技能 数据科学 研究可复现性标准指南

研究可复现性标准指南

v20260724
edbt-reproducibility
本指南提供了学术论文(尤其数据库系统领域)提升研究可复现性的最佳实践。它指导作者如何记录数据来源、环境参数、实验流程和证据链,确保论文中的每一个结论都可以被其他研究人员通过可运行的代码和数据重新验证,提高研究的严谨性和透明度。
获取技能
372 次下载
概览

EDBT Reproducibility

Use this before submission and again before camera-ready. EDBT's community has a reproducibility-forward culture, and the published record is open access on OpenProceedings — so an inspectable, re-runnable package raises a paper's standing and, for an Experiments & Analysis paper, is the contribution. The goal is that a competent reader could rebuild your measurements and reach your conclusions.

Evidence map

  • Map each claim, mechanism, and reported number to a verifiable location — a paper section, a table generated from a logged run, or a script in the artifact.
  • For a mechanism, give enough of the algorithm, data structures, parameters, and system integration that a reader could reimplement or rebuild it.
  • For an evaluation, report workloads and their derivation, dataset versions and sources, the measurement harness, metrics, and the analysis scripts.
  • Keep the availability statement truthful and specific: what is shared, the workloads and data, the hardware assumptions, and — if something cannot be shared — exactly why.
  • Keep the paper and the artifact consistent: a number in the PDF that no script produces is the contradiction reviewers read as carelessness.

Availability statement audit

Claim in the paper Weak availability answer EDBT-ready answer
"We evaluate on workload W" "Data available on request" Archived workload/query-log derivation + the extracted data or a documented access path
"Our operator lowers latency" "Code will be released" Runnable system/prototype with a build, a demo run, and the config
"We compare N systems" (E&A) Numbers with no harness The full comparison harness that regenerates every table
"On a 128-node cluster" Nothing about environment Hardware/cluster spec, engine build/commit, and how to size a smaller reproduction

"Available on request" is treated as not available; convert every such line into a concrete package or an explicit, justified exception (licensing, confidentiality).

Provenance pinning (database-systems flavor)

[Data]       pin dataset versions and sources; archive the derived workload/query-log, not just a
             description; document filtering and sampling
[System]     record the engine/prototype build or commit; ship a build recipe or container
[Environment] state hardware, memory, network, and node counts; note what a smaller reproduction changes
[Harness]    the measurement scripts that produce each table/figure, with fixed configuration
[Randomness] log seeds for any stochastic step; say what is and is not deterministic

Degrees of reproducibility (state the one you achieved)

  • Turnkey: one documented command (or container) regenerates each table/figure from a run or from logged results.
  • Scripted: scripts exist but require documented manual steps, a specific cluster, or external data access.
  • Descriptive: prose detailed enough that a competent reader could rebuild the pipeline.

For EDBT, aim turnkey for anything a reviewer might re-run quickly (a demo run on a small workload, a plot from logged results); large-cluster or licensed-data experiments may stay scripted with the environment and access clearly documented. Stating the achieved level honestly beats promising turnkey behavior that fails on a clean machine.

Vignette: a distributed-operator study

Consider an operator evaluated on a cluster. Its reproducibility spine: a container or build recipe for the engine plus the operator; the workload-derivation scripts with pinned dataset versions; the measurement harness that runs the operator and the tuned baseline across node counts; the logged raw results; and the analysis notebooks that turn them into the paper's tables — plus one honest sentence about the parts (the full 128-node run, a licensed dataset) that a reader reproduces at reduced scale and why.

Consistency and camera-ready pass

  • Before submission: every reported number traces to the artifact; the availability statement matches reality; if the cycle is double-blind, the artifact carries no identity strings.
  • Before camera-ready: deposit the package in a DOI-issuing archive (Zenodo, figshare, Software Heritage) with an OSI-approved license, replace any anonymized links with the permanent ones, and make the statement consistent with the open-access OpenProceedings record (edbt-artifact-evaluation, edbt-camera-ready).

Output format

[Claim inventory] <claim -> evidence location>
[Availability] concrete / vague / missing
[Provenance gaps] <dataset versions / engine build / environment / seeds>
[Reproducibility level] turnkey / scripted / descriptive, stated honestly
[Paper fixes] <must appear in the PDF>
[Artifact fixes] <additions before upload>
信息
Category 数据科学
Name edbt-reproducibility
版本 v20260724
大小 5.12KB
更新时间 2026-07-28
语言