BiTA: New Temporal Graph Network Framework for Alert Prediction

A research paper proposes BiTA, a Bidirectional Gated Recurrent Unit-Transformer Aggregator, for alert prediction in computer networks. The framework enhances Temporal Graph Neural Networks (TGNs) by redesigning the temporal aggregation function to capture bidirectional sequential dependencies and long-range contextual relations. This addresses limitations of existing TGN methods that rely on unidirectional or single-mechanism aggregation, which fail to model recursive, multi-scale temporal patterns in real-world attacks. BiTA does not increase model depth or capacity but improves temporal reasoning. The paper is available on arXiv with ID 2604.22781.