andresnowak/gcib-icml2026-reproduction-evidence
GCIB ICML 2026 reproduction evidence Machine-readable evidence for a scaled reproduction of GCIB: Graph Contrastive Information Bottleneck for Multi-Behavior Recommendation. The audit uses the processed Yelp matrices from the official implementation at akajinchen/GCIB@a887988. The substantive local run sampled 1,024 evaluation users (10.24% of the released Yelp test users), 3,000 candidate items, 14,585 target-training edges, and 13,767 auxiliary edges. It tested the intended… See the full description on the dataset page: https://huggingface.co/datasets/andresnowak/gcib-icml2026-reproduction-evidence.
GCIB ICML 2026 reproduction evidence
Machine-readable evidence for a scaled reproduction of GCIB: Graph Contrastive Information Bottleneck for Multi-Behavior Recommendation.
The audit uses the processed Yelp matrices from the official implementation at `akajinchen/GCIB@a887988`. The substantive local run sampled 1,024 evaluation users (10.24% of the released Yelp test users), 3,000 candidate items, 14,585 target-training edges, and 13,767 auxiliary edges. It tested the intended differentiable edge mask, the released detached-mask implementation, component ablations, 20% tip-edge noise, β sensitivity, and global-encoding depth.
Hugging Face GPU Jobs were attempted but unavailable: the authenticated canary was rejected with 403 missing job.write. The published Trackio logbook records the exact scope, commands, results, and limitations.
Files:
gcib_large_mps_results.json: primary MPS histories and best metrics.paper_claim_audit.json: deterministic Table 2/4/5/6 arithmetic and source audit.- CSV files: raw data underlying the logbook figures.
repro_gcib.py: independent scaled mechanism implementation.audit_paper_claims.py: deterministic claim/source audit.make_figures.py: figure and raw-table generator.