Modeling Trust for Recommender Systems using Similarity Metrics 107 niques are not suitable to handle. This is because existing solutions for encoding trust either deal with data in an unsuitable form(see[ 14] beta pdf) or do not provide links to similarity. In our opinion it is not appropriate for users to be asked to pro- vide trust measures for others, mainly because this requires skills and adequate ex- perience that not all users have 4 Our Approaches In general, trust models are used to enable the parties involved in a trust relationship to know how much reliance to place on each other. Our model aims to provide a method for estimating how much trust two entities can place in each other, given the similarities between them. The problem that emerges when Trust is to be used in a recommender system is the fact that the entities invol ved usually provide their views in the form of ratings about items and not as their trust estimates about other entities. That means to bene fit from such model it is required that all user ratings be transformed into trust val ues. We are contributing to sol ving this issue by proposing and comparing various formulas for encoding direct trust. The first formula we propose in paragraph 4 has already been used in an experimental P2P recommender system which has been studied in [18]. The other new modeling approaches we propose are extensions of the same idea. The significant difference, though, between the existing and the new approaches is found in the way we model the uncertainty property. In all the new approaches we keep the main method of modeling uncertainty the same but we change the way that the remaining properties(belief and disbelief) are shaped 4.1 The existing appro Unlike the other modeling concepts we discussed above, such as beta pdf modeling, in our first approach we use both quantitative and qualitative criteria on the evi dence to derive uncertainty. In order to achieve this, we consider the ratings that us- rs have given to items as the behavioral data required for the composition of opi- nions. In order to capture this requirement in our model we assume that the level of trust that develops between every pair of entities is based on how similar they perce ive each others choices to be. We used the pearson coefficient, as this is the best known and most suitable coefficient for this type of application. It can take values between-I and I where two entities are considered as having higher similarity when their Pearson values are close to I and as completely dissimilar when the Pearson Coefficient is-1. A value of0 would mean that there is no relationship between the two entities at all. Bearing in mind the idea that those entities whose ratings can be accurately predicted should be considered as trustworthy sources of information, the uncertainty in such relationships should be lower Thus, in this approach we have re-defined the perception of Uncertainty as the inability of some entity to make accurate predictions about the choices of the otherniques are not suitable to handle. This is because existing solutions for encoding trust either deal with data in an unsuitable form (see [14] beta pdf) or do not provide links to similarity. In our opinion it is not appropriate for users to be asked to pro￾vide trust measures for others, mainly because this requires skills and adequate ex￾perience that not all users have. 4 Our Approaches In general, trust models are used to enable the parties involved in a trust relationship to know how much reliance to place on each other. Our model aims to provide a method for estimating how much trust two entities can place in each other, given the similarities between them. The problem that emerges when Trust is to be used in a recommender system is the fact that the entities involved usually provide their views in the form of ratings about items and not as their trust estimates about other entities. That means, to bene￾fit from such model it is required that all user ratings be transformed into trust val￾ues. We are contributing to solving this issue by proposing and comparing various formulas for encoding direct trust. The first formula we propose in paragraph 4.1 has already been used in an experimental P2P recommender system which has been studied in [18]. The other new modeling approaches we propose are extensions of the same idea. The significant difference, though, between the existing and the new approaches is found in the way we model the uncertainty property. In all the new approaches we keep the main method of modeling uncertainty the same but we change the way that the remaining properties (belief and disbelief) are shaped. 4.1 The existing approach Unlike the other modeling concepts we discussed above, such as beta pdf modeling, in our first approach we use both quantitative and qualitative criteria on the evi￾dence to derive uncertainty. In order to achieve this, we consider the ratings that us￾ers have given to items as the behavioral data required for the composition of opi￾nions. In order to capture this requirement in our model we assume that the level of trust that develops between every pair of entities is based on how similar they perce￾ive each other’s choices to be. We used the Pearson coefficient, as this is the best known and most suitable coefficient for this type of application. It can take values between -1 and 1 where two entities are considered as having higher similarity when their Pearson values are close to 1 and as completely dissimilar when the Pearson Coefficient is -1. A value of 0 would mean that there is no relationship between the two entities at all. Bearing in mind the idea that those entities whose ratings can be accurately predicted should be considered as trustworthy sources of information, the uncertainty in such relationships should be lower. Thus, in this approach we have re-defined the perception of Uncertainty as the inability of some entity to make accurate predictions about the choices of the other Modeling Trust for Recommender Systems using Similarity Metrics 107