Integrity Review Report: BACH: Bridging Adjacency List and CSR Format using LSM-Trees for HGTAP Workloads (DOI: 10.14778/3718057.3718076)

Academic fraud report · Geng Detector

Summary

Verdict: Cleared (No evidence of academic misconduct detected). This PVLDB 2025 systems paper by Huang, Cao, Ren, Wu, and Miao was reviewed across timeline consistency, code/reproducibility transparency, dataset arithmetic, figure plausibility, and internal methodological logic. No concerns were identified. Key supporting points: (1) timeline is internally consistent with 2025 publication and 2024-era references; (2) the authors released source code under a public GitHub repository (https://github.com/2600254/BACH) and the paper carries PVLDB artifact availability marking, which materially raises the cost of data fabrication in a systems paper; (3) dataset statistics in Table 3 (e.g., Datagen-8_5-fb: 332.0M edges / 4.6M vertices ≈ 72.17 avg degree vs. reported 72; UK-2007-05 consistent with undirected double-counting) are arithmetically self-consistent; (4) numerical distributions in Figures 6–8 exhibit realistic long-tail variance rather than fabricated uniformity. Limitations: pixel-level forensic analysis of figures was not possible due to text-only input.

Verdict

Cleared. No indicators of data fabrication, image manipulation, or methodological misconduct were identified. The paper passes timeline, reproducibility, dataset-arithmetic, and internal-logic checks.

Key findings

Timeline & hardware consistency: 2025 PVLDB publication aligns with cited 2024 literature (e.g., ChunkGraph, Spruce). The Intel Xeon E7-8860 platform is older but plausibly representative of academic server-class experiments, and is paired with 384 GB DDR4 and 2 TB SSD — a coherent configuration.
Reproducibility / open-source code: A public GitHub repository is provided (https://github.com/2600254/BACH) and PVLDB records source-code availability. Systems papers with artifact-evaluated code have substantially higher fabrication costs.
Dataset arithmetic check (Table 3): Datagen-8_5-fb: 332.0M / 4.6M ≈ 72.17 vs. reported avg degree 72. UK-2007-05: 3.3B / 105.9M ≈ 31.1; doubled to ≈62.2 for undirected storage matches the reported 62.8 within reasonable rounding tolerance.
Numerical realism in Figures 6–8: Reported P99 latencies (e.g., 23, 93, 1371) and throughput numbers show genuine long-tail variance inconsistent with the uniform patterns typical of fabricated datasets.
Internal methodological honesty: The authors openly exclude LLAMA from deletion-heavy comparisons, citing documented limitations ("inability to accurately process edge deletions on disk"), and present balanced tradeoffs between Tiering, Leveling, and Elastic merge policies — consistent with rigorous systems-paper practice rather than selective reporting.

Evidence highlights

DOI: 10.14778/3718057.3718076 (PVLDB 2025).
Table 3 dataset cross-checks:
Datagen-8_5-fb (DA): 4.6M vertices, 332.0M edges → computed avg degree 72.17 vs. reported 72.
UK-2007-05: 105.9M vertices, 3.3B edges → computed 31.1; undirected doubling ≈ 62.2 vs. reported 62.8.
Artifact availability statement and GitHub URL: https://github.com/2600254/BACH.
Hardware configuration (Section 8.1): Intel Xeon E7-8860, 384 GB DDR4, 2 TB SSD.
Justified comparison exclusions (Section 8.2/8.3) referencing LLAMA's documented disk-deletion handling.

Notes

Limitations of this review:
Figure forensics (PS-level noise analysis, splicing detection, clone-region detection) on Figures 6, 7, and 8 could not be performed because only text was available; conclusions on figures are based on reported numerical distributions only.
No statistical re-analysis (e.g., digit-preference, Benford-style checks on full raw tables) was attempted beyond the targeted arithmetic spot checks above.
Reviewer is an AI assistant; results are advisory and do not constitute a formal institutional investigation.
No further action recommended.