It's time to talk time series

Times series data is everywhere! I mean, seriously, try to think of some data that isn’t a time series. You have stock prices and weather data, which are the classics, but you also have a time series of images on your phone, time series log data coming off of your servers, and much more. In this episode, Anais from InfluxData helps us understand the range of methods and problems related to time series data. She also gives her perspective on when statistical methods might perform better than neural nets or at least be a more reasonable choice. Sponsors: DigitalOcean https://do.co/changelog – DigitalOcean now offers three managed databases https://www.digitalocean.com/products/managed-databases/ — PostgreSQL, MySQL, and Redis. Get started for free with a $50 credit. Learn more at do.co/changelog https://do.co/changelog. KubeCon + CloudNativeCon (San Diego) https://events.linuxfoundation.org/events/kubecon-cloudnativecon-north-america-2019/?utm_source=changelog&utm_medium=newsletter&utm_campaign=kcnapracticalai19 – The Cloud Native Computing Foundation’s flagship Kubernetes community conference which gathers adopters and technologists from leading open source and cloud native communities. Learn more and register https://events.linuxfoundation.org/events/kubecon-cloudnativecon-north-america-2019/?utm_source=changelog&utm_medium=newsletter&utm_campaign=kcnapracticalai19 — get 10% off with the code KCNAPRACTICALAI19 Feel free to use the Convince Your Boss https://events.linuxfoundation.org/events/kubecon-cloudnativecon-north-america-2019/attend/convince-your-boss/ letter in part or in full so you can your team can attend. The Brave Browser https://brave.com/cha913 – Browse the web up to 8x faster than Chrome and Safari, block ads and trackers by default, and reward your favorite creators with the built-in Basic Attention Token https://basicattentiontoken.org. Download Brave https://brave.com/cha913 for free and give tipping a try right here on changelog.com https://changelog.com. Featuring: Anais Dotis-Georgiou – X https://x.com/anaisdotisChris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaShow Notes: InfluxData https://www.influxdata.com/InfluxDB https://github.com/influxdata/influxdbM-Competition benchmarks https://en.wikipedia.org/wiki/Makridakis_CompetitionsUpcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Being with your change log is provided by Fastly. Learn more at fastly.com. We move fast and fix things here at Change Law because of Rollbar. Check them out at rollbar.com. And we're hosted on Leno cloud servers. Head to leno.com slash change log.

• 0:16Speaker 00

  This episode is brought to you by Digital Ocean, the simplest cloud platform out there. And we're excited to share they now offer dedicated virtual droplets. And unlike standard droplets, which use shared virtual CPU threads, their two performance plans, general purpose and CPU optimized. They have dedicated virtual CPU threads. This translates to higher performance and increased consistency during CPU intensive processes. So if you have build boxes, C I C D, video encoding, machine learning, ad serving, game servers, databases, batch processing, data mining, application servers, or active front end web servers that need to be full duty CPU all day, every day. Then check out Digital Ocean's dedicated virtual CPU droplets. Pricing is very competitive, starting at 40 bucks a month.

• 1:01Speaker 00

  Learn more and get started for free with a $100 credit at do.co slash change log. Again, do.co slash change log.

• 1:21Speaker 00

  Welcome to Practical AI, a weekly podcast about making artificial intelligence practical, productive, and accessible to everyone. This is where conversations around AI, machine learning, and data science happen. Join the community and slay with us around various topics of the show at change.com slash community. Follow us on Twitter. We're at Practical AI FM. And now into the show.

• 1:47Daniel Whitenack?

  Welcome to another episode of Practical AI. This is Daniel White neck. I'm a data scientist with S I L International. And I'm joined by my co-host, Chris Benson, who is a principal AI strategist with Lockheed Martin. How are you doing, Chris?

• 2:03Chris Benson?

  Doing great. How's it going, Daniel?

• 2:05Daniel Whitenack?

  It's going good. It's looking a little bit more like fall around here, which is a really nice time of year. So I'll have to get out the leaf blower soon. But otherwise, doing pretty good.

• 2:15Chris Benson?

  Yep. It's finally starting to cool down here in the south. So I'm looking forward to cooler weather.

• 2:20Daniel Whitenack?

  Yeah, nice, nice. Well, speaking of times changing, we have a guest today that's gonna help us dive into some things about time series data and other related things. We have Anais Dotis Georgiou from Inflex Data with us. Welcome, Anais.

• 2:39Speaker 03

  Hi, thank you so much. It's nice to be here.

• 2:42Daniel Whitenack?

