Dynamic Trust Production Based on Based on Interaction Sequence

New emerging applications such as e-commerce and peerlo-peer systems require autonomous agents (0 dynamically determine trust in open environments. Trust is a uustec's opinion towards a lruslee established lhrough direct interactions or recommendations. Many research efforts are on trust propagation via recommendations. We investigate lhe trust production problem that determines trust from a sequence of interactions (i.e. raled interactions). This paper presents fOUf trust production rules: "equal-weight-interaction" I "dilute-with-time", "sIOW-lrust-quick-distrust". and "slow-trust-quickdistrust-with-supervision". These rules can be implemented by using an update or an analysis algorithm. An update algorithm only maintains a trust state, which is modified when a new interaction becomes available. An analysis algorithm maintains the whole interaction sequence and evaluates trust based on it. Four types of user behaviours, "stable user", "repenting user", "cheater" and "smarter cheater", are modelled. The proposed rules are studied via simulation by using these models. The results demonstrate that these rules make similar judgments on a stable user who behaves consistently well or badly. All rules except "equal-weight-interaction" are able lo react to sharp changes in behaviours (a "repenting user" or a "cheater") within 10 interactions. The "slowtrust-quick-distrust-with-supervision" rule can even catch a "smart cheater", who repeatedly misbehaves in short periods and then tries to cover the bad effect by good behaviours.