Infosec, meet data science

I know you’ve been friends for a while, but I hear you’re starting to get closer, and maybe there are some things you need to know about each other. And since part of my job is using my data skills to help secure information assets, it’s time that I put some thoughts down on paper… er… pixels.

Infosec and data science have a lot in common: they’re both about really really understanding systems, and they’re both about really understanding people and their behaviors, and acting on that information to protect or exploit those systems.  It’s no secret that military infosec and counterint people have been working with machine learning and other AI algorithms for years (I think I have a couple of old papers on that myself), or that data scientists and engineers are including practical security and risk in their data governance measures, but I’m starting to see more profound crossovers between the two.

Take data risk, for instance. I’ve spent the past few years as part of the conversation on the new risks from both doing data science and its components like data visualization (the responsible data forum is a good place to look for this): that there is risk to everyone involved in the data chain, from subject and collectors through to processors and end-product users, and that what we need to secure goes way beyond atomic information like EINs and SSNs, to the products and actions that could be generated by combining data points.  That in itself is going to make infosec harder: there will be incursions (or, if data’s coming out from outside, excursions), and tracing what was collected, when and why is becoming a lot more subtle.  Data scientists are also good at subtle: some of the Bayes-based pattern-of-life tools and time-series anomaly algorithms are well-bounded things of beauty.  But it’s not all about DS; also in those conversations have been infosec people who understand how to model threats and risks, and help secure those data chains from harm (I think I have some old talks on that too somewhere, from back in my crisismapping days).

There are also differences.  As a data scientist, I often have the luxury of time: I can think about a system, find datasets, make and test hypotheses and consider the veracity and inherent risks in what I’m doing over days, weeks or sometimes months.  As someone responding to incursion attempts (and yes, it’s already happening, it’s always already happening), it’s often in the moment or shortly after, and the days, weeks or months are taken in preparation and precautions.  Data scientists often play 3d postal chess; infosec can be more like Union-rules rugby, including the part where you’re all muddy and not totally sure who’s on your side any more.

Which isn’t to say that data science doesn’t get real-time and reactive: we’re often the first people to spot that something’s wrong in big streaming data, and the pattern skills we have can both search for and trace unusual events, but much of our craft to date has been more one-shot and deliberate (“help us understand and optimise this system”). Sometimes we realize a long time later that we were reactive (like realizing recently that mappers have been tracking and rejecting information injection attempts back to at least 2010 – yay for decent verification processes!). But even in real-time we have strengths: a lot of data engineering is about scaling data science processes in both volume and time, and work on finding patterns and reducing reaction times in areas ranging from legal discovery (large-scale text analysis) to manufacturing and infrastructure (e.g. not-easy-to-predict power flows) can also be applied to security.

Both infosec and data scientists have to think dangerously: what’s wrong with this data, these algorithms, this system (how is it biased, what is it missing, how is it wrong); how do I attack this system, how can I game these people; how do I make this bad thing least-worst given the information and resources I have available, and that can get us both into difficult ethical territory.  A combination of modern data science and infosec skills means I could gather data on and profile all the people I work with, and know things like their patterns of life and potential vulnerabilities to e.g. phishing attempts, but the ethics of that is very murky: there’s a very very fine line between protection and being seriously creepy (yep, another thing I work on sometimes).  Equally, knowing those patterns of life could help a lot in spotting non-normal behaviours on the inside of our systems (because infosec has gone way beyond just securing the boundaries now), and some of our data summary and anonymisation techniques could be helpful here too.  Luckily much of what I deal with is less ethically murky: system and data access logs, with known vulnerabilities, known data and motivations, and I work with a most wonderfully evil and detailed security nerd.  But we still have a lot to learn from each other.  Back in the Cold War days (the original Cold War, not the one that seems to be restarting now), every time we designed a system, we also designed countermeasures to it, often drawing on disciplines far outside the original system’s scope.  That seems to be core to the infosec art, and data science would seem to be one of those disciplines that could help.

