Algebraic Operations on Multilevel Explanations and Quantification of Their Uncertainty

This paper presents a scoping map of approaches aimed at improving the reliability of machine-learning model explanations through ensembling, aggregation, and multimodal fusion, and it systematizes evaluation metrics and sources of explanation variability. The study was conducted as a scoping review and reported in accordance with PRISMA-ScR. The search covered publications from 2015–2026 in Russian and English in Scopus, Web of Science, IEEE Xplore, ACM Digital Library, OpenAlex, arXiv, OpenReview, and CyberLeninka; citation chasing of key papers was also performed. The final corpus comprised 106 sources. The results show that explanation variability is driven by differences among explainers, stochasticity and hyperparameter settings, model uncertainty, and data-related effects. Three levels of ensembling and major classes of aggregators are discussed, including rank-based, trust-weighted, and probabilistic schemes; for heterogeneous representations, multimodal fusion and cross-modal agreement assessment are considered. Risks of false consensus, manipulation, adversarial fragility, and computational constraints are discussed. Engineering recommendations for building an explanation pipeline with uncertainty quantification are proposed, and directions for future work toward formal guarantees and standardized protocols are outlined.