Simple Versus Composed Temporal Lag Regression with Feature Selection, with an Application to Air Quality Modeling

Anthropogenic environmental pollution is a known and indisputable issue, and the need of ever more precise and reliable land use regression models is undeniable. In this paper we consider two years of hourly data taken in Wrocław (Poland), that contain the concentrations of NO2 and NOx in the atmosphere, and, along these, traffic flow, air pressure, humidity, solar duration, temperature, and wind speed. In the quest for an explanation model for the pollution concentrations, we improve and generalize the simple temporal lag regression model, and introduce a composed temporal regression model that entails a transformation of the data to improve the effectiveness of classical learning algorithms. We show that using the latter we obtain more accurate and better interpretable explanation models than using the former, and also than using the original, non-transformed data.