Gai Kun has a PhD in machine learning and artificial intelligence from Tsinghua University. He has published several papers in top journals and conferences (NIPS/CVPR/AAAI/TPAMI, etc.) and a few years ago proposed the Mixture of Logistic Regression (MLR) algorithm to improve the accuracy of Alibaba’s CTR prediction.
When he first joined Alibaba, Gai worked in the advertising technology department of Etao in advertising related technology. The first killer application of machine learning in the Internet industry was AD click rate estimation, so Gai Kun was in charge of the development of machine learning algorithms on alibaba Group’s machine learning platform, although he was in the advertising department. On the business side, it is equivalent to applying the algorithm on the platform to the core technical modules of advertising click estimation and conversion rate estimation, which is a three-end connected work from the platform to the model and then to the business.
At that time, Gai Kun was in the basic research team and supported several different advertising business lines horizontally. His platform also supports some non-search, non-advertising businesses, such as recommendations and search.
Related Manufacturer content
QCon in October to get technical masters clearance strategy!
Luji Thinking Go language micro-service transformation complete process
Future IT architecture model for Netflix: Serverless
Ali Cainiao global cross-domain RPC architecture practice pain points and difficulties
The hybrid architecture practices of alliance chain and public chain are analyzed from the perspectives of consensus mechanism, rights management, privacy protection, smart contract and governance mechanism
Related Sponsors
“Our team will gradually become a vertical business team, responsible for targeted advertising. Precision targeted advertising is the ability to generate ads that interest users in the absence of search terms.” “Gai kun told TGO.
At the 2018 New Zhiyuan Industry · Leapfrog AI Technology Summit, Gai Kun made a speech on the topic of “The Application of Deep learning in advertising Business”, talking about alibaba’s exploration and practice of applying deep learning to advertising business.
Here are excerpts from Gai Kun’s remarks:
One of the core applications in advertising, and the first killer application of machine learning, is CTR estimation for advertising. CTR estimates can be generalized to many estimates, including conversion rate estimates. Both merchants and platforms care about conversions and transaction amounts as well as clicks. So this estimate is directly related to the platform’s revenue, because CBC doesn’t charge for clicks.
In the past, the general performance was large-scale linear model, but now we have proposed a nonlinear model called MAR. From the nonlinear model, we have transformed it into a standard deep learning with embedded and multi-layer perceptron. The user’s interest is not single, we want the expression of interest to the user to be diverse, mathematically we think of it as a multi-modal distribution.
It’s also inspired by actual data — when we shop for things ourselves, there’s never just one or a group of things in the cart. Even if you only buy one item for the day, there is more than one item in the cart, so users’ interests are diverse. In deep learning, if users are grouped into a point of interest, they cannot be fully expressed. We hope to outline the diversity of users’ internal interests with models. After the standard deep learning, we also proposed a feature called Deep interest Networks to describe the coexistence of multiple interests of users, which has achieved good results in business effects.
In terms of the input and perception ability of deep learning, we will introduce an understanding of the essence of what users really care about in addition to their behaviors. For example, if A user is looking at A and B, if many users look at A and then look at B, A is closely related to B, which can be captured by traditional behavior-based models. What is the essence behind A and B, we want the model to really understand. If a lot of people are looking at these two things at the same time, we know they’re related, but we don’t know what they are. Therefore, we add the image information into deep learning behind the scenes, connect the image model with the CTR prediction model, and turn it into a CTR prediction model with perception image. The model is very large, and we have also made an innovation in the computing framework that allows these two models to work together end-to-end.
This also includes our recognition of deep learning — end-to-end learning is the key to make the most of deep learning. Here’s a counter proof: We’ve tried a lot of non-end-to-end approaches, and they don’t work very well. For example, an image, the easiest way to do this is to extract features from the image and plug them into a model of CTR or conversion rate. Moreover, deep learning would not have flourished in recent years if the method of extracting features and adding a model for each feature worked. It turns out that I can take the output of one shallow model as the input of the next shallow model, and make it better step by step with the shallow model, and finally become a deep calculation. In practice, the output of these models must be coordinated with subsequent models rather than directly fitting the final goal, so end-to-end is important. This is both a challenge to the framework and an opportunity.
