Multiplicative Forests for Continuous-Time Processes.

Area of Science:

• Machine Learning
• Time Series Analysis
• Causal Inference

Background:

• Learning temporal dependencies in continuous time is crucial for understanding dynamic systems.
• Continuous-time Bayesian networks (CTBNs) model these processes but face scalability issues due to exponentially growing parameter spaces.
• Existing methods struggle with the complexity of high-dimensional, continuous-time data.

Purpose of the Study:

• To develop a scalable and efficient method for learning temporal dependencies in continuous time.
• To overcome the limitations of traditional continuous-time Bayesian networks.
• To introduce a novel representation for modeling complex temporal relationships.

Main Methods:

• Developed a partition-based representation using regression trees and forests.
• Parameter space grows linearly with the number of node splits, enhancing scalability.
• Utilized a multiplicative assumption for closed-form updates of forest likelihood.
• Efficient model updates achieved through analytical solutions.

Main Results:

• Multiplicative forests demonstrate significant performance gains.
• The method is scalable and effective even with few temporal trajectories.
• Reduced computational complexity compared to traditional CTBNs.
• Accurate learning of temporal dependencies from sparse data.

Conclusions:

• Multiplicative forests offer a computationally efficient and scalable solution for continuous-time dynamic systems.
• This approach enhances the practical applicability of Bayesian networks for temporal data.
• The method provides a robust framework for analyzing complex temporal dependencies.