Causal Autoregressive Attention and Automated Causal Inference (CARAT-CI)
Open Access DepositedThe complexity of modern wireless networks has exceeded what traditional monitoring and correlation-based machine learning systems can diagnose. Root cause analysis (RCA) involves identifying causal mechanisms in high-dimensional, nonlinear, and evolving telemetry data. This is a task that manual methods cannot manage at scale. This Praxis introduces Causal Autoregressive Attention with Automated Causal Inference (CARAT-CI). The two-stage model integrates transformers, variational autoencoders, and causal discovery to automate RCA.CARAT-CI utilizes a transformer variational autoencoder (T-VAE) to encode multivariate time series into a structured latent space in the first stage of the model. This captures hidden confounders. In the second stage of the network, a Directed Acyclic Graph (DAG) learning model identifies both instantaneous and lagged causal effects while adhering to acyclicity constraints. This architecture supports counterfactual simulation and interventional reasoning, enabling CARAT-CI to distinguish between direct and indirect causes, and to produce actionable, causally grounded diagnostics. Evaluation used the Solenix Controlled Anomalies Time Series (CATS) dataset as well as the Tennessee Eastman Process (TEP) benchmarks. On CATS, CARAT-CI identified root causes with 87.6% top-3 accuracy, which is higher than leading models MPGE and DECI. For TEP IDV1, the model reached 93.3% overall accuracy and achieved 100% edge-based top-3 accuracy. Performance dropped on the more challenging TEP IDV8 benchmark, though CARAT-CI still reached 80.0% top-3 accuracy and outperformed the leading models. Taken together, these findings suggest that the two-stage design improves stability and diagnostic accuracy, while the inclusion of counterfactual reasoning provides stronger explanatory insight. This work advances causal representation learning by connecting generative modeling techniques with structural causal inference, offering an algorithm that enables both discovery and inference. CARAT-CI enhances the development of scalable and interpretable RCA systems, laying a solid foundation for autonomous, self-healing network management.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.