Some thoughts on the job-hopping season
Here is what you saw on Zhihu:
Every February and March, around the Spring Festival, is the time when companies make money. But there is a kind of year-end bonus, called "the year-end bonus of other people's family."
You may have seen such a news a year ago: The game company Lilith has given out a bonus of 190 million for the 90 members of the "Sword and Crusade" project team.
At that time, many people thought that this was a joke, but unexpectedly, it was true.
The Paper reported that the dividend of 190 million yuan was true, except that there were fewer than 90 people who divided the money, which was divided according to the proportion of "shares." This means that each of them can get more than 2.11 million rewards.
In addition to the sub-infrared of the "Sword and Expedition" project team, Lilith also has annual conference prizes such as Sony TV, PS5, iPhone 12 Pro Max and iPad Air...
The year-end awards of other people's homes are so unpretentious and boring. Ordinary workers, working hard for a year, have to endure the company's postponement of the year-end bonus until March.
Even if you finally receive the money, when you finish reading the amount of the bank card, you may only have one sentence left to say:
"Sorry, I want to quit."
In the golden three silver four job hunting season, are you feeling restless and looking forward to changing jobs and filling up your bank card balance?
Today, I will take you from the overall picture to give you an overview of what is going on with advertising on the Internet and how the advertising system works. As an advertising system, what is its goal? Know that these may be useful when you are facing a job search or job-hopping.
In the personal window of the blogger at the end of the article, there is the book "Market and Technology of Computing Advertising Internet Commercial Monetization", which contains a detailed systemic description of the Internet advertising system, and small partners who need it can purchase it by themselves.
Here to feed yourself a bag of salt, visualization (tableau) and graduation design (matlab) partners, can subscribe to the following two columns carefully organized by the blogger.
Tableau Visual Data Analysis Advanced Tutorial
https://blog.csdn.net/wenyusuran/category_9596753.html
MATLAB in-depth understanding of advanced tutorial (with source code)
https://blog.csdn.net/wenyusuran/category_2239265.html
There are also application example source codes of various algorithms in the blogger's resources, and small partners who need it can pick it up.
Advertising and bidding
You may more or less feel that advertising is very important to Internet companies, and it is also the profit model of many Internet companies. Because most Internet companies don't directly make money from users , but the platform needs to make money , so advertising is a very good model. In fact, good advertising wins more. Merchants and platforms make money, and users buy products they like. Users who have not spent money also enjoy the services of the platform, and because of this, advertising accounts for a very large proportion of Internet companies' revenue, even more than half.
As users, we browse advertisements when surfing the Internet. These advertisements are not fixed or random. They are also retrieved from the ad library through similar search engines. In order to maximize the benefits of the platform, advertising will be conducted in the form of bidding. For a simple example, for example, when we open a web page, there is only one advertisement field, but there are 3 candidate advertisements, such as shampoo, car, and medical beauty. There are three advertisement candidates, which one should I choose? It's very simple, of course, it is whoever bids high for the platform.
In other words, within the advertising platform, each exposure of a display advertisement is actually equivalent to an auction . Advertisers will give their own ads a bid, the logic of this bid is different in different scenarios. The best thing is to be able to dynamically bid based on the advertiser’s current budget, the amount of traffic already obtained, and the expected traffic, but this is more difficult to achieve, and not all systems have this capability. The simpler is a fixed bid. We don't need to go into details here, we just need to understand it roughly.
Let's assume that all advertisements have fixed bids, such as 3 yuan for shampoo, 10 yuan for cars, and 50 yuan for medical beauty. Then obviously it must be the victory of medical beauty, so this advertisement will show medical beauty. But there are two problems here, let's talk about them one by one.
The first question is: What is the world of high bidders?
The profit of the medical beauty industry is relatively large, so the advertising bids are high, so that it is not possible to display shampoos and cars? This is not entirely true, because the auction is conducted multiple times, based on the recalled candidate advertisements. Although shampoo bids are low, it does not necessarily compete with medical beauty advertisements every time. So it is also possible to be displayed.