Notes from John Sarapata’s talk on online responses to organised adversaries

John Sarapata (@JohnSarapata) = head of engineering at Jigsaw  (= new name for Google Ideas).  Jigsaw = “the group at Google that tries to help users facing organized violence and oppression”.  A common thread in their work is that they’re dealing with the outputs from organized adversaries, e.g. governments, online mobs, extremist groups like ISIS.
One example project is, which looks for people who are searching for extremist connections (e.g. ISIS) and shows them content from a different point of view, e.g. a user searching for travel to Aleppo might be shown realistic video of conditions there. [IMHO this is a useful application of social engineering in a clear-cut situation; threats and responses in other situations may be more subtle than this (e.g. what does ‘realistic’ mean in a political context?).]
The Jigsaw team is looking at threats and counters at 3 levels of the tech stack:
  • device/user: activities are consume and create content; threats include attacks by governments, phishing, surveillance, brigading, intimidation
  • wire: activities are find then transfer; threats include DNS hijacking, TOR bridge probes
  • server: activities are hosting; threats include DDOS
[They also appear to be looking at threats and counters on a meta level (e.g. the social hack above).]
Examples of emergent countermeasures outside the team include people bypassing censorship in Turkey by using Google’s public DNS, and people in China after the 2008 Szechwan earthquake posting images of school collapses and investigating links between these (ultimately leading to finding links between collapses, school contractors using substandard concrete and officials being bribed to ignore this) despite government denial of issues.  These are about both reading and generating content, both of which need to be protected.
There are still unsolved problems, for example communications inside a government firewall.  Firewalls (e.g. China’s Great Firewall) generally have slow external pipes with internal alternatives (e.g. Sino Weibo), so people tend to consume information from inside. Communication of external information inside a firewall isn’t solved yet, e.g mesh networks aren’t great; the use of thumb drives to share information in Cuba was one way around this, but there’s still more to do.  [This comment interested me because that’s exactly the situation we’ve been dealing with in crises over the past few years: using sneakernet/ mopeds,  point-to-point, meshes etc., and there may be things to learn in both directions.]
Example Jigsaw projects and apps include:
  •, still in beta: creates a knowledge graph when Google scans news stories (this is language independent). One of the cooler uses of this is being able to find things that are reported on in every country except yours (e.g. Russia, China not showing articles on the Panama Papers).
  • Anti-phishing: team used stuff from Google’s security team for this, e.g. using Password Alert (alerts when user e.g. puts their company password into a non-company site) on Google accounts.
  • Government Attack Warning. Google can see attacks on gmail, google drive etc accounts: when a user logs in, Google displays a message to them about a detected attack, including what they could do.
  • Conversation AI. Internet discussions aren’t always civil, e.g. 20-25 governments including China and Russia have troll armies now, amplified by bots (brigading); conversation AI is machine classification/detection of abuse/harassment in text; the Jigsaw team is working on machine learning approaches together with the youtube comment cleanup team.  The team’s considered the tension that exists between free speech and reducing threats: their response is that detection apps must lay out values, and Jigsaw values include that conversation algorithms are community specific, e.g. each community decides its limits on swearing etc.; a good example of this is Riot Games. [This mirrors a lot of the community-specific work by community of community groups like the Community Leadership Forum].  Three examples of communities using Conversation AI: a Youtube feature that flags potential abuse to channel owners (launching Nov 2016). Wikipedia: flagging personal attacks (e.g. you are full of shit) in talk pages (Wikipedia has a problem with falling numbers of editors, partly because of this). New York Times: scaling existing human moderation of website comments (NYT currently turns off comments on 90% of pages because they don’t have enough human moderators). “NYT has lots of data, good results”.  Team got interesting data on how abuse spreads after releasing a photo of women talking to their team about #gamergate, then watching attackers discuss online (4chan etc) who of those women to attack and how, and the subsequent attacks.
  • Firehook: censorship circumvention. Jigsaw has the Uproxy plugin  for peer-to-peer information sharing across censorship boundaries (article), but needs to do more, eg look at the whole ecosystem.  Most people use proxy servers (e.g. VPNs), but a government could disallow VPNs: we need many different proxies and ways to hide them.  Currently using WebRTC for peer to peer proxies (e.g. Germany to Turkey using e.g NAT hole punching), collateral freedom and domain fronting, e.g. GreatFire routing New York Times articles through Amazon and GitHub.  Domain fronting (David Fyfield article) uses the fact that e.g. CloudFlare hosts many sites: the user connects to an allowed host, https encrypts it, then uses the encrypted header to go to a blocked site on the same host.  There are still counters to this; China first switched off GitHub access (then had to restore it), and used the Great Cannon to counter GreatFire, e.g. every 100th load of Baidu Analytics injects malware into external machines and creates a DDOS botnet. NB the firewall here was on path, not in path: a machine off to one side listens for banned words and breaks connections, but Great Cannon is inside the connection; and with current access across the great firewall, people see different pages based on who’s browsing.
  • DDOS: (with Arbor Networks), shows DDOS in real time. Digital attacks are now mirroring physical ones, and during recent attacks, e.g. Hong Kong’s Umbrella Revolution, Jigsaw protected sites on both sides (using Project Shield, below) because Google thinks some things, like DDOS, are unfair.  Interesting point: trolling as the human equivalent of DDOS [how far can this comparison go in designing potential counters?].
  • Project Shield: reused Google’s PSS (Page Speed Service) to protect news sites, human rights organization etc from DDOS attacks. Sites are on Google cloud: can scale up number of VMs used and nginx allows clever uses with reverse proxies, cookie challenges etc.  Example site: Krebs on Security was being DDOSed (nb a DDOS attack on a site costs about $50 online), moved from host Akamai to Google with Project Shield. Team is tracking Twitter user bragging about this and other attacks (TL;DR: IoT attack, e.g. baby monitors; big botnet, Mirai botnet source code now released, brought down Twitter, Snapchat).  Krebs currently getting about 5 attacks a day, e.g. brute-force, Slowloris, Hulk (bandwidth, syn flood, post flood, cache busting, WordPress pingback etc), and Jigsaw gets the world’s best DDOSers hacking and testing their services.
Audience questions:
  • Qs: protecting democracy in US, e.g. botnets, online harassment etc.? A: don’t serve specific countries but worldwide.
  • Q: google autofill encouraging hatespeech? A: google reflects the world as it is; google search reflects what people do, holds a mirror back up to you. Researching machine learning bias on google suggest results and bias in training data.  Don’t want to censor, but don’t want to propagate bad things.
  • Q: not censor but inform, can you e.g. tell a user “your baby monitor is hacked”? A: privacy issue, eg g connecting kit and emails to ip addresses.
  • Q: people visiting google site to attack… spamming google auto complete, algorithms? A: if google detects people gaming them, will come down hard.
  • Q: standard for “abusive”? how to compare with human?. A: is training data, Wikipedia is saying what’s abusive. Is all people in the end.
  • Q: how deal with e.g. misinformation? A: unsolved problem, politically sensitive, e.g. who gets to decide what’s fake and true? censorship and harassment work will take time.
  • Q: why Twitter not on list of content? A: Twitter might not want this, team is resource constrained, e.g. NYT models are useless on youtube because NYT folks use proper grammar and spelling.
  • Q: how to decide what to intervene in?A:  e.g. Google takes sides against ISIS, who are off the charts on eg genocide.
  • Q: AWS Shield; does Google want to commercialise their stuff? A: no. NB CloudFlare Galileo is also response to google work.
  • Q: biggest emerging tech threat online? Brigading, e.g. groups of people and bots. This breaches physical and online, includes eg physical threats and violence, and is hard to detect and attribute.
Other references:

Why am I writing about belief?

