[Paper Review] ruptures: change point detection in Python

Charles Truong, Laurent Oudre+1 more|arXiv (Cornell University)|Jan 2, 2018

Statistical Methods and InferenceMathematics10 references19 citations

TL;DR

ruptures is a comprehensive, modular Python library for offline multiple change point detection in univariate and multivariate signals. It provides a consistent, easy-to-use interface with support for parametric and non-parametric cost functions, multiple optimization methods (e.g., dynamic programming, binary segmentation), and flexible regularization, enabling fast, reproducible analysis of non-stationary time series across diverse applications.

ABSTRACT

ruptures is a Python library for offline change point detection. This package provides methods for the analysis and segmentation of non-stationary signals. Implemented algorithms include exact and approximate detection for various parametric and non-parametric models. ruptures focuses on ease of use by providing a well-documented and consistent interface. In addition, thanks to its modular structure, different algorithms and models can be connected and extended within this package.

Motivation & Objective

• To address the growing need for accessible, fast, and extensible tools for detecting changes in non-stationary signals across domains such as finance, bioinformatics, and system monitoring.
• To provide a unified, consistent, and well-documented interface that lowers the barrier to entry for non-experts while supporting advanced users.
• To enable seamless integration and extension of new algorithms and cost functions through a modular software architecture.
• To support both exact and approximate detection methods under various constraints (e.g., fixed number of change points, cost budget, penalty terms).

Proposed method

• The library implements a general optimization framework where change point detection is formulated as minimizing a contrast function $ V(\mathbf{t}, y) = \sum_{i} c(\{y_t\}_{t_i+1}^{t_{i+1}}) + \text{pen}(\mathbf{t}) $, with $ c $ as a cost function and $ \text{pen} $ as a regularization term.
• It supports multiple search algorithms: dynamic programming (exact), binary segmentation, bottom-up segmentation, window-based segmentation, and $ l_0 $-constrained detection.
• Cost functions include parametric models (e.g., for mean, scale, linear relationship, autoregressive coefficients) and non-parametric models (e.g., kernel-based or Mahalanobis-type metrics for distribution shifts).
• The library uses Numpy arrays as input and integrates with Scipy for efficient numerical operations, with Matplotlib for visualization and Pytest for unit testing.
• A caching mechanism is implemented to speed up repeated runs with different parameters on the same signal, and sub-sampling and minimum distance constraints are supported for scalability.
• The modular design allows users to plug in custom cost functions and optimization methods by adhering to a standardized interface.

Experimental results

Research questions

• RQ1 How can a unified, user-friendly, and extensible Python library for change point detection be designed to support both non-experts and advanced users?
• RQ2 What combination of optimization methods and cost functions enables robust detection of diverse change types (e.g., mean shifts, scale changes, distributional shifts) in multivariate signals?
• RQ3 To what extent can modularity and caching improve performance and usability in iterative signal exploration workflows?
• RQ4 How does the integration of both exact and approximate algorithms in a single framework enhance the practical applicability of change point detection?
• RQ5 Can a single library effectively support a broad range of change detection tasks beyond simple mean shifts, including non-parametric and multivariate changes?

Key findings

• ruptures is the first comprehensive Python library dedicated to multiple change point detection, offering a unified framework that integrates multiple algorithms and cost functions not previously available in a single Python package.
• The library supports both parametric and non-parametric cost functions, enabling detection of changes in mean, variance, linear relationships, autoregressive parameters, and distributional shifts.
• The use of dynamic programming ensures exact detection for small to moderate signal sizes, while approximate methods like binary segmentation and window-based segmentation allow scalable analysis on longer signals.
• The modular architecture enables seamless integration of new algorithms and cost functions, reducing implementation overhead and promoting reproducibility and extensibility.
• Caching and sub-sampling features significantly reduce computational cost during parameter tuning, enhancing usability in exploratory data analysis.
• The library is available under the BSD license, with full documentation and support for cross-platform use (Windows, Linux, macOS), ensuring broad accessibility and integration into scientific workflows.