The graphs are nice. But I think raw histograms of the most used commands, and a table with the highest correlations per given command would also useful and more digestible.
If you still have the per user data around, I'd also love to see the most often used commands by the community that I never or very rarely use. Maybe I'll find a gem of a unix command I didn't know about yet.
I've spent quite some time looking at those histograms, but I haven't found a good way to display it usefully yet.
The most popular commands are 'git', 'ls', and 'cd', and the table of most used commands is too long to be able to easily digest. Any suggestions would be appreciated, though!
It'd be awesome to see a dump of the data so that we can examine it on our own - many people may think of ways to use it that you haven't, or draw different conclusions from it. :D
Yeah! This is something that I'm interested in, but there are a bunch of privacy issues with a raw data dump (references to specific files/URLs and possibly passwords), and I haven't gotten around to making sure it's safe yet.
(note: I haven't actually seen any passwords, but that doesn't mean there aren't any)
Oh, that totally skipped my mind. Maybe run something to recognize URLs (replace with $URL, or some kind of placeholder), and then try to obfuscate filenames and the like in a similar way? By normalizing the data like this, you could get much better results with regards to command line switches and the like.
I would have to guess that the first "word" with the command only is clean? Have you seen any evidence to the contrary? That data set alone would tell an interesting story.
In general, a very nice thing and thanks for sharing! But...
1. Are you sharing the raw data? (It would be great!)
2. It would be useful to see frequencies (e.g. as sizes of nodes).
3. Why we cannot see 'git', 'ls', 'cd'?
4. At the first glance things like "pytho", "sourc", "worko" look like a glitch.
Also, when there is a cluster of commands starting with the same thing, the subgraph is hardly informative. How about scaling text (instead of cutting them)?
5. When it comes to a measure of co-occurrences, a nicer quantity than correlation is the following - http://stats.stackexchange.com/questions/6047/does-this-quan... with a direct interpretation of "how does the observed coincidence rate correspond to the expected one for independent variables".
I used it a few times (after testing other measures of co-occurrences (also: conditional probability) and being dissatisfied by results, especially ones favouring edges for big or small nodes). Examples (with their recipes) below:
If you still have the per user data around, I'd also love to see the most often used commands by the community that I never or very rarely use. Maybe I'll find a gem of a unix command I didn't know about yet.