mlr3changepoint: A Unified Framework for Weakly Changepoint Detection

Offline change-point detection is a widely used statistical technique for identifying structural breaks in sequential data, with applications in fields such as signal processing, finance, and bioinformatics. Traditional methods, such as prune exact linear time (Killick et al., 2012) and binary segmentation (Hocking, 2024), are unsupervised, modelling shifts in the underlying data distribution without requiring prior knowledge or labels. While effective, they cannot leverage domain knowledge when it is partially available. Weakly supervised change-point detection methods bridge this gap by incorporating small amounts of coarse, inexact, or noisy labels to guide detection. Unlike unsupervised methods, many of which are efficiently implemented with consistent interfaces in packages such as changepoint (Killick et al., 2024) and rupturesRcpp (Nguyen et al., 2025), implementations of weakly supervised approaches are fragmented across packages, with no standardised tooling or shared interface currently available. This project aims to address this gap by extending the mlr3 ecosystem (Lang et al., 2019), a modular and extensible machine learning framework in R that provides unified interfaces for model training, evaluation, and benchmarking, to support weakly supervised change-point detection. It will also establish common abstractions for tasks, learners, and evaluation protocols tailored to change-point problems within the mlr3 design. The resulting framework will be rigorously tested and benchmarked, benefiting practitioners across various fields, enabling systematic comparison across methods, and providing a foundation for future extensions.