AI Summit San Francisco 2019
September 26 2019, San Francisco, CA, United States
AI Summit San Francisco 2019
Video
Algorithmic Retail Using AI to Create Advantage in a Digital World - Publicis Sapient
Available
In cart
Free
Free
Free
Free
Free
Free
Add to favorites
63
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Rashed Haq
Vice President of Robotics at Cruise

Currently making self-driving vehicles, the hardest engineering challenge of our generation, a reality. To address the AV challenge, we have to combine AI and Robotics in a way that hasn’t been done before. It is pushing the boundaries of both disciplines.

View the profile

About the talk

Share

Good afternoon. So I'm going to talk a little bit about algorithmic retail. And this was about how AI is being used across the whole retail 00:05 organization. About to start with I'll go back to a little bit of prehistory, which is back in 2013 Mark 00:15 Andres and had said software eats retail and what he was supposedly talking about those the advantages of e-commerce over traditional retail 00:24 and it became very notorious for that within the retail space except what he didn't say and 00:34

maybe he didn't realize it at the time is that retail also eats software. And what retailers are doing using software and 00:44 data and AI is pushing back through these three things that they're doing. So the first one is around friction a frictionless 00:54 dropping where they're improving the convenience for the customer. This is both in the e-commerce channels, but also across the 01:03 physical stores. The second one is improving the operational efficiency in the supply chain process across the organization and the 01:13

third is creating new experiences both through content digitally, but also within stores 01:23 Not all retail companies are doing that. So there is a set of companies that are on the left side that are still thinking about 01:33 what they're going to do and how they're going to do it and they're playing around with a few things to try out. 01:42 The majority of companies are in the metal bucket where they're taking a i and applying it within some of the 01:53 departments in retail. So that might be and marketing it might be an e-commerce or in supply chain and improving 02:03

their performance for the organization within those groups. And then there's a set of company is about 18% of 02:12 retailers from surveys that have been done recently that are transforming themselves and they're doing that in two ways. One is their building 02:22 AI in all parts of the organization at scale. That's one part. And then the other part is in the process of doing that they're 02:32 breaking down the silos between the different department. So you can see companies that are that were historically doing demand forecast for the 02:41

supply chain bass stuff historical sales that they're now augmenting that with clickstream data or ad Impressions data and 02:51 dramatically improving the accuracy of those demand forecasts. So in the transformation process, they're breaking down the silos across these 03:00 organizations. So now the companies that are thinking and some companies not thinking if they're taking too long 03:10 there. We've seen some examples historically over the last few years of the disadvantage. But for the 03:20

companies that are using algorithmic retail and the patterns from algorithmic retail are getting a huge Advantage 03:29 financially from those and this is a business case that's showing the different areas where they're creating advantage and the 03:38 column second from the right is showing what algorithmic retail is changing within those organizations. This is a model 03:48 based off of 10 million 10 million dollars and revenue overall 1.2 of that coming from their eCommerce 03:58

Channel and you can see that even on the lower end of those percentages. It's locked significant numbers 04:08 and in some cases. It's going to be higher than that. But it's not just about the advantage from 04:17 one-time advantage that. They're getting they're also getting a benefit from the cumulative effect of what you can 04:27 see on this graph from Mackenzie on the left side. When you look at what's going on in inside of that it's that ask the 04:36 become more algorithm driven that's improving their business, which is allowing them to collect more data and better quality data 04:46

and that allows them to create better algorithm. So that virtuous cycle is giving them a huge Runway to 04:55 create Advantage for themselves. So those are some of the benefits on the types of things these companies are doing a very quickly go through 05:05 some use cases and then talk about what are the organizing principle behind behind algorithmic retail? 05:15 If you look at the the operations and supply chain part of the retail business there many use cases. Some of them are listed out at 05:25

the bottom here where it is being used in various companies. And if you look at a customer intelligence for marketing to 05:35 sales to e-commerce and recommendation engines and applying that physically inside of stores as well. You can see many use cases 05:45 around that and I'm just going to dive into an example of that first one. So in this 05:54 use case, this was a company that was on their online Channel. They were anybody who came to browse the channel 06:03 browse 30 Commerce site who are Anonymous Prospect? They're not customers yet when they left the site that they would be tracked and they would be 06:13

sent a digital ad retargeting at and in that process they were doing fairly. Well, they're getting a 2% conversion 06:23 from the from the retargeting but it was standard retard targeting based off of the people that are visiting. 06:33 Then the things we added there are some of these new data sets on the left side. So they were already using the clickstream data for retargeting and 06:43 we added ad Impressions data and then product search data from from the e-commerce site and most importantly 06:52

collecting in-store behaviors of customers. So when customers go in the one that connect to the Wi-Fi in the store, then you 07:01 have their device ID. So you can start to connect the dots between those Sony or aggregate those you can make much better behavioral predictions 07:11 using that we try to understand the customer and in particular. One of the things to do is look at which of those are likely to 07:20 convert and become buyers and which ones will not become buyers and then for the buyer so we decide that in the next step we decide what 07:30

