Continuous Management of ML-Based Applications

Modern software systems increasingly integrate Machine Learning (ML) components, whose non-deterministic behavior presents significant challenges across the application life cycle—from design and development to deployment and operation. A key concern lies in maintaining stable non-functional properties (NFPs), such as fairness, privacy, confidentiality, and explainability, especially as ML models evolve and are retrained. While current approaches predominantly address functional performance, such as classifier accuracy or algorithm optimization, they fall short in ensuring long-term stability of NFPs under dynamic, real-world conditions. In this paper, we propose a multi-model architectural and methodological approach designed to ensure consistent non-functional behavior in ML-based applications. The core idea is to monitor, verify, and maintain NFPs over time by comparing multiple models exhibiting similar non-functional characteristics and selecting the one that best supports stable performance amid contextual changes. Our approach is ML algorithm-agnostic and includes a two-step operational process: (1) Model Assessment, which evaluates non-functional properties of models chosen during the development phase, and (2) Model Substitution, which dynamically selects and replaces models to sustain stable NFPs during application runtime. We validate our method through experimental evaluation in a real-world scenario, focusing on fairness as a representative non-functional property, demonstrating the viability and effectiveness of our solution.

Keywords: Machine Learning, Non-functional Properties, Fairness, Model Assessment, Model Substitution, ML-based Applications, Explainability, Privacy, Confidentiality, Context-aware Systems.