tasks/ml/manifests/ML19.yaml

# ============================================================================
# T19 — Semi-supervised strategies on small tabular datasets
# ----------------------------------------------------------------------------
# Unlike paper_replication's P01-P07, the "synthesis" here frames a research
# QUESTION rather than a known paper's method. The model must design the
# experiment (conditions, metrics, datasets) — we only commit to what a
# competent study of this topic would include and what the rubric expects.
# ============================================================================

id: ML19
title: "Comparing graph- and pseudo-label-based semi-supervised learning on small tabular datasets"
arxiv_id: null
venue: "ARC-Bench 2026"
paper_asset: null

# The "synthesis" plays the role of the upstream briefing: research question,
# background, why the question matters, what "a reasonable experiment" looks
# like. It deliberately does NOT pre-specify a single method to reproduce.
synthesis: |
  Semi-supervised learning is attractive when labels are scarce but unlabeled
  examples are cheap. In sklearn, LabelPropagation and LabelSpreading provide
  graph-based transductive approaches, while SelfTrainingClassifier offers a
  pseudo-labeling wrapper around a standard supervised base learner. These
  methods rely on different assumptions: graph smoothness over neighborhoods
  versus confidence-thresholded iterative self-labeling. On small tabular
  datasets, their relative behavior can change quickly as the labeled fraction
  drops from moderate (20%) to very low (10%).

  A useful CPU-scale study should compare these methods under the same splits,
  feature preprocessing, and random seeds, with explicit control of labeled
  fraction. Because the key claim of semi-supervision is label efficiency, the
  experiment should include a supervised-only baseline trained on the same
  labeled subset and evaluated on a fixed held-out test set. This makes gains
  attributable to unlabeled-data usage rather than favorable splits.

  The study should report not only test accuracy but also balanced accuracy and
  macro-F1, since small tabular datasets can have class imbalance and accuracy
  alone may hide minority-class degradation. Repeating runs across several seeds
  is necessary because which points are labeled strongly affects outcomes at low
  label rates.

  Practically, the benchmark should stay lightweight: sklearn datasets only,
  no downloads, and model choices that complete within minutes on a single CPU
  core. The focus is not SOTA performance but robust relative comparisons across
  label regimes and datasets.

  *How do LabelPropagation, LabelSpreading, and SelfTrainingClassifier compare in label-efficiency versus a supervised-only baseline when only 10–20% of training labels are available on small sklearn tabular datasets?*

hypotheses:
  - id: H1
    statement: "At 10% labeled data, at least one semi-supervised method (LabelPropagation, LabelSpreading, or SelfTrainingClassifier) achieves test accuracy at least 3 percentage points higher than the supervised-only baseline on at least 2 of 3 datasets, averaged over ≥5 seeds."
    measurable: true
  - id: H2
    statement: "Across datasets at 10% labeled data, LabelSpreading with RBF kernel attains mean test accuracy greater than or equal to LabelPropagation with RBF kernel in at least 2 of 3 datasets (seed-averaged)."
    measurable: true
  - id: H3
    statement: "For SelfTrainingClassifier, increasing labeled fraction from 10% to 20% improves mean test accuracy by at least 1 absolute percentage point on at least 2 of 3 datasets."
    measurable: true

experiment_design:
  research_question: "How do graph-based (LabelPropagation/LabelSpreading) and pseudo-labeling (SelfTrainingClassifier) methods compare to supervised-only training under 10% and 20% labeled-data regimes on small tabular benchmarks?"
  conditions:
    - name: "supervised_logreg_labeled_only"
      description: "LogisticRegression trained only on the labeled subset of the training split; unlabeled points ignored."
    - name: "label_propagation_rbf"
      description: "LabelPropagation with RBF kernel; unlabeled training labels set to -1 and transductive predictions used for test inference."
    - name: "label_spreading_rbf"
      description: "LabelSpreading with RBF kernel under the same masked-label protocol."
    - name: "self_training_logreg"
      description: "SelfTrainingClassifier wrapping LogisticRegression with confidence threshold (e.g., 0.8), fit on mixed labeled/unlabeled training data."
  baselines:
    - "supervised_logreg_labeled_only is the primary baseline"
    - "10% labeled regime serves as a lower-label baseline versus 20% within each method"
  metrics:
    - name: "test_accuracy"
      direction: "maximize"
      description: "Classification accuracy on held-out test split, averaged over seeds."
    - name: "balanced_accuracy"
      direction: "maximize"
      description: "Balanced accuracy on held-out test split, averaged over seeds."
    - name: "macro_f1"
      direction: "maximize"
      description: "Macro-averaged F1 score on held-out test split, averaged over seeds."
  datasets:
    - name: "breast_cancer"
      source: "sklearn.datasets.load_breast_cancer"
    - name: "wine"
      source: "sklearn.datasets.load_wine"
    - name: "digits"
      source: "sklearn.datasets.load_digits"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 420

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML19.json"