The Future Of Advertising Platforms In LLM! New Business Models And The Potential For New Personalized Advertising
3 main points
✔️ Proposal for an advertising system utilizing large-scale language models: Proposes multiple frameworks for an advertising system utilizing large-scale language models and compares the advantages and disadvantages of each approach. Provides an opportunity to discuss future online advertising.
✔️ Examining the basic requirements of an advertising system: basic requirements that a practical large-scale language model-based advertising system must meet, including protecting privacy, ensuring reliability, minimizing latency, maximizing user and advertiser satisfaction, and monetizing the provider of large-scale language models through advertising. Presented by.
✔️ Validation of Dynamic Creative Optimization (DCO): discusses the potential for personalized advertising beyond traditional DCO techniques by using large-scale language models.
Online Advertisements with LLMs: Opportunities and Challenges
written by Soheil Feizi, MohammadTaghi Hajiaghayi, Keivan Rezaei, Suho Shin
(Submitted on 11 Nov 2023 (v1), last revised 14 Feb 2024 (this version, v2))
Subjects: Computers and Society (cs.CY); Artificial Intelligence (cs.AI)
code：
The images used in this article are from the paper, the introductory slides, or were created based on them.
Summary
In the world of online search engines, advertising plays an integral role in shaping the digital experience. This market has reached trillions of dollars and continues to grow. The economic importance of advertising is even greater. With a wide range of information and services available for free, advertising has become an important pillar of support. The advertising revenue model is democratizing access to education and information, enabling people to enjoy content for free, and contributing significantly to the development of the digital ecosystem. The synergies between advertising and content creation are driving economic growth, and with companies like Netflix introducing ad-supported plans, the advertising market is only growing in importance.
In recent years, the provision of large-scale language models, such as ChatGPT, has become widespread, assisting users with a variety of functions such as question answering and content generation. It is said that the spread of these technologies could replace traditional search engines as a means of information retrieval, and the possibility of advertising monetization by providers of large-scale language models is also being considered. Against this backdrop, discussions have begun on how to implement online advertising in large-scale language models and generate profits.
This paper examines the challenges and possibilities of applying online advertising and auction formats to large-scale language models. In particular, it focuses on the benefits and challenges when advertising is incorporated into the unstructured output of a large-scale language model and examines ways to meet the needs of advertisers and users. We also examine the issue of how advertisers interact with and bid on the system. We also discuss how large-scale language models can provide ad content that is better suited to individual users and generate more attractive ads, so-called "dynamic creative optimization" (DCO) and "responsive advertising.
The study spotlights the future of advertising and digital content and presents a new approach that will benefit both advertisers and users.
Requirements for a large scale language model advertising system
The paper also discusses the requirements for a large scale language model advertising system. The requirements include the following
Protecting Privacy: The most important aspect of user and advertiser interaction is to protect the confidentiality of privacy. This is essential to keep users' information and data secure and avoid the risks posed by accidental leaks.
Ensuring Reliability: As with online advertising in general, a trusting relationship must be established with advertisers. To maintain this trust, the system must closely monitor advertiser behavior to ensure consistency and reliability of the system.
Minimize latency: Users expect fast service, and while there is no small amount of latency in displaying ads, this latency should be minimized as much as possible so as not to compromise the user experience.
Maintaining User Satisfaction: Content quality should be kept high even when advertising is incorporated into the output of a large-scale language model. Excessive advertising can detract from the user experience and significantly reduce user satisfaction, especially if it does not match the user's search and interests.
Ensure advertiser satisfaction: Advertisers expect their own ads to be appropriately exposed. Ads should be presented in an attractive and engaging way. This increases profitability for the advertiser.
Ensure profitability for providers of large language models: the provision of advertising services is intended to generate revenue. It should be noted that the addition of advertising may cause a decrease in the number of users, which could be counterproductive to providers of large language models. The cost of advertising should be designed to be fully compensated by revenues from advertisers.
By meeting these requirements, the advertising system provides a valuable service for both users and advertisers, while ensuring profitability for the providers themselves.
Overview of Advertising Systems for Large Language Models
This section provides an overview of an advertising system that utilizes large-scale language models (hereafter referred to as "LAS"). The user inputs prompts to the large-scale language model, which processes the output to see how it can be connected to advertisements. The LAS takes into account various contextual factors, such as the user's past search history, and recommends ads.
LAS consists of four modules: Modification of Output (Modification), Bidding, Prediction, and Auction. Each module works in concert to find the optimal combination of user input to the large-scale language model and the optimal advertisement for that input.
