Fuzzy models for group decision making and applications to e-learning and recommender systems

Abstract

The work presented in the Ph.D. thesis deals with the definition of new fuzzy models for Group Decision Making (GDM) aimed at improving two phases of the decision process: preferences expression and aggregation. In particular a new preferences model named Fuzzy Ranking has been defined to help decision makers express fuzzy statements on available alternatives in a simple and meaningful form, focusing on two alternatives at a time but, at the same time, without losing the global picture. This allows to reduce inconsistencies with respect to other existing models. Moreover a new preference aggregation model guided by social influence has been described. During a GDM process, in fact, decision makers interact and discuss each other exchanging opinions and information. Often, in these interactions, those with wider experience, knowledge and persuasive ability are capable of influencing the others fostering a change in their views. So, social influence plays a key role in the decision process but, differently from other aspects, very few attempts to formalize its contribution in preference aggregation and consensus reaching have been made till now. In order to validate the defined models, they have been instantiated in two application contexts: e- Learning and Recommender Systems. In the first context, they have been applied to the peer assessment problem in massive online courses. In such courses, the huge number of participants prevents their thorough evaluation by the teachers. A feasible approach to tackle this issue is peer assessment, in which students also play the role of assessor for assignments submitted by others. But students are unreliable graders so peer assessment often provides inaccurate results. By leveraging on defined GDM models, a new peer assessment model aimed at improving the estimations of student grades has been proposed. With respect to Recommender Systems, the group recommendation issue has been tackled. Instead of generating recommendations fitting individual users, Group Recommender Systems provide recommendations targeted to groups of users taking into account the preferences of any (or the majority of) group members together. The majority of existing approaches for group recommendations are based on the aggregation of either the preferences or the recommendations generated for individual group members. Customizing the defined GDM models, a new model for group recommendations has been proposed that also takes into account the personality of group members, their interpersonal trust and social influence. The defined models have been experimented with synthetic data to show how they operate and demonstrate their properties. Once instantiated in the defined application contexts, they have been also experimented with real data to measure their performance in comparison to other context-specific methods. The obtained results are encouraging and, in most cases, better than those achieved by competitor methods. [edited by author]