Rainfall-induced clustered landslides pose severe hazards in the hilly and mountainous regions of southern China. Landslide susceptibility assessment serves as a pivotal support for disaster prevention and reduction; however, its accuracy is directly constrained by the scientific rationality of evaluation models and the selection of negative samples. Taking the rainfall-induced clustered landslides in Xinyi, Guangdong Province in October 2023 as the research background, this study aims to explore the impacts of different negative sample sampling strategies and machine learning models on assessment accuracy. Landslide positive samples were acquired via remote sensing image interpretation, and three types of negative sample datasets were constructed based on factor constraints (low slope), buffer random sampling, and unsupervised clustering. Subsequently, susceptibility assessments were conducted by integrating these datasets with ensemble machine learning modeling. The results indicate that while ensemble machine learning models inherently possess high baseline accuracy, the negative sampling method significantly influences the final precision. Specifically, the model utilizing unsupervised clustering sampling achieved the optimal accuracy, followed by buffer random sampling, whereas the low-slope constraint sampling yielded the lowest accuracy. The unsupervised clustering negative sample sampling method is well-adapted to the Xinyi study area, and its combination with ensemble machine learning can further enhance assessment accuracy. This study provides valuable references for sample selection and model construction in the susceptibility assessment of rainfall-induced clustered landslides in the hilly and mountainous regions of southern China.