SoroushVahidi/mwfas-heuristic-metrics

MWFAS Heuristic Metrics

MWFAS Heuristic Metrics v1 is a project-generated, metrics-only dataset (7,371 rows across 14 tables) reporting how heuristic algorithms perform on the Minimum Weighted Feedback Arc Set (MWFAS) problem: given a directed graph whose arcs each have a weight, find the minimum-weight set of arcs whose removal makes the graph acyclic. The problem is directly connected to ranking from pairwise comparisons — a weighted directed edge can represent one item being preferred over another, so minimizing the removed weight corresponds to finding an ordering with the smallest total weighted disagreement. This release covers run-level and aggregate outcomes (objective quality, runtime, robustness, and ablations) for the IPSNS/LR-TA/WMSF heuristic family and several baselines, evaluated on graph instances from the third-party alidasdan/graph-benchmarks collection and other third-party sources (not redistributed here).

This dataset does not include raw third-party graph benchmark files, graph edge lists, adjacency lists, raw ranking vectors, manuscript drafts, logs, or machine-local execution artifacts.

Canonical Hugging Face repository: SoroushVahidi/mwfas-heuristic-metrics.

One row represents: one sanitized experimental outcome — but the exact grain (a single algorithm run, a single instance-level ablation, or a pre-aggregated summary) differs by config. See the table below and the per-config descriptions further down.

Why this dataset exists

Reproducing the full MWFAS heuristic-evaluation pipeline (parameter tuning, sensitivity sweeps, stochastic robustness runs, MIP baselines, and ablations across dozens of benchmark instances) is compute-intensive. This dataset publishes the resulting run-level and aggregate metrics directly, so researchers can audit, reanalyze, or build on the reported heuristic behavior without rerunning the experiments.

What is included

• 14 Parquet configs of project-generated run-level and aggregate metrics for MWFAS heuristic experiments.
• Public-safe graph structural metadata (n_vertices/n, n_edges/m, density) per instance.
• Sanitized instance and configuration identifiers, algorithm/variant labels, seeds, and objective/runtime outcomes.

What is not included

• Raw benchmark graph files, edge lists, or adjacency matrices (third-party; not redistributed).
• Temporary ranking vectors, raw output paths, local filenames, machine names, CPU/environment details, PIDs, or process metadata.
• Manuscript drafts, reviewer/editorial material, logs, and private build-audit files.
• The separate Ranking-by-FAS/GNNRank result matrices — those are a different project, published separately as SoroushVahidi/ranking-fas-results.

Dataset size

Config	Rows	Meaning
run_metrics	1,290	Stage-2 IPSNS parameter-grid tuning/holdout validation runs (COAP study).
sensitivity	140	Stage-1 one-at-a-time IPSNS parameter-sensitivity runs.
robustness	3,720	Repeated stochastic robustness runs (1,860 IPSNS + 1,860 DRMacIver/FAS) on a common sparse instance subset.
aggregate_summary	96	Aggregates (objective/runtime summaries, success/validation rates) derived only from the three configs above.
core_benchmark	369	Full DIMACS-style reproducibility run across the core benchmark suite; per-instance objective/runtime by algorithm. (described in AUDIT.md as the "123-instance DIMACS reproducibility run"; not detailed in the prior live README)
ablation	80	Per-instance runs varying algorithm_variant, recording backward_weight/runtime_seconds and status. [inferred from schema — verify exact ablation dimension with the author before publishing]
exact_small	67	Small instances with an exact solver objective (exact_bw) alongside LR-TA/WMSF/IPSNS objectives, gaps, and optimality flags — supports measuring heuristic optimality gap where exact solving is tractable. [inferred from schema]
external_baselines	984	Runs of external/third-party baseline algorithms (source_type, external_version_or_commit columns) for comparison against this project's heuristics. [inferred from schema]
lolib_dense	400	Heuristic runs on dense instances from the third-party LOLIB benchmark family (instance_family column; raw LOLIB files not redistributed). [inferred from schema]
budget_curve	120	Runs across a swept budget (iteration/time budget) parameter, showing objective/runtime as a function of budget. [inferred from schema]
plain_local_search	60	Plain local-search baseline runs (no SCC-local destroy-and-repair), with seed method, improvement-over-seed, and pass/move counters. [inferred from schema]
medium_mip_baseline	15	Medium-sized instances solved with a MIP solver (solver, mip_objective, mip_dual_bound_bw, mip_gap_pct, proven_optimal) compared against heuristic objectives. [inferred from schema]
application_case	6	A small set of application-motivated case-study instances (top_n column) with per-algorithm outcomes. [inferred from schema]
topological_extraction_sensitivity	24	Sensitivity of a topological-order-extraction rule (extraction_rule column) on the resulting backward weight/gap. [inferred from schema]
Total	7,371

Rows/configs marked [inferred from schema] were not described in the live README's prose and are not independently documented in the local audit trail (AUDIT.md covers only run_metrics, sensitivity, robustness, and core_benchmark in detail) — the descriptions above are inferences from column names only and should be checked against the source experiment scripts in ~/projects/minimum-weighted-fas-heuristics before publishing.

