Classificação de produtos utilizando técnicas few-shot learning

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. Retraining with these new classes is a necessity, as incorrect categorization of products in marketplaces can lead consumers to unpleasant experiences in the purchase process. 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. The proposal presented in this work is the use of product classifiers that use few-shot learning algorithms, which are capable of being trained with one or few samples per class. These have rapid training and need a small-scale database. The algorithms tested were: k-nearest neighbors (KNN), Matching Networks (MN) and DPGN (Distribution Propagation Graph Network). The proposed algorithms for product classification use characteristics previously extracted from the transfer learning process, except for encoder matching networks containing a Bi-LSTM network that received data in natural language extracted by embedding algorithms. The algorithms were tested with leave one out and k-fold cross validation. The selection of the best characteristics of the data base was also carried out, making it possible to reduce their dimension, facilitating training of neural networks with few-shot learning. Two databases were used for the tests, one containing 34 classes and 394 samples and the other containing 312 classes and 3120 samples. KNN was used as a baseline for the project and, despite its simplicity and no need for training, it presented satisfactory results. The matching and DPGN networks both presented results with 96.85% accuracy, managing to overcome the KNN using the database with 34 classes and for the database with 312 classes, the best result was obtained by matching with 93.78% accuracy. The proposed approach for classifying products belonging to new classes contributes to the correct categorization and maintenance of the accuracy required in marketplaces, without the need for constant retraining of the classifiers currently in operation. This can bring significant cost reduction of cloud server usage and better shopping experiences for customers.