# What’s in that Alteryx Predictive Install? - Inspire 2017

Take a fast paced tour of the Alteryx advanced analytics functionality from Clustering to Predictive Modeling to Optimization. We'll slow down only for a moment to discuss why you'd use a certain predictive algorithm over another. This session is excellent in helping you understand when you'd use one model over another, and get more comfortable with the predictive models available.

**Video Transcription**

**Neil Ryan:**

Thanks everyone for coming to Las Vegas, coming to the conference and thanks for coming to this session. What's in the Alteryx predictive install? So I'm gonna give you a high level tour of all the additional tools you get when you do the Alteryx predictive install. I'm Neil Ryan, I'm a product manager at Alteryx, I work out of the Broomfield, Colorado office, that's our development headquarters. My area of responsibility and product management is the advanced analytics functionality in Alteryx, so pretty much exactly what I'm going to be talking about today.

I've been at Alteryx for about three years, before Alteryx I did ten years of data analytics in various places, building models for banks and insurance companies and government agencies. So I've been doing this for a little while.

We went public a couple months ago, so we have to put these slides in all our presentations, the lawyers tell me how to say "What you hear today may contain forward looking statements that are subject to risks and uncertainty. Do not use any forward looking statements or product descriptions to make purchasing decisions".

Okay. So, this might be one of the first things you see when you download, install Alteryx. Some of you may have seen this in the last couple weeks, some of you may have seen this years and years ago. Basically, this is the first choice you're presented with. Do you want to install just the typical Alteryx or do you want to install the Alteryx with the R-based predictive tools. And these are the advanced tools, it says. So that should make you feel pretty good.

So what we're gonna talk about today is if you go with that second option, what do you get? Well, you get 50 additional tools, a lot of tools. I think Alteryx designer comes with about 150 so we're adding a lot of tools to the mix here. They're spread across a bunch of different categories that you see at the top of Alteryx designer. Six or seven categories and so that's what we're gonna talk about today. We're gonna go through each of these new categories of tool that get added when you install the predictive tools.

But if we take a step back for a second and go back to that download screen, I just want to hone in on the predictive part of the predictive tool. So, could consider it a bit of a misnomer, you actually get a lot more out of your predictive tools than just predictive analytics functionality. I just wanted to spend a minute talking about the wording that people throw around these days when it comes to data science and statistics.

First of all, you hear about predictive analytics versus machine learning. What's the difference there? Is there a overlap? So Garth, he's actually a colleague from Alteryx, so he's got his opinion what the difference between machine learning and predictive analytics is. What else can we add into the mix? Data science versus statistics. What is the difference between data science and statistics?

**Crowd question:**

[inaudible 00:03:15]

**Neil Ryan:**

That's one take on it. AI you gotta through AI into the mix when we talk about this stuff. Where does AI fit into machine learning, this fit into data science, fits into statistics. So here's my take on it. So for me, statistics is where it all started, a branch of mathematics, dealing with data analysis. That's pretty much as simple as it gets in MYAB. You add data science into the mix, just kind of a more modern term for statistics to me. If we're gonna force a kind of distinguishing characteristic on it, I'll say data science is statistics that you do on computers. You can do a t-test with pen and paper, you're not gonna be able to train a random force model on pen and paper, it'll take you a while anyway.

Machine learning is definitely a subset of data science, this is the modern learning algorithms that make up a big part of what data science is today. I can certainly put a artificial intelligence into this mix, I'd say it's definitely a subset of data science but I'd say it actually, it could technically fall outside of machine learning when you get into expert systems. So expert systems, you take experts, data scientists, scientist, business experts, you program rules into a data base and you know, based on those rules interacting, you can get some level of human... something that looks like human intelligence coming out of a machine without having to train a machine learning model.

But within machine learning and artificial intelligence, I'm sure you've all heard a lot about these deep learning models coming out of the likes of Facebook and Google. To this point, mostly what these are really useful for is image recognition, video, speech recognition, translation, things like that. Things that maybe you guys don't do day to day in your jobs but one day, these deep learning algorithms, you know there is a lot of research today that's working towards applying these to more common business problems. So these deep learning algorithms, one day you might see them coming into the mix in Alteryx.

But what we're focusing on today is over here. This intersection of statistics, data science and machine learning across descriptive, predictive and prescriptive analytics. So descriptive analytics, answering the question what's happened, what's going on in my data now or in the past? Predictive, what's gonna happen? And prescriptive, well, what should I do?

So that's for the most part where these predictive tools or these data science tools that you get in the predictive install fit in. We've got data investigation tools for exploring your data, especially visually. And we've got predictive grouping tools, this is where our clustering tools sit in terms of the categories across Alteryx designer. So unsupervised learning for example. These are tools in our descriptive toolset.

Moving on to predictive analytics, this kind of... it's called the predictive install but it's really just one part of it, we got a predictive category, our machine learning algorithms, our learners for classification and regressional. Define those terms later. Time series also is another type of predictive analysis. It's similar to our predictive tools except this one, there's a time series component when you have monthly or yearly data for instance.

Then we get up into some of the more advanced stuff in the prescriptive category, what should you do? AB Testing, we've had this category of tools for a few years now actually. It's when you set up an AB experiment, so you want to test a price change, propose a price change with one of your products for instance. See how it affects profitability or a revenue but you don't want to do it across your entire business because it's risky, you want to just do a small test, see if it helps your business or not before you roll it out across the entire business. So we have tools to help with that kind of testing.

And then a little more recently, about a year ago I think it was version 10.6 of Alteryx, we added the prescriptive category. This category includes simulation and optimization tools.

So when you install the predictive tool set, you don't get just predictive tools. You get tools across all these areas of data science. All right, so let's start going through these, one by one. I'll give you some examples.

Start with the data investigation category. So this is where you really should always start. You want to get a feel for your data before you even necessarily decide what questions you want to ask of your data. With version 11 of Alteryx, we actually built data profiling into the browse tool so that's a great place to start with your data investigation. So every time you have a browse tool, you can click through the different columns and see plots of how the data is distributed. But we have a bunch of tools with the predictive install that give you some more advanced ways to explore your data visually.

So, the data investigation category actually already exists before you even install the predictive tools. It comes with base designer. These are just the additional tools that you get with the predictive install because they're based on our... so there's a bunch of them here, about nine. I'll take you through just a few examples, starting with the frequency table tool.

So if you have categorical fields in your data, categorical columns, you know, A, B, C, 1, 2, 3, here we're looking through some vehicle data so we're looking across how many gears different vehicles you have or how many cylinders. So these are all categorical fields. So basically what you get here is a visual depiction of the distribution in your data across the different categories for a given field. So here, cylinders. So vehicle data, your cars can have four cylinders, six cylinders or eight cylinders, how many of each do we have in our data set. So simple stuff but important, you gotta get a feel, say you find out if your data is dirtier than you expect to be then you need to go back to the source and fix it.

Next one I'll talk about, next tool is contingency table tool. It's like the frequency table tool but takes it to the next level where you can, instead of looking one column at a time, you can compare the columns against each other to see how the data's distributed across fields, across these categories. So you can also compare up to four fields at the same time. More than that it gets a little complicated so we limited it to four. And then we have a different visualizations, bar charts or heat plots to help you look across this.

Here is an example, sticking with vehicle data. We're comparing cylinders to horse power. So you have horse power on your x axis here, goes from 52 to 335, then you got cylinders on your y axis, four, six and eight and you can just see what makes sense, at a glance, is that when you have fewer cylinders, you get less horse power, more cylinders, more horse power. So, you're confirming... maybe you're just confirming things about your data but if you're less familiar with the data, maybe you're finding out new things here. You know, the visual side is all about pattern recognition and machines are getting really good at pattern recognition but I think that we're still pretty good at it, too.

