# ============================================================================
# T15 — Filter vs embedded feature selection under irrelevant-feature injection
# ----------------------------------------------------------------------------
# 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: ML15
title: "Filter-based vs embedded L1 feature selection with injected noise features"
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: |
  Feature selection on tabular classification problems is often presented as a
  choice between simple univariate filters and embedded sparse models.
  Univariate filters such as chi-squared, ANOVA F-score, and mutual
  information are computationally cheap and easy to apply in a pipeline, but
  they evaluate each feature independently and can miss interactions. Embedded
  L1-regularized logistic regression performs selection during model fitting,
  potentially yielding feature subsets that better align with predictive
  structure when many irrelevant variables are present.

  A practical stress test is to inject synthetic irrelevant features into
  otherwise standard benchmark datasets. This setup creates controlled feature
  dilution while preserving original labels and base difficulty. Under such
  dilution, methods that can ignore noise should maintain accuracy and avoid
  selecting many synthetic features. Conversely, methods sensitive to spurious
  univariate associations may degrade as noise dimension increases.

  A credible CPU-only study should compare at least three filter selectors
  (mutual_info_classif, chi2, f_classif) and an embedded L1 selector in a
  shared downstream classifier pipeline, evaluate on multiple sklearn
  classification datasets, and report both predictive performance and selection
  quality. Including multiple random seeds is important because both noise
  injection and train/test splits can change apparent selector rankings.

  The key aim is not to reproduce a single published table, but to determine
  whether embedded sparsity offers robustness advantages over filters as
  irrelevant dimensions grow, and whether that robustness generalizes across
  datasets with different feature scales and class structures.

  *Does embedded L1-logistic feature selection retain predictive performance and reject injected irrelevant features more reliably than univariate filter methods under controlled feature-noise injection?*

hypotheses:
  - id: H1
    statement: "With 5x injected irrelevant features (relative to original feature count), L1-logistic embedded selection achieves higher mean test_accuracy than each filter method (mutual_info_classif, chi2, f_classif) on at least 2 of 3 datasets, averaged over >=5 seeds."
    measurable: true
  - id: H2
    statement: "At fixed selected-feature budget k=20 (or all features if p<20), L1-logistic selects a lower fraction of injected irrelevant features than the average of the three filter methods on all evaluated datasets."
    measurable: true
  - id: H3
    statement: "From no-injection to 5x-injection settings, mean test_accuracy drop for L1-logistic is <= 3 percentage points on at least 2 of 3 datasets."
    measurable: true

experiment_design:
  research_question: "Does embedded L1-logistic feature selection outperform univariate filter methods in robustness to injected irrelevant features on sklearn tabular classification datasets?"
  conditions:
    - name: "filter_mutual_info_k20"
      description: "SelectKBest(mutual_info_classif, k=20 or p if p<20) followed by LogisticRegression classifier."
    - name: "filter_chi2_k20"
      description: "MinMax scaling to nonnegative domain, then SelectKBest(chi2, k=20 or p if p<20) followed by LogisticRegression classifier."
    - name: "filter_f_classif_k20"
      description: "SelectKBest(f_classif, k=20 or p if p<20) followed by LogisticRegression classifier."
    - name: "embedded_l1_logistic"
      description: "L1-penalized LogisticRegression (saga/liblinear) used as embedded selector; keep top-20 by absolute coefficient magnitude (or nonzero if <=20), retrain LogisticRegression on selected subset."
  baselines:
    - "filter_f_classif_k20 as a standard univariate linear-statistic baseline"
    - "filter_mutual_info_k20 as a nonlinear dependency baseline"
  metrics:
    - name: "test_accuracy"
      direction: "maximize"
      description: "Held-out test accuracy averaged over >=5 seeds for each dataset and injection level."
    - name: "noise_feature_selection_rate"
      direction: "minimize"
      description: "Fraction of selected features that come from injected irrelevant columns."
    - name: "accuracy_drop_0x_to_5x_pp"
      direction: "minimize"
      description: "Absolute percentage-point drop in test_accuracy from 0x to 5x injection."
  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/ML15.json"