FMI pegasus network
Eric Lu, VICE President of Technology at AdMaster: Digital marketing practices in the field of ARTIFICIAL intelligence
When it comes to digital marketing, I believe many people are very familiar with it, but most of the digital marketing we have come into contact with in the past is the application in the field of big data. In the age of AI, digital marketing is also eye-catching. Today, we are going to talk about the practice of digital marketing in the field of artificial intelligence introduced by Lu Yi Lei, vice president of AdMaster technology, at the FMI Artificial Intelligence Summit of Pegasus!
Lu Yilei, Vice President of AdMaster Technology, senior expert of big data, member of CCF Big Data Committee. Deep understanding and practical experience in distributed storage and distributed computing, super cluster, big data analysis, etc. More than 10 years of experience in cloud computing, cloud storage and big data.
“Artificial intelligence is hot right now, but it’s easy to misunderstand the concept of intelligence,” Lew said. Lew told us that ai, as we’ve developed it so far, has two big chunks: the first is symbolic intelligence, traditional AI. The second is computational intelligence. In fact, this is what we practice most at present, which is the point of our fastest development.
The basic technologies of artificial intelligence can be divided into five parts:
* First, a large knowledge base. In fact, most of the time, if you want to do artificial intelligence applications, it is difficult to achieve without a database.
* The second search technology, no matter foreign Google, or China’s Baidu, it raises more, it naturally applied search engine to the inside of artificial intelligence.
* The third reasoning technique, personal refutation, is deductive.
* The fourth is knowledge acquisition technology, which is also an important direction in the future. If you don’t have any way to acquire knowledge, the development of artificial intelligence is still limited.
* Fifth application scenario.
At present, we focus on the direction of artificial intelligence has two: the first, is the understanding of natural language, this is the direction of the future, now see a lot of such, robots, it is the understanding of language. Second, computer vision, Lu Yi Lei told us that computer vision is a point he values very much, at present, a lot of robot image recognition technology is still relatively weak, if we can make this breakthrough, it will be very promising.
Luigley said that without data, other developments are slow, such as machine intelligence, data intelligence, computational intelligence, and finally brain-like intelligence, which is the future.
Next, Lew introduces six algorithmic models and their applications in the digital marketing industry.
* The first one is SVM, which mainly evaluates the general analysis, judging your gender and age, and whether your advertisement is safe, for example, whether it is placed on some pornographic websites.
* The second part is natural language processing. We will judge these emotions and determine whether the sentence is neutral, positive or negative for readers and writers. We will do an in-depth analysis. We’ll make a category.
* The third piece of cluster analysis, we make cluster analysis of the existing crowd.
* The fourth regression analysis, CTR, and comprehensive effect of historical data, which platforms and which media have the best effect, we will do in-depth analysis.
* The fifth block, GBDT, is mainly used to judge whether the device ID belongs to the same person, and another block is used for CTR estimation.
* the sixth time series analysis, a lot of people to do good, in laboratory effect is very good, but one to the actual environment effect is not good, he is easy to forget the time sequence, we have put this man tell his gender, age, including personal interest in judgment is very accurate, but into effect is the opposite, actually he is not considering the time factor.
In addition, in terms of deep learning, Lew told us based on Hidden Nodes, it’s not that the more Hidden nodes you have, the better. This is where you have to make some trade-offs. Because the more layers you do, the more you end up with, the more you end up with. So we can adjust this parameter in other ways.
TensorFlow is based on Spark, and you actually changed the code no more than 10 lines. The first one, the new version is called RDMA, and that’s when we learned that in machine learning, if you’re network-based, actually your performance drops off dramatically. How do you get through that? We can quickly improve your computing performance by directly using RDMA. Based on this, we did some machine learning, equivalent to relatively fast can be used.
In addition, Lu Yi lei introduced several common cases to us. The first one is cross-device identification to people, to tell us how to separate the device ID information from the same person?
The answer can be identified through algorithms including behavioral analysis.
And the way of recognition is divided into three levels:
▲ The first is the identification stage, the identification of a family. We identify individuals by specific behaviors. For the same computer, I may be able to identify you as different people in different time periods. For example, if you surf the Internet, or your child plays computer during the day, after the evening, it may be parents surfing the Internet. Because everyone has different access habits, the same device can be identified in different scenarios.
▲ The second combines behavioral attributes through IP. In fact, IP it also points the scene, we know that there are university IP, company IP, base station IP, and even family IP, divided into a lot of categories, you do the category, then the intelligent identification will be very good.
▲ Third, we need to combine human behavior. By family groups, by individuals, actually by families, when you go home, you always have a router, and your router is connected to your phone, or your PC, or your computer, or even your TV box.
