Influence of gap length on GAN-imputed missing dengue time series data
Abstract
Missing data in dengue surveillance systems pose a challenge for disease monitoring and response. In this paper, we explored the use of Generative Adversarial Networks (GANs) in imputing missing dengue time series data. GAN simulations were conducted using the real-world missing data distribution from OpenDengue dataset. We incorporated a bidirectional LSTM in the GAN architecture to capture temporal dependencies. An RMSE below 0.10 was achieved for short gap lengths. However, increased gap lengths were associated with higher RMSE values. Visual analysis further supported the deteriorated imputation at longer gap lengths, where imputed data struggled to follow the true data. Iteration loss suggests that this framework could still be enhanced for a more accurate missing dengue time-series data imputation.