Interrogating Ground Truth Assumptions for Human Factor Software Engineering through Local Interpretable Decision Tree eXplanations

Ground Truth (GT) quality is critical in supervised machine learning as it directly influences model performance and the validity of research outcomes. This study investigates the reliability of GT assumptions for temperament classification, comparing diverse annotation assumptions with a Behavioural Anchored-based Clustering (BAbC) paradigm, derived from the baseline K-Prototypes. Generated labels are evaluated through decision tree modelling, feature-path alignment, and model-agnostic interpretability eXplanations to interrogate bases of the GT assumptions. Experimental result show that BAbC consistently improved clustering quality and stability over K-Prototypes, reducing within-cluster dissimilarity and avoiding poor local optima. Its structural analysis revealed strong decision-path and subtree overlaps with rule-based paradigm, indicating shared logical hierarchies despite different derivation methods. K-Prototypes returned the most reliable GT decision rule with AUC weighted average of 0.7965. Based on its ground truth paradigm, a use case diagram is employed to visualize design functionalities that would better endear personalities towards the anchor centroid. Findings underscore the trade-offs between stability, behavioural completeness, and predictive separability amongst the paradigms, advocating for hybrid GT establishment that leverages unsupervised clustering for structural consistency and interpretable rules for reliability.