Entity Alignment in Product Knowledge Graphs: Techniques and Applications

Background : In e‑commerce procurement and operations, aligning product information across multiple platforms is essential for real‑time price monitoring. Different platforms have heterogeneous category systems and attribute schemas, requiring unified entity alignment at billion‑scale product volumes.

Technical Progress : Knowledge‑graph‑based entity alignment is introduced, where two graphs KG1 and KG2 contain entities with partial correspondences. Standard datasets such as DBP15K and DWY100K are referenced. Early methods relied on TransE embeddings; recent work adopts graph neural networks (GNNs) for richer representation, combining structural and textual features.

Entity Alignment Algorithms : Two main trends are highlighted—vector alignment using L1 distance with embedding loss, and more sophisticated models that incorporate neighbor node information via attention, multi‑head mechanisms, and kernel pooling. The pipeline includes recall, coarse‑ranking, fine‑ranking, and post‑processing steps, with coarse models using BERT‑based text similarity and fine models leveraging GNN embeddings.

Online Prediction Workflow : The system treats product matching as a search problem, using the smaller external catalog as query and the larger JD catalog as candidate set. Recall rules filter candidates, coarse ranking reduces the set to ~100 items, and fine ranking applies deep GNN‑based alignment models. Evaluation uses product detection rate and detection accuracy rather than traditional Top‑K hit rates.

Data Construction and Evaluation : Positive pairs are obtained via unsupervised similarity followed by crowdsourced verification; negative pairs are generated from SPU/SKU variations to create hard negatives (ratio 1:10). Metrics include detection rate, detection accuracy, and online A/B testing.

Product Attribute Extraction : A two‑step NER approach extracts potential attribute values from titles and then classifies their categories, improving graph completeness and alignment performance, especially for sparse or noisy third‑party listings.

Scalability and Large‑Scale Graph Handling : Inductive GNN models are employed to train on sampled neighborhoods, enabling efficient updates for newly added or removed products without retraining the entire graph.

Future Work : Plans include enriching graph relations with general knowledge, incorporating multimodal signals (e.g., images), better integration of attribute extraction features, improved handling of low‑resource categories, and further optimization of inference efficiency.

Q&A Highlights : Answers address schema alignment (unsupervised with optional manual review), sample balancing, attribute embedding initialization via BERT, impact of structural information (150%+ detection rate boost), and handling of long‑tail entities through attribute extraction and multimodal cues.