The second question is: What if the user community is not interested?
The user may be a male, who is not interested in medical beauty, and he will not consume it no matter how you make it. Wasn’t the money for such advertisements wasted for nothing? So here is the issue of the choice of advertisements . When we choose advertisements , we can’t just choose based on bids, because even if users do not order, the platform will not make money . How can the platform make more money? Of course, we must expect the greatest. What is the expectation? Equal to bid x click-through rate, so the core is click-through rate, with click-through rate we can estimate its expectations. Where does the click-through rate come from? Of course, it is predicted by the model.
So this thread is stringed together. Why does the advertising scene need click-through rate? Because you need to calculate the expected revenue that advertising brings to the platform.
Second bidding strategy
There is also a very interesting strategy in the advertisement, called the second bidding strategy . The meaning of this strategy is also very simple. Let's use the example of shampoo, car and medical beauty advertisements. The shampoo cost 3 yuan, the car cost 10 yuan, and the medical beauty cost 50. Obviously we are inverted, and the medical beauty is the most, followed by the car and finally the shampoo. We do not consider the impact of click-through rate on the advertisement, and first consider the bidding alone. This advertisement should show medical beauty.
General bidding is the first bidding strategy, which means that the highest bidder gets the price. This is very easy to understand, and it is the same in daily auctions. But in the second bidding strategy, the highest bidder gets the same, but the person with the highest bid does not need to pay his own bid. He only needs to pay the second highest price plus a base figure , such as 1 cent or 1 cent. Money (the base depends on the specific scenario). This is called the second bidding strategy. In the example just now, although the medical beauty advertisement called 50, it only cost 10.1 yuan.
But it seems that we will definitely feel wrong. Isn't this less money? Is the platform silly? Obviously, you can collect more money and have to make less?
In fact, the platform is not stupid at all, and the one you buy is never the one you sell. The reason here is very simple. Let’s think about it. Suppose you are an advertiser of medical aesthetics and bid for an advertisement at a price of 50 yuan. Will you still bid 50 in the next bid? Obviously not, the reason is very simple. You competed for the advertisement with the price of 50. Although you don't know how much other people paid, it must be lower than you. Then you will definitely try to lower the price next time. Can you use less money to get ads? In fact, all advertisers will do this. If everyone lowers the price together, is this auction still called an auction? How can the platform make money?
In order to prevent people from lowering prices, the platform adopted a second bidding strategy. In this way, even if advertisers take the advertisement, they cannot keep the price down . Because the second advertiser’s price also fluctuates, it may also increase the price. And the person with the highest bid does not have the need to lower the price, because even if the price is lowered, the price he gets for the advertisement is still the second-place price, and there is no benefit to him. In this way, although it seems that the platform has not maximized the benefits, it can actually avoid the situation where everyone is squeezing prices together . This kind of profit will be higher.
At present, basically mainstream algorithm platforms adopt such strategies, such as certain search advertising, Taobao product advertising and so on.
Advertising logic
For advertising in the Internet age, it is very different from traditional advertising. For example, traditional flyers or TV advertisements, sign advertisements, etc., these advertisements are fixed, and there is no user-customized function. No matter who gets the flyer, the content is the same. It's impossible to say that different people get different results when they get the flyer. Therefore, traditional advertisements are all made by Haitou, the conversion rate is very low, and it is often difficult to count conversions . For example, we put up an advertising sign in a certain place to sell a certain household appliance. It is difficult for us to count how much sales of this household appliance are contributed by this advertisement.
But advertising in the Internet industry is different. Since all our content is electronic, it can make great progress. The first one is personalization , which can push advertisements to us according to our needs, instead of randomly pushing or pushing according to the needs of advertisers. The second is that it can count conversions and subsequent effects very well , because all the information in the Internet is connected in series, and it can be tracked from users' views, clicks to purchases. We can also train advanced models to predict user preferences, so as to screen out high-quality advertisements for users and earn benefits for the platform.
