data.bythebay.io: Zachary Taschdjian, Data Design Challenges for Enterprise IoT Applications
so this is data design challenges for Enterprise iot applications I um I decided to drop the semantic ontology stuff in the interest of time if um anybody's here specifically about semantic ontologies then uh you can throw tomatoes or we can talk in the hall afterward um so my name is Zach I work for GE digital I'm a data viz guy a interaction designer and a product person um and I'm going to talk about basically what we're working on um in abstracted terms so I don't violate any agreements um but we're working on Internet of Things applications and we B we build stuff like um jet engines trains uh the technical term would be locomotives and not chooo as sometimes we call them um and other things of that nature we call it the internet of big things um so what I'm going to talk about today is is is how designers approach data and the need to approach data intelligently um when we are designing data visualization applications um so so this this applic or this presentation is geared toward designers um and like a lot of designers I came from an arts background and the thought of programming and math strikes fear into my heart uh to this day um so I think really a better term for this is probably go home data you're drunk because that's really how A lot of designers feel about data it's something that's sort of scary and foreign um and go figure you know a lot of the data that we work with is it's quantitative it's mathematically driven um it's uh programmatically created and it's basically as far from design as you could imagine so um when you get past that though I think that designing with data in mind and what I'm going to start calling data first design um not a term that I've coined by the way uh it'll put design in a whole new light uh and my hope is that this presentation will um get rid of some of the fear but uh judging by the looks of you guys I don't think any of you are especially afraid of data you don't look like a design audience raise your hand if you're a designer anybody a designer in here okay all right thank you I'll take it so we're going to go from go home data you're drunk to data first design and to be clear I don't mean that um we should all become data scientists or statisticians which probably most of you already are um data first design to me is not about data above everything else it's about prioritizing data exploration the same way that we would prioritize say user research or usability testing um after all we do them for the same reason and that's to ultimately make a better product so what this presentation is not it's not a Kai paper Kai was last week um I'm going to reference some research but all the stuff that I'm going to talk about in here is really um you know anecdotal and it's my own my own experience on it so um also these are challenges that are the subject of a lot of ongoing visualization research um so there's really no Silver Bullet I'm merely suggesting that doing data exploration as a designer is something that's valuable um but it will solve all of our problems in making um you know complex data driven products so what am I going to talk about I'm going to talk about a quick working definition of what a designer is and what data is um then I'm going to show a little bit of research about how designers have traditionally approached data um and then I'm going to lay out some examples from my own experience and hopefully those will suggest some solutions along the way so uh who are you as we've established only one of you is a designer um so this is a presentation about designers so I don't have any hard and fast rule about um what you are but you're probably skewed more toward a designer at an agency or excuse me at an Enterprise rather than an agency um and you're probably working on applications for professional users you're not working on consumer facing applications and probably there's not a lot of Mobile in your life either at least there's not in mine um but our Challenge and I'm speaking of myself here is that um we're working with data that's kind of out of our league we're not doing infographics here so um with that in mind you know what is an infographic in in this uh respect I think that um this has been covered pretty well by a lot of other speakers in a lot of other areas um but I think that there's a big difference between visualization and infographics and infographics are going to be very much data driven um they're going to be uh highly abstracted the data is not going to be um represented in nearly the complexity and level of detail as it would be in a visualization there's really no explicit data exploration step there's little data wrangling and again these aren't hard and fast these are just kind of what I've seen in my experience um finally it's made with tools like uh illustrator and other design tools as opposed to being uh programmatically encoded um and the flip side of that is that it's actually going to be manually encoded so some designer is going to go in there and actually make dots uh in the application you to show where where data will be so um in contrast the uh a data visualization is going to be far more data driven it's not going to be abstracted you're going to see um much higher level of granularity with the data um it's going to be made with tools like at the very least Excel but probably something like mat lab or r or you know other packages like that um the tool is going to do the data encoding so nobody's going to hand encode data um it could be highly dimensional data it could be um lots and lots of data uh and there could be static or streaming data um and there's usually an explicit data wrangling step that's involved as well um and then there's going to be uh you know it could be it could be relational data it could be um graph database it could be unstructured or structured Etc um essentially it could be almost anything I'm just talking here more about the data that I've encountered in my day-to-day work so um put simply the differences on where the emphasis lies so does it focus on Aesthetics or does it focus on analysis and of course not mutually exclusive I think good design should do both but in general the motivations for an infographic are you know there's going to be some narrative aspect to it it's going to tell a story um it's going to be interested in Persuasion entertainment uh potentially marketing another form of persuasion whereas um data visualization is more geared toward analysis um identification experimentation uh or externalizing a concept that's um highly abstract and complex uh for example a um a cost function in logistic progression which is what this visualization is um so just to be clear uh v data visualization is not a single monolithic subject there's lots of different subdisciplines uh and I'm not suggesting that the compelling aspects of infographics should be left out of data visualization they all need to tell stories they all need to persuade uh it's just the call to action that comes out of that is probably um more around uh you know actionable data rather than a purchase decision for example so um I bring this up to highlight the fact that most designers uh the skill set that most designers