Trend Drift Discovery for Individual Highway Drivers through Ensemble Learning

Abstract

Inter-city transportation plays an important role in modern smart cities, and has accumulated massive spatio-temporal data from various sensors in IoT (Internet of things). Current travel characteristics and future trends of highway traffic are valuable for traffic guidance and personalized service. As a routine domain analysis, trend drift discovery for highway drivers faces challenges in processing efficiency and predictive accuracy. Sensitive privacy of business data has to be considered, executive latency on huge data is hard to guarantee, and correlation among spatiotemporal characteristics cannot be fully employed. In this paper, a travel-characteristic based method is proposed to discover the potential drift of payment identity for individual highway drivers. Considering time, space, subjective preference and objective property, monthly travel characteristics are modeled on toll data from highway toll stations, and predictive error for those trends can be reduced dramatically through gradient boosting classification technology. With real-world data of one Chinese provincial highway, extensive experiments show that our method has second-level in executive latency with more than 85% F1-score for predictive accuracy.