Codes for the SIGSPATIAL 2021 GISCUP ETA Contest

• WDR
• DeepFM

• Baseline: Wide-Deep-Recurrent (WDR) from Didi's paper Learning to Estimate the Travel Time

• WDR

  1. Bidirectional LSTM (Bi-LSTM)
  2. Initialize hidden states of Bi-LSTM using embeddings of non-link-cross features, which is also the input to "deep"
  3. Axuiliary LSTM loss (predict link status at arrival time)

• DeepFM

  1. Bidirectional LSTM (Bi-LSTM)
  2. Initialize hidden states of Bi-LSTM using embeddings of non-link-cross features, which is also the input to "deep"
  3. Axuiliary LSTM loss (predict link status at arrival time)

For all models, we train in 2020.08.16 - 2020.08.27, evaluate in 2020.08.28 - 2020.08.29, and test in 2020.08.30 - 2020.08.31

The reported MAPE on the test set