kind of offer to recommend to them and then Create an Oculus tomorrow and finally showed that some sort of prospects 07:40 digital ad. Nephew look at this picture on the left. This is showing the the customer segments. Sorry. 07:50 This is showing the customer Journey classification based on their likelihood of converting becoming buyers. So the green or likely the blue or 08:00 unlikely the orange or in between so that that's one of the first steps and then the second step. Is creating an uplift 08:09

model which is this picture on the right side where the vertical axis is showing? What is the probability of conversion for each of 08:19 those customers if you make no offers. And the horizontal is showing if you make an offer. What is the probability? How is the probability 08:29 changing? So then you can see that at the bottom ride the green part. I think it's a green those are customers who are 08:38 likely to increase their conversion or their probability of conversion increases so you should definitely show them the ads 08:47

the strip in the middle of our customers who are not going to be moved by the offer. So if you do show them the ad you're wasting your marketing 08:57 dollars, I'm done surprisingly enough. There's a set of customers on the left which are who are likely to buy but now that you've shown that my 09:07 digital offer they're less likely to buy so there you're not only waste your marketing dollars, but you lose Revenue because you've turned away some 09:17 customers. So going going through the process increase the conversion from 2% to 6% and then it also decreased 09:26

because you're not using adds to Target. Most people but they're much smaller subset. It's decreasing the cost by 43% 09:36 So between those two it's a 5x performance gain in customer acquisition night. So. That's just one example, but 09:46 then the question is if you have a hundred or so use cases like that sometimes more. How do you get organized to be 09:56 able to do that across the whole company? And as you would expect there are four key things that we will look at one is 10:06

collecting more data building a platform and then integrating that into the existing 10:16 Journeys weather that's in the supply train or the customer Journeys and together with those you also have to reorganize 10:25 your company on your team's so that they're more oriented towards doing more and more experiments from which they can learn which Jose 10:34 talked about very eloquently this morning. so we've done a number of projects like these with 10:44 retailers and over the last five years and I'll share some of the lessons. We took away from from those in in the next few slides. 10:54

So this first one is when we're collecting the data often people collect data that you know, whatever is available. That's how we started or 11:05 a little bit more on an ad-hoc basis, but eventually what we realize that we're trying to do is create a digital twin of the business. 11:15 So it may not be a perfect win. That digital send me to but you want to collect as much data to be able to understand what you're doing in the 11:25 business and that means understanding the customers and their intent and behavior getting a descriptive digital twin of 11:34

all your products that you're selling our marketing and then a digital to enough all the processes that are going on within the organization. 11:44 Pain on the left side or some of some examples of the data that we're collecting. So with that you can start 11:54 to understand what's going on in the business and then make decisions around those. The other thing is when you start collecting the 12:04 stay there that it's coming from many sources and you have to make sure that the data is connected. So when you got the Wi-Fi device ID, if a customer 12:14

who was in the store and add impression you have to connect those two so that you understand the behavior of that customer across different channels. 12:23 They have to create a customer graph. You have to create a product graph. You have to create process graph or in some cases a few process graph so 12:32 that you know how all these pieces are connecting and then you can make decisions around it. And the other part is being able to carry through 12:41 data from one process to another and that's the picture of the bottom of showing the Adam presents data for customers on the left side from a DMP 12:50

and it's showing a customer data platform where you have the customer online Behavior transactions physical behavior in store since 04 13:00 And we think many companies collect a lot of Intelligence on the left side in marketing and when the customer 13:11 converts and starts buying their on the right side, but all the intelligence from the left side is now not carried over to the right side. So you lose 13:20 a lot of intelligence using which you could do personalization and customization that sort of keeping the lineage of the 13:29

data. This next one was about using the modern and emerging AI techniques 13:39 in dr. Are available now and if you look at how a i yai capability has improved since 13:49 2012, which is why I was calling 2013 prehistoric times in terms of AI the capability Improvement. It has been three hundred thousand 13:58 fold. Particularly with deep learning and if you look at our entire history of anything with ever done as a species there has 14:08 nothing has improved 300,000 fold in the span of seven years and then some people talk about but more slow, but if you look at more slow, it's 14:18

almost flat line compared to this Improvement and more slide South is an exponential gain. So this is exponentially faster 14:27 Improvement than Moore's Law. Sew-in using these modern techniques 14:37 particularly in deep learning the main reason to use that as a few compare its performance to classical machine 14:47 learning you got that 10 or 15% gain and performance in terms of accuracy and often much more depending on what the 14:56 previous model look like the details of this but 10 or 15% gain in performance, for example in a 15:06

recommendation engine or knot ad targeting on demand forecast has huge business upside when it's implemented 15:15 into the business work clothes. With those types of advanced model. The other thing we 15:25 realized overtime is when we're going into doing a use case not to think of each use case as being solved by one model 15:34 but thinking about using multiple models interacting with each other to be able to solve that problem. So in this example, the 15:44 left column is showing something like a product recommendation or content recommendation. The second column is talking about I want 15:54

to show the customer that recommendation weather that's time or point point in customer journey, and the third column is 16:04 looking at using what channel we should be communicating and the the advantage when we get these experiments over time and in 16:13 different types of retail settings, we found that your recommendation might give you a to 3% increase in conversion. 16:23 But if you combine it with the right time and the right channel we can often see that go from 3% to 12% 16:33

