-
Aleson G. S. Chaves
Programa de Pós-Graduação em Engenharia de Sistemas e Automação, Universidade Federal de Lavras
-
Bruno H. G. Barbosa
Departamento de Automática, Universidade Federal de Lavras, CP 3037, 37200-900, Lavras/MG
-
Danton D. Ferreira
Departamento de Automática, Universidade Federal de Lavras, CP 3037, 37200-900, Lavras/MG
-
Paulo R. Silva
Omnilogic Inteligência S/A
-
Sinval T. Nascimento
Omnilogic Inteligência S/A
Keywords:
Natural Language Processing, E-commerce, Machine Learning, Few-Shot Learning, DPGN, Matching Networks, Artificial Intelligence
Abstract
E-commerce platforms (marketplaces) receive daily thousands of products belonging to new classes that have not participated in the training process of the algorithm responsible for automating the classification of products. However, it is difficult to constantly update the system with these products, because the cost of retraining the classifiers currently in operation is high due to the large size of the databases. In this sense, the use of product classifiers that use few- shot learning algorithms is an interesting option, as they are capable of being trained using only new classes containing one or few samples per class. Therefore, the k-nearest neighbor (KNN), Matching Networks (MN) and DPGN (Distribution Propagation Graph Network) algorithms were compared in the product classification problem using a database, with 312 classes and 3120 samples, from a marketplace. The tests were performed with k-fold cross-validation, of which the matching networks presented the best result with 93.78% accuracy.
