Understanding Search Ads: A Machine Learning Problem

The world of search ads is a fascinating realm where machine learning plays a vital role. To understand why, let's dive into the basics of how search ads work and the challenges that come with them. In the context of search ads, the agency like Google or Bing or Amazon gets paid whenever a customer clicks on an ad. This means they want to predict what is the probability of a click for a given user and for a given search query and for a given ad.

Predicting Probability of Click

Imagine you have ad one, and you're a user with a search query let's call it Q 1. Now, suppose you have ad two with the same user but on a different search query, say Q 2. The agency wants to know which ad is more likely to be clicked by this user given that their query is Q 1 versus Q 2. This is known as predicting the probability of click given a specific user and query, written as P(user ad | query). If we denote the probability of click given the user and query one as P(click|user ad 1, query 1), and the probability of click given the user and query two as P(click|user ad 2, query 2), then we want to determine which one is greater: P(click|user ad 1, query 1) versus P(click|user ad 2, query 2). The decision is based on whether it's more likely that this user will click on ad one rather than ad two.

Conditional Probabilities

The problem can be posed as a standard classification problem. In industry, most companies use techniques like logistic regression or boosting to solve this problem. However, there are also deep learning models that can be used, such as neural networks with multiple layers, which learn complex patterns in data and perform well on tasks like click-through rate prediction.

A Real-World Example: Click-Through Rate Prediction

For those who may know some machine learning, this problem can be framed as a standard classification task. However, for those new to the field, this problem is an excellent way to get started with machine learning using real-world data and examples from the industry. One such example is the Click-Through Rate (CTR) prediction problem hosted on Kaggle. A popular start-up called Vasowise provides historical click data on advertising campaigns.

Solving CTR Prediction

If you know some machine learning, including basics of logistic regression and boosting, then taking up this problem can be a great way to apply your skills to real-world data. The CTR prediction problem is well-documented on Kaggle, with details on the dataset, evaluation metrics, and models that have been tried in the past. By working with this dataset, you can gain hands-on experience applying machine learning techniques to solve a complex real-world problem.

Case Study: Applying Machine Learning to Predict Click Probability

In our own course at Applied AI Academy (AAA), we offer an exhaustive and thorough course on machine learning, including probability and statistics as well as deep learning. In the context of this course, we have a case study where students apply machine learning techniques, including logistic regression, to predict the probability of click in search ads using real-world data. This is just one example of how our course applies theoretical concepts to practical problems.

Getting Started with Machine Learning

For those new to machine learning who want to learn it, there is no better place to start than Applied AI Academy's course on machine learning. Our course covers everything from the basics of programming and Python programming to probability and statistics up to deep learning models. We believe in making the course as applied as possible by introducing real-world examples like search ads. By taking our course, you'll gain a comprehensive understanding of machine learning concepts and how they can be applied to solve complex problems in various fields.

