Authors: Parth Iramani, Sittun Swayam Prakash, Atrima Ghosh, Zoe Masood, Jenna Gottschalk, Mugundhan Murugesan
__DESCRIPTION:
The aim of our project is to detect fake reviews on Amazon using the review text. Our approach combines semi-supervised learning and transformer models to identify fake Amazon reviews. The Datasets used in our projects are: Labelled Dataset from Kaggle - https://www.kaggle.com/lievgarcia/amazon-reviews Unlabelled dataset from Amazon - https://s3.amazonaws.com/amazon-reviews-pds/tsv/amazon_reviews_multilingual_US_v1_00.tsv.gz
Following are the steps involved in creating and evaluation of the Model to predict fake reviews using review text: (1) We split the original labeled dataset into four parts: 70% training set, 10% first validation set to compare initial supervised classification models, 10% second validation set to compare the updated classification models, and 10% test set to evaluate the final classifier’s performance. (2) We trained the initial model on the training set of labeled data. (3) We generated pseudo-labels by using the initial model to classify the unlabeled data set. (4) The most confidently predicted pseudo-labels above a specific threshold became training data for the next step. (5) We updated the initial model using the pseudo-labels as training data. (6) Finally, we tested the final classifier on the test set of the original labeled data.
Modules used: Pandas, Numpy, scikit-learn - Commonly used Machine Learning python libraries transformers - Transformers provides thousands of pretrained models to perform tasks on texts such as classification, information extraction, question answering, summarization, translation, text generation and more in over 100 languages. datasets - Huggingface Datasets is a library for easily accessing and sharing datasets, and evaluation metrics for Natural Language Processing (NLP), computer vision, and audio tasks.
The following are the steps and respective code (.py files) invloved in obtaining the final model file that is used for prediction.
- Modeling - Sentence Transformer and getting embeddings for review text |___ 11initial_cleaning.py
- Logistic Regression with Labelled Dataset |___ 21initial_modelling.py
- Generation of Pseudo Labels |___ 31clean_final_dataset.py |___ 32get_partitions.py |___ 33get_embeddings_for_partition.py |___ 34pseudo_labels.py
- Unsupervised Learning using Stochastic Gradient Descent (SGD) |___ 41SGD_log.py
Analysis using BERT (Bidirectional Encoder Representations from Transformers) model:
- Supervised Learning |___ 1_6242_GROUP_PROJECT_BERT_training_for_a.py
- Generation of Pseudo_labels |___ 2_6242_GROUP_PROJECT_Generate_BERT_Pseudo_Labels.py
- Unsupervised Learning |___ 3_6242_GROUP_PROJECT_BERT_training_for_b.py
- Compilation of final dataset for visualization |___ 4_6242_GROUP_PROJECT_compiling_final_dataset
The public Tableau dashboard (https://public.tableau.com/app/profile/zoe.masood/viz/Book1_16383996536860/Dashboard1#1) is accessible for analysis of fake reviews.
__INSTALLATION:
The python code files include commands that will install all the required libraries.