TCMaster: Confidence-Aware Querying and Workload-Guided Physical Design for Multi-Source Traditional Chinese Medicine Knowledge Graphs
Zheng Chen, Yuzhu Li, Haoxuan Li, Haoxuan Li, Zhongde Zhang, Lianshun Jin, Peiwu Qin, Peiwu Qin
cs.AI
Quick verdict
TCMaster: confidence-aware querying + skew-aware physical design on a 221K-node TCM KG; honest, modest gains; 20pp QA win rests on a self-built benchmark.
Summary
TCMaster is a property-graph query substrate over a Traditional Chinese Medicine knowledge graph built from pharmacopoeias, HERB 2.0/SymMap molecular data, 11,845 prescriptions, and DeepSeek-V3-extracted micro-semantics (221,225 entities, 722,671 base edges, five layers). Every edge carries a source-level confidence (0.95 authoritative down to 0.30–0.60 predicted) plus a provenance label; Cypher templates are rewritten with confidence predicates, and multi-hop paths are ranked under PRODUCT, MIN, or weighted-average aggregation. Workload-guided physical design exploits ontology skew via traversal-direction selection, 28-bit herb-attribute bitmaps, and 28.75M materialized Prescription→Target shortcut edges. On Neo4j 4.4, latency across seven clinical workloads is 5.2–12.0 ms; direction selection speeds attribute lookup 1.47×, shortcuts speed high-fanout target counting 4.42× (51.19→11.57 ms), while bitmaps (1.03×) and shortcut-plus-confidence lookup (0.65×, a slowdown) show limits the authors report candidly. Filtering at θ=0.18 removes 39.3% of paths, but only on the heterogeneous 3-hop workload. Downstream KG-RAG with DeepSeek-V3 improves accuracy on the authors' own TCMbench v3 split (2,000 questions) by 20.0 pp (72.5% vs 52.5%), with Wilson intervals. Main limitations: hand-set, uncalibrated confidence constants; the QA benchmark is self-built and named confusingly like the external TCMbench; latency lacks variance reporting; shortcut storage costs (roughly 40× base edges) are unquantified; evaluation uses a single engine on one machine.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Per-edge source-level confidence with transparent Cypher predicate injection supports confidence-bounded multi-hop path queries at interactive latency (5.2–12.0 ms across seven clinical workloads, Table V).
- Exploiting ontology skew pays off selectively: direction selection gives 1.47× on attribute lookup and 28.75M materialized shortcuts give 4.42× on high-fanout target counting (51.19→11.57 ms), while bitmaps and shortcut-plus-confidence composition do not beat Neo4j's planner (Table VI).
- Threshold filtering is workload-specific: at θ=0.18 only the heterogeneous 3-hop workload Q3 loses paths (39.3%), while authoritative-only workloads are unaffected (Table IX).
- KG-RAG over TCMaster improves accuracy on the authors' TCMbench v3 split by 20.0 pp (72.5% vs 52.5%) using DeepSeek-V3 with full-recall θ=0 retrieval (Table VIII).
Strengths
Candid reporting of negative and boundary results
Table VI shows bitmap pruning at only 1.03×, shortcut top-k lookup at 1.17×, full enumeration at 1.19×, and a 0.65× slowdown (10.73→16.45 ms) for shortcut-plus-confidence composition; §VII-K explicitly frames the system boundary instead of claiming a universal optimizer.
Strong reproducibility engineering
§VI and the Artifact section specify fixed seeds (42/123/456), pinned query templates, MD5-seeded subsampling, checksums, a Neo4j 4.4 dump including the 28.75M shortcuts, and short verification scripts decoupled from the 36-hour ETL, with a README mapping tables to commands.
Operational clarity of the confidence semantics
Quantified data-quality auditing
Workload design resists cherry-picking
Weaknesses
Headline QA gain rests on a self-built benchmark
Uncalibrated, hand-set confidence constants
Weak baseline and single-engine evaluation
Physical-design costs under-quantified
Thin statistical treatment of latency
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TCMaster is deliberately positioned as a data-management substrate, not a clinical model or probabilistic database: a five-layer TCM property graph (221,225 entities, 722,671 base edges; §II, Table II) built from pharmacopoeias, HERB 2.0/SymMap snapshots, 11,845 prescriptions, and DeepSeek-V3-extracted micro-semantics. The paper claims three things: (1) confidence-bounded path querying — per-edge source-level confidence (Table III: 0.95 AUTHORITATIVE down to 0.30–0.60 PREDICTED) exposed via transparent Cypher predicate injection (§IV-C) and PRODUCT/MIN/WAVG path ranking (§IV-D); (2) workload-guided physical design for ontology skew — direction selection, 28-bit herb bitmaps, 28.75M materialized REACHES_TARGET shortcuts (§V); (3) an auditable, reproducible artifact. Headline numbers: 5.2–12.0 ms across seven workloads (Table V), 1.47× direction selection, 4.42× shortcut counting, 39.3% filtering at θ=0.18 (Q3 only, Table IX), +20.0 pp QA accuracy (Table VIII). Scope statements are unusually disciplined — the paper repeatedly states what the system is not.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TCMaster?
TCMaster is a query layer over a Traditional Chinese Medicine knowledge graph (about 221K entities and 723K base edges from pharmacopoeias, molecular databases like HERB 2.0 and SymMap, 11,845 prescriptions, and LLM-extracted facts). Every edge carries a source-level confidence score and provenance label; the system rewrites Cypher queries with confidence predicates and applies skew-aware physical design (traversal-direction selection, herb-attribute bitmaps, and 28.75M materialized Prescription→Target shortcut edges) on Neo4j 4.4.
How much faster is TCMaster than plain Neo4j?
It depends on the workload, which the paper reports openly. Direction selection speeds attribute lookup 1.47× (9.40→6.41 ms). Materialized shortcuts speed counting of all reachable targets for high-fanout prescriptions 4.42× (51.19→11.57 ms), but only 1.17× for top-k lookup and 1.19× for full enumeration; adding a confidence predicate to shortcut lookup is slower (0.65×); bitmaps give about 1.03×. Overall latency for the seven tested workloads is 5.2–12.0 ms.
Does confidence filtering lose useful results?
Mostly no for trusted data: at threshold θ=0.18 only the heterogeneous 3-hop workload (Q3, mixing curated, molecular, and predicted evidence) loses paths — 39.3% of them — while workloads traversing authoritative or curated relations keep all results. Above θ=0.35 all Q3 paths are removed. The downstream QA experiment used θ=0 (no filtering), so filtering and answer accuracy are evaluated separately.
Is the claimed 20-point QA improvement on a standard public benchmark?
No. It is measured on 'TCMbench v3', the authors' own 2,000-question clinical evaluation split built from their artifact data, with DeepSeek-V3 as the reasoning model: 72.5% for KG retrieval versus 52.5% for the no-retrieval baseline (Wilson intervals reported). External TCM benchmarks are cited for context only, which a rigorous reader should weigh when interpreting the headline gain.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–6 are described in captions and prose but not analyzable from the extracted text; all figure-derived numbers were taken from the surrounding narrative.
- The supplemental artifact (Neo4j dump, scripts, checksums) is referenced but not verifiable from the text; reproducibility is assessed as stated, not as tested.
- The paper appears complete (abstract through references), but supplementary appendices with extended logs may exist outside the extracted text.
- arXiv preprint (v1, 22 Sep 2026), not yet peer reviewed; author affiliations and AI-use disclosure noted but not evaluated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.