  Yeah, we're so happy to have you. We saw your talk listed on the All Things Open website. A couple of people from the change log were there, and I'm really excited to dig into a few of those details and other things. But before we do that, why don't you just give us a little bit of an intro about how you got into data things and eventually ended up at Inflex Data?

• 3:05Speaker 03

  Sure. So originally, my background is in chemical engineering, at least that's what I went to school for. And when I got straight out of school, I thought maybe that I wanted to go into biotech and do research. And I spent some time working with a liquid handling robot. And after a little while of just kind of being in this sterile environment where my only friend was this robot.

• 3:31Daniel Whitenack?

  So liquid handling robot, like a robot that handles hazardous chemicals? Is that the idea?

• 3:37Speaker 03

  Well, it didn't handle hazardous chemicals. It essentially was like a micropipetter and could execute protocol in like a larger scale. So it could replicate a single experiment like, yeah, more efficiently.

• 3:52Daniel Whitenack?

  Yeah, so it was more like a automation thing versus like, first, when you were talking about that, I had like those like OpenAI robot hands in mind, like carrying like test tubes around or something.

• 4:04Speaker 03

  Yeah, that would be really cool. This carried around tiny volumes of liquid, but nothing quite like that.

• 4:11Daniel Whitenack?

  We're recording this around Halloween and so I was just thinking, you know, liquid nitrogen bubbling over the side. It could be a lot of fun.

• 4:17Speaker 03

  I honestly, maybe I would have stayed longer if I had been messing with a robot that was handling liquid nitrogen, so.

• 4:23Daniel Whitenack?

  Well, maybe someday.

• 4:25Speaker 03

  Maybe.

• 4:26Daniel Whitenack?

  We can all aspire.

• 4:27Speaker 03

  Right. But yeah, so I did that and I decided I didn't like being in a sterile room with only a robot to talk to. And I got to work with some data scientists who were actually creating the detection algorithms for the work that I was doing. And basically all the data that I was collecting and they were trying to detect autism, prenatal, do prenatal testing for autism. And I felt like the data scientists were the ones that really got to have all the fun because they didn't have to do the same procedure over and over again, just to collect the data. They got to actually take the results and then derive meaning out of it.

• 5:07Speaker 03

  So I decided maybe I should go down that path and that way maybe I'll even get to talk to humans more, which I don't know, is kind of funny when I think about it now. Because people usually think about tech being a little bit, I mean, like more sterile or less like people facing. But especially in the role that I have now, it's extremely people facing and I really enjoy that. And I was missing that from biotech.

• 5:30Daniel Whitenack?

  Your role now is developer advocate, is that right?

• 5:33Speaker 03

  Yeah. So are you familiar with what a developer advocate does?

• 5:37Daniel Whitenack?

  So I am, but I expect that there's a lot of confusion out there. So maybe it would be good to have a developer advocate define it for us.

• 5:45Speaker 03

  Okay. Well, every developer advocate seems to have a different answer for what developer advocacy or developer relations is.

• 5:53Daniel Whitenack?

  Well, for the record, your definition will stand at Practical AI because you're the first one to define it. So. It's canonical for us.

• 6:01Speaker 03

  Oh, no. Okay. Well, I would just define it as, and I think this is borrowing from a lot of other developer advocates. But basically a way to connect the company to the community and the community back to the company. So our role is to facilitate the use of our product, but also to bring product feedback. So kind of just establishing that bridge. And that looks like, takes the form of giving presentations.

• 6:32Speaker 03

  Going to meetups, having meetups, writing blogs and tutorials, maybe contributing to documentation or the product itself, and hopefully having interactions with product to help guide the product in the direction that the community needs.

• 6:47Daniel Whitenack?

  Was that a sort of natural transition for you? Did you do like data science kind of as a title, you know, data scientist for a while? Or did you kind of immediately want to get into this or more of a developer-focused side of things? Because I definitely think that, you know, both are important. It sounds like the kind of developer-facing side was really important to you.

• 7:10Speaker 03

  Yes, I wasn't ever a data scientist. I took a data science bootcamp, the University of Texas in Austin. And I headed towards data science because I did a lot of math as an undergraduate and I really enjoy math. And data science felt to me like an opportunity where I'd get to use math and get to think about math. So that's what attracted me to it. And also the fact that you get to solve problems by thinking critically and looking at data and trying to uncover solutions and also reevaluating your biases and stuff, so.

• 7:44Daniel Whitenack?

