Constructiveness-Based Product Review Scoring Using Machine Learning

Abstract

To make the internet a more productive environment, it is vital to promote constructiveness in online discussion forums. Customers are regularly offered the chance to share their thoughts and experiences with a product on online marketplaces. Generally, online products have fewer constructive reviews, and some of them are unrelated to the product. Existing approaches focus on textual features to classify a product's constructiveness and ignore semantic and contextual information about the reviews. The directed graph model has been utilized in this study to represent information about the product. Also, the node and graph level features like average in-degree, out-degree, and clustering coefficients are used to model constructiveness in product evaluation to encourage the most informative reviews. Graph embedding techniques are used to depict each node as a vector into low-dimensional space and preserve the structure and properties of the graph as well. The topic modeling approach has been used to contextualize the reviews with the appropriate product. Additionally, we employed logistic regression, random forest, Gaussian naive Bayes, support vector machine (SVM), and Gradient Boosting Machine models trained on Amazon product reviews and constructive news corpus for constructiveness. These ML models outperform the baseline approach, achieving a 90% F1-Score.