SIGSPATIAL_2021_GISCUP_ETA_WDR_DEEPFM

Codes for the SIGSPATIAL 2021 GISCUP ETA Contest

Models

• WDR
• DeepFM

Experiments

• 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

Results

The reported MAPE on the test set

	WDR	DeepFM
no Bi-LSTM	0.1295	0.1323
Bi-LSTM (1)	0.1318	0.1325
Bi-LSTM + initialize LSTM (2)	0.1290	0.1292
Bi-LSTM + initialize LSTM (2) + auxiliary loss (3)	0.1279	0.1277