  Developer advocate, I think, is misunderstood by a lot of people, but even more so probably like data scientist and what that entails. And like being a developer advocate for a data-focused company, it seems really needed thing right now when like so much is misunderstood about how people are processing data, like what data science is in general, what we should, what we shouldn't be doing, what developers actually want to do. So yeah, I could see how it could be a very challenging position, but definitely very valuable.

• 8:16Speaker 03

  Yeah, it's definitely challenging. You get, I tend to get two camps of people, people who are either just breaking into the field or people who seem to have like several years' experience and PhDs and spend most of the time educating me, which is, I mean, I get to learn a lot, so I'm always grateful. But yeah, I think it can definitely be tricky because I sort of try to position myself towards helping. People break into data science and also data analytics. And so there's a lot of opinions about how one should handle their data. And so it can be tough trying to, you know, talk to the both extremes of that audience because they have such different needs and such vastly different knowledge base.

• 8:59Speaker 03

  It's really hard to talk about math with people who are just learning about math and also people who have PhDs in math in the same room. So.

• 9:09Chris Benson?

  Fair enough. And I think that of the three of us, I'm probably the weakest in math. And so I'll probably have all sorts of questions as we go through the conversation for you. I actually wanted to start off by asking, through the course of the beginning of this conversation, you know, the phrase time series has come up several times. And I was wondering if you would, for those who are maybe just getting into it or not previously familiar, if you would kind of tell us what time series is and means and give us a little background on that.

• 9:38Speaker 03

  Sure. So time series is just any data that has a timestamp attached to it. So probably the most common example is stock price. And another really tangible one is weather or temperature data, right? Because you have your temperature and that that temperature happens at a certain time. But what we're finding out, or I think what people are finally coming to recognize is that time series data is present in almost every industry. So if you think about industrial IOT or any sort of industry where you have chemical industry, petroleum, et cetera, biotech, it doesn't matter. You have a lot of sensors, you're monitoring your environments.

• 10:19Speaker 03

  You need to find out the pressure and temperature of maybe a pipe or a heat exchanger, and you need to collect that data to make sure that your process is running smoothly and that you're not going to have any risk of explosions or any sort of damage to your process and the people that work there. You have examples of time series also in DevOps monitoring, continuous integration, application monitoring is a big one. So obviously it exists all throughout tech. We can also think about time series also existing for patients in healthcare where you need to monitor maybe the different attributes of their health over time. And we also have customers using.

• 11:03Speaker 03

  Influx for monitoring the growth of their farms or their greenhouses. We have customers using us to monitor solar panels. CERN used influx to monitor all of their particle accelerators and help them find the God particle. So really, time series and probably because we live like we are in spacetime. It's like time series data exists everywhere. So people are finally coming to realize that that data is valuable and that they probably could benefit from actually trying to store it.

• 11:39Chris Benson?

  So I guess you really have demonstrated that it is just about universal. You know, there's an application for it in, I guess, most any industry. I am curious, just as a follow-up to your own background, what is it about working with time series data that has attracted you personally? And of that, do you have a particular use case that you've worked on that was the most interesting to you?

• 12:02Speaker 03

  Yeah, I really like the CERN use case. I like it for two reasons. One, because as a developer advocate, I help the open source users primarily. And so anytime that I have an open source user who's doing something really cool with the product, it makes me happy. And they were able to monitor all of their particle accelerators using the open source. I think that's pretty cool.

• 12:23Chris Benson?

  And by the way, you mentioned God particle a moment ago. I'm assuming that you're talking about the Higgs boson. And for just, I know this isn't a physics thing, but if you would take just one second for anyone that hadn't heard that and might think it's a religious connotation rather than a scientific one, could you tell us for just two seconds what the Large Hadron Collider is doing in that project that attracted your attention and what the Higgs boson is? Kind of short answer?

• 12:48Speaker 03

  Sure. So basically, they're colliding atoms into each other to try and figure out all of the subatomic particles. And there's one particle that was called the God particle, and it is actually known as the Higgs boson. So that's like a subatomic, like, micro particle. I don't actually know the right word because I'm not a physicist.

• 13:08Chris Benson?

  And Daniel, you should hop in as well.

• 13:10Speaker 03

  Yeah, please.

• 13:11Daniel Whitenack?

  You're doing great, probably better than I could do, actually.

• 13:14Speaker 03

  So yeah, so basically, after two protons collide, then this is a byproduct, this Higgs boson, which is sometimes referred to as the God particle. And they were trying to find that and because I think it was at the time, and maybe it still is, I don't really know where the phrase comes from, but I think it was like the smallest particle in the universe. So the idea being that maybe that's where everything else came from in the universe. And so if we can find or prove that the Higgs boson exists, then we can find like the most fundamental building block of the universe. And that could be referred to as the God particle.

