Modeling Trust for Recommender Systems using Similarity Metrics 105 cause not much evidence can be gathered to support a recommendation. This is usually because users themselves are unwilling to invest much time or effort in rat ng items. In existing CF systems users can only be correlated through their com- mon experiences, so in the presence of limited data they turn out to be unable to make accurate predictions. The idea to enhance the neighboring base of users, by using the potentially developed trust relationships between them, could make it possible to reach other members of the community through them. Assuming that the potential trust between the users could help in reducing the number of empty cells in the matrix by allowing missing values to be predicted from xisting ones, finding a way of computing that trust from the existing data(user ex- periences) might help to alleviate the problem. For such an idea to be applicable, it is necessary that, somehow, users must be able to place trust on their neighbors. In some centralized consumer opinion sites [11 it is a requirement that this trust measure should be provided by the users them selves. However, this requires that users should have developed some instinct in judging things accurately, and this cannot be assured. Poor judging abilities intro- duce the danger of establishing relationships with wrong counterparts. Our approach to this issue is to introduce a technique for mapping between similarity measures and trust, and which will be done automatically on behalf of the users In our model we use ordinary measures of similarity taken from CF to form the potential trust between the correlated entities which would be propagated in a simi lar way to the word-of-mouth scheme. In that scheme the trust that the first entity should place on the distant one is derived through a trust graph. Finally, by trans forming the value back into similarity measure terms it could be made appropriate for use in CF algorithms. However, to our knowledge, today there is no standard pproach for modeling trust from such type of existing evidence. In this work as well as in a previous one [9] we express trust in the form of opinions as they are modeled in Subjective Logic [10]. In this theory trust is considered as a subjective measure and introduces the important idea that there is always imperfect knowledge when judging things. The latter is expressed with a notion called uncertainty and is present when trust is based on user observations. Another interesting point of sub- jective logic is that it provides an algebra for combining direct and indirect trust along chains of users. Direct trust is considered the trust that is built upon first hand evidence or else derived from experience with the trustee. Indirect trust is built upon recommendations from others when first hand evidence is not present The use of trust in transitive chains requires the existence of a common [8] which needs recommender trust to be derived or given from a specific tra chain. This has either to be modeled from relevant evidence or. somehow must be enabled to derive it from past experiences Our work in this paper is concerned with the construction of trust re Ising first hand evidence, which in our case is the users'ratings. More specifically we try various similarity-to-trust transformation formulas with the purpose of find- ng the most suitable one. In the future we aim to evaluate the accuracy of a whole recommender system that employs the proposed transformation formulacause not much evidence can be gathered to support a recommendation. This is usually because users themselves are unwilling to invest much time or effort in rat￾ing items. In existing CF systems users can only be correlated through their com￾mon experiences, so in the presence of limited data they turn out to be unable to make accurate predictions. The idea to enhance the neighboring base of users, by using the potentially developed trust relationships between them, could make it possible to reach other members of the community through them. Assuming that the potential trust between the users could help in reducing the number of empty cells in the matrix by allowing missing values to be predicted from existing ones, finding a way of computing that trust from the existing data (user ex￾periences) might help to alleviate the problem. For such an idea to be applicable, it is necessary that, somehow, users must be able to place trust on their neighbors. In some centralized consumer opinion sites [1] it is a requirement that this trust measure should be provided by the users them￾selves. However, this requires that users should have developed some instinct in judging things accurately, and this cannot be assured. Poor judging abilities intro￾duce the danger of establishing relationships with wrong counterparts. Our approach to this issue is to introduce a technique for mapping between similarity measures and trust, and which will be done automatically on behalf of the users. In our model we use ordinary measures of similarity taken from CF to form the potential trust between the correlated entities which would be propagated in a simi￾lar way to the word-of-mouth scheme. In that scheme the trust that the first entity should place on the distant one is derived through a trust graph. Finally, by trans￾forming the value back into similarity measure terms it could be made appropriate for use in CF algorithms. However, to our knowledge, today there is no standard approach for modeling trust from such type of existing evidence. In this work as well as in a previous one [9] we express trust in the form of opinions as they are modeled in Subjective Logic [10]. In this theory trust is considered as a subjective measure and introduces the important idea that there is always imperfect knowledge when judging things. The latter is expressed with a notion called uncertainty and is present when trust is based on user observations. Another interesting point of sub￾jective logic is that it provides an algebra for combining direct and indirect trust along chains of users. Direct trust is considered the trust that is built upon first hand evidence or else derived from experience with the trustee. Indirect trust is built upon recommendations from others when first hand evidence is not present. The use of trust in transitive chains requires the existence of a common purpose [8] which needs recommender trust to be derived or given from a specific transitive chain. This has either to be modeled from relevant evidence or, somehow, trustors must be enabled to derive it from past experiences. Our work in this paper is concerned with the construction of trust relationships using first hand evidence, which in our case is the users’ ratings. More specifically we try various similarity-to-trust transformation formulas with the purpose of find￾ing the most suitable one. In the future we aim to evaluate the accuracy of a whole recommender system that employs the proposed transformation formula. Modeling Trust for Recommender Systems using Similarity Metrics 105