Scam Product Detector

Research Questions

• In light of the increased usage of online webpages to purchase products, you aren't sure are just sometimes which are legitimate products are something that was made to have been given 5 stars from bot rating accounts.
• Is it possible to train a model to predict whether a product crosses a given legitimacy threshold.

Prediction Methods

• The validation data required to compute a supervised learning model is a feature which could possibly annotated?
• Can possibly start with an unsupervised model. Feature weights will give an idea of what constitutes a legitimate product.
• Or possible to instead construct an algorithm that predicts bot reviews
• Translate reviews to English test
  • Deepl.com
    • Google translate api/plugin

Data Collection

• Download all reviews for a given set of products
• Use amazon certified products as a positive validation set
• Negative validation set ○ Certified scam list ○ Sketchy websites § Ebay § Craigslist § Amazon may have some too

Data Annotation

Native features:

• Start rating distribution (mean, CI, variance)
• Number of reviews
• Avg. review length

Generated Features

• Proportion of review from bots
• Need a model to predict which reviews are bots
• Adjusted non-bot star rating, num of reviews, avg review length
  • Occurrences of words from scam dictionary
    • i.e. {'fake', 'scam', '…' }

Techniques

• Natural Language processing ○ API? ○ Python package? § NLTK

Applications/UI

• Chrome extension