A Cross-Lingual Real-Time E-Commerce Recommendation Method Based on Siamese Graph Convolutional Network and Bilinear Attention

Abstract

To balance low latency and high accuracy in cross-lingual real-time recommendation, we propose a two-stage method combining offline high-precision entity representation with online low-latency feature interaction. First, the cross-language scenario is abstracted as a heterogeneous graph. Then, a Siamese Graph Convolutional Network is utilized for entity representation learning. Finally, an efficient bilinear attention mechanism is employed for deep feature interaction to output predictions. After conducting experiments on the cross-border e-commerce dataset, it was found that the model performed well in entity representation learning. When the recommendation list length was 30, the normalized discounted cumulative gain value of the Siamese graph convolutional network was stable at more than 7.8%, which was more than 20% higher than other models. Regarding feature interaction, the bilinear attention mechanism showed superior convergence. Its mean average value reached 12.7% in the 100th round, 1.9 percentage points higher than the bilinear mechanism. In the scenario of increasing the sales rate of “long-tail products,” the hit rate of the recommendation method proposed by the study reached 46.5%. In summary, the proposed method demonstrates excellent accuracy and efficiency, proving its potential for real-time cross-language applications.

Keywords

cross-border e-commerce cross-domain recommendation cross-language processing graph convolutional networks low latency recommendation system

Overview

Cross-border e-commerce (CBEC) has become a key driver of global trade.^1–3 As an important extension, the conversational business model is accelerating the reconstruction of the interaction logic between users and various links in the supply chain.^4,5 However, compared with the mature domestic e-commerce environment, the CBEC SC shows complexity and uncertainty.^6,7 It involves international logistics, tariffs, overseas warehousing, and complex settlement systems.^8,9 Connecting diverse global users creates cross-lingual challenges. Semantic gaps and cultural differences often lead to mismatches in machine translation and user intent. Furthermore, data sparsity and the “cold start” problem are exacerbated in new markets, severely limiting recommendation accuracy and conversion rates.¹⁰ Current Region Server（RS）RSs, like collaborative filtering algorithms, primarily employ the direct interaction matrix between users and objects for prediction in an attempt to address the issue of data sparsity.^11,12 This type of method performs well in data-dense scenarios, but in an environment such as CBEC, where interaction records are extremely sparse, it is difficult to generate meaningful recommendation results.¹³ This failure stems from the inability to effectively utilize upstream supply information to compensate for the lack of downstream user data. Therefore, an advanced algorithm integrating complex business networks is required.

Recommendation models based on Graph Convolutional Networks (GCNs) offer powerful solutions for complex interaction data. To address privacy leakage, Liu et al. suggested a privacy-preserving model based on GCN. By creating a subset reflecting user preferences, the method proved effective on two public datasets.¹⁴ Addressing sequential recommendation, Huang et al. proposed a location-enhanced GCN model. This model simulated user-item interaction through a bipartite graph to achieve high-order connectivity.¹⁵ To solve error propagation, Ma et al. proposed a method combining knowledge graphs and GCNs. This approach utilized path-level self-attention to enhance interpretation.¹⁶ Yan et al. proposed a cascaded residual GCN model to mine connections between behaviors. Experiments showed that the average relative gain of the Hit Rate (HR) reached 24.76%, 27.28%, and 25.10% on three datasets.¹⁷

Recent research on e-commerce recommendations has focused on the retail sector. Li et al. suggested an approach based on image processing. Experimental findings indicated increased user happiness and accuracy.¹⁸ To address consumer behavior in CBEC, Chen et al. established a framework for systematic literature evaluation. The results provided a solid foundation for future research.¹⁹ Furthermore, Jian et al. suggested a model based on multilevel determinants. Analysis of 278 consumers revealed that logistics services positively impacted purchase intention, improving the conversion rate.²⁰ To provide scientific guidance, Lu et al. suggested a paradigm merging information systems with technological acceptance models. Results revealed that this model provided effective suggestions for optimizing strategies.²¹

In summary, although current research has progressed, significant challenges remain in large-scale real-time systems. The inference phase typically involves multi-hop graph traversal, leading to high latency. Therefore, the study proposes a method based on Siamese GCN and Bilinear Attention (BLA). The innovation lies in an end-to-end framework integrating representation learning and interaction prediction. By introducing a Siamese network, the matching relationship is directly optimized. The BLA mechanism is designed to conduct deep feature interaction. This research aims to provide a solution balancing high accuracy and low latency.

