Research area

User reactivation, generative recommendation, Semantic-ID interface diagnostics, dynamic retrieval, graph embedding, counterfactual watch-time prediction, and efficient CTR modeling form the main technical line.

Publication record

Authors

Jiandong Ding, Heng Chang, Huijie Qin, Tianying Liu

Venue

Proceedings of the ACM International Conference on Information and Knowledge Management (CIKM '26)

Year

2026

Area

Recommender Systems

Problem

Exported item-to-code mappings from Semantic-ID tokenizers usually lack a shared inspection interface, so coverage gaps, aliasing, weak prefixes, tail compression, and fan-out problems tend to surface only after downstream recommendation training has already paid for them.

Approach

SIDInspector is mapping-first. An adapter contract covers item mappings, metadata, interactions, and generator traces, and probes report utilization, aliasing, neighborhood alignment, popularity allocation, and structural cost. It inspects the interface itself rather than a trained recommender.

Public evidence

The release evaluates four tokenizer families, including GRID-style RQ-KMeans, ReSID/GAOQ, LETTER, and LC-Rec artifacts. Reported measurements show D2 aliasing dropping substantially under added capacity and reaching zero for ReSID/GAOQ, while D3 stays below a deterministic category-prefix control. These are diagnostic measurements of the mappings, not evidence of recommendation-quality gains.

Takeaway

Capacity, addressability, and prefix alignment are distinct, measurable properties of a Semantic-ID interface. Checking them before method-level evaluation or reuse avoids attributing downstream failures to the wrong layer.