momo-ml 1.0.1

A production-oriented model monitoring library for detecting performance drift, data drift, and prediction drift, with automated reporting for ML systems.

momo-ml

momo-ml (MOdel MOnitoring for ML) is a production‑oriented library for systematically, scalably, and automatically monitoring model quality.
It covers key monitoring dimensions such as model performance drift, data drift, and prediction drift, and also supports automatic generation of visual reports.

This project is suitable for data science, ML engineering, MLOps, and other scenarios – it can be integrated into various monitoring and governance workflows after model deployment.

---

📌 Features

1. Performance Drift

Evaluate how model performance changes over time by comparing a reference dataset with a current dataset.
Supports both classification and regression tasks with a full suite of metrics, enabling detailed drift detection across time windows.

Key capabilities include:

• Classification metrics such as AUC, Accuracy, Precision, Recall, F1, and KS (supports binary & multiclass).
• Regression metrics including RMSE, MAE, R², SMAPE, P90/P95 Error, and Huber Loss.
• Automatic task‑type detection (classification vs. regression) based on label/prediction structure.
• Flexible configuration for label and prediction columns.

Use this to monitor whether a model’s predictive quality remains stable in production.

---

2. Data Drift

Detect how input feature distributions shift between a reference and a current dataset.
Works for both numeric and categorical features, with multiple statistical measures to quantify distributional changes.

Key capabilities include:

• For numeric features: PSI, KL divergence, JS divergence, KS statistic, and Wasserstein distance.
• For categorical features: PSI, KL divergence, JS divergence, and Wasserstein distance.
• Automatic separation of numeric/categorical fields.
• Safe handling of incompatible feature types with clear warnings.

This module helps identify upstream data issues such as schema drift, feature instability, or gradual population changes.

---

3. Prediction Drift

Monitor changes in the distribution of model predictions, independent of labels.
Useful for real‑time systems, unlabeled production environments, and early anomaly detection.

Key capabilities include:

• Summary statistics for numeric predictions (mean, std, quantiles) and categorical predictions (proportion changes).
• Drift metrics including PSI, KL, JS (all prediction types) and KS for continuous predictions.
• Histogram‑based L1/L2 distance and quantile/decile shift for continuous outputs.
• Automatic distinction between numeric and categorical prediction behavior using a configurable threshold.

Ideal for detecting unexpected output shifts before they escalate into performance degradation.

4. Report Generation

Generate human‑readable reports from model monitoring results, making drift analysis accessible to both technical and business stakeholders. Reports can be exported in multiple formats for different use cases.

Key capabilities include:

• Markdown reports – Clean, structured tables with visual risk badges (🟢🟡🔴) for quick drift assessment.
• JSON exports – Machine‑readable format for integration with dashboards, CI/CD pipelines, or downstream analytics.
• Customizable drift thresholds – Define your own risk levels for PSI, KL, JS, and other metrics to match your domain requirements.
• Comprehensive summaries – Performance drift tables, data drift feature‑wise metrics, prediction distribution shifts, and decile analyses all in one place.

Use this component to transform raw drift metrics into actionable insights, helping teams make informed decisions about model retraining or deployment.

---

🔧 Installation

pip install momo-ml