search

swangcs / BusGPS

This project aims at the big data challenges for predicting bus arrival time using GPS datasets.
13 stars 7 forks source link

Some problems of LSTM RNN model #16

Open Ruixinhua opened 4 years ago

Ruixinhua commented 4 years ago

I have met some problems when implements the RNN-LSTM model, the error is very huge when the part of known trips is short, for instance, only the arrival time of the first stop is known, the predictions of the next a few stops are very terrible and then becomes better. I think that's why the final MAE is 408. The solution to this problem might be due to the lack of iteration for the beginning steps, so I increase the time of epochs. Another problem is the initial weight before training which may affect the performance of the model according to the paper. By the way, I found when the point is mapped to the nearest stop, there are some stops share the same point which means the arrival time is the same at different stops which is not reasonable.

swangcs commented 4 years ago

@Ruixinhua For your last point:

there are some stops share the same point which means the arrival time is the same at different stops which is not reasonable.

I have uploaded a simple script to check both stop-based and meter-based trips data that you gave me. Haven't found any case as you described. Please have a look and let me know if I missed anything d8a34c5d63491e651df5af02c2eb15be7b93f8df.