An Ensemble Approach to Multi-Source Transfer Learning for Air Quality Prediction

Declining quality of air is one of the major challenges faced by cities across the globe today. A product of rapid industrialization and urbanization, air pollution has quickly manifested itself as an existential threat to human life in major urban centres of the world. Developing economies such as India and China are at the forefront in bearing the brunt of this phenomenon. Intelligent predictive analysis is critical for framing policies that can help in controlling the severe effects of air pollution. Although Deep Learning architectures have proven to be effective in reliably forecasting the concentration of hazardous air particles like PM2.5, their predictive capabilities are significantly reduced when the amount of data available is not enough for effective training. Newer monitoring stations often lack the resources and/or personnel to reliably collect meteorological and environmental data, thus suffering from crippling data-insufficiency issues which hinder the applicability of forecasting models. This paper proposes an ensemble approach for Multi-Source Transfer Learning, with the aim of mitigating the data shortage issue. The proposed method generates a cumulative prediction by transferring the knowledge learned from multiple source stations to a given target station, thereby better utilizing the data that is readily available from nearby stations to boost prediction performance.