Collaborative Recommender Agents Based On Case - Based Reasoning and Trust

of the Thesis The Artificial Intelligence (AI) community has carried out a great deal of work on how AI can help people to find out what they want on the Internet. As a result, the idea of recommender systems has been widely accepted among users. The main task of a recommender system is to locate items, information sources and people related to the interest and preferences of a single person or a group of people. This involves the construction of user models and the ability to anticipate and predict user preferences. This thesis focusses on the study of AI techniques which improve the performance of recommender systems. Initially, a detailed analysis of the current state-of-the-art in this field has been carried out. This work has been organised as a taxonomy where existing recommender systems on the Internet are classified into 8 general dimensions. This taxonomy provides us with an indispensable knowledge base from which to design our proposal. Secondly, this thesis proposes a new CBR approach to recommendation. Casebased reasoning (CBR) is a paradigm for learning and reasoning through experience suitable for recommender systems due to its being based on human reasoning. We provide a forgetting mechanism for case-based profiles that controls the relevance and age of past experiences. Experimental results show that this proposal better adapts the profiles to users and solves the utility problem of CBR systems. Thirdly, this thesis proposes the “agentification” of recommender systems in order to take advantage of interesting agent properties such as proactivity, encapsulation or social ability. Recommender systems sharply improve the quality of results when information about other users is utilised when recommending a given user. Collaboration among agents is performed with the opinion-based filtering method and the collaborative filtering method through trust. Both are based on a social model of trust making agents less vulnerable to others while collaborating. Experimental results show that our collaborative recommender agents improve the performance of the system while preserving the privacy of the user’s personal data. Finally, this thesis also proposes an evaluation procedure for recommender systems that allows a scientific discussion of the results. This proposal simulates the users behaviour over time based on real user profiles. We hope this new evaluation methodology will contribute towards the progress in this area of research.