Predicting the directed acyclic graph based on feature extraction.

Area of Science:

• Causal inference and machine learning
• Time series analysis and forecasting
• Network science and graph theory

Background:

• Directed acyclic graphs (DAGs) are crucial for causal discovery.
• Existing methods primarily focus on estimating DAGs from static or time-series data, neglecting DAG prediction.
• There is a need for advanced methods to forecast dynamic causal structures.

Purpose of the Study:

• To develop a novel method for predicting directed acyclic graphs (DAGs).
• To transform the DAG prediction problem into a matrix prediction task.
• To enable forecasting of future causal relationships and network structures.

Main Methods:

• The proposed method reframes DAG prediction as a matrix prediction problem.
• Causal order and conditional independence are extracted via demixing and correlation coefficient matrices.
• Future DAGs are predicted by modeling these extracted matrices, integrating various time series forecasting techniques.

Main Results:

• Numerical simulations confirm the method's effectiveness in predicting both feature matrices and final DAG structures.
• The approach successfully predicted changes in risk spillover relationships within financial market data.
• The method demonstrates strong performance in forecasting dynamic causal networks.

Conclusions:

• The novel matrix prediction framework offers a versatile approach to DAG prediction.
• This method provides significant implications for forecasting dynamic relationships in economics, management, and social sciences.
• The ability to predict future causal structures enhances understanding and decision-making in complex systems.