Abraham is currently a Senior Director of Applied Research at eBay. He leads the Search Ranking and Monetization team, bringing together product search and advertising. Prior to eBay he was doing computational advertising at Amazon. Abraham has authored over a dozen papers in computational advertising and co-chairs the AdKDD working a top workshop in the field. He is also a part-time faculty member at Northeastern University in Silicon Valley.View the profile
About the talk
Digital advertising is estimated at $87B in 2020 in the US alone. Ubiquitous and highly visible, advertising interacts with nearly everyone. From selecting ads to calculating costs to deciding where and how to display them, billions of computational decisions are made every day in this challenging field. In this talk, we will discuss some of the terms and solutions for computational advertising.
It's good to be here. Thank you again for inviting me to give this talk. I'll be talking about various problems in artificial intelligence and machine learning for computational advertising will give an overview of computational advertising is and we're different learning. So first of all, what is advertising undoubtedly some of you have seen advertising and maybe you'll like it if you don't have a lot of work goes into it and I will try to open that
open up the hood and try to understand what's going on. Digital advertising for any form of advertising really focuses on as far as her problems giving the right options like information or opportunity to get someone to do something to the right person. Hopefully you the right contacts when when you're interested or in the market for that particular option and at the right time getting all of these things, right? I would be ideal for advertising and that's what advertising is trying to do
different terms that you may have heard about probably the best-known are real-time bidding or at exchanges that you may have heard their auctions going on and in advertising targeting privacy and then such things, they're all they're all really important to understand a few of them. today so as as in machine learning problem, we can view advertising as solving an optimization problem. So as a relation problem, it is strictly simple to write we want to find the ad that you ignored
by a witch from the set of all possible as a region show to a person maximizes some sort of the volume a hundred billion times a day and less than a hundred milliseconds is in very frequently salt optimization problem. And what's interesting about it is despite. The fact is relatively simple to write down a lot of different things go into it. These are the different types of learning algorithms and Data Systems, the disciplines that go into advertising we have
not just what happened in the past, but also it was likely to come in the future how much add traffic or user interest. Will there be in a particular product that helps us calculate the value that we talked about and also helps us with advertisers to tell them your ad is going to be very valuable in the future because a lot of people are next month going to look at our content because it's Christmas time. Classification algorithms also help determine that the valuation add an instantaneous moment. So to know how likely a person is to purchase a product
or to respond to an advertisement that the advertiser in Posh formal take then there are regression problems and I'll tell you pull in in particular. Is it really what the advertisers States as their stated by you or is it something else next mechanism design there are multiple parties in advertising. There's the advertiser in advertising. Platform or a publisher you want there to be many more hats. So each of them has a different value. How do we are betrayed between these values of how
do we decide the best allocation for everybody mechanism design once advertisers work within their business if they're thinking about how much money can I spend how do I allocate advertising budgets across different campaigns marketing initiatives and then also non-convex optimization problems, but is the convex lens. Sorry to hear what's the time? So if if we didn't have to focus in on the purpose of advertising we will see that there is some important components one is that there is a call to action
which is somebody's asking you the receiver of the add to do something I eat. It could be by a product to be look at content that you're just being watched this video. But there's generally an action that's requested and that is a paid action which means that the platform on which are viewing is not the only stakeholder and getting you to complete the action to pay the platform to give get you to complete the action. Yeah, that doesn't have to sound mysterious
advertising your friend Ryan recent times when some digital ad spend went down there still growing and various other parts. And You're my Wonderwall. How is it that advertising be such a big business and going in in a people find out a little bit annoying. Why do people keep doing it as a simple answer it works and it works in ways that are not always obvious because people as they seek to buy things and are are looking for different content really know
exactly what it is that they want all the time. I think you do but oftentimes you're at your decision to buy can be influenced by content. You see or videos you've watched how many of you have watched a video about a particular hobby like Woodworking and all the sudden you find yourself wanting to buy a saw that's sort of advertising and it sort of visit. Like the advertising does have a part to play movies in particular on Market. There's no good way to find out about movies people don't just walk in front of the movie theater every day. So
you hear about them may be heard from your friend holiday hear about it. Cars phones the big things that people buy in life are often required either directly or indirectly through advertising from the perspective of a statistics for mathematics. We can see that advertising lowers the cost of discovering a product or service right you could go and look at every single phone that that exist in the world and make it clear determination or like technical journals about all these different phones and make their decision or
you can see an ad story great way for you to find out about something. So that lowering of the cost of discovering and other products that I've sort of the key motivation advertising from them for each perspective user consumer. There are two general classes of advertising. There is the offline case which up until recently was the only way to do advertising in which somebody puts up a sign or an indication somewhere. We called that an impression you you see this billboard and it's a little bit of an old but you probably see
lots of signs like this and at some point in the future people by things did they convert 2 from a passive Observer to a customer they call that a conversion process and most the time. It's a purchase by sometimes it's something else. Now the challenge with offline advertising is that there is no direct relationship between these two things. How is it that we know that buy the phone wasn't influenced by the sign you put up the sign people buy iPhones. How do you know that the sign made them by the phone? Well, it's
