Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)¶
L1 mechanism diagnostic. User-approved academic/Evolve exception on 2026-10-02; no online A/B or industrial evidence claim.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv v1 |
| Affiliations | Huawei Ireland Research Centre |
| First public release | 2026-10-01 (arXiv v1) |
| Authors' code | graph-informed-sids |
| Adapter / method | gris |
| Local code | src/auto_research/reproductions/gris/ |
Original paper¶
Background and contribution¶
Build semantic IDs by hierarchical graph partitions. Nodes carry content semantics and edges collaboration. Empty graphs reduce to content quantization.
flowchart LR
I[Content and graph] --> P[Graph-aware binary partitions] --> S[Semantic IDs]Key figure from the paper¶
Original GrIS method figure. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\(SID(i)=RecursivePartition(G,X)_i\). Each layer combines smoothed semantics and adjacency.
Paper results¶
Up to 52% higher Hit@10 than CF-aware state of the art across real datasets, as reported by the authors.
Local reproduction¶
Compare disabled/default versus enabled. Three-seed diagnostics verify transitions and invariants, not paper-scale effects (diagnostic_only=true).
Reproduction limits¶
The graph-informed binary hierarchy executes. The exception is not production A/B evidence.