[Cross-post from LinkedIn]

I’ve been meaning to write a set of sessions on computational belief for a while now, based on the work I’ve done over the years on belief, reasoning, artificial intelligence and community beliefs. With all that’s happening in our world now, both online and in the “real world”, I believe that the time has come to do this.

We could start with truth. We often talk about ‘true’ and ‘false’ as though they’re immovable things: that every statement should be able to be assigned one of these values. But it’s a little more complicated than that. What we see as ‘true’ is often the result of a judgement we made, given our perception and experience of the world, that a belief is close enough to certain to be ‘true’.

But what is there are no objective truths? In robotics, we talk about “ground truth” and the “god’s eye view” of the world: the knowledge of the world that our robots (or computer vision or reasoning systems) would have if they had perfect information about the world. We talk about things like the “frame problem”, where a system’s ability to reason and act is limited by the “frame” that it has around the world, and the “naughty baby problem” of outside influences that it has no awareness of and cannot plan for. We accept that a robot’s version of “truth” is limited to what it can perceive. But humans being are also limited by their perceptions of the world, by the amount of information available to them. Without going all “Matrix” on you, is it possible that we too are wrong about our “truths”, and we’re not truly objective in reasoning about them because there is no “God’s Eye View” that we can access?

For now, let’s put aside Godel’s theorem and the ‘undecidable’ sentences like “this sentence is false” that can’t be assigned a true or false value, and think about what happens in a world where we all have only perception and consensus agreements on ‘reality’, and nobody has perfect information. One of the things that happens is that we stop talking about “true” and “false”, and start talking about perception: what we can reasonably believe to be true or false (or undecidable), the uncertainty we have about those beliefs, influence and what it might take in terms of evidence or new information to change them. In psychology, that gets us into the theories of mind and reasoning like cognitive psychology and studies of people like Aspergers individuals who process ‘facts’ differently; in design, into things like social engineering and the theory of change. In maths and AI, that gets us into territories like multi-state logics (which beliefs are possible, necessary etc) and both frequentist (‘what happened”) and Bayesian (“what if”) statistics. We might also shift our focus, and talk not of beliefs, but of what we are trying to achieve with them, getting us into theories of actions, influence and decisions (hello, robotics and operational research). There are many theories of uncertainty, but for now probability theory is dominant, so its good to spend time with and understand how that works under the hood.

Although this may all seem abstract hand-wavey, late-night-discussiony “are we living in the Matrix already”, theories of belief have many practical applications. They’re used heavily in data science, and underpin decisions on things like technology designs (via e.g. AB testing), political information release and propaganda (which are usually not the same thing). We need to talk about that too, because the tools and the terrain available (e.g. the internet) are already more powerful than their current uses. One thing we’re becoming painfully aware of now is how belief functions in groups; the use of influence, multiple separate sources of information and repetition to spread beliefs that compete with each other, and the role of things like desire in those beliefs. We’re also learning that systems we build based on human outputs (e.g. Internet-based AI) have the same biases in belief as the humans; an obvious-but-not-obvious thing that we need to recognise and handle. There are useful theories for this too, ranging from the maths of multiple viewpoints to techniques used to both create and make sense of competing views (ACH, information incest detection, phemes). There are also theories of how humans think in groups, and how they can be persuaded to or do change their minds, both rapidly and slowly over time (e.g. game theory, creativity theory and the study of both human and scientific revolutions).

I’ve spent a lot of my life thinking about and applying the theories above, but I’ve never really got round (apart from the odd note on intelligence, data science or the risks inherent in processing data about people) to writing it all down. The session notes are, I hope, a start on this, and even if nobody else reads them, it’ll be fun to do some targeted thinking around them.

Data Science Tools, or what’s in my e-backpack?

One of the infuriating (and at the same time, strangely cool) things about development data science is that you quite often find yourself in the middle of nowhere with a job to do, and no access to the Internet (although this doesn’t happen as often as many Westerners think: there is real internet in most of the world’s cities, honest!).

Which means you get to do your job with exactly what you remembered to pack on your laptop: tools, code, help files, datasets and academic papers.  This is where we talk about the tools.

The Ds4B toolset

This is the tools list for the DS4B course (if you’re following the course, don’t panic: install notes are here).

Offline toolset:

Online toolset (ie. is useful, but not available when you’re offline):

Python and R both come with a bunch of useful libraries in Anaconda (here’s the Python and R lists).  The (Anaconda 4.0) pre-installed libraries used in DS4B are:

  • Python: Basemap, BeautifulSoup, Dask, Matplotlib, NetworkX, Nltk, Numba, 
    Numpy, Pandas, Requests, Scikit-image, Scikit-learn, SeaBorn, Shapely, Sqlite3
  • R: ggplot2, frame, lm

The course also uses some Python libraries that don’t come with the Anaconda standard install. These are:

  • Csv, DateTime, Facepy, Fiona, Gdal, Gdalconst, Geopy, Googlemaps, 
    Json, Ogr2ogr, Osgeo, Pyspark, Re, Twitter

How to get each of these is listed in the DS4B course instructions, but usually isn’t more onerous than typing “pip install libraryname” in the terminal window.  Fiona might give you some trouble (“cannot find the gdal.h library”): if this happens, there are notes in the install instructions.

