Dataset Doctor

Dataset Doctor

Input

Parse $ARGUMENTS for:

• $0 — train/main dataset file path (required). If not provided, ask.
• $1 — optional test dataset file path.
• --target — target column name.

If --target is not specified, attempt to infer: look for columns named target, label, y, class, outcome, or is_*. If none found, ask the user.

Supported formats: CSV, Parquet, Excel, JSON.

Step 1: Detect Environment

Follow [environment-detection.md](../../lib/environment-detection.md) to:

1. Detect the package manager.
2. Check for required dependencies: pandas, matplotlib, scipy, sklearn.
3. Optional: statsmodels (for VIF calculation — if unavailable, use manual formula).
4. If any required packages are missing, ask the user before installing.

Step 2: Read the Data

Load the train dataset (and test dataset if provided). Examine shapes, columns, dtypes, and basic statistics to inform the health report.

python -c "
import pandas as pd
df = pd.read_csv('TRAIN_PATH')
print('Train shape:', df.shape)
print('Dtypes:', dict(df.dtypes))
print('Nulls:', dict(df.isnull().sum()))
print('Describe:')
print(df.describe())
"

If a test file is provided, repeat for the test set.

Step 3: Generate the Health Report Notebook

Follow [notebook-format.md](../../lib/notebook-format.md) for structure.

Section 1: Header (markdown)

# Dataset Health Report: {dataset_name}

*Generated on {YYYY-MM-DD} by dataset-doctor skill*
*Train file: `{train_path}`*
*Test file: `{test_path}` (if provided)*
*Target column: `{target}`*

Section 2: Health Score (markdown)

Calculate a health score out of 10:

• Start at 10.
• Deduct 2 per critical issue.
• Deduct 1 per warning.
• Deduct 0.5 per info item.
• Minimum score: 0.

Critical issues (deduct 2 each):

• Any column with >30% missing values.
• Data leakage detected (row overlap or suspiciously high correlation with target >0.95).
• Class imbalance >95% (majority class).
• VIF >10 for any feature.

Warnings (deduct 1 each):

• Any column with >5% missing values.
• Moderate class imbalance (>80% majority class).
• VIF >5 for any feature.
• Duplicate rows >1%.

Info items (deduct 0.5 each):

• Mixed dtypes in a column.
• Constant or near-constant columns (nunique 99% same value).
• Minor formatting inconsistencies.

Display: "Dataset health: X/10. N critical issues, M warnings."

List each issue with severity and brief description.

Section 3: Data Quality (code + markdown)

Code cells:

• Duplicate rows: df.duplicated().sum()
• Mixed dtypes detection: df[col].apply(type).nunique() for each column.
• Inconsistent formatting: for string columns, compare df[col].str.lower().nunique() vs df[col].nunique().
• Constant/near-constant columns: `nunique() 99%.
• Suspicious value ranges (negative values where unexpected, future dates, etc.).

Markdown cell summarizing findings.

Section 4: Missing Data (code + markdown)

Code cells:

• Percentage missing per column.
• Bar chart of missing percentages.
• Missingness correlation matrix (which columns tend to be missing together).

Markdown cell with recommendation per column:

• Drop column if >50% missing.
• Impute if 10:1): suggest SMOTE, class weights, or undersampling.
• If moderate (>3:1): suggest class weights or stratified splitting.
• If balanced: note that no action needed.

Section 6: Multicollinearity (code + markdown)

Code cells:

• VIF calculation for numeric features:

from sklearn.preprocessing import StandardScaler
import numpy as np

# If statsmodels available:
from statsmodels.stats.outliers_influence import variance_inflation_factor

numeric_df = df.select_dtypes(include='number').dropna()
scaler = StandardScaler()
scaled = scaler.fit_transform(numeric_df)

vif_data = pd.DataFrame({
    'Feature': numeric_df.columns,
    'VIF': [variance_inflation_factor(scaled, i) for i in range(scaled.shape[1])]
}).sort_values('VIF', ascending=False)

• Flag pairs with VIF >5.
• Correlation matrix for top offenders.

Markdown cell explaining which features are problematic and suggesting removal/combination.

Section 7: Data Leakage (code + markdown)

Only include if BOTH train and test files are provided.

Code cells:

• Row overlap: check for identical rows between train and test.
• Distribution comparison: KS-test per numeric feature between train and test. Flag p-value 0.95.

Markdown cell explaining any leakage found and its impact.

Section 8: Data Drift (code + markdown)

Only include if BOTH train and test files are provided.

Code cells:

• PSI (Population Stability Index) per feature:

def calculate_psi(expected, actual, bins=10):
    breakpoints = np.linspace(min(expected.min(), actual.min()),
                               max(expected.max(), actual.max()), bins + 1)
    expected_counts = np.histogram(expected, breakpoints)[0] / len(expected)
    actual_counts = np.histogram(actual, breakpoints)[0] / len(actual)
    expected_counts = np.clip(expected_counts, 1e-4, None)
    actual_counts = np.clip(actual_counts, 1e-4, None)
    psi = np.sum((actual_counts - expected_counts) * np.log(actual_counts / expected_counts))
    return psi

• Flag PSI > 0.2 (significant drift), > 0.1 (moderate drift).
• Side-by-side distribution plots for top drifting features.

Markdown cell summarizing drift findings.

Section 9: Recommendations (markdown)

Numbered priority list. Critical issues first, then warnings, then info.

Each recommendation is specific and actionable:

1. "Drop column id (constant, provides no signal)."
2. "Investigate leakage in total_spend (r=0.99 with target)."
3. "Apply log transform to income (skewness=3.2)."

Step 4: Write the Notebook

Write to ./notebooks/data_health_{dataset_name}_{YYYY-MM-DD}.ipynb.

Create the ./notebooks/ directory if it doesn't exist.

Step 5: Report

Tell the user:

• Where the notebook was saved.
• The health score.
• List of critical issues (if any).
• Top 3 recommendations.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: juanlurg
• Source: juanlurg/data-science-claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.