AWS ML Summit 2021 | Transform your business with data-driven insights
[music playing] Welcome! Thank you so much for being here! My name is Karl Albertsen and I lead Product Management for AWS’s AI services. Today I'm really excited to speak with you about how you can be applying machine learning to your business analytics. By the end of this session we’re going to demystify what AI and ML mean here at AWS and you’ll be able to know how you can start applying our AI services into your business. Ultimately, this is going to help you better manage your core operations, save time while doing so and be more responsive to your customer needs. So, let’s first talk about what do we mean by business analytics? This is a really broad space, so we’re going to focus on a couple of areas where analysts frequently need to take decisions, or analyze and take decisions in near real-time, manage large complex data sets and those data sets are frequently being updated. This often requires historical trend analysis, anomaly detection or validating individual transactions.
By no means is this restricted to retail or ecommerce, but using that as an example can start tying these together. So, we think about how do we better manage and predict customer demand so we have the stock in the right place at the right times? How do we better manage a website's customer experience in the page views, clicks, add to carts so customers have a great experience? Or third, how do we think about the individual transactions or the views to make sure that they are real and accurate and not fraudulent? So, why are we talking about business analytics in the context of machine learning? I’ll ask a couple of questions to see if this sounds familiar to you. So, one have you invested quite a bit of time and money instrumenting your business from all across different sites, functions or regions, but find yourself left with just more and more dashboards and more and more reports, but still unclear on where you should be prioritizing limited resources? Or two, have you invested in rule based systems, but as they’ve evolved over time they’ve gotten more and more complex and become more and more rigid and unable to be agile for how customers change? Or third, maybe your customer base is growing. It’s a great problem to have, but it can also be a bit scary, because you recognize that the methods you have in place might not scale up very well to meet those customer needs.
So, this is where machine learning can help. ML methods thrive on very large volumes of data. They can pick up subtle trends that most humans can’t and they will be able to adapt and be agile to as those customer needs or demands or what they’re doing changes over time. So this can give you a huge advantage when you’re thinking about predicting upcoming needs, staying agile or spotting issues in real-time. So, let me give you two examples to help tie this together.
So one is imagine you’re at your local grocery store, the one in your neighborhood. They need to restock thousands of items every single day and they need to make sure that they have say, plenty of foods and vegetables in stock for you when you arrive, but also not too many so they don’t waste the food and it perishes. So, imagine you’re in the position where you need to decide how many avocados or apples do you want to purchase for the next day. Now imagine that you need to do this for 20,000 SKUs every single day across all your stores. And to make it a little bit more complex, now imagine you have an upcoming holiday and some of your stores are based in the city, some are based by the beach and while you think it’s going to be a sunny weekend ahead, there’s actually a chance of thunderstorms coming.
Oh, by the way, you also have some promotions and so forth. So, this is a hugely complex problem, but this is where machine learning can start to come in and now we’ll start to automatically learn the patterns, not only of historical purchases for individual items, but how those items relate to each other and how they relate to promotions or other different data sets, which can be predictive in terms of what customers will want to purchase. So, that’s a case of applying ML in terms of a forecasting problem. Let’s take one more example. Imagine that you’re running an online direct to consumer website. Each day, thousands of customers are signing up, leaving reviews or making transactions.
But this is again, how do you make sure that these reviews are accurate and those accounts are not fraudulent? In traditional methods, you may look at the individual sign-ups reviews, apply rules and then deep dive areas that look suspect. That can be very difficult to scale. ML takes this approach slightly differently. It looks at the individual transaction were fraudulent and picks up patterns across multiple attributes and what can be leading to that.
Then it generalizes those patterns, which can be applied more broadly to different regions, different languages, or do you think about more and more attributes coming into play, doing that in a much faster and agile way than humans can. Okay, now we’ve established the growing challenges of business analytics and how ML can be applied. You’re not alone facing these challenges. Here at Amazon we face these exact challenges as well.
Each day we deliver millions of packages across 160 countries. Additionally, we support the website that makes this all happen. Every second, customers are logging in, posting reviews, adding to cart and checking out.
Furthermore, we support not only Amazon, but many of the websites across the internet through AWS. And that has similar operational needs and need to be monitored on a second by second basis. So, in order to do this effectively, we developed machine learning techniques.
