Abstract

Modern systems such as cybersecurity networks and autonomous platforms generate large-scale relational data that evolve over time, requiring robust, adaptive, and scalable AI/ML modeling approaches. In this talk, I present a research program centered on graph-based representation learning for dynamic network streams under real-time and memory constraints. My recent work develops real-time, unsupervised, and adaptive methods for anomaly detection in evolving graph streams, including Adaptive-DecayRank, a Bayesian temporal ranking framework for tracking node-level importance under concept drift, and Adaptive-GraphSketch, a probabilistic multi-tensor sketching approach for scalable edge-stream modeling with bounded memory, processing up to 20 million edges in under 3.4 seconds. These methods demonstrate how principled algorithmic design improves temporal modeling with low-latency inference in high-velocity graph streams. Building on this foundation, I further show how graph-based approaches extend to autonomous systems through ViGAT, a real-time vision-to-graph framework in which visual data are transformed into graph representations for unsupervised and scalable anomaly detection under edge deployment constraints. I conclude by outlining future directions in scalable and dependable AI systems and graph-based representation learning for dynamic and security-critical environments.

About the Speaker

Anthony (Ocheme) Ekle is a Ph.D. candidate in Computer Science at Tennessee Technological University. His research focuses on graph-based machine learning, real-time anomaly detection in dynamic graph streams, and scalable AI systems for security-critical environments, including cybersecurity and autonomous vehicles. He has published in ACM TKDD, IEEE venues, and Applied Sciences, and received a Best Paper Award at Tennessee Tech. His work aims to develop efficient and dependable AI systems for real-world applications.