The calculation mode of deep learning framework has evolved from the original classical parameter partial service mode to the distribution mode of Advanced Model Server. In the future, models will become larger and larger, and different data sources will be input together. Multiple models will be integrated together to conduct large-scale collaborative training. The distribution mode of Advanced Model Server may be the direction of future framework evolution. This is a series of advances in deep learning on predictive models.
The second line is on the match. For a traffic-based business like the Internet, there is an online engine, and there is a limit to what is delivered to the model to be estimated by the online engine, so there are retrieval or matching modules in front of it. For example, when you come to Taobao, we first select 10,000 items that you may be interested in in a rough way. We use deep learning to estimate the click rate, conversion rate and interest rate of each of these 10,000 items, and finally select the best ones to display. But a lot of times, if you don’t have what you really want in 10,000 items, then the model doesn’t matter how much you measure it, so matching or retrieval is a bottleneck in the whole system, and we’ve made some improvements on that.
There is A classic recommendation algorithm called collaborative filtering: A and B if A lot of people look at A and B at the same time, we think A and B are similar. This is called collaborative filtering. Collaborative filtering is used for recommendations, where related items that you have seen historically are used as a candidate set for recommendations, and the model evaluates which items to recommend to you. But collaborative filtering directly limits the candidate set, which is a set of products that you’ve already done that are very similar to each other, which makes users feel less innovative or less intelligent. So we want machine learning to be able to better predict what users are interested in, and we want to remove this limitation in the matching process, and let machine learning play a role here.
There is a simple method in the industry, when the machine learning model still stays in the very simple internal machine model, that is, the limited retrieval method, using the model to measure the interest degree of each product, rather than using a certain rule or some statistical formula to define first, so that the quality of the whole recommendation will be greatly improved.
Second, the industry also has a hand behind it. Facebook opened source a FASS framework last year, which may not be as well known as Tensor Flow, but it’s been used internally by many companies to improve their search systems. But it doesn’t have to be Facebook, and some of the best teams are doing it themselves. But having an open source framework in place makes development easier and takes less time. Tensor Flow hasn’t been developed by every company yet, limited retrieval has been developed by itself.
But there is one limitation to the model — it must be internal. We were wondering if an arbitrary high order deep learning model could measure the reputation of users and products, could we have a way to pick out the ones you’re most interested in out of a sea of products. Therefore, we made a tree structure search engine, the search engine is to build the whole commodity into a hierarchical category tree, built with a binary tree, each node under only two nodes. Category tree from top to bottom: if the top node is not the most interesting node, all subsequent nodes do not need to calculate, the engine has mathematics or basic data to ensure this thing, so you can choose the best product in a very large range, will become a very fast calculation from top to bottom.
For example, if a billion goods is 30 layers, that’s 30 judgments, not 1 billion judgments. So it allows deep learning to approximate the entire library, to pick out what deep learning thinks is the best commodity. Deep learning has really transformed from a prediction and measurement problem into a retrieval optimization problem, which is our progress in matching.
Going forward, we believe that in many scenarios of Alibaba, there will be great changes in the quality of recommendations, the level of interest of users, and the novelty of continued recommendations. We are also trying some scenes, some scenes have been online, but ali’s scenes are many, we still have a lot of work to do in the future, this is the second development of deep learning in matching, retrieval.
The third line is intelligent decision making. At present, intelligent decision making mainly uses reinforcement learning. We can divide it into two parts:
The first is the platform. There are many links and a lot of parameters in our online advertising system, recommendation system and search system. When the system is powerful, it has many parameters and modules. How to adjust these parameters has a great influence on the final result, so we have changed from trying to adjust many parameters by people to using reinforcement learning to automatically adjust parameters according to the final effect, and the effect is very obvious.
The second piece is for advertisers to make decisions. For the merchants, how can we help them choose the crowd, make the price, and even help them to evaluate each flow, each flow can make a customized bid, which is a special ability given by the platform to the merchants. But businesses cannot control it themselves, neither the cost of control nor the confidentiality of data is allowed. Businesses can’t see a specific user’s data, so if our platform has this system, we can give this ability to businesses, but the data is not completely accessible to businesses, only the data related to themselves.
