20.3.2.2 Automatic Trust Generation
The algorithms discussed in the previous section require explicit trust input from the users. As a consequence, the applications that use such an algorithm must provide a means to obtain the necessary information; think e.g. of FilmTrust or Moleskiing. However, this might not always be possible or feasible. In such cases, methods that automatically infer trust estimates, without needing explicit trust information, might be a better solution. An example of such a system can be found in [47].
Most commonly, these approaches base their trust generation mechanism on the past rating behaviour of the users in the system. More specifically, deciding to what degree a particular user should participate in the recommendation process is influenced by his history of delivering accurate recommendations. Let us exemplify this with the well-known approach of O’Donovan et al. [46].
Profile- and item-level trust Our intuition tells us that a user who has made a lot of good recommendations in the past can be viewed as more trustworthy than other users who performed less well. To be able to select the most trustworthy users in the system, O’Donovan introduced two trust metrics, viz. profile-level and item-level trust, reflecting the general trustworthiness of a particular user u, and the trustworthiness of a user u with respect to a particular item i, respectively. Both trust metrics need to compute the correctness of u’s recommendations for the target user a. In particular, a prediction pa,i that is generated only by information coming from u (hence u is the sole recommender) is considered correct if pa,i is within ? of a’s actual rating ra,i.
The profile-level trust tP
u for u is then defined as the percentage of correct recommendations that u contributed. Remark that this is a very general trust measure; in practice it will often occur that u perfoms better in recommending a set of specific items. To this aim, O’Donovan also proposed the more fine-grained item-level trust tiu, which measures the percentage of recommendations for item i that were correct.
Hence, in such automated approaches, trust values are not generated via trust propagation and aggregation, but are based on the ratings that were given in the past.
Remark that O’Donovan’s methods are global trust metrics. The way the values are obtained can be seen as probabilistic.
Trust-based filtering Similar to other trust-enhanced techniques, the values that are obtained through the trust metric are used as weights in the recommendation process. Just like Massa, O’Donovan et al. focus on trust-based adaptations of collaborative filtering. In [46] they investigate several options, such as combining the obtained trust values with PCC information. An alternative to this scheme is to use trust values as a filter, so that only the most trustworthy neighbours participate in the recommendation process. This strategy is called trust-based filtering, see Formula (20.5) in which wa,u denotes the PCC and RT+ = RT ?R+.