Problem background

A feedback arc set is a set of directed arcs whose removal makes a directed graph acyclic. In the weighted version, each arc has a weight, and the objective is to minimize the total weight of arcs that point backward under the produced ordering — lower is better. This connects to ranking from pairwise comparisons: a weighted directed edge can represent one item being preferred over another, and minimizing backward weight corresponds to finding an ordering with small weighted disagreement.

Important fields

• benchmark_resource: upstream benchmark collection referenced by the row. Raw benchmark files are not redistributed here.
• instance_family: sanitized family/group inferred from the source benchmark path (e.g. core, core-bad, iscas).
• instance_id: public benchmark-local instance identifier — not an edge list or graph encoding.
• n_vertices, n_edges (or n, m in some configs): graph structural metadata.
• density: directed density, n_edges / (n_vertices * (n_vertices - 1)), when applicable.
• algorithm / algorithm_variant: heuristic or baseline identifier, e.g. IPSNS or DRMacIver/FAS.
• objective_weight (or backward_weight/total_weight in some configs): weighted backward-arc objective; lower is better.
• normalized_objective: objective divided by total edge weight, when available.
• runtime_seconds (or runtime): measured runtime in seconds — implementation/environment-dependent, not a hardware-independent complexity measure.
• improvement_absolute / improvement_relative: improvement over the initial incumbent objective, where available.
• validated: boolean validation flag derived from source status and (for robustness rows) ordering/objective/acyclicity checks.
• source_commit: source repository commit associated with the record, where recoverable.

Full per-config column lists are in metadata/schema.json in the dataset repository.

Quickstart

from datasets import load_dataset

# Each table is a separate config — pick the one you need:
ds = load_dataset("SoroushVahidi/mwfas-heuristic-metrics", "run_metrics")
print(ds)

# Other available configs: "sensitivity", "robustness", "aggregate_summary",
# "core_benchmark", "ablation", "exact_small", "external_baselines", "lolib_dense",
# "budget_curve", "plain_local_search", "medium_mip_baseline", "application_case",
# "topological_extraction_sensitivity"

This dataset is small (7,371 rows total across all configs); no streaming is needed.

Research use cases

• Reproducing and auditing MWFAS heuristic experiment summaries without rerunning them.
• Analyzing runtime/objective tradeoffs across graph instances and configurations.
• Studying parameter sensitivity (sensitivity, topological_extraction_sensitivity, budget_curve) and stochastic robustness (robustness).
• Comparing heuristic objectives against exact-solver or MIP bounds where available (exact_small, medium_mip_baseline).
• Meta-analysis or meta-learning over algorithm outcomes using structural graph metadata as features.

Non-intended uses

• A replacement for the upstream graph benchmark datasets.
• A source of graph edge lists or raw ranking data.
• A universal benchmark of all FAS/MWFAS algorithms.
• A hardware-independent runtime leaderboard.
• Evidence that one method dominates outside the stated benchmark families and protocols.

Generation methodology

Results are drawn from git-tracked experiment directories in ~/projects/minimum-weighted-fas-heuristics (source repo, HEAD 40209c26966247d9bf9ad34764de4ac4181f98c2): experiments/coap_ipsns_holdout/results/runs.jsonl → run_metrics; experiments/coap_ipsns_sensitivity/summary/canonical_runs.csv → sensitivity; experiments/exp10_stochastic_robustness/summary/run_level_results.csv → robustness; results/tables/unified_reproducibility_summary.csv (123-instance DIMACS run, completed 2026-06-06) → core_benchmark; aggregate_summary is derived only from the public primary configs. The remaining 9 configs are built from other experiment directories in the same repository; see the note under Dataset size above.

Provenance and ownership

Artifact	Classification
All 14 published Parquet configs	Created/generated by Soroush Vahidi (derived run-level metrics from own experiments)
alidasdan/graph-benchmarks instances (referenced by instance_id/benchmark_resource)	Third-party upstream data; not redistributed
LOLIB dense instances (lolib_dense config)	Derived from third-party source data; raw LOLIB files not redistributed

Author: Soroush Vahidi, New Jersey Institute of Technology. ORCID: 0000-0003-1934-6282 (from this project's own CITATION.cff).

