Supply Chain Disruption Risk Prediction Based on Hypergraph Representation and Dynamic Relational-Attentive
A
JinlongWang1Emailwangjinlong@qut.edu.cnEmail202323050901@stu.qut.edu.cn
QixinZhao1Emailzzzqx0707@163.com
YingminLiu1
PengjunLi2Emailqdlipengjun@163.com
YuanyuanZhang1Emailyyzhang1217@163.com
XiaoyunXiong1Emailxxyqd@126.com
1School of Information and Control EngineeringQingdao University of Technology266520HuangdaoQingdao, ShandongChina
2China Mobile Communications Corporation Shandong Co., Ltd. Jinan Branch266034ShizhongJinan, ShandongChina
Jinlong Wang1 · Qixin Zhao1 · Yingmin Liu1 · Pengjun Li2 · Yuanyuan Zhang1 · Xiaoyun Xiong1
Abstract
Traditional supply chain risk prediction methods, relying on historical data, expert judgment, scenario analysis, and simulation, exhibit limitations in handling sudden events and high uncertainty within complex systems. Typically leveraging historical semantic links in knowledge graphs, these methods forecast future relational facts among companies. To address these shortcomings, we construct a supply chain risk knowledge graph integrating multi-dimensional enterprise features. We propose a novel Hypergraph Dynamic Graph Attention Neural Network (HG-DRA) for disruption risk prediction. HG-DRA employs hypergraph representation learning and a dynamic relational attention mechanism. Experiments demonstrate that HG-DRA, by effectively integrating operational features, cluster characteristics, and complex heterogeneous graph relationships, outperforms existing machine learning and graph representation learning approaches in identifying supply chain disruption characteristics.
Keywords:
Supply chain
Disruption risk prediction
Knowledge graph
Hypergraph Representation Learning
Dynamic relational attention
∗ Jinlong Wang
wangjinlong@qut.edu.cn
Qixin Zhao
zzzqx0707@163.com
Yingmin Liu
202323050901@stu.qut.edu.cn
Pengjun Li
qdlipengjun@163.com
Yuanyuan Zhang
yyzhang1217@163.com
Xiaoyun Xiong
xxyqd@126.com
1 School of Information and Control Engineering, Qingdao University of Technology, Huangdao, Qingdao 266520, Shandong, China
2 China Mobile Communications Corporation Shandong Co., Ltd. Jinan Branch,Shizhong,Jinan 266034,Shandong,China
A
1 Introduction
Supply chain strength serves as a crucial indicator of a nation or region's technological advancement and overall competitiveness. The establishment of supply chain risk assessment models to effectively manage and mitigate potential risks has become increasingly vital. The complexity of supply chains increases with the number of companies involved, making the critical relationships within the supply chain more intricate and difficult to measure. This underscores the importance of extracting relevant information from graph data. In recent years, in addition to early methods based on graph theory for constructing supply chain risk assessment models, machine learning approaches have been proposed for tasks like link prediction or clustering different labels[1]. Existing research has demonstrated that in a supply chain network, all enterprises are interconnected; once a small or medium-sized enterprise defaults on a loan, the impact can ripple through the entire supply chain [2, 3]. Concurrently, the use of knowledge graphs (KG) or graph neural networks (GNN) for knowledge inference in supply chain link prediction has gained increasing popularity [4, 5].
As a large-scale semantic association network, knowledge graphs [6] have demonstrated exceptional capabilities in knowledge association and inference. They can model the complex interest relationships among companies within a supply chain, injecting various risk indicators as entity attributes and attributes like supply amounts and sales amounts as inter-company supply relationship attributes into the supply chain disruption risk network. By combining the topological structure information between enterprise nodes with fine-grained risk attribute information, the disruption risk prediction model's ability to identify high-risk companies is enhanced. However, research on supply chain enterprise risk knowledge graphs primarily focuses on themes like bankruptcy and financial fraud [7–9], which cannot be directly applied to the prediction of supply chain disruption risks.
A
Considering the heterogeneous multi-source nature of internal risk data and the complexity of inter-company association risks within the supply chain, we integrate internal supply chain risks with association risks among companies. By leveraging comprehensive financial and non-financial features, we identify internal risks, while a heterogeneous hypergraph neural model captures association risks, such as buyer-seller relationships. This model explores hyperedges in the supply chain knowledge graph (e.g., same industry or region) to enhance enterprise risk prediction. For instance, during COVID-19, medical companies producing masks and vaccines thrived, whereas the catering industry faced severe supply chain disruptions.To address the aforementioned issues, this study proposes a knowledge graph-based method for predicting supply chain disruption risks. It leverages a supply chain risk knowledge graph to integrate various association information of companies within the supply chain. Using a hypergraph neural network model, we apply risk factors related to the same industry or region of enterprises to the task of supply chain disruption prediction. Simultaneously, we employ a hierarchical dynamic relational attention aggregation mechanism to capture the complex relationships among enterprises in the supply chain risk knowledge graph.
