Build Predictive Analytics in Tableau Using Machine Learning Models Deployed on Amazon SageMaker

Key Takeaways

hi everyone and welcome to today's aws tech talk where we will go over how you can achieve a male powered analytics within tableau by leveraging the power of machine of amazon stage major and machine learning kostyvas lagaikis and i lead the go-to-market strategy of sage maker as it pertains to all integrations and partnerships and today i'm joined by nathan mannheimer and holt calder nathan how about introducing yourself hey everyone great to be here uh like kosty said my name is nathan manheimer i'm director of data science and machine learning products at tableau that means i get to lurk on a lot of really cool technologies and techniques where we're helping our customers and our partners like sagemaker integrate predictions and advanced analysis into tableau and use tableau as a pathway to get those insights into the hands of the business awesome thanks nathan holt how about you yourself hello everyone thank you for coming my name is holt calder i'm a data architect with inner works and here today to talk to you about amazon sagemaker for tableau as the lead developer on that group so really excited to be here thank you for joining awesome awesome thanks gentlemen now let's see what we will cover today so the we will start with um first of all with the promise and the fundamental issues of enabling machine learning power predictive analytics across the organization and there has been a lot of discussion of how customers can do that and there has been also a discussion about the impediments of our customers not being able to really fully materialize that vision so we'll cover those two and then we will walk you through our amazon sagemaker for tableau solution how it works and what are some of the most common use cases we have here from our customers using it for then i will go through a demo of the three steps that are required to make use of our integration first of all we'll go through the deployment of the integration where a halt will uh guide us then we'll build a model with amazon sagemaker autopilot or go through a demo of autopilot and then nathan will use the model i'll create an autopilot and the integration that hold will deploy right within the tableau environment to make ml powered predictions and lastly we will offer some guidance on how you can get started and how you can get help and just keep in mind we're aiming to have an about 40 minute presentation and then we'll open it up for questions but as you see at the bottom of the screen there is a q and a area where we have live moderation throughout the session so the q a will happen in the end through the same live moderation on the q and a box below so feel free to use it at the beginning or at the end we have our aws employees addressing questions on the fly nathan how about we kick it off and we how about you help us level set on what we believe is the true promise of the ml power predictive analytics absolutely thank you kosty so over the last 10 or so years we've really seen a revolution in how businesses consume share and make decisions on information and on the data that they collect and generate business intelligence and especially visual analytics has really democratized access to that data and allowed orders of magnitude more people to understand their work understand their business understand their customers and really drive more effective more informed business decision making this is where tableau has been a leader in the space of helping people better access their data helping more people get involved get their hands on it better understand the state of the world and the way things have looked historically and start to interrogate process understand and present and share and communicate the insights that they've discovered the meaningful value they've generated on top of the raw data that they have at their disposal and increasingly as organizations get comfortable with that wider access to information that more robust ability to explore to ask new questions to socialize and communicate answers they look to the future and want to take the same ability to explore and process and communicate and ask forward forward-looking questions ask things about what might happen as conditions change as the world evolves as new data becomes accessible can we better understand what might happen next and use that information to plan to make better decisions to better allocate resources to ensure a better faster customer experience and so into the space we have the real power of machine learning artificial intelligence and statistical modeling to help us take the data that we have better understand it and then use it to answer questions about what might happen in the future and this has been driven by a really powerful surge in the technology the algorithmic understanding and the scale at which we can operate in the computational space so we have technology like gpu compute improved algorithms better access to data and really even just more and more raw data at our disposal which is combined to create this environment where generating machine learning models has become more realistic for more and more organizations more and more companies people who want to drive better business better serve their customers and so when we pair the ubiquity of business intelligence and tableau visual analytics with the customer need to ask questions about the future to better understand their world and make predictions and the power and the technology that we have in aws sagemaker and really the ability to do machine learning at scale we start to develop this really powerful promise for how we can deliver the next stage of analytics and make that something that is accessible