Researchers propose PaletteID to boost AI recommendation accuracy
New method PaletteID improves ad click prediction by 20% in tests using semantic prototypes...
PaletteID (PID) is a prototype-based semantic identifier for multimodal click-through rate prediction. It addresses key limitations of existing residual semantic identifiers by constructing a compact prototype palette via Semantic Quality-Aware Determinantal Point Process (SQ-DPP), which jointly captures local content density and global semantic diversity. For each item, PID retrieves and aggregates semantically related prototypes, enabling richer and more complementary representations. Experiments on two public datasets show that PID consistently improves CTR prediction—especially for long-tail items—while producing more robust identifier assignments and more interpretable token semantics than existing residual SID methods.
- PaletteID (PID) improves CTR prediction accuracy by up to 20% on public datasets compared to existing residual SID methods
- Uses SQ-DPP (Semantic Quality-Aware Determinantal Point Process) to construct prototype palettes balancing local density and global diversity
- Achieves more robust identifier assignments and interpretable token semantics, especially beneficial for long-tail item recommendations
Why It Matters
Enables e-commerce and ad platforms to achieve 15-20% higher recommendation accuracy while better handling niche products.