Open Nullees opened 4 years ago
Yes, I did. It is unreasonable to use Bi-LSTM in knowledge tracing task. Because the output of each time step is the prediction of the score of next time step, if you adopt Bi-LSTM, the score of next time step will be visible to the model, the model will predict perfectly.
Thanks for your implements of DKT model.I change your model as LSTM and other version.They are all work well. But when i change LSTM function(only change Bidirectioanl=True and *hidden_size 2**). It's AUC value reached about 1.0. This is unbelievable. Did you know how to change as Bi-directional LSTM successfully?