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:
- Interaction Neglect: Failure to capture the interplay between an entity's attributes (e.g., color) and its relations (e.g., co-purchased items).
- Data Heterogeneity: Significant variation in how product information is presented across platforms.
- 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.
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.
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