to more people than ever before and so costi we talked a lot about the promise but i'm going to turn it back over to you because it's a grand vision but there are certainly some challenges along the way and let's look a little bit more into those certainly certainly and 100 agree with the promise here it's a very big but as you may already know um ml power predictive analytics are not that easy to materialize and there is quite some reason about it but first and foremost i'd like to start with the the fact that we have noticed there's in many organizations a gap between the data scientist persona and the business users even if those two roles might be in the same line of business we have noticed that the data scientists are not necessarily directly partnering with all the stakeholders from the business side and usually those two personas work with completely different set of tools so for the data scientists to expose their models or predictions they would have to either store their predictions in a database that the business has access to or they would have to create a completely separate application or even actually connect their models into completely manually and in an internal application that the business consumes and we have also heard from our customers that due to this gap of technologies the partnership between the two personas has been getting very limited in with regards to ad hoc request only and that's predominantly because there is a lead time from request to request because data scientists have to go through model development prediction and then the predictions have to be stored either to the database we discussed or to be connected with the application that the business consumes and this is not a matter of minutes or hours it's a matter of days or weeks so we hardly see in our customers any real time or even same day predictions coming into the business powered by ml and lastly it all comes down to the fact that the data scientists in a way these days are quite scarce of a resource making uh the creation of models an exercise that is compound in the amount of the data scientist a team actually employs but because we believe in the full promise of ml powered predictive analytics we actually decided to uh to partner with tabloid we aim to help alleviate many of those impediments that i discuss described here with a solution that we're about to present so how about you help describe your audience the solution we brought to market and which one that we will be also demoing in live in in a couple of minutes absolutely and so to understand the solution here we really want to look at three different components of this process first we have a real goal in this work of making it accessible and easy to get started with even though we're plugging together components that are doing some really interesting and sometimes very complex work we want to make the configuration and the ability to get started as smooth and seamless and push button as possible and so to do that we're using the aws technology quick start this allows us to essentially combine in the most effective way a number of different amazon web services into one package that we can initiate very easily with only a minimal configuration by the user so the user can essentially take this quick start off the shelf and get going with just the push of a button and a little bit of extra information provided and by initializing the quick start they're able to configure sagemaker configure all of the services that are required for authentication for logging for storage and provision all of that on amazon so that they can then start connecting to models start generating predictions and then deploy those models and plug them into tableau and so to do that they're working on amazon stage maker which is a platform for deploying predictive modeling at scale that can be done with data science teams using code or now increasingly for more business analysts users embedded with the actual business functions to start building models of their own using automated machine learning or automl and this sagemaker autopilot technology allows a real democratization of access to the ability to build test and deploy predictive machine learning models to understand those models to be able to share and inform data science teams if they have questions about the deployment but to really be able to go through that process from business problem to idea to data set to model uh and eventually to consumption much more quickly and when they're able to deploy those models into production they can then consume them directly in tableau so all they have to know is the name of the endpoint that they've deployed and they can connect it right into a tableau dashboard or visualization provide the appropriate data from tableau and then drop the predictions from that model into a visualization just like they'd use any other calculation in tableau so leveraging tableau's analytics extensions api we can communicate directly between stage maker models and tableau visualizations and really augment take visualizations in tableau to the next level by using real-time predictions on live data that can be connected right into a tableau visualization so anything that you can view in tableau today you can connect to predictions and start running those and then use tableau's own ability to scale up using tableau server or online to publish those dashboards to hundreds or thousands or tens of thousands of users and allowed them all to ask questions to explore to get their hands on predictions and really start using those as they would any other visual analytics to really run their business better make better decisions make those faster and access them where they're accessing their business information today whether it's