The assumption we have made in our modeling that the value of trustworthiness should correspond to a similarity value of zero seems to be not an optimum choice, therefore more investigation is needed towards finding the optimum fitting. As men- tioned in paragraph 4. 1 and 4. 2 about the complexity of the formulas used, there is a ptential problem since these computations need to be repeated for every user in the system. So, whenever a new user joins the system the trust and similarity computa tions will have to be redone against all existing users. As the system grows the com- putation time will increase significantly raising a scalability issue. A possible sol tion to this problem is to restrict the correlation process to a subset of participants rather the whole world. A focus of a future research is to investigate if clustering [22], a technique that is used for tackling a similar problem in recommender sys tems, can be quite effective here We intend to apply our technique to a real recommender system, with the expec tation that it will improve the quality of the derived recommendations. Another idea is to make use of the web-of-trust that could evolve from the establishment of direct trust relationships between users. Our aim is to improve recommendations by ex- ploiting the experiences of any entities not neighboring the querying one but which can be reached via the web-of-trust. The question that arises from this is how accu- rate these predictions can be Short tests we performed, showed a significant in- crease in the coverage, which translates into reduced sparsity, without significant impact on the error in predictions. Our short-term plans include a thorough study and analysis of the various parameters that may have some impact on the results as well as a performance analysis of the resulting system. The long-term plans include the deployment of a totally distributed recommender system Recommender trust can be derived in a similar way to that described for function- al trust in this paper. The basic idea is that someones(lets call it the trustee)re- ommender trust can be estimated by some other entity(lets call it the trustor)by comparing any recommendations that trustee has provided in the past about state- ments for which the trustor also maintains its own evidence. Then the trustor, by comparing its relevant personal experiences with the trustees recommendations, will be able to estimate how good in doing recommendations the trustee has been Similarly to direct trust, recommender trust is a subjective measure, which means every trustor has to maintain its own picture of the community No matter the quality of the recommendations such architecture can provide there are weaknesses concerning security for the recommender systems that must al so be covered. In particular, any deployed solution must be resistant to attacks from users that try maliciously to influence the system. 7 Conclusion We presented an empirical technique for modeling the trustworthiness of entities ng evidence that describe their rating behavior. The novelty comes from way that trustworthiness is derived from Similarity using a non-linear mapping.The assumption we have made in our modeling that the value of trustworthiness should correspond to a similarity value of zero seems to be not an optimum choice, therefore more investigation is needed towards finding the optimum fitting. As mentioned in paragraph 4.1 and 4.2 about the complexity of the formulas used, there is a potential problem since these computations need to be repeated for every user in the system. So, whenever a new user joins the system the trust and similarity computations will have to be redone against all existing users. As the system grows the computation time will increase significantly raising a scalability issue. A possible solution to this problem is to restrict the correlation process to a subset of participants rather the whole world. A focus of a future research is to investigate if clustering [22], a technique that is used for tackling a similar problem in recommender systems, can be quite effective here. We intend to apply our technique to a real recommender system, with the expectation that it will improve the quality of the derived recommendations. Another idea is to make use of the web-of-trust that could evolve from the establishment of direct trust relationships between users. Our aim is to improve recommendations by exploiting the experiences of any entities not neighboring the querying one but which can be reached via the web-of-trust. The question that arises from this is how accurate these predictions can be. Short tests we performed, showed a significant increase in the coverage, which translates into reduced sparsity, without significant impact on the error in predictions. Our short-term plans include a thorough study and analysis of the various parameters that may have some impact on the results as well as a performance analysis of the resulting system. The long-term plans include the deployment of a totally distributed recommender system. Recommender trust can be derived in a similar way to that described for functional trust in this paper. The basic idea is that someone’s (lets call it the trustee) recommender trust can be estimated by some other entity (lets call it the trustor) by comparing any recommendations that trustee has provided in the past about statements for which the trustor also maintains its own evidence. Then the trustor, by comparing its relevant personal experiences with the trustee’s recommendations, will be able to estimate how good in doing recommendations the trustee has been. Similarly to direct trust, recommender trust is a subjective measure, which means, every trustor has to maintain its own picture of the community. No matter the quality of the recommendations such architecture can provide, there are weaknesses concerning security for the recommender systems that must also be covered. In particular, any deployed solution must be resistant to attacks from users that try maliciously to influence the system. 7 Conclusion We presented an empirical technique for modeling the trustworthiness of entities using evidence that describe their rating behavior. The novelty comes from way that trustworthiness is derived from Similarity using a non-linear mapping. 116 G. Pitsilis et al