Figure 8
The framework centres on prior rationales for M L techniques. Data-related rationales include handling nonlinear and complex relationships or patterns, specific and mixed data types, noisy data, big data, and data scarcity and imbalance. Related needs include structural relationships in infrastructure networks, sequential data with temporal relationships, informal language, visual data, complex optimisation under uncertainty and operation without prior data annotation or labelling. Modelling technique rationales include computational efficiency for real-time inference and training or retraining, model size for deployment versatility, interpretability for confidence and reliability, model robustness, modelling accuracy, and technique maturity and reliability. Computational efficiency connects to low-cost and fast training or retraining and rapid real-time inference. Model robustness connects to reduced reliance on assumptions about predictor-target relationships and reduced sensitivity to outliers and data uncertainties.Aggregated a priori rationales of HSCML methodology consideration
Source: Authors’ own work