through the browser on desktop on mobile so together these three technologies the quick start that ties everything together sagemaker which allows for scalable and accessible building of machine learning models and tableau which brings data and brings insight to people all across the business from a ceo to a development manager to a frontline sales rep we have this really powerful ability to actually take data turn it into predictive analytics and then democratize and share that analysis across an organization and really go from a data set to a working solution in a matter of hours or days so we have this really powerful potential that is addressing that last mile gap getting the predictions to the people who need them is allowing the ability to iterate and develop machine learning solutions quickly get predictions out in real time on live data and really help alleviate some of that challenge in a limited access to data science resources means more people can get involved start working but still have a completely auditable and accessible trail of their work to share if it needs to be reviewed by a data scientist or data science team or leadership so let's take a look now at a just a quick sneak peek of how this all works so here you're going to see us build a tableau calculation that's calling out to a predictive endpoint in sagemaker passing live data from tableau we can just add that calculation to the visualization and in just a second or two start to see predictions in this case whether customers will churn or stay with the business added to the visualization in real time and now we've added this to a dashboard that allows us to interact with and explore and find interesting trends or patterns based on those predictions in this case how international calling and the expense of that impacts the likelihood of a customer to churn and then publish and share that dashboard out with a large number of users who have an interest in its results and use that dashboard essentially as a front-end application for that predictive model so now i'm going to hand it back over to coste to talk about some of the customer scenarios where we see real value with this solution absolutely thanks nathan and one quick point i want i wanted our audience to be very clear on is that the art of possible with this solution and the number of use cases one can solve with this is really really vast i'm going to describe a couple of them but at the end of the day any problem that you have that is has tabular data set and is can be represented in the tableau dashboard can be enabled with the solution to embed ml power predictions using um what we're describing here and as you think um about for example propensity to buy a product which resembles very much the lead quality scoring use case that we will be showing and that can help any sales team no matter the industry to pre prioritize their attention on the opportunities that make the most sense but the list doesn't stop here for example you will see today term prediction or in other ways called predictive engagement are very common use cases our customers have been bringing up and are indeed applicable to a big variety of industries so prediction is um quite applicable in from the financial services all the way to healthcare all the way to a software intern and retail so we're excited to actually be covering this today and we'll dive deeper on how you can build it yourself then you can think of the credit risk prediction where our customers can evaluate an individual's risk profile completely on the fly or even the whole portfolio's risk just embedded into the visualization have male power predictions of how my book is doing or how this individual how what is the rich profile that my model is predicting then there is the demand and revenue forecasting where our customers can predict what is the anticipated revenues or what is the anticipated cost and what will the future will be in terms of the sales of a specific product or a whole line of business or there is also the hospital capacity and the patient readmission problem our customers from the healthcare space have been bringing up quite frequently where with this integration they can quickly analyze and optimize how many beds would be needed by when all within a single dashboard and as i said the art of possible is very vast and would love to know how actually this integration can help you so feel free to just think us below or email us after to understand how this might solve your own problems but now let's shift towards the demo and the the three steps that that are included in in in for you to be able to use the end to end solution so here we demonstrate the on the high level the three steps that we will cover today and in step one where halt will uh help us here we'll go through the integration deployment that happens via very easy to use aws quick start this step will allow deploy to allow us to deploy certain resources resources and actually help hold will be able to describe what are some of the tips and tricks that you need to know as you are going to deploy the solution and this is what will pretty much allow us to connect sage maker and tableau but important to note here is the fact that this step is needed only once so once an organization deploys it in the environment you have your models on then you don't need to deploy it again and then going towards the step two the step involves the creation of the machine learning model within sagemaker and this is again an optional step if you already have models deployed in sagemaker but in case you don't we'll walk you through how to leverage our automl capability that is called sagemaker autopilot to create an ml model even if you're new to data science and finally all the magic will come in step three we'll do a demo and actually nathan will do a demo on how you can use the model that i create in sagemaker to make predictions within tableau via the deployed integration that hulk would deploy and keep in mind that you can use either the tableau server the desktop or even the online version to get your real-time predictions lastly in today's demo we'll focus on a very ubiquitous problem i just described which is the customer turn prediction and spotting unhappy customers and offering them the right engagement and the incentives is a a quite dubious problem so today we'll show you how you can proactively go about predicting those turn moments with a tableau dashboard integrations now i'll pass it on to halt that will dive deeper on the quick start based integration but before we dive deeper on the deployment hold do you mind going over the reference architecture of the solution so for the technical folks out there to be able to deeper understand how this works absolutely thanks kosty as kosty just mentioned we're going to walk through architecture really quickly to just look at how the amazon sagemaker for tableau integration is actually deployed into your aws account looking at the reference architecture the quick start is going to launch a cloudformation template into your aws account the customer and end user will enter through a calculated field in the tableau desktop or tablet server application and from there the data will be sent into a suite of resources that create the integration and produce inference results from amazon sage maker if you look at the architecture diagram here everything contained in the quick start resources pane is going to be the primary api using the analytics extension api framework that powers the integration after data is sent through the amazon api gateway service it is translated using lambda functions to the necessary format to invoke your sagemaker endpoints these things all communicate via a vpc endpoint that is deployed into your account and basic authentication is handled using amazon cognito and a lambda authorizer through the api gateway service if you look at the solution and how the quick start is deployed the necessary im permissions are created that allow the integration to connect to any sagemaker model endpoint in your aws account and it allows you to govern and authorize users using cognito by signing up and signing in different user accounts for different users who will be utilizing the integration and producing dashboards with sagemaker insights now let's see how we can actually go and deploy this into our aws account we're going to jump into our demo we're going to get started on the quick start page for the solution here we have some information about amazon sagemaker for tableau we have a description of everything that's going to be launched through the quick start an architecture diagram the deployment options along with some information about the costs and licenses that could be associated with launching the solution what we're going to focus on here is the deployment options so tableau amazon stagemaker for tableau is actually available in three different launch configurations we could deploy the solution into a new vpc that the cloudformation template would create for us we could actually deploy the solution into an existing vpc if your organization already has a vpc in your amazon web services account that is a kind of the standard you're able to utilize that existing vpc and then we could also deploy the solution not into a vpc at all so for this example we are going to deploy into an existing vpc i'm just going to select this hyperlink and it will actually launch the cloudformation console inside of my aws account so here we have the template for our amazon stagemaker for tableau solution i'm going to go ahead and hit next and we can go ahead and start filling in the parameters for our launch configuration so i'm going to go ahead and give this a different name i'm just going to call this demo usc s1 and then we will go in and configure our vpc parameters so what we're going to do is select that we would like to launch this into an existing vpc or into a vpc and what we'll do now is we'll select the vpc id that we would like to launch the solution into kind of building out from here what we'll do is actually open up our vpc console so that we can retrieve the subnet ids that we would like to launch the solution into so i'm going to go ahead and open the vpc console and we will go and grab the subnet ids here in a little bit so we'll just select this one we want to get the id here go to subnets and we just want to confirm that the subnets we select in our cloudformation console are for the correct vpc so we have a fd here we'll want to select the uh we'll just pick some subnets here just make sure that they're from the correct console so we'll select that so 982 you can see that that's for the correct vpc we'll do a 937 as well great so that will be our subnet configuration and what we will do now is actually move into the part of the parameters that are specific to amazon stage maker for tableau what we're going to do is select a hosted zone id so the two requirements for amazon stage maker for tableau are that you need to have a domain name that is hosted in amazon route 53 and you need to have a certificate attached to that domain and that certificate needs to be provisioned in the us east one region so here is where we will actually pass that information into our quick start cloudformation template so i'm just going to go ahead and select the hosted zone id and this is going to be the domain that i have in my aws account for the solution and then what i'm going to do is select a fully qualified domain name so i'm actually just going to type this here with a prefix that i