A multi-source window-dependent transfer learning approach for COVID-19 vaccination rate prediction

Timely vaccination for new respiratory infectious diseases like COVID-19 is crucial for controlling pandemics. Accurate vaccination rate prediction is essential for aiding decision-makers and vaccine manufacturers in vaccine production and distribution planning, particularly in developing countries with limited resources. However, insufficient historical vaccination records in these countries challenge traditional machine learning models. This study proposes a multi-source window-dependent transfer learning (WDTL) approach integrated with a Convolutional Neural Networks with Long Short-Term Memory (CNN-LSTM) model, enabling target countries with limited vaccination record to learn from multiple source countries with similar COVID-19, policy, and economic factors within a temporal window. A case study is conducted on three developing countries from different continents, each with limited resources and significant challenges regarding vaccine supply and hesitancy during the early stages of the pandemic. The model's effectiveness was tested against a CNN-LSTM model and a multi-source TL model without window-dependent similarity evaluation, showing significant improvements with average Mean Absolute Percentage Error (MAPE) reductions of 45% and 19%, respectively. These results underscore the importance of selecting appropriate source countries across temporal windows, considering the evolving COVID-19 situation over time.