Skills Artificial Intelligence Ensuring Reproducibility In AI Research

Ensuring Reproducibility In AI Research

v20260724
cikm-reproducibility
This comprehensive guide outlines best practices for achieving rigorous reproducibility in complex AI and knowledge graph research. It details how to pin every component—from tokenizers and KGs to LLM prompts and environment settings—ensuring that results are fully documented and verifiable by peers, even when proprietary data is involved.
Get Skill
176 downloads
Overview

CIKM Reproducibility

Reproducibility at CIKM has a venue-specific difficulty: the typical paper chains components from different communities — an index, a graph, a model, a ruleset — and each link has its own silent-divergence habits. A reader who cannot rebuild the chain cannot attribute the result, and a blended review panel contains someone able to notice each weak link.

Divergence map for chained pipelines

Chain link How results silently drift Pin
Text preprocessing / indexing Tokenizer versions, stopword lists, index-time defaults differ across toolkits Name toolkit + version + config file in the artifact
KG snapshot Public KGs (Wikidata-class) change daily; entity counts drift Freeze and state the dump date; ship the extracted subgraph if licensable
Candidate generation Recall stage caps and thresholds rarely reported Report every cutoff; they bound the final metrics
Training Seeds, hardware nondeterminism, early-stopping criteria Seed policy + selection rule in the protocol paragraph
Evaluation Metric implementations disagree at tie-breaking and cutoffs Name the evaluation library version; never hand-roll silently
LLM components Model version/API drift; prompts unlogged Pin model identifiers and dates; log prompts verbatim in the artifact

The discipline: for each link, either the artifact pins it or the paper states it. A link pinned nowhere is where a failed replication will land.

Unreleasable data, releasable knowledge

CIKM's KM lane routinely involves enterprise corpora, clickstreams, or proprietary KGs that cannot ship. The venue-honest pattern:

  • Describe the unreleasable data statistically (size, schema, class balance, collection window) at a level where a reader could construct a synthetic analog — then actually provide that analog generator when feasible.
  • Run the public-data variant of every headline experiment, even if the effect is smaller; a result that exists only on invisible data asks the panel for faith.
  • State the release position explicitly in the paper ("logs cannot be released; the sampling script and schema are in the artifact") rather than leaving the reader to discover the gap.

GenAI disclosure as a reproducibility document

CIKM 2026's mandatory GenAI Usage Disclosure covers code and data, not just prose (source map, 2026-07-08). Treat it as part of the methods record: if evaluation scripts, synthetic data, prompts, or labels were generated with AI assistance, the disclosure plus the artifact should together let a reader judge what that implies for the result. A disclosure that says "AI used for coding" while the artifact contains unexplained generated labels is an inconsistency automated compliance checks — which the conference reserves — or reviewers can catch.

Environment capture

Chained pipelines multiply environment surface, so capture it in layers:

Layer Capture mechanism
OS + system libraries Container image or a documented base image tag
Language environments Lockfiles (exact versions), not loose requirement ranges
Toolkit configs The actual config files, committed — not "default settings" prose
Data inputs Checksums + download scripts, or the frozen extraction (see KG row)
Hardware assumptions GPU/CPU class and memory floor stated where results are timed

The test is transferability: a lab-mate on a clean machine, without the authors in the room, reaches the headline table. Running that internal replication before submission is the single highest-yield reproducibility exercise — it finds the unpinned link while it can still be pinned.

Where reproducibility pays at this venue

Three concrete CIKM payoffs beyond principle. First, the blended panel: whichever lane doubts the result will probe its own link of the chain, so pinning every link is defense in all three directions. Second, resource-track reviewers and readers judge adoptability, which is reproducibility wearing its public face (cikm-artifact-evaluation). Third, follow-up work: CIKM's back catalog shows methods becoming standard baselines (DRMM, BERT4Rec); papers get that afterlife only when third parties can run them — the reproducible version of a method is the one that accumulates citations as a baseline.

Release timeline

Anonymized review artifact during submission (see cikm-supplementary for what the budget permits); public repository at camera-ready, with license, versioned release tag, and the exact commit that produced the proceedings numbers. The 2026 notification-to-camera-ready window is thirteen days — build the release during the review wait (cikm-workflow Mode A), not inside that window.

Honest-failure disclosure

Chained pipelines rarely reproduce perfectly, and the venue-credible move is to say so first: a REPRODUCING.md that states which numbers regenerate exactly, which vary within a stated tolerance (GPU nondeterminism, sampling), and which depend on restricted inputs and therefore only regenerate in public-analog form. Declared tolerance reads as competence; discovered variance reads as concealment. The same document is where to state known environment sensitivities ("results verified on CUDA X; version Y shifts Table 3 by ±0.2") — the sentence that saves a replicator a week is the sentence that earns the citation.

One-command bar

# The replication target for a CIKM chained pipeline:
git clone <repo> && cd <repo>
make setup          # pinned environment, data download or synthetic analog
make table2         # rebuilds the headline table end-to-end from the frozen inputs

If make table2 cannot exist because data is restricted, the repo must say so at the top and offer the public-variant target instead. Silent partiality — a repo that looks complete but is not runnable — costs more reviewer goodwill than an honest scope statement.

Output format

[Chain audit] <link → pinned where (paper / artifact / nowhere)>
[Data position] <releasable / described+analog / public-variant-only>
[Disclosure consistency] <GenAI section vs. artifact contents>
[Release plan] <review artifact state → camera-ready repo state, dated>
[Weakest link] <the divergence a replicator would hit first>
Info
Name cikm-reproducibility
Version v20260724
Size 6.57KB
Updated At 2026-07-28
Language