Choosing the course toolset

For the course, we deliberately biased towards tools that were:

  • Free (because they need to be accessible for everyone, not just people who can pay),
  • Open source with good communities (because you can ask for help, go in and see how and why things work when you need to, and things get fixed a lot faster when you’re not limited by a company’s resources/ will).
  • Accessible on most platforms (e.g. Windows, Mac and Linux of varying versions of operating system)
  • Easy to install. Because nothing puts you off a tool quite as much as watching a build fail again and again, and having to learn all about the technologies its built on and their variants before you can do even the simplest thing with it </rant>.
  • Stable. It’s cruel to tell people without tech support to install a tool that regularly crashes on them. Even if you’re a techie, it’s still very annoying…

I do a lot of basic numbers work in Microsoft Excel (add things up, do sorts and filters of data, click on numerical columns to averages etc at the bottom of the screen) and the calculator.  Some data scientists (although a dwindling number of them) only use Excel to process data, but in development data science we’re often handling very unstructured or messy data (e.g. ‘who knows how many columns this thing really has?’ type data) and need to produce results that we can both repeat, and trace back step-by-step to the raw data (and sometimes we get lucky and have a dataset bigger than Excel’s limits of 1,048,576 rows by 16,384 columns: NB Excel will silently fail and only give you the first rows/columns if you do this).

That means using a coding language (yes, yes, SPSS, SAS, Matlab, Pentaho etc, but a) those aren’t free, and b) did I mention the really messy inputs?).

R and Python are coding languages that turn up a lot in data science (sql is used a lot too, for database data).   R is a beautiful language for doing statistical things: it was built by statisticians, has packages that aren’t in other languages yet, and has some lovely lovely visualisation libraries.  But we needed to pick one language for the course to minimise the amount of code needed to illustrate each step in data science without causing cognitive dissonance in students’ heads (the previous version of the course taught both Python and R, but was much more code-focussed), and chose Python.

Why Python? Basically three things:

  • its ability to deal with really nastily messy data,
  • its great libraries for things outside statistical analysis (e.g. natural language processing, web scraping, content management systems etc)
  • not having to rewrite code when you start working with developers building applications and websites (although it’s possible to call both Python and R from each other using e.g. the rpy2 package, adding extra languages makes the code that much harder to maintain).
  • it’s much easier to learn (and teach) than people think

There have been many debates on R vs Python for data science. “R vs Python” isn’t really the right question to be asking; the question is “what tools will work for me”, and the answer for many data scientists is “both R and Python, and sometimes neither”: you use what works for you, and you use what’s appropriate for the task that you have at hand.

The other tools included are less controversial (!), and more specialised.

  • OpenRefine is a great tool for summarising and cleaning unruly row-column data without coding (and has provenance tracing built in)
  • Tabula is the most popular open-source pdf processing tool (CometDocs works on messier documents, but is online-only)
  • QGIS is a popular open-source GIS (maps etc) visualisation tool; CartoDb (now called Carto) is a beautiful online GIS visualisation tool
  • The GDAL toolkit is great for command-line processing of GIS data
  • Python includes visualisation libraries, but if you really want something special, D3 is a good offline tool (Tableau Public is good online, but as the name implies, your visualisations will be public).
  • Pen and paper are useful for doing mockups and calculations whilst saving your machine’s battery, and is flexible and portable to boot.

That should be enough to get most people started, and has already been field-tested by both myself and various ex-students (guys: I love it when you send me notes about what you’re doing with data, from the field!).

What’s in my own backpack?

Screen Shot 2016-08-24 at 2.20.36 PM

It’s only fair for me to open up my Mac and show you what I’ve got hiding on my own machine.  That’s my first Launchpad screen above. It’s the usual suspects: Excel, R, OpenRefine, Tableau, Calculator, with a few other things:

  • SQL tools: MySQL workbench, Postgres.
  • Readers for some common proprietary tool formats: SPSS Smartreader, Pentaho Kettle (“Data Integration”).
  • More data tools: Weka, Gephi, Dato, Trifecta. Weka is the granddaddy of data mining tools, and is still worth having in your ebackpack. Gephi is great for visualising and playing around with graph data.
  • Disk cleaners. OMG can I frag a disk on deployment, which is why I have Ccleaner and Disk Inventory X installed. Because sometimes you just need that extra 2Gb of space.
  • Tools for keeping my code contained: Github desktop, VirtualBox.
  • A Java development environment (IntelliJ) because, contrary to popular opinion, I don’t just write code in Python.

The things you can’t see here include:

  • Mapping tools: OpenStreetMap, LeafletJS, MapBox
  • Visualisation tools: Highcharts
  • Machine learning tools: specialist tools, as and when I need them
  • Ideation tools: freemind
  • Text tools: Acrobat reader, sublime text
  • Losing the minimum amount of stuff if my laptop dies in the humidity tools: Dropbox, Evernote, Google Drive, biggest portable drive I can find (2 of)
  • Giving me textbooks to read on the road tools: iBooks, Safari Books

That’s my ebackpack. I’d be interested to see what other people pack, and if there’s anything useful that I’ve missed from the lists above.

