ProductNet: a Collection of High-Quality Datasets for Product Representation Learning

Chu Wang, Lei Tang, Yang Lu, Shujun Bian, Hirohisa Fujita, Da Zhang, Zuohua Zhang, Yongning Wu

TL;DR

This paper introduces ProductNet, a high-quality, curated dataset for product representation learning, inspired by ImageNet. It tackles the challenge of building such datasets by proposing an iterative approach where a multi-modal deep learning model (master model) learns product embeddings, which then actively accelerate the human annotation process for further dataset refinement.

What Problem Does This Paper Address?

The paper addresses the critical need for high-quality datasets to enable effective product representation learning in e-commerce. Traditional approaches often rely on massive, noisy datasets, which can hinder the development of robust models. The authors highlight two main challenges: (1) the difficulty of curating high-quality product datasets with a well-defined taxonomy, and (2) the challenge of learning product embeddings with satisfactory transferability across various downstream tasks. They aim to overcome these by focusing on data quality over quantity and by creating a symbiotic relationship between dataset construction and model learning.

Methods

The core of ProductNet's methodology is an iterative, symbiotic approach between dataset curation and representation learning.

  • Multi-Modal Master Model: A deep neural network, termed the "master model," is trained to learn product embeddings. This model leverages both product images and catalog information (title, description, brand, keywords) using a fusion mechanism. For image processing, they employ Inception-v4, and for text, they use a combination of DPCNN, A-LSTM, and fastText, depending on the text field.
  • Active Learning for Annotation: The learned product embeddings are then fed into an active learning framework. This framework intelligently suggests products for human annotators to label, significantly accelerating the annotation process and improving data quality. This is achieved through "local models" that can be KNN search engines, keyword-based search engines, or binary classifiers trained on positive and negative samples.
  • Function-Based Taxonomy: To reduce annotation ambiguity, they adopt a function-based taxonomy for non-media products.

This iterative loop allows the model to improve with better data, and the improved model to guide the data collection more effectively.

Datasets

The paper introduces ProductNet, a newly curated collection of high-quality product datasets.

  • ProductNet: This dataset is built with a focus on quality and a carefully chosen taxonomy. It currently comprises 3900 categories of non-media products, with approximately 40-60 products per category. The dataset contains a significant portion of products with both text and image information.
  • Existing Datasets (for inspiration/comparison): The authors mention ImageNet as a key inspiration for their approach to building high-quality image datasets. They also reference Google's Product Taxonomy as a source for product categorization.

Key Insights and Ideas

  • Quality over Quantity: The paper strongly advocates for prioritizing data quality and a well-designed taxonomy over sheer data volume for effective product representation learning. This is a departure from many large-scale dataset efforts.
  • Symbiotic Learning Loop: The core innovation is the iterative feedback loop between model learning and active learning-driven annotation. This creates a virtuous cycle where better data leads to better models, and better models lead to more efficient and accurate data collection.
  • Active Learning for Efficiency: Active learning, guided by learned embeddings, is shown to be a powerful tool for dramatically accelerating the human annotation process, making the creation of large, high-quality datasets more feasible.
  • Multi-Modal Fusion is Key: Combining image and text information in a multi-modal deep learning architecture is crucial for achieving high accuracy in product understanding and representation.

Key Results

  • High Classification Accuracy: The master model achieves a top-1 accuracy of 94.7% for classifying products into 1240 classes. This high accuracy is attributed to both the multi-modal model and the quality of the ProductNet dataset.
  • Annotation Acceleration: The active learning approach demonstrates a significant speedup in labeling. Compared to vanilla labeling, the acceleration factor is estimated to be around 20x for finding positive data points for a leaf node. After the master model is trained, this factor is estimated to be 80x.
  • Effective Product Embeddings: The learned product embeddings are shown to be useful for various transfer learning tasks and can be used as search indices, partition keys, and input features for machine learning models.

Limitations and Future Work

  • Limited Scope of Taxonomy: While the current taxonomy covers 3900 categories of non-media products, the authors acknowledge that a more comprehensive taxonomy might be needed for broader applications.
  • Focus on Non-Media Products: The current dataset is focused on non-media products. Future work could extend to other product domains.
  • Further Exploration of Active Learning Strategies: While active learning is shown to be effective, further research could explore more advanced sampling strategies within the active learning framework.
  • Generalization to Diverse Business Tasks: While the embeddings are stated to be useful for transfer learning, further empirical studies could demonstrate their effectiveness across a wider range of specific business tasks.

Summary updated

Abstract

ProductNet is a collection of high-quality product datasets for better product understanding. Motivated by ImageNet, ProductNet aims at supporting product representation learning by curating product datasets of high quality with properly chosen taxonomy. In this paper, the two goals of building high-quality product datasets and learning product representation support each other in an iterative fashion: the product embedding is obtained via a multi-modal deep neural network (master model) designed to leverage product image and catalog information; and in return, the embedding is utilized via active learning (local model) to vastly accelerate the annotation process. For the labeled data, the proposed master model yields high categorization accuracy (94.7% top-1 accuracy for 1240 classes), which can be used as search indices, partition keys, and input features for machine learning models. The product embedding, as well as the fined-tuned master model for a specific business task, can also be used for various transfer learning tasks.
BibTeX
@misc{wang2019productnet,
  author = {Chu Wang and Lei Tang and Yang Lu and Shujun Bian and Hirohisa Fujita and Da Zhang and Zuohua Zhang and Yongning Wu},
  title = {{ProductNet: a Collection of High-Quality Datasets for Product Representation Learning}},
  year = {2019},
  eprint = {1904.09037},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1904.09037},
}

Loading related content...

Figure