INTERPRETABLE DEEP LEARNING FOR HEALTHCARE
For rule list part, we directly use the code from: https://github.com/nlarusstone/corels
For word2vec, we use the code (https://github.com/Adoni/word2vec_pytorch)
src/neural_net-v1.py: first version rule-list + lstm + prototype + binary classify
src/neural_net-v2.py: second version rule + bi-lstm + prototype + binary classify
src/neural_net-v3.py: bi-lstm + full-connect + binary classify
v4: final state => hidden state
v5: full-connect => conv1d
v6: embedding => one-hot add dim of "time stamp", optimal AUC = 0.64,
52.3seconds/epoch, totally epoch
v7: modify based on "v2", it doesn't work now
v8: based on v6: one-hot+timestamp => embedding, optimal AUC = 0.61, converge very slowly.
v9: embedding+lstm+prototype+softmax, don't success
v10: based on v8, embedding+lstm => embedding+cnn+lstm.
AUC = 0.672+ => 0.684+ (lr=1e-1,version 13) 20sec/epoch
v11: based on v10, embedding+cnn+lstm+full-connect => embedding+cnn+lstm+highway+full-connect.
AUC = 0.671+
v12: based on v11, embedding+cnn+lstm+highway+full-connect => embedding+cnn+lstm+highway+prototype+full-connect, doesn't work.
v13: see v10. same with v10
v14: based on v10/13, embedding + cnn + lstm + full-connect, then save the output of LSTM in ./data/
v15: use the output of v14, prototype. AUC=0.676 (highway=1:0.676,2:0.676,3:0.678+, 4:0.636+). don't BP grad for prototype.
v14 + v15:
v16: same with v15, but do BP gradient for prototype. it's very slow, AUC=0.672.
v17: it is based on v15, add some test procedure
v18: encoder-decoder seq2seq
v19: classify + interpretable_key_factor, kmeans is time-consuming
v20:
v20_0 as a reference (only classify, optimal AUC: 0.681+-0.001),
v20: classify + interpret layer, optimal AUC 0.690+-0.004
./run.sh