want to deploy the amazon sagemaker for tableau solution to so i'm going to call this is tech talk us east demo and i'm going to add a period so that will be my prefix and then i'm going to type that domain name that we see below now that we have added the domain name what we're going to want to add is the arn of our certificate which is hosted in amazon certificate manager i'm just going to type certificate manager above and we will open that console in a new tab and we'll go and retrieve that arn of the certificate that we've provisioned in us east one so we will select the certificate and what we're going to want to do is just grab this ar in here so i'm going to copy it and bring it back over to our cloud formation console i'm going to give that arn to the certificate arn parameter and uh from there we can leave these quick start parameters alone these are just going to be the default parameters that point to the the assets in the quick start which are in a publicly accessible s3 bucket so now that we've filled in all of our parameters the vpc parameters as well as our route 53 domain name and our certificate arn we can move through the rest of the steps to deploy the stack so i'm going to go ahead and hit next here and we're on the review screen what i'm going to do is just scroll down and select that all of these settings and capabilities are are allowed so i'm just going to acknowledge these give them the check mark and then i'll select create stack and it will actually begin to deploy into my aws account now so amazon sagemaker for tableau is kicking off i'll pause and resume the recording as soon as it's complete great and as you can see our solution has just finished deploying so on the left hand side when we look at our stacks we have that flag that says create was complete we can look at the details of the cloud formation templates outputs and resources so we can look at everything that was created as a part of the quick start this is all going to correspond one to one to the architecture diagram on the quick start home page and what we'll really want to focus on now are the outputs here so we do have a link to a post deployment steps this is going to be the quick start deployment guide for this example what we're going to want to do is actually sign up a user so we will open this link in a new tab it's going to take us to a cognito url where at this point we could sign up a user and that user would be able to access and use the solution and connect to sagemaker from their tableau dashboards jumping back into the cloud formation console this url that we specified when we launched the cloud formation template is going to be the url that we're able to access sagemaker from within tableau so now that we've deployed the integration i want to call out a few tips and tricks that will make this process easier for you as you get started with amazon sagemaker for tableau before launching the quick start review the deployment guide so that you ensure that you have all of the necessary prerequisites and understanding of the services that are going to be deployed a quick tip that will help you from getting stopped in the deployment process is to make sure that you provision your acm certificate in the u.s east 1 region and make sure that you launch the integration into a vpc that is accessible from your tableau environment and is able to access your sagemaker environment verify that the security group created in the quick start conforms to your vpc security policies and a tip that i like to call out for customers is to deploy the solution to a subdomain as it will allow you to use a wildcard cert on a domain that your organization already owns instead of procuring a specific domain just for the integration now i'll pass on nakasi to talk us through how you can create an ml model using the automl capabilities of sagemaker autopilot awesome awesome thank you halt great demo so with regards to amazon sage maker in general i want to make sure that our audience out there has a level set understanding of what sage maker is and how you can go about using it especially with regards to building some models and building some models quickly so amazon sage maker is the most complete end-to-end service for machine learning it is a managed service for data scientists citizen data scientists data engineers an ml operations teams that helps remove all the addition forecasted heavy lifting associated with machine learning and first off as you see on the upper part of the screen sagemaker provides users with an integrated warplanes of tools brought together in one single pane of glass which is called sagemaker studio and users can launch jupyter notebooks and jupyter lab environments instantly through the save maker studio and from the same environment users can get access to a complete center feature around data preparation model development training and tuning experiment management debugging deployments and pipeline automation to actually help all the personas in the ml lifecycle be a lot more productive secondly though and this is where our focus will be today sagemaker offers an automl capability called autopilot now why people love autopilot it's because it automatically builds trains and tunes the best machine learning models based on your data while allowing you to maintain full control and visibility building a machine learning model requires you to manually prepare features test multiple algorithms and optimize hundreds of different model parameters in order to find the best feature engineering pipeline as well as model for your own data however this approach usually requires deep ml expertise and that's where the war paints i just described described above fits right in if you don't have the expertise though you could use an automated