Data Science Ethics [DS4B Session 1e]

This is what I usually refer to as the “Fear of God” section of the course…


Most university research projects involving people (aka “human subjects”) have to write and adhere to an ethics statement, and adhere to an overarching ethics framework, e.g. “The University has an ethical commitment to minimize the risks to research subjects and to ensure that individuals who participate in research projects conducted under its auspices… do so voluntarily and with an informed understanding of what their involvement will mean”.  Development data scientists are not generally subject to ethics reviews, but that doesn’t mean we shouldn’t also ask ourselves the hard questions about what we’re doing with our work, and the people that it might affect.

At a minimum, if you make data public, you have a responsibility, to the best of your knowledge, skills, and advice, to do no harm to the people connected to that data.  Data science projects can be very powerful (especially if we’ve designed them well), development data science can affect many people, and we need to be mindful of who we’re affecting and the risks we might unintentionally cause them with our work.  With that power comes responsibility, and a sometimes-difficult balance between making data available to people who can do good with it, and protecting that data’s subjects, sources, and managers.

I start by asking these two questions:

  • Could this work increase risk to anyone?
  • How will I respect privacy and security?


Risk is defined as “The probability of something happening multiplied by the resulting cost or benefit if it does”.  There are three parts to this: cost/benefit (what might happen), probability (how likely that is to happen) and subject (who the risk is to). For example:

  • Risk of: physical, legal, reputational, privacy harm
  • Likelihood (e.g. low, medium, high)
  • Risk to: data subjects, collectors, processors, releasers, users

So, basically, make a list of what bad things might happen, who to, how likely these bad things are, and what you should be doing to prevent or mitigate them, up to and including stopping the project or making sure that anyone at risk is aware of that risk and can consent to be subject to it.  It doesn’t have to be a down-to-the-tiniest-thing detailed list, but you do have to think about who could be harmed by your work and how.

And this isn’t a one-time thing.  New risks can occur as new data becomes available, or the original environment around your project changes: you need to be thinking about potential risks right from the planning phase of your project through to when you’re sharing data or insights (and beyond, if stale data or old results in changed contexts could also cause harm).

Risks can be obvious (e.g. try not to get your sources or subjects targeted by making them easy to find), but they can also be subtle.  Some of the more subtle ones include:

You don’t have to make this list on your own. Groups including the Responsible Data Forum have been doing great work on data risk (and I’ve been lucky to be part of that hivemind too): try reading those groups’ articles to kickstart your thinking about your project’s potential risks.

Privacy and Security

That last risk on the list (accidentally making data subjects easy to identify), is important.  Privacy and security is a complex topic, but at a minimum you should be thinking about PII (Personally identifiable information). PII is this: “any data that could potentially identify a specific individual. Any information that can be used to distinguish one person from another and can be used for de-anonymizing anonymous data can be considered PII.

When you think about PII risk, think beyond name, address, phone number, social security number. Think about things like these (which can all accidentally release PII):

  • Unique identifiers: Names, addresses, phone numbers etc
  • Locations: lat/long, GIS traces, locality (e.g. home + work as an identifier)
  • Members of small populations (e.g. there may be only a small number people fitting this profile in the given geographical area)
  • Untranslated text (e.g. text that your team can’t read and understand)
  • Codes (e.g. “41”, especially if you don’t have a codebook telling you what they mean)
  • Slang terms
  • Can be combined with other datasets to produce PII (aka reidentification)

There is much literature on how easy reidentification can be easy from very small amounts of data, so you will have to think of it also in terms of risk: how likely is it that someone will want to reindentify, how easy is it for them to do this, and what can you do to make it harder.

But Don’t Panic

Don’t panic. When you first start thinking about risk in data science, the usual reaction is an “OMG I can’t do anything without causing harm” one.  That’s not the right answer either.  Be aware of the risks.  Take a deep breath, and remember that risk is cost times probability, and that if you give them the chance, people will often make a surprising but informed decision about their own risks.

Sometimes, the only answer is to walk away from the project. If it isn’t, and risk is high, consider mitigations like releasing data and results to a smaller group of people (e.g. academics, direct responders, people in your organisation, data subjects) with the caveat that once you release, you no longer have control of that data and have implicitly trusted other people to do the right thing too.  Consider releasing data at a different granularity, e.g. use town/district instead of street, and/or a subset or sample of the data ‘rows’ and ‘columns’. Look in places like the RDF website to see what other people have done as mitigations.

Be aware of ethics. Be as honest and cynical about data and results as you can. And don’t start a development data science project without doing a basic risks check.

[Image by Steve Calcott, licensed under cc-by-nc 2.0]

Writing a problem statement [DS4B Session 1d]

Data work can sometimes seem meaningless.  You go through all the training on cool machine learning techniques, find some cool datasets to play with, run a couple of algorithms on them, and then.  Nothing. That sinking feeling of “well, that was useless”.

I’m not discouraging play. Play is how we learn how to do things, where we can find ideas and connect more deeply to our data.  It can be a really useful part of the “explore the data” part of data science, and there are many useful playful design activities that can help with “ask an interesting question”.   But data preparation and analysis takes time and is full of rabbitholes:  interesting but time-consuming things that aren’t linked to a positive action or change in the world.

One thing that helps a lot is to have a rough plan: something that can guide you as you work through a data science project.  Making a plan for data science has much in common with making plans for Lean and design thinking: you’re putting in effort, so you need to be:

  • focused on the change you want to make in the world, e.g. solve a problem or change people’s minds (there’s no point doing analysis if you don’t do anything with it),
  • pragmatic about the work you need to do (sometimes the answer is a piece of paper, e.g. much simpler than the beautiful concept you had in your mind, but much more likely to be used in the current context) and
  • realistic about the problem you’re trying to solve and the resources you have around to do that (e.g. use what works for you, not an unachievable ‘best’ technology).