The last one I'm gonna show in the data investigation category is the association analysis tool. So here again, we're comparing distributions between combinations of variables. This time, though, we're looking at continuous numerical variables. So not categorical data anymore, we're looking at numbers. So on the x axis, you get a list of all the variables that you want to analyze, and same thing, on the y axis, the same variables so that you can compare them against each other. In the grid, the dark red signifies highly positive correlation, dark blue highly negative correlation. And then you can click across this grid and get the scatter plot comparing those two variables.

So an example if I freeze the animation, now we've clicked on the comparison of miles per gallons, still vehicle data to weight and it's dark blue so we know it's negatively correlated, but we can see it in the scatter plot that the heavier the vehicle gets from two, to three, to four to five times the lower the gas mileage gets. So when you got a two ton vehicle, can be over 30 mpg, you get to the heavier ones, it's practically nothing. It's probably old data, you can really get better miles per gallon than that these days.

Okay, so sticking with descriptive analytics but moving to the next category within Alteryx, the predictive grouping category. So this is an entirely new category that you get with the predictive install and there's kind of two main types of analysis you can do in this tool set. Market basket analysis and clustering.

So there's a bunch of tools to help you with clustering and market basket analysis, I'll focus on two here. So in terms of cluster analysis, we have the K-Centroids Cluster Analysis tool. It's a bit of an intimidating name but basically this is when you have lots of numerical data, say on your customer base and so you have say demographic data like age, and then say you have behavioral data like how many times they have done clicks on your website, how often they visit your website, things like this. So cluster analysis is great for for example customer segmentation. So you have all your customer data, you want to do some marketing, you don't want to blast everybody with the same marketing email, you want to target it to the types of customers you have. So you can do cluster analysis, break it up into a reasonably sized groups of data and target your marketing towards those groups.

So this is unsupervised learning, meaning you don't have a target variable, you don't have something you're trying to predict, you're just trying to understand your data better and classify it. Oftentimes, once you've got your clusters here, you might want to go back to the data investigation tools to try to understand better the clusters that you've formed using this cluster analysis to help inform that marketing campaign you're gonna be using tailored to the types of customers you have.

The next set of tools we have in this predictive grouping category, the market basket analysis tools. So with these tools you can for example feed in transaction data that contains details of what products people have bought across various transactions and what you get is what types of products people buy in conjunction with each other.

So one of the classic example of what this could be used for is deciding what to put near each other on grocery store shelves. So a classic example is when you do the market basket analysis, you see that often in terms of when you buy milk, you buy diapers, if you put those two together on the end of your shelf, you get more people buying them, you get more revenue, more profit. What's interesting about this MB Affinity tool that I have highlighted here is that it can not only help with market basket analysis but you can kind of extend its use and so a classic example of this... so this tool creates what's called a co-sign similarity matrix where you have your products, it's kind of like the correlation matrix where you have your products along those axes and then in the grid it shows how they're related together.

And a classic example of this, outside of the classic market basket analysis example, is recommendation engines. So Netflix uses or has used a co-sign similarity type algorithm to help build their recommendation engine. So the analogy to market basket analysis is they feed in all the movies that people have watched and then based on comparing that to what other people have watched, you can make recommendations similar to products. So this is a pretty interesting tool, you can help it to build a recommendation engine if you're interested in building one of those.

So that wraps up descriptive analytics. Taking it a run up the ladder, we're going to predictive, what the install is named after. So predictive analytics, this is when you have historical data with various attributes and you have a known outcome. So say whether somebody bought your product or not. Or whether somebody clicked on a button on your website. And based on all the other attributes, you want to try to predict that so you train the model and then when new data comes in, you're able to leverage that to predict whether that thing is gonna happen or not. It's the predictive category in a nutshell.

Now this is our biggest category of tools in the predictive install. There's a bunch here. I'll try to break it down a little bit. The first eleven here, these are our learning algorithms. So we have a pretty suite, we have some old tried and true algorithms, Linear Regression, Logistic Regression, we have some more modern ones that are really robust. Boosted Model, Random Forest, these are the ones that you actually use to train the models. The rest are kind of supporting and miscellaneous. I'll go into these in a little bit more detail. These are used in conjunction with those learning algorithms.