approach like automl and that's where the sagemaker autopilot really signs it eliminates the heavy lifting of building ml models by providing such oitml capabilities allowing pretty much everyone out there to build ml models with just a few clicks but let's see how it works exactly so as i mentioned amazon sagemaker autopilot makes the ml um process a lot easier and a lot faster so you can build classification and regression models by just providing two things a tabular data set and a target attribute and that's it it's that simple then what autopilot does it automatically analyzes your data set and like any data scientist would do an exploratory data analysis and then explores all the machine learning solutions with different combinations of data preprocessors algorithms and and hyper parameters to find the best combination of those as well as the best model and hyper parameters instead of requiring you actually to decide what feature engineering to do what algorithm do you use and what hyper parameters will work best for your model autopilot does all of those things that the data science would do completely automatically with just one click and in the end it will present you with a leaderboard of all the combinations it actually considered and will show you the best performing models at the top then you can quickly um you can quickly evaluate if you want to deploy the best or second best you can evaluate them all and then with a single click of a button you can deploy it to an endpoint so i will showcase in just a second you'll be able to see all the different combinations with run how they performed based on the objective metric that decided it is appropriate as well as if you want you can even see all the code and inspect it but of course that's only if you want so for our citizen data scientist and ml users out there today i'll showcase how to trigger an autopilot job and then how to quickly deploy the best performing model so let's switch to our aws console here we first enter the platform and we navigate to a sagemaker and more specifically we will navigate to the sagemaker studio then we will which will take us to the jupiter lab environment that you get for the studio experience and in today's session we'll go over and search for within the amazon sagemaker examples for the autopilot examples that is focusing on customer return so if you open the notebook environment on customer churn here the first thing you'll see is the description of the problem and of course losing customers is very costly and divorce and usually it's better to identify those customers early on and give them the right incentives instead of actually spending money to acquire new customers so here we're going to focus on this customer sure use case and we'll first set up the environment we'll run our first line of code to get access to the sagemaker libraries and then we'll get access to a couple of standardized pandas and numpy libraries as well to be able to download the data here we're downloading some um open source uh discovery knowledge uh in data data set where you can always access also uh for free and then if we um open the data set we just download it you'll see in a descriptive analysis here we can see the state we can see the account length the area code how many minutes uh during the day the customer spoke how many calls per day how many evening calls etc which we will use in the model but in the end you'll also see the churn uh variable where we're gonna try to predict today both with the autopilot but also in the tableau integration so we'll go down and we'll split the data set in 80 20 so that we have 80 percent of the data set for training and 20 of the data set to be able to test this within tableau here we i'm doing the randomizer on 80 of the data and then i'll upload those into an s3 bucket which i'm just going to copy so i'll take this training data set and i'll navigate towards the autopilot experience so we go here on the left of the studio we'll open the sagemaker components and under experiments and trials you'll see create experiment we'll go create an experiment and we'll name this we'll give it a name then we'll use the data the training data i just created in this s3 bucket so we'll copy the s3 bucket in the in give it as a data set and then we'll also give it the target attribute which is the turn question mark so we give it a training data set give it a target variable and we just tell it to okay use this s3 bucket for output you can give any s3 bucket here you also see that you can select if you want autopilot to go through an automatic identification of the problem or you say oh this is actually a binary classification which actually is certain yes or no true or false or it is a regression or remove the class classification but because we believe autopilot is smart enough we'll use the auto and of course there is other uh more advanced capabilities with regards to giving certain permissions selecting how many combinations you you want this autopilot to select but will go with the vanilla case here so we'll press the exp create experiment and all we need to do then autopilot will go through the pre-processing understanding what is the right kind of defini uh that it should explore go through feature engineering and then it will start producing different models that it actually has tuned and it has actually optimized for the right hyper parameters so in the end we'll get a list of a leaderboard of all the different options it considered as well as a list of the performance it actually achieved so here we achieved 0.96 in a binary f beta score which means it understood that it's a binary classification problem although we didn't state that so we'll go in the best option the best alternative and here you have a couple of