Hierarchical Positional Approach for ETA Prediction

Abstract

The GISCUP 2021 focuses on estimated time of arrival (ETA) which is widely used in various industries such as Transportation and Mobility. In this paper, we describe the 6th-place-solution that uses positional features hierarchically from wide to narrow and other statistical features for predictions with GBDT. Especially for narrow features, graph-embedding features are generated by extending node2vec to make it easier to handle large amounts of data. This solution got MAPE score of 12.478 as the final score.