Representing Context for Multiagent Trust Modeling

We present a universal mechanism that can be combined with existing trust models to extend their capabilities towards efficient modelling of the situational (context-dependent) trust. The mechanism describes the similarity between the situations using their distance in a metric space and defines a set of reference contexts in this space to which it associates the trustfulness data. The data associated with each reference context is updated and queried with the weight that decreases with distance between the current situation and the reference context. In the presented mechanism, we use Leader-Follower clustering to place the reference contexts to be representative of the data. In an empirical test, we show that context-aware models easily outperform the general trust when the situation has an impact on partner trustfulness and that their performance and efficiency is comparable with general trust models when the trustfulness is independent of the situation. Multi-context nature of the model also expands its use towards more advanced uses, allowing policy/norm learning from at the trust model at runtime, as well as reasoning based on uncertain identities.