Companies want to obtain the largest possible amount of profit while spending minimum money independently from the type of advertising. Compared to offline advertising—print media, TV, etc.—online advertising enables more instant feedback, like when a user clicks on an advertisement or a link. Thereby, brands can obtain extensive information about their customers, such as previous behavior, demographic, and geographic information (Pandey et al., 2011).
To illustrate, imagine a brand selling self-care products. They could advertise their products on TV during popular and relevant shows or programs being aired, aiming at the audience who are interested in self-care to notice. On the other hand, if they do online advertising, they could put their ads next to blog posts or articles about beauty, wellness, or self-care. But the real strength of advertising online is being able to target people who searched with words like “self-care products” or visited similar sites. This unlocks new potential to engage with the right audience, optimizing ads. This is called targeting (Pandey et al., 2011).
Targeting (or behavioral) advertising is becoming a more significant source of income for both advertisers and advertising organizations in the rapidly expanding fields of mobile services and in-app advertising. The foundation of targeted advertising is big data analytics, which gathers and processes user personal data for targeting and profiling (Ullah et al., 2014).
In the current digital marketing realm, it is vital for marketers to leverage targeting techniques for providing tailored experiences to users. Nevertheless, preexisting strategies for user targeting often lack adaptability across various domains and have difficulties in making predictions about future user behavior. To solve this problem, Dou et al. introduce a new technique called FIND (Forecastable and Industrial-grade Network for user targeting). This model overcomes the problems of existing methods, making it a groundbreaking model in the industry.
What is FIND?
FIND is an advanced foundation model built for efficient user targeting. It not only combines multiple user data types from different sources but also is able to specify target audience with single-sentence demand, enhancing usability and effectiveness.
User data types that FIND combines:
- Transaction records—past purchasing activities; particular company, product, and category; amount of money, etc.
- Browsing history—used keywords, clicked links, and visited websites, etc.
- Financial data—activity of an account, etc.
Key Developments of FIND: Two-Stage Process of Pre-Training
- Self-Supervised User Modeling: In this stage, FIND learns to identify universal user representations by examining interactions of users. Ensured by a contrastive learning approach, similar types of users are grouped together, while the distinction between various user groups is preserved.
- User-Text Alignment: In the second stage, FIND analyzes the user’s previous behaviors and generates simple, readable texts that are able to predict future user behaviors. This enables companies to input just a simple inquiry such as “Find users who are interested in self-care products,” and the relevant group will be automatically identified by FIND.
What are the practical applications of FIND?
There are two main strategies that FIND offers for efficient targeting:
- Zero-shot Transfer: With this strategy, FIND specifies the right users without further fine-tuning, allowing companies to fill out just one sentence that defines their target group. This method is specifically advantageous for marketers, either seeking to enter new markets or to adjust quickly to new marketing efforts.
- Few-shot Targeting via Prompt-tuning: There might be scenarios where only a limited amount of labeled data is available. But the real strength of FIND shows itself here; it is capable of refining its targeting prompts by using a small number of seed users. This is very efficient, considering conventional models would not be as successful in these circumstances.
Real-World Scenarios
An extensive series of experiments has been conducted to demonstrate the effectiveness of the FIND foundation. To test real-world circumstances, researchers leveraged data from the Alipay platform. FIND consistently surpasses the performance of existing baselines in various user targeting tasks, including security risk control, marketing, and recommendation systems. Its efficiency in making precise predictions about user behavior makes it a highly remarkable tool for marketers aiming to enhance the effectiveness of their marketing campaigns.
Furthermore, FIND has already been implemented in Alipay, and there was a notable success. It has enhanced the performance of key metrics such as click-through rates (CTR). For example, it has led to a 30% increase in CTR for users interested in “Used Car Selling Preferences” and a 13.2% increase for those engaging with “Video Watching in Alipay.”
Takeaways
FIND represents an important progress in the field of user targeting. By addressing the limitations of traditional methods—such as poor adaptability across fields and inadequate forecastability—FIND provides a more robust and adaptable solution for digital marketers. Its ability to combine diverse types of data and predict future user behaviors makes it a valuable tool for businesses aiming to optimize their advertising efforts.
As digital marketing continues to evolve, models like FIND will play an increasingly important role in helping brands connect with their target audiences more effectively.
References
Dou, B., Wang, B., Zhu, Y., Lin, X., Xu, Y., Huang, X., … & Hong, C. (2024). Transferable and Forecastable User Targeting Foundation Model. arXiv preprint arXiv:2412.12468.
Pandey, S., Aly, M., Bagherjeiran, A., Hatch, A., Ciccolo, P., Ratnaparkhi, A., & Zinkevich, M. (2011, October). Learning to target: what works for behavioral targeting. In Proceedings of the 20th ACM international conference on Information and knowledge management (pp. 1805-1814).
Ullah, I., Boreli, R., Kaafar, M. A., & Kanhere, S. S. (2014, April). Characterising user targeting for in-app mobile ads. In 2014 IEEE Conference on computer communications workshops (INFOCOM WKSHPS) (pp. 547-552). IEEE.
