Trust and distrust with strength of relations
Trust and distrust with strength of relations
In our extension of our PST 2013 paper, we investigate the impact of relation strength in trust decisions. This work is accepted to appear in ACM TWEB journal. The software for this work is available at: github.com/rpitrust/structuralbalance, and a preliminary version of the paper can be found here [YiAdali_TWEB2014.pdf].
In trust prediction, there is a distinction between trust and distrust. When a person is not trusted and not distrusted, it is possible to develop trust for them over time as the individual acts trustworthy. This is not true for distrust. A distrusted individual may never gain one’s trust even if he acts trustworthy. It is possible that such actions are viewed even as a manipulative way to gain trust. In short, the evidence regarding trust and distrust should be handled differently.
Another important aspect of trust relations is the level of strength. Even though structural balance theory is based on the level of comfort a person feels in relationships, it does not describe how this comfort changes based on the level of strength in relationships. Strong trust relationships can be considered to model close friendships and trust based one’s reliability in high risk situations. While a weak trust relationship may model lower risk decisions and those that involve one’s ability to make correct recommendations in such decisions. Equivalent notions of strong and weak trust can also be considered. This duality between competence and trustworthiness based trust is discussed in many different fields from social and cognitive psychology to information sciences, and discussed in my book in detail as well as in recent agent models.
In this paper, we show that we can achieve much better prediction of whether a link is likely to be a trust or distrust link by taking into account relation strengths in multiple networks. For example, in Epinions, we consider bidirectional edges with the same sign as strong links (Bi) and single directional edges as weak links (S). Our algorithm only looks at the existence of edges, not the reviews that accompany the trust and distrust votes. We run our algorithm to guess the strength of the links after convergence as modeled by our method.
Given our notion of convergence, we look at the average rating for each link, in other words, if A trusts B, what is the rating that A gives to the reviews of B? It is higher for trusted edges than distrusted edges of course, and the average rating goes up over time, but more so for trusted edges. However, we look at the edges that become positive or negative after convergence according to our algorithm. Recall that this models the social pressure to become more positive or negative based on neighbor’s opinion of a person. We see that for edges that became positive according to our algorithm, the average rating went up considerably. This effect is not visible in competing methods. Our method manages to capture the impact of this type of social pressure on trust accurately!
Armed with this strong result, we are looking for new applications of our algorithm. The take home message is that having negative links is very important to ascertain true negative opinions in networks, but even such negative opinions may change over time due to the opinions of others.
March 2014