Speaking of which, just to mention briefly, why there are so many e-commerce companies in China, and only Taobao is making money, and it makes a lot of money? Because Taobao has always been a platform, what it plays is traffic, and what it earns is traffic money . Taobao buys traffic from other platforms and then sells these user traffic to merchants. What we buy on Taobao is not Taobao's own goods, but the goods of the merchants in it. To sell goods, merchants need to publish advertisements on the platform to obtain traffic. The money for selling traffic is the source of income for Taobao, not the money for buying things. It is not easy to find Taobao for buying things, and Taobao is never favoritism, because Taobao does not make money by selling goods, and users are the source of traffic, so Taobao will never favor merchants and can basically do notarization. For example, a certain Dong has always been focusing on self-employment, which is to make money by selling goods. In the first two years, there have been problems such as adulteration and refurbished machines. This series of problems will continue in the future, because in essence, certain Dongs rely on selling goods. To make money, it has to compete with Taobao at low prices. In order to make money, it is bound to be difficult to maintain high quality.
As we said earlier, in order to maximize benefits, it is necessary to put in the products with the greatest expectation of benefits. We all know that expectation is equal to probability multiplied by value. We don't know the probability here, and no one can accurately evaluate it. We can only use some machine learning or deep learning models to predict the click-through rate as closely as possible . The click-through rate is the number of clicks divided by the number of exposures. The English is Click Through Ratio, or CTR for short. So we can write the platform's revenue as a formula:
The R here refers to Revenue, which is the return of income, and bid is the merchant’s bid. This formula has a premise, that is, ads are deducted based on clicks . However, the advertisements in the platform may not be charged according to clicks. For example, some advertisements in good positions may be charged according to exposure (viewing). For example, the location of Taobao homepage:
We see that there are also advertisements in the top section of the homepage. These advertisements are in the best position and are must-see content for users. These high-value locations may not necessarily be charged based on clicks, but based on exposure. What is the advantage of this? In addition to making more money, it can also force these advertisers to improve the quality of their ads . This is the same reason that the advertisements in YouTube can be skipped. It is precisely because the advertisements can be skipped that advertisers are forced to improve the creativity and quality of the advertisements, so that the audience is willing to watch it through.
When we search on Taobao, there are advertisements in the search results. For example, when I searched for a mobile phone, the first one was an advertisement. The advertisements in the search results do not have their own unique copywriting or special pictures. Generally speaking, they are charged according to the click.
There are also some ads in hidden corners, which are in a worse position, and may be charged according to the conversion, that is, the user's purchase. The platform can make money only if the user does purchase the product. Such as some Moments of Friends, website small advertisements are generally like this, because the conversion rate is not high, if you charge according to exposure or click, advertisers are not happy.
For advertisements that are charged according to transaction, we only predict the CTR is not allowed. Because some products may have a high click-through rate, but the conversion rate is not necessarily high . Let me give you an example. Because I bought clothes for my wife, Taobao will push me some sexy ladies. The click-through rate of these obviously eye-catching pictures or titles is often very high, because people are visual animals and may be tempted to click them. Anyone who has ever advertised knows that the click-through rate of these beautiful pictures is very high. In this case, we ca n’t just estimate the click-through rate, but also estimate its conversion rate .
The scientific name of the conversion rate is CVR (Conversion Rate), which is the probability of placing an order for each click. The advertising revenue at this time is:
That is, the probability that the user clicks is multiplied by the probability that the user will buy after clicking, which is equal to the probability that the user may buy when they see it. For the advertising team, in addition to the need to estimate CTR, there is also an additional task of predicting CVR .
to sum up
Let's briefly summarize that the logic of the advertising system is similar to the recommendation system, except that the target is different from the recommendation and search. Because the target involves calculating revenue, there will be an operation that is multiplied by the bid to calculate the expectation. To calculate accurate expectations, two core values are required, one is CTR and the other is CVR. Relatively speaking, CTR is more widely used, and the estimation logic of CVR and CTR is basically the same. So like recommendations, the core of the advertising team is also CTR estimation. As an algorithm engineer, the daily work is actually very, very close.
One-click triple connection for your favorite friends!