MEP-3M: A large-scale multi-modal E-commerce product dataset

The product categories are vital for the E-commerce platforms due to the core applications on auto-matic product category assignment, personalized product recommendations, etc. In this paper, we con-struct a large-scale Multi-modal E-commerce Products classification dataset MEP-3M, which is large-scale, hierarchical-categorized, multi-modal, fine-grained, and long-tailed. Statistically, MEP-3M consists of over 3 million products, thus achieves the largest data scale in comparison to the existing E-commerce product datasets. The products in MEP-3M are represented in three modalities: image, textual descrip-tion, and OCR text, and labeled with tree-like labels. The third level labels are extremely fine-grained. In addition, we exploit four novel practical tasks on this dataset, Product classification, Hierarchical Product Classification, Fine-grained Product Classification, and Product Representation Learning. For each task, we present some image-only, text-only, and multi-modal baseline performances for further researches. The MEP-3M dataset will be released at https://github.com/ChenDelong1999/MEP-3M . ?? 2023 Elsevier Ltd. All rights reserved.