the most widely known technique used in Recommender systems and is based or the idea of making predictions using similarity metrics to correlate users However, Recommender Systems and particularly Collaborative Filtering are not perfect and as it is well known that they appear to have weaknesses such as a low uality of predictions( known as the false negatives and false positives problems 5D), caused by sparsity in the dataset. Also, the architectural characteristics of CF are known to be vulnerable to attacks from malicious and libelous users. CF systems employ statistical techniques to develop virtual relationships between users, and in this way, neighborhoods of users can be formed consisting of those who have a his- tory of agreeing and who are thus assumed to be similar. The virtual relationships are built upon a metric that is used for correlating the users based on their expe- riences and is called Similarity. In order to know how similar two users are with each other, a number of common experiences must exist Trust has been investigated by many researchers of recommender systems in the past [23] and proposed also as a potential solution to alleviate the previously men- tioned problems of recommender systems [6, 7]. Trust can also express integrity in relationships between entities and so can be used to express the quality of service providers. So, service consumers should be able to assess reliably the quality of ser- vices before they decide to depend on a particular instance. In order to know the trustworthiness of a service provider evidence needs to be provided to potential con- sumers from which they can derive their own trust for the provider Under appropriate circumstances(with regard to a common purpose), trust rela- tionships can also support transitivity [8] whereas similarity generally does not. In order to benefit from the special characteristics of trust such as the ability to propa gate along chains of trusted users, a formula for deriving it from similarity and vice versa is needed. In this way user entities that cannot be correlated due to lack of common experiences can benefit from each other and thus extend the and/or the quality of predictions they can make about their future choices tribution to this research problem is the provision of appropriate formulas be used for converting trust to similarity The rest of the paper is organized as follows. In the next section, there is a more detailed description of the problem. Section 3 includes related work in the field and section 4 we analyze our approach to the problem, showing the formulas we have introduced. Next in section 5 we present the evaluation we performed and some comparative resul ts which show the best candidate. Finally, in section 6 we discuss some future issues concerning the applicability of the proposed method 2 Motivation The main idea of collaborative filtering is to make predictions of scores based on the heuristic that two people who agreed (or disagreed) in the past will probably agree agree)again. A typical collaborative filtering system runs as a centralized ser- vice and the information it holds can be represented by a matrix of users and items Each value of the matrix represents the score that a particular user has given to some item. The number of empty cells is known as sparsity and as we mentioned in the previous section, it is the main reason that recommender systems behave poorly, bethe most widely known technique used in Recommender systems and is based on the idea of making predictions using similarity metrics to correlate users. However, Recommender Systems and particularly Collaborative Filtering are not perfect and as it is well known that they appear to have weaknesses such as a low quality of predictions ( known as the false negatives and false positives problems [5]), caused by sparsity in the dataset. Also, the architectural characteristics of CF are known to be vulnerable to attacks from malicious and libelous users. CF systems employ statistical techniques to develop virtual relationships between users, and in this way, neighborhoods of users can be formed consisting of those who have a his￾tory of agreeing and who are thus assumed to be similar. The virtual relationships are built upon a metric that is used for correlating the users based on their expe￾riences and is called Similarity. In order to know how similar two users are with each other, a number of common experiences must exist. Trust has been investigated by many researchers of recommender systems in the past [23] and proposed also as a potential solution to alleviate the previously men￾tioned problems of recommender systems [6,7]. Trust can also express integrity in relationships between entities and so can be used to express the quality of service providers. So, service consumers should be able to assess reliably the quality of ser￾vices before they decide to depend on a particular instance. In order to know the trustworthiness of a service provider evidence needs to be provided to potential con￾sumers from which they can derive their own trust for the provider. Under appropriate circumstances (with regard to a common purpose), trust rela￾tionships can also support transitivity [8] whereas similarity generally does not. In order to benefit from the special characteristics of trust such as the ability to propa￾gate along chains of trusted users, a formula for deriving it from similarity and vice versa is needed. In this way user entities that cannot be correlated due to lack of common experiences can benefit from each other and thus extend the quantity and/or the quality of predictions they can make about their future choices. Our con￾tribution to this research problem is the provision of appropriate formulas that can be used for converting trust to similarity. The rest of the paper is organized as follows. In the next section, there is a more detailed description of the problem. Section 3 includes related work in the field and in section 4 we analyze our approach to the problem, showing the formulas we have introduced. Next in section 5 we present the evaluation we performed and some comparative results which show the best candidate. Finally, in section 6 we discuss some future issues concerning the applicability of the proposed method. 2 Motivation The main idea of collaborative filtering is to make predictions of scores based on the heuristic that two people who agreed (or disagreed) in the past will probably agree (disagree) again. A typical collaborative filtering system runs as a centralized ser￾vice and the information it holds can be represented by a matrix of users and items. Each value of the matrix represents the score that a particular user has given to some item. The number of empty cells is known as sparsity and as we mentioned in the previous section, it is the main reason that recommender systems behave poorly, be- 104 G. Pitsilis et al