Attention-enhanced wasserstein GAN for agricultural market data imputation

Abstract

Crop price prediction in ASEAN markets is hindered by incomplete and inconsistent data, making data imputation essential. This study introduces the Wasserstein generative adversarial imputation network with attention (WGAIN+Att) to improve data quality for forecasting. Four configurations—GAIN, GAIN+Att, WGAIN, and WGAIN+Att—were evaluated on rice, corn, and soybean datasets (1961–2023). Results show that WGAIN+Att, particularly when attention is applied across all matrices (x, m, and z), achieved the best imputation performance, minimizing mean absolute error (MAE) and preserving statistical distributions, with optimal results at a 0.1 missing rate and 0.9 hint rate. In predictive tasks, GAIN-based imputations combined with convolutional neural networks (CNN)-long short-term memory networks (LSTM)-gated recurrent units (GRU) models consistently outperformed others in forecasting accuracy, achieving lower MAE and root mean squared error (RMSE). The findings highlight the role of attention in stabilizing imputation and ensuring realistic reconstructions, while also showing that aligning imputation with forecasting objectives improves agricultural price predictions.