Predicting silver economy demand during population aging transition via federated learning-based multi-platform behavioral data collaboration
TL;DR
The Federated Learning-based Silver Economy Prediction Framework (FL-SEPF), which is a collaborative framework for privacy-preserving prediction in the silver economy, features a four-layer architecture with an adaptive weighted federated aggregation strategy that dynamically computes platform-specific weights based on data volume, data quality, and local validation loss to address Non-IID heterogeneity.
Abstract
The accelerating global population aging has fueled a surge in financial demands in the silver economy, making it critical to forecast elderly financial demands accurately for product allocation and risk management in institutions. The study proposes the Federated Learning-based Silver Economy Prediction Framework (FL-SEPF), which is a collaborative framework for privacy-preserving prediction in the silver economy. FL-SEPF features a four-layer architecture with an adaptive weighted federated aggregation strategy that dynamically computes platform-specific weights based on data volume, data quality, and local validation loss to address Non-IID heterogeneity. Local models employ BiLSTM with attention mechanisms, and a cross-platform attention fusion module integrates multi-dimensional features for multi-task prediction covering demand type classification and intensity regression. Differential privacy and Top-K gradient sparsification ensure privacy protection and communication efficiency. Experiments on four-platform datasets, covering 185,000–203,000 elderly users, demonstrate that FL-SEPF achieves an F1-score of 0.8312±0.0047 and AUC-ROC of 0.8927±0.0038, outperforming FedAvg by 3.5% and XGBoost by 11.4% in F1-score, with only a 1.7% gap compared to centralized Transformer training. Ablation studies confirm that adaptive weighted aggregation contributes the largest performance gain (3.34% F1 drop upon removal). Under extreme Non-IID conditions, FL-SEPF shows only 5.5% F1 degradation versus 11.3% for FedAvg, and at a privacy budget ε=1.0, performance loss is limited to approximately 1.0%. SHAP analysis reveals that financial behavior features, particularly portfolio diversity index and credit utilization rate, are the dominant predictors. This study provides a systematic federated learning solution for silver economy demand prediction under privacy-compliant conditions.