Methods

To address data sparsity and complex relationships, a two-stage method is proposed to balance accuracy and efficiency. First, the Siamese GCN architecture is used to learn high-quality entity representations from cross-domain heterogeneous graphs. Second, a computationally efficient prediction layer incorporating BLA generates accurate recommendation scores, improving real-time response.

Design of SC cross-domain entity representation method based on Siamese GCN

CBEC scenarios involve unique entities like overseas warehouses and customs.^22,23 In this structure, cross-border sellers act as “bridges” connecting markets. However, information barriers between upstream supply and downstream retail make it difficult to utilize rich supply-side data for sparse user scenarios.²⁴ Studying an effective representation method is a prerequisite for improving recommendation performance. Existing recommendation technologies, such as traditional collaborative filtering algorithms, mainly rely on direct interaction between users and products, and their effects are significantly attenuated when faced with sparse data.²⁵ To address this, the study introduces GCN as a basic technology. By aggregating neighborhood information, GCN learns vector representations that integrate node attributes and structural information. To apply GCN to complex cross-lingual scenarios, the business logic is formally modeled, as shown in Figure 1.

FIG. 1.

Cross-lingual and cross-domain recommendation scenario modeling.

In Figure 1, the source domain contains a dense interactive network between diverse users and suppliers. The target domain represents CBEC platforms with sparse data. Cross-border sellers act as a bridge to transfer knowledge from the source to the target domain. To model this, the business scenario is abstracted into a heterogeneous graph, as shown in Figure 2.

FIG. 2.

Schematic diagram of cross-border e-commerce supply chain heterogeneity construction.

As shown in Figure 2, the heterogeneous graph formally expresses key business processes by capturing multilevel relationships (e.g., component sourcing to final sales) often missed by traditional models. This provides a structural foundation for GCN representation learning. GCN aggregates neighborhood information to translate the graph’s topology and node properties into low-dimensional vectors, as illustrated in Equation (1).²⁶

H^{(l + 1)} = σ ({\tilde{D}}^{- \frac{1}{2}} \tilde{A} {\tilde{D}}^{- \frac{1}{2}} H^{(l)} W^{(l)})

(1)

In Equation (1), $H^{(l + 1)}$ and $H^{(l)}$ represent the characteristic matrices of nodes in the $l$ layer network. $\tilde{A}$ represents the self-joining adjacency matrix. $\tilde{D}$ represents the degree matrix of $\tilde{A}$ . $σ$ stands for nonlinear function. $W^{(l)}$ stands for weight matrix (WM). While this process aggregates neighborhood features, standard GCNs optimize for node representation rather than the dynamic association strength between entity pairs. To address this, we introduce a Siamese GCN architecture to explicitly model these “relational matching relationships.” Its structure is shown in Figure 3.

FIG. 3.

Structure of Siamese GCN. GCN, Graph Convolutional Network.

In Figure 3, the Siamese GCN comprises two weight-sharing subnetworks. During training, user and product neighborhood subgraphs are input into respective branches. Through multilayer graph convolution, node representations are propagated and updated. The final entity vectors are generated by aggregating the output of the last GCN layer. The matching degree is then calculated via vector inner product, as shown in Equation (2).

{\hat{y}}_{u i} = e_{u}^{T} e_{i}

(2)

In Equation (2), $u$ displays the user, $i$ represents the product, and ${\hat{y}}_{u i}$ represents the matching score. Final entity representation vectors $e_{u}^{}$ and $e_{i}^{}$ are output by the Siamese GCN model, where a larger ${\hat{y}}_{u i}$ indicates a higher predicted user preference. To optimize ranking beyond simple matching scores, a Bayesian personalized ranking loss function $L_{BPR}$ is employed, as shown in Equation (3).

