讲座时间：2026年9月3日13:00-18:00
讲座地点：管理楼G216
讲座主题：From Raw Data to Explainable Decisions: A Hybrid AI Framework for Decision Support
讲座嘉宾简介
Valeriya V. Gribova（瓦列里娅·V·格里博娃）教授，俄罗斯科学院通讯院士，俄罗斯科学院远东分院自动化与控制过程研究所（IACP FEB RAS）副所长兼智能系统实验室主任，俄罗斯人工智能协会副主席，远东联邦大学（FEFU）与符拉迪沃斯托克国立大学教授，发表学术论文400余篇，担任10余种国际和俄罗斯期刊编委。格里博娃教授在声明式知识库系统、AI组件自动化开发、智能用户界面及混合AI方法等方面取得一系列原创性成果。她领衔开发的IACPaaS云平台已成功应用于医学决策支持、激光增材制造、技术诊断、交通建模等多个领域，是俄罗斯人工智能产学研结合的标杆人物。
讲座摘要
The development of intelligent decision support systems (DSS) for real-world applications is hampered by two persistent obstacles: the overwhelming volume of unstructured, heterogeneous data (free text, logs, PDFs, mixed formats) and the complexity of formalizing domain knowledge encoded in clinical guidelines, technical standards, and regulatory documents. Traditional AI paradigms address these challenges only partially: data-driven machine learning suffers from opacity and data quality issues; knowledge-based systems face high acquisition and maintenance costs; and case-based reasoning alone struggles with scalability and similarity assessment in high-dimensional spaces.
This report presents a hybrid AI framework that integrates ontological engineering, knowledge graphs, large language models, and case-based reasoning within a unified architecture. The methodology comprises four key stages: (1) automated knowledge extraction from unstructured sources using LLMs and cascade prompting; (2) construction of formal, machine-executable knowledge bases (SMART standards) grounded in domain ontologies; (3) a two-step similarity engine—LocalSim for parameter normalization using reference ontologies and GlobalSim for weighted precedent retrieval via a modified k-nearest neighbours algorithm; and (4) generation of transparent, expert-oriented explanations based on ontological inference.
The framework is instantiated on the IACPaaS cloud platform and validated across diverse domains: clinical DSS for COVID-19 and stroke rehabilitation, additive manufacturing process optimisation, autonomous underwater vehicle mission planning, and construction compliance monitoring. Across all cases, the hybrid system demonstrates superior interpretability, adaptability to evolving knowledge, and robustness to incomplete or noisy input. We further discuss the use of ontological patterns to reduce development effort and the integration of LLMs for dynamic knowledge graph refinement. The results confirm that hybrid AI, rooted in explicit knowledge representations, provides a scalable, trustworthy, and explainable foundation for next-generation DSS in safety-critical and knowledge-intensive environments.
