Context Email archives are a great source of information about the real-world social networks people are generally most involved in. Although sharing of full email exchanges is almost never a good idea, flow metadata (i.e. who sent a message to whom, and when) can be **anonymized** quite effectively and still carry a lot of information. I'm sharing over 10 years of flow metadata from my work and personal email accounts to enable data scientists experiment with their favourite statistics and social network analysis tools. A getting-started notebook is available [here](https://www.kaggle.com/emarock/getting-started-with-email-flows). For anyone willing to extract similar datasets from their own email accounts, the tool I put together for producing mine is available at [https://github.com/emarock/mailfix](https://github.com/emarock/mailfix) (currently supports extraction from Gmail accounts, IMAP accounts and Apple Mail on macOS). Content This dataset contains two files: - `work.csv`: email flow metadata from my work account (~146,000 emails, from 2005 to 2018) - `personal.csv`: email flow metadata from my personal account (~41,000 emails, from 2006 to 2018) As one should expect from any decade long archive, the data presents some partial corruptions and anomalies, that are however time-confined and should be easily identified and filtered out through basic statistical analysis. I will be happy to discuss and provide more information in the comments. Inspiration Basic exploration: - Who am I? - Who's human and who's not? How different are attention-seekers from mailing list engines? - How did my communication patterns change over time? Did they change in the same way in and out of work? - Did my social network grow? Did it shrink? - Who's my boss? Who were my former ones? Who'll be the next one? I will be also available to extend the dataset with additional data for training advanced classifiers (e.g. lists of actual humans, mailing lists, VIPs...). Feel free to ask in the comments. Anonymization and Privacy Note The anonymization function (code [here](https://github.com/emarock/mailfix/blob/master/lib/anonymizer.js), tests [here](https://github.com/emarock/mailfix/blob/master/test/anonymizer.js)) is based on [djb2 string hashing](http://www.cse.yorku.ca/~oz/hash.html) and on a [Mersenne Twister pseudorandom generator](https://en.wikipedia.org/wiki/Mersenne_Twister), implemented in the [string-hash](https://www.npmjs.com/package/string-hash) and [casual](https://www.npmjs.com/package/casual) node.js modules. It should be practically irreversible, modulo implementation defects. However, if you've ever been involved in email exchanges with me, you can work your way back to the anonymized address associated to your actual address by comparing the message timestamps. Similarly, with a little more guesswork, you can discover the anonymized addresses of those who were also involved in those exchanges. Since that is also true for them in respect to you, if that is of any concern just reach out and I'll censor the problematic entries in the dataset.