This study proposes a self-supervised representation learning-based risk prediction model to address the challenges of label scarcity, structural complexity, and high heterogeneity in Electronic Health Records (EHR) data for risk assessment tasks. The model combines masked reconstruction and context prediction to automatically learn temporal dependencies and latent semantic structures in EHR data under unlabeled conditions, thereby generating robust and generalizable patient health representations. The overall architecture includes a feature embedding layer, a temporal encoding module, an attention-based context aggregation module, and a risk decoding layer, which together capture long-term dependencies and achieve hierarchical feature modeling from multidimensional medical information. Multiple experiments were conducted on the MIMIC-III dataset, and comparisons with LSTM, BiLSTM, Transformer, GRU, and GAT models show that the proposed method achieves significant improvements in Accuracy, F1-Score, Precision, and AUC, verifying the effectiveness of the self-supervised mechanism for EHR representation and risk assessment. In addition, sensitivity analyses of key hyperparameters such as learning rate, hidden layer dimension, and noise interference level demonstrate that the model maintains stable performance under various training conditions, reflecting strong structural robustness and noise-resistant feature extraction ability. Overall, this study achieves the goal of learning high-quality health representations from unlabeled EHR data and provides an efficient and scalable technical framework for intelligent medical risk prediction.

This paper proposes a deep risk discrimination framework that integrates a Siamese network structure with a federated optimization mechanism to address challenges in financial transaction environments, including multi-source heterogeneity, data silos, and unclear decision boundaries. The method preserves data privacy by deploying structurally symmetric and parameter-sharing Siamese subnetworks across multiple clients. It performs contrastive encoding on transaction sample pairs and uses a contrastive loss function to pull normal samples closer and push anomalous ones apart in the embedding space. To improve the detection of low-frequency, high-impact behaviors, the model incorporates an entropy-aware regularization term and an embedding compression module. These components guide local training to suppress redundant representations and enhance embedding compactness. During the federated optimization process, the classic FedAvg strategy is used to aggregate models from different clients, improving global semantic alignment and generalization. Experiments on real transaction datasets demonstrate the superior performance of the proposed method across multiple metrics. The model shows particularly strong results in embedding consistency and anomaly boundary detection, confirming its modeling value and discriminative power in financial risk control.

This study addresses the challenges of scarce fraudulent samples, complex data distributions, and the limited adaptability of traditional methods in financial statement fraud detection by proposing a self-supervised learning algorithm. The approach first standardizes multidimensional financial indicators to mitigate scale differences, then employs an encoder to construct latent representations that capture high-order nonlinear relationships among indicators. A reconstruction task is introduced as an auxiliary signal, where a decoder approximates the input and minimizes reconstruction error to enhance the fidelity of representations. In parallel, a classification module distinguishes normal from fraudulent statements, with the model jointly optimizing reconstruction and classification losses to improve both feature completeness and discriminative ability. Experiments on a public financial fraud dataset show that the proposed method significantly outperforms existing baselines on Precision, Recall, F1-Score, and AUC, with particular strength in minority class recognition under imbalanced and limited data. Additional sensitivity experiments demonstrate that the method remains stable and robust across variations in optimizer type and imbalance ratios, confirming its effectiveness in complex financial environments. Overall, the algorithm provides an efficient and reliable pathway for fraud detection and exhibits distinctive advantages in accuracy and adaptability.