have is not really sufficient for doing what's on the right so that's the designers uh a little bit and that's the data a little bit or excuse me that's the um let's talk about the data a little bit so uh I've mentioned it's it could be structured or unstructured it could be time series which is where I've spent a lot of my time um it could also be um graph which is another place I've spent a lot of time um if it's time series that might not be of constant data density or scale um and may or may not have any data metadata with it may or may not have unique names units of measure might be different um there could be really complex architextures that are kind of a dependency uh to be able to surface it in a UI at all or run any kind of analytic on it um again it could be highly dimensional or it could be big tall or fast so um accessibility and honestly this is the biggest challenge that I've faced in in working with any kind of data um being able to surface it at all is uh there's a big dependency on getting it all into one unified integrated place so there's you know huge demand for systems integrators in this space um but you know you can have workarounds for design but if you don't have uh accessible data there's no real workaround for that um so to that point our biggest challenge is just getting it all in one place having it accessible having it integrated having it cleaned um it's very uh it's a non-trivial tasks and each one of these um each one of these steps presents a point of failure a potential point of failure in any visualization application or analytic for that matter so um data accessibility huge issue so I want to Pivot a little bit uh to the designers here um and talk a little bit about how design ERS have approached data traditionally uh this is a really great paper about how designers work uh from University of Utah that came out maybe four or five years ago now um they determin that a lot of designers approach data uh based on their skill set and the tools that they're comfortable with so um the tools that you're comfortable with if you're comfortable with the tools on the left there things like illustrator hype uh various design tools you're probably going to be less familiar with the ones on the right if you're a pro with both of these then you're a purple spotted unic unicorn as someone once told me I don't know anybody that can do mat lab as well as they can do You Know sketch for example and knows both of those well um so uh big big um big divide in terms of skill sets there um another thing that they came up with was that designers inferences about data are usually not usually they're occasionally divorced from the actual data in other words designers will look at the problem and then start designing based on their Assumption of what the problem is rather than doing any kind of um formative data exploration step which is what I'm calling data first design so finally they also determined that um a lot of how comfortable you are is correlated with uh excuse me the amount of data exploration you do is correlated with how comfortable you are with data in the first place so going back to what I opened with a lot of designers are scared shitless of data so there's not a lot of data exploration because they just don't really know how to approach it and that's not universally true but it's occas true unfortunately this all starts falling apart when you start looking at Big complicated and Abstract data so I I would contend that you reach this point pretty quickly and as the amount of data that we work with grows uh that point is going to um become sooner and sooner in the design process so uh you know when you start working with this stuff you have to really if if it's for placement only in design terms where you're roughing something and you're mocking something up and you can't even do that um without you know having some understanding of the data anymore so um there's a point at which you can't actually do this without having some visualization package or at least being able to understand the scale of the data you're talking about so I'm going to switch to tactics here I'm going to make things a little bit more concrete um with some examples that have come out of my work so let's say that we're talking about a tool for an analyst the use case is monitoring assets uh to analyze and ultimately prevent failures and this is this is a real use case that we Face pretty much overarching use case at GE digital um so she's monitoring a fleet of assets one of ge's uh verticals is jet engines so let's say she's working on jet engines and she needs to understand not only why an asset has failed but also and more importantly to prevent it from failing in the first place ultimately so how does she do this well the engine has sensors on it um like a lot of uh instrumented iot objects um and Those sensors collect data from the engine so in this case um for example your body temperature 98.6 Fahrenheit um there's a thing called exhaust gas temperature which is a pretty good coraly so things like that RPMs altitude Etc so the analyst needs to compare these variables and identify the critical X's this is kind of a common pathway for analysts to do this they're going to lay out a whole bunch of different variables um and then go through them and use human preattentive cognition to say oh hey I instantly recognize that in you know quadrant 1B or whatever there's a um there's a problem there's an issue there so then they're going to start drilling into that uh to dig down and start plotting out time series and independent um uh singular variables rather than looking at these matrixes so um she's going to start with a chart like this and she's going to compare those X's she's going to scan it um and let's say that we're just going to deal with a single asset here because uh in the real world they're going to be comparing many many assets um you know fleets of jet engines 40 50,000 Each of which have you know 4 to 800 sensors on them and then they're going to generate time series based on that um so let's say that we're coming back down to just one asset uh and the takeaway is that without first examining and understanding the data it would be really easy to think that just comparing the time series for these is pretty trivial um so let's say that these are the variables that she wants to compare here so um analysts might know that this event occurred for C under a certain set of operating conditions um and let's say that the asset the jet engine that she's looking at is a 20-year-old engine so it's it's kind of old uh it's been pushed pretty hard so just like um you know if you drive your car really really hard and you Redline those RPMs a lot uh it's going to impact the way it performs in your life on it so it's been a abused um let's say it's also been flown in place like Dubai where there's sandstorms there's hot and harsh conditions it's 100° Fahrenheit or more throughout a large part of the year um so she's going to be able to take that kind of environmental um data about the asset and then that's one thing she's going to plot but she's also going to look at things like air speed exhaust gas temperature uh altitude um and all these things and she's going to line these things up in time um to try and