L_{BPR} = - \sum_{(u, i, j) \in D_{S}} \ln σ ({\hat{y}}_{u j} - {\hat{y}}_{u k})

(3)

In Equation (3), $D s$ represents the sampling data set. $j$ represents the positive sample products that the user has interacted with. $k$ represents negative sample products that the user has not interacted with. $(u, i, j)$ represents the training triplet. To prevent overfitting and improve generalization, this study applies an L2 regularization penalty to the model parameters. The final optimization objective function, shown in Equation (4), combines the BPR ranking loss with the L2 regularization penalty.

L = L_{BPR} + L_{reg} = - \sum_{(u, i, j) \in D_{S}} \ln σ ({\hat{y}}_{u i} - {\hat{y}}_{u j}) + λ \sum_{θ \in Θ} || θ {||}_{2}^{2}

(4)

The model is trained using the objective function in Equation (4). This process involves sampling neighborhood subgraphs for ranking triplets, computing embeddings via the Siamese GCN, and updating weights via backpropagation until convergence. The final entity representation is denoted as shown in Equation (5).

e_{v} = H_{v}^{(L)}

(5)

In Equation (5), $e_{v}$ represents the high-dimensional embedding vector finally output by node $v$ . $H_{v}^{(L)}$ is the feature representation of node $v$ in the $L$ layer GCN network. In summary, cross-lingual scenarios are abstracted into heterogeneous graphs, and a Siamese GCN with a ranking-oriented loss function is constructed. This generates informative vector representations for each entity, laying the foundation for accurate Collaborative Deep Ranking (CDRs).

Cross-domain recommendation algorithm design integrating BLA

High-quality vector representations are obtained through Siamese GCN. However, relying on simple combinations may capture only broad similarities. A complex interaction layer can become a bottleneck for latency.²⁷ Deep-seated preference information is embedded within the relational attributes. Lack of mining this information leads to decreased accuracy.^28,29 To solve this, a dedicated feature interaction layer is built. The study introduces the bilinear (BL) mechanism to build this layer. The core operations consist of matrix multiplication and dot products. Accordingly, the study presents an attention method based on BL to create BLA, enabling the model to dynamically focus on important interactions. Its structure is shown in Figure 4.

FIG. 4.

Architecture of bilinear attention fusion and prediction mechanism.

In Figure 4, the product and user representation vectors enter independent attention branches. These branches dynamically learn discriminative patterns. After attention weighting, the enhanced representations are fed into a BL interaction module to generate the final interaction vector. Finally, the prediction score is calculated through an MLP. The calculation process of the weight is shown in Equation (6).

a_{u i} = softmax (w_{2}^{T} ReLU (W_{1} [p_{u}; p_{i}] + b_{1}) + b_{2})

(6)

In Equation (6), $a_{u i}$ represents the attention weight vector. $p_{u}$ and $p_{i}$ represent new representations of users and products projected into the interaction space, respectively. $T$ represents the transpose operation of a vector. $W_{1}$ represents the WM. $w_{2}$ , $b_{1}$ , and $b_{2}$ represent the WM and bias term of the fully connected layer, respectively. It incorporates attention-weighted BL interaction, namely BLA. The final interaction vector $z_{attn}$ is shown in Equation (7).

z_{attn} = a_{u i} ⊙ (p_{u} ⊙ p_{i})

(7)

Through Equation (7), the model dynamically amplifies crucial interactive features. For instance, activating specific user preference dimensions increases the weight of related product attributes while suppressing noise. This results in a discriminative vector $z_{attn}$ . Converting this into prediction scores requires an MLP, as shown in Equation (8).

h_{1} = ReLU (W_{h 1} z_{attn} + b_{h 1})

(8)

In Equation (8), $h_{1}$ is the output vector of the MLP hidden layer. $W_{h 1}$ and $b_{h 1}$ are the WM and bias term of the hidden layer. $h_{1}$ captures the deep pattern of interactive features and maps it to the final prediction score through an output layer. The output layer uses Sigmoid function (Sigmoid) as the activation function, as shown in Equation (9).

{\hat{y}}_{u i_{f} i nal} = σ (W_{out} h_{1} + b_{out})

(9)

In Equation (9), ${\hat{y}}_{u i_{f} i nal}$ represents the final prediction score of the model. $W_{out}$ and $b_{out}$ represent the weight and bias parameters of the output layer. To align predictions with user behavior, the difference between the score and the actual label is measured using a binary cross-entropy loss function. The specific expression is shown in Equation (10).