options either to inspect the model and understand what exactly it produced and why and which model is created or you can decide to just simply deploy the model and for convenience today here we just go in the deployment states here you can use it either for real-time prediction or bus predictions but given we want to offer real-time predictions in tableau will go with a real-time option so in the real-time option i'll give it a name to the real-time endpoint and remember that end point the customer turn endpoint is what actually nathan will use in the calculated field to make the predictions uh in within tableau so we name the endpoint customers endpoint we'll give an instant a small instance because it's a small data set and then we just deploy the model so it's going to kick off the deployment process and if we go in the endpoints we will see the customer endpoint being created so we'll refresh and now the customer endpoint is in service and we can proceed with actually using at end point within the tableau environment but before we get there i'd love to give some tips and tricks about going about using autopilot so here you um i have some uh of the most common advice we give to our customers first of all always split your data set between 80 20 70 30 90 10 but always make sure that you're using a test data set that your models have not seen also within autopilot you can use the explanability report to start building some trust and so that you're able to understand what exactly the model is um seeing as more important features than others also keep note of the schema and the endpoint name i just described because you will need those two within the calculated field in tableau and if you use a vpc ensure that both the endpoint and the deployed integration that walk us through are under the same vpc so let's now go to the third step nathan how about showcasing our audience how to use the model i just built to make ml part predictions within tableau awesome to do that let's start by jumping into tableau desktop so here we are in tableau desktop and we're looking at a normal tableau visualization in this case a scatter plot where we're comparing individual customers in our telecom data set by the number of international calls they've made and the charge that they received for those calls we can see that generally there's a trend downward which means that as people have made more calls they are getting charged less on average but now let's add predictions to this data set so we can better understand what's really going on here and how it's impacting our business to do this i would first go to configure the analytics extension for stage maker so i go to help settings and performance and manage analytics extension connection to connect to sagemaker i would choose the analytics extension api connection type and then enter the information for my connector in this case a hostname and a port and a username and password credential in this case we're both requiring authentication through aws and securing the connection with ssl so now i'm ready to go and actually start using predictions in this analysis of this data set so i've created this calculated field churn prediction and if i open it up here i can see that i'm using what's called a script calculation in tableau and script calculations allow us to do a lot of really powerful and flexible things by connecting to analytics extensions and so here because we're connected to the sagemaker analytics extension it means that using a script calculation allows us to return predictions from the model that we've just deployed in sagemaker all i have to do is reference that customer churn endpoint and then pass the appropriate fields with the right level of aggregation from the tableau dataset i'm connected live to a version of this data set and i'm actually able to pass data in real time but i could also be connected to an extract or published data source in tableau and this would work just the same and by passing these fields i don't have to pre-load them into another data source outside where they natively reside i'm able to pass them in real time to sagemaker and get inference from the model and return it to tableau so i've selected the appropriate fields and chosen my endpoint so now i'm going to click ok and just take this predictive calculation and place it on color and we're actually going to pass that data to the model and return the predictions whether the customers are at risk of churning in orange or whether they're at low risk in gray and so here we see something interesting pop up that we saw that general trend down for customers overall and that seems to hold for customers who aren't likely to churn but the relationship is much more flat for customers who are predicted to be at risk of churning so it seems like on average they're spending more money on international calls as the number of calls goes up so there's an interesting beginning of an insight here i'm going to take this visualization and now put it in my dashboard that provides a historical and predictive view of the risk of churn in our data set and now if i select customers in that sort of upper right hand section of the visualization those customers who spent more on international calls as their number of calls went up we see that they're generally more at risk of churning this includes customers who had an international plan as well as those who didn't so we can see that in this case based on our predictive model having a plan doesn't have a significant impact but that we have customers who are at risk in both groups and will want to develop a plan for making them allowing them to have an easier path to stay with the business and potentially mitigate some of those increased costs to help them stay with our business so i built