- NARM.py -- Main code -- Run this to train models - Needs Theano
- data_process.py -- Is called by NARM.py to load data from the .pkl train and test dataset files and to convert that data to numpy arrays and lists
- example_prepocess.py -- Standalone piece of code used to obtain the .pkl train and test dataset files from the original dataset published at http://cikm2016.cs.iupui.edu/cikm-cup/
All of the data preprocessing code is agnostic to Theano except https://github.com/prashant-jayan21/sessionRec_NARM/blob/master/data_process.py#L32. You may want to edit it as per your needs when you switch to using TensorFlow.
The final train and datasets to be used are:
- data/digi_train.pkl
- data/digi_test.pkl
(The validation set is created on-the-fly in the code.)
Also, the original raw dataset can be found at data_raw/dataset-train-diginetica.
The items feature matrix can be found here: https://drive.google.com/drive/folders/1uLCH3xGEu6LA0AC6weyHsRZqEeM4DiFF?usp=sharing. Simply load it as a numpy array using numpy's load function.
Format:
- #rows = #items (NOTE: row index = item id - 1)
- #columns = feature vector size = #unique prices + #unique categories
Each feature vector is a concatenation of the following two vectors:
- A one-hot vector for price info (size = 12)
- A one-hot vector for category info (size = 995)
You can also experiment with picking just one of these feature vectors instead of using the concatenated one. You'll need to split the numpy matrix accordingly.
Please do not push to master if your changes are unstable. Always work on a separate branch and merge to master whenever ready.