• 13:50Daniel Whitenack?

  Yeah, so there's like, they call it the standard model, I think. And this is outside of my domain as well. But yeah, it was like a missing piece of that standard model that could help them really put all the pieces together of how things were formed. So yeah, it's super exciting. Were you able to go visit CERN during that project? Or just, just talk to people?

• 14:10Speaker 03

  No, I actually didn't talk to people and I wasn't at the company when this was happening. I heard about it afterwards. Yeah.

• 14:16Daniel Whitenack?

  It's just cool, isn't it?

• 14:17Speaker 03

  Yeah, exactly.

• 14:18Daniel Whitenack?

  Well, maybe you'll get assigned to that project at some point. That'd be cool. I'm sure that they still have other time series that they need to analyze.

• 14:26Speaker 03

  I hope so. That would be cool.

• 14:28Daniel Whitenack?

  Yeah, so I mean, I was actually trying to think about like, while you both were talking, I was trying to think about, what is like not a time series that I work with? And there's certainly things that aren't, you know, time stamped that I work with. But pretty much any data could be time stamped, right? Like I was thinking of, of images like I'm taking with my phone, right? And like, an image is that sort of, you know, matrix representation of reality. But actually, you know, as I scroll through my phone, you know, it says photos from today or yesterday or before. So there's actually a time series of photos on my phone. So it's really kind of all encompassing. And I guess it's time series data could be like a timestamp

• 15:13Daniel Whitenack?

  matched with any type of data, not just like a numerical type of data. It could be, it could be other, other forms of data too, right?

• 15:22Speaker 03

  For sure. There are a ton of papers out there that incorporate the use of LSTMs, long, short-term memory networks for image classification, like you're talking about. Because it turns out that if we wanted to classify, like, let's say the scenes of The Breakfast Club, and we took any random Still from that movie, it would probably contain four kids and it would be really hard to classify. But if we use a temporal element to those images and we look at the images that came before and the images that came after as an effort to classify various stills from that image, then all of a sudden we've provided context, right, for those images and it vastly improves the classification of images.

• 16:07Speaker 03

  So yeah, you're completely right. Like even though Influx isn't really a platform to storing that type of data specifically, that's definitely also can be thought of as time series.

• 16:27Speaker 00

  This episode is brought to you by CoopCon/CloudNativeCon and you are invited to attend this flagship conference from the Cloud Native Computing Foundation. CoopCon/CloudNativeCon North America 2019. It's happening November 18th to the 21st in San Diego, California. This conference gathers adopters and technologists from leading up the source and cloud native communities. Use the code KCNA PracticalAI19, again KCNA PracticalAI19, to get 10% off registration or check the show for a special link to register and a link to the Conventure Boss Letter. Again, check the show for links to learn more and register.

• 17:17Chris Benson?

  So wanted to start off the next section by asking about what InfluxDB is. Could you give us a little bit of background about what Influx is and kind of what it's trying to solve?

• 17:29Speaker 03

  Sure. So InfluxDB is a time series database and it is trying to solve the problem of being able to store or ingest time series data. And what makes time series data unique is that there's usually, you usually need to be able to write huge, huge volumes. So Influx has been made as an append-only database to prioritize really high writes to allow you to ingest data at the nanosecond precision and also be able to then query that data in real time.

• 18:11Daniel Whitenack?

  So like if I'm just thinking of stock price or some like common time series example, you have like your timestamp and you have like the the stock price. Maybe in a simple example, it's just those those two things. So you could want to store those very quickly over time, right? And then what is a kind of query like that you might make on that stock price data? Is it like, I want to see the stock price from this time to this time? Or I want to see like, what was the average stock price during this time? What are the sorts of operations that you might want to do on time series data in a time series database like InfluxDB?

• 18:52Speaker 03

  For sure. So it uses two languages. It depends what what version you're using. If you're using 2.0, then we have created a functional query language and scripting language called Flux. And it's kind of JavaScript-esque. It has a lot of piped forwards, which to me help increase the readability of it. And that would be like from this bucket called, you know, stock price. You know, I want to filter by this particular ticker and I want to specify my range as having this start time and this end time. And then you can apply a whole bunch of different functions to it, whether or not that's in

• 19:33Speaker 03

  the case of stock prices, applying various sort of analytics to those stock prices, like the Shanday momentum oscillator. Or maybe you want to do things like apply the the average or find the derivative or look at the standard deviation for a group of time series. Yeah, there's a bunch of different functions you can do. And then if you're using the 1.x line, then you can use InfluxQL, which is like SQL. It's very similar. And so you do select maybe all from this particular stock for from the last five minutes or whatever.

