TRGIR-Rec

The code to reproduce the experimental results for "A Text-based Deep Reinforcement Learning Framework using Self-supervised Graph Representation for Interactive Recommendation".

Datasets

The data pre-processing codes is also included. You could download Amazon data from here, and run the dataProcessing.py first, then run the dataPrepare.py.

Runtime Environment

The code has been tested under Windows 10(version 1909) and Ubuntu 16.04 with TensorFlow 1.15.0 and Python 3.7.6.

Support independent training with CPU and joint training with CPU and GPU when CUDA is available.

Resource

You can download and add these resource to this project under the folder ./resource.

The pre-trained word vectors is available on GloVe.6B, which was trained on Wikipedia2014 and Gigaword 5.

The Long Stopword List can be obtained here.

Model Training

Take Digital_Music for example. After getting the source data, you should run data process first:

# Digital_Music in Self-supervised Graph Representation

python dataProcessing.py -d Digital_Music -dm sg
python dataPrepare.py -d Digital_Music -dm sg

To train our DQN model on Digital_Music:

python TRGIR-DQN.py 

or train our DDPG model on Digital_Music:

python TRGIR-DDPG.py 

You can modify the source codes to run other datasets. For other embedding methods, you should change the input by modify 'method' from 'mf' to 'sa' and 'sg''.