Learning Penalty for Optimal Partitioning via Automatic Feature Extraction

TL;DR

This work tackles the challenge of selecting the OPART penalty $\lambda$ for changepoint detection by learning it directly from raw sequences using recurrent networks. By integrating feature extraction and $\lambda$ prediction into a single framework, the approach aims to outperform traditional feature-engineering pipelines that rely on hand-crafted statistics. The authors demonstrate that recurrent architectures, particularly GRUs, can produce highly informative sequence features, leading to improved partitioning accuracy on 20 genomic benchmark datasets. Although the method offers higher accuracy, it requires substantial training time and memory, especially on long sequences, highlighting a trade-off between performance and practicality. Overall, the study introduces a robust, data-driven pathway to penalty learning that enhances changepoint detection while outlining directions for efficient deployment and future methodological refinements.

Abstract

Changepoint detection identifies significant shifts in data sequences, making it important in areas like finance, genetics, and healthcare. The Optimal Partitioning algorithms efficiently detect these changes, using a penalty parameter to limit the changepoints count. Determining the optimal value for this penalty can be challenging. Traditionally, this process involved manually extracting statistical features, such as sequence length or variance to make the prediction. This study proposes a novel approach that uses recurrent networks to learn this penalty directly from raw sequences by automatically extracting features. Experiments conducted on 20 benchmark genomic datasets show that this novel method generally outperforms traditional ones in changepoint detection accuracy.

Figures (8)

• Figure 1: Literature review of this study: (i) Selecting a detection method, (ii) Choosing a regularization method to control the changepoints number, (iii) Selecting a model type to predict the penalty, (iv) Set up the penalty prediction supervision, and (v) Choosing the model architecture.
• Figure 2: Example of how $\lambda$ affect the OPART changepoints. The sequence contains 4 labels: 3 positive (1 changepoint regions) and 1 negative (no changepoint region). Two types of errors are considered: false positives (fp), where extra changepoints are detected in any labels, and false negatives (fn), where no changepoint is detected in a positive label. In this examlpe, $\lambda=10$ is optimal.
• Figure 3: Diagram of the methods - Instead of performing feature extraction and the learning model separately, the proposed method uses a recurrent network to combine both into a single operation.
• Figure 4: Example: features vs. targets from the dataset detailed. The upper plots: the length or variance vs. targets, as these features are used on $\lambda$ prediction unsupervised methods. The lower plots: automatically extracted features vs targets, using a GRU model with two extracted features. Since we consider the predicted $\lambda$ has a linear relationship with the features, a good feature should exhibit a fairly linear pattern in the plot.
• Figure 5: The accuracies for each fold are presented for datasets without any pooling preprocessing. The cancer dataset has 2 folds, while the other two datasets each have 6 folds. The Constant and RNN models have been excluded due to their small accuracies to better highlight the differences and improve visualization. Overall, the GRU model achieves the highest accuracy.