With the rapid development of the Internet and social platforms, the problem of spammer detection has become a major technical challenge in building a harmonious Internet environment. However, user data collected from social platforms are often subject to issues such as missing information and data noise. Therefore, in graph-based learning models for bot army detection, methods that use point estimation as weights fail to express uncertainty in regions with sparse or missing data. A graph neural network model for bot army detection, VRGAT, integrating variational inference, was proposed. It introduces a probability distribution for the weights and derives a variational approximation of the true posterior. By applying different convolution operations to the mean and variance, the model more accurately captures the variability in the data. Simulations based on the Twibot-20 dataset show that, compared to the best existing benchmark for bot army detection (F₁ = 88.12), VRGAT achieved an improved performance with an F₁ score of 89.64. In robustness experiments, when random noise was added at varying levels, the accuracy drop for VRGAT is significantly slower than for other baseline models, demonstrating its superior noise resistance. The experimental results demonstrate that the introduction of variational inference can enhance the effectiveness of spammer detection and improve the model's robustness against noise.