Predicting the spatial distribution of rock mass properties in cave mining is a critical challenge. In this study, we introduce a novel machine learning-based model to advance our comprehension of rock behaviour. This model specifically focuses on the spatial distribution of rock mass and mechanical properties, a crucial aspect of our understanding of rock mechanics. Our model is inspired by the recognition that neighbouring information serves as crucial prior knowledge for predicting the spatial distribution of these properties, wherein closer neighbours are likely to exhibit similar characteristics. To effectively capture and model these neighbour relationships, we design a neural network-based model to map the spatial locations within the data set into a new highdimensional space. Within this transformed space, we utilise Euclidean distance to identify N neighbours. Leveraging these neighbours, we introduce a straightforward yet highly effective approach—label propagation—to construct a graph that illustrates neighbour relationships. This graph not only facilitates the prediction of the spatial distribution of rock mass properties but also offers a measure of uncertainty for these predictions. We validate the proposed method using a real rock data set and showcase its efficacy in capturing intricate relationships and distributions within rock properties, enabling precise predictions across diverse spatial scales. Our results underscore the potential of machine learning-based approaches to revolutionise the field of rock mechanics, providing valuable insights applicable to geotechnical engineering, mining, and other sectors where an accurate understanding of rock properties is paramount. This model has direct practical implications, offering a new tool for geotechnical engineers and mining professionals to enhance their understanding and prediction of rock properties, thereby improving safety and efficiency in their operations.