Abstract
Modern Machine Learning Systems (MLSystems) must operate reliably in dynamic environments, yet current approaches offer limited support for system-level integrity analysis, runtime adaptation, and transparent component interactions. This paper introduces a generalizable Knowledge-Driven Architecture (KDA) that operationalises MLSystems by embedding semantic models into a Knowledge Graph (KG), reducing component complexity, formalising relationships, and enabling system-aware behaviour across ML lifecycles.
The KDA incorporates declarative probe definitions for component input and output streams, allowing automated instantiation of integrity-monitoring modules that detect behavioural degradation and trigger adaptive reconfiguration using the MAPE-K methodology. Semantic job descriptions further enable the KDA to optimise data flows and maintain consistent, traceable, and reconfigurable processing pipelines.
We evaluate the approach in a synthetic manufacturing setting through adaptive model switching and dynamic lifecycle management scenarios, achieving 0.56-second adaptation latency and 97% consumption reliability. The results demonstrate that combining semantic abstractions with knowledge-driven operational logic provides an effective foundation for building reliable, transparent, and self-managing MLSystems.
Why this matters
The EU AI Act introduces a risk-based regulatory framework that mandates continuous monitoring, transparency, and post-market oversight for high-risk MLSystems. The KDA shows how an architecture can maintain traceability, enforce semantic constraints, and support ongoing operational governance across ML lifecycles – exactly the kind of property regulators are now asking for.
Citation
Bæk-Petersen, A. L., Moghaddam, M., Hviid, J., & Kjærgaard, M. B. (2026). Operationalize Self-Managing Machine Learning Systems with a Generalizable Knowledge-Driven Architecture. In 2026 IEEE 23rd International Conference on Software Architecture Companion (ICSA-C). IEEE.