examine where um basically she's tring to narrow down what caused this so this is FMEA it's failure mode analysis basically why an engine failed um so she's also going to know what a healthy engine of this model and year and under those conditions would look like uh and she's going to compare those two types of data so even if we have one asset that we've narrowed this down to it's entirely likely that she's going to be comparing it to either data for that asset in the like historical data or um abstract the data for that class of asset so um you know your Honda Civic how how your Honda Civic would compare to uh you know the exact same conditions everybody else's Honda Civic under those conditions if that makes sense um so at the end of her for investigation she's going to want to use that data to forecast failure and other similar assets under similar conditions so let's say that she wants to plot these um these five pieces of data so if we're going to take a design first approach we might start with something like this so we've got um you know X along the x axis would be time the Y AIS would then be the the the scales for the different variables and we need to have five we've got five different variables there one's got um wind speed might be in in miles per hour kilometers per hour um stop doing that uh temperature is going to be Fahrenheit or Centigrade oil pressure might be pounds or kilogram per inch um altitude feet or meters air speed again uh could be knots and then RPMs is rpm so we've got a bunch of different kinds of data there um hence the need for a bunch of different y- axes so we might experiment with our design and do something like this where we start stacking them up because pretty quickly you get um if you're looking at something like this it gets ugly real fast ditto with this this isn't going to scale very well you got four there we didn't manage to get five uh in this View and then if you want to get fancy you might do something like this where you're OB fisca data based on like interaction like you're hovering and showing a y- AIS based on what you hover and everything else is grayed out uh you might be able to do something like this where you're you know packing lots of stuff into the view um but eventually you're going to re reach a point at which the data is too dense and overplotted to really be usable so the problem is clear that none of these uh design first Solutions really accommodate the actual use case um remember that we've simplified this down to looking at a single asset we're not looking at fleets of assets so this is going to there's going to be a huge scalability problem here and I know the last last presenters mentioned that as well and data scaling should come as no surprise at a conference like this it's going to be a big problem so nail it before you scale it um and this also opens up a whole debate on the lean process and the MVPs which I'm going to just s side step but um we need to nail down our use case before we start designing and looking at any kind of interaction features or things like that um and then one other caveat you know these like I said these are ongoing problems so there's no there's no Silver Bullet from um you know data visualization Community about how to do this but let's um contrast with a a data first approach so if we looked at the data first we might look at things like scale um we'd look at the uh the units of measure and we'd know that um by looking at those units of measure we'd see that um you know they're going to be very different we've got like I mentioned we've got RPMs we've got temperature we've got speed and different units of measure things like that um take temperature for example it could be Centigrade Fahrenheit ranking Kelvin um all of these are going to have different scales um and also Al this could be coming from multiple sensors so you might have 10 or 15 or 100 different thermometers in the engine um and the data that you're actually plotting might look like uh might be averaged and normalized down to a point where um you know you're losing um anomalies and outliers from the data so uh things to know about beforehand secondly you'd see that um data is going to behave a little differently in the wild so um you need to run a use case like your user is going to do it so um the other thing is is for example how many times uh what's your sampling rate of the data for example um if you're looking at something like um uh data that's sampled over the course of an 8 hour flight maybe there's things that are sampled five times like once you know uh like throttle changes for example if the plane climbs or declines in altitude that might might trigger a data point so there might be five of those and but something like um exhaust gas temperature might be sampled five times a second so you've got 180,000 data points that you're comparing to five so you're going to have to do some kind of time warping to compare those things otherwise if you end up having a ukian you know straight line comparison it's not going to work because you're not going to be able to line those things up so again very non-trivial challenges for uh visualizing this data uh so out of this come a few assertions um understand the use case uh minimally viable is fine but again nail it before before you scale it um understand the data uh look very closely at it streaming uh static time series graph Etc uh units of measure scale um ontologies that are underlying it what the relationships are uh understand the architecture um is it cloud or on premises or is it distributed or unified uh will the architecture support the necessary data performance because um things that I've run into are things like search you know can you you know do you have the uh the throughput to support search of of massive amounts of data um finally don't be afraid to uh to do research and know the state-of-the art and bring in experts if you need to so thanks for not running screaming I appreciate your uh interest um and just to sum up um I I think that we need to do more data design first excuse me data first in our design um and it's going to lead to better products so thank you very [Applause] much there's a bit of time for any questions sir um actuallya C however um it's actually harder for me to think um you only focus on data itself when you're interest in data I think you will have something that you want to do actually drive 5 years looking for all the and uh the last side you you point you lay out I think those are other dimensions of of a design Tas you must consider you have to consider uh what task you want to perform as well as the data available to use theistic and the process and so on but still you want to deliver the value act to the people who's going to use it and that's still going to be yeah I agree so just to restate your point it sounds like you're emphasizing the need to solve legitimate user problems and to have a kind of a pro a product focused approach to this I I couldn't agree more I think that's something that gets lost uh very regularly and at least where I work um people tend to get um Tunnel Vision on on their little part of the product uh you know because they're experts um and that's what they focus on so um I think we're at time thank you very [Applause] much