"WEBVTTKind: captionsLanguage: enhi folks on this channel we have we are planning to launch a series of videos it's a web series of videos on this YouTube channel where we'll publish videos on regular basis on the topic of applications of machine learning and AI bicep locations I mean applications of machine learning which are already being used like which are currently these are all current applications of machine learning to various domains from healthcare travel tourism internet search etc also future applications we will also cover some future applications of machine learning in AI which we can see in the next three to five years right so if you are interested to follow applications of machine learning or to understand how machine learning is already being applied or how it could be applied in the near future it will be good for you to subscribe to this channel and follow it as we'll constantly or regularly publish great content and for this video if you could leave a comment on what other topics are what are the applications you want us to cover we will prioritize those applications based on popular demand right the first application in this current video that we'll discuss is on internet right on internet we see lot of ads right so we'll specifically focus on one type of ad called search ads which is literally the bread and butter of companies like Google and Bing again Bing is a search engine by Microsoft for those of you who don't know or who haven't used it much so search ads generate billions of dollars of revenue for companies like Google Bing and it's also estimated that it generates a few billion dollars even for Amazon in one of their recent this is this is again a guesstimate Amazon never publicly announces how much money they make through ads but investors assume that they're generating billions of course Google publicly announces how much money they make through ads now let me show you a few examples so that sort of it clarifies what's happening for you right suppose suppose if I go to Google so here if you notice I'm on google.com in this is the India version of Google search and if I search by medicines online imagine if I'm trying to buy medicines online which is legal in this part of the world if I go and search buy medicines online I get a bunch of ads actually it says clearly here that this is an ad this is an ad by a advertiser or a company called net plus smart right this is also an ad this is also an ad this is also an ad these four are ads actually all these four are actually ads and if you scroll down from here you start seeing the regular results so because it doesn't say add here the results from here on are called organic search results right because these are regular search results and everything above this are ads these ads are often referred to as sponsored sponsored search ads because these are ads which are sponsored by this for example this ad is sponsored by a company called 1mg comm and the way they work is so here let's look at it let's look at what all the players there is Google and there is a customer the customer is looking to buy or the person who is searching here that is me it is planned to buy medicines online so there are some companies like net plus smart like 1mg calm like farm easy . in like smart medics dot co all these companies sell or provide medicines online right now these companies would want a customer like me who is interested to buy medicine online they would attract customers like me to their websites because since i am looking to buy medicines online and since these people these companies are these these these institutions or these agencies are the startups sell medicines online it makes complete sense for them to attract a customer like me to their website because there is a very high chance that will actually use their service or their products right so now what happens is when it's an ad Google gets paid a small amount whenever I click on this suppose if I click on this suppose if I click on this suppose if I click on this and if I go to med plus matcom metal s dot Mentalist matcom pays a small money to Google and if if an ad is at the top of the page these are called top of the page ads right because they're the start most people do not even scroll on Google such they don't they're even blind to the fact that this is an ad they just go and click on it right and Google make some money Google makes billions be billions with a be billions of dollars billions of dollars of revenue every year just through search ads and I know this very well and I've worked on this problem because my first job at Yahoo labs India my first jobs out of out of my college was working on rena who was a big search engine yahoo also had a big search engine around 2008-2009 and then they sold this search search inventory to do microsoft at that point of time so i was working in the very early i was working very early in my career on predicting which ad to show for search results for Yahoo for Yahoo search for countries like Australia and Southeast Asia again it changes by language it changes by there lots there are lots of there lots of details based on language geography etc so my team was working on Southeast Asia and and it's called a pack right asia-pacific regions so I have first-hand experience on how this works and how this can really make billions of dollars to companies having said that the same thing you can see on bing also so if you go to Bing calm so you see the same you see the same thing this is an ad on being this is also an ad on bing this is also an ad in bring this is also an ad in being below that organic results and you see the similar type of companies you see 1mg dot-com advertising here you see or hear or surprisingly Amazon wrote in also advertisers here but I don't think Amazon in India sells medicine song name I think it just because by online for these keywords Amazon got triggered as an ad right so having said this so it's very obvious when Google and Bing do it but let's look at how Amazon in the US so this is amazon.com their Us website right here if I search for vitamin supplements right when I search for vitamin supplements like I might I might be interested in buying some vitamin supplements