Integration Model for Estimated Time of Arrival

Abstract

Estimated Time of Arrival (ETA) plays a vital role in many application scenarios. For example, in various scenarios such as online car-hailing order distribution, price estimation, mid-trip estimation, and route decision-making. Accurate arrival time estimation can help the platform improve efficiency. However, accurate arrival time estimation is affected by static information and dynamic information, and the estimated arrival time has high technical difficulties and challenges. The organizers of this competition provided departure time and date, itinerary, road conditions, as well as topological structure data and weather information of the city's road network. At the same time, according to the characteristics of the given data, rich feature processing methods such as statistical features, category features, graph features, embedding features, and sequence features are used to provide massive feature information for model learning. One of the most important points is the application of "future data". Of course, in addition to the features, a lot of work has been done on the model structure and model fusion through the combination of machine learning and deep learning, ensuring the accuracy and stability of the model. The ETA is a typical time series problem. Therefore, In the deep learning section, we choose DCN [3] model and WDR [4] model as the basis, and the model distillation is combined on this as the deep learning part of the integrated model. At the same time, traditional machine learning is also used as a part of the integrated model, through a large number of different dimensions of feature engineering, to make up for the machine learning model's inability to better express the time series problem, and build a machine learning model with higher accuracy. Finally, through the fusion of the deep learning model and the machine learning model, extremely high accuracy is achieved in the ETA problem.