Fraudulent E-Commerce Websites Detection Through Machine Learning

Experimental Evaluation of Possible Feature Combinations for the Detection of Fraudulent Online Shops

“…The work most similar to our study is Sánchez-Paniagua et al [38]. The authors of [38] also proposed to use features obtained from the source of the website and third parties, including Trustpilot [31] and WHOIS [33], but in addition they use data from social network platforms together with features obtained from the metadata of the website.…”

“…The work most similar to our study is Sánchez-Paniagua et al [38]. The authors of [38] also proposed to use features obtained from the source of the website and third parties, including Trustpilot [31] and WHOIS [33], but in addition they use data from social network platforms together with features obtained from the metadata of the website. These features include high discounts, social media footprint, domain age, registration date, SSL names, country and issuer, Trustpilot score and review, e-commerce technologies and policies.…”

“…Training and testing sets were used to train and evaluate machine learning classifiers, namely Decision Tree, Random Forest, Stochastic Gradient Descent, Logistic Regression, Gaussian Naive Bayes, Multilayer Perceptron, and XGBoost. The reason for the selection of these classification models was their diversity and frequent use in previous studies demonstrating good classification results [17,37,38].…”

“…Seven of the most significant features ensured a classification accuracy of 0.9605, while the best accuracy of 0.9693 was achieved using 13 and 15 features. Therefore, our research allowed us to identify the small set of the most significant features, which allows one to detect fake shops with high accuracy, avoiding complex, time-consuming computations and analysis of web shop content, such as, for example, comparing shop prices with other known legitimate e-shops [38].…”

“…They used features including tokenized HTML, CSS and JavaScript text, comments and individual tags, tag-attribute-value patterns, and the HTML tree structure, reaching the accuracy of up to 0.97 and a very high degree of certainty in classifying fraudulent e-commerce. Sánchez-Paniagua et al [38] proposed to use features obtained from third-party sources, including Trustpilot [31] and WHOIS [33] together with features obtained from the website source and metadata. These features include high discounts, social media footprint, domain age, registration date, SSL names, country and issuer, Trustpilot score and review, e-commerce technologies and policies.…”