news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Synced

Author: Zhang Qian

Being able to "bid" is also a very practical skill.

Among the many top AI conferences, what rank does NeurIPS belong to? Someone put it into the universe of "Empresses in the Palace" and made a picture: it is probably worthy of the "empress position".

Image source: Xiaohongshu user @云卷月舒

This ranking may be controversial, but there is no doubt that NeurIPS has always been among the top three AI conferences, and has long been in the top ten of all academic journals and conferences in the world on Google Scholar.

Therefore, being able to publish papers in this conference is a common goal for many AI researchers. The groundbreaking AlexNet, Transformer, and GPT-3 papers are all accepted papers in this top conference.

But it is worth noting that papers cannot represent the full value of NeurIPS. Some competitions held during the conference may be more suitable for researchers and engineers who focus on AI practice. Even NeurIPS officials said, "These competitions play an important role in researching and solving complex problems.」。

So where can you find these competitions? In fact, NeurIPS officially released a blog in June that specifically gave a list of these competitions.

The list contains a total of 16 competition topics, each of which has been carefully screened and has the "Extensive scientific research value」。

These competition topics are collected by NeurIPS. Based on the experience of previous years, most of the competition topics that can be selected are from universities, research institutions or foreign technology companies such as Google, OpenAI, Meta, etc. The chances of being selected by domestic industrial competition topics are extremely low. This year, due to the popularity of the large model track, the competition is more intense.

But surprisingly, in such a fiercely competitive environment, some people in the domestic industry still stand out.Auto-Bidding in Large-Scale Auctions: Learning Decision-Making in Uncertain and Competitive Games(Automatic Bidding in Large-Scale Auctions: Learning Decisions in Uncertain and Competitive Games)" is the competition topic they submitted.

Competition official website: https://tianchi.aliyun.com/specials/promotion/neurips2024_alimama#/

The competition topic was produced by the decision intelligence cooperation team of the Peking University-Alimama Artificial Intelligence Innovation Joint Laboratory (PAAI). The Alimama decision intelligence technology team, as the first unit, jointly submitted the topic with the research team of Professor Deng Xiaotie and Professor Lu Zongqing of Peking University. After the topic was selected, Alimama won the right to host the competition, becoming the only organization in the domestic industry to win the right to host the NeurIPS competition this year.

The competition topic revolves around the problem of "automatic bidding in large-scale auctions". This problem is closely related to the interface of searching and browsing products when we open shopping apps every day, and it contains huge research and commercial value.The NeurIPS expert judges commented that the topic is "Practically important, well organized, and well tested."。

So, what is the problem of "automatic bidding in large-scale auctions"? Why was it proposed by the Peking University-Alimama Artificial Intelligence Innovation Joint Laboratory and received such high praise? What exactly do the contestants have to do? Synced will explain it one by one in this article.

What is "Automatic Bidding in Massive Auctions"?

To understand what “automatic bidding in large-scale auctions” is, let’s first recall the experience of opening Taobao and starting shopping.

Every time you enter a keyword, the system will pop up a product page. In fact, this page is very particular: which products will appear and which product is ranked first are the result of precise calculations by the system, and the advertisements are the result of the platform running an auction mechanism to allocate advertising space to advertisers.

The operating logic of the whole process is as follows: First, the platform will build a user profile by analyzing the user's interests and behavior patterns. When a user searches or browses products on Taobao, the platform will immediately start the advertising auction process in the background. Advertisers participate in this auction through a bidding mechanism, hoping to get their ads displayed. The automatic bidding system plays a core role in this process. It comprehensively considers the user's profile, behavioral data, advertiser's promotion goals, budget constraints, and various factors in the auction environment to calculate the most optimized bidding strategy in real time. Based on this data and calculation results, the platform will select the advertisement with the highest bid and the most relevant to the user's needs. These advertising results will be displayed to users together with the natural results. The whole process is fully automated and can be completed in a very short time.

From this process, we can see thatThrough the automatic bidding system, advertisers can greatly simplify the advertising process and use artificial intelligence technology to achieve precise marketing, thereby saving time and energy.

In 2023, the global online advertising market has reached $626.8 billion. Automatic bidding technology is essential to driving its continued growth. Similar research issues include advertising delivery strategies and other mechanism designs, which all fall within the scope of decision intelligence research. Research related to decision intelligence can bring new operating methods to enterprises, reduce dependence on people in decision-making mechanisms, and thus significantly increase the growth rate of corporate revenue and enhance corporate growth space.

However, it is not easy to do the "automatic bidding" in the competition. Because the automatic bidding system needs to deal with a huge and complex data torrent, covering multi-dimensional information such as user behavior data, advertising data, bidding data, and these data will be updated in real time. In addition, the system needs to make decisions in a game environment full of uncertainty, and it is impossible to obtain complete information on all influencing factors. Therefore, the system can only rely on currently available data and accumulated historical experience, and make predictions and decisions through intelligent algorithms, striving to make the best bidding choices in a rapidly changing market environment.

The road to optimizing automatic bidding

From reinforcement learning to generative AI

Overall, the bidding industry has gone through four generations of evolution. Alimama has also conducted many years of research on the optimization of automatic bidding strategies.

