Flexible Sliding Windows for Kernel Regression Based Bus Arrival Time Prediction

摘要: Given a set of historical bus trajectories D and partially observed trajectory S up to position l on the route, kernel regression KR is non-parametric approach which predicts arrival time at location $$l+h$$$$h>0$$ by averaging times same in past. The method does not weights data equally but it gives more preference similar data. This has been shown outperform baseline methods such as linear or k-nearest neighbour algorithms for prediction problems [9]. However, performance very sensitive evaluating similarity between trajectories. General algorithm looks back entire similarity. In case prediction, this work well when outdated part reflect most recent behaviour buses. order solve issue, we propose an that considers only sliding window them. introduces parameters corresponding lengths every along route determining how long should look into past These are automatically learned from training Nevertheless, parameter learning time-consuming process given large least quadratic size. Therefore, proposed approximation with guarantees error bounds learn efficiently. magnitude faster than exact algorithm. experiment real-world application deployed Dublin city, our significantly reduced compared state art