ABSTRACT
This paper introduces TrajectoryCausal, a framework for causal discovery from observational text data by leveraging anomalous trajectories. It combines trajectory representation, anomaly detection, and causal structure learning to infer causal relationships, achieving state-of-the-art performance in causal discovery tasks.
PAPER · PDF
Loading PDF…
Key findings
TrajectoryCausal framework integrates trajectory embedding, anomaly detection, and causal structure learning.
Achieves F1 scores of 0.82 on benchmark datasets, outperforming existing methods.
Opens new avenues for extracting causal knowledge from unstructured text.
Limitations & open questions
The framework's effectiveness may be limited in domains with less predictable language patterns.