In this way, we let businesses in the advertising link, whether it is the selection of the crowd, fine bid for each flow and how to show personalized creativity and other different links can be connected, become a real intelligent advertising system. It ultimately uses reinforcement learning to optimize business-defined goals. For example, the largest number of stores, the highest input-output ratio, the largest number of transactions, and so on, we optimize a series of decision-making processes involved in the business under this goal.
Of course, there’s a part of it that’s semi-automatic. For example, creativity, now can not completely let the algorithm automatically generated from nothing, really in business, there have been a lot of intelligent drilling exhibition customers in use. Users can upload a lot of materials in the creativity. After uploading such elements and materials, we will automatically help users to assemble them, and automatically show the appropriate elements to the appropriate users. For example, some users are more sensitive to price, so more relevant information will be displayed. Some users are sensitive to style, so the whole idea is organized differently. In fact, this is the ultimate goal of attracting more users according to the effect, which can greatly improve the effect of the final advertising business, and can reduce the difficulty of the operation of advertising. So this is the third line, where we use intelligent decision making to solve a lot of the decision problems on the platform and on the advertiser side.
So now everyone is going to reuse these three lines:
- Predictive model. This is the most classic advertising core algorithm;
- Retrieve and match. When the latter model is more powerful, it is a bottleneck;
- Platform and advertiser decision – making problem.
After the speech, reporters from TGO Kunpeng interviewed Mr. Gai Kun at the conference site. The following content is based on the interview of the day, some of which are not changed.
TGO Kunpeng: What did your team mainly do?
Gai Kun: My team is responsible for targeted advertising. The specific products are the diamond booth (now called smart Diamond exhibition) and the directional function in the through train.
There are two key elements: one is the display or recommendation position, which is distinct from the search position, and this position will be more and more in the future, because search engines are still run by a few big companies. While almost all companies now have recommendations, not all companies have strong search. The expansion of different user scenarios including ali will also bring about non-search. In fact, the location of such content display is increasing. The second is our advertising. Unlike brands, advertising is responsible for the effect, which is related to click-through rate and conversion rate.
TGO Kunpeng club: What system do you have in terms of specific effect evaluation?
Gai Kun: There are many effect evaluation systems. In terms of advertising flow, we will have display volume, click volume, click rate, conversion rate, purchase volume, transaction volume, transaction amount and so on. We have also made a lot of perspective on the real value of advertising: on the one hand, most of the time, the essential purpose of advertising is not just to see the value of advertising traffic, but to achieve their own store growth through advertising channels and grasp.
For example, customers will not buy a product if there are few reviews from merchants. Because many activities have thresholds, some businesses may not participate in the activities; Many places also don’t get traffic because it doesn’t have a proven history of being a good seller. In fact, commercialization provides such a channel for businesses. After buying traffic, if its goods are really good and have a way of operation, it can accumulate things for its future growth. Therefore, In the broad SENSE of ROI, Alibaba also measures other values brought by advertising, its own growth value, including the incremental value brought by some full-stack equity.
For merchants, the operation process is the whole link process of each consumer — from unknown to contact with merchants, consumers gradually increase their interest through many interactions in the contact process, and finally reach purchase. After purchase, they buy back, and become continuous regular customers, and even help merchants to spread. In the future, we will not only see advertising now prominent window, Alibaba covered ali system. Alibaba has partnerships with many media, and with these media, we will find ways to help merchants do a better and more efficient job in the whole consumer’s whole link operation.
This includes several new aspects:
- How to establish a scientific measurement system for the whole link operation;
- How to optimize efficiency;
- If we can really make better scientific insight and optimization, will we be able to make more intelligent decisions in some links and give birth to a new way of operation?
The above three points are the joint efforts of various business segments of Alibaba, and are also the direction of future research and development. We have an advantage in the coverage of the whole link of merchants’ consumers. We recommend positions in various places, including taobao home page and external advertising.
TGO kunpeng would: I have a question, for example, after I mistakenly ordered a product on Taobao, the system would recommend some similar products to me. Is there any solution to this problem?
Gai Kun: This is something that alibaba is developing internally. It has several aspects:
The first level is the problem of misalignment between target and data. Many consumers complain that they still recommend things they’ve already bought. Right now, the main problem with user experience is that the goals of model optimization are not aligned with the experience. There is a technical problem here — machine learning must have a clear goal, and a large amount of target data, in order to work well.