1.
1.We designed a supply chain risk knowledge graph for supply chain entities and their relationships, constructed the schema layer of the knowledge graph, and completed the data layer mapping based on various information such as basic information of each company, financial indicators, inter-company supply relationships, and sales relationships. This enables effective modeling of supply relationships among companies in the supply chain and various risk correlations.
2.
2.Based on the various relationships within the supply chain risk knowledge graph, such as the same industry or region, we constructed a hypergraph of supply chain enterprise associations. We propose the extraction of clustering risk features of supply chain enterprises using a hypergraph neural network, integrating graph features related to regulatory risks, clustering risk features, and basic information of enterprises to enhance the comprehensive performance of the enterprise regulatory risk prediction task.
3.
3.To differentiate the degree of association among different suppliers within the supply chain risk knowledge graph, we employ a hierarchical dynamic relational attention aggregation mechanism [10]. This mechanism adaptively computes different attention coefficients for each enterprise node in the supply chain, thereby better capturing the complex relationships in the heterogeneous graph.
2 Related Work
The modern supply chain is increasingly confronted with various uncertainties, including economic globalization, volatile international dynamics, new regulatory policies, and the strong interconnectivity among companies within the supply chain. These factors make supply chains more susceptible to natural and human-induced disruptions, as well as potential risks. Risks in the supply chain can be broadly categorized into two types: the first type is disruption risk caused by unpredictable external events (e.g., natural disasters, the COVID-19 pandemic). Although such risks have a low frequency of occurrence, they can lead to prolonged production interruptions, resulting in significant economic losses and societal impacts. The second type is operational risk, which typically arises from internal factors within supply chain enterprises (e.g., demand, supply, and cost fluctuations) and occurs with greater frequency[11]. The increasing complexity of supply chains has significantly amplified the importance of studying disruption risks in this domain.
2.1 Traditional Methods for Supply Chain Disruption Risk Prediction
Traditional methods for supply chain risk prediction are primarily based on statistical analysis and machine learning models. These approaches rely on historical data and probabilistic models, using mathematical modeling to quantify supply chain risks. Statistical analysis methods typically assume that the data meet specific distributions or independence conditions, which are often challenging to satisfy in highly complex and dynamic supply chain environments.
With advancements in computational capabilities, machine learning models have been widely adopted in supply chain risk prediction. For example, techniques such as support vector machines, random forests, and neural networks have been employed to predict the likelihood of supply chain disruptions. Accurate prediction of credit risk for small and medium-sized enterprises (SMEs) is critical to the sustainability of the entire supply chain and its participants, such as core enterprises and financial service providers. [12] proposed a novel approach combining an improved sparrow search algorithm (ISSA) and Light Gradient Boosting Machine (LightGBM) to address the limitations of traditional machine learning algorithms in handling high-dimensional and complex nonlinear data, which often result in suboptimal classification performance. [13] introduced a new method for SME credit risk prediction in supply chain finance by integrating multi-source information and employing an imbalanced sampling strategy. Their study demonstrates that financial information is the primary source for credit risk prediction; however, other sources, such as operational information, innovation metrics, and adverse events, also significantly impact credit risk assessment.
2.2 Supply Chain Disruption Risk Prediction Method Based on Knowledge Graph
Knowledge graphs have increasingly become a research focus in the field of supply chain risk prediction due to their powerful capabilities in knowledge representation and reasoning. By constructing semantic networks of entities and their relationships, knowledge graphs can intuitively represent the complex structures of supply chains, enabling unified modeling and analysis of multi-source heterogeneous data. Graph neural network (GNN) [14–17] technology is continuously improving the performance of relational data modeling. In particular, knowledge graph-based methods have performed well in many real-world scenarios [18, 19], which provide us with experience in exploring enterprise supply chain relationships. [20] propose a neural symbolic machine learning method combining graph neural network (GNN) and knowledge graph (KG) is proposed to solve the hidden risk problem in supply chain risk management.
The correlation between enterprises in the supply chain is also worthy of attention. There are multiple connection paths between two given enterprises in the supply chain risk knowledge graph, and these connection paths convey a variety of semantic information, Summarizing these path connection patterns can make more accurate predictions [21]. In addition, regional economic imbalance and diversified industrial structure have led to diversified geographical distribution of supply chains, which means that the geographical information of small and medium-sized enterprises should also be considered in the prediction of supply chain disruption risks [22]. [23] constructed an enterprise network based on supply chain relationships and utilized event mining and graph neural networks (GNNs) to detect enterprise risks. [24] proposed a fraud detection method based on a dynamic attention mechanism, which demonstrated excellent performance in fraud detection tasks on heterogeneous graphs. In summary, supply chain risk prediction methods have evolved from traditional statistical analysis and machine learning models to cutting-edge techniques based on knowledge graphs.