So it can be hard to decide when to use which algorithm. There are a bunch of them, the names are a little opaque, what does Boosted Model even mean, what does Support Vector Machine even mean? So introducing the Alteryx predictive cheat sheet, I'll talk about this a little. There's actually a version of this printed out in the solution center if you go to the community area, so you can take home a copy of this. Obviously these slides will also be posted afterwards but this the Alteryx predictive cheat sheet and it's meant to help you understand when to use one learning algorithm over another when you start down the path of building your predictive models.

So the first thing to talk about is whether you're trying to predict a category A, B, C, true, false, that's called a classification algorithm or a classifier. And if you're trying to predict a number, say total sales or how many times somebody is gonna visit my store, that's a regression and those learning algorithms are called regressors.

So depending on what you're trying to do, it's gonna limit which ones of these tools you're gonna try to use. So on the top in the greenish, we have algorithms that only work for regression, trying to predict numbers. In the bottom, we have algorithms that only work for classifications, trying to predict categories. In the middle we have some of these more modern machine learning algorithms which can actually do both.

So couple other nuances here. The logistic regression tool is only gonna work when you're trying to predict one or two possible outcomes, a binary classification. Whereas all the other classifiers can actually work with three, four, five and two classes, so those are called multinomial classifiers. So that's kind of... there's like this little pattern around the tool. I tried to make... it says it's limited to two categories but it's hard to see.

Couple other nuances, the Gamma Regression tool only works when your target variables are Gamma distributed. The Count Regression is only when you're trying to predict a count, like the example I gave before how many times somebody's gonna visit your store. All the rest of these are pretty flexible, so really what you want to do when you're trying out these predictive algorithms is do a whole bunch of them on your data and then compare them to see which ones work best. Cause, really, oftentimes you don't know which of these algorithms is gonna be the most accurate for your data until you actually train it and test it. So you want to try a whole bunch.

And to do that, to compare the models-

**Crowd question:**

[inaudible 00:22:04]

**Neil Ryan:**

So, the Logistic Regression does not support multinomial, trying to predict more than two possible outcomes. So you would use any of these other ones that support classification.

**Crowd question:**

[inaudible 00:22:20]

**Neil Ryan:**

Yeah, Decision Trees, Spline Model, Boosted Model, Force Model, Neural Network, Support Vector Machine and Naïve Bayes can all do multinomial classification. You're welcome.

So, once you built up multiple predictive models, how do you compare them? Well one tool to help with that is one of these supporting tools, the Lift Chart tool. This works for comparing binary classification models. To compare regression models or multinomial classification models, actually I will direct you to our gallery, gallery.alteryx.com if you go to the predictive district, there's a model comparison tool that you can download and install for free. That's excellent for other types of models.

Then of course you have the Score tool so once you've trained the model on historical data, now you're gonna want to use it in production, you want to deploy it and score, get predictions for new data where you don't know the outcome. That's where the Score tool comes in, this works in conjunction with all of the other learning algorithms that we saw in the last slide.

So that's predictive analytics in a nutshell in terms of four offering in this category. I'll throw one extra in here, I'm not really sure if this belongs in the predictive category, but it's cool so I wanted to show it. The network analysis tool, so this is when you have network data, think social network, anything that might have nodes and edges in it. So nodes being entities, edges being relationships between entities. So here, we got some data from IMDB about Indiana Jones, the characters in Indiana Jones and when they appear in scenes together. We fed that into the network analysis tool and made a cool graph. You can use this for more businessy stuff but this is kind of fun.

