A Novel Idea of Classification of E-commerce Products Using Deep Convolutional Neural Network

Abstract

The high numbers of non-food products and categories on today E-commerce sites render validation of the data as labor intensive and expensive task. Therefore, there is a recent push to automate validation of correct placement of product in category. The French E-commerce company CDiscount has launched Kaggle competition, sharing huge dataset of over 7 million products, to solve the very problem. The goal is to classify products containing multiple images into one of 5270 categories. This paper proposes, implements and experimentally evaluates deep neural network architecture for classification of non-food E-commerce products. To tackle the complexity of the task on available hardware, hierarchical architecture of neural networks that exploits existing category taxonomy is proposed. The hierarchical architecture achieved the Top-1 accuracy of 0.61061. It has been found, that specific networks in hierarchical architecture can be successfully transferred onto similar datasets, by transferring network that learned on books onto different book dataset. The transferred model performed better than the same model pre-trained on ImageNet dataset.