• 20:10Daniel Whitenack?

  So we're kind of starting to get into, I guess, these are just kind of query operations or queries that you might perform on time series data. Maybe you could give us a little bit of a sense as well about like the term time series analysis. And you know, you have time series data, let's say you have time series data, it's stored in nicely in influx DB, you can kind of query it in these ways to kind of get the data back in various different ways that you might be interested in. What is this whole range or this whole topic of time series analysis about? And kind of what buckets of analysis might you want to do? Like I'm thinking of forecasting, for example, it might be one type of thing, but maybe

• 20:54Daniel Whitenack?

  there's a whole bunch of different things. Could you let us know what those things are?

• 20:58Speaker 03

  Yeah, so I mean forecasting is one big bucket. Of course, that's why people collect time series data is because they want to try and predict what's going to happen in the future. But another is anomaly detection and trying to figure out if your environment is running smoothly or your plant is running smoothly and trying to protect against failures. And then beyond just forecasting, which is extremely complicated, you might need to look into the different statistical elements of your time series in order to find out which forecasting method you should use and which anomaly detection method makes the most sense.

• 21:34Chris Benson?

  So I'm kind of curious, I'm going to approach it from the side of if you're a developer who's getting into time series data and you may or may not have done anything in the AI space, does influx data automatically provide you a set of tools for which functions you might use? If I was a developer and had a use case in mind, how might I know what functions would be appropriate to apply and how would I go about doing that?

• 22:00Speaker 03

  So out of the box, influx offers triple and double exponential smoothing. And that's a statistical forecasting method. So it doesn't involve any machine learning, any neural nets really. And so that's all that comes out of the box with influx. Of course, there are client libraries, so you can always, you know, use some Python library or R library if you're choosing that you're more familiar with. The act of figuring out which forecasting method you should use for your time series data is extremely complicated. It can be almost as complicated As you want it to be. Every forecasting method, every classical forecasting method and neural net has some underlying statistical assumptions about your data.

• 22:46Speaker 03

  So sort of one of the first steps that you can take is making sure whether or not your data violates one of those assumptions. Or on the other side, matches the assumptions that maybe it contains like, for example, for Holt winters or triple exponential smoothing, one of the assumptions, two of the assumptions is that your data is non-stationary, meaning that it has trend. And the second assumption is that seasonality is present. So if your data doesn't have seasonality or it doesn't have trend, then you don't want to use Holt winters to generate a prediction or forecast. So that's kind of the short answer. Does that help?

• 23:21Chris Benson?

  Yeah, that helped a lot. I appreciate that.

• 23:22Daniel Whitenack?

  Yeah, so it's like, I don't know about maybe other people. I definitely get what you mean by there's so many things at play here. So like, I kind of, whenever I look into time series, and I have a couple of times in the past, especially when I was working with a telecom startup and doing some monitoring stuff, but there's like all of these elements of it, it's like, how many like lags in your data is important? And like moving average and seasonality, like trends, all of these things. For like people getting into this, do you have any suggestions for like a starting place or maybe like a starting type of data that they could

• 24:06Daniel Whitenack?

  kind of experiment with to kind of learn a little bit about all of these different elements? Because I definitely see what you're saying. It can be overwhelming.

• 24:13Speaker 03

  I actually recommend that people, like I try and identify the problem that they want to solve first, because I think if you have a real a problem, rather than just exploring theoretical data sets, you're a little bit more tied to the problem and you're a little bit more motivated to dive into the different attributes that your time series has. And also, hopefully if you picked the data set, then you have some domain expertise about that data set and you understand it better. So I always, you know, recommend looking at a data set that you're familiar with. And then beyond there, in terms of good tools, I mean, I use scikit-learn. I'm a Pythonista, so I will probably use scikit-learn to do sort of.

• 24:57Speaker 03

  Initial discovery about my dataset and dive into the different attributes of it, looking at things like you're talking about like lag, auto-correlation, correlation between other datasets, all the statistical analysis, standard deviation, et cetera. Just to get a feel for the attributes of my dataset. And then whether or not, I think the next step, or the very first step maybe is to determine whether or not your problem is univariate versus multivariate. So multivariate meaning that you have multiple time series that you want to account for when you're making a prediction or an anomaly detection. And then the second, univariate is where you just have one time series. And the reason why it's important to identify whether or not

• 25:43Speaker 03

  your problem requires multivariate analysis or univariate analysis is because the way that you handle those two cases is entirely different. Turns out that if you are looking to do, for example, univariate time series forecasting, statistical methods work extremely well.

