In e-commerce, product reviews play a critical role in consumers’ purchasing decisions. However, fake opinions generated by fraudsters to promote their products and defame competitors can directly influence e-commerce platform reputation and revenue.
Rochester Institute of Technology assistant professors Ali Tosyali and Gijs Overgoor have developed a method for detecting five-star fakes using network structure. Their model flags highly connected and clustered products that likely buy fake reviews.
With the rapid growth of online shopping, product quality information becomes a critical source to shape buying decisions. However, the reliability of product quality information can be undermined by review manipulation through artificial positive reviews, resulting in misrepresentation of products and sellers.
To help combat this, a new approach to identify fake reviews and their buyers has been developed by Rochester Institute of Technology assistant professors Ali Tosyali and Gijs Overgoor. Their method uses network structure to identify fakes by detecting patterns of connections between products and reviewers who share them.
The authors’ model identifies cluster data of reviews that are likely to be fake, using unsupervised learning techniques such as K-means and hierarchical clustering. It then identifies reviewers with whom those clusters are most associated, based on the number of common ties between them. Those reviewers are then investigated to determine the likelihood that they are the authors Is Legit or Scam Reviews. The results show that product characteristics mediate the relationship between authorial writing styles and reviewing behavior, and rating moderates this relationship.
The holidays are a busy time for online shopping and shoppers often turn to customer reviews to separate solid offerings from the duds. But fake reviews are a real problem, damaging trust and sales for both sellers and buyers.
Fake reviewers can be spotted by looking for clues in their profile, such as using a generic name (e.g., John Smith) or a picture that is used elsewhere on the Internet. They also tend to write reviews quickly, in a few days or less.
For many sellers, reducing the number of fake reviews is an important objective. One way to do so is to ensure that all buyers who write reviews have verified purchases of the products. Various tools exist for checking this, such as cross-referencing a buyer’s email address with purchase history or sign-ins to other online accounts. Another tool is fraud-detecting artificial intelligence. However, these methods can be subject to false positives or even be manipulated by malicious agents seeking to promote their own work or products.
Many e-business platforms are vulnerable to fraudulent product reviews that mislead customers and harm the reputation of products or platforms. Identifying these fake reviews is challenging as they are hard to distinguish from truthful ones and cannot be easily manipulated with traditional discrete manual characteristics.
To solve this problem, various machine learning models have been proposed. Among them, deep learning models have shown the best performance. [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
Neural network-based text classification methods use word embedding to represent review text as a vector that preserves contextual information. The resulting vectors are fed into the classifier, which categorizes the review as either fake or truthful.
Goswami et al.  collected and preprocessed data from the Yelp website, extracted social and behavioural indications, and used a backpropagation neural network to detect fake reviews. Their model achieved a detection accuracy of 95%. Ren et al.  combined a CNN with a gated recurrent neural network to detect opinion spam in both in-domain and cross-domain reviews, achieving an accuracy of 83%.
Online reviews can play a vital role in influencing customers purchasing decisions. However, some reviews are fraudulent and misguide buyers. A new study from Rochester Institute of Technology assistant professor Ali Tosyali and colleagues aims to identify the buyers behind phony reviews.
Although the impact of writing style and authorial linguistics on identifying fake reviews have been explored extensively, the impact of such features in nudging customer buying preferences remains to be fully understood. In order to fill this gap, this research uses a natural language processing technique to uncover latent factors underlying both genuine and fake reviews.
Results indicate that genuine reviews have profound informational content pertaining to product specific attributes that nudge consumer buying intentions. Furthermore, whereas ratings moderate the relationship between these characteristics and buying persuasion for genuine products, this effect is absent for fake products. Specifically, for the sports products category, the underlying characteristics that influence buyer buying preferences are utility, weight, material, quality and style.