Selective Inference for Change Point Detection by Recurrent Neural Network.

Area of Science:

• Time series analysis
• Machine learning
• Statistical inference

Background:

• Recurrent Neural Networks (RNNs) excel at identifying complex patterns in time series data.
• However, RNNs risk mistaking noise for genuine change points (CPs).
• Quantifying the statistical reliability of detected CPs is crucial for robust analysis.

Purpose of the Study:

• To develop a method for statistically validating change points (CPs) detected by Recurrent Neural Networks (RNNs).
• To rigorously control the risk of false positive CPs.
• To provide theoretically sound p-values for RNN-identified CPs.

Main Methods:

• Implementation of a novel method based on the framework of selective inference (SI).
• Applying the SI framework to RNN-based change point detection.
• Addressing the technical challenge of characterizing RNN's CP selection process.

Main Results:

• Demonstrated the validity of the proposed method using artificial data.
• Validated the effectiveness of the method through real-world data experiments.
• Successfully provided statistically reliable p-values for detected CPs.

Conclusions:

• The proposed selective inference (SI) method effectively quantifies the statistical reliability of RNN-detected change points (CPs).
• This approach mitigates the risk of false detections in complex time series.
• The method offers a theoretically sound way to validate CPs identified by machine learning models.