Prediction of Ground Transportation Arrival Time

Vadim Porvatov25 May 2024

Problem

A taxi or navigation app shows an arrival forecast, and the user plans around it. A large error breaks decisions that have no time margin. Map and logistics companies need a model that takes a ready route and returns its duration in seconds. Research on this is thin: few strong papers, Chinese models often come with weak code or none, and no public path-aware benchmarks existed at the end of 2022.

Example from the article: a user sees a 40-minute forecast for a ride to the airport and just makes the flight check-in, but the trip takes 140 minutes and the plane leaves.

Solution in brief

GCT-TTE represents every point of a route twice: as a node of the road graph (GCN encoder, pretrained with Deep Graph InfoMax) and as a map patch (RegNet pretrained on ImageNet). It concatenates both embeddings with weather and time features and feeds the sequence of points into a vanilla Transformer encoder. On 2GIS trips from Omsk and Abakan it beat five baselines on MAE and RMSE.

Questions this case answers
  • How to predict the duration of a path from a variable-length sequence of location points, combining a graph of the underlying network with map imagery
  • How to pretrain a node encoder on a large network graph without labels and reuse it in a regression model
  • How to predict the travel time of a given route from its GPS trajectory
  • How to combine a road graph and map images in one ETA model

Tools and methods

  • GCT-TTE — the model: point representation block plus a Transformer route block.
  • GCN and Deep Graph InfoMax — road-graph encoder and its pretraining.
  • RegNet — encoder for map patches.
  • Baselines — WDR, DeepTTE, DeepIST, DeepI2T, GridLSTM.
  • OpenStreetMap — source of the map patches.
Continue reading — it's free

Google or a one-time email code. No password.