Formalized Heuristics for Revealing Causal Relationships to Support Trust in Intelligent Systems

The effectiveness of modern high-performance AI applications in a number of critical areas – medicine, security, transportation, etc. – is limited by the inability to explain the results of their work. The lack of explanation in decision-making undermines users' trust in digital products and services. One of the fruitful approaches to building explainable AI and maintaining trust in intelligent systems is the creation of data analysis methods that study cause-and-effect models and provide interpretation of the results. Problematic for the effective application of widespread machine learning methods for this purpose is the need for a sufficiently large representative dataset. The paper proposes an approach that logically implements plausible reasoning for inductively generating causal relationships in limited (but potentially replenishable) datasets. Empirical induction is represented by formal refinements and extensions of J.S. Mill's inductive methods based on the analysis of the algebraic similarity of objects with common properties. The stability of the generated empirical regularities is analyzed when expanding the databases of facts, depending on the formalized heuristics used. Ways to reduce the difficulty of detecting various types of such regularities are being considered. The concept of trust under consideration is based on the condition of falsifiability of results, as well as the operationally defined and verifiable concept of causality. The identified causal relationships can serve as a basis for explaining the mechanisms of the observed phenomena, which is especially important for making pragmatic decisions in poorly formalized areas