All right, so sticking with predictive analytics in general but moving on from the predictive category of tools in designer, we have the time series tools. So like I mentioned before, this is-

**Crowd question:**

A lot of my colleagues are using logistic regressions with a weight of evidence algorithms, first to help them define... A lot of my colleagues use logistic regression with a weight of evidence analysis first to help them define which variables to use which is something that's very easy and built into Sass and something that I haven't been able to tell them how to easily do in Alteryx yet. Do you know, am I missing-

**Neil Ryan:**

Yeah, there's a bunch of different ways you can do. So that's what's referred to as... that's choosing the variables essentially that you want to use to build a model. So there's a bunch of different ways you can do that, the Stepwise tool can be used in conjunction with linear and logistic regression to help choose which variables to use. There's also a tool on the gallery again, you can download and install for free Important Weights where you can help decide which are the most important variables. In fact, a lot of these more advanced algorithms... sorry, I'm flipping around like crazy... a lot of these more advanced ones actually do that for you. So while it's training the model, it automatically decides which variables are most important, so you can, for instance, Boosted Model and Random Forest Model, just throw in all your data and it'll pick the most important predictors for you.

So Timeseries, it's pretty similar to predictive except you're adding a time series component, so you're not just predicting whether something's gonna happen or not or how much. But how much over time, over the months, coming months, over the coming years. But, besides that, they actually behave similarly, you use them similarly to the predictive tools in that we have learners to train the model and those are up top. So instead of eleven, we have two, ARIMA and ETS. Once you build the model, and again, you should really try to train models... train both ARIMA models and ETS models and then compare them with the TS Compare tools, similar to the lift chart tool in the predictive category and then once you've got your model, you can use the forecast tool, the TS Forecast tool to get those predictions similar to the score tool with the predictive category.

And then finally, the TS Plot tool, produces some nice visualizations to help you understand the models that you are producing. So you can get the predictions, you can zoom in on specific areas of time to drill down, you can get it across years and you can look at it how it gets broken down by the overall seasonal-, how does seasonality affect the prediction, how does trends affect the prediction. So a lot of good stuff in the visualizations here.

So, that is about it for the predictive tools. Okay, let's move on to prescriptive analytics. So this is getting into not just what's happened, what's going to happen but what should we do? So this is kind of near the top of the pyramid. So, two categories for this AB testing and prescriptive. We'll start with AB testing. I've mentioned before, you know, overall what AB testing is, you want to test a change that you want to make with the business on a subset of your business so you don't risk screwing things up across your entire business. But the reason we have tools for that is because it's a little more complicated than it sounds at first. You want to compare apples to apples when you create your subset. You might pick a subset of your business to test out this change on and then it might just be a unique subset of business, if it's really profitable, it might not generalize the rest of your business if it fails, it's possible across the rest of your business, it might have done well. So AB testing is a process that helps you set up those... that test subset to your control subset and control for things like trend and seasonality.

So, we have four tools to help out in this process. AB trend tool, to control for trend and seasonality, you know, maybe that in your test set, you had much faster growing stores than in your control set, what you're comparing it to, so you need to control for that trend. This helps you do that. Then you need to select your test stores, your treatment stores. Okay, so I'm saying stores because I'm assuming... in my head, the use case is you have a bunch of franchises and you want to charge six dollars instead of five dollars for your burgers and see if that drives people away and your profits go down or if they keep buying and you get more profit. But another thing you can do with AB testing is on your website, you want to make a slight design change on your website and you want to see if it improves click rate or traffic. So two different types of AB testing, two different use cases, I'm thinking the stores and the burgers and fries.

So you need to select a subset of your stores to actually perform the test on to change the price from five to six dollars. That's selecting treatment units, treatment stores, test stores and then you need to select a bunch of stores to compare those test stores to in your AB analysis, that's selecting the control units. And so you want to compare apples to apples, so these tools help you select control units, stores that are similar to your test units that you've already selected.

And then finally you use the AB analysis tool, once you've already selected your treatment and control units and we have, again, a nice visualization to help you understand the analysis that you just performed. So here we're comparing the treatment units to the control units, the treatment units are pink and the control units are blue. So you can see how they perform store by store on this bottom graph. You can see how they perform over time on these top line plots and then these overall summary matrix up here are great because you can see, they're green. So we color coded. So if it's green, good job, you can move forward, you should propagate this test to the rest of your business. We show you the expected lift, the expected impact and kind of most importantly, the significance of this test. So it could be that the test is inconclusive, maybe you need to do another test, maybe it's kind of a wash, whether you change the price or not but if you get good lift and you get a good impact and you have a high significance level here, then you can feel confident to either go forward if it's green or don't do it if it's red. And say, we just saved you a bunch of money.