Relationship to related datasets

This is a graph-optimization / weighted-feedback-arc-set experimental metrics dataset. It is distinct from SoroushVahidi/lafc-evict (cache-eviction candidate supervision), SoroushVahidi/module-intervention-credit and SoroushVahidi/llm-serving-scheduler-baselines (LLM-serving scheduler data), SoroushVahidi/consistency-aware-judgments (IR pairwise LLM judgments), SoroushVahidi/frontier-allocation-metrics (budgeted LLM inference outcomes), SoroushVahidi/scidocs (a third-party BEIR mirror), and SoroushVahidi/lafc-evict-sample (a synthetic workflow artifact).

Most directly related — and most easily confused with — SoroushVahidi/ranking-fas-results:

	mwfas-heuristic-metrics (this dataset)	ranking-fas-results
Problem framing	Minimum Weighted Feedback Arc Set on general weighted digraphs	Ranking from pairwise comparisons
Benchmark ecosystem	alidasdan/graph-benchmarks DIMACS-style directed weighted graphs, LOLIB, others	GNNRank pairwise-comparison benchmarks (tournament-style data)
Methods measured	IPSNS / LR-TA / WMSF / DRMacIver-FAS heuristics, plus MIP/exact baselines	OURS_MFAS-family vs. classical ranking baselines vs. GNNRank neural methods
Primary metric	Feedback-arc-set objective weight and runtime	Ranking "upset"/violation rate
Row granularity	One heuristic algorithm run on one graph instance (varies by config)	One (dataset × method × config) comparison result

Neither dataset supersedes the other; there is no row-level overlap.

Limitations

• Runtime values depend on implementation and execution environment.
• The released metrics are historical results from specific code/protocol versions.
• run_metrics and sensitivity emphasize IPSNS parameter behavior, not a full cross-algorithm benchmark.
• robustness is quality-focused, not an equal-time comparison.
• Raw graph inputs are excluded; users must fetch upstream graph benchmarks separately to rerun algorithms from scratch.
• 10 of the 14 configs (see Dataset size above) lack prose documentation in the previously-published README; this draft's descriptions for them are schema-inferred and pending author verification.

Version history

• v1 (current), immutable data revision 5621684d3c03138d5b2ebe544e91ae5698e67b5a; repository metadata HEAD 653b415dcf53281730fd476c5307a371d1a792e2 (2026-08-18). This revision already supersedes an earlier 4-config baseline (2caabbe...) via 10 additive new configs — no further versioning action needed beyond the task_categories metadata fix documented here.

Related resources

• Source code repository: https://github.com/SoroushVahidi/minimum-weighted-fas-heuristics
• Primary associated paper: Soroush Vahidi and Ioannis Koutis, "Minimum Weighted Feedback Arc Sets for Ranking from Pairwise Comparisons," arXiv:2412.16181.
• Related supporting work (IPSNS/SCC-neighborhood component, not the primary dataset citation): Soroush Vahidi, "Incumbent-Protected SCC-Neighborhood Search for the Weighted Feedback Arc Set Problem," SSRN abstract 6281222.
• Related Hugging Face dataset: SoroushVahidi/ranking-fas-results (see "Relationship to related datasets" above).

Citation

Cite the dataset when using the released metrics. Cite the paper when discussing the methodology or scientific findings. Cite both when using the data and materially relying on the associated methodology/results.

Vahidi, S. (2026). MWFAS Heuristic Metrics: Run-Level Outcomes for Minimum Weighted Feedback Arc Set Experiments (v1) [Data set]. Hugging Face. https://huggingface.co/datasets/SoroushVahidi/mwfas-heuristic-metrics

@dataset{vahidi2026mwfasheuristicmetrics,
  title        = {MWFAS Heuristic Metrics: Run-Level Outcomes for Minimum Weighted Feedback Arc Set Experiments},
  author       = {Vahidi, Soroush},
  year         = {2026},
  version      = {v1},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/SoroushVahidi/mwfas-heuristic-metrics},
  note         = {Derived metrics dataset; no dataset DOI assigned}
}

@article{vahidi2024mwfas,
  title   = {Minimum Weighted Feedback Arc Sets for Ranking from Pairwise Comparisons},
  author  = {Vahidi, Soroush and Koutis, Ioannis},
  journal = {arXiv preprint arXiv:2412.16181},
  year    = {2024},
  doi     = {10.48550/arXiv.2412.16181}
}

No dataset DOI has been assigned. For exact reproducibility, cite the immutable Hugging Face revision 5621684d3c03138d5b2ebe544e91ae5698e67b5a.

License

CC BY 4.0 for the released project-generated metrics, summaries, metadata, and documentation, to the extent controlled by this project. This does not relicense upstream graph benchmark datasets, third-party software, or papers referenced by identifier or citation.