L_{rec} = - \sum_{(u, i) \in D} [y_{u i} \log ({\hat{y}}_{u i_{f} i nal}) + (1 - y_{u i}) \log (1 - {\hat{y}}_{u i_{f} i nal})]

(10)

In Equation (10), $L_{rec}$ represents the loss value of the recommendation task. $y_{u i}$ stands for authentic label. ${\hat{y}}_{u i_{f} i nal}$ represents the final predicted probability. Combining the loss functions of representation learning and recommendation tasks, the total loss consists of a weighted loss for the learning task and a cross-entropy loss for recommendation, as shown in Equation (11).

L_{total} = L_{rec} + ε L_{BPR}

(11)

In Equation (11), $L_{total}$ represents the final optimization goal. $ε$ is a hyperparameter. In summary, entity representation learning and feature interaction prediction are integrated to build a CDR method based on Siamese GCN and BLA. The details are shown in Figure 5.

FIG. 5.

Cross-domain recommendation framework for Station Computer (SC) based on Siamese GCN and BLA. GCN, Graph Convolutional Network; BLA, bilinear attention.

In Figure 5, interaction data are constructed into a cross-domain heterogeneous graph to generate training triples. These are fed into the Siamese GCN module to learn entity representations, guided by an auxiliary loss. Subsequently, the vectors enter the BLA module for deep feature fusion. The resulting interaction vector is mapped to a prediction score via an MLP and Sigmoid activation, enabling candidate ranking. In summary, by integrating graph modeling, Siamese GCN, and the BLA layer, an end-to-end method capable of handling cross-lingual scenarios is established.

Results and Analysis

The efficacy of the proposed approach is systematically verified. First, the superiority of Siamese GCN for entity representation is demonstrated against other networks. Second, the advancement of the BLA mechanism is verified against other interaction layers. Finally, ablation experiments confirm the necessity of each component.

Performance verification of Siamese GCN cross-domain entity representation learning

To verify performance, the Siamese GCN model is compared with GCN, GraphSAGE, and Relational Graph Convolutional Network (RGCN), which are representative baselines for large-scale and heterogeneous graphs. A public dataset from Alibaba International (B2B source domain) and AliExpress (B2C target domain) is used. The preprocessed dataset comprises 52,516 users, 65,832 products, 5128 suppliers, and 1,245,310 interactions. The data are randomly split 6:4 into training (747,186 records) and test (498,124 records) sets. Table 1 displays the experimental environment and key parameters.

Table 1.

Experimental environment and key parameter configuration

Experimental environment	Configuration information
Operating system	Ubuntu 20.04
Memory	32GB
Graphics Processing Unit (GPU)	Intel Core i7-12700H
Graphics Processing Unit (GPU)	NVIDIA GeForce RTX 3060
Deep learning framework	PyTorch
Programming language	Python 3.9
Initial learning rate	0.001
Embedding vector dimensions	64
Number of training rounds	100
Batch size	1024

The study uses HR and Recall Rate (RR) as core indicators. Scenario 1 is a sparse data environment, while Scenario 2 is a sufficient data environment. In Figure 6(a), the HR and RR of Siamese GCN in Scenario 1 reach 12.6% and 8.9%, respectively. The HR and RR of the RGCN model are 10.7% and 7.5%. In Figure 6(b), under Scenario 2, the HR of the Siamese GCN reaches 20.6%, surpassing the 19.5% of the RGCN model. Therefore, the Siamese GCN demonstrates significant performance advantages in complex cross-domain scenarios.

FIG. 6.

Comparison of HR and RR in different scenarios. HR, hit rate; RR, recall rate.

To further evaluate stability, the study compares the NDCG and Mean Average Precision (MAP) under different Recommendation List Lengths (RLL). In Figure 7(a), when K = 20, the NDCG of Siamese GCN reaches 7.8%. The NDCG values of the RGCN and GCN models are 6.9% and 6.4%, respectively. In Figure 7(b), the MAP value of Siamese GCN reaches 4.4%, while RGCN and GraphSAGE are 3.7% and 3.2%. This proves the robustness of the Siamese GCN model.