so one thing you'll notice is you have lot of this search results the first search result it says that this is answered clearly right so when I click on this and go and when I click on this the seller in this case I think it is emergency is the seller who sells this product on Amazon this seller gives a cutter to Amazon whenever I click on it and whenever I click on it whenever I click on it and go to the go to the landing page and purchase this product sometimes it could be based on whether I purchase or not sometimes it could be based on whether I click on this ad or not right so similarly here is another ad because it's a sponsored right again this is this ad is sponsored by M you need three Inc right because these people are selling this product and they want to be at the top of the search results so that their products get sold more similarly now if you look at this this person this ad here this is not an ad sorry this search result here does not have sponsored so this is an organic search result so this is an organic search result right imagine at the scale of a company like Amazon but millions of people visit every day or every week to purchase their items being at the top of the search results makes a huge impact on how much you can sell right so companies like Amazon so companies like Amazon a companies like Amazon Bing which is on the Microsoft and Google actually make billions of dollars through search ads now you might ask since you understand since we understood a little about search ads what are the big machine learning problems there are multiple machine learning problems in the search ad space or the sponsored search ads one of the problems is basically predicting the probability of a click because the the the agency like Google or Bing or Amazon gets paid whenever a customer clicks on an ad right so they want to predict what is the probability of a click for a given user and for a given search query and for a given ad right because when they have these probabilities they can arrange or they can they can sort which ad to show to which user because for a given ad and a query and a user if let's say if let's say let's take an example right suppose if I have ad one right I am a user and I have a search query let's call it Q 1 so same user if I have ad 2 and on the search query if the probability of click the probabilities written as P if the probability of click given the user add one and query one if it is greater than the probability of click given a user add and query right if this probability is more if this probability is more than this probability where if it is more likely that this user will click on this add one rather than on add 2 so it is beneficial for a company like Google or Amazon or Amazon to show the ad for add one rather than add two at the top of the search results so if they have four ads that they can show it's better to put ad 1 in this first slot and add 2 in the second slot because the probability or the chance that ad one will be clicked by this user you given that his given query 1 is higher than ad right so one of the big machine learning problems were to solve is to compute these probabilities is to compute the probability of click given this is written as given this is called conditional probabilities if you're interested to learn more about how these models are built how some things like this code build please check out our course at applied AAA course.com where we have an exhaustive a very very thorough course explaining everything from basics of probability up to state-of-the-art deep learning models having said that now you might ask what is the type of problem for those of you who may know some machine learning this problem can be posed as a standard classification problem and typically in industry most companies use techniques like logistic regression or boost addition trees or gradient booster decision trees to solve this problem having said that now if you want to try if you know some machine learning if you know a little bit of district regulation GBD t if you want to try to solve a problem which is similar to this I would suggest you to go to Kaggle and there's a very nice data set so if you go to Kaggle I'll provide I'll provide you the link for this but you can just simply search for if you want to search for it you can just search for a bazoo ABA z ZD u kaggle if you search for this on google this will be one of the first links that you will get so a vas who is is I believe a small start-up which provides you data which is an advertising company which provides you data on on on past historical clicks here you have to predict something called click-through rate so click-through rate is nothing but it's it's very related to the probability of clipper it's very related to the probability of clicks so if you really know some machine learning and if you know basics of logistic regression and gptt and if you want to solve a problem of search ads you could just take up this problem on goggle there is there's a good data set on this so this is a problem hosted by a vessel it's called the click-through rate prediction go through the description of the problem or how to evaluate it what models have people tried and stuff like that and you have a data set which you can download you can just log into this website download this data set and play with it and understand how to solve or understand or apply techniques that you know like logistic regression to a real-world problem those of you who do not know machine learning and who want to learn it please go to apply to a course comm where we offer to the best of our knowledge when the best courses without any prerequisites where we cover everything from basics of programming basics of Python programming probability statistics up to deep learning actually in our own course we solve this problem of predicting the probability of click on search ads using a value data as a case study we do this as a case study at apply da course for our AI / ml course because our course is designed as the name suggests it's very very applied in nature and we envy in addition to teaching all the core theoretical concepts we make the course very applied by introducing lot of case studies on