• First generation: classic control type.Indirectly transform the optimization problem of maximizing the effect into the control problem of budget consumption. Calculate the consumption curve based on business data, and control the budget to consume as much as possible according to the set curve. PID and related improvements are commonly used control algorithms in this stage. When the value distribution of bidding traffic is stable, this type of algorithm can basically meet the effect optimization requirements at the beginning of the business launch.
• Second generation: planning solver type.Compared with the first generation, the planning and solving (LP) algorithm directly aims at maximizing the target. The current future traffic set can be predicted based on the previous day's competing traffic to solve the bidding parameters. The automatic bidding problem becomes a new sub-problem based on the current data that has been delivered, so this method can be used to solve it repeatedly, that is, Online LP. This type of method relies on accurate prediction of future competing traffic, so when it is implemented in actual scenarios, more work needs to be done on predicting the quality and quantity of future traffic.
• The third generation: reinforcement learning.In the real world, the online bidding environment is very complex and dynamically changing, and it is difficult to accurately predict future traffic sets. The entire budget cycle must be coordinated to maximize the effect. As a typical sequential decision problem, the third stage uses reinforcement learning methods to optimize the automatic bidding strategy. The iterative process starts with the implementation of the early classic reinforcement learning method, and then further approximates the "data distribution of the online real environment" based on the Offline RL method, and then to the end, it approaches the essence of the problem based on the Online RL method to achieve interactive learning with the real bidding environment.
• Fourth generation: Generate model class.Generative big models represented by ChatGPT have arrived with a surging momentum and have shown amazing results in many fields. New technical concepts and technical paradigms may bring revolutionary upgrades to automatic bidding algorithms. The Alimama technical team made early arrangements and reshaped the technical system of advertising intelligent marketing with the intelligent marketing decision-making big model AIGA (AI Generated Action) as the core, and derived automatic bidding strategies represented by AIGB (AI Generated Bidding).

When the latest research in the industry is at the third generation (2022),Peking University-Alimama Artificial Intelligence Innovation Joint Laboratory (PAAI) was establishedThis laboratory brings together many industry and academic experts: Dean of the School of Intelligence at Peking UniversityZhu SongchunProfessor leads the academic guidance, Peking University Chair ProfessorDeng Xiaotie, Associate Professor, School of Intelligence, Peking UniversitySong Guojieand Alimama’s technical directorZheng BoThey are all core members of the laboratory. Several experts led the laboratory to continue researching decision-making intelligence issues such as automatic bidding based on existing achievements.

During the research, they found that the original reinforcement learning method had some limitations. For example, in the long sequence decision-making scenario of automatic bidding, there would be a problem of excessive accumulation of training errors. At the same time, ChatGPT is verifying the powerful capabilities of generative AI in many fields. So, the team began to think about what generative models can bring to automatic bidding strategies? In the end,They proposed a bidding strategy optimization solution based on generative model construction - AIGB (AI Generative Bidding)。

Specifically, AIGB considers related indicators such as bids, optimization goals, and constraints as a joint probability distribution, thus transforming the bidding problem into a conditional distribution generation problem. Unlike the perspective of reinforcement learning, it directly links decision trajectories and reward information (as shown in the figure below), which can avoid the accumulation of training errors and is more suitable for long-sequence decision scenarios. This is the first attempt by the joint laboratory to apply generative large models to the field of smart bidding.Related papers have been accepted by the top international conference KDD 2024。

Of course, decision-making intelligence problems such as "automatic bidding in large-scale auctions" are far from being solved, and the application of generative AI in these problems has just entered the trial stage. Therefore, the Joint Laboratory submitted relevant competition topics to NeurIPS, hoping to leverage years of research accumulation and gather the power of the community to jointly promote in-depth research and solutions to these problems.

In addition to commercial value, these problems themselves also have high research value. This is because decision intelligence integrates disciplines such as artificial intelligence, data science, and game theory, and provides a systematic framework to solve complex problems. This promotes interdisciplinary integration and promotes innovation and cooperation in fields such as computer science, statistics, and economics.

AIGB and GM

Registration is open

The "Automatic Bidding in Large-Scale Auctions" competition is divided into two tracks:

• AIGB track: Using generative models to learn automatic bidding agents
• General track: Automatic bidding with uncertainty

In the AIGB track, contestants need to think about how to make accurate bidding decisions for long sequences. As mentioned earlier, traditional reinforcement learning methods are limited by factors such as error accumulation when facing this problem, and their performance is limited, while generalized generative models have shown great potential in this task. Therefore, this track requires contestants to adopt a wide range of generative models, such as Diffusion Models, Transformers, etc., to meet this challenge. If you have a research or professional background in Diffusion Models, Transformers, Foundation Models, Large Language Models (LLMs), and other generative methods, you can consider signing up for this track.

In the general track, contestants are challenged to make effective bidding decisions in large-scale auctions, which requires effective perception of changes in competitor strategies. The complex ad auction environment in the real world brings additional challenges, namely uncertainty. Participants must consider the randomness of consumer arrival, the variance of conversion behavior prediction, data sparsity, and other factors. If you have a research or working background in reinforcement learning, optimization, machine learning, game theory, and data science, you can consider signing up for this track.

Participating in this competition will bring many benefits. NeurIPS has a very high international influence. Achieving excellent results in the competition will undoubtedly add points to your resume and will be very helpful for your future career development. In addition, the winners will have the opportunity to win a $6,000 competition prize, as well as internship opportunities at Alibaba, campus recruitment green channel, and visiting scholar qualifications.

This competition will also make public about 500 million pieces of game data and the corresponding training framework for the first time. Such large-scale game data is very rare in the industry. This is a good practice and research opportunity for researchers and practitioners in the fields of decision intelligence, reinforcement learning, game and generative models.

The competition schedule and award settings are as follows:

• AIGB track registration: https://tianchi.aliyun.com/competition/entrance/532236
• General track registration: https://tianchi.aliyun.com/competition/entrance/532226

Competition official website: https://tianchi.aliyun.com/specials/promotion/neurips2024_alimama#/