very difficult to do and it's not obvious answer people spend a lot of time building models that affect the part. Now in digital advertising we don't have as much of a problem about this. We didn't send feedback we can track from time the person sees the ad they purchase this is really helpful for advertisers to give some great information. And it also subsidizes most of the free content that that they see and it allows us to be highly personalized we can take into
account that the desires or the interest of consumer mostly for their benefit to the Show Samantha that makes sense to them and therefore reducing the cost to advertisers. Sophie conversion funnel as it's called in digital advertising introduces a lot of extra events that we can measure Amtrak between the impression and the purchase so you can add your microphone be on your go-to. It's called the landing page, which is the place that you go after coming out of the advertisers opportunity to really
sort of kid shoes if their particular product or service. Then you might go to a purchase. You might fill out a form you might be something else if you favor and all these different arrows that you see here are trackable. Thanks to the benefits of digital advertising used to identify the impression which led to the purchase and then assign credit backwards. That it was an armed president Atul in the history of advertising that there may digital advertising so attractive. Now there are many different forms that advertising takes that that we can
see the first is sponsored search is probably the most people are experiencing today. I was just you type in a search and into your favorite search engine and adds up here and these ads usually have something to do with the query is sometimes they don't click on them. They go to the advertisers page that's called The Landing and if that works out for you, you'll convert drive you towards that conversion. So there is an explicit intent you type in a fiery and
that kind of focus has the advertiser on trying to drive you towards the conversion and the conversion van depends on the advertiser someone to sign up for a credit card. So I want you to buy a specific. Another very common advertising for Ms. Display advertising. So these are pictures typically or videos that you see on various. You're not searching for something specifically and to see the under you might watch a video and I know it appears your might go to website and see a banner ad
in any of these cases be intent is not obvious. You're watching a video about something or you're looking for an article. You're not interested in buying something. So then the ad has to really draw your attention hopefully in a positive way and assume or inferred that you might be interested for a particular product or service. It's the same process that exists in sponsored search, but it's it's different because we attempt. And another reason forms of e-commerce advertising most notable is the Amazon sponsored
products in which you are already very close to buying something but we might want you to buy something slightly different so you can search for Saya teddy bear and maybe there's like a hundred million teddy bears on Amazon or Ebay and you want to find a good one and it kind of gives you some strong confident you click on and there's a direct and straightforward process so that the purchase funnel is a little bit different in an emphasis that is generally the same
pattern. So to understand more about all of these things fit together, let's take a look at the different optimization problems that each of our components have the user has as love to give her money to spend a day. They love their certain content and they want to spend their money on things that will bring die and sterilize a romantic view of this user casting itself out into the vast internet. The Advertiser has things that they think that people would like to buy it's only one problem. They're sitting off someplace in the insurance and they cannot find their happy customers. They're not
they're not coming to their site in our coins their store. They need to find them somehow and show them to their products are really very nice. And then we have the publisher for sure is where the content is. The publisher is the customers are stopping so did publisher has the customers but most of the time the publisher is not charging the customer to to participate in their continents will giving it away for free. So the user is providing their love to the publisher. They love the content. They're going to keep on watching it. Otherwise, they wouldn't be watching it.
The so how does how do I connect these two sides? There is something called a platform. The demand-side platform connecting find users who are interested to them and they kind of give a specification. It's called an ad campaign that says you're kind of customer that I think would make sense of my product. But I'd like to show them how much I like to pay in order for that to happen has the job of finding the customers. How do they do this? They interact with there are probably less than a hundred at exchanges in the world, maybe a thousand
gsp's and millions of advertisers. So the exchanges take the job of connecting to each of millions of different Publishers and by pulling the connection information, they form a task to love of Paws. All users as they're attracting with content and its eldest to to the advertisers. So in the average data exchanges allow are the dsp's to buy impression opportunities to say, okay, here's a stop on this page this particular side connection works, but still
how does the publisher make money today? Just put money into the publisher. No user brings the money back to the advertiser. So this makes us a virtuous cycle because the publisher is getting money in from the yard exchanges advertisers are only putting their their money into this system when there's a return on investment. So if the user the customer is buying the product from The Advertiser, then the advertiser will put more money and continue to love the
content. They're going to keep on interacting with it and find products that they loved it. So it's nice happy virtuous cycle. So probably sure then is incentivize to provide strong enriching content to our users and in by process getting some money in exchange. So how do all these different systems work? There's a lot of data processing change or DSP. You have to interact with Publishers and how does it all work? Well, it all comes together serving components. That is the dsp's
interaction with the exchange and a very similar process happens with the exchange in order to in an hour. Dial Phil time server needs several pieces of information in particular needs a predictive model, which is a bunch of data in real-time data attend Publishers interactions with each one of these interactions generates a long event, which set of data per day into a lock system and then a model training process transform those logs and signals into both feature look up set. Go back to the real time in service and also updated model.
