Diffusion models represent a novel approach to data generation. However, existing diffusion model-based methods for tabular data generation face challenges in adaptively learning feature relationships and enabling mutual learning between mixed features. This paper integrates graph neural networks to advance research on diffusion models. First, leveraging the graph topology of tabular features, we propose an adaptive edge weight computation method for feature extraction in diffusion models. Second, we analyze the limitations of independently training diffusion models for different feature types and introduce a collaborative evolution method for mixed-type tabular data generation by combining mutual interference between discrete and continuous diffusion models with pseudo-label fusion of generated features. Finally, we present a graph-augmented diffusion model-based algorithm for mixed-type tabular data generation. Experimental results demonstrate that, compared to existing algorithms in the literature, our method achieves superior performance in terms of AUC metrics, effectively reducing the distribution gap between generated and real data while enhancing the predictive performance of downstream models.