The Modification (Modification) module outputs customized advertisements based on user input. Two approaches are being considered: the Advertiser Modification Model, in which advertisers themselves customize the output, and the LAS Modification Model, in which LAS directly modifies the output.
The Bidding module determines the advertiser's bid amount based on the modified output. In the dynamic bidding model, the advertiser determines the bid amount by providing the modified output and related information for each input. In the static bidding model, on the other hand, bids are based on predefined keywords.
The Prediction module calculates user satisfaction and ad CTR and evaluates the quality of the final ad output based on these metrics. This coordinates the improvement of user experience and maximization of LAS revenue.
The Auction module ultimately determines which ads are displayed to users and the amount charged to advertisers. This module determines the best ad serving and appropriate price based on bid amount, satisfaction, and CTR.
The goal of LAS is to maximize advertising revenue while balancing short-term revenue and long-term user retention. To this end, each module is precisely designed to allow advertisers the flexibility to adapt to their own strategies. Ultimately, the appropriate auction format is selected, such as a first-price auction or a second-price auction.
Dynamic Creative Optimization (DCO) using large-scale language models
Here we introduce the potential of Dynamic Creative Optimization (DCO), a technology that leverages large-scale language models to dynamically adjust combinations of ad assets to build ads that best match customer preferences, often used in the traditional search and display advertising markets DCO is a technology used in the traditional search and display advertising markets. This technology significantly improves the quality and relevance of the ad to the user by tailoring the content of the ad to the individual. Many online advertising platforms use DCO-based advertising, so-called responsive advertising.
In a typical DCO framework, a static ad consists of a single image file that is only displayed in a specific ad channel; in contrast, a dynamically optimized ad consists of a single ad template and multiple assets to fill that template. This is done by allowing advertisers to register multiple asset options for an ad template, and the system selects the best combination, with the goal of increasing CTR and SR and improving the efficiency of the online advertising ecosystem and the ad platform's revenue.
However, the advent of large-scale language models, such as ChatGPT-4, which support image output, could potentially replace the role of traditional DCOs in online advertising platforms by introducing a dynamic process of ad creation, unlike traditional DCO technologies. For example, based on user input and context, platforms can send queries to large-scale language models to customize ad images that attract users by capturing their preferences. The figure below illustrates two scenarios of personalized advertising.
A more advanced modification module is needed to incorporate responsive ads. This module should not only modify the original response to include ads, but also provide the ability to incorporate ads based on user preferences. This includes leveraging user context, such as gender, location, and device used for the query, into the ad generation process. This will make ads more engaging by taking user preferences into account. Leveraging the language model can prompt consideration of various factors related to user context during the generation of modified output. As shown in the previous figure, information about user context can be leveraged to create more engaging ads. In addition, the predictive module could work with the modification module to commit to modified output that maximizes the user experience.
Unlike traditional DCO techniques, which simply determine the efficient combination of ad assets with respect to cost-sharing models, large-scale language model-based responsive advertising can create truly new content. However, this process involves increased use of computational resources, especially as more queries are made to the large-scale language model to customize the ad. This section discusses cost-sharing models that LAS can employ.
The simplest model is to bill the advertiser each time content is dynamically modified. For example, an advertiser and LAS could enter into a contract for the amount to be paid for each responsive ad, and LAS would bill the advertiser each time the ad content is modified to be responsive. The contract could specify when the advertiser's ads will be changed to responsive. LAS may also provide options for the advertiser to determine how often and how heavily the ad should be modified to be responsive.
These processes require additional computational resources and can affect latency when incorporating dynamic advertising. Online advertising typically requires that the overall latency of the real-time bidding process be within 100 milliseconds, but there can be a delay of a few seconds to correct the image output. Balancing ad quality and latency is technically interesting and can be done in simple ways, such as using caching, or by having the advertiser bear some of the financial burden.
Summary
This paper addresses the prospects and challenges faced by online advertising systems that utilize large-scale language models. It lists the basic requirements that a practical advertising system must meet and presents a framework for meeting them. It also examines its effectiveness. It compares the advantages and disadvantages of the frameworks proposed in this paper and discusses the technical and practical challenges involved in designing an effective online advertising system. In addition, the paper discusses the usefulness of advanced dynamic creative optimization (DCO), which can be achieved by utilizing large-scale language models. The paper provides an opportunity to explore the role and potential of online advertising in future large-scale language models.
Categories related to this article
