tasks/ml/manifests/ML08.yaml

# ============================================================================
# T08 — Class-imbalance handling strategies for binary classification
# ----------------------------------------------------------------------------
# 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: ML08
title: "Impact of imbalance-handling strategies on sklearn binary classifiers"
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: |
  Class imbalance is one of the most common practical failure modes in binary
  classification: a model can achieve high raw accuracy by predicting the
  majority class while missing the minority class almost entirely. In
  scikit-learn workflows, practitioners often choose among simple data-level
  resampling (random oversampling, random undersampling), algorithm-level
  weighting (class_weight="balanced"), or synthetic methods like SMOTE.
  However, these strategies trade off minority recall, precision, and overall
  discrimination differently, and their behavior varies with dataset geometry.

  A compact CPU-friendly study can still be rigorous if it compares multiple
  imbalance-handling strategies under the same base learner and preprocessing
  pipeline, evaluates with metrics appropriate for skewed labels (balanced
  accuracy, F1, PR-AUC), and averages over several random seeds. The focus is
  not on maximizing one benchmark score but on understanding whether methods
  that improve minority detection do so consistently across datasets, and at
  what precision cost.

  To keep implementation feasible in a short autonomous run, the experiment can
  use sklearn-native datasets and synthetic imbalance generated from a standard
  source dataset, with a shared train/test protocol and fixed model family
  (e.g., logistic regression). SMOTE can be implemented in a lightweight way
  using sklearn.neighbors to interpolate minority-neighbor pairs, avoiding any
  external dependency.

  The key analytical goal is to test whether weighted learning or synthetic
  oversampling provides the most reliable gain in balanced accuracy over a
  no-treatment baseline, and whether naive random oversampling tends to hurt
  precision-oriented metrics relative to class weighting.

  *Which imbalance-handling strategy (SMOTE, random oversampling, class weights, undersampling) gives the most consistent balanced-accuracy improvement for binary sklearn classification under a fixed model and protocol?*

hypotheses:
  - id: H1
    statement: "At least one of {class_weight_balanced, smote} achieves higher balanced_accuracy than no_handling by >=0.03 absolute on at least 2 of 3 datasets (mean over >=5 seeds)."
    measurable: true
  - id: H2
    statement: "Random undersampling yields the highest minority recall among compared strategies on at least 2 of 3 datasets, but does not achieve the best PR-AUC on those same datasets."
    measurable: true
  - id: H3
    statement: "Random oversampling does not outperform class_weight_balanced in balanced_accuracy on more than 1 of 3 datasets (seed-averaged comparison)."
    measurable: true

experiment_design:
  research_question: "Which imbalance-handling strategy provides the most consistent gains in balanced accuracy for binary classification with fixed sklearn models under controlled train/test splits?"
  conditions:
    - name: "no_handling"
      description: "Baseline logistic regression trained on imbalanced data with no resampling and default class weights."
    - name: "class_weight_balanced"
      description: "Same logistic regression with class_weight='balanced'."
    - name: "random_oversampling"
      description: "Training-set minority class randomly oversampled with replacement to match majority count."
    - name: "random_undersampling"
      description: "Training-set majority class randomly undersampled without replacement to match minority count."
    - name: "smote_k5"
      description: "Training-set minority class augmented with synthetic samples via kNN interpolation (k=5) until class balance is reached."
  baselines:
    - "no_handling is the primary baseline"
    - "class_weight_balanced serves as an algorithm-level baseline against data-level resampling"
  metrics:
    - name: "balanced_accuracy"
      direction: "maximize"
      description: "Mean of sensitivity and specificity on held-out test split, averaged over seeds."
    - name: "minority_recall"
      direction: "maximize"
      description: "Recall for the positive/minority class on test data."
    - name: "average_precision"
      direction: "maximize"
      description: "Area under precision-recall curve (PR-AUC / AP) on test data."
    - name: "f1"
      direction: "maximize"
      description: "Binary F1 score for the minority class at default threshold 0.5."
  datasets:
    - name: "breast_cancer_imbalanced"
      source: "sklearn.datasets.load_breast_cancer with induced imbalance by downsampling positive class in training folds"
    - name: "make_classification_90_10"
      source: "sklearn.datasets.make_classification(n_samples=4000, n_features=20, weights=[0.9,0.1], random_state=seed)"
    - name: "make_classification_95_05"
      source: "sklearn.datasets.make_classification(n_samples=5000, n_features=25, weights=[0.95,0.05], class_sep=1.0, random_state=seed)"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 360

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML08.json"