Bayesian online learning of the hazard rate in change-point problems

Area of Science:

• Time-series analysis
• Statistical modeling
• Machine learning

Background:

• Change-point models identify critical events in time-varying data, such as stock market shifts or brain state changes.
• Accurate change-point detection is challenging in noisy data streams.
• Prior Bayesian methods required pre-specifying the hazard rate (h), limiting adaptability to time-varying hazard rates.

Purpose of the Study:

• To develop a novel hierarchical Bayesian model for online change-point detection.
• To overcome limitations of previous models by inferring the hazard rate directly from data.
• To improve the identification of change points in complex, noisy time-series data.

Main Methods:

• Developed a hierarchical extension to existing change-point models.
• The model infers the hazard rate (h) from the data, rather than requiring it as a prior.
• Applied the model to both simulated (toy) and real-world datasets.

Main Results:

• The hierarchical model effectively identifies change points even with complex, time-varying hazard rates.
• The approach demonstrates superior performance compared to models with fixed hazard rates.
• The model serves as an ideal-observer for behavioral studies with rapidly changing inputs.

Conclusions:

• The proposed hierarchical model offers a more robust and adaptable method for change-point detection in time-series data.
• Inferring the hazard rate from data enhances the accuracy and flexibility of change-point analysis.
• This framework has potential applications in neuroscience, finance, and behavioral science.