Up to this point in the book, you have become quite familiar with our mbox-short.txt and mbox.txt data files. Now it is time to take our analysis of e-mail data to the next level.
In the real world, sometimes you have to pull down mail data from servers and that might take quite some time and the data might be inconsistent, error filled and need a lot of cleanup or adjustment. In this section, we work with an application that is the most complex so far and pull down nearly a gigabyte of data and visualize it.
You can download this application from:
We will be using data from a free e-mail list archiving service called www.gmane. org. This service is very popular with open-source projects because it provides a nice searchable archive of their e-mail activity. They also have a very liberal policy regarding accessing their data through their API. They have no rate limits, but ask that you don’t overload their service and take only the data you need. You can read gmane’s terms and conditions at this page:
It is very important that you make use of the gmane.org data responsibly by adding delays to your access of their services and spreading long-running jobs over a longer period of time. Do not abuse this free service and ruin it for the rest of us.
When the Sakai e-mail data was spidered using this software, it produced nearly a Gigabyte of data and took a number of runs on several days. The file README.txt in the above ZIP may have instructions as to how you can download a pre-spidered copy of the content.sqlite file for a majority of the Sakai email corpus so you don’t have to spider for five days just to run the programs. If you download the pre-spidered content, you should still run the spidering process to catch up with more recent messages.
The first step is to spider the gmane repository. The base URL is hard-coded in the gmane.py and is hard-coded to the Sakai developer list. You can spider another repository by changing that base url. Make sure to delete the content.sqlite file if you switch the base url.
The gmane.py file operates as a responsible caching spider in that it runs slowly and retrieves one mail message per second so as to avoid getting throttled by gmane. It stores all of its data in a database and can be interrupted and re-started as often as needed. It may take many hours to pull all the data down. So you may need to restart several times.
Here is a run of gmane.py retrieving the last five messages of the Sakai developer list:
How many messages:10http://download.gmane.org/gmane.comp.cms.sakai.devel/51410/51411 9460 email@example.com 2013-04-05 re: [building ...http://download.gmane.org/gmane.comp.cms.sakai.devel/51411/51412 3379 firstname.lastname@example.org 2013-04-06 re: [building ...http://download.gmane.org/gmane.comp.cms.sakai.devel/51412/51413 9903 email@example.com 2013-04-05 [building sakai] melete 2.9 oracle ...http://download.gmane.org/gmane.comp.cms.sakai.devel/51413/51414 349265 firstname.lastname@example.org 2013-04-07 [building sakai] ...http://download.gmane.org/gmane.comp.cms.sakai.devel/51414/51415 3481 email@example.com 2013-04-07 re: ...http://download.gmane.org/gmane.comp.cms.sakai.devel/51415/51416 0 Does not start with From
The program scans content.sqlite from one up to the first message number not already spidered and starts spidering at that message. It continues spidering until it has spidered the desired number of messages or it reaches a page that does not appear to be a properly formatted message.
Sometimes gmane.org is missing a message. Perhaps administrators can delete messages or perhaps they get lost. If your spider stops, and it seems it has hit a missing message, go into the SQLite Manager and add a row with the missing id leaving all the other fields blank and restart gmane.py. This will unstick the spidering process and allow it to continue. These empty messages will be ignored in the next phase of the process.
One nice thing is that once you have spidered all of the messages and have them in content.sqlite, you can run gmane.py again to get new messages as they get sent to the list.
The content.sqlite data is pretty raw, with an inefficient data model, and not compressed. This is intentional as it allows you to look at content.sqlite in the SQLite Manager to debug problems with the spidering process. It would be a bad idea to run any queries against this database as they would be quite slow.
The second process is to run the program gmodel.py. This program reads the rough/raw data from content.sqlite and produces a cleaned-up and well-modeled version of the data in the file index.sqlite. The file index.sqlite will be much smaller (often 10X smaller) than content.sqlite because it also compresses the header and body text.
Each time gmodel.py runs - it deletes and re-builds index.sqlite, allowing you to adjust its parameters and edit the mapping tables in content.sqlite to tweak the data cleaning process. This is a sample run of gmodel.py. It prints a line out each time 250 mail messages are processed so you can see some progress happening as this program may run for a while processing nearly a Gigabyte of mail data.
Loaded allsenders 1588 and mapping 28 dns mapping 11 2005-12-08T23:34:30-06:00 firstname.lastname@example.org 2005-12-22T10:03:20-08:00 email@example.com 2006-01-12T11:17:34-05:00 firstname.lastname@example.org 2006-01-24T11:13:28-08:00 email@example.com...
The gmodel.py program handles a number of data cleaning tasks.
Domain names are truncated to two levels for .com, .org, .edu, and .net. Other domain names are truncated to three levels. So si.umich.edu becomes umich.edu and caret.cam.ac.uk becomes cam.ac.uk. Also e-mail addresses are forced to lower case and some of the @gmane.org address like the following
are converted to the real address whenever there is a matching real e-mail address elsewhere in the message corpus.
If you look in the content.sqlite database there are two tables that allow you to map both domain names and individual e-mail addresses that change over the lifetime of the e-mail list. For example, Steve Githens used the following e-mail addresses as he changed jobs over the life of the Sakai developer list:
We can add two entries to the Mapping table in content.sqlite so gmodel.py will map all three to one address:
firstname.lastname@example.org -> email@example.com@cam.ac.uk -> firstname.lastname@example.org
You can also make similar entries in the DNSMapping table if there are multiple DNS names you want mapped to a single DNS. The following mapping was added to the the Sakai data:
iupui.edu -> indiana.edu
So all the accounts from the various Indiana University campuses are tracked together.
You can re-run the gmodel.py over and over as you look at the data, and add mappings to make the data cleaner and cleaner. When you are done, you will have a nicely indexed version of the e-mail in index.sqlite. This is the file to use to do data analysis. With this file, data analysis will be really quick.
The first, simplest data analysis is to determine ”who sent the most mail?” and ”which organization sent the most mail”? This is done using gbasic.py:
How many to dump? 5Loaded messages= 51330 subjects= 25033 senders= 1584 Top 5 Email list email@example.com firstname.lastname@example.org email@example.com firstname.lastname@example.org email@example.com 1184 Top 5 Email list organizationsgmail.com 7339umich.edu 6243uct.ac.za 2451indiana.edu 2258unicon.net 2055
Note how much more quickly gbasic.py runs compared to gmane.py or even gmodel.py. They are all working on the same data, but gbasic.py is using the compressed and normalized data in index.sqlite. If you have a lot of data to manage, a multi-step process like the one in this application may take a little longer to develop, but will save you a lot of time when you really start to explore and visualize your data.
You can produce a simple visualization of the word frequency in the subject lines in the file gword.py:
Range of counts: 33229 129Output written to gword.jsonp
This produces the file gword.jsonp which you can visualize using gword.htm to produce a word cloud similar to the one at the beginning of this section.
A second visualization is produced by gline.py. It computes e-mail participation by organizations over time.
Loaded messages= 51330 subjects= 25033 senders= 1584Top 10 Oranizations['gmail.com', 'umich.edu', 'uct.ac.za', 'indiana.edu','unicon.net', 'tfd.co.uk', 'berkeley.edu', 'longsight.com','stanford.edu', 'ox.ac.uk']Output written to gline.jsonp
Its output is written to gline.jsonp which is visualized using gline.htm.
This is a relatively complex and sophisticated application and has features to do some real data retrieval, cleaning and visualization.