2020/03/27 by Eric L. Manibardo, Manibardo, Eric L., Ibai Laña +5
Computer Science · Environmental Science · Engineering · #Data Stream Mining Techniques #Air Quality Monitoring and Forecasting #Traffic Prediction and Management Techniques
paper · pdf · doi:10.48550/arxiv.2003.14304
This work focuses on classification over time series data. When a time series\nis generated by non-stationary phenomena, the pattern relating the series with\nthe class to be predicted may evolve over time (concept drift). Consequently,\npredictive models aimed to learn this pattern may become eventually obsolete,\nhence failing to sustain performance levels of practical use. To overcome this\nmodel degradation, online learning methods incrementally learn from new data\nsamples arriving over time, and accommodate eventual changes along the data\nstream by implementing assorted concept drift strategies. In this manuscript we\nelaborate on the suitability of online learning methods to predict the road\ncongestion level based on traffic speed time series data. We draw interesting\ninsights on the performance degradation when the forecasting horizon is\nincreased. As opposed to what is done in most literature, we provide evidence\nof the importance of assessing the distribution of classes over time before\ndesigning and tuning the learning model. This previous exercise may give a hint\nof the predictability of the different congestion levels under target.\nExperimental results are discussed over real traffic speed data captured by\ninductive loops deployed over Seattle (USA). Several online learning methods\nare analyzed, from traditional incremental learning algorithms to more\nelaborated deep learning models. As shown by the reported results, when\nincreasing the prediction horizon, the performance of all models degrade\nseverely due to the distribution of classes along time, which supports our\nclaim about the importance of analyzing this distribution prior to the design\nof the model.\n