To identify whether or not review sentiments are a good indicator of book approval/popularity. To check whether they are as good or better than star ratings as a predictor.
As such, I focused on thre questions:
Q1 - Do reviewer's ratings of the work differ from their sentiments?
Q2 - Do review sentiments change over time? I.e. will reviews given up to 6 months after publications differ in sentiment from those given after 6 months for the same book?
Q3 - Which feature, ratings or reviews, is a better predictor of overall popularity (approximated by number of reviews)?
tensorflow
keras
scikit.learn
scipy.stats
UCSD Bookgraph
Stanford AI Lab IMDB dataset
A BERT model was trained on the IMDB reviews, this can be found in the training folder.
This model was then applied to the Goodreads data. The data was prepared as per the data_preparation folder. The files c_splitter and e_combiner only serve to break the data into chunks for easier processing.
The three questions are answered by analyses in the following files:
Q1 - rev_rating_diff
Q2 - early_late_review
Q3 - feature_imp