Biography: Dr. Nashid Shahriar is an Associate Professor in the Department of Computer Science at the University of Regina. He holds a PhD from the University of Waterloo and an MSc from Bangladesh University of Engineering and Technology. His research focuses on security, reliability, and performance management of computer and communication networks and critical infrastructure such as nuclear reactors and data centers. He received the University of Regina’s 2025 Faculty of Science Award for Research Excellence (Early Career Researcher category) and the IEEE Communications Society’s 2023 Young Professional Award (Network Operation and Management area). His research has earned multiple international honours, including Best Paper Awards at IEEE PST (2025, 2026) and IEEE/ACM/IFIP CNSM (2017, 2019), and the Best PhD Dissertation Award at IFIP/IEEE IM (2021).
Abstract: Root cause analysis (RCA) is a critical capability for network observability, allowing operators to move beyond symptom detection to accurate, interpretable, and actionable fault diagnosis. However, modern communication networks generate heterogeneous observability data at increasing scale, making manually designed RCA rules difficult to maintain and purely data-driven or large language model (LLM)-based methods insufficiently grounded for reliable operational decision-making. In this talk, I will present our proposed neuro-symbolic root cause reasoning approach that integrates model-derived symbolic rules with LLM-based reasoning. Instead of requiring computationally expensive task-specific LLM training or fine-tuning, the proposed approach constructs a lightweight rule-driven knowledge base (KB) using traditional machine learning models and leverages it as external symbolic knowledge to guide the LLM toward evidence-based root causes. This design combines the transparency and traceability of symbolic rules with the flexibility and explanatory capability of LLMs, enabling more accurate and auditable RCA. We evaluate the proposed approach on two publicly available RCA datasets, where rule-grounded reasoning improves accuracy by up to 45.52 percentage points on average over the corresponding baselines. We further assess the generated reasoning using an LLM-as-a-judge strategy with a carefully designed rubric, showing that the proposed approach improves the relevance, rule alignment, and operational usefulness of RCA explanations across different model settings.