One-Step MAML Algorithm with AdaBelief Optimization of Stroke Incidence Prediction

This paper provides a reliable prediction of stroke incidence based on meteorological factors, and due to the lack of sufficient stroke case data in some areas, the AdaBelief Optimized One-step Model Agnostic Meta Learning (MAML) training algorithm is adopted. Learning with this algorithm, the random forest prediction model is better adapted to new small samples data. We take the resolution of One-step MAML method into account to solve the shortages of MAML and MAML++, which have complicated training process, low efficiency, and large memory occupation. Considering One-step MAML updates the parameters only once in the inner loop, which reduces the influence of support set and increases the role of query set with small data samples. One-step MAML training method is more suitable for our application scenario. Optimization performance are presented and evaluated by MSE, RMSE, MAE evaluation indicators. In order to achieve significant predictive performance, AdaBelief is used to optimize One-step MAML method, for the purpose of solving overfitting problem in training process, and help the random forest prediction model converge more quickly to improve the prediction performance.