Predicting users’ review helpfulness: the role of significant review and reviewer characteristics

Abstract

There are thousands of reviews constantly being posted for popular products on e-commerce websites. The number of reviews is rapidly increasing which creates information overload problem. To solve this problem, many websites introduced a feedback mechanism to vote for a review (helpful or not). The attracted votes reflect the review helpfulness. This study addresses the review helpfulness prediction problem and investigated the impact of review, reviewer and product features. Multiple helpfulness prediction models are built using multivariate adaptive regression, classification and regression tree, random forest, neural network and deep neural network approaches using two real-life Amazon product review datasets. Deep neural network-based review helpfulness prediction model has outperformed. The results demonstrate that review-type characteristics are most effective indicators as compared to reviewer and product type. In addition, hybrid combination (review, reviewer and product) of proposed features demonstrates the best performance. The influence of product type (search and experience) on review helpfulness is also examined, and reviews of search goods show strong relationship to review helpfulness. Our findings suggest that polarity of review title, sentiment and polarity of review text and cosine similarity between review text and product title effectively contribute to the helpfulness of users’ reviews. Reviewer production time and reviewer active since features are also strong predictors of review helpfulness. Our findings will enable consumers to write useful reviews that will help retailers to manage their websites intelligently by assisting online users in making purchase decisions.