Cross-platform Product Matching Based on Entity Alignment of Knowledge Graph with RAEA model

Wenlong Liu, Jiahua Pan, Xingyu Zhang, Xinxin Gong, Yang Ye, Xujin Zhao, Xin Wang, Kent Wu, Hua Xiang, Houmin Yan, Qingpeng ZhangarXiv 2025 · DOI · arXiv

Identifying identical products across diverse e-commerce platforms is a critical yet challenging task for market analysis and price comparison. This paper addresses this challenge by framing product matching as an Entity Alignment (EA) problem within Knowledge Graphs (KGs). The authors introduce RAEA (Relation-aware and Attribute-aware Graph Attention Networks), a novel framework designed to fuse structured graph data with unstructured textual descriptions effectively.

The Challenge of Cross-Platform Matching

In the global e-commerce landscape, products are often listed on platforms like Amazon and eBay using disparate languages, category systems, and description styles. Current Entity Alignment research faces three major hurdles:

  1. Interaction Neglect: Failure to capture the interplay between an entity's attributes (e.g., color) and its relations (e.g., co-purchased items).
  2. Data Heterogeneity: Significant variation in how product information is presented across platforms.
  3. Information Silos: Difficulty in integrating structured graph data with unstructured semantic text.

Proposed Methodology

To address these issues, the authors propose a comprehensive two-stage pipeline designed for scalability and precision.

The proposed two-stage pipeline for cross-platform product matching.

1. Rough Filtering

Given the massive scale of e-commerce data, the first stage employs rule-based matching. By applying regular expressions to product titles and categories, the system narrows down millions of potential matches to a manageable set of candidates.

2. Fine Filtering: The RAEA Model

The core innovation lies in the RAEA model, which utilizes a multi-channel Graph Attention Network (GAT) to perform deep semantic alignment.

Detailed architecture of the Relation-aware and Attribute-aware Graph Attention Network (RAEA).

Key components of the RAEA architecture include:

  • Graph Partitioning: The KG is split into four distinct channels: Literal (text), Digital (numbers), Name, and Structure.
  • MPnet Attribute Encoder: The model uses MPnet (pre-trained with SimCSE) to generate high-quality semantic embeddings for product descriptions, outperforming standard BERT in capturing fine-grained similarities in noisy text.
  • Relation-aware GAT: This component iteratively updates entity and relation representations. It maps entities to relations, then relations back to entities, and finally performs "entity enhancement" to capture two-hop neighborhood information.
  • Pre-weighted Ensemble: Instead of simple averaging, the model weights each channel's importance based on its individual performance, ensuring the most informative signals drive the final decision.

Experimental Evaluation

The model was evaluated on standard benchmarks (DBP15K, DWY100K) and a real-world eBay-Amazon dataset containing nearly one million items.

Key Results

  • Benchmark Superiority: On the DBP15K dataset, RAEA achieved a 6.59% average improvement in Hits@1 over 12 state-of-the-art baselines, including AttrGNN and MultiKE.
  • Real-World Performance: On the eBay-Amazon task, the model demonstrated high effectiveness, achieving a Recall of 0.862, indicating its ability to surface correct matches within the top results.
  • Ablation Insights: Studies revealed that the "Name" and "Relation" channels are the most critical. Removing the Name channel resulted in the largest performance drop, confirming that entity labels remain primary anchors for alignment.

Conclusion and Future Work

RAEA demonstrates that treating product matching as a multi-channel KG alignment task significantly improves accuracy by capturing the complex interplay between attributes and relations. While highly effective, the model currently relies on supervised "seed" entities. Future research will explore semi-supervised or unsupervised versions of RAEA to reduce manual labeling requirements and further optimize the rough filtering stage for even larger datasets.

Summary updated

Abstract

Product matching aims to identify identical or similar products sold on different platforms. By building knowledge graphs (KGs), the product matching problem can be converted to the Entity Alignment (EA) task, which aims to discover the equivalent entities from diverse KGs. The existing EA methods inadequately utilize both attribute triples and relation triples simultaneously, especially the interactions between them. This paper introduces a two-stage pipeline consisting of rough filter and fine filter to match products from eBay and Amazon. For fine filtering, a new framework for Entity Alignment, Relation-aware and Attribute-aware Graph Attention Networks for Entity Alignment (RAEA), is employed. RAEA focuses on the interactions between attribute triples and relation triples, where the entity representation aggregates the alignment signals from attributes and relations with Attribute-aware Entity Encoder and Relation-aware Graph Attention Networks. The experimental results indicate that the RAEA model achieves significant improvements over 12 baselines on EA task in the cross-lingual dataset DBP15K (6.59% on average Hits@1) and delivers competitive results in the monolingual dataset DWY100K. The source code for experiments on DBP15K and DWY100K is available at github (https://github.com/Mockingjay-liu/RAEA-model-for-Entity-Alignment).
BibTeX
@misc{liu2025cross,
  author = {Wenlong Liu and Jiahua Pan and Xingyu Zhang and Xinxin Gong and Yang Ye and Xujin Zhao and Xin Wang and Kent Wu and Hua Xiang and Houmin Yan and Qingpeng Zhang},
  title = {{Cross-platform Product Matching Based on Entity Alignment of Knowledge Graph with RAEA model}},
  year = {2025},
  eprint = {2512.07232},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.07232},
}
Figure