Características de personalidade identificadas via textos aplicadas à recomendação por filtragem colaborativa.

Abstract

The amount of digital resources tends to increase with the growing use of information and communication technologies in several domains, such as e-commerce, e-learning, and tourism. In this context, there are Recommender Systems (RS) aiming, for example, to direct the most appropriate resources to users. Some researchers consider the user’s personality when designing RS strategies. Many of these researchers consider the premise that “people with similar personalities tend to prefer similar items”, and apply personality in RS via Collaborative Filtering (CF). However, we have not found research investigating whether, with collaboration focused on the opinion of users with a similar personality, the accuracy of any CF algorithms would, at a minimum, be maintained. Besides, most researchers do not apply APR (Automatic Personality Recognition) and do not consider that the different personality components can influence the recommendation process in different ways. Furthermore, these researchers commonly focus on specific scenarios (such as cold-start situations). Therefore, this thesis presents an investigation to improve RS accuracy in general scenarios by applying information related to human personality obtained without traditional personality questionnaires. With the experimental study conducted, (i) we reinforced results from initial studies that used questionnaires to identify personality; (ii) we realized that the accuracy of a CF algorithm could be affected when the algorithm disregards the opinion of less similar users in terms of personality (tending to improve the accuracy of memory-based algorithms); (iii) we analyzed that the components used to define personality influence the recommendation accuracy differently, although it is inappropriate to focus on specific components and ignore others; and (iv) we proposed and evaluated new strategies for RS applying the personality of users. The results obtained in this thesis highlight the relevance of the use of personality characteristics of CF-based RS users (without the need to complete specific questionnaires to identify such characteristics), especially in hybrid recommendation strategies, since, although it is valid to consider personality in a personalized recommendation process, this factor is not the only important aspect.