Towards a Collaborative Filtering Method in an Open World

MIQUEL MONTANER; BEATRIZ LÓPEZ; SILVANA ACIAR; ESTEVE DEL ACEBO

Barcelona, España

Workshop; Workshop on Security, AgentCities Third Information Day (iD3); 2003

AgentCities, Red de Agentes

One of the most emergent popular services developed for a city is restaurant  recommenders.Several approaches are followed, from individual recommender agents to distributed systems in which an agent for each restaurant is implemented.Our approach is also a distributed approach but regarding the collaborative  recommendations in an open environment: users profiles are not kept in a server but distributed in each individual user computer. A hybrid procedure of collaborative and content-based filtering is defined, in which a trust mechanism has been developed. When an agent  has a lack of information situation in order to recommend some item to the user, it looks for the opinion on the item to other trustworthy agents.Trust, then, it is not due to reliability issues, but due to preference similarity among users. Similarity is not based on profile exchange but on an opinion that comes from a question-answering process between agents, keeping user data on private. In order to achieve our objective in a open environment as Agentcities, we have deployed a Multi Agent system upon the JADE platform. Then, our restaurant recommender service consists of service agents and personal agents. Service agents offer information about restaurants and the personal agents (named restaurant server agent  and personal agent facilitator, correspondingly). Personal agents are in charge of recommending restaurants to their users based on both, information on restaurants and interaction with  ther friendly personal agents. Personal agents interact with the restaurant server agent in order to know about the restaurants, the personal agent facilitator agent in order to know about other personal agents in the system, and the personal agents in order to find similar users and take  advantage of their opinions and advice. All agents are currently available through  Agentcities at http://arlab.udg.es/, and the system developed,  GenialChef is registered on the Agentcities competition.