00mjk / spamgan

Generative Adversarial Network-based classification for opinion spam

ARCHIVED VERSION, NOT SUPPORTED. For most recent SpamGAN work, see: https://github.com/YankunShen/spamGAN

SpamGAN Code

GANs for Semi-Supervised Opinion Spam Detection Code from: https://arxiv.org/abs/1903.08289

Instructions

You will need to install Texar: https://texar.io/

The code was developed in Tensorflow 1.12.0, with Python 3.6.7.

Data

The cleaned original data is present in the data folder, with opspam_reviews.txt and opspam_labels.txt including the 1600 raw reviews and labels from the Ott dataset. The opspam_train_reviews.txt and associated labels include the 1276 base training examples (post-removal of the 4 duplicate reviews) and the opspam_test_reviews.txt includes the common test set for all the experiments. The various chicago_unlab_reviews files include the extracted unlabeled reviews from TripAdvisor

The labeled training examples and the unlabeled examples are further chopped up for the experiments to identify variation in performance for the model depending on amount of labeled/unlabeled data used.

In the base folder, the minrun series of files are one of the best performing model configurations: spamGAN-50. This should serve as a minimal working example (hopefully!). The full experiment code and such is in the experiment_code folder, and the data processing code is in the data folder.

Usage

Once you have texar installed and have cloned the project, try:

python spamGAN_train.py spamGAN_config_minimal

Which will run one epoch of each of the components.

Running

python spamGAN_train.py spamGAN_config_smallunsup

will run a full training process for the data in the base folder.

Adjusting the model file or the model configuration file should provide a springboard for exploration. If you wanted to precisely replicate the experiments in the paper, the code from the data and experiment_code folder can do that, albeit requiring some tweaking.