This notebook shows how BigBang can help you explore a mailing list archive.
First, use this IPython magic to tell the notebook to display matplotlib graphics inline. This is a nice way to display results.
In [1]:
%matplotlib inline
Import the BigBang modules as needed. These should be in your Python environment if you've installed BigBang correctly.
In [2]:
import bigbang.mailman as mailman
import bigbang.graph as graph
import bigbang.process as process
from bigbang.parse import get_date
#from bigbang.functions import *
from bigbang.archive import Archive
Also, let's import a number of other dependencies we'll use later.
In [4]:
import pandas as pd
import datetime
import matplotlib.pyplot as plt
import numpy as np
import math
import pytz
import pickle
import os
pd.options.display.mpl_style = 'default' # pandas has a set of preferred graph formatting options
Now let's load the data for analysis.
In [4]:
urls = ["http://mail.scipy.org/pipermail/ipython-user/",
"http://mail.scipy.org/pipermail/scipy-dev/",
"http://mail.scipy.org/pipermail/scipy-user/",
"http://mail.scipy.org/pipermail/numpy-discussion/"]
archives = [Archive(url,archive_dir="../archives") for url in urls]
activities = [arx.get_activity() for arx in archives]
This variable is for the range of days used in computing rolling averages.
In [5]:
window = 100
For each of the mailing lists we are looking at, plot the rolling average of number of emails sent per day.
In [6]:
plt.figure(figsize=(12.5, 7.5))
for i, activity in enumerate(activities):
colors = 'rgbkm'
ta = activity.sum(1)
rmta = pd.rolling_mean(ta,window)
rmtadna = rmta.dropna()
plt.plot_date(rmtadna.index,
rmtadna.values,
colors[i],
label=mailman.get_list_name(urls[i]) + ' activity',xdate=True)
plt.savefig("activites-marked.png")
plt.show()
Now, let's see: who are the authors of the most messages to one particular list?
In [7]:
a = activities[0] # for the first mailing list
ta = a.sum(0) # sum along the first axis
ta.sort()
ta[-10:].plot(kind='barh')
Out[7]:
This might be useful for seeing the distribution (does the top message sender dominate?) or for identifying key participants to talk to.
Many mailing lists will have some duplicate senders: individuals who use multiple email addresses or are recorded as different senders when using the same email address. We want to identify those potential duplicates in order to get a more accurate representation of the distribution of senders.
To begin with, let's do a naive calculation of the similarity of the From strings, based on the Levenshtein distance.
This can take a long time for a large matrix, so we will truncate it for purposes of demonstration.
In [9]:
import Levenshtein
distancedf = process.matricize(a.columns[:100], lambda a,b: Levenshtein.distance(a,b)) # calculate the edit distance between the two From titles
df = distancedf.astype(int) # specify that the values in the matrix are integers
In [10]:
fig = plt.figure(figsize=(18, 18))
plt.pcolor(df)
#plt.yticks(np.arange(0.5, len(df.index), 1), df.index) # these lines would show labels, but that gets messy
#plt.xticks(np.arange(0.5, len(df.columns), 1), df.columns)
plt.show()
The dark blue diagonal is comparing an entry to itself (we know the distance is zero in that case), but a few other dark blue patches suggest there are duplicates even using this most naive measure.
Below is a variant of the visualization for inspecting the particular apparent duplicates.
In [11]:
levdf = process.sorted_lev(a) # creates a slightly more nuanced edit distance matrix
# and sorts by rows/columns that have the best candidates
levdf_corner = levdf.iloc[:25,:25] # just take the top 25
In [12]:
fig = plt.figure(figsize=(15, 12))
plt.pcolor(levdf_corner)
plt.yticks(np.arange(0.5, len(levdf_corner.index), 1), levdf_corner.index)
plt.xticks(np.arange(0.5, len(levdf_corner.columns), 1), levdf_corner.columns, rotation='vertical')
plt.colorbar()
plt.show()
For this still naive measure (edit distance on a normalized string), it appears that there are many duplicates in the <10 range, but that above that the edit distance of short email addresses at common domain names can take over.
In [13]:
consolidates = []
# gather pairs of names which have a distance of less than 10
for col in levdf.columns:
for index, value in levdf.loc[levdf[col] < 10, col].iteritems():
if index != col: # the name shouldn't be a pair for itself
consolidates.append((col, index))
print str(len(consolidates)) + ' candidates for consolidation.'
In [14]:
c = process.consolidate_senders_activity(a, consolidates)
print 'We removed: ' + str(len(a.columns) - len(c.columns)) + ' columns.'
We can create the same color plot with the consolidated dataframe to see how the distribution has changed.
In [15]:
lev_c = process.sorted_lev(c)
levc_corner = lev_c.iloc[:25,:25]
fig = plt.figure(figsize=(15, 12))
plt.pcolor(levc_corner)
plt.yticks(np.arange(0.5, len(levc_corner.index), 1), levc_corner.index)
plt.xticks(np.arange(0.5, len(levc_corner.columns), 1), levc_corner.columns, rotation='vertical')
plt.colorbar()
plt.show()
Of course, there are still some duplicates, mostly people who are using the same name, but with a different email address at an unrelated domain name.
How does our consolidation affect the graph of distribution of senders?
In [17]:
fig, axes = plt.subplots(nrows=2, figsize=(15, 12))
ta = a.sum(0) # sum along the first axis
ta.sort()
ta[-20:].plot(kind='barh',ax=axes[0], title='Before consolidation')
tc = c.sum(0)
tc.sort()
tc[-20:].plot(kind='barh',ax=axes[1], title='After consolidation')
plt.show()
Okay, not dramatically different, but the consolidation makes the head heavier. There are more people close to that high end, a stronger core group and less a power distribution smoothly from one or two people.