I like Max Shron’s CoNVO for planning small data science projects. In my dayjob, I work a lot with Lean Enterprise and Design Thinking techniques to achieve similar results at scale, but at the least you should have an A4 piece of paper somewhere with this on it, to refer back to:

  • Context: who needs this work, and what are they doing it for?
  • Needs: what are you trying to fix?
  • Vision: what do you expect your final result to look like?
  • Outcome: how do you get your results to the people who need them? What happens next?

The other question you’re going to want to ask is “is it worth me doing this work”.  It’s okay to say “yes, because I learn from it”, or “it’s fun”, but data science is work, and you don’t want to feel like you put in a ton of effort and late nights only to realise that effort was wasted.  I like the DrivenData competition guidelines for helping with this thinking:

  • Impact: “… clear win for the organisation in terms of effective planning, resources saved or people served… good story around how they generate social impact…”
  • Challenge: “… challenging enough for a rich competition…”
  • Feasibility: “….the right kind of data to answer the question at hand… does it have enough signal to be useful?…”
  • Privacy: “… can answer this question while protecting the privacy of individuals in the dataset and the operational privacy of an organisation…”

I haven’t yet mentioned the thing that your work will focus on: the “interesting question(s)” that you’re trying to answer.  There are several contexts you might find yourself in here, from a business team bringing you a well-defined question (at which point you start with the “what is the question you’re really trying to answer here” discussion), to having complete freedom over the questions you’re asking and the ways they could be turned into action.

One way to get better at asking good questions is to see what other people ask.  Look at your subject area: find other projects and questions in it, and see how they’re asked (and answered).  Look at existing data science projects for inspiration, e.g. Kaggle (and their UseCase list), DrivenDataDataKind and the projects listed in the course reading list, then design your questions.  Asking questions about your questions can help here, e.g.:

  • Is the question concrete enough? Is it solving a real problem, or just a symptom of that problem? (e.g. “what are the barriers to people engaging with us” vs “how can we get more people to call”)
  • Can you translate the question into an experiment? g. can you ask something like “I believe people have more phones than toilets” and start proving (or, more generally, disproving) that.
  • Is it actionable? And what actions will be taken given the answer?
  • What data is needed to do the analysis? At this point, datasets could be anything – tables, images, maps, sensor feeds; anything. Be aware that although data access can limit what you can do, data is just a support here, and focusing on the question can help you think about other ways that you might be able to answer it.

As you do this, you’ll find yourself questioning the meaning of many of the components of your original question: I have a longer blogpost on that, but this is where the plan becomes useful: instead of focusing on “what actually counts as a toilet” (yes, that really is a difficult thing to define), go back to your notes about who this is important to, why, and what they could do about it.  You’ll also find that a seemingly-simple initial question will generate a whole bunch of other questions you’ll need to answer too (questions are like bunnies: they breed).  Again, use your plan as a guide, and accept that there will usually be several parts to each project.  You’ll also find that several of these questions could be answered without using available data (e.g. you might be able to get a strong enough ‘signal’ from surveys that an action is worth further investigation): that too is a useful thing to know.

Plans rarely survive contact with your datasets, users etc.  They’re not about forcing you to produce things a specific way: they’re there to make you think about what you’re doing, and stop you from making newbie mistakes like fitting the questions to the data you have available or falling into data rabbitholes.  You might want to go down at least one of those rabbitholes: a fascinating piece of data that you want to explore for fun, or a bunch of other questions that look like really fun things to answer; they’re not necessarily bad things, and can be really valuable in themselves, but you do need to be aware that you’ve done this, and of any impacts it might have on your original goals.  Planning might seem a distraction from getting on with the data analysis, but it does help to have a guidestar, something to go back to and think “is this valuable to the people that I’m trying to help here?”.

Some short exercises

We ran 3 small exercises in class, to get people thinking about project design. Each of these was time-limited to 3 minutes, to make people concentrate hard on what might be needed, and to hit issues quickly so they could be discussed in class before students tried this at home.

Exercise 1: Ask some interesting questions. Either your own questions, or pick an existing question and think about how it might have been formed.

  • Questions that data might help with
  • Stories you want to tell with data
  • Datasets you’d like to explore (where ‘datasets’ could be anything – tables, images, maps, sensor feeds, etc)
  • Competition questions: Kaggle, DrivenData
  • A data science project that interested you

Exercise 2: Get the data.  Pick one of your questions:

  • List the ideal data you need to answer it
  • List the data that’s (probably) available

Think about what you’ll do if the data you need isn’t available:

  • What compromises could you make
  • Where would you look for more data
  • Are there proxies (other datasets that tell you something about your question)
  • Are there ways to get more data (surveys, crowdsourcing etc)

Exercise 3: Design your communications. List the types of people you’d want to show your results to.

  • How do you want them to change the world? Can they take actions, can they change opinions etc
  • Describe the types of outputs that might be persuasive to them – visuals, text, numbers, stories, art… be as wild with this as you want

Data Science is a Process [DS4B Session 1c]

People often ask me how they can become a data scientist. To which my answers are usually ‘why’, ‘what do you want to do with it’ and ‘let’s talk about what it really is’.  So let’s talk about what it really is.  There are many definitions of data science, e.g.:

  • “A data scientist… excels at analyzing data, particularly large amounts of data, to help a business gain a competitive edge.”
  • “The analysis of data using the scientific method
  • “A data scientist is an individual, organization or application that performs statistical analysis, data mining and retrieval processes on a large amount of data to identify trends, figures and other relevant information.”