That's at 34426 Sox game 1 of use multiple models. And when you start using multiple models, you have to 16:43 now you're going to run into a scaling problem. If you don't step back and think about how you're going to manage those models. So this is the picture 16:53 where the bottom row is showing the tables inside of a customer data platform each of the black circles as a ml model 17:02 and then the white circles or how it's getting activated inside if a customer journey and the lines are showing how does interact with each other. 17:11

So so you might have five of these black circles being used to activate something with a customer like like the previous 17:20 example of the recommendation and the timing and the channel But each black circle may be more than one model. So if you take 17:30 even something as simple as a recommendation engine, if you build one recommendation engine that works across all your products and customer you might 17:40 get a certain level of performance. But if you build multiple Ones based on customer segments or product types and suffer the 17:50

performance within each of those segments is very likely going to be much higher than one generic model send that to each 18:00 black circle, maybe 10 models and then you have to keep in mind when you're going through the AI life cycle, you're doing 20-30 18:10 experiments / model that you actually put into production night. So now you're getting 5 * 10 * 20 * 1000 18:19 models to make recommendation engines really work and gain gain that huge. increase in conversion and 18:29

the way to deal with that kind of modeling scale is to build an AI platform that has a lot of cell service 18:39 and automation built into it for the machine learning scientist some on the left here. I'm sure thing some of the key components of 18:49 a platform. So everything from a date of marketplace on a feature Marketplace from where you can collect data during the 18:59 going through the model making process and testing it. Invalidating isn't done deploying us and then measuring performance in production. So you have 19:08

to think of it as a newer model Fleet and how it's running when you build these platforms, you know, 19:18 if you start with one of the cloud platforms where they have strong ml capabilities there still lots of gaps 19:28 in. That you have to fill in to make it work smoothly into an and you can see huge gain in terms of performance, which is shown on the right 19:37 side you move from the red line to the blue or green line. So the number of models that the data science team 19:47

can put into production per month when we've done before and after measurements, we can see approximately 5 X game. So if you have 19:57 20 data scientist and a data platform the twenty-plus platform acts like a team of hundred data 20:06 scientist, and that gives you a significant gain in terms of thinking about the multimodal approach. 20:16 The end of the modeling scaling modeling is not the only issue. So if you think about you've got a thousand models that your daughter exposed at 20:27

apis and you have many places across your work clothes or customer Journeys, where does need to be integrated? If you try to integrate 20:36 those as point-to-point connections very quickly. It's not going to be manageable anymore. So you have to step back and think about how to architect 20:46 that and how to orchestrate between the models and the endpoints from where those are getting a activated. And the last 20:56 flight I have here is talking about model risk governance. So anytime that you're putting into production at to 21:06

make a recommendation to make a decision and so far there's some risk involved and it's because the the model is 21:16 not trained on every possible combination of every possible data type. Among other issues. So so you have to make 21:26 sure that the test at least test for one that works on when it's going to have issues. But what we 21:36 discovered through building a lot of these types of platforms and engines is that the tests will fail 60 or 70% of the 21:45 time. So what do you need to do is start that much earlier than the modeling process soap. So for example of bias every time you test a model for 21:55

bias, there's bias and it's you have to start from scratch and set to go back to the beginning and Lookouts. Is there bias in the data there usually 22:05 is then you go through a data transformation and or data augmentation to remove Tobias and them during the modeling process most deep 22:14 learning is based off of a machine learning is based off of minimizing the error rate on the model, but you have to Then regular ice 22:24 for bias also, so you said it minimize the error rate plus the bias and those two together. We'll keep the bias low and then 22:33

obviously once that's done then you still tested. So let me end with that and close their. Thank you. 22:43

Cackle comments for the website

Buy this talk

Access to the talk “Algorithmic Retail Using AI to Create Advantage in a Digital World - Publicis Sapient”
Available
In cart
Free
Free
Free
Free
Free
Free

Video

Get access to all videos “AI Summit San Francisco 2019”
Available
In cart
Free
Free
Free
Free
Free
Free
Ticket

Buy this video

Video

Access to the talk “Algorithmic Retail Using AI to Create Advantage in a Digital World - Publicis Sapient”
Available
In cart
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
512 conferences
19971 speakers
7276 hours of content