ReviewRadar

Inspiration

Google reviews are critical for businesses and consumers alike, but they're often flooded with irrelevant posts and spam advertisements. This makes it hard for consumers to trust the reviews, which in turn can impact businesses negatively. I wanted to build a model that attempts to separate genuine reviews from unhelpful ones to give consumers greater insight on a location.

What it does

ReviewRadar classifies reviews into 4 categories:

Normal - genuine reviews
Advertisement - promotional or spammy reviews
Irrelevant - Off topic reviews
Non-visitor - Reviews from people who did not visit the location

How we built it

Collected and labelled review data (with a synthetic generator)
Extracted features from reviews using TF-IDF and custom features such as review length and link detection
Trained an XGBoost Classifier
Validated performance on test data and new unseen reviews
Created a review generator to generate new reviews efficiently

Challenges we ran into

Creating synthetic data As I could not find a dataset with a substantial number of reviews that go against policies, I had to use advanced models like GPT to create synthetic data. Initially, I did not prompt engineer it enough to give me varied data, which resulted in the XGBoost model overfitting and performing poorly when I tested it on external data.
Large datasets Handling large datasets (big JSON files) and finding a way to upload them onto Github.

Accomplishments that we're proud of

Built a working multi-class classification pipeline that flags ads with links
Achieved rather high performance on test data
Created a data generator script so that the whole workflow can be reproduced easily