技能 数据科学 学术研究可复现性评估指南

学术研究可复现性评估指南

v20260724
vldb-artifact-evaluation
本指南详细介绍了为学术会议(如VLDB)准备可复现性研究成果的方法论。它指导作者系统地构建包含原型代码、输入数据、工作负载和分析脚本的完整技术包。目的在于确保外部评估者能够独立、准确地重现并验证研究中的所有核心结果,从而提高论文的学术可信度和可复现性。
获取技能
307 次下载
概览

VLDB Artifact Evaluation

Use this once a PVLDB paper is accepted (or when an EA&B submission is being planned, since participation is not optional there). The pVLDB Reproducibility Evaluation — run jointly with SIGMOD's effort since the 2018 push — has committee members rebuild your experiment from your package. Two distinct prizes exist: the ACM availability badge for sharing, and the Reproducible outcome (with a Best Reproducible Paper Award) for surviving an independent rerun.

Who must play, who should

Situation Obligation
EA&B paper Required: release all data and software, submit to evaluation
Regular research paper Optional but strongly encouraged; badge on offer
Industrial paper with proprietary core Availability of what can be shared; document the rest
Vision paper Rarely applicable

The four surfaces evaluators rebuild

The committee's published expectations decompose an artifact into four layers. Package each one explicitly:

  1. Prototype — source code, build environment, configuration. A container image plus the Dockerfile that produced it is the community's default.
  2. Input data — the datasets themselves, or deterministic generators with pinned seeds and a size knob, plus download scripts for public corpora.
  3. Workload — the exact queries, client configuration, thread counts, and run durations behind every experiment, not a representative sample.
  4. Analysis — scripts that transform raw measurements into each numbered figure and table in the PDF. This layer is the one authors most often skip and evaluators most often need.

Design for a stranger's machine

  • Assume the evaluator has no access to your cluster. Provide a scaled-down mode that demonstrates every claim's shape on one commodity machine, and document how the full-scale numbers were obtained.
  • Pin everything: base images, package versions, competitor-system commits. "Latest" is where reruns go to die.
  • Emit expected outputs and tolerances. A rerun that produces a plot is only useful if the evaluator can tell whether the plot is right — state which qualitative relationships must hold even when absolute numbers shift with hardware.
  • Time-box honestly: state wall-clock cost per experiment so the committee can schedule, and mark the one experiment that best represents the paper if resources run short.

Minimal package skeleton

artifact/
  README.md          # claims map: figure/table -> command -> expected shape
  Dockerfile         # or image reference + build recipe
  data/get_data.sh   # fetch or generate, seeded
  workloads/         # exact configs per experiment
  run_one.sh <exp>   # single experiment, scaled-down default
  run_full.sh        # full-scale protocol, hardware stated
  plots/make_all.sh  # raw results -> paper figures

Award positioning

Winning packages read like engineering products: one command to a first result, claims mapped to figures, failures anticipated. If the evaluation report will say "worked on the first try," you are in contention; if it says "worked after correspondence with the authors," you got the badge and lost the award. Current-cycle evaluation logistics and criteria wording: 待核实 on vldb.org/pvldb/reproducibility before packaging.

Output format

[Track] EA&B-mandatory / voluntary / availability-only
[Surface coverage] prototype / data / workload / analysis — gaps listed
[Stranger test] scaled-down mode exists / cluster-only (risk)
[Pinning] images, versions, competitor commits — unpinned items
[First-command experience] <what happens>
[Fixes before submission to the committee] <ordered>
信息
Category 数据科学
Name vldb-artifact-evaluation
版本 v20260724
大小 4.01KB
更新时间 2026-07-29
语言