driftguard-prathvi 0.1.0

A robust data drift detection and schema validation library for machine learning pipelines.

DriftGuard

DriftGuard is a lightweight, robust, and publish-ready Python library designed to automate dataset validation and detect statistical data drift in machine learning pipelines.

By comparing new incoming datasets against a trusted baseline reference dataset, DriftGuard alerts you to schema changes, increases in missing values, or shifts in feature distributions before they affect downstream model performance.

---

Features

• Schema Validation: Detect missing columns, new columns, and data type mismatches.
• Null Rate Analysis: Monitor and flag columns where the rate of missing values increases beyond a configurable threshold.
• Statistical Drift Detection:
  • Kolmogorov-Smirnov (KS) Test (scipy.stats.ks_2samp) for numerical columns.
  • Chi-Square Test (scipy.stats.chi2_contingency) for categorical columns.
• Severity Tagging: Categorizes issues as INFO, WARNING, or CRITICAL for pipeline routing or CI/CD gate checks.
• Interactive Reports:
  • An ASCII summary table output directly to console.
  • Machine-readable JSON output for automated pipelines.
  • A beautiful, self-contained interactive HTML dashboard with per-column breakdowns and interactive searching/filtering.

---

Installation

pip install driftguard

Note: Depends on numpy, pandas, and scipy only.

---

Quickstart

Validate your production features in real-time or as part of a batch training/inference pipeline:

import numpy as np
import pandas as pd
import driftguard as dg

# 1. Create a reference dataset (baseline)
np.random.seed(42)
ref_data = {
    "age": np.random.normal(35, 10, 1000),
    "income": np.random.uniform(30000, 120000, 1000),
    "city": np.random.choice(["New York", "Chicago", "San Francisco"], 1000),
    "target": np.random.choice([0, 1], 1000, p=[0.7, 0.3])
}
reference_df = pd.DataFrame(ref_data)

# 2. Create a new dataset (with some drift and schema issues)
new_data = {
    "age": np.random.normal(38, 10, 1000),                 # Slight shift
    "income": np.random.uniform(30000, 120000, 1000),
    "city": np.random.choice(["New York", "Chicago", "Boston"], 1000),  # "Boston" is a new category
    "target": np.random.choice([0, 1], 1000, p=[0.7, 0.3]),
    "extra_col": np.random.random(1000)                    # New column
}
new_df = pd.DataFrame(new_data)
# Add some null values to 'income'
new_df.loc[np.random.choice(1000, 150, replace=False), "income"] = np.nan

# 3. Instantiate Validator and run checks
# p_threshold matches the significance alpha for KS/Chi2
# null_threshold is the maximum allowed null rate increase
validator = dg.Validator(reference_df, p_threshold=0.05, null_threshold=0.10)
report = validator.check(new_df)

# 4. Consume the report
# Prints a formatted ASCII table of all issues
report.summary()

# Export interactive HTML dashboard (saves report.html)
report.export("html")

# Export machine-readable JSON (saves report.json)
report.export("json")

---

License

This project is licensed under the MIT License - see the LICENSE file for details.