tasks/ml/manifests/ML18.yaml

# ============================================================================
# T18 — Post-hoc probability calibration for sklearn classifiers
# ----------------------------------------------------------------------------
# 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: ML18
title: "Post-hoc calibration methods for sklearn classifiers on tabular benchmarks"
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: |
  Many widely used scikit-learn classifiers optimize discrimination rather
  than probability quality. Random forests, gradient boosting models, and
  RBF-kernel SVMs can achieve strong accuracy while output probabilities are
  miscalibrated, especially on small or moderately imbalanced tabular data.
  This matters in decision settings where thresholds, risk ranking, or cost-
  sensitive actions depend on reliable confidence estimates.

  Post-hoc calibration methods are attractive because they can be layered onto
  existing models without retraining the base learner. Platt scaling fits a
  sigmoid map, isotonic regression fits a non-parametric monotone map, and
  temperature scaling applies a single-parameter logit rescaling (implemented
  for binary tasks via optimization on held-out logits/probabilities). These
  methods differ in flexibility and overfitting risk, so their relative value
  may depend on classifier family and dataset size.

  A credible CPU-scale study should compare uncalibrated outputs versus at
  least three calibrators across multiple sklearn tabular datasets, using a
  proper train/calibration/test protocol and repeated seeds. Because this is a
  calibration-centric question, expected calibration error (ECE) and log loss
  should be primary metrics, with accuracy used as a guardrail to ensure that
  calibration does not degrade classification utility.

  The goal is not to reproduce a single paper result but to test whether any
  one calibrator is consistently superior across heterogeneous models and
  datasets under tight compute constraints. The key uncertainty is whether
  method ranking is stable enough to justify a default choice.

  *Which post-hoc calibration method (Platt, isotonic, temperature) most reliably improves ECE across RF/GBM/SVM-RBF on small tabular sklearn benchmarks without materially harming accuracy?*

hypotheses:
  - id: H1
    statement: "At least one post-hoc calibrator (Platt, isotonic, or temperature) reduces ECE by at least 10% relative to the uncalibrated model for at least 2 of 3 classifiers on at least 2 of 3 datasets, averaged over >=3 seeds."
    measurable: true
  - id: H2
    statement: "Isotonic regression achieves lower mean ECE than Platt scaling on at least 2 of 3 datasets when calibration-set size is >=150 samples."
    measurable: true
  - id: H3
    statement: "Applying post-hoc calibration changes test accuracy by no more than 1.0 absolute percentage point (mean over seeds) for each classifier-dataset pair."
    measurable: true

experiment_design:
  research_question: "Which post-hoc calibration method (Platt, isotonic, temperature) most consistently improves probability calibration for RF/GBM/SVM-RBF on sklearn tabular datasets while preserving accuracy?"
  conditions:
    - name: "uncalibrated"
      description: "Base classifier probabilities/decision scores used directly with no calibration."
    - name: "platt_scaling"
      description: "Sigmoid calibration fit on a held-out calibration split (CalibratedClassifierCV with method='sigmoid' or equivalent)."
    - name: "isotonic_regression"
      description: "Isotonic calibration fit on a held-out calibration split (CalibratedClassifierCV with method='isotonic' or equivalent)."
    - name: "temperature_scaling"
      description: "Single temperature parameter optimized on calibration split to minimize NLL; applied to logits/scores before probability mapping."
  baselines:
    - "uncalibrated outputs are the primary baseline"
    - "platt_scaling serves as a classic parametric calibration baseline"
  metrics:
    - name: "ece"
      direction: "minimize"
      description: "Expected Calibration Error (10-15 bins) on the held-out test split, averaged over seeds."
    - name: "log_loss"
      direction: "minimize"
      description: "Negative log-likelihood on test probabilities."
    - name: "test_accuracy"
      direction: "maximize"
      description: "Classification accuracy on held-out test split to verify discrimination is preserved."
  datasets:
    - name: "breast_cancer"
      source: "sklearn.datasets.load_breast_cancer"
    - name: "wine"
      source: "sklearn.datasets.load_wine"
    - name: "digits_binary"
      source: "sklearn.datasets.load_digits with target transformed to binary (e.g., digit<5 vs >=5)"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 480

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML18.json"