Idea
CTR prediction model improving accuracy and efficiency by integrating sequential and set-wise user behavior contexts.
Research Paper
Core Innovation
This paper introduces HoMer, which aligns sequence side features with non-sequential features for fine-grained interest representation, shifts prediction from point-wise to set-wise to capture cross-item interactions, and employs a unified encoder-decoder architecture for computational efficiency and scalability. This approach addresses feature, context, and architecture heterogeneities in CTR prediction.
Why It Matters
Accurate CTR prediction is critical for recommender systems to increase user engagement and revenue. HoMer addresses key heterogeneities that degrade prediction quality, enabling better user interest modeling and cross-item context understanding. Its scalable and efficient design supports industrial deployment, improving business metrics and reducing computational costs.
Market Size (TAM)
$10–20B TAM for recommender system software; $2–5B SAM from e-commerce and online advertising platforms. Driven by growth in digital advertising spend and demand for personalized user experiences.
Potential Customers & Pain Points
- E-commerce platforms – Need improved CTR prediction accuracy
- Online advertising networks – Require efficient large-scale user behavior modeling
- Streaming services – Seek better recommendation relevance
- Recommender system providers – Struggle with integrating heterogeneous data and scaling models.
Business Model
Licensing the HoMer model as a SaaS API or on-premise solution for recommender system providers and large digital platforms, with tiered pricing based on usage and scale.
Competitive Landscape
- YouTube DeepCTR
- Alibaba DIN
- TikTok Recommendation System
- Amazon Personalize
Implementation Challenges
- Integration complexity with existing pipelines
- Data privacy and compliance concerns
- Competition from established CTR models
Validation Strategy
- Pilot deployment with select e-commerce and advertising partners
- Benchmarking against existing CTR models on real-world datasets
- Measuring impact on online business metrics such as CTR and revenue
- Optimizing engineering for resource efficiency and scalability
Research Paper Overview
HoMer: Addressing Heterogeneities by Modeling Sequential and Set-wise Contexts for CTR Prediction
Summary
HoMer improves click-through rate prediction by aligning sequence and non-sequential features, shifting from point-wise to set-wise prediction to capture cross-item interactions, and unifying architecture for efficiency and scalability. It outperforms industrial baselines in accuracy and online business metrics while reducing GPU resource usage, making it practical for large-scale recommender systems.