Trust-Based Analysis of E-Commerce Reviews Using Sentiment Similarity

Abstract

Consumer reviews in e-commerce systems serve as important resources that significantly affect user purchasing decisions, experience, feelings, and willingness to buy products. These reviews contain consumers' views on products that express interests, sentiments, and opinions, providing valuable information for both potential buyers and sellers. Research in social psychology has consistently shown that people are more likely to trust and follow recommendations from individuals who share similar attitudes and preferences toward similar products. This paper proposes that seeking and accepting sentiments and suggestions in e-commerce systems implies a form of trust between consumers during shopping, and develops a sentiment similarity-based trust analysis framework. Trust is categorized into two types: direct trust, computed from sentiment similarity between pairs of users who have reviewed common products, and propagation trust, computed through trust transitivity across the reviewer network using an improved shortest path algorithm. An entity-sentiment word pair mining method is developed for extracting similarity features from review text, where entities represent product aspects and sentiment words capture the polarity and intensity of opinions. The propagation trust is calculated according to the transitivity feature of trust relationships, where trust can flow through chains of mutually trusting users. Using the proposed trust representation model, the shortest path in the trust graph describes the tightness of trust relationships. Evaluation on a large-scale e-commerce dataset containing 50,000 reviews from 8,000 users across 5,000 products demonstrates 88.6% trust prediction accuracy and 0.86 correlation with actual user trust ratings, significantly outperforming collaborative filtering approaches (0.72 correlation).