Identification of Explainable Structures in Data with a Human-in-the-Loop

Abstract

Explainable AIs (XAIs) often do not provide relevant or understandable explanations for a domain-specific human-in-the-loop (HIL). In addition, internally used metrics have biases that might not match existing structures in the data. The habilitation thesis presents an alternative solution approach by deriving explanations from high dimensional structures in the data rather than from predetermined classifications. Typically, the detection of such density- or distance-based structures in data has so far entailed the challenges of choosing appropriate algorithms and their parameters, which adds a considerable amount of complex decision-making options for the HIL. Central steps of the solution approach are a parameter-free methodology for the estimation and visualization of probability density functions (PDFs); followed by a hypothesis for selecting an appropriate distance metric independent of the data context in combination with projection-based clustering (PBC). PBC allows for subsequent interactive identification of separable structures in the data. Hence, the HIL does not need deep knowledge of the underlying algorithms to identify structures in data. The complete data-driven XAI approach involving the HIL is based on a decision tree guided by distance-based structures in data (DSD). This data-driven XAI shows initial success in the application to multivariate time series and non-sequential high-dimensional data. It generates meaningful and relevant explanations that are evaluated by Grice’s maxims.