So, in doing so we’ve learned what works, what doesn’t, and what’s needed and frankly, building ML capabilities does require a unique set of skills. It requires the research and development of sophisticated algorithms and techniques, it requires building at the infrastructure to manage both very large data sets and multiple customers that scale across the globe. It requires the ML operation to train and maintain the models and then finally, it needs an interface that can allow it to be easily used through other applications and operations. So, we recognize that all these skillsets are needed, they’re not common and frankly nor should they be unless you're really specializing in the space. So, there’s a need out there for how do we bring ML into businesses across many industries and functions that don't have these special skills? And that’s how we thought about our AI services. So we took a step back and thought about our experience, what are the functions that are most critical to businesses and where can machine learning really benefit those? Those happen in a couple of areas.
So, one we thought about forecasting. Two, we thought about anomaly detection and three we thought about fraud detection. So, for each of these spaces we developed a purpose built AI service, Amazon Forecast, Lookout for Metrics and Amazon Fraud Detector. While these services are all purpose built and different, there are some similarities, mainly that we’ve abstracted a way, all of that undifferentiated heavy lifting such as managing he infrastructure, managing the pipeline and developing the algorithms to provide you the customer with the ability to bring these ML services directly into your business operations.
So I'm going to go through each of these in a bit more detail. So, starting with Amazon Forecast, I'll share two examples of how this is actually being used of customers. And then talk a little bit about how the service works. So, first if I take you back to that grocery example that we started with, that’s actually not hypothetical, that’s a real customer. That’s More Retail, who’s one of the largest grocery and hypermarket chains of India with almost 700 stores and thousands of SKUs per store.
They were dealing with the problem of how do they reduce their perishable food waste? This could be fruits, vegetables, meats, produce, etc., while also making sure they maintain very high in-stock rates. So, the key for them in using machine learning and specifically Amazon Forecast was that the machine learning techniques don’t learn patterns based on individual items, but they’re actually learning how those individual items correlate and work together. So, frequently when you go grocery shopping, you buy a basket of goods. By understanding how these different goods are purchased together, how they relate to different times of the year and holidays, or how they relate to promotions was key into them improving their accuracy.
So, ultimately they brought Amazon Forecast in to automate their repurchasing and in doing so were able to reduce food wastage by about 30% while maintaining and both improving their in-stock rate. The second example I want to talk about is with Anaplan. This is a very different use case. Anaplan is not using Amazon Forecast as an individual end user, they’re actually building their own AI and ML platform designed to help with business analytics overall. And so, they’ve designed a UX and capability specific to industries and functions, which they’re calling PlainIQ, but they need to do the sophisticated ML technology under the hood that can scale with their needs. So, we have partnered with Anaplan as they’re building out their PlainIQ platform and Amazon Forecast is that engine underneath the hood, powering those ML workloads.
As a result, they’re going to be launching PlainIQ later this quarter and be able to bring machine learning technologies to all of their customers. So, let me talk a little bit now about what is Forecast and how it works. So, Amazon Forecast is a fully-managed, time-series forecasting service. It extracts away all that infrastructure ML Ops, allowing you to focus on where forecasting can improve your business and demand planning needs.
So, how does it work? You can access it through either the AWS Console or through the API and it’ll bring your historical data in Forecast. So, for example, this could be your sales history, by item by store location and then you can also bring any other related data that goes with that. So, that might be pricing or promotions or a third party dataset which you think correlates well with demand. Behind the scenes, Forecast is going to be doing a lot. It’s going to be ingesting your data, cleaning it up, grouping, do a lot of pre-processing steps.
Then, it actually is going to be a layering on the internal datasets that we have that can help improve Forecast’s accuracy. A great example is what we call the Amazon Forecast Weather Index, which we launched at the end of 2020. That’s actually going to look at weather forecasting two weeks out across all of your locations and whether weather can improve demand planning accuracy in terms of what is going to be the overall foot traffic or what is going to be, certain items that will demand increases, it will apply the weather forecast automatically to those SKUs or items. After that, Forecast is going to train, optimize and host your model allowing you to create forecast and then export it or access it directly through the API and use it however you like.
So, a couple of key benefits with the service is one: it's going to bring you highly accurate machine learning techniques and technologies. Two, it’s going to allow you to import not only your historical demands or periods, but also related data sets which can help. And third, the output is going to be a probabilistic forecast distribution. This is a subtle but really important point, is that it allows you then to map your forecasting problem back to your business problem. So, for example, items have different cost of over versus under forecasting, meaning that if you over forecasted a perishable good, you’d have to then, there will be cost to the inventory either being held or being wasted, versus the cost of under forecasting, which might result in a missed sale or a lower customer satisfaction. So, ultimately we talked a lot about demand planning and physical goods, but forecasting in general spans multiple different use cases, so you’re thinking about waste and markdowns, think about improving in-stock rates, you’re thinking about how do you actually want to staff accordingly to meet demands that fluctuates by week, by day or by hour, these are all cases where time-series forecasting can help.
