A Robust Collective Classification Approach to Trust Evaluation

In this paper, we present a collective classification approach for identifying untrustworthy individuals in multi-agent communities from a combination of observable features and network connections. Under the assumption that data are organized as independent and identically distributed (i.i.d.) samples, traditional classification is typically performed on each object independently, without considering the underlying network connecting the instances. In collective classification, a set of relational features, based on the connections between instances, is used to augment the feature vector used in classification. This approach can perform particularly well when the underlying data exhibits homophily, a propensity for similar items to be connected. We suggest that in many cases human communities exhibit homophily in trust levels since shared attitudes toward trust can facilitate the formation and maintenance of bonds, in the same way that other types of shared beliefs and value systems do. Hence, knowledge of an agent’s connections provides a valuable cue that can assist in the identification of untrustworthy individuals who are misrepresenting themselves by modifying their observable information. This paper presents results that demonstrate that our proposed trust evaluation method is robust in cases where a large percentage of the individuals present misleading information.