Spam review detection

Abstract

As more people depend heavily on the information presented on the web, user generated content like reviews could easily influence the purchase decisions of other consumers. As such, multiple fake reviews have been frequently posted to various popular online review websites to mislead the consumers.

Several studies have also been made in spam review detection. However, most research focus on specific review websites such as either Amazon or Yelp. Therefore, this raised a question whether these observed features suggested in these research papers could perform equally well in other domains such as TripAdvisor.

In this project, a series of progressive phases were employed to implement algorithm that would detect these spam reviews with referenced to the suggested set of features and procedures. In total, three different types of features, N-Grams features, review centric features and user behavior features were chosen for the study.

From the experiments, N-Grams features generally generate a better accuracy than review centric features with a difference in accuracy ranges from 10% to 30%. User behavior features consistently outperforms the other two sets of features with an average accuracy of 60% and above.

Despite the limitations in this project, it is evident from the findings that the features relating to user behaviors gives the best accuracy among the rest which means that it is more versatile.