Methodological Aspects of the Transition from Accuracy Metrics to Risk Modelling in the Design of Hybrid Intelligent Systems

( Vol-12,Issue-2,March - April 2026 )

Author(s): Bezhentsev Yurii

Download Full Text PDF
Total View : 146
Downloads : 7
Page No: 120-127
DOI: 10.22161/ijaems.122.14

Keywords:

hybrid intelligent systems, risk-oriented evaluation, accuracy metrics, risk modelling, cost-sensitive classification

Abstract:

The article examines the methodological transition from the use of standard accuracy metrics to risk-oriented modelling in the design of hybrid intelligent systems embedded in a management and control loop. The relevance of the approach stems from the fact that, in real processes, the quality of a solution is determined by the probability of undesired events, the magnitude of their consequences, and the speed of error detection and reversibility in the operational environment. The aim of the work is to translate the evaluation of intelligent models from the plane of numerical indicators into the language of systemic risk associated with people, infrastructure, and organisational procedures. The scientific novelty lies in integrating cost-sensitive error assessment, class imbalance analysis, and data shifts with an architectural description of hybrid (neuro-symbolic) systems, in which risk is distributed along the entire chain: data – model – rules – human – action. Accuracy metrics are proposed to be treated as particular input characteristics within a more general scheme for managing undesired events, defined by loss functions, barrier architecture, traceability, explainability, and controlled degradation modes. It is shown that, under class imbalance, label defects, and drifting data, the choice of metric and thresholds becomes a methodological decision that directly influences actual damage rather than a technical detail of the experiment. A conclusion is formulated on the necessity of shifting acceptance criteria from maximising aggregated metrics to constraining expected loss and ensuring risk controllability at the level of the hybrid system as a whole. The article is intended for researchers and engineers developing and deploying risk-sensitive intelligent systems in safety-critical and regulated domains.

Received: 20 Feb 2026; Received in revised form: 22 Mar 2026; Accepted: 26 Mar 2026; Available online: 01 Apr 2026

Cite This Article:

Share: