python3, numpy, pandas, scikit-learn, seaborn, mxnet
Based on the paper: Dynamic Key-Value Memory Networks for Knowledge Tracing Source code: https://github.com/jennyzhang0215/DKVMN
Simple RNN version of DKT, based on the code of DKVMN above.
Different from the original DKT:
- original DKT: output n_questions predict probability;
- my DKT: input the current question and previous <question, answer>, then output a single probability.
- try GRU here.
- augment the help of positional infos, e.g. the frequency of the word at each time step appearing before. (seems no help for the performance)