n in nodesets[k] ]
    print k, len(ns)
    net.draw_networkx_nodes(core, pos, nodelist=nodesets[k], node_size=ns, node_color=colours[i], alpha=alphas[k])
    colourmap[k] = colours[i]
    i += 1
print 'colourmap: ', colourmap

# draw edges
net.draw_networkx_edges(core, pos, width=0.5, alpha=0.5)

# draw labels
alphas = { 'TED': 1.0, 'Not TED': 0.5}
for k in nodesets.keys():
    for n in nodesets[k]:
        x, y = pos[n]
        plt.text(x, y+0.02, s=n, alpha=alphas[k], horizontalalignment='center', fontsize=9)
view raw

visualize.py hosted with ❤ by GitHub

Python file: visualize.py

• Line 7 Load edge data from disk
• Line 9-13 Create a directed graph from the edge data and populate a dictionary with the followers count data
• Line 18 Centre and restrict size of graph around the SEED node (TEDxSingapore)
• Line 20-29 Method to prune the network graph by eliminating nodes that don’t meet filter criteria
• Line 31-41 Method to prune the network graph by eliminating edges that don’t meet filter criteria
• Line 44, 46 removes nodes and edges from the network that don’t meet the filter criteria
• Line 67-73 For each nodeset draw the nodes, the size of each node is based on the log value of the followers count
• Line 76 Draw network edges
• Line 80-83 Draw network labels, use matplotlib directly to do this rather than net.draw_networkx_labels() method.

Output from running script in IPython Notebook

g:  119567
core after node pruning:  958
core after edge pruning:  198
Not TED 38
TED 160
colourmap:  {'Not TED': 'red', 'TED': 'green'}

twitter network

See Also:

• NetworkX library
• Social Network Analysis for Startups by Maksim Tsvetovat; Alexander Kouzetsov
• Snowball Samping

Referensi

• http://mark-kay.net/2014/08/15/network-graph-of-twitter-followers/