Noise Reduction Approach for Decision Tree Construction: A Case Study of Knowledge Discovery on Climate and Air Pollution

Data mining is more effective on noisy time series with appropriate data pre-processing. Singular spectrum analysis (SSA) is explored as the noise reduction approach for a decision tree classifier for noisy data. SSA provides groups of additive components, from low to high frequency, by decomposing the noisy time series. In this study, the noisy climate data is decomposed by SSA and is used to construct decision trees to predict the carbon monoxide (CO) air pollution levels. Analysis shows that separating out seasons from the annual data helps the algorithm; the classification accuracy improvements vary by season, with the maximum improvement (from 60.7% to 77.3%) found in summer by removing 6.42% of the high frequency signals, while autumn showed no improvement. Examining decision tree structures provides threshold climate values that impact on different CO levels, e.g., a light wind speed of les 2.5 m/s and any level of temperature inversion formation is found to associate with the high CO level (> 0.70 mg/m³). Overall, data pre-processing using SSA is encouraging to improve the results of any time series data mining approach. Examining decision trees of the climate and air pollution helps increase knowledge about the data, and the studied approaches can be adaptable for various future environmental studies