We can spend hours debating which definition is ‘right’, or we could spend those hours looking at what data scientists do in practice, getting some tools and techniques under our belts and finding a definition that works for each one of us personally.

My own working definition of data science is “a process that helps people gain understanding through using data”.  So let’s look at some of that process.  The scientific method, mentioned above, is a process (O’Neill & Schutt, “Doing data science”):

  • Ask a question
  • Do background research
  • Construct a hypothesis
  • Test your hypothesis by doing an experiment
  • Analyse your data and draw a conclusion
  • Communicate your results

Most of science works like this: it’s all about creating explanations that fit our knowledge of the world, then testing those explanations with experiments.  I like the cynicism embedded in this, the acknowledgement that everything is a working hypothesis that might turn out to be false, or false in different circumstances (see under Newton/Einstein), and those are all good things, but they don’t quite cover what data scientists do all day.

One of the process descriptions that data scientists use for themselves is the OSEMN process (Obtain-Scrub-Explore-Model-Interpret, pronounced ‘awesome’: your data is safe with data scientists, but not your acronyms…):

  • Obtain datasets
  • Clean, combine, transform data
  • Explore the data
  • Try models (classification, machine learning etc)
  • Interpret and communicate your results

This is less about experiments, and more about the things that you need to do with data, but it loses what, to me, is the most important part: asking an interesting question.  Data science isn’t about data – it’s about people, their problems and questions, and informing, persuading or entertaining them with your results (I’m with Sarah Cohen when she says “every good story starts with an idea, a question or an observation”, and with anyone who says that a visualization isn’t always the answer). So the process for these sessions, the thing we’ll be working through slowly, is:

  • Ask an interesting question
  • Get the data
  • Explore the data
  • Model the data
  • Communicate and visualize your results

with a healthy dose of cynicism, e.g. sanity-checking your results, in the context they’re relevant to, which is especially important in a development data context, where data is hard to come by and may be erroneous, miss geographical areas or demographics and potentially be older than it looks.

Note that none of the quotes say “enormous amounts of data”.  We’ll touch on big data in a later session (session 10), but most development data scientists work with small datasets, and that’s nothing to be ashamed of: I’d rather have relevant, information-rich datasets than huge amounts of data that tells me almost nothing.

That process again

  • Ask an interesting question. Write hypotheses that can be explored (Do people have more phones than toilets?, How is Ebola spreading? Is using wood fires sustainable in rural Tanzania? Can we feed 9 billion people?). Make them simple, actionable, and incremental (e.g. you can test different parts of the question separately).
  • Get the data. There are many different data sources (e.g. datafiles, databases, APIs, text, maps, images, social media, people). Some of them are harder to get information out of than others, but they all contain data. Which means you’ll often be extracting datasets from those sources (e.g. 80-page PDFs), and cleaning it.  By cleaning, I mean getting the data into a shape that can be used by algorithms: dealing with file formats (pdfs!), badly-specified locations, human errors and differences between standards (e.g. “Tanzania” vs “Republic of Tanzania”).  Although cleaning takes a lot of time, it’s also time spent getting to know your datasets: what’s in them, what’s missing, what’s strange, what potentially got lost in translation. Which leads us to:
  • Explore the data. Once you’ve got the dataset in machine-readable form, you can start looking for more issues to deal with (e.g. these issues with different placename standards in Tanzania) and, eventually, for potentially interesting patterns. Eyeballing (looking at) your data is usually a good place to start; often you have to take a subset of the whole dataset to do this, but it’s usually worth it to get a better feel for what’s contained in each dataset.  Doing quick but ugly visualisations of your data is also a good thing to do.
  • Model the data. Modelling is where we look for patterns and insights hidden in the data.  It’s where machine learning comes in. We’ll look later at how to learn relationships between numbers, categories and graphs.
  • Communicate and visualize your results. We want to get this effect: ”I already knew that increased incarceration didn’t lower crime, but I wasn’t sure of the statistics. To see it on the graphs is really eye opening.” (Pandey et al, The Persuasive Power of Data Visualisation), using whatever’s appropriate (which might or might not be visualisations).

What we’re aiming at is simple: “ask good questions, tell good stories” – if you can do this with data, you’ve won. 

Data Scientists

Data scientists are the mysterious beasts people who do data science.  The data science Venn diagram basically says that to be a data scientist, you need to know statistics, your business area and be able to code.  But it’s not quite like that.  Although it’s heresy to say this, many good data scientists don’t code at all, and you can useful on a data science team without knowing everything, e.g. build insightful visualisations without using statistics, or specify a data science problem well enough for hardcore machine learning specialists to develop good algorithms for it.

That said, it does help a development data scientist to have expertise in development and statistics (these sessions were originally designed for people who had these skills: a statistics session is already in the pipeline…), and being familiar with data science skills and having the coding skills to get, clean and explore data will help you even if you never want to do anything more than create and be the ‘client’ for a data science problem specification.

At this point, you might be asking two questions:

  • How do you become a data scientist?, and
  • Should you become a data scientist?

