Enhanced Multi Feature based Machine Learning Techniques for Identification of Online Spam Reviews

Abstract

In any online e-commerce business, online reviews are powerful medium for expressing opinions, where the ratings and customer feedbacks are the centerpiece. In the past decade, spam reviews have trouble the e-commerce sector worldwide. These reviews can be used to control the sentiment of a product in a negative manner and therefore, are considered extremely harmful to online businesses and individuals alike. In order to detect and remove these harmful reviews, this research work proposes spam detection system that consists of two steps, namely, feature engineering and classification. newlineThe research methodology was designed in three phases, where the first phase focus on feature engineering, while the second and third phases focus on the classification step. Phase I (feature engineering) performs two tasks, namely, feature extraction and optimal feature vector construction. Task 1 extracts three types of features, namely, review centric features, reviewer centric features and product centric features. A total of 50 features were extracted. In order to handle the problem of high dimensionality, a feature selection algorithm that combined enhanced maximum relevant minimum redundant filter based algorithm with ant colony optimization algorithm combined with genetic algorithm, was proposed. newlinePhases II and III used this optimal feature vector. Phase II focused on designing an enhanced ensemble system. The base classifier used was support vector machine which was enhanced to improve its accuracy and reduce its time complexity. A hyperplane construction method that is based on Mahalanobis distance measure was proposed to improve the accuracy of the Support Vector Machine (SVM) classifier. Similarly, a speed optimization procedure that removed irrelevant support vectors was also proposed. The ensemble system was constructed by differing the kernel function used by the SVM classifier. In Phase III, hybrid spam review detection systems that combined clustering, classification and the enhanced ensemble system from Phase II was