Okay, that's AB testing. Finally, prescriptive. So this is simulation and optimization. So let's talk about optimization first. Optimization is essentially optimizing for a value given certain constraints. So I'll just explain it with a few examples. So grocery store shelf optimization. You probably want a minimum amount of variety on the shelves, that's constraint. You have a limited amount of shelf space, that's a constraint but you get different profit margins on different products you sell and you want to maximize your profit. So you want to maximize for profit, given the constraints of shelf space and minimum level of variety, that's an optimization problem.

Classic one is the mixing problem. You want to mix ingredients together to achieve optimal nutritional value but the ingredients cost different amount so you want to minimize for cost. It's another example. Workforce scheduling. How many restaurant servers should start their shift at what time, you want to minimize salary but maximize customer service. These are all optimization problems that our optimization tool can help you with.

Simulation, this is kind of... you know, you've heard of Monte Carlos simulation, this where this comes in. So let's take an example. You're a hardware manufacturer, you make harddrives. Hard drives can fail, they can fail on arrival, on delivery from Amazon, they can fail after one year, after two year, after three year. Also, you are selling warranties, you want to sell two year warranty, three year warranty, four warranty. So let's say, you want to sell a new type of hard drive, but you don't know which warranty would be most profitable. But you have done some testing so you know the probability that the harddrives are gonna fail. So what you can do is run simulations to project when these harddrives might fail versus is it gonna be within two years, within three years, within four years, then you can combine that with the sales models you have from other hard drive products that you've already put to market. Combine those two models and figure out which of these warranties you should be selling with this hard drive.

So, a couple tools to mention here, Simulation Scoring. This is a cool tool, because it works with the predictive models you've made with those brown tools, the predictive tools. So you create a predictive model with one of those learning algorithms we talked about. When you use the score tool, the brown score tool to score that, there's an underlying probability distribution when you're making that prediction and we just give you the mean value when we use the score tool. You know, you made a predictive model to predict sales and we give you our best guess as to what that sale's gonna be for that situation.

Now if you combine that predictive model with the simulation scoring tool instead, it looks at that underlying distribution and say you do a 1,000 Monte Carlo trials in your simulation, instead of just giving the mean every time, it's gonna take a sample from that probability distribution. So you can actually see what the chances are of the worst scenario happening or the best case scenario, not just the average scenario. So that's a cool tool.

So in this hard drive selling scenario you might use a sales model from your warranties for another product and you might simulate sales for that. Now, we also have the Simulation Sampling tool, which allows you to run simulations without having had a previously trained predictive model and you can do it in a few ways. It's kind of cool looking since I have the animation here but if you know the underlying probability, so we've done tests, so we know the chances that this hard drive is gonna fail on arrival, DOA. In this case, it's like 5.8%, you can put that five percent in to the model and get that Monte Carlo simulation. If you don't know the probability but you know the underlying distribution, you can tell it, I know it's a binomial distribution but I don't know the probability, look at my data, figure out the probability and run a simulation. And finally, if you don't know the underlying distribution or the underlying probability, you can have it look at the data and figure that out for you and do a Monte Carlos simulation.

So in this hard drive example, we can run a simulation for DOAs, we know it's a 5.8% probability, the hard drive's gonna fail on delivery, we can do one for whether it's gonna fail in the first five years or we can do one that can try to simulate what month it's gonna fail in. You can combine these three simulations along with the sales and then we can figure out which of those warranties is gonna be the most profitable.

So that's prescriptive. We went through all the categories, there's one more. The developer category. With the predictive install, we add the R-tool to the developer category. It's a powerful tool, you can write any R-script you want and embed it into your workload. This powers all of our other predictive tools. I'm doing a session tomorrow, called Predictive Wrappers Delight where we talk about how Alteryx is code friendly, you can do R-code in your work flows.

