Date of Award

Spring 5-2021

Document Type

Thesis (Master's)


Department of Computer Science

First Advisor

Venkatramanan Siva Subrahmanian

Second Advisor

Soroush Vosoughi

Third Advisor

Mauro Conti


With the popularity of e-commerce and review websites, it is becoming increasingly important to identify the helpfulness of reviews. However, existing works on predicting reviews’ helpfulness have three major issues: (i) the correlation between helpfulness and features from review text is not clear yet, although many standard features are proposed, (ii) the relations between users, reviews and products have not been considered, (iii) the effectiveness of the existing approaches have not been systematically compared. To address these challenges, we first analyze the correlation between standard features and review helpfulness that are widely used in other work. Based on this analysis, we propose an end-to-end neural network architecture, the Global-Local Heterogeneous Graph Neural Networks (GL-HGNN). It consists of the graph construction and learning nodes representations both globally and locally. The graph is composed of three types of nodes including users, reviews and products, as well as four link types to build connections among these nodes. To better learn the feature representations, we employ a global graph neural network (GNN) branch and a local GNN branch on the whole graph and associated subgraphs to capture graph structure and information propagation. Finally, we provide an empirical comparison with traditional machine learning models training on hand-crafted features as well as four state-of-the-art deep learning models on eight Amazon product categories.