• 26:00Daniel Whitenack?

  And by statistical methods, you're kind of meaning non-what we would consider like machine learning or AI methods in... I know that's like a very convoluted thing, like drawing the line there. But that's kind of the sense you mean, like statistical as in non-machine learning, I guess.

• 26:18Speaker 03

  Right, yeah, like no neural nets, I guess, is how maybe I would describe it. Yeah, everyone has a different description for machine learning. I hear sometimes people consider linear regression as technically machine learning because it like uses a machine to make like a forecast. But I'm like, I disagree. I'm like, no, I don't think we can call it linear regression, machine learning.

• 26:38Daniel Whitenack?

  I'm with you on that.

• 26:39Speaker 03

  I guess I kind of make the distinction that neural nets, I'd say.

• 26:43Chris Benson?

  So I have a quick question for you there. As we, you know, we started talking about neural nets a little bit. I'm kind of curious, how does time series data and a database, in this case InfluxDB, how does that fit into a workflow? If you're starting to think about neural network training, or deployment, or whatever, and you're kind of trying to put everything together that you and your team may need, where does this fit into that process?

• 27:07Speaker 03

  We don't have, I mean, I don't know very many people who are actually employing like online machine learning with neural nets. So a lot of people are find that using really simple methods, like certain standard deviations away from the mean to define an anomaly works just fine for their use case. And so they don't bother with really fancy tools and methods. I guess I would probably, if I were building one myself, I would probably look into using h2o.ai with influx together, or maybe like BigQuery.

• 27:40Daniel Whitenack?

  And when you're talking about like, online versus offline, am I correct in, so like, online, you're kind of monitoring a stream of data that's coming in, a stream of time series data that's coming in in some way and applying some method. Offline would be like, oh, you pull a bunch of, maybe you have influx DB and it's storing time series data, and then you like make a query and pull some data out and then like load it into scikit-learn or something like you're talking about and do some sort of retrospective or historical analysis on it. Is that the sort of distinction?

• 28:20Speaker 03

  I think you basically touched upon it. I would consider online machine learning to be when you need to update your training. So, training, especially for neural nets, can be pretty expensive and time consuming. But if your data is changing a lot, then you might need to update your model. And so that would require retraining your model. And offline would essentially be that you've already trained your model, you only need to do that once for whatever reason, which unless your data is extremely consistent, and if it is that consistent, then maybe you can just use statistical methods. So, I'm going off track, but online, updating your training, training again. And then for me, offline is maybe training just once.

• 29:18Speaker 04

  This episode is brought to you by Brave. The Brave team is on a mission to fix the web by building an open source, privacy-focused, and performance-oriented browser. Browse the web up to eight times faster than Chrome and Safari. Block ads and trackers by default. And reward your favorite creators with the built-in basic attention token. Yes, you heard that right, a real-world use case for blockchain. Download Brave for free using the link in the show notes and gift tipping a try on changelog.com.

• 29:59Daniel Whitenack?

  So we're just getting into kind of statistical versus machine learning and also online versus offline. But if we kind of go back to the statistical versus machine learning side of things, I know that earlier in our conversation and also in some of your talks, you've talked about when you might want to go after statistical methods versus machine learning and neural nets and when you might want to do the opposite. Could you dive into that a little bit more specifically around time series? Like, what are the signs maybe in your data that you should be looking for when statistical methods are enough? And maybe they're better in terms of interpretability or efficiency or whatever? And what are

• 30:44Daniel Whitenack?

  the signs that maybe you need to do something a little bit more or maybe pull in a neural net?

• 30:49Speaker 03

  Yeah, so my answer, and you know, everyone has a different opinion, but for me, I think it makes sense to use statistical methods when you are only dealing with univariate time series data and use neural nets if you're using multivariate time series data and you're looking to do forecasts. There are some pretty efficient ways to do anomaly detection with multivariate data that are statistical or really simple. But yeah, so I'd say if you're looking at a group of time series, then use machine learning. Otherwise, use statistical methods. And I came to this conclusion because are you familiar with the Makitaiki's comps or M comps?

• 31:31Daniel Whitenack?

  Uh, no.

