Structured Data --- Data Quality Diagnostics & Baseline Modeling

Table of Contents

• Scope Boundary
• Workflow
• Decision Tree
• Required Inputs
• Code Structure
• Reporting Standards
• Models Available
• Example Dataset
• Configuration Defaults
• Why These Defaults
• Minimal Smoke Test
• Cross-Skill Interface

Open-source skill.

Scope Boundary

Use this skill when:

• The task is supervised prediction from row-column data with a single target.
• A strong tabular baseline is needed before trying a larger model battery.

Do not use this skill when:

• The main problem is causal inference, survival analysis, time series, panel data, or repeated measures.
• The feature space is mostly free text or images rather than tabular columns.
• The dataset is so small or leakage-prone that a simple baseline can only be interpreted as exploratory.

Workflow

Read each step file in workflow/ before executing that step.

Step	Responsibility	Executor	Document	Input	Output
Collect	Collect Inputs	Main Agent	workflow/step01-collect-inputs.md	User input	Structured input summary
Diagnose	Check Distribution	Main Agent	workflow/step02-check-distribution.md	Prior step output	PART 1 code block
Baseline	Run Primary Test	Main Agent	workflow/step03-run-primary-test.md	Prior step output	PART 2-3 code blocks + T1 track artifacts

Decision Tree

1. TASK TYPE
   ├── Classification (target is categorical) → F1 + AUC metrics
   └── Regression (target is continuous) → RMSE + R² metrics

2. CHECK CLASS BALANCE (classification only)
   ├── Balanced (minority ≥ 10%) → standard methods
   └── Imbalanced (minority < 10%) → class weighting, note power limits
   # 10% threshold: widely used heuristic for "mild" vs "severe" imbalance in tabular ML
   # (matches sklearn's `class_weight='balanced'` becoming materially different below this mark).
   # Not a strict rule — rare events < 1% may warrant additional treatments (SMOTE, Firth).

3. FEATURE TYPES
   ├── All numeric → standard scaling
   ├── Mixed (numeric + categorical) → encoding + scaling
   └── High cardinality categorical → target encoding or frequency encoding

Required Inputs

Role	What to collect
Target column	Column name, type (classification/regression)
Feature columns	List or "all except target"
Task type	Classification or regression
ID column	To exclude from features (optional)

Code Structure

PART 0: Setup & Data Loading
PART 1: Data Quality Diagnostics   → plot_01_data_overview.png
PART 2: Baseline Classification    → plot_02_confusion_roc.png
PART 3: Recommendation Block       → text pointing to analysis workflow

Reporting Standards

1. Classification: weighted F1 and macro AUC (OVR) --- always with 95% bootstrapped CIs
2. For imbalanced classes (minority < 10%): also report PR-AUC (Average Precision) alongside F1 and ROC-AUC. PR-AUC is the primary discrimination metric under severe class imbalance because it is not inflated by true negatives.
3. Regression: RMSE and R² with 95% bootstrapped CIs; also report MAE (equally primary — more robust to outliers than RMSE when the loss is symmetric in absolute error).
4. Format: "F1 = 0.XXX, 95% CI [0.XXX, 0.XXX]"
5. Feature importance: top 20 features by gain, normalized 0-100
6. Decimal places: 3 for metrics, 1 for percentages, 0 for counts
7. Sample size: report final analytic N (train/val/test split sizes)
8. Non-significance: "not statistically significant at alpha = .05" --- never "no difference"

Models Available

Status	Models
Implemented in this skill	LightGBM baseline for classification or regression
Implemented downstream in vera-ai-structured-generating	Logistic Regression / Ridge, SVM / SVR, Random Forest, XGBoost, LightGBM, CatBoost, MLP, TabNet, stacking, weighted voting
Not implied by this open-source build	Any method not named above should be treated as out of scope until a corresponding module exists in src/

Example Dataset

sklearn built-ins: load_breast_cancer() (binary), load_iris() (multi-class), load_wine() (multi-class). Python: from sklearn.datasets import load_breast_cancer

Configuration Defaults

Pipeline constants live in config/default.json. Read it before generation. Key knobs:

• alpha, ci_level — inference thresholds
• plot_dpi, plot_width, plot_height — figure dimensions
• bootstrap_iterations — bootstrap CIs for metrics
• split.{test_size, val_size_of_train, random_state} — data-split reproducibility
• baseline_model.* — LightGBM baseline hyperparameter grid
• imputation.{numeric, categorical} — default imputation strategy (median/mode)
• class_balance_threshold — minority-class rare-event threshold
• task_type — "classification" or "regression" (affects metrics and models)

To override: create config/local.json; the runtime merges local over default.

Why These Defaults

• LightGBM is the baseline because it handles mixed tabular features well, gives a credible first-pass nonlinear model, and stays much cheaper than a full model battery for open-source smoke tests.
• Median/mode imputation is the default because it is transparent, deterministic, and easy to keep inside a train-fold-only preprocessing pipeline.
• PR-AUC is promoted under class imbalance because ROC-AUC can remain deceptively high when the negative class dominates.

Minimal Smoke Test

• Classification smoke test: "Run vera-ai-structured-reviewing on load_breast_cancer() with the target as label and all other columns as features. Produce the standard T1 artifacts."
• Regression smoke test: "Run vera-ai-structured-reviewing on load_diabetes() with the target as outcome and all other columns as features. Produce the regression T1 artifacts."
• Expected pass condition: the run writes one runnable Python script, methods.md, results.md, tables/, references.bib, and either the classification or regression baseline plot set.

Cross-Skill Interface

Output:
├── code_python      → .py script
├── methods_md       → methods.md baseline fragment
├── results_md       → results.md baseline results fragment
├── tables/          → Markdown/CSV tables for diagnostics + metrics
├── figures/         → 2 PNGs (data overview + confusion/ROC)
├── references_bib   → .bib with baseline/evaluation citations
└── recommendations  → text block (what analysis workflow produces)

When this skill is used as T1_baseline inside vera-ai-application-pipelining, the standardized track artifacts above are REQUIRED, not optional.