International agricultural trade data indicators often vary due to differences in statistical standards, trade policies, and climatic conditions across countries, and their selection can be influenced by subjective factors. To address the domain generalization problem in agricultural trade time series data, this paper proposes an anomaly detection method that handles the diversity and complexity of feature distributions while capturing unique patterns in agricultural trade sequences. The features extracted by a recurrent neural network are treated as learned knowledge. A standard classifier aligns the marginal distribution in the feature space, and a hidden variable autoregressive model is then applied for advanced prediction, thereby improving forecast accuracy. Furthermore, a hidden variable regression model is constructed for trade risk identification. By capturing the distribution characteristics within agricultural trade time series data, the model identifies potential risks. Experimental results confirm the validity of the proposed approach.
Cui et al. (Mon,) studied this question.