Abstract: Online reviews are becoming the most important way for buyers to determine whether to buy a product or use an online service. Spammers have been promoting bogus reviews on the internet more and more to trick customers into buying things, with the goal of promoting businesses, products, and services to boost sales and marketing, frequently at the expense of quality and after-sales care. This study analyzes a comprehensive feature engineering approach for detecting fraudulent reviews by employing linguistic, statistical, semantic, and behavioral features. A number of machine learning classifiers have been trained, verified, and tested, including ensemble models on benchmark datasets viz. “Fake Reviews Dataset” and “Yelp Labeled Dataset”. Focus has been on how the engineering features affect the classification performance. Experimental results demonstrate that models with numerous attributes outperform baseline techniques significantly. The mentioned two datasets have been used for testing to make the results more reliable. The ensemble model is the best performer on the leaderboard, with F1-score of 89.10% for the Fake reviews dataset and 85.72% for the Yelp labeled dataset. The ranking of the top 20 features that help find false reviews for each dataset has been stressed to show how important feature engineering is for finding phony reviews. The results have been emphasized through implication of spam score calculations leading to revised star rating of product and online service reviews, and the predictions were consolidated with Explainable AI (XAI) assimilation enhancing model interpretation. These consequences provide substantial information regarding the primary indicators of fraudulent reviews.

Key words: feature engineering, natural language processing, machine learning, ensemble models, text classification, explainable AI