FIG. 7.

Comparison of model performance with different RLLs. RLLs, recommendation list lengths.

Precision and coverage are compared across various models. As shown in Figure 8(a), the Siamese GCN model maintains the highest accuracy, reaching 4.5% at K = 15, surpassing RGCN (3.9%) and GCN (3.2%). In Figure 8(b), at K = 20, its coverage reaches 18.8%, exceeding the RGCN model by 1.6 percentage points and surpassing GCN (14.3%). In summary, the Siamese GCN model achieves the best performance in balancing accuracy and diversity.

FIG. 8.

Comparison of precision and coverage with different RLLs. RLLs, recommendation list lengths.

Comparative analysis of BLA feature interaction mechanism

To verify the BLA mechanism, it is compared with the BL, CNIM, and Gated mechanisms. Figure 9 displays the performance. In Figure 9(a), at the 100th round, the NDCG of BLA reaches 23.8%, while the BL mechanism is 22.2%. In Figure 9(b), the MAP value of the BLA mechanism reaches 12.7%, exceeding the 10.8% of the BL mechanism. This nearly 2 percentage point increase in MAP usually translates into increased platform revenue, possessing significant business value.

FIG. 9.

Comparison of NDCG and MAP for different interaction mechanisms during the training process.

HR and accuracy are compared under standard (Scenario 1) and sparse (Scenario 2) conditions. As shown in Figure 10(a), BLA achieves 28.9% HR and 3.4% accuracy, surpassing BL (27.4%, 3.1%) and CNIM (26.2%, 3.0%). In Scenario 2 (Fig. 10b), BLA leads with 30.5% HR and 2.6% accuracy, significantly outperforming the Gated mechanism, which reaches only 24.9% HR and 1.2% accuracy. This proves the robustness of the mechanism.

FIG. 10.

Comparison of four mechanisms under standard and sparse data conditions.

Figure 11 displays ROC and RR performance. In Figure 11(a) (K = 20), BLA achieves 15.5% RR, exceeding BL (14.5%) and the Gated mechanism (12.5%). In Figure 11(b), at a 0.2 false positive rate, BLA achieves a ∼0.95 true positive rate, significantly higher than BL (0.85), CNIM (0.74), and Gated (0.63). This confirms BLA’s superior performance in user preference discrimination.

FIG. 11.

Comparison of performance of each mechanism in the recommendation task.

Verification of cross-domain recommendation method based on Siamese GCN and BLA

An ablation experiment is carried out to confirm the thorough performance of the CDR approach based on Siamese GCN and BLA suggested in the study. Table 2 displays the findings. The ablation experiment in Table 2 confirms the contribution of each component. Starting from the baseline, introducing heterogeneous graph modeling, Siamese GCN, and the BLA mechanism progressively improves performance, with the final NDCG@20 reaching 18.62%. In summary, as shown in Table 2, the introduction of each component resulted in a stable performance increase, indicating the effectiveness of these contributions. These components address business ecosystem structure modeling, cross-level information transfer, and accurate matching of relational connections. All components are essential.

Table 2.

Results of ablation experiment

Method	+Heterogeneous architecture modeling	+SiameseGCN	+Bilinear interaction	+Attention mechanism	NDCG@20 (%)	Recall@20 (%)
Baseline	×	×	×	×	14.33	15.88
Improved 1	√	×	×	×	15.32	17.14
Improved 2	√	√	×	×	16.77	19.03
Improved 3	√	√	√	×	17.75	20.31
Research method	√	√	√	√	18.62	21.40

GCN, Graph Convolutional Network.

To further verify the practical application effect of the method proposed in the study (Method 1) in different business scenarios, the study selected two typical real-life scenarios in CBEC platforms for special testing: one is to accurately recommend “high-value potential users,” and the other is to increase the sales rate of “long-tail products.” The study compares Method 1 with the CDR method based on GCN and BL mechanism (Method 2), the CDR method based on GraphSAGE and self-attention splicing (Method 3), and the CDR method based on matrix factorization and MLP (Method 4). Using NDCG and HR as core evaluation indicators, the comprehensive performance of the four methods is shown in Figure 12. In Figure 12(a), Method 1 achieves the best NDCG (16.8%), compared with Method 2 (15.1%), Method 3 (13.9%), and Method 4 (11.8%). Similarly, in Figure 12(b), Method 1’s HR reaches 46.5%, surpassing Method 2 (45.1%) and Method 4 (42.3%). These results demonstrate superior commercial application.

