Forest fire, as a significant ecological disturbance factor under the background of global change, not only affects the structure and function of forest ecosystems but also poses threats to regional climate, carbon cycling, and human activities. Therefore, accurately predicting the probability of forest fire occurrence and identifying its dominant influencing factors are essential for effective prevention and mitigation strategies. This study focuses on Heilongjiang Province, a fire-prone region in northeastern China, and aims to explore the spatial heterogeneity and driving mechanisms of forest fires using machine learning models.Based on forest fire point data from 2006 to 2020, we integrated 15 influencing factors encompassing five categories: meteorological conditions, topography, vegetation, anthropogenic activities, and social infrastructure. Two models were developed and compared: the global Gradient Boosting Machine (GBM) and the Geographically Weighted Gradient Boosting Machine (GWGBM). While the GBM model assumes a globally stationary relationship between variables, the GWGBM incorporates spatial weighting to capture local variations in the relationships between predictors and fire occurrence probability.The results demonstrate that the GWGBM model significantly outperforms the GBM model in terms of predictive accuracy and the ability to account for spatial heterogeneity. Specifically, the GWGBM model achieves a 25% improvement in test set accuracy compared to the GBM, highlighting its superior capability in capturing the nonlinear and spatially varying relationships between influencing factors and fire occurrence. Furthermore, the spatial distribution of predicted fire risk probabilities aligns closely with historical fire hotspot areas, demonstrating the practical relevance and reliability of the model outputs.Analysis of variable importance across models reveals that climatic conditions and vegetation-related variables are the primary driving factors of forest fires in Heilongjiang Province. These factors exhibit marked spatial variability in their influence, which is effectively captured by the GWGBM model. In contrast, the global GBM model fails to reflect such spatial nuances due to its assumption of spatial homogeneity.This study confirms the feasibility and effectiveness of integrating geographically weighted approaches into gradient boosting frameworks for spatial probability modeling of forest fires. The GWGBM model not only enhances predictive performance but also provides interpretable insights into the spatial dynamics of fire drivers, making it a valuable tool for forest fire risk management. In practical applications, the GWGBM can support the development of region-specific and risk-level-specific fire prevention and control strategies.Future research should consider incorporating the temporal dimension to develop spatiotemporally weighted models, enabling the exploration of dynamic forest fire risk evolution under changing climate conditions. Additionally, further analysis of model sensitivity across different spatial scales is warranted to improve the comprehensiveness and precision of forest fire management and decision-making support systems.