Our target data does not have the actual thoughts of users, our data is click and buy behavior. Therefore, many models we build will eventually come back to the prediction of whether users click or not, and whether users click more times or even buy after optimization. However, the probability of clicking and buying is not exactly equal to the quality of the user experience, just as clicking and buying are not equal.
For example, when we stop recommending something a user has bought, we find that the results are much worse than when we recommend something he has bought. And to many people’s surprise, many people would think that we should make a rule and filter it out directly, but we found that the filtering effect also decreased significantly. I am a living example of this myself. Some time ago I bought an electric refrigerator, which I thought I would never buy again, but it turned out that I bought another one for my parents. Another example is water purifiers. When I bought them, I thought I would never buy them again, but I ended up buying three. But it still doesn’t feel right when it continues to recommend me after I’ve bought it.
This shows a problem, even in statistics, people who have bought electricity are not less likely to buy it again. Anyone who has actually done drainage knows that the average drainage efficiency is a few percent or a few thousandths. Buy the person of big home appliance so, if be taller than this probability, he can go buying repeatedly, tell from probability, show this kind of thing to him to click rate character, it is better choice.
TGO Kunpeng: Actually, sometimes I want to check the price increase.
Gai Kun: Yeah, sometimes the hits are not low. Statistically speaking, our optimization goals are now all about clicks and conversions, not user experience. On the user experience side, our biggest problem is the lack of data. First, we don’t know how to distinguish between those who will repeat purchase and those who won’t. So, we’re just going to have to mix them up and do it the average best way.
The second is what the experience really is. If we put a lot of weight on negative experiences, we can just leave out all the negative propensity patterns. In fact, we don’t experience positive or negative data. So there’s a big challenge here: in the future, how to capture the perception of user experience.
There are several directions for the future:
One is whether it is possible to derive it automatically with algorithms; The second problem is data annotation. For example, images or search engines may have relevancy tagging teams dedicated to identifying which pages are relevant. There are a lot of problems without CTR optimization, but in order to control the experience, it all has relevant labeling, which is the way to get the data manual. For example, the image data is relatively good, if the public data set is not well marked, you have to find their own manpower to mark.
The third way is to design some user feedback channels on the product, so that users can express and collect data. So there are some internal considerations and options, but it’s not online yet. Therefore, we may comprehensively look at how to get experience data so that the algorithms and models behind can really have a grip to optimize this matter.
Technically, there are two other things: first, data acquisition is difficult and of low quality, so how can the model better depict the subtle feelings of users’ psychology through insufficient data? For example, we have done more interest in characterization before, and later we will do more refined characterization on fewer data, which is the direction that the model itself needs to explore.
Second, if our model is capable, we must not get stuck in some kind of rhythmic way, and the whole tree search we just talked about might help a lot in this matter. So there are a number of things that are going on technically, but also in tandem with changes on the business side, that we might think about as a whole how to improve on.
TGO Kunpeng: As I said just now, user experience is also the challenge facing advertising business. Besides this, are there any other challenges or problems?
Gai Kun: The recommendation business is more, but the advertising business may be ok, because the commercial proportion of advertising business in various scenes is not very high, so the proportion of advertising traffic in each scene is relatively small, and the probability of real advertising bringing negative experience is relatively small.
For example, if only one out of hundreds of feed streams is an AD, most users might not be able to tell which of alibaba’s scenes are ads at all. One is less, and the other is that Ali is really a place called native advertising. The concept of native advertising may have been mentioned in other companies, where the format of the AD is more similar to the mainstream results, so that users feel more natural. But Ali doesn’t have a form. Ali is a thing. I’m using recommendations and search to offer products that don’t pay for them. Essentially, the content is the same, except for the distribution of traffic. So it should be said that What Ali really provides is a native advertising experience. I’m sure many of you don’t know where advertising is unless you have done a through train and diamond show.
The related content
Baidu’s “deep Learning open course” is full for the first time, and developers reap the most advanced “Treasure Book”
Build a distributed advertising event prediction system
Massive practice of social advertising system with tens of billions of flows on Qzone platform
Application of deep learning in image understanding
Application of deep learning in CTR
Hello, friend!
Register an InfoQ account
The login
Get more experiences from InfoQ.