SoroushVahidi/ranking-fas-results

Ranking FAS Results

Ranking FAS Results v1 is a compact, metrics-only dataset of project-generated ranking-method comparison results for ranking from pairwise comparisons. This is the canonical public Hugging Face repository for version v1.

The released rows summarize experiments from the ranking-by-feedback-arc-set project comparing feedback-arc-set / acyclic-graph-construction ranking methods against classical ranking baselines and GNNRank-family neural methods on directed weighted pairwise-comparison benchmark instances.

This release does not contain raw graph files, adjacency matrices, edge lists, node identifiers, raw benchmark records, GNNRank preprocessed objects, result arrays, notebooks, prompts, private logs, or manuscript/reviewer material.

Why this dataset matters

Researchers can use this dataset to inspect machine-readable method comparisons, runtime/quality tradeoffs, timeout and missingness behavior, classical-vs-neural-vs-training-free comparisons, and the table-level evidence behind the associated paper. It avoids rerunning many classical and GNN baselines while preserving non-winning methods, missingness, timeouts, and runtime tradeoffs.

Configurations

Config	Rows	Meaning
method_results	1,468	Primary per-dataset × method × configuration result matrix.
dataset_inventory	81	Canonical benchmark-instance inventory with public-safe graph-size metadata.
method_summaries	109	Full-suite and compute-matched per-method/config aggregate summaries.
missingness	56	Runtime/metric coverage, timeouts, and finance-status evidence by method/config.
best_in_suite	81	Per-dataset best OURS/classical/GNN comparisons.
runtime_tradeoff	79	Per-dataset runtime, speedup, and Pareto comparison evidence.

The canonical suite contains 81 dataset identifiers in the current source set. Historical alternate/root outputs and archived legacy files in the source repository are intentionally excluded from v1.

Method families

• ours_fas_style: Soroush Vahidi's feedback-arc-set / acyclic-graph-construction ranking methods: OURS_MFAS, OURS_MFAS_INS1, OURS_MFAS_INS2, OURS_MFAS_INS3.
• classical_ranking: classical ranking baselines such as SpringRank, syncRank, serialRank, btl, davidScore, PageRank, rankCentrality, SVD_RS, and SVD_NRS.
• gnnrank_family_neural: neural methods from the GNNRank experimental framework, including DIGRAC, ib, and btlDIGRAC.

Original method IDs are preserved. Normalized family fields are added only for readability and do not replace the experimental implementation names.

Metric definitions

• upset_simple: primary upset/violation metric from the source pipeline. Lower is better.
• upset_naive: alternate upset metric from the source pipeline. Lower is better.
• upset_ratio: ratio-style upset metric from the source pipeline. Lower is better. Missing values are preserved where the source has no valid value.
• runtime_sec: measured runtime in seconds for the corresponding method/configuration aggregate where available. Lower is faster.
• timeout_flag: whether the row corresponds to a timeout or missing-runtime condition in the leaderboard source.
• coverage: source coverage string such as 77 / 81.
• coverage_n and coverage_denominator: parsed numeric coverage components added for easier analysis.
• valid_metrics: number of datasets with valid metric values for a method/configuration in the missingness table.
• valid_runtime: number of datasets with valid runtime values.
• timeouts_or_missing_runtime: count of timeout or missing-runtime cases.
• speedup_baseline_over_ours: baseline runtime divided by OURS runtime in the runtime-tradeoff table. Larger means the listed baseline is slower relative to OURS; values below 1 mean the baseline is faster.
• ours_not_dominated: Boolean indicator from the source runtime/Pareto table; true means the listed OURS result is not dominated under the table's quality/runtime comparison. The original source table stores this as a 0/1 indicator; the public Parquet schema represents it as Boolean because that is the column's logical type.

The summary configs report medians and means exactly as in the canonical source tables. Missing values and timeout-related rows are preserved because they are scientifically meaningful.

Data dictionary

method_results

Column	Meaning
dataset	Canonical benchmark dataset identifier from the source leaderboard.
family	Benchmark family from the canonical inventory.
method	Original experimental method ID.
method_family	Normalized method family: ours_fas_style, classical_ranking, gnnrank_family_neural, or other_or_unknown.
config	Original source configuration string. Preserved for provenance.
upset_simple, upset_naive, upset_ratio	Ranking upset metrics; lower is better.
runtime_sec	Runtime in seconds where available.
timeout_flag	Source timeout indicator.
source	Public-safe source identifier; v1 rows come from result_arrays summaries, not raw arrays.

All-null source columns kendall_tau, seed, and trial_id were omitted from method_results because they contained no information in the canonical table.

dataset_inventory

Column	Meaning
dataset	Canonical benchmark dataset identifier.
family	Benchmark family.
n_nodes	Number of graph vertices/nodes when available.
m_edges	Number of directed weighted comparison edges when available.
density_directed_no_self_loops	m_edges / (n_nodes * (n_nodes - 1)) when computable.
notes	Public-safe source note, if present.

