Idea
User foundation model improving ad bidding and click prediction with fragmented open-web browsing data.
Research Paper
Core Innovation
This paper introduces a user foundation model pre-trained with masked language modeling and sequence-level contrastive learning on fragmented browsing histories. It uniquely addresses the open-web's non-persistent user identity challenge by exploiting sequential structure in short sessions, outperforming traditional aggregated counter-based methods in production ad bidding and click prediction tasks.
Why It Matters
Open-web advertising faces challenges from fragmented, non-persistent user identities and limited browsing history due to privacy constraints. This model enhances ad targeting and bidding efficiency by leveraging short, disjointed user sessions, increasing click-through rates and reducing costs. It scales across diverse open-web environments, improving revenue and user experience for advertisers and platforms.
Market Size (TAM)
$20–50B TAM for digital advertising and real-time bidding; $5–10B SAM from ad tech and e-commerce platforms. Driven by increasing demand for privacy-compliant user modeling and efficiency in programmatic advertising.
Potential Customers & Pain Points
- Ad tech companies – Struggle with fragmented user data
- Real-time bidding platforms – Need better click prediction
- E-commerce platforms – Require improved user targeting
- Digital marketers – Seek cost-effective ad spend
Business Model
Licensing the user foundation model as an API or SDK to ad tech companies and real-time bidding platforms, with usage-based pricing tied to prediction improvements and cost savings.
Competitive Landscape
- Google Ads
- The Trade Desk
- Criteo
- Amazon Advertising
Implementation Challenges
- Data privacy regulations limiting user data availability
- Integration complexity with existing ad tech infrastructure
- Competition from established large-scale ad platforms
Validation Strategy
- Deploy model in live A/B tests with partner ad platforms to measure CTR and eCPC improvements
- Benchmark against existing user modeling approaches on real-world RTB datasets
- Collect feedback from early adopters on integration ease and performance gains
Research Paper Overview
Building a User Foundation Model for the Open Web
Summary
User foundation models improve e-commerce and social recommendations but struggle with fragmented, non-persistent user identities on the open web. This paper presents a Transformer-based user foundation model trained with self-supervised learning on browsing histories, enhancing click prediction and bidding performance in real-time bidding environments. The model shows measurable gains in production metrics and live A/B tests, proving its effectiveness in open-web scenarios with limited historical data.