A federated learning framework for cross-domain e-commerce behavior prediction based on adaptive aggregation and differential privacy

Cross-domain e-commerce behavior prediction, inferring what a shopper will click, search for, or buy as they drift across several platforms, sits at the heart of personalization and conversion, yet the data that makes it valuable is locked inside competing platform silos. Three obstacles recur, and, to our reading, they are coupled rather than separate. Heterogeneity across platforms skews local updates and biases any single global model; the privacy-utility trade-off forced by data-protection law means the noise needed for a tight budget erodes the very signals personalization rests on; and non-IID data drags federated convergence out, occasionally to the point of divergence. We propose AAPD-FedRec (Adaptive Aggregation and Privacy-Differential Federated Recommender), whose name marks its two pillars, and which tackles the three obstacles together because, in practice, fixing one in isolation tends to reopen another. Three modules share the work: an adaptive aggregator answers heterogeneity by weighting client updates through a composite signal of data quality, gradient direction consistency, and historical update stability, displacing the volume-only rule of FedAvg; a sensitivity-aware scheduler answers the privacy-utility tension by spreading a fixed total budget non-uniformly across rounds and clients, paired with gradient clipping and Gaussian perturbation, while the moments accountant tracks cumulative privacy loss with closed-form \((ε ,δ )\) guarantees. A third module, a domain-aware embedding alignment layer, reconciles heterogeneous feature spaces across platforms without exchanging raw representations, so the three components attack the three obstacles in concert rather than one at a time. Experiments on Amazon and Taobao cross-domain partitions show that AAPD-FedRec converges in roughly half the rounds required by competing baselines, recovers AUC at \(0.7714\) under \(ε =1\) where DP-FedAvg drops to \(0.6921\) , and limits membership-inference attacker advantage to \(0.041\) at the same budget. Ablations confirm that the three contribution dimensions act complementarily rather than redundantly. The framework offers a practical path toward joint behavior modeling for cross-border platforms operating under strict regulatory constraints, with potential transfer to financial risk control, medical recommendation, and educational analytics.