sjy1203 / gamenet Goto Github PK

i also get the opensource original theano version of LEAP kdd2017:
https://github.com/neozhangthe1/AutoPrescribe
I have noticed that you two papers have the same cooperation authors. Can i just use your LEAP implementaion version instead of the theano version to test the leap model on my data?
Wait for your response, thank you!

i want to reproduce LEAP and GAMENet, my scripts:
%run train_Leap.py
%run train_Leap.py --resume_path final.model --eval

%run train_GAMENet.py --model_name GAMENet-ddi --ddi
%run train_GAMENet.py --model_name GAMENet-ddi --ddi --resume_path final.model --eval

i choose final.model to do the test predict. Here is my performances:

• LEAP重现Test DDI Rate: 0.0699, Jaccard: 0.4438, PRAUC: 0.6362, AVG_PRC: 0.6364, AVG_RECALL: 0.6158, AVG_F1: 0.6064 avg med 19.
• GAMENET Tets(no DDI) DDI Rate: 0.0778, Jaccard: 0.5151, PRAUC: 0.7644, AVG_PRC: 0.6748, AVG_RECALL: 0.6943, AVG_F1: 0.6704
• GAMENET Tets(with DDI)DDI Rate: 0.0775, Jaccard: 0.5153, PRAUC: 0.7678, AVG_PRC: 0.6765, AVG_RECALL: 0.6923, AVG_F1: 0.6705

They are not same as your paper, but have the similar trend, is it normal or have some problems?

Wait for your responses, Thank you!

I am confused about the train and test process, why do you use c^t_d, c^t_p, c^t_m to predict c^t_m?

@sjy1203
EDA.ipynb
I tried running EDA.ipynb but getting this error 'ndc_rxnorm_file' is not defined. Upon further investigation I figured out there is a mismatch in file name. Please note that.

I can't find the file: data_gamenet.pkl, can you provide a copy? Thank you!

dropbox seem to be walled in China ,can I have some other way to download this file,if you have downloaded this file ,would you like to send me a copy ,my email is [email protected] you!

1. have you tried mini-batch in your code?
2. is the LEAP baseline implement the no-reinforcement learning version?
3. when tuning the LEAP baseline, what are the important hyper parameters ? can you list the grid-search range for LR and EPOCH?

Thank you!

Hello！Thank you for sharing the code. When reading your implementation code of Retain and DMNC methods, I find that dnc.py file is missing. Can you send me a copy of this file? My email is [email protected]. Thank you very much! Looking forward for your replay. With best regards!

In your paper, the coefficient of the multi-label margin loss is 0.1, however, in the code it's 0.01. I am wondering which one is the best parameters.

GAMENet/code_hybrid/models.py", line 185, in
) for _ in range(K - 1)])
TypeError: init() got an unexpected keyword argument 'input_size'

The length of the file I generated by running EDA.py is inconsistent with that given records_final.pkl in the data folder. May I ask how to generate a file with the same length as the file?Thank you for your advice.

Thank you very much for the code!
Regarding the data source, I am a bit confused. [Drug code mapping files] already exists under the [data] directory. Can the source of these data be provided?

In your implementation, for the baseline Leap, the groudtruth medicine rank is random or some kind of special rank?

I can find this file data_gamenet.pkl
Please help me in this one

Hi I am unable to get mimic dataset. Please help me on this

1. i read your paper and code, is the measures: average precision + average recall + F1 actually the same as the traditional "precision" "recall" "F1" in top-k item recommendation scenario?

2. i want to run the LEAP baseline to compare with other top-k medicine recommendation method(similar to the traditional top-k item recommendation design). In LEAP and your paper, the prediction medicine length is automatically decided by the model, how should i compare the results between the LEAP kind of models and the fixed top-K length models?
   Do you think it is fair to just choose the best top-k results to compare with the LEAP kind of models?

Wait for your response and thank you!

Why has been this line of code in data/EDA.ipynb:

ddi_most_pd = ddi_most_pd.iloc[-40:,:]

been used for getting the top-40 severity DDIs?
This is the code for getting the side effects with the least number of occurrences in the DDI file. I wanted to know how it corresponds to the top severity DDIs.