Gaussian Process State-Space Models with Time-Varying Parameters and Inducing Points.

Area of Science:

• Machine Learning
• Statistical Modeling
• Time Series Analysis

Background:

• Gaussian process state-space models (GPSSM) are powerful tools for analyzing sequential data.
• Standard GPSSMs assume static model parameters, limiting their ability to capture dynamic changes in underlying processes.
• Extracting information from complex, evolving datasets requires models that can adapt over time.

Purpose of the Study:

• To propose novel time-varying Gaussian process state-space models (TVGPSSM).
• To enhance the flexibility of Gaussian process models for estimating time-varying functions.
• To improve the extraction of information from observed data exhibiting temporal dynamics.

Main Methods:

• Developed TVGPSSM with time-varying hyper-parameters to capture evolving data characteristics.
• Employed time-varying inducing points within the inference approach to adapt to functional changes.
• Utilized hierarchical importance sampling for efficient model inference.

Main Results:

• Demonstrated superior performance of the proposed TVGPSSM compared to standard Gaussian process models.
• Showcased the model's ability to effectively estimate time-varying functions.
• Validated the adaptability of the inference approach using time-varying inducing points.

Conclusions:

• TVGPSSM offers a more flexible and powerful framework for analyzing data with temporal variations.
• The proposed inference strategy effectively handles dynamic changes, outperforming traditional methods.
• This advancement enables more accurate information extraction from complex, time-dependent datasets.