DriftImpact Library

A comprehensive Python library for detecting Data Drift and Concept Drift, and analyzing their impact on Model Performance.

🧠 Methodology

The library analyzes changes in datasets across three main layers:

1. Data Drift

Compares feature distributions between training data (Train) and scoring data (Score).

• Numerical Variables: Uses KS Test for distribution differences, Mann-Whitney U for central tendency shifts, and Levene Test for variance changes.
• Categorical Variables: Frequency distribution changes are measured using the Chi-Square test.
• P-Value Standardization: All test results are standardized according to the selected threshold. Metrics with detected drift are reported in red, and those without in green.

2. Concept Drift

Analyzes how the relationship between features and the target variable changes over time.

• Fisher Z-Test: Detects significant changes in the correlation of numerical features with the target variable.
• Cramer's V: Measures changes in the strength of association between categorical features and the target variable.
• Temporal Analysis: Finds the point where the relationship begins to deteriorate (First Drift Point) by dividing data into time windows.

3. Performance Impact Analysis

Analyzes whether drift is just a statistical deviation or if it actually degrades model performance.

• Calculates a "Weighted Drift Score" using model feature importances.
• Compares the drift score with model performance metrics (Accuracy, F1, RMSE, etc.) across time periods.
• Visualizes whether performance drops as drift increases, helping you determine the right time for intervention.

---

✨ Key Features

• Interactive HTML Report: A modern dashboard with drag-and-drop support and modal zooming.
• P-Value Color Coding: Immediate visual feedback based on drift status.
• Automated Dashboards: Detailed visual analyses and summary tables with a single command.
• AI Advisor (LLM): Actionable recommendations to fix drift, supporting local (Ollama, vLLM) and cloud (OpenAI) models.
• Flexible Integration: Works directly with scikit-learn compatible models.

---

🚀 Installation

To install the library locally:

pip install .

---

🛠 Usage

1. Full Drift Analysis and Reporting

from driftimpact import DriftAnalyzer
import pandas as pd

# Initialize Analyzer
analyzer = DriftAnalyzer(target_col='churn', threshold=0.05)

# Load data
train_df = pd.read_csv('train.csv')
score_df = pd.read_csv('score.csv')

# Run all analyses (Data Drift + Concept Drift + Temporal)
results = analyzer.full_analysis(train_df, score_df, time_col='date')

# Create Interactive HTML Report
analyzer.generate_html_report(results, save_dir='./reports')

2. Model Performance Impact Analysis

To measure your model's resilience against drift:

# Run analysis by providing the model, score data, and actual labels
impact_df = analyzer.analyze_performance_impact(
    model=my_trained_model,
    score_df=score_df,        # Features + Time column
    y_true=score_df['label'], # Realized labels
    time_col='date',
    drift_results=results.get('temporal_drift')
)

# Visualize the relationship between Drift and Performance
analyzer.visualize_performance_impact(impact_df, metric_name='accuracy', save_dir='./reports')

3. AI Advisor (Actionable Insights)

Get technical roadmap to fix detected drift using LLMs. Supports OpenAI-compatible APIs (Ollama, vLLM, OpenAI, etc.):

Option A: Local LLM (e.g., Ollama)

# Default is configured for local Ollama with qwen2.5
advice = analyzer.get_ai_advice(results, language='tr')

Option B: Cloud API (e.g., OpenAI, Azure)

# Pass your API key and base URL
advice = analyzer.get_ai_advice(
    results, 
    base_url="https://api.openai.com/v1",
    api_key="your-api-key-here",
    model="gpt-4",
    language='en'
)

Generate report with AI insights included

analyzer.generate_html_report(results, ai_advice=advice)

---

## 📁 Project Structure

- `driftimpact/`: Core library code.
- `reports/`: Generated HTML reports and PNG charts.
- `test.py`: Testing the entire flow using an example Churn model.

## 📄 License

MIT License