You become a data scientist through learning and practice (that never stops: I’m still working on it myself).  Yes, you need to learn a bunch of theory, but there’s nothing like learning data science by doing it: you’ll handle issues you didn’t know existed, and learn many details about techniques by using them on non-sanitised (uncleaned) data.  Good places to practice exploring and modelling data include:

  • Kaggle – online datascience competitions
  • Driven Data – social good datascience competitions
  • Innocentive – some datascience challenges
  • CrowdAnalytix – business datascience competitions
  • TunedIt – scientific/industrial datascience challenges

Good places to practice asking good questions, getting data, communication and visualizing results include:

  • Your own projects
  • Data science for good groups (e.g. DataKind)

The answer to “should you become a data scientist” is “not necessarily”.  There are lots of data science students desperate for good problems to work on, so you might want to become someone who can work with data scientists; which means learning how to specify data problems well.   One of the places to see the work of these people who can specify problems (“problem owners”) is the competition sites listed above.  The problem owner doesn’t have to do data modelling or machine learning themself, but they do need to be able to specify a problem well, find and clean data related to that problem so that competitors can access it easily (and all have the same starting dataset), and specify how the competition results will be marked (e.g. by accuracy on an unseen ‘test’ dataset).   Go look at some of the problems listed on these sites, and think about how you would have done this yourself.

(session 1 slideset is here; cover image is from the Pump it Up challenge)

About these sessions [DS4B Session 1b]

First, there are no prerequisites for these sessions except a curiosity to learn about data science and willingness to play with data. Second, the labs are designed to cover most of what you need to talk to data scientists, to specify your project, and to start exploring data science on your own.  They’re specifically aimed at the types of issues that you’ll see in the messy strange world of development, humanitarian and social data, and use tools that are free, well-supported and available offline.

This may not be what you need, and that’s fine. If you want to learn about machine learning in depth, or create beautiful Python code, there are many courses for that (and the Data Science Communities and Courses page has a list of places you might want to start at).  But if you’re looking at social data, and to quote one of my students “can see the data and know what I want to do with it, but I don’t know how to do that”, these sessions might help.

There are a bunch of resources that go with these sessions: they’re all available on the sessions’ github page, and slideshare pages.

There are 10 sessions in all, grouped into themes: people (designing a project, communicating results), tools, getting data, special data types (text, GIS, big) and learning from data (there are also bonus sessions in the works, as the chats with Quito continue); the session order is designed to get people doing useful things as quickly as possible, and give breathing space between difficult topics.

Each session concentrates on 5 to 7 concepts within a single topic (e.g. machine learning), and plays with apps or code related to that topic.  Install instructions for each tool used in a session are in the tool install instructions; pre-session required reading, and post-session reading for fun are in the course reading list, and python code used in or related to the session are in the ipython notebooks.

There’ll be several blogposts here for each of the sessions, loosely arranged around those 5-7 concepts and homework.

DS4B: The Course [DS4B Session 1a]

[Cross-posted from LinkedIn]

I’ve designed and taught 4 university courses in the past 3 years: ICT for development and social change (DSC, with Eric from BlueRidge), coding for DSC, data science coding for DSC and data science for DSC (with Stefan from Sumall).  The overarching theme of each of them was better access and understanding of technical tools for people who work in areas where internet is unavailable or unreliable, and where proprietary tools are prohibitively expensive.

The tagline for the non-coding courses was the “anti-bullsh*t courses”: a place where students could learn what words like ‘Drupal’ and ‘classifier’ mean before they’re trying to assess a technical proposal with those words in, but also a place to learn how to interact with data scientists and development teams, to learn their processes, language and needs, and ask the ‘dumb questions’ (even though there are no dumb questions) in a safe space.

Throw a rock on the internet and you’ll hit a data science course. There are many of these, many covered by the open data science masters, and many places to learn the details of machine learning, R, business data science et al.  What there isn’t so much of on the internet is courses on how to specify a data science problem, how to interact with the people coming off those machine learning courses, or how to do data science work when you’re days from the nearest stable internet connection.  The latest course, Data Science for Development and Social Change, was designed for that.  It changed a bit from its inception: we found ourselves heavily oversubscribed with students from not just International Development, but also Journalism, Medicine and other departments, and adjusted from a 12-person hands-on intensive lab format (e.g. the coding course included D3 down to the CSS, HTML, SVG and Javascript level) to a 30+ person lab lecture with ipython notebooks.

The slideset shared here is from the latest inception of the lab part of that course (the non-lab half showed social data science in different contexts): a set of sessions that started as a weekly one-hour chat between myself and the Quito ThoughtWorks team, and grew into a multi-city weekly data science session.

The way to approach these slides is to read the slide notes in each session’s powerpoint file (you’ll probably have to download the files to see this: I’ll slowly upload the notes if needed), then try the exercises in the ipython notebooks for that session.

The first session is an introduction, a ‘why are we doing this’, with a bunch of downloads and setups (student- and developer- tested!) as an exercise.  I’ll write about that soon.  The downloads and setups are integral to these sessions (to leave someone with enough tools, notes and information on their machines to be able to do basic data science, and understand what a data scientist does all day, even where there’s no hope of stable Internet (yet)), but though the learning is much better if you try the examples, the slides can be mostly followed without them.

Releasing my course materials [DS4B]

[cross-post from LinkedIn]

I teach data science literacy to different groups, and I’ve been struggling with a personal dilemma: I believe strongly in open access (data, code, materials), but each hour of lecture materials takes about 15-20 hours of my time to prepare: should I make everything openly available, or try to respect that preparation time?

Which is basically the dilemma that anyone open-sourcing goes through. And the answer is the same: release the materials.  Made stronger by the realisation that I teach how to do field data science (data science in places without internet connectivity) and people in places without internet connectivity are unlikely to be dropping into New York for in-person sessions.

Starting this week, the materials are going online in github; I’ll be backing this up by restarting the I Can Haz Datascience posts (posts on development data science, with Emily the cat) to cover the topics in them (designing and scoping a DS project, python basics, acquiring data, handling text, geospatial, relationship data etc).  Hopefully they’ll be of some use to people.