<p>Unsupervised deep learning has become a powerful tool for geochemical anomaly recognition, particularly in greenfield exploration, due to its ability to perform mineralization-agnostic analysis. This category of unsupervised deep learning methods follows a reconstruction-based paradigm, where deep learning models reconstruct the geochemical background and detect anomalies via reconstruction errors. However, two major challenges persist: (1) anomalies and noise introduce disruptive variance that obscures the underlying background, and (2) limited training data leads to overfitting, causing models to memorize anomalies rather than generalize background patterns. To address these issues, deep learning is cast as a Maximum Likelihood Estimation (MLE) problem with latent variables, where the geochemical background is treated as latent variables that guide the learning process. An Expectation–Maximization (EM) algorithm is presented to address the MLE problem, enabling the iterative capture of the underlying background representation. It is demonstrated that this EM process can be reformulated as a self-distillation process in deep learning—a training strategy in which a network model refines its predictions by learning from its own prior outputs. This results in a mathematically interpretable model that progressively refines its understanding of background patterns while generating self-supervised signals that help mitigate overfitting to local variations such as anomalies. On this basis, the Transformer is adopted as the backbone network within this self-distillation framework, with its self-attention mechanism leveraged to capture long-range geochemical dependencies during background reconstruction and to support random data masking, which further reduces overfitting and improves generalization. The proposed approach is validated on geochemical data from the geochemically heterogeneous northwest Jiaodong gold province in Eastern China characterized by complex gold mineralization. Results demonstrate our model's enhanced robustness to anomalies and noise, its superior performance over state-of-the-art models, and the effectiveness of EM-derived self-distillation in improving anomaly recognition performance under limited data conditions.</p>