tidychangepoint: a unified framework for analyzing changepoint detection in univariate time series

Abstract: We present tidychangepoint, a new R package for changepoint detection analysis. Most R packages for segmenting univariate time series focus on providing one or two algorithms for changepoint detection that work with a small set of models and penalized objective functions, and all of them return a custom, nonstandard object type. This makes comparing results across various algorithms, models, and penalized objective functions unnecessarily difficult. tidychangepoint solves this problem by wrapping functions from a variety of existing packages and storing the results in a common S3 class called tidycpt. The package then provides functionality for easily extracting comparable numeric or graphical information from a tidycpt object, all in a tidyverse-compliant framework. tidychangepoint is versatile: it supports both deterministic algorithms like PELT (from changepoint), and also flexible, randomized, genetic algorithms (via GA) that -- via new functionality built into tidychangepoint -- can be used with any compliant model-fitting function and any penalized objective function. By bringing all of these disparate tools together in a cohesive fashion, tidychangepoint facilitates comparative analysis of changepoint detection algorithms and models.

• The paper presents a unified framework integrating diverse changepoint detection algorithms, including PELT, Binary Segmentation, and genetic methods.
• It consolidates these techniques within a tidyverse-compliant interface, enabling straightforward comparative analysis and enhanced workflow consistency.
• Application to MLB data validates the package's effectiveness and highlights its potential for extension to multivariate and complex datasets.

"tidychangepoint": A Unified Framework for Analyzing Changepoint Detection

The paper presents the "tidychangepoint" package for R, a versatile framework designed for changepoint detection in univariate time series. The package consolidates existing algorithms and model-fitting procedures into a tidyverse-compliant toolset, providing a unified interface for intricate changepoint analysis, enhancing user experience and comparative algorithm analysis.

Introduction

Changepoint detection is crucial for understanding structural changes within time series data. It involves identifying points in time where statistical properties of the data sequence change. The paper introduces "tidychangepoint," an R package leveraging tidyverse principles to harmonize various algorithmic approaches to changepoint detection, including deterministic methods like PELT and flexible genetic algorithms.

Figure 1: Difference in home runs per plate appearance between American and National Leagues in MLB.

The Framework of "tidychangepoint"

The package integrates well-known changepoint detection methods such as PELT, Binary Segmentation, and Wild Binary Segmentation, alongside genetic algorithm-based approaches. It supports various parametric models, penalized objective functions, and graphical diagnostics. This facilitates the comparison of different algorithms and models, providing insights into their efficiency and suitability for various datasets.

Changepoint Detection Algorithms

The paper elaborates on the efficient algorithms implemented within "tidychangepoint." PELT leverages dynamic programming for linear time complexity, making it optimal under specific conditions. Genetic algorithms offer flexibility in model specification, allowing for the exploration of large solution spaces through evolutionary principles, though at a higher computational cost.

Figure 2: Comparison of changepoint sets returned by PELT and Shi's genetic algorithm.

Application to MLB Data

The package demonstrates its capability through analysis of MLB data, detecting known changepoints, such as rule changes affecting game dynamics. Through various algorithmic applications, the true structural changes in the data are discerned, even if not perfectly aligned, showcasing the nuance and power of the package's built-in algorithms.

Architecting "tidychangepoint"

"tidychangepoint" employs S3 object-oriented programming to structure its operations, seamlessly integrating with R's ecosystem. It extends functionality by incorporating methods for object manipulation, visualization, and detailed statistical summaries, ensuring comprehensive analytics and comparability across different algorithmic outputs and datasets.

Figure 3: Diagnostic plot showing the time-series segmented by identified changepoints.

Implications and Future Work

Changepoint detection using "tidychangepoint" has significant implications for time series analysis across domains, such as climate studies, financial market analysis, and quality control in manufacturing. The package's design is conducive to extension, inviting contributions for additional algorithms, models, and performance optimizations through C integration.

Conclusion

The "tidychangepoint" package represents a substantial step forward in unifying changepoint detection methodologies within the R programming language. It facilitates robust comparative analysis and model fitting across diverse datasets, offering a powerful toolset for researchers engaged in time series analysis.

Future work aims to further extend the package's capabilities, incorporating more changepoint detection algorithms, improving computational efficiency, and exploring multivariate time series analysis to broaden its applicability and precision across complex datasets.