Relation- & Attribute-Aware Graph Attention Networks

• The paper introduces a unified framework that utilizes hybrid attention mechanisms and multi-channel encoding to improve cross-platform product matching.
• It employs specialized attribute-aware encoding by partitioning knowledge graphs into four channels, leveraging MPNet and SimCSE for robust entity representations.
• Empirical results show significant Hits@1 improvements on benchmarks such as DBP15K and DWY100K, demonstrating superior effectiveness in real-world eBay-to-Amazon matching.

Relation-aware and Attribute-aware Graph Attention Networks for Entity Alignment (RAEA) constitute a unified neural architecture for the product matching problem, treated as an instance of Entity Alignment (EA) over large-scale knowledge graphs (KGs). RAEA was introduced for fine-grained matching across cross-platform product catalogs, particularly between eBay and Amazon, leveraging both attribute triples and relation triples with explicit modeling of their interactions via a hybrid of attention mechanisms and multi-channel encoding. The RAEA pipeline demonstrates significant improvements in alignment tasks on both benchmark datasets and real-world product matching scenarios (Liu et al., 8 Dec 2025).

1. Two-Stage Pipeline for Cross-Platform Matching

The RAEA approach is positioned within a two-stage matching pipeline. In the initial “rough filter” phase, a set of manually crafted, rule-based regular expressions operates over eBay product categories and title keywords to retrieve a candidate subset of Amazon products for each eBay entry. For example, products in the eBay category "rock climbing → anti-skating claw" are matched to Amazon entries with titles containing "climbing.*crampons".

The “fine filter” applies the RAEA model. For each candidate subgraph pair (one from eBay-KG, one from Amazon-KG), RAEA computes dense entity embeddings and constructs a similarity matrix. Entities are then aligned by ranking candidates according to embedding similarity. This two-stage design enables efficient large-scale matching while maximizing precision in the final ranking step.

2. Attribute-aware Entity Encoding

A distinctive feature of RAEA is the partitioning of the KG into four disjoint channels—Literal, Digital, Name, and Structure-only—where each channel contains only the corresponding attribute type triples. For each channel, entities are represented by encoding their attribute predicate-value pairs with a pre-trained MPNet model (768 dimensions), further refined via SimCSE for improved sentence similarity.

Let $a_j\in\mathbb{R}^{d_a}$ and $v_j\in\mathbb{R}^{d_v}$ denote the embeddings for predicate $p_j$ and value $v_j$. Each entity $e$ has an initial state $\boldsymbol{h}_e^{(0)}\in\mathbb{R}^{d_e}$. Attribute aggregation proceeds via a single-headed attention mechanism, where the entity state and each predicate embedding are concatenated and scored:

$$o_j = \mathrm{LeakyReLU}\left(u^\top\,[\boldsymbol{h}_e^{(i-1)}\,\|\, a_j ]\right),\quad \alpha_j = \frac{\exp(o_j)}{\sum_{k=1}^n \exp(o_k)}$$

The updated representation is

$$\boldsymbol{h}_e^{(i)} = \sigma\left(\sum_{j=1}^n \alpha_j W_i [a_j \| v_j]\right)$$

with learnable $u$, $W_i$, and $\sigma$ as ELU. Stacking one or two layers yields an attribute-aware embedding $\boldsymbol{x}_e^{attr} = \boldsymbol{h}_e^{(L)}$. This architecture enables rich modeling of heterogeneous attribute signals across different attribute types.

3. Relation-aware Graph Attention Networks

After entities are assigned attribute-aware embeddings, RAEA introduces a three-stage mechanism for relation-aware enhancement.

3.1 Entity-to-Relation Representation

For each relation type $r_k$, all triples $(e_i, r_k, e_j)$ induce “head view” and “tail view” representations: $$\alpha_{ijk} = \frac{\exp\!\bigl(\mathrm{LeakyReLU}(a_r^\top[\boldsymbol{x}_iW^h\;\|\;\boldsymbol{x}_jW^t])\bigr)}{\sum_{(i',j')\in\mathcal{T}_{r_k}}\exp\!\bigl(\mathrm{LeakyReLU}(a_r^\top[\boldsymbol{x}_{i'}W^h\;\|\;\boldsymbol{x}_{j'}W^t])\bigr)}$$

$$\boldsymbol{r}_k^h = \mathrm{ReLU}\left(\sum_{(i,j)\in\mathcal{T}_{r_k}} \alpha_{ijk} \boldsymbol{x}_iW^h\right),\quad \boldsymbol{r}_k^t = \mathrm{ReLU}\left(\sum_{(i,j)\in\mathcal{T}_{r_k}} \alpha_{ijk} \boldsymbol{x}_jW^t\right)$$

$$\boldsymbol{r}_k = \boldsymbol{r}_k^h + \boldsymbol{r}_k^t$$

Here, $\mathcal{T}_{r_k}$ is the set of triples with relation $r_k$, and $a_r, W^h, W^t$ are learnable.