The source adj_artifact path field is not published because it is a local/source artifact path and raw adjacency files are excluded.

method_summaries

Rows summarize method/config performance across either full_suite or compute_matched scope. compute_matched applies the source runtime filter runtime_sec_mean <= 1800. Lower upset metrics are better. coverage, coverage_n, and coverage_denominator describe how many datasets contribute valid results.

missingness

Rows report valid-metric count, valid-runtime count, timeout/missing-runtime count, and finance-specific status for each method/configuration. This config exists to make failures and exclusions visible rather than silently dropping them.

best_in_suite

Rows compare the best OURS, classical, and GNNRank-family methods per dataset according to upset_simple. Lower values are better. Missing values indicate no valid method result for that group on that dataset.

runtime_tradeoff

Rows compare a selected OURS method against a baseline method on each dataset, including runtime, speedup, and not-dominated indicator. Runtime is in seconds.

Ownership and provenance

Soroush Vahidi is the author/releaser of this derived metrics dataset. The contribution released here consists of experiment execution/curation, OURS/MFAS-style method results, result aggregation, paper-facing comparison tables, and this public metrics release.

This dataset builds on the GNNRank framework and benchmark ecosystem. GNNRank, its neural baselines, code structure, and preprocessed benchmark ecosystem are third-party work by Yixuan He and collaborators and must be cited separately where relevant.

The original benchmark datasets remain the work of their respective creators. This release does not claim ownership of those raw datasets and does not redistribute or relicense them.

Associated papers

Primary associated paper

Soroush Vahidi. “Learning-Free Ranking from Pairwise Comparisons via Feedback-Arc-Set Pruning and Add-Back.” Research Square preprint, DOI: 10.21203/rs.3.rs-9281720/v1.

The public Research Square PDF currently uses the title variant “Scalable and Training-Free Ranking from Pairwise Comparisons via Acyclic Graph Construction.” This dataset supports the ranking-method comparison results and metrics associated with that work.

Related paper

Soroush Vahidi and Ioannis Koutis. “Minimum Weighted Feedback Arc Sets for Ranking from Pairwise Comparisons.” arXiv:2412.16181. DOI: 10.48550/arXiv.2412.16181.

This is related MWFAS/ranking work, not the primary dataset citation for this release.

Upstream framework paper

Yixuan He, Quan Gan, David Wipf, Gesine Reinert, Junchi Yan, and Mihai Cucuringu. “GNNRank: Learning Global Rankings from Pairwise Comparisons via Directed Graph Neural Networks.” ICML/PMLR, 2022.

Citation policy

Cite the dataset when using the released result rows or metrics.

Cite the primary associated paper when discussing or relying on the methodology or scientific results described there.

Cite both when using this dataset and materially relying on the associated methodology/results.

Cite GNNRank and relevant upstream benchmark papers when using or discussing the upstream framework/data context.

Recommended dataset citation:

Vahidi, Soroush. Ranking FAS Results, version v1. Hugging Face dataset SoroushVahidi/ranking-fas-results, 2026. License: CC BY 4.0. Source repository: https://github.com/SoroushVahidi/ranking-by-feedback-arc-set.

BibTeX:

@dataset{vahidi2026rankingfasresults,
  title        = {Ranking FAS Results},
  author       = {Vahidi, Soroush},
  year         = {2026},
  version      = {v1},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/SoroushVahidi/ranking-fas-results},
  note         = {Derived metrics dataset; no dataset DOI assigned}
}

License

This release uses CC BY 4.0 for the project-generated result metrics, derived summaries, release metadata, and documentation to the extent controlled by this project.

CC BY 4.0 does not relicense GNNRank, original ranking datasets, graph/preprocessed benchmark files, third-party code, or upstream source material.

How this dataset differs from existing resources

GNNRank and related repositories provide code, preprocessed data, and experiment infrastructure. Upstream benchmark repositories provide ranking/graph data. Papers usually report selected aggregate tables.

This release provides a compact machine-readable result matrix of project-generated comparisons across OURS/FAS-style, classical, and GNN methods, including runtime and missingness evidence. It is not a raw benchmark-data mirror.

Difference from SoroushVahidi/mwfas-heuristic-metrics

SoroushVahidi/mwfas-heuristic-metrics covers graph-optimization heuristic runs on a different benchmark ecosystem. Ranking FAS Results covers ranking-from-pairwise-comparisons experiments and comparisons against classical/GNN ranking methods. Neither dataset supersedes the other.

Reproducibility

• Source repository: https://github.com/SoroushVahidi/ranking-by-feedback-arc-set
• Source commit: 706b21771be3af61c8fd4dd22723d0d5611fa042
• Canonical inputs: listed in metadata/provenance_summary.json.
• Raw graph/benchmark data and raw result arrays are excluded from this release.