|
Nome do aluno
|
Diego Corrêa da Silva
|
|
Título do trabalho
|
Exploiting Calibration Settings toward Fairness in Recommender Systems
|
|
Resumo do trabalho
|
Collaborative Recommender Systems learn the users' preferences through their interaction history and use it to create personalized recommendation lists. Most recommender systems create the recommendation list based on the items most relevant/similar to the user's preferences. However, the emphasis on relevance can cause problems such as super specialization, popularity bias, or imbalance among the preferred classes. All these problems create a recommendation list with items that belong to only a few areas/classes of the user's preferences, i. e., generating a list that is not fair to the user's preferences. In this regard, calibrated recommendation systems have attracted attention as a means of ensuring a degree of fairness. This type of recommender system is designed to recommend items that are simultaneously relevant and fair. Fairness is measured by comparing two distributions, one extracted from the user's preferences and another from the recommended list. Usually, as the distributions are less divergent, the recommendation becomes fairer. In this sense, this study benchmarks 57 fairness measures to determine the best one. The differences in the measure's functionality lead us to introduce a new way of extracting the distributions and measuring the relevance as infer changes on the trade-off balances proposed by the state-of-the-art. In total, our implementation produces 5928 unique calibrated recommendation systems. Each system is implemented to use two datasets, Movielens and Yahoo Movies, running 35 times each system with each dataset. The results show that our proposals increase the precision by 28% and 23%, respectively. For both datasets, four fairness measures reached the best positions.
|
|
Orientador
|
Frederico Araújo Durão (PGCOMP/UFBA)
|
|
Membro externo 1
|
Gabriel de Souza Pereira Moreira (NVIDIA/ITA)
|
|
Link para o curriculum lattes
|
http://lattes.cnpq.br/
|
|
Membro interno 1
|
Tatiane Nogueira Rios (PGCOMP/UFBA)
|
|
Link para o curriculum lattes
|
http://lattes.cnpq.br/
|
|
Suplente do membro externo
|
Marcelo Garcia Manzato (ICMC/USP)
|
|
Link para o curriculum lattes
|
http://lattes.cnpq.br/
|
|
Suplente do membro interno
|
Danilo Barbosa Coimbra (PGCOMP/UFBA)
|
|
Link para o curriculum lattes
|
http://lattes.cnpq.br/
|
|
Data do exame
|
16 Nov, 2022
|
|
Horário do exame
|
9:00 AM
|