real-world data setshi folks on this channel we have we are planning to launch a series of videos it's a web series of videos on this YouTube channel where we'll publish videos on regular basis on the topic of applications of machine learning and AI bicep locations I mean applications of machine learning which are already being used like which are currently these are all current applications of machine learning to various domains from healthcare travel tourism internet search etc also future applications we will also cover some future applications of machine learning in AI which we can see in the next three to five years right so if you are interested to follow applications of machine learning or to understand how machine learning is already being applied or how it could be applied in the near future it will be good for you to subscribe to this channel and follow it as we'll constantly or regularly publish great content and for this video if you could leave a comment on what other topics are what are the applications you want us to cover we will prioritize those applications based on popular demand right the first application in this current video that we'll discuss is on internet right on internet we see lot of ads right so we'll specifically focus on one type of ad called search ads which is literally the bread and butter of companies like Google and Bing again Bing is a search engine by Microsoft for those of you who don't know or who haven't used it much so search ads generate billions of dollars of revenue for companies like Google Bing and it's also estimated that it generates a few billion dollars even for Amazon in one of their recent this is this is again a guesstimate Amazon never publicly announces how much money they make through ads but investors assume that they're generating billions of course Google publicly announces how much money they make through ads now let me show you a few examples so that sort of it clarifies what's happening for you right suppose suppose if I go to Google so here if you notice I'm on google.com in this is the India version of Google search and if I search by medicines online imagine if I'm trying to buy medicines online which is legal in this part of the world if I go and search buy medicines online I get a bunch of ads actually it says clearly here that this is an ad this is an ad by a advertiser or a company called net plus smart right this is also an ad this is also an ad this is also an ad these four are ads actually all these four are actually ads and if you scroll down from here you start seeing the regular results so because it doesn't say add here the results from here on are called organic search results right because these are regular search results and everything above this are ads these ads are often referred to as sponsored sponsored search ads because these are ads which are sponsored by this for example this ad is sponsored by a company called 1mg comm and the way they work is so here let's look at it let's look at what all the players there is Google and there is a customer the customer is looking to buy or the person who is searching here that is me it is planned to buy medicines online so there are some companies like net plus smart like 1mg calm like farm easy . in like smart medics dot co all these companies sell or provide medicines online right now these companies would want a customer like me who is interested to buy medicine online they would attract customers like me to their websites because since i am looking to buy medicines online and since these people these companies are these these these institutions or these agencies are the startups sell medicines online it makes complete sense for them to attract a customer like me to their website because there is a very high chance that will actually use their service or their products right so now what happens is when it's an ad Google gets paid a small amount whenever I click on this suppose if I click on this suppose if I click on this suppose if I click on this and if I go to med plus matcom metal s dot Mentalist matcom pays a small money to Google and if if an ad is at the top of the page these are called top of the page ads right because they're the start most people do not even scroll on Google such they don't they're even blind to the fact that this is an ad they just go and click on it right and Google make some money Google makes billions be billions with a be billions of dollars billions of dollars of revenue every year just through search ads and I know this very well and I've worked on this problem because my first job at Yahoo labs India my first jobs out of out of my college was working on rena who was a big search engine yahoo also had a big search engine around 2008-2009 and then they sold this search search inventory to do microsoft at that point of time so i was working in the very early i was working very early in my career on predicting which ad to show for search results for Yahoo for Yahoo search for countries like Australia and Southeast Asia again it changes by language it changes by there lots there are lots of there lots of details based on language geography etc so my team was working on Southeast Asia and and it's called a pack right asia-pacific regions so I have first-hand experience on how this works and how this can really make billions of dollars to companies having said that the same thing you can see on bing also so if you go to Bing calm so you see the same you see the same thing this is an ad on being this is also an ad on bing this is also an ad in bring this is also an ad in being below that organic results and you see the similar type of companies you see 1mg dot-com advertising here you see or hear or surprisingly Amazon wrote in also advertisers here but I don't think Amazon in India sells medicine song name I think it just because by online for these keywords Amazon got triggered as an ad right so having said this so it's very obvious when Google and Bing do it but let's look at how Amazon in the US so this is amazon.com their Us website right here if I search for vitamin supplements right when I search for vitamin supplements like I might I might be interested in buying some vitamin supplements so one thing you'll