FIG. 12.

Comparison of NDCG and HR performance in different scenarios. HR, hit rate.

The study verifies accuracy under different RLLs. In Figure 13(a), at K = 20, Method 1 reaches 7.4%, exceeding Method 2 (6.9%) and Method 4 (5.8%). In Figure 13(b), for “high-value potential users” (k = 15), Method 1 achieves 12.5%, surpassing Method 2 (11.8%), Method 3 (10.8%), and Method 4 (9.5%). In addition, at RLL k = 15, Method 1 maintains 7.4% accuracy, higher than Method 2 (6.8%), Method 3 (6.3%), and Method 4 (5.9%). In summary, the proposed method achieves the best balance of precision and recall, demonstrating robustness in sorting performance.

FIG. 13.

Precision of each method at different RLLs. RLLs, recommendation list lengths.

To evaluate robustness in data-sparse scenarios, the study analyzes cold-start and long-tail distributions, as shown in Figure 14. Figure 14(a) indicates that Method 1 performs best in cold-start scenarios. With an average user interaction of only four times, Method 1 achieves a hit rate of 13.8%, surpassing Method 2’s 11.0%. Figure 14(b) shows Method 1’s significant advantage in long-tail recommendation, consistently outperforming comparison methods. Furthermore, as shown in the ROC curve in Figure 14(c), Method 1 achieves the highest Area Under the Curve (AUC), demonstrating superior ranking robustness.

FIG. 14.

Performance and robustness evaluation in key cross-domain scenarios.

Discussion

To address performance bottlenecks, a two-stage method is proposed. Experimental results verified the effectiveness of this method. In the entity representation stage, the Siamese GCN model was superior to baselines like GraphSAGE and RGCN, especially in sparse scenarios. Jin et al. showed that standard GCNs struggle to distinguish different relationships.³⁰ while Ye et al. pointed out that RGCN parameters increase linearly with relationships, leading to overfitting.³¹ The success of the Siamese GCN is attributed to two factors. First, the heterogeneous graph incorporates key ecosystem entities like suppliers and overseas warehouses. Second, the Siamese architecture directly optimizes the “UI” matching relationship. Unlike standard GCNs focusing on global representation, the Siamese GCN employs a dual-branch structure for “pairwise learning.” By minimizing the distance between positive samples and maximizing it for negative samples, this paradigm optimizes relative ranking, improving discriminative power.

In the feature interaction prediction stage, the BLA mechanism showed the best performance, with final NDCG and MAP values reaching 23.8% and 12.7%, respectively. Relevant research notes that simple splicing fails to capture second-order interactions, while cross networks lack explicit modeling between feature groups.³² The success of the BLA mechanism lies in its attention-enhanced interaction. First, BL interaction performs a dimension-wise combination of vectors. Second, the attention mechanism addresses the issue of treating all features equally by dynamically assigning weights to crucial information. Ablation experiments demonstrated that removing the attention mechanism caused NDCG@20 to drop by nearly 1 percentage point. This “focused” interaction enables the model to accurately locate purchase intentions. This efficient design, combined with offline representation learning, balances accuracy and low latency. Furthermore, the framework offers scalability by integrating non-product data, such as social relationships, to capture richer user preferences.

Summary and future work

Addressing the dual challenges of low latency and high accuracy in cross-lingual real-time recommendation, an efficient method with three core contributions is proposed. First, a two-stage framework balances performance by placing high-precision entity representation offline and low-latency feature interaction online. Second, a Siamese GCN method alleviates data sparsity by directly optimizing user-item matching relationships. Finally, a feature interaction mechanism integrating BLA improves accuracy by focusing on second-order interactions. Results show that in the entity representation stage, the Siamese GCN achieved HR and RR of 12.6% and 8.9% in sparse scenarios, respectively, outperforming baselines like GCN, GraphSAGE, and RGCN. In the feature interaction stage, the BLA mechanism exceeded traditional methods, achieving NDCG and MAP values of 23.8% and 12.7%, respectively. Ablation experiments confirmed the necessity of each component, improving NDCG@20 from a baseline of 14.33% to 18.62%. For “high-value potential users,” the method achieved 12.5% accuracy with a recommendation list length of 15. In summary, the proposed method effectively addresses data sparsity and complex relationships by constructing an end-to-end framework. Despite good results, limitations remain. Future research will explore three directions: introducing dynamic graph models for real-time data, combining reinforcement learning to optimize feedback loops, and utilizing advanced architectures like Graph Attention Networks to enhance expressive power.

