Interpretable and explainable rule-based modeling: analysis, design, and evaluation in the framework of Granular Computing and federated learning

点击翻译按钮获取中文摘要

Introduction With the heightened complexity of real-world systems and their advanced models, distributed sources of data, and increased expectations of the users, the quest for interpretable and explainable models. It becomes of paramount importance exhibiting tangible benefits, especially in system modeling and decision making in critical environments. Explainable system modeling is about developing models aimed at producing transparent knowledge about relationship existing in data/models and helping explain and audit prediction/classification results. The proposed Discovery Grant program is aimed at addressing the challenges of federated learning of explainable models with distributed data, analyzing innovative ways of evaluation of their performance and developing associated original methodologies and solutions. Objectives and originality We pursue a research program aimed at building a novel and efficient environment for designing and analyzing interpretable models such that they (i) possess sound and legitimate abstraction capabilities, (ii) exhibit the logic fabric of their structures, and (iii) are endowed with semantically sound mechanisms of evaluation of performance of the interpretable models and their results. We analyze how that these essential modeling features are efficiently realized with the aid of information granules (sets, fuzzy sets, rough sets) and how interpretable models can be constructed, studied and evaluated with the aid of Granular Computing. We design and evaluate quality of information granules regarded as building blocks fundamental to the formation of interpretable models. We develop new generalized techniques of data clustering essential to the formation of information granules. We construct and analyze innovative and efficient ways of federated learning of rule-based models. We establish efficient ways of assessing quality of interpretable models by engaging information granules of higher type. We open pioneering and far-reaching avenues of exploration of ways of assessing relevance of models through mechanisms of Granular Computing. Significance The research program will deliver substantive outcomes: Advancing fundamental knowledge via (i) breakthrough discoveries in Granular Computing and granular modeling. These discoveries open a new uncharted territory of the theory of advanced system modeling, concepts of explainability and a comprehensive evaluation environment, (ii) establishing new directions of granular explainable modeling, federated learning and evaluation of performance of models, (iii) exploring novel and transformative concepts of information granules, and (iv) analyzing and delivering quantifiable ways of evaluating granular results. Forming a range of original and transformative algorithms and design guidelines to tackle practical problems in system modeling, prediction, decision-making, and classification thus addressing the timely quest for advancements in complex system modeling.