Adaptive Multi-Task Learning for Urban Air Quality Assessment
Accurate air pollution forecasting is vital for protecting the environment and public health. However, predicting air pollution continues to be a challenge due to complicated relationships between meteorological variables, pollutant concentrations, and time-dependent characteristics. This paper proposes a Temporal-Aware Multi-Task Learning (TMTL-AQI) framework to assess urban air quality via structured data. The model simultaneously performs three tasks: Air Quality Index category classification, PM2.5 and PM10 particulate matter regression, and auxiliary AQI value prediction. The framework consists of data preprocessing, cyclical temporal encoding, a shared backbone based on hard parameter sharing, and three task-specific output heads optimized jointly. It uses cyclical transformations to encode temporal traits that capture periodic patterns of the environment and produces a common neural representation through hard parameter sharing. The TRAQID dataset was used for empirical testing. The model outperforms single-task and baseline multi-task models with an accuracy of 0.7605 and an F1-score of 0.7422 for AQI classification. Although based solely on structured input (no image-based features), the model performed competitively with a state-of-the-art approach due to the reduced prediction error in regression tasks (MAE values of 13.34 μg/m³ for PM2.5 and 22.06 μg/m³ for PM10).