Now, next we’re going to talk about Amazon Lookout for Metrics and that’s our managed service for anomaly detection. So, I'll start again with two examples to make this a bit more real. So, first is a company called Digitata. They’re a mobile telecom operator that manages pricing and subscriber engagement for their customers. So, Digitata has complex pricing scenarios and they deploy Lookout for Metrics to make sure as they make changes there are not unintended consequences elsewhere in the system. This can happen and normally they’re able to identify and fix the problem within a day.
But with Lookout for Metrics they’re actually quite surprised and pleased that running Lookout for Metrics identified a price change causing downstream impact within minutes and they were able to deep dive that root cause and implement a fix within two hours. Ultimately, the speed helped them avoid about a seven or eight percent impact to their daily revenue, which is huge. Playrix is another example. So, Playrix is a digital mobile game developer and they’re focused on how do they keep the customer experience at many points of the game, downloading, install and usability up to par. So, they’re tracking multiple different pieces of data. This could be installs, this could be ad clicks, this could be user activity, all this by a region of the different systems that support it.
So, they needed something that could scale and be agile with these different needs. Currently they were using manual thresholds but that was nearly impossible to manage and to scale and adjust for different seasonal impacts or weekly trends. So by building Lookout for Metrics into the processes, it allowed them to really step back and apply a generalized solution across all of their needs. Lookout for Metrics is able to understand, move away from these manual thresholds that were fixed to understand kind of daily, weekly trends by region along with other subtleties and really improve their anomaly detection accuracy. Better yet, it reduced a lot of the false alarms and frustrating deep-dives and investigations that took hours.
So again, let’s talk about what Lookout for Metrics is and how does it work. So, just like Forecast, we’ve abstracted away all of those complex issues with managing a machine learning model and pipeline, allowing you to focus on what it is that you want to measure for anomalies. So in order to use it, you’ll connect Amazon Lookout for Metrics through connectors that we have built in or directly into your data sets that you want to be monitoring. As soon as you start ingesting those, it’s going to start learning.
So, it’s going to look at those patterns over time and from those start identifying what’s anomalous or not. Now, instead of just flagging this as anomalous, it’s going to give you both an impact summary, which is going to give you the highlight, the severity of the potential issue and give you not only the issue itself, but the potential root cause. This can be huge, as customers want to think about, or excuse me, want to action very quickly ways to improve or remove the anomaly entirely.
Another key feature of Amazon Lookout for Metrics is the fact that it incorporates human feedback. So over time, you can flag if this was an anomaly it’s not something that you should be worried about over time or in some cases, it’s things that it did not pick up initially and by adding that feedback over the system, it’s able to learn over time, improve and make more and more effective anomaly detection flags. And then finally, it has notifications. So you can build notifications directly into whatever service that you like, whether it automates next steps, or just notifies your operations teams to take next steps, but those notifications allow you to move very quickly to root cause the analysis, to root cause the anomaly, excuse me and take action. So, some of the key benefits here are its ease of use through the connectors that are built in.
it provides actionable results at scale, so not only it will provide anomaly, but it will help group related anomalies together and rank them in order of severity, so you’re not responding to a thousand anomalies all based on one issue, a whole group does together, you can respond to one of the potential root causes. And its third is that it continues improvement, allowing you as the user, who knows your use case the best to provide input to make it better over time. This is going to be really effective in processes that impact your revenue, such as order rates, shopping cart metrics, PHD telemetrics, can be used with marketing, downloads, installs, page views turn rates, customer experience as well as operations, this might be such as CPU utilizations, error rates, etc. Okay, third we’re going to talk about Amazon Fraud Detector in a couple of use cases associated with that. So first, we have CDKeys.
This is an ecommerce website specializing in digital products and sales and they pride themselves on being trustworthy, reliable and have a fast purchasing experience for customers. But like any ecommerce site, they’re dealing with fraudulent transactions at times. So, today they have, excuse me, prior to using Amazon Fraud Detector they had many different rules in place to identify what could potentially be a fraudulent transaction, but those rules were becoming more and more complex and harder to make adjustments to and act on in real time. So, they were looking for a solution which can generalize the results more easily, especially as expanded new product lines and new customer regions. And so Amazon Fraud Detector by not taking that rule-based approach, by learning from what was actual fraud or actual real transactions enabled them to do that generalization and be much more effective. So, once they implemented Amazon Fraud Detector they actually saw about a 6% reduction in the overall fraudulent transactions, but more importantly, all of these transactions that were not fraudulent, they were able to move through without any slowdown or manual debug, really improving the customer experience.
