Research on the Adaptation Algorithm of Preschool Science Education Resources Based on Federated Learning and Particle Swarm Optimisation

To address the challenges of privacy protection and data heterogeneity in adapting preschool science education resources, this paper proposes the EduSwarm-Optimised Federated Aggregation Mechanism (ESO-FAM). It is evaluated on BB-Sci2025, a de-identified dataset of 120,000 interaction records from 8,000 users over 40 virtual kindergarten nodes, in which heterogeneity is controlled by a Dirichlet concentration parameter over twelve science-topic categories, swept from α = 1.0 (near-IID) to α = 0.05 (strongly non-IID). Each node's model update is treated as a particle steered by a contribution-aware fitness function, while a proximity-gated local update separates short-lived curiosity from genuine cognitive development. Implementation uses PyTorch 2.1 with a Ray backend on four A100 GPUs, with λ = 1×10⁻⁴ and μ = 0.01. Resource-adaptation accuracy reaches 91.12%, 16.84 percentage points above FedAvg and 6.47 points above the strongest recent baseline, SCAFFOLD; convergence requires 90 rounds against 180 for FedAvg, and the communication load of 62.1 MB is 32.8% below PSO-FL. Under strong heterogeneity (α = 0.05) accuracy remains 82.5% with an average trajectory error of 0.08, whereas FedAvg falls to 42.1%. Every recommendation carries a predictive uncertainty estimate (expected calibration error 0.031), and only drift classified as substantive triggers re-aggregation. Against the four pathways reviewed here, whose accuracies remain within a 74–84% band, this work is the first to embed population-based search inside the federated aggregation step itself.