this dashboard the dashboard is acting as a front end to the predictive model but it's really not doing me any good unless i can communicate this insight out to my colleagues in the business and so to do that i'm going to jump over to tableau online and i've published this dashboard up there where i can share it with my colleagues and so here we're in tableau online and as a site administrator in online i've gone to the settings section of my site and i've gone to extensions and then configured the sagemaker extension via the analytics extensions api connection so we're in good shape here the connection is set up on my online site so then i could publish that dashboard up and consume it directly through the browser and share it out to wherever people are consuming it whether it's another browser or on mobile and start to interact with it and share these insights so we see here i've highlighted that same selection that i made in desktop and i've commented with a snapshot of the dashboard allowing me to share that with my colleagues jim and julie discussing the problem or the interesting insight that i've uncovered based on the augmentation of visualization with predictions and then starting the process of coming up with a plan to take action so in just a few clicks i brought a predictive model into my visualization uncovered an insight and then shared that to where any number of people can see what i've seen or even go further and ask questions of their own so this covers our entire end-to-end workflow in tableau of integrating the model publishing and sharing the results and collaborating with teammates so let's cover some tips and tricks for how to consume these predictions in tableau because of the script calculation type we're passing data from the tableau visualization in real time to sagemaker so it means that we need to construct a visualization that's at the right level of detail or the correct granularity of the data so that when we're passing it to the model it's in the right format in this case we're predicting on individual customers who are characterized by a phone number so when i broke the visualization up by phone number it provided that correct level of detail for me to accurately apply the predictions we also need all the variables in the training set to be used in the same way when we're actually making predictions the only thing we don't need is the actual outcome because we don't know that yet but all of the inputs to the model should be provided in the calculated field like you saw me create so that we can pass them to the model to generate that inference but we don't necessarily need them in the visualization itself this means that you can construct a visualization like you saw with not that many variables in it but still pass a larger number through the calculation to the model the type of calculations that we used script calculations are table calculations in tableau which means they can be controlled in terms of how they partition and address the visualization so as a recommendation we would suggest to set all of the dimensions in the view onto addressing this means going in and modifying the table account to ensure that there are no partitions because of the way that the script calcs work they'll make a separate call to sagemaker for each different partition and in the case of inference here where we're just returning predictions from fit model there's really no advantage to doing that and setting everything on addressing will improve the computational performance of the calculation and finally to actually deploy the results and share those out with your colleagues and teammates you can deploy workbooks using analytics extensions in both tableau server whether it's deployed on premise or in a cloud provider like aws or via tableau online all you have to do is go in and configure the connection like you saw me do and then you can publish workbooks to enable consumption of predictions by the business at scale so now i'm going to pass it back over to costi for some tips and advice on how to get started awesome awesome thanks nathan so how can you get started it's quite easy here in the first link you can find the quick start itself that will help you deploy the integration with just a few clicks that hold showcased and if you're part of the business or a data scientist that doesn't have the privileges to deploy the integration do get in touch with your it admins to help and as we discussed this integration only needs to be deployed once and if you are the it admin or you actually have the privilege to deploy this you have a lot of supporting material other than the deployment guide that you will find in the quickstart we have built blogs we have recorded demos and you can find them in the second link that you see in the screen in the tableau.com solutions aws and last but not least we'd love to hear from you so um if you have feedback or you actually need help from our technical teams feel free to reach out at the sagemaker for tableau at amazon.com and we will help guide you with the right technical resources as well as help guide you with any other additional questions you might have in terms of what use cases to use it to or anything else you might need help with and of course don't forget in the chat would be below you can use um you can use it and you can ask any questions you have we have live moderators responding throughout the hour and will stay till the top of the hour so that takes me to a big thank you to all of you for joining us we'd love to hear from you again and either use our direct emails or get back to us on the email i just sent before and i'd like to really thank nathan and hall for joining us today it was great to have you and we're very excited about what we built together thank you