And this is the basic feedback loop for our machine learning in advertising systems. So how do all these things start to work together? The publisher then has a learning problem which is to to lay out the specific content on the page and the ants that make sense to the different Advertiser they decide when and where and how to allocate various problems with one of them is layout optimization, which sword is decides. Well, should I put five ads at the top or four? Should I put a component over here or not? Each of them makes a different value to the advertisers and drives continued love and
support from the customers and it needs to balance those two things. In yelled optimization we're interested in just the answer. Okay. This is the space you have for ads. How many ads should you show at a particular Moment In Time? Next is forecasting. So advertisers want to buy specific has appointed time and I want to know how much it's going to cost. So often provide forecast. How many users are going to be on the side, but will those users be wanting these time series models
minimum value that's required by odds of the site so that they don't sell content at a very low price and risk-reducing the engagement from the users. On the allocation side most Publishers integrate with multiple different at exchanges and the most common approach is called a water flow a waterfall talk to a first exchange Forex a high-value, but low quantity of ads these would get first preference and then goes on to the next one and the next one isn't as efficient as running in an auction for example, but it has the benefit of providing strong reducing business risk,
so that it works on high quality different ads in the Publishers can plan around. So the publisher then has this nice virtuous cycle, right? It has it gets money for for users through advertising. It channels that money back into building at site better and also driving traffic and perhaps even running ads on its own. And as long as they're the rate at which it makes money for users is greater than the rate at which expensive it becomes a profitable business The
Advertiser then has to invest in in its users and it hopes for a return on investment. So ideally it expects more money for the products than it spends on Advertising. Are the advertisers and choir traffic to their content or to their products in a variety of ways, they spend money to use their products, which is very efficient, or they can give you know, maybe some free money to to loyal customers or they can provide ads which will intensify to buy. So that the key question that an
Advertiser has to answer is what is their marketing hypothesis? Who are your customer? How do you find them? And what do you want them to do? And the customer development process is where advertisers kinds of think through what makes their product. There's lots of ways you can Define who are your customers and where do you find them? And those are the different types of advertising that you might choose to use? As a business grows, you get better and better at refining your marketing hypothesis.
You'll see a lot of start of snooping around free t-shirts and talking a lot about their app, but you see very large Brands focusing on brand advertising because they have very different view on where their customers are an extra ticket. Demand-side platform and is where all of these things get put together the ad exchanges where the traffic comes in there is a float called bidding for the advertisers need to show ads to to customers is transformed into a value on the Prairie impression basis
then, and if it's a slow process of the PSP sort of his intersecting this floor, that's where he can drive now. I am I being smart to figure out where good and good diets avocados come from The key component that at the Espy offers advertisers is targeted. It learns where the users are likely to be and how you are to find them. And I think of targeting as a production problem find the ads which have a highly likely to find users that are of interest and try not to
spend so much money by limiting your Beach traffic users that makes the most sense for your head product. Inside the bidding system the formulate that value is a response production problem. We want to predict whether someone will respond to an ad given that were available from logs are other data set that we have most of research. Once you have a response prediction model you put that into a bidding process by sinking of the expected value. That is all the different outcomes in
that conversion funnel. Each of them has a value at The Advertiser and possibly other entities and we need to tell you put together all the different components in the advertising process. You have the targeting that forms that the set of possible as you have the bids and response prediction value and the Ants election which comes through the ESPYs and exchanges drive. He adds to the customers of Interest. And how does the scale we have hundreds of billions of adoptions per day that is more than all stock transactions on all exchanges
everyday and this happens within a hundred milliseconds from the time that use it looked at the content to the time that they see an AT-AT General traffic time. We're processing 400,000 request for second a typical at exchange that's more than all the credit card transactions in the whole world and you know this about a hundred terabytes of data that being generated on a daily basis just to power these models and there's about a hundred different players in different companies that that may participate in a single request the remote calls. So it's it's a very challenging as
well. Listen to challenge for our big data and machine learning because of the rate of very few people click on ads 2.1% It is equipped for Impressions after a quick there's about a 1% chance that people will buy something. So if you distill that you have a big data problem is you have very few positive examples of how do you get all those different things to work with in the latest and they're still quite a few on the solve problems and advertising today. How do we
balance Ali Zafar request? How do we do targeting and protect privacy? How do we predict whether people are going to respond in town for that about you so that people get higher-quality yet? That's that's it. That's how we put it all together all the different components within the system.
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.