Kuaishou HubMixer: Efficient feature interaction via latent hubs
_reachsumit · x · 2026-08-31
Kuaishou proposes HubMixer, a parameter-efficient architecture for feature interaction in recommendation systems.
- Context: Directly mixing heterogeneous tokens (user profiles, item attributes, etc.) is parameter-inefficient.
- Method: Introduces learnable "latent hubs" to organize interactions via an induction--interaction--readout paradigm.
- Benefit: Achieves higher accuracy with fewer parameters by mixing in a cleaner latent space.
More from Research
- NVIDIA's Kimodo Turns Text Into Full-Body 3D Human and Robot Motion — maier_ak · 2026-08-31
- GEPA optimization boosts Molmo performance by 122% — lateinteraction · 2026-08-31
- Study finds 90% of biomedical papers show signs of AI writing assistance — skdh · 2026-08-31
- ABot-Recon: Long-Horizon Streaming 3D Reconstruction with 12-Frames Context — zhenjun_zhao · 2026-08-31
- WilLaGS: Robust Gaussian Splatting in-the-Wild via Latent Appearance Fields — zhenjun_zhao · 2026-08-31
- Code as Worlds: Agentic Discovery of Executable World Representations — _akhaliq · 2026-08-31