But just to wrap things up, say you're a eyeglass wholeseller. Every once in a while, the eyeglass manufacturers are going to tell you they got a new product they're trying to sell you and they want you to distribute it to the eyeglass retailers. You have a business decision to make, is it worth it to buy these and is it worth it to market it to my retailers. You know, you have a limited marketing budget, so you know, you have Alteryx, so you're not just gonna use your gut, you're not just gonna pick the glasses that look prettiest to you, you're gonna prep and blend the data. You want to get into some descriptive analytics, maybe you'll make a tableau dashboard and slice and dice the data and look at kind of trends over time and use your judgment.

Might want to take a step up, do some cluster analysis, see which of your retailers behaves similarly to each other, look at the demographics around those stores, figure out how it matches up to your inventory. You can take it a step further, do some predictive analysis based on historical data, which of my retailers do I think are gonna purchase these if I purchase these? Not only, which, that's classification, will they or won't they? How much will they buy? So regression. You can combine those two. You can do some time series, figure out whether it's better to buy in the summer or the winter. You could take it up to prescriptive and do an AB test. Try out buying this one time, this one time, you can do an optimization. You can say, turn into a shelf optimization problem. I have a certain number of warehouses, I have only so much space, I know these eyeglasses are gonna be more profitable than the others but I want an underlying base variety, so turn it into an optimization problem.

The point here is that as you get more and more sophisticated analytics, you're delivering more and more value. It's hard running a business, it's just an optimization problem. Optimization problem, you have constraints, you have to pay your employees, you have a limited number of employees but you want to maximize profit, growth, shareholder value. If you can turn your business problem into an optimization problem and solve that optimization problem, that's the holy grail. So I just... in closing, I'd like to encourage you, if you're using Alteryx now for data prep and blending, try descriptive tools. If you're using descriptive tools, try predictive, if you're using predictive, try the prescriptive. If you've aced the prescriptive tools, you should come back to Inspire next year and do a presentation to everybody.

Thanks.

Everybody please, do the survey. You can do it through your app.

**Speaker**:

We'll start Q&A now, if you wouldn't mind giving me two seconds to run to you so we can hear your question that would be great.

**Crowd question:**

A quick question on predictive tools. Do you have any recommendations for textual predictions, if I have user input basically, it's freehand text, freestyle text. It can have abbreviations, spelling mistakes, stuff like that. I want to pick out the key word that possibly can describe that entire sentence in one word. Is there a recommendation you have for how to do that?

**Neil Ryan:**

Yes, I have two answers to that question. One is, I mentioned the gallery. Gallery.alteryx.com. We put up some additional tools that you can download and install for free. One that's up there is called the Microsoft Cognitive Services Text Analytics Tool. Or something like that. So basically, it reaches out to Microsoft API, so it pushes your data up there, it does the text analytics in the cloud and then returns the answers. That covers sentiment analysis, key phrase extraction, topic detection, and I think it also does language detection.

The second answer to the question is make sure you go to Jay's and [inaudible 00:42:18] Keynote tomorrow morning.

**Crowd question:**

[inaudible 00:42:27]

**Neil Ryan:**

The question was, does the cognitive services text analytics API require a [inaudible 00:42:33] license? Not an [inaudible 00:42:35] license but you do have to pay and you get... you buy a subscription to this API specifically. I don't know if it's... it's not like buying a license, it's... you pay for the number of records you consumer in a month.

**Crowd question:**

A lot of these cool tools have cool visualizations that come out of them. I guess, I'm wondering how you free them from the workflow. Does that make sense?

**Neil Ryan:**

Yes. So you can use reporting tools to help do that. So when you have an output anchor on a tool where the interactive visualization comes from. To just view it, you attach a browse tool. But, if you want to export it outside of Alteryx, you can use a combination of a layout tool which is in the reporting category and a render tool and then you can actually write out that visualization to an html file and take that html file with you wherever you want to go and as an html file, it maintains that interactivity.

Any other questions? All right, thanks again and don't forget to do the survey.