Footnotes

Author Disclosure Statement

No competing financial interests exist.

Funding Information

The research is supported by 2024 Philosophy and Social Science Research Project of Jiangsu Universities, “Research on the Advantages, Challenges, and Paths of Digital Empowerment in the Governance of Harmonious and Beautiful Villages in Jiangsu” (Project No. 2024SJYB1728).

Abbreviations Used

References

1. He

, Wang

, Liu

, et al. Construction and implementation of cross-border e-commerce supply chain system under the background of green and low-carbon. JIDE 2024;4(1):1–11; doi: 10.1108/JIDE-01-2024-0002

2. Zhang

, Jia

, You

. Striking a balance between supply chain resilience and supply chain vulnerability in the cross-border e-commerce supply chain. Int. J. Logist. Res. Appl 2023;26(3):320–344; doi: 10.1080/13675567.2021.1948978

3. Dai

, Min

. Can cross-border e-commerce reform reduce supply chain risks? J Knowl Econ 2023;15(3):14998–15026; doi: 10.1007/s13132-023-01689-9

4. Han

, Rizwanullah

, Luo

, et al. The role of cross-border e-commerce on the export of goods and services. Electron Commer Res 2024;24(2):1367–1384; doi: 10.1007/s10660-024-09818-5

5. Tang

, Li

. Enhancing competitiveness in cross-border e-commerce through knowledge-based consumer perception theory: An exploration of translation ability. J Knowl Econ 2023;15(3):14935–14968; doi: 10.1007/s13132-023-01673-3

6. Haider

, Zeb

, Islam

AKMM

, et al. Enhancing IoT security with resource-efficient cryptography: A comprehensive review of lightweight and hybrid algorithms. Comput Sci Rev 2026;59:100861; doi: 10.1016/j.cosrev.2025.100861

7. Haider

, Zeb

, Rahman

, et al. A survey on anomaly detection in IoT: Techniques, challenges, and opportunities with the integration of 6G. Comput Netw 2025;270:111484; doi: 10.1016/j.comnet.2025.111484

8. Du

. Innovation of cross-border e-commerce supply chain management mechanism under digital background. IJNVO 2025;32(1/2/3/4):291–312; doi: 10.1504/IJNVO.2025.145394

9. Goswami

, Mondal

, Sarkar

, et al. Artificial intelligence-enabled supply chain management: Unlocking new opportunities and challenges. AIA 2024;3(1):110–121; doi: 10.47852/bonviewAIA42021814

10.

10. Omidvar

, Tran

. Tackling cold-start with deep personalized transfer of user preferences for cross-domain recommendation. Int J Data Sci Anal 2025;20(1):121–130; doi: 10.1007/s41060-023-00467-9

11.

11. Mandalapu

, Narayanan

, Putheti

. A hybrid collaborative filtering mechanism for product recommendation system. Multimed Tools Appl 2023;83(5):12775–12798; doi: 10.1007/s11042-023-16056-8

12.

12. Widayanti

, Chakim

MHR

, Lukita

, et al. Improving recommender systems using hybrid techniques of collaborative filtering and content-based filtering. J Appl Data Sci 2023;4(3):289–302; doi: 10.47738/jads.v4i3.115

13.

13. Huang

, Jia

, Zuo

. Improved collaborative filtering personalized recommendation algorithm based on k-means clustering and weighted similarity on the reduced item space. MMC 2023;3(1):39–49; doi: 10.3934/mmc.2023004

14.

14. Liu

, Zhou

, Kou

, et al. Privacy-preserving point-of-interest recommendation based on simplified graph convolutional network for geological traveling. ACM Trans Intell Syst Technol 2024;15(4):1–17; doi: 10.1145/3620677

15.