And then finally, with that they were able to reduce the number of transactions that they mainly expected by over 90% which freed up their resources and time to do things that would add more and more value to the customer. Truevo is another example. They’re a payments processor and they were looking for fraud detection solution. They were debating building out a bespoke custom model and map that out and realized that was going to take multiple months. So they looked to see if Amazon Fraud Detector was something that could provide them capabilities they needed much faster. So ultimately what they realized is that with Amazon Fraud Detector the minute they put in their historical data and train the model, they’re able to deploy something that was functional and quite accurate within 30 minutes.
Not only that, with the integrations in CloudWatch and AWS CloudTrail, they estimated that they could be saving about three to six months in their overall development time period in going with Fraud Detector versus the bespoke custom solution that they developed themselves. So let’s talk a little bit about Amazon Fraud Detector, what it is and how it works. So, just like the other services, we’ve abstracted away that ML pipeline, the algorithms themselves, the ML operations, the infrastructure, allowing you as the customer to focus on the underlying transactions and reviews or what it is that you want to review to see if there is fraudulent behavior.
So how does it work? You’ll start with what it is that you want to be measuring, those could be transactions themselves, they could be reviews, they could be new account setups and you’re going to put that in the data stream of what those transactions are, along with the files, the ones that you found to be fraudulent in the past and not fraudulent. Fraud Detector is going to be enriching that data, it’s going to train and optimize the ML model and it’s going to host the model and provide you with actionable insights. Users can review and visualize the score and distribution of what is potentially fraudulent or not and set the sensitivity based on how mission-critical that process is, what the resources are to deep dive and over time improve from there. After that the model can be deployed directly and they can automate the overall processes. So, some of the key capabilities that we want to highlight with Fraud Detector is one: it's approachable, so the graphical views or the model performance, it allows customers to kind of intuitively understand the trade-offs they’re making in terms of how sensitive do they want the models to be. Second, they will operate in real time or near real time, which is going to allow customers to actually flag and remediate those potentially fraudulent transactions before the issue becomes too big.
Third, it’s able to operate at scale, can adjust the end transaction volume or velocity that customers may have, allowing them to grow with their business. And finally, like the others, it’s a managed service, so we’ve done all the heavy lifting on the Amazon side. You can use it for new account fraud, you can use it for checkout fraud and you can use it for promotion abuse among many other use cases. So let’s take a step back and recap the services we talked through and the problems that we’ve talked through today. So, first off we’ve talked about the AI services that we internally have created that address some of the thorniest business analytic problems that customers face today where machine learning can be most applicable.
That’s forecasting, anomaly detection and fraud detection. And some of the key benefits we’re excited to share with you as we built these services ìs the ability to provide accurate and fast results that will help you drive business actions and help improve your customer experience. It’s going to leverage the past years of investment that we’ve made at Amazon in these problems and as we continue to invest going forward, which brings the most cutting edge ML techniques to your business. Ultimately we’re very excited this is going to save you time, it’s going to save you cost and it’s going to improve your customer experience overall.
So with that you might be asking how can you get started? How can you begin today? And coming out of this discussion I think there’s really two paths that a lot of folks and customers like to go down. One is how do they learn more. So you’ve got tons of information on our AWS website, each of these services has its own individual detail page that will contain overview of the service, the features, getting started material. We also have tons of blog posts, not only on individual features, but customer use cases and how they’re implemented, so you can understand and learn from other customers what problems did they solve and how they solved them with these services. Third, we have videos showing these implementations, it may be walking through the console itself or different examples through the CLI or other notebook examples.
And then finally, we actually have GitHub examples, in some cases example data and code snippets so you can be deploying directly into your business. If you’re good at that and you’re ready to start building, this is why I say start with our AWS Console. You can start building there. These services can also be accessed directly through APIs themselves and then to make things even faster, we have AWS CloudFormation templates for many common use cases, so you can skip ahead if you know that your business matches other, you know has a common need that we’ve seen across customers as well. And then if you have a more complex or larger implementation and you want to speak with us, feel free to reach out to your Amazon, or excuse me, your AWS Account Manager or potentially we can put you in touch with our qualified partners through the Amazon Partner Network.
So, with that a big thank you. I want to thank you for joining us today at the ML Summit and welcome you to begin building with AWS’s AI services. And also I do want to point out before we go that there’s over 65 purses cross, you know, not available at the ML Summit itself, but across our web and when you’re ready you can feel free to prepare for the AWS Certified Machine Learning Certificate that will include a specialty exam and it’s going to validate your skills and provide industry recognized credentialing.
And all assets mentioned during this talk and others are going to be downloadable as an attachment throughout the ML Summit. So again, thank you very much and I appreciate our time today.