Experiments on bike traffic-flow prediction

 NYCChicagoDC
HM6.797344.680783.66747
ARIMA5.604773.797393.31826
HMM5.420303.797433.20889
XGBoost5.320693.751243.14101
LSTM5.133073.698063.14331
CG-GCLSTM4.643753.382552.87655
DG-GCLSTM4.671693.512432.98404
IG-GCLSTM4.778093.456252.68370
ST_MGCN (Multi-Graph)4.41732  
AMulti-GCLSTM4.226403.023012.58584

Add trend and period into feature:

(C6-P7-T4) means the length of closeness, period and trend are 6, 7, and 4 respective.

Default C6-P7-T4

 NYCChicagoDC
HM (C6-P7-T4)4.554743.285852.74502
HM (C0-P7-T4)4.278443.182902.68013
XGBoost4.149093.025302.73286
GBRT3.943482.858472.63935
LSTM3.927462.926632.65197
DG-GCLSTM3.885723.000552.60095
IG-GCLSTM3.791872.977072.58739
CG-GCLSTM3.774222.987972.59339
AMulti-GCLSTM3.734642.794752.47565
NYC CitySingleGraph-GCLSTM(Average)AMulti-GCLSTMST_MGCN
Number of trainable variables1974961993249245
Converaged training time2 hours6 hours51 hours

Use the ./Experiment/AMultiGCLSTM_Master_Bike.py to train the model or view evaluation results.

 

Back To HomePage