Rice Safety Risk Data Generation Model Based on Gated Feature Transformation and Adaptive Diffusion

Rice safety risk assessment relies on complete monitoring data. However, limited by the cost of laboratory testing, severe data scarcity is often encountered in practice, which greatly hinders the effective implementation of food safety evaluation. Focusing on chemical contaminants in rice, this paper proposes a tabular data generation model based on adaptive diffusion (RADiff) to generate high-fidelity rice risk data under small-sample conditions. First, to address the traditional fixed parameter setting, an adaptive layer normalization (AdaLN) mechanism is employed to adaptively adjust parameters according to the varying noise levels across different denoising stages. Second, a residual gated cross transformation (RGCT) network is designed to capture the complex statistical associations between heavy metals (lead [Pb], cadmium [Cd], inorganic arsenic [iAs]) and mycotoxins (aflatoxin B₁, AFB₁, benzo[a]pyrene, BaP). Finally, the reverse diffusion process is optimized through skip sampling and an x₀-prediction paradigm to improve sampling efficiency and generation quality. The model is validated using empirical data from a major rice-producing province in China from 2022 to 2023, covering five key risk indicators: Pb, Cd, iAs, AFB₁, and BaP. The results indicate that the generated data by RADiff is highly consistent with real data in statistical distribution. Its Wasserstein distance (0.0029) and KL divergence (0.0061) significantly outperform SMOTE and other baseline models. Further food safety risk assessment experiments show that the model achieves superior performance in classification tasks after incorporating the generated data. This verifies that the generated data can effectively alleviate the problem of rice sample scarcity and provide reliable data support, thereby better serving downstream food risk assessment tasks.