-
Fernanda C. e Silva
Programa de Pós-graduação em Engenharia de Sistemas e Automação, Universidade Federal de Lavras
-
Danton D. Ferreira
Departamento de Automática, Universidade Universidade Federal de Lavras, CP 3037, 37200-900, Lavras/MG
-
Bruno H. G. Barbosa
Departamento de Automática, Universidade Universidade Federal de Lavras, CP 3037, 37200-900, Lavras/MG
-
Paulo R. Silva
Omnilogic Inteligência S/A, Belo Horizonte/MG
-
Sinval T. Nascimento
Omnilogic Inteligência S/A, Belo Horizonte/MG
Keywords:
One shot learning, Siamese neural networks, Natural language processing, Classification, Machine learning, E-commerce
Abstract
Applying artificial intelligence to e-commerce is a trend, however, algorithms that deal with small databases are lacking. This work proposes a model with Siamese neural network in a few shot learning context to classify new samples from an online store. In order to choose the best representative sample to be used in the Siamese network, the random choice and centroid calculation approaches were tested. In addition, an ensemble structure of Siamese network considering 3 and 5 representants was also proposed. Finally, the k-Means algorithm was used to calculate the centroids of the clusters of each class and use them as representative samples in the ensemble. The best result was with the centroid, improving performance from 93% to over 98% accuracy. Different structures were also proposed for the internal network, the best being with three hidden layers and applying the DropOut technique. In addition, it is possible to develop other alternatives to choose representatives and improve feature extractors, improving data quality, being useful to reduce the size of the analyzed data to reduce the complexity of the network in order to reduce costs with the use of servers and data storage in the cloud.
