Recomendação de artigos científicos: um foco na integração de perfis de usuários.

Abstract

The personalized Recommender Systems have emerged as a possible solution to the information overload problem. However, their quality is related to the user profile and generate a profile with quality is not a trivial task. Consequently, the user that does not receive good recommendations may lose interest and confidence in the system. Our research presented here addresses this problem by proposing an approach to paper Recommendation Systems focusing on the integration of user profiles. The profiles were constructed fromthree sources: CV Lattes, Mendeley and LinkedIn. The integration of user profiles was performed by linear combination and we proposed three strategies: (i) equal importance (Igual); (ii) quantity of items (Quant); and (iii) user activity on the source (Ativ). To validate the profile models, we performed an experiment in which the participants evaluated the relevance of 50 papers, we used the metric NDCG@5. We performed two evaluations, the first only in Lattes, we used the strategy of building profile as a factor and evaluated the following strategies: terms (LT); concepts (LC) and Lopes strategy. The proposed strategies provided the best results, according to the Wilcox’s test (α = 0.05): Alternative Hypothesis (HA) = LT > Lopes (p-value = 0.01543) and HA = LC > Lopes (p-value = 0.04292). In the second evaluation, with the integrated profiles, we used two factors: profile representation (terms and concepts) and integration strategy (Igual; Quant; Ativ). The integrated profiles did not provide better results than non-integrated profiles, according to the Friedman’s test (α = 0.05): HA = There is difference (p-value = 0.9971). Based on the results, we can conclude that themodel provided satisfactory results in the Lattes platform, which can be characterized as an important contribution, given the importance of this platform for Brazilian researchers. Concerning the profiles integration, we did not achieved the expected results. In this sense, we verify that the integration model needs further investigation, whether conducting an experiment with more factors or with a larger sample of users.