Causal Structure Learning via Temporal Markov Networks.

Area of Science:

• Causal Inference
• Machine Learning
• Time Series Analysis
• Biomedical Informatics

Background:

• Dynamic Bayesian Networks (DBNs) are widely used for causal discovery in time series data.
• However, the combinatorial complexity of DBN structure learning hinders accuracy and scalability.
• This limits their application in complex, real-world scenarios.

Purpose of the Study:

• To develop a more accurate and scalable method for causal discovery in time series data.
• To address the limitations of traditional dynamic Bayesian network (DBN) structure learning.
• To introduce log-linear temporal Markov networks (TMNs) as an alternative approach.

Main Methods:

• Proposed learning structure with log-linear temporal Markov networks (TMNs).
• Replaced combinatorial optimization with a continuous, convex optimization problem solvable via gradient methods.
• Leveraged feature representation for modeling irregular, sparse, or noisy event sequences.

Main Results:

• TMNs demonstrated faster computation times compared to representative DBN structure learners.
• TMNs achieved comparable accuracy to DBNs on synthetic datasets.
• TMNs performed accurately on a real-world causal discovery task in electronic medical records.

Conclusions:

• Log-linear temporal Markov networks (TMNs) offer a scalable and efficient alternative to DBNs for causal discovery.
• TMNs effectively handle complex time series data, including irregular, sparse, or noisy events.
• The proposed method shows promise for causal discovery in biomedical informatics and other fields.