A Multi-Aspect Feature Dependency Screening for Augmentation of Derivative Feature Space Boosting Deep Learning Training: A Case of Air Quality in Subtropical Areas

This study examines the application of deep learning (DL) in predicting multivariate time-series (MTS), emphasizing the significance of novel feature exploration for intensive model re-training. This approach is essential for adaptive feature engineering, as the numerical properties of the foundational features differ from their statistical attributes. Elements such as the population density, resident count and land area, as well as foreign exchange rates, goods and hedging indicators, illustrate this disparity. The augmentation of feature space aims to improve the model’s efficiency without compromising performance by unnecessarily expanding the dataset dimension. This objective can be achieved by implementing an adaptive screening approach, specifically a multi-aspect feature dependency screening (MFDS), for the augmentation of derivative feature space (DFS). This can be accomplished through an adaptive screening process, an MFDS for augmentation of DFS. The methodology includes literature review and experimental modelling. The application of the Variance Inflation Factor (VIF) as an adaptive weighting method to evaluate redundancy is crucial for establishing effective screening mechanisms, enabling them to self-adjust to subtle changes and emerging patterns that may otherwise go unnoticed, and for ensuring sustainable advancement and enhancing dynamic training processes within the DL domain. Those multi statistical indicators include: cross-correlation for monitoring movements, Granger causality for evaluating predictor-predictand effectiveness, R-squared for assessing coincidence, amplitude-square coherence for estimating response degree, and Pearson coefficient for measuring linear dependence. In a seasonal context, formulating air temperature and humidity into water vapour pressure deficit facilitates a comprehensive analysis of critical features. Furthermore, the efficiency of predictive models is enhanced through the use of this adaptive screening and auto augmentation, as these methodologies incorporate new essential feature spaces. This research effectively combines physical knowledge with statistical screening techniques of feature dependency weighted mean (FDWM), resulting in a significant reduction in loss for DL models. Additionally, the distinct spatial-temporal and physical properties of the six air pollutants exhibit considerable variations, while also sharing certain commonalities.