In scope of this research project, we have experimented with state of the art approaches for Natural Language Processing, in order to identify fake product reviews at one of the leader Virtual Merchants of the world. All of our models are trained on the amazon_reviews dataset, which explicity has fake reviews labeled.
1 - Transfer Learning with Tensorflow Hub Model
- Experiments were conducted on a simple model by using a pre-trained model provided through tensorflow hub. Weight were frozen and unfrozen to observe the different training behaviour.
2 - Fake review detection review using biLSTM
- We use keras functional API to build a architecture with 3 heads to process 3 different features. The head that is used to process the 'text' has the biLSTM. The result of the heads are concatenated and processed further. In this notebook we show how keras functional API could be used to build complex architectures.
3 - Explainable AI(XAI)
- In this notebook we show how a smart feature engineering could help us build a simple model that is interpretable. We encode the information from categorical features into the text. The text is then process by a model build using the Keras Sequential API. 'LIME' method of explanations are then used to explain the predictions of the model.
4 - Fake review detection review using biLSTM and Word2Vec
- In this notebook we have the same model architecure as it is presented on section 2 -"Fake review detection review using biLSTM". The difference is that we are importing the Word2Vec pretrained model from the gensim library and we finitune it on our current dataset.
- Also we use two hyperparameter tuning algorithms (Bayesian and Hyperband) in order to get the best accurasy results based on the hyperparameter tuning space that we have defined.
5 - Transformer Model of distilled Bert
- We enrich the review text with the information of whether a purchase was verified or not and the rating given before feeding it to the transformer network. The network is then fine-tuned to be able to detect fake amazon reviews.
6 - Data Visualization and unsupervised learning
- We use Doc2Vec model to encode review text into 300 dimensional vector. The goal was to find any statistical difference between fake and authentic reviews or good and bad reviews to automatically detect them. However, after reducing the dimesnion to 2D using PCA and t-SNE, we discover that there is no significant statistical difference.
- Inside the folder Anomaly-Detection there are scripts that try to fit normal distribution to the embeddings, as well as an AutoEncoder architecture that tries to learn latent lower level representation of the embeddings.