3.2 Relation-to-Entity Representation

Each entity $e_i$ aggregates incoming relation representations via attention: $$\beta_{ik} = \frac{\exp\!\bigl(\mathrm{LeakyReLU}(a_e^\top[\boldsymbol{x}_i^{attr}\;\|\;\boldsymbol{r}_k])\bigr)}{\sum_{r_{k'}\in\mathcal{R}_{e_i}} \exp\!\bigl(\mathrm{LeakyReLU}(a_e^\top[\boldsymbol{x}_i^{attr}\;\|\;\boldsymbol{r}_{k'}])\bigr)}$$

$$\boldsymbol{x}_i^r = \mathrm{ReLU}\left(\sum_{r_k\in\mathcal{R}_{e_i}} \beta_{ik}\, \boldsymbol{r}_k\right),\quad \boldsymbol{x}_i^{rel} = [\boldsymbol{x}_i^{attr}\;\|\;\boldsymbol{x}_i^r]$$

where $\mathcal{R}_{e_i}$ denotes all relation types incident on $e_i$.

3.3 Entity Representation Enhancement

A final graph attention (GAT) layer propagates signals over the graph: $$\gamma_{ij} = \frac{\exp(\mathrm{LeakyReLU}(a_g^\top[\boldsymbol{x}_i^{rel}\|\boldsymbol{x}_j^{rel}]))}{\sum_{j'\in\mathcal{N}(i)} \exp(\mathrm{LeakyReLU}(a_g^\top[\boldsymbol{x}_i^{rel}\|\boldsymbol{x}_{j'}^{rel}]))}$$

$$\boldsymbol{x}_i^{out} = [\boldsymbol{x}_i^{rel}\ \|\ \mathrm{ReLU}(\sum_{j\in\mathcal{N}(i)}\gamma_{ij} \boldsymbol{x}_j^{rel})]$$

These outputs serve as the final per-channel entity embeddings.

4. Channel-wise Ensemble for Similarity Computation

Similarity between eBay and Amazon entities is computed channel-wise as the cosine similarity between the resulting embeddings. To aggregate these into a single cross-KG similarity matrix, RAEA applies a data-driven pre-weighting strategy: channel $k$ is weighted by its Hits@1 performance, $$\mathrm{weight}_k = \mathrm{Hits@1}_k / \sum_{k'} \mathrm{Hits@1}_{k'}$$, yielding

$S = \sum_k \mathrm{weight}_k\, S_k$

where $S_k$ is the similarity matrix for channel $k$. This approach leverages the empirical discriminative power of each attribute type in the ensemble.

5. Margin-based Hard Negative Training Objective

For each channel, RAEA employs a margin-based ranking loss using hard negatives. Let $\Psi^s$ be the set of supervised aligned seed pairs $(e, e')$. For each, up to 15 hardest negatives $e_-, e'_-$ are sampled: $\mathcal{L}_k = \sum_{(e,e')\in\Psi^s}\left[ \sum_{e_- \in NS(e)} [d(e^k,e'^k)-d(e_-^k,e'^k)+\gamma]_+ + \sum_{e'_- \in NS(e')} [d(e^k,e'^k)-d(e^k,e'_-^k)+\gamma]_+ \right]$ with Euclidean distance $d(\cdot, \cdot)$ and typical margin $\gamma=1.0$. The overall loss sums across all channels.

6. Implementation Configurations

Key hyperparameters include MPNet embedding dimensions ($d_a=d_v=768$), intermediate GAT dimensions ($d_e=d_r=256$), optimizer (AdaGrad), learning rate grid $\{1e^{-3}, 5e^{-3}, 1e^{-2}, 1.5e^{-2}, 2e^{-2}\}$, and $\ell_2$-regularization grid $\{0, 1e^{-4}, 1e^{-3}, 1e^{-2}\}$. Negative sampling is fixed at 15 per seed for DBP15K, 5 for DWY100K. Pre-training of MPNet utilizes “paraphrase-multilingual-MPnet-base-v2” with batch size 512 and SimCSE objective. Training uses early stopping (no Hits@1 improvement in 50 epochs, max 1,500) on hardware comprising an NVIDIA RTX 3090 (24 GB) and 12-core CPU.

7. Empirical Performance and Applications

On the cross-lingual DBP15K EA benchmark, RAEA with pre-weighted ensemble achieves:

• zh–en: Hits@1 = 86.28% (best baseline 79.60%)
• ja–en: Hits@1 = 88.48% (best baseline 78.50%)
• fr–en: Hits@1 = 94.97% (best baseline 91.85%) with an average improvement of approximately 6.59% in Hits@1 over 12 baselines.

On the monolingual DWY100K dataset, RAEA attains Hits@1 of 97.35% (DBP–WD, best=98.10%) and 99.82% (DBP–YG, best=99.89%).

In practical eBay-to-Amazon matching, RAEA’s fine filter yields NDCG=0.566 and MRR=0.345, outperforming simpler ensemble baselines. This suggests that the joint modeling of attribute and relation signals, combined with channel-weighted similarity and hard negative loss, is particularly effective for real-world cross-platform product alignment. The RAEA source code is available for public use.

---

RAEA introduces a unified EA framework that synthesizes attribute-aware encoding (via MPNet and attention), explicit aggregation of relation types through RGAT, multi-channel ensemble scoring grounded in empirical discriminative strength, and discriminative margin-based training. These architectural features result in superior alignment accuracy on diverse datasets and improved product matching performance in operational settings (Liu et al., 8 Dec 2025).