A Review Helpfulness Modeling Mechanism for Online E-commerce: Multi-Channel CNN End‑to‑End Approach

Abstract

Review helpfulness prediction aims to provide helpful reviews for customers to make purchase decisions. Although many studies have proposed prediction mechanisms, few have introduced consistency between the review text and star rating information in the review helpfulness prediction task. Moreover, previous studies that have reflected such a consistency still have limitations, including the star rating facing information loss, and the interaction between review text and star rating not extracted effectively. This study proposes the CNN-TRI model to overcome these limitations. Specifically, this study applies a multi-channel CNN model to extract semantic features in the review text and convert star ratings into a high-dimensional feature vector to avoid information loss. Next, element-wise operation and multilayer perception are applied to extract linear and nonlinear interactions to learn interaction effectively. Results measured by real world online reviews collected from Amazon.com show that CNN-TRI significantly outperforms the state-of-the-art. This study helps e-commerce websites with marketing efforts to attract more customers by providing more helpful reviews and thus, increasing sales. Moreover, this study can enhance customers’ attitudes and purchase decision-making by reducing information overload and customers’ search costs.