Forecasting trend of agricultural talents flow by spatio-temporal graph neural network and LightGBM

Current agricultural talent flow prediction mainly uses single models (e.g., linear regression, ARIMA, LSTM), which fail to capture non-Euclidean spatial-temporal relationships and automatically extract complex spatio-temporal interactions, limiting accuracy and interpretability.This paper proposes a hybrid framework integrating a spatio-temporal graph neural network (STGNN) and LightGBM.Using 2010-2020 data from 17 cities in Henan Province, a spatio-temporal graph is built with city nodes and geographic-threshold edges.STGNN combines graph convolution and temporal convolution (TCN) to automatically learn spatio-temporal features, while LightGBM regresses lag and socio-economic indicators for interpretability.Benchmark comparisons with ARIMA, LSTM, and LightGBM, plus ablation and sensitivity tests, confirm the hybrid model's superiority.It reduces error by 10%-14% versus standalone STGNN/LightGBM, achieving under 12.3% overall error, with significantly improved accuracy and stability.