[Paper Review] The Klarna Product Page Dataset: A Realistic Benchmark for Web Representation Learning.

Alexandra Hotti, Riccardo Sven Risuleo+3 more|arXiv (Cornell University)|Nov 3, 2021

Web Data Mining and Analysis40 references4 citations

TL;DR

This paper introduces the Klarna Product Page Dataset, a large-scale, realistic collection of 51,701 manually labeled e-commerce product pages from 8,175 real websites, and evaluates Graph Neural Network (GNN)-based methods for representing DOM tree elements. It demonstrates that Graph Convolutional Networks produce superior embeddings for web element prediction tasks compared to state-of-the-art methods.

ABSTRACT

This paper tackles the under-explored problem of DOM tree element representation learning. We advance the field of machine learning-based web automation and hope to spur further research regarding this crucial area with two contributions. First, we adapt several popular Graph-based Neural Network models and apply them to embed elements in website DOM trees. Second, we present a large-scale and realistic dataset of webpages. By providing this open-access resource, we lower the entry barrier to this area of research. The dataset contains $51,701$ manually labeled product pages from $8,175$ real e-commerce websites. The pages can be rendered entirely in a web browser and are suitable for computer vision applications. This makes it substantially richer and more diverse than other datasets proposed for element representation learning, classification and prediction on the web. Finally, using our proposed dataset, we show that the embeddings produced by a Graph Convolutional Neural Network outperform representations produced by other state-of-the-art methods in a web element prediction task.

Motivation & Objective

• To address the lack of realistic, large-scale datasets for web element representation learning in real e-commerce environments.
• To evaluate the effectiveness of graph-based neural networks in learning meaningful representations of DOM tree elements for web automation tasks.
• To lower the barrier to entry for research in web representation learning by releasing a publicly accessible, browser-renderable dataset of real product pages.
• To establish a benchmark for evaluating and comparing methods in web element representation and prediction.
• To demonstrate the superiority of Graph Convolutional Networks over other state-of-the-art models in predicting web elements using learned embeddings.

Proposed method

• Adapting popular Graph Neural Network architectures to learn representations of elements in website DOM trees.
• Representing webpages as directed, hierarchical graphs where nodes correspond to DOM elements and edges represent parent-child or sibling relationships.
• Training GNN models on the Klarna Product Page Dataset to learn dense vector embeddings for each DOM element.
• Using the learned embeddings to perform downstream web element prediction tasks, such as identifying relevant elements for data extraction.
• Evaluating model performance using standard metrics on a held-out test set of real product pages.
• Leveraging the dataset’s realism and diversity by ensuring all pages are fully renderable in a web browser, enabling integration with computer vision and web automation pipelines.

Experimental results

Research questions

• RQ1 Can graph-based neural networks effectively learn meaningful representations of DOM tree elements in real-world e-commerce webpages?
• RQ2 How do GNN-based embeddings compare to other state-of-the-art methods in predicting relevant web elements for automation tasks?
• RQ3 To what extent does the realism and diversity of the Klarna Product Page Dataset improve the generalization and performance of representation learning models?
• RQ4 Can the proposed dataset serve as a robust benchmark for future research in web representation learning?
• RQ5 What is the impact of dataset scale and real-world rendering capability on model performance in web element prediction?

Key findings

• The Klarna Product Page Dataset contains 51,701 manually labeled product pages from 8,175 real e-commerce websites, offering high realism and diversity.
• All pages in the dataset are fully renderable in a web browser, enabling compatibility with computer vision and web automation applications.
• Graph Convolutional Networks (GCNs) outperform other state-of-the-art methods in a web element prediction task using the proposed dataset.
• The GNN-based approach achieves superior performance on the prediction task, demonstrating the effectiveness of graph-based representations for DOM elements.
• The release of this open-access dataset lowers the barrier to entry for researchers in web representation learning and enables reproducible benchmarking.
• The dataset's real-world nature and browser-renderability make it particularly suitable for training and evaluating models in realistic web environments.