Explainable multi-attribute machine learning via a hierarchical nature-inspired system toward predicting geological hazards

Explainable learning represents an efficient solution to enhancing computational cost and accuracy simultaneously in multivariant feature engineering with real-world conditions, particularly due to its extensive applicability across diverse decision-making problems. To address this issue within the hazard prediction systems, the computational time is also an important factor besides the computational cost and accuracy, but a multi-objective approach is not an effective technique due to high complexity. To overcome this problem, a collective framework is introduced to identify interpretable multi-aspect features that utilize a multi-layer optimization approach to enhance robust feature learning for real-world problems incorporated with a forecasting model, which proficiently anticipates hazard characteristics toward reliable decision-making. This multi-stage explainable algorithm is based on the layers of feature extraction, computational cost, computational time, hybrid computational accuracy, and a decision-making layer alternatively. This study presents a multi-stage knowledge-informed machine learning method fused with collective biological swarm intelligence-inspired optimization to reliably forecast varied parameters within a joint layout. Through numerical validation utilizing a real condition, the constructed model demonstrates an effective capability in predicting the hazard dynamic response of pile-reinforced slopes considering soil spatial variability with nature-inspired learning, exploiting novel assessment indices toward efficient decision-making. Finally, the proposed methods are compared with each other through some novel criteria such as the operational underfitting index, operational overfitting index, computational cost, computational time, and the computational accuracy criterion as performance metrics in both phases of training and testing. Furthermore, this numerical-theoretical approach exhibits the potential to supplant state-of-the-art deep learning and analytical models with machine learning alternatives, thereby creating a more rapid and cost-effective accurate approach for real-world applications that require high degradability of uncertainty and nonlinearities quantification.