• 31:32Speaker 03

  Okay, so they are the benchmark for time series forecast methods. Now, unfortunately, they only evaluate univariate time series data, but they take 100,000 time series and they invite Researchers from all over the world to participate and try and come up with the best forecasting method. And this, this event happens every year and then the results are published.

• 31:56Chris Benson?

  Sounds like a party.

• 31:57Speaker 03

  Yeah, right. Um, last year, I think is like in June, they just released the most recent results. And what they found was that a hybrid method of um, an RNN and exponential smoothing outperformed every other model. But if we looked or evaluate just the statistical methods versus machine learning methods, just the pure statistical machine learning, the statistical methods outperform the machine learning methods. So while there might be some combination methods and some hybrid methods that outperform some statistical methods in univariate time series forecasting, really statistical methods, if you're just trying to like not generate your own forecasting method because you don't have that time or that resources and you're

• 32:41Speaker 03

  looking to pick between one or the other, it makes sense to use statistical for univariate time series data. That being said, we have the luxury now of monitoring a lot of different things, getting a lot of different data. And, you know, depending on the cost benefit to your business and the type of business decisions you'll be making based off of your forecast, it might make sense to go and spend the extra effort to create multivariate time series forecasting and incorporate neural nets and tackle that problem, which is a lot harder.

• 33:14Chris Benson?

  So got a question, and I remember actually uh, watching your talk on YouTube, you covered that as well and talked about that comparison between statistical and machine learning and the fact that the statistical came. What I am wondering is, could you take that and put it into more of a kind of a real life example, just to make it tangible on where you might see that in reality come about? And it doesn't have to necessarily be a real event that you were part of or something, but just how you might think of it that way. So that if someone's struggling to follow why and they hear you say that the statistical outperformed, kind of explain why that's the case.

• 33:49Speaker 03

  Why? I mean, the simplest answer for me is that a lot of neural networks, like that are commonly used for time series data like RNNs and LSTMs, and I'm not talking about hybrid methods, just plain ones. They operate on the assumption that your data Or the evaluation of the forecast operates on the assumption that your data doesn't exhibit autocorrelation. And autocorrelation is when a portion of your time series data is correlated to another portion of it in an earlier time. And that's often the case in the world where like if we monitored my hunger levels throughout the day,

• 34:32Speaker 03

  because I live a very regular lifestyle and I am a creature of habit, I tend to be hungry at really predictable times of the day. And so you'd find out that my hunger levels on today will be highly correlated with my hunger levels a month ago. And like you'll find that I'm hungry at the same times a month ago. And so this pattern that exhibits that would be present in my like hunger data violates an assumption of how RNNs and LSTMs are often evaluated and it causes overfitting of the models. And so that's kind of like the shortest answer I can provide.

• 35:15Daniel Whitenack?

  That was a good answer. Thank you very much. So in the case of like the multivariate data, it's more of a like there's more data, there's more complexity going on. And so it may be harder to overfit and like neural networks are thus more, you know, appropriate, I guess. Would that be a reasonable statement kind of generally? It's a fairly general statement.

• 35:38Speaker 03

  For sure. I like that.

• 35:40Daniel Whitenack?

  And when you're talking about these sort of multivariate scenarios, I was just kind of curious from your experience working with developers, what's the sort of range of number of time series that people are putting together in these multivariate models? Is it like a whole bunch, like hundreds? Or is it generally like a handful of time series? Like, oh, you've got three different sensors and you're putting those together or something like that?

• 36:07Speaker 03

  Yeah, it can range from both of those extremes.

• 36:10Daniel Whitenack?

  So in the hundreds case, I imagine that these are pretty computationally expensive things and there's a lot of complexity in terms of the model and all that. Are there, like when you're working with time series, I'm trying to make the connection with some of the things I'm familiar with, like sequence to sequence models for text and that sort of thing. When you have like a whole bunch of different Time series, is it just kind of that, but on steroids, I guess?

• 36:42Speaker 03

  Yeah.

• 36:42Daniel Whitenack?

  In terms of like how you prepare the data and like the types of, you, you mentioned RNNs and LSTMs and that sort of thing. So it's kind of similar. It's just, you know, kind of on steroids, I guess.

• 36:53Speaker 03

  Yeah. I'm not sure I understood your question, but I agree.

• 36:57Daniel Whitenack?

  Yeah. Okay. No, it wasn't really a question. I was just trying to kind of get a mental model of it in, in my mind, I guess. But, uh, yeah, I, I was wondering, you know, we've talked a lot about a ton of different methods and you've mentioned and kind of described influx DB and what's available there. As a developer advocate, I wanted to give you the chance also to kind of share a little bit about how people might get started with influx DB and like what they might need to get spun up, or if they can test it on their local machine and maybe, you know, put some of their time series data into it and that sort of thing. How can people get started?