notice is you have lot of this search results the first search result it says that this is answered clearly right so when I click on this and go and when I click on this the seller in this case I think it is emergency is the seller who sells this product on Amazon this seller gives a cutter to Amazon whenever I click on it and whenever I click on it whenever I click on it and go to the go to the landing page and purchase this product sometimes it could be based on whether I purchase or not sometimes it could be based on whether I click on this ad or not right so similarly here is another ad because it's a sponsored right again this is this ad is sponsored by M you need three Inc right because these people are selling this product and they want to be at the top of the search results so that their products get sold more similarly now if you look at this this person this ad here this is not an ad sorry this search result here does not have sponsored so this is an organic search result so this is an organic search result right imagine at the scale of a company like Amazon but millions of people visit every day or every week to purchase their items being at the top of the search results makes a huge impact on how much you can sell right so companies like Amazon so companies like Amazon a companies like Amazon Bing which is on the Microsoft and Google actually make billions of dollars through search ads now you might ask since you understand since we understood a little about search ads what are the big machine learning problems there are multiple machine learning problems in the search ad space or the sponsored search ads one of the problems is basically predicting the probability of a click because the the the agency like Google or Bing or Amazon gets paid whenever a customer clicks on an ad right so they want to predict what is the probability of a click for a given user and for a given search query and for a given ad right because when they have these probabilities they can arrange or they can they can sort which ad to show to which user because for a given ad and a query and a user if let's say if let's say let's take an example right suppose if I have ad one right I am a user and I have a search query let's call it Q 1 so same user if I have ad 2 and on the search query if the probability of click the probabilities written as P if the probability of click given the user add one and query one if it is greater than the probability of click given a user add and query right if this probability is more if this probability is more than this probability where if it is more likely that this user will click on this add one rather than on add 2 so it is beneficial for a company like Google or Amazon or Amazon to show the ad for add one rather than add two at the top of the search results so if they have four ads that they can show it's better to put ad 1 in this first slot and add 2 in the second slot because the probability or the chance that ad one will be clicked by this user you given that his given query 1 is higher than ad right so one of the big machine learning problems were to solve is to compute these probabilities is to compute the probability of click given this is written as given this is called conditional probabilities if you're interested to learn more about how these models are built how some things like this code build please check out our course at applied AAA course.com where we have an exhaustive a very very thorough course explaining everything from basics of probability up to state-of-the-art deep learning models having said that now you might ask what is the type of problem for those of you who may know some machine learning this problem can be posed as a standard classification problem and typically in industry most companies use techniques like logistic regression or boost addition trees or gradient booster decision trees to solve this problem having said that now if you want to try if you know some machine learning if you know a little bit of district regulation GBD t if you want to try to solve a problem which is similar to this I would suggest you to go to Kaggle and there's a very nice data set so if you go to Kaggle I'll provide I'll provide you the link for this but you can just simply search for if you want to search for it you can just search for a bazoo ABA z ZD u kaggle if you search for this on google this will be one of the first links that you will get so a vas who is is I believe a small start-up which provides you data which is an advertising company which provides you data on on on past historical clicks here you have to predict something called click-through rate so click-through rate is nothing but it's it's very related to the probability of clipper it's very related to the probability of clicks so if you really know some machine learning and if you know basics of logistic regression and gptt and if you want to solve a problem of search ads you could just take up this problem on goggle there is there's a good data set on this so this is a problem hosted by a vessel it's called the click-through rate prediction go through the description of the problem or how to evaluate it what models have people tried and stuff like that and you have a data set which you can download you can just log into this website download this data set and play with it and understand how to solve or understand or apply techniques that you know like logistic regression to a real-world problem those of you who do not know machine learning and who want to learn it please go to apply to a course comm where we offer to the best of our knowledge when the best courses without any prerequisites where we cover everything from basics of programming basics of Python programming probability statistics up to deep learning actually in our own course we solve this problem of predicting the probability of click on search ads using a value data as a case study we do this as a case study at apply da course for our AI / ml course because our course is designed as the name suggests it's very very applied in nature and we envy in addition to teaching all the core theoretical concepts we make the course very applied by introducing lot of case studies on real-world data sets\n"