15. Huang

, Ma

, Liu

, et al. Position-enhanced and time-aware graph convolutional network for sequential recommendations. ACM Trans Inf Syst 2023;41(1):1–32; doi: 10.1145/3511700

16.

16. Ma

, Huang

, Lu

, et al. Kr-GCN: Knowledge-aware reasoning with graph convolution network for explainable recommendation. ACM Trans Inf Syst 2023;41(1):1–27; doi: 10.1145/3511019

17.

17. Yan

, Cheng

, Gao

, et al. Cascading residual graph convolutional network for multi-behavior recommendation. ACM Trans Inf Syst 2024;42(1):1–26; doi: 10.1145/3587693

18.

18. Li

, Cai

, Hanafiah

, et al. An empirical study on personalized product recommendation based on cross-border e-commerce customer data analysis. J. Organ. End User Comput 2024;36(1):1–16; doi: 10.18280/ts.410615

19.

19. Chen

, Lan

, Chang

. Consumer behaviour in cross-border e-commerce: Systematic literature review and future research agenda. Int. J. Consum. Stud 2023;47(6):2609–2669; doi: 10.1111/ijcs.12969

20.

20. Jian

, Lv

, Wan

. Exploring consumer purchase intention in cross-border e-commerce: Evidence from ‘belt and road’ countries. APJML 2023;35(3):625–644; doi: 10.1108/APJML-12-2021-0934

21.

21. Lu

, Yeh

, Liau

. Exploring the key factors affecting the usage intention for cross-border e-commerce platforms based on DEMATEL and EDAS method. Electron Commer Res 2023;23(4):2517–2539; doi: 10.1007/s10660-022-09548-6

22.

22. Yang

. The improved BP neural network supply chain performance evaluation model under the cross-border e-commerce marketing chain. JCM 2024;24(3):1727–1740; doi: 10.3233/JCM-247290

23.

23. Wang

, Sun

. Evaluation method of cross-border e-commerce supply chain innovation mode based on blockchain technology. IJITM 2024;23(3/4):261–277; doi: 10.1504/IJITM.2024.139572

24.

24. Chen

, Gan

, Wu

, et al. Data scarcity in recommendation systems: A survey. ACM Trans Recomm Syst 2025;3(3):1–31; doi: 10.1145/3639063

25.

25. Shao

, Wu

, Zhang

, et al. Average user-side counterfactual fairness for collaborative filtering. ACM Trans Inf Syst 2024;42(5):1–26; doi: 10.1145/3656639

26.

26. Li

, Xiao

, Chen

, et al. An explicitly weighted GCN aggregator based on temporal and popularity features for recommendation. ACM Trans Recomm Syst 2023;1(2):1–23; doi: 10.1145/3587272

27.

27. Sharafian

, Monirul

, Mokarram

, et al. Adaptive fuzzy controller for chaos suppression in nonlinear fractional order systems. ICCK Transactions on Advanced Computing and Systems 2024;1(1):5–18; doi: 10.62762/TACS.2024.318686

28.

28. Zheng

, Qu

, Cui

, et al. AutoML for deep recommender systems: A survey. ACM Trans Inf Syst 2023;41(4):1–38; doi: 10.1145/3579355

29.

29. Zhang

, Xie

, Hou

, et al. Recommendation as instruction following: A large language model empowered recommendation approach. ACM Trans Inf Syst 2025;43(5):1–37; doi: 10.1145/3708882

30.

30. Jin

, Zhu

. A systematic study and analysis of graph neural networks under noise. ACM Trans Knowl Discov Data 2025;19(6):1–20; doi: 10.1145/3733605

31.

31. Ye

, Bao

, Xu

, et al. RD-GCN: A role-based dynamic graph convolutional network for information diffusion prediction. IEEE Trans Netw Sci Eng 2024;11(5):4923–4937; doi: 10.1109/TNSE.2024.3403652

32.

32. Brynildsen

, Rajan

, Henderson

, et al. Network models to enhance the translational impact of cross-species studies. Nat Rev Neurosci 2023;24(9):575–588; doi: 10.1038/s41583-023-00720-x