• 37:39Speaker 03

  Yeah. So we just released a cloud offering and there's a free tier. So that's probably the easiest way because you just have to create a signup and then you're good to go. Otherwise you can install the platform as a single binary and then I recommend checking out Telegraph, even if you are not interested in influx DB. So Telegraph is a collection agent. It's a single binary and it's plugin driven and it's database agnostic. So it's by far our most popular tool. There are over 180, 190 plugins. So if you're looking for a way to collect data and you haven't found something that you like, I recommend looking into, into Telegraph any which way. And it's completely open source.

• 38:22Daniel Whitenack?

  So when you're, yeah, when you're talking about collect, that could be like, oh, I have a Raspberry Pi. I have a sensor or something and I want to put something on that to get the sensor data back to my laptop. That's the sort of collection we're talking about, or is it different than that?

• 38:40Speaker 03

  Yeah. You can be collecting data from a sensor. You can collect data from, I mean, there's so many input plugins. You can collect data from any other database. You can collect data from. CSV or JSON, you can collect data from Jenkins or M Q T T or, I mean, like, like I said, there's like over 180 input plugins. So if you can think of it, there's probably a way to collect data from that source.

• 39:08Daniel Whitenack?

  And if there's not, I'm sure that you would welcome contributions.

• 39:11Speaker 03

  Oh, for sure. Yeah.

• 39:13Daniel Whitenack?

  Yeah. So on that front, what is the kind of open source community like around influx DB? I find this interesting because I worked on like an open source data platform and our users were like data scientists and other people, but then like the developer community were primarily like backend sort of distributed systems people. Is it the same with influx DB in the sense that you've got kind of like separate developer and user communities where the users are like sort of analytics people and then developers are most of the time like database and backend people? Or is there a lot of overlap there?

• 39:51Speaker 03

  Um, there's some overlap. I think like you don't need to be a DBA to use influx. Like a lot of people, because I mostly talk to open source users. So I have a lot of people that I talk to that are just like, hey, I'm using influx to monitor my vegetable garden or like my barbecue or they just have a home project or they're trying to like make their house a smart house. And so sometimes they're just developers that are curious about getting into data science and data analytics. And so they just want to have a fun project. And so they really, they come from all over.

• 40:24Daniel Whitenack?

  Cool. Yeah, that's great. And I'm sure it sounds like I personally need to try out influx and it sounds easy to get spun up and start using. And this whole conversation, I've been thinking of all the time series data that I'm probably not leveraging in my own world. I don't know about you, Chris.

• 40:42Chris Benson?

  No, I'm actually, while you guys were talking in the last few minutes, I was looking at influx DB on the webpage and looking at telegraph. And when you were talking about doing things around the house and stuff, I have a bunch of ideas right now. My wife has a new little garden area that we had this summer and I'm thinking maybe we can monitor data in the garden. I can actually get her wrapped up in that. So I am very thankful for that suggestion because I think I'm always looking for ways to make this AI stuff practical, not just for myself, but for my family who are not technical at all.

• 41:12Daniel Whitenack?

  Yeah, I'm excited to hear more about Chris's vegetable monitoring and I'm really excited that you were able to join us on Ace and share with us a little bit about time series and about some of the things you've been working on and your perspective on statistical methods versus machine learning, all that was really useful. And I hope we'll for sure put links in our show notes to influx DB and the other things mentioned. But thank you so much for for joining us. It was a great conversation.

• 41:46Speaker 00

  All right. Thank you for tuning into this episode of Practical AI. If you enjoyed the show, do us a favor, go on iTunes, give us a rating, go in your podcast app and favorite it. If you are on Twitter or social network, share a link with a friend, whatever you gotta do, share the show with a friend if you enjoyed it. And bandwidth for change log is provided by Fastly. Learn more at fastly.com. And we catch our airs before users do hear change law because of roll bar. Check them out at rollbar.com slash change log. And we're hosted on Leno cloud servers. Head to leno.com slash change law. Check them out. Support this show. This episode is hosted by Daniel Whitenack and Chris Benson. The music is by Breakmaster Cylinder. And you can find more shows just like this at changelog.com. When you go there, pop in your email address, get our weekly email, keeping you up to date with the news

• 42:31Speaker 00

  and podcasts for developers in your inbox every single week. Thanks for tuning in. We'll see you next week.