Computational Trust Modeling in Social Networks and e-Commerce Environments

Abstract

Computational trust applies the notion of human trust to the online world and has been studied in several computer science fields such as social networks and e-commerce. A variety of approaches to computational trust have been proposed for different application scenarios. In this dissertation, we focus on promoting online trust in social networks and e-commerce like movie websites and e-marketplaces.

In social networks, the problem of computational trust is to determine those users who are not directly connected to a target user in a social network but the target user is able to trust. This problem is referred to as trust prediction. In this dissertation, we classify previous trust prediction studies into two categories: one-way trust prediction and two-way trust prediction, and propose a method for one-way prediction and another method for two-way trust prediction. First, we develop a new one-way trust prediction method that collectively exploits two kinds of information, explicit trust information and user interactions, which have been treated separately in previous researches. The excellence of the proposed method is demonstrated with a real-life data set in comparison with previous methods. Second, we identify underlying significant factors for two-way trust formation (trust factors) and develop two-way trust prediction methods that use ranking and classification based on these factors. The experimental results show that most of our trust factors are useful in predicting two-way trust relationships by our ranking approaches and the proposed classification approaches using our trust factors significantly improve the performance (up to 37% by F1-Measure) of two-way relationship prediction compared with previous approaches.

In e-commerce environments, reputation systems can encourage trust by providing reliable reputation scores of entities (sellers/products/services) based on others’ opinions. In open e-commerce environments, however, it is not easy to build reliable reputation scores because some online users intentionally provide misleading opinions about sellers/products to promote them (ballot-stuffing) or to bad-mouth others. In this dissertation, we refer to this as false reputation problem and classify the situations where false reputation problems occur into two categories: false reputation of products and false reputation of online sellers. We propose two methods to solve the false reputation problems for products in movie websites and online sellers in e-marketplaces. First, we propose TRUE-REPUTATION, a method that iteratively adjusts a product reputation based on the confidence of customer ratings, which is determined by adopting three key factors of activity, objectivity, and consistency and define these factors in the context of online ratings. The effectiveness of TRUE-REPUTATION is shown through extensive experiments in comparison with the state-of-the art methods. Second, we claim that there is no guarantee of the reliability of a seller reputation and even it could be distorted when (1) some buyers intentionally give unfair ratings to a seller and (2) the computation process of a seller reputation does not realize the fact that a buyer’s rating is a compound score of the capability of a seller and the quality of an item. We propose RS&BSD, a robust reputation system that provides a seller reputation correctly based on the seller's score by repetitive separations and reduces the influence of unfair ratings. In our experimental evaluation, RS&BSD outperforms existing reputation systems significantly.