<p>This paper presents a comparative analysis of various machine learning (ML) models used for predicting pillar collapses in a cave mine located in Northern Chile. The study’s primary objective is to evaluate the performance of these models by integrating geological, mine design, and operational parameters that cannot be accounted for using traditional methods such as numerical stress analysis and empirical approaches. Four different tree-based models (Random Forests, Extremely Randomized Trees, XGBoost, and Explainable Boosting Machines) and Support Vector Machines are employed for the ML analysis. The dataset was built using annotated records and image sets in the format of geological maps, for which an automated processing step was implemented to extract geospatial information, resulting in a tabular dataset. Each model was evaluated, and the best performing models, achieving accuracy rates of 85% and above, were selected for further analysis of feature attributions. The outcomes indicate that the intact rock strength of the pillars, column height, the number of open drawpoints ahead of the cave front, and cumulative oversize of extracted material are the most significant predictors of collapses for the case study. This work presents a workflow for data acquisition and representation, ML modeling application and knowledge extraction that can be used for complex geomechanical phenomena through the identification of critical factors related to the development of such instabilities.</p>