Another scenario, deep learning, image recognition, why does the advertising industry do image recognition? In fact, very simple, we in the image recognition, because our company is to do a third party effect monitoring company, we have to judge whether the advertising is cheating. The proof was when I took a screenshot of it, so I could show it to advertisers and tell them you had a problem with it. Having high requirements for image recognition.
For image recognition, you can not simply recognize the image, you need to be around it, including related knowledge base, you can recognize this well.
Once the recognition of this level, for our AI implementation, quite high requirements, the first image recognition do well, second, knowledge base is good, third, political things, have to have. At this time you need to automatically identify the picture, image recognition, you also identify his logo, including his content is what kind of, what kind of state is displayed at that time, one is image recognition, including sequence matching. Third, you need a very strong knowledge base to do this.
What is the third scene? We have a lot of behavioral data, how to use the massive behavioral data through our machine learning to catch cheating behavior?
Ruleley says this is a place where you can do this by constantly adjusting your model, including the parametric characteristics of your behavior, and you have to be strong enough to do it. For example, if you are currently swiping AD traffic, you can change your ID, including the length of your UI, and you can make all kinds of adjustments.
Third party monitoring effect evaluation company, exposure and click, including time, region, including audience coincidence, such a dimension, we through manual way to achieve this implementation, manual report to advertisers later. So when you do that it’s actually quite expensive. Is there a way to do this? Lu billion lei said, will automatically help you choose what media, what advertising space, what time period, what region, what crowd, I gave him all intelligent into a package, automatically put out after, after casting, will automatically do a new choice.
In fact, based on these scenes, natural persons can make the effect of the real efficiency of intelligent, Lu Yi Lei told us a word called marketing automation, now also called marketing intelligence, is also this reason. “The development trend of our digital marketing in artificial intelligence is not just to put the AD out and OK, we hope that after putting this AD out, no one will be disgusted,” Lew said.
First, the search revolution. Let’s say I’m looking for a place to eat. This restaurant is delicious. In fact, what we’re going to do is we’re going to do a search intent-based algorithm, and in fact, when you don’t type in, I’m probably going to predict what you’re going to do, which is infinite search.
Second, search engine intelligence. Now AI, including advertising, people are all personalized, how do you divide each person into segments, make each point into precise delivery of people, personalized content recommendation. Luyi lei believes that the final “private customization” is the best result of search engine intelligence.
Three, TPU. TPU, in fact, is specifically based on linear calculations. At present, it is good for servers or desktops to reach several terabytes or ten gigabytes of computing capacity, while TPU is 180 terabytes, so this computing capacity is very important in the era of big data.
In the end, Luyilai ended his share with a sentence: the future of digital marketing, the real intelligence, is to let everyone on the advertising is not so annoying, so that users feel that the advertising is really what they want to see.
Here’s luigley’s live Q&A session:
Q: Currently, precision marketing based on mobile devices involves data that users do not want to be mined. Have you considered the next step of interaction with legal boundaries when doing research? How to protect privacy? A: It is true that people are very concerned about data privacy. On June 1 this year, the Supreme People’s Procuratorate, the people’s Court and the people’s Procuratorate promulgated the Double High Court Law. If I remember correctly, the law stipulates that more than 50 articles of privacy are involved in the criminal law, which is quite serious. We also deeply studied the content of advertising, which has two points:
The first point, PI data, these personal things, like the phone network, including your name, these are PI data, these are PI data, which you absolutely cannot do, this is a very, very important issue.
The second PI data is irreversible if it is encrypted. In fact, you can do data exchange, and the same is true when we do personal IP identification. For example, device ID, our ADFA, is itself an irreversible thing. We do not touch other IP, such as id number, will not touch, there will be no such problem. Our side actually do is put the ID information in a later, he is not a probability problem, it is a portrait, it is not a delay IP, because we are a weak IP system, is not a strong IP system, nature can achieve 20%, 30%, do good, can achieve 80%, 90%, which is based on the algorithm to do. The second use of the scene, I put the PERSONAL information of the IP, just when the user touches you, you give him push, this we really we will not touch the red line.
Q: What you mentioned just now is that advertising is pushed based on people’s interests. However, we know that when evaluating the effect of advertising, the content and region are generally pushed, and then the effect of advertising can be inferred based on the consumption situation in this region. What do you do about this part?
Lew: Like advertising, how does it do when it wants to advertise? In the first point, he will decide on a certain city, a certain period of time, and even if he decides to vote, he will decide. But after he decides, he still cares about the crowd.
Second, when we are doing this, if you try to know the person first, he is in the industry, if you know in advance, after the prediction, you can put the advertisement, the effect will be twice as effective with half the effort.
Live streaming:
