Decoder Transfer Learning for Predicting Personal Exposure to Air Pollution

Personal air quality is an important indicator when assessing the impact of air pollution on personal health. Because personal air quality data are collected manually, it difficult to collect such data in large quantities. The main challenge facing personal air quality predictions is building an effective prediction model with a small amount of training data. Moreover, public atmospheric monitoring stations in urban areas have collected large quantities of air quality data. Therefore, we focus on using atmospheric monitoring data with a transfer-learning method to predict personal air quality. In this paper, we design a transferlearning framework based on an encoder-decoder structure. This transfer-learning framework uses the Wasserstein distance to match the heterogeneous distribution of the source domain (the data from the atmospheric monitoring stations) and the target domain (the personal air quality); we refer to this as decoder transfer learning (DTL). We use data from public atmospheric monitoring stations, collected by the Atmospheric Environmental Regional Observation System (AEROS) of Japan, as the source domain dataset and private datasets collected in Fujisawa, Japan, and Tokyo, Japan, as the target domain datasets to evaluate this approach. The experimental results demonstrate that compared with the inverse distance weighting (IDW), IDW with linear regression, and typical transfer-learning models, the proposed DTL framework demonstrates a significant improvement in prediction performance.