3 Problem Formulation
Definition 1
Supply Chain Risk Knowledge Graph
Supply Chain Risk Knowledge Graph is defined as
,where
represents the set of nodes of the knowledge graph;
represents the set of edges of the knowledge graph;
represents the type of edge in the knowledge graph.
Definition 2
Supply Chain Enterprise association Hypergraph
Supply Chain Enterprise association Hypergraph is defined as
,
represents the set of enterprise nodes in the supply chain association hypergraph;
represents the hyperedge set of the enterprise association hypergraph, which contains three types of hyperedges;
represents the set of hyperedges of each type in the hypergraph. The function
represents the mapping from the index of each hyperedge in the hypergraph to its corresponding type in
.
is the diagonal matrix composed of the weights of each hyperedge.
Definition 3
Supply Chain Multi-Relation graph
Supply-Sales Multi-Relation graph is defined as
where
is a set of node,
is a set of relations,
,
,
represents node i and node j and their relation.
is the feature matrix for the nodes, and
is the dimension of each feature vector.
Problem Supply Chain Disruption Risk Prediction
The purpose is to construct a supply chain risk graph
and a hypergraph
based on the basic information and relationships of enterprises in the supply chain every year, and at the same time aggregate the multi-layer representation
of the target enterprise to capture the correlation between supply chain enterprises and predict whether the target enterprise has risks. If there are risks, the supply chain may be interrupted.
4. Methodology
A
In this section, we will provide a detailed introduction to the proposed knowledge graph-based method for predicting supply chain disruption risks, as shown in Fig. 1. Our model primarily consists of four steps: (1) Data preprocessing and construction of the supply chain risk knowledge graph; (2) Construction of the supply chain associated enterprise hypergraph; (3) Multi-layer dynamic relational attention aggregation; (4) Prediction of supply chain disruption risk. First, we build a supply chain enterprise risk knowledge graph based on the basic information, supply, and sales information of related enterprises in the supply chain. Then, using the Hypergraph Neural Network (HGNN) representation model, we extract common risks faced by associated enterprises in the same industry or region. Subsequently, the relational attention mechanism dynamically identifies important suppliers and buyers, aggregating their information, while the multi-head attention mechanism further enhances the model's expressive capability. Our method embeds node features of each company into the risk knowledge graph, effectively integrating basic enterprise information, operational metrics, risk indicators, clustering features of associated enterprises, and multi-layer information aggregation of the target enterprise, to predict whether there is a risk of supply chain disruption.Fig. 1
Overall architecture diagram of the supply chain disruption risk prediction model
4.1 Data Collection and Preprocessing
Predicting regulatory violations of related enterprises in the supply chain typically requires the annual economic information, audit information, and legal records of enterprises as data support. We referred to related studies on predicting violations by listed companies [25–27] and collected basic operational information of related enterprises in the supply chain from multiple dimensions. We arranged the absolute values of Pearson correlation coefficients [28] corresponding to each company's basic information indicators in descending order. Finally, we retained the top 11 indicators. Combining existing research on credit risk evaluation indicators in supply chain finance, this study selects financial risk indicators for manufacturing supply chain enterprises from four aspects: solvency, operational capability, profitability, and risk level, and selects two types of non-financial indicators as supplements: Annual margin and audit opinion category. The specific indicators are shown in Table 1.
First-level indicator | second-level indicators | Specific indicators | Type |
|---|---|---|---|
Financial indicators | solvency | Cash ratio | Positive |
quick ratio | Positive | ||
Capital fixed ratio | Negative | ||
operational capability | Inventory turnover rate | Positive | |
Total asset turnover ratio | Positive | ||
profitability | Return on equity | Positive | |
Operating Profit Ratio | Positive | ||
risk level | Financial leverage | Negative | |
Operating leverage | Negative | ||
Non-financial indicators | basic information | Annual margin | Positive |
Audit opinion category | Positive |
We performed data cleaning on the collected data: for handling missing values, considering the significant differences in individual continuous operational performance indicators, the mean was not reflective of the median level of the related dimension. Therefore, we used the median instead of the mean to fill in the missing values in the raw data. Since discrete variables had almost no missing data, we used this mode to fill in the missing values at a very low rate.
To evaluate the performance of the proposed bankruptcy prediction model, we manually collected and pre-processed a real-world dataset of listed companies. Specifically, we selected information on listed companies from the China Stock Market & Accounting Research (CSMAR) database.
4.2 Supply Chain Risk Knowledge Graph Construction
Considering that the data sources for this paper include a large number of basic information tables of listed companies and inter-company association information tables, which meet the requirements of a top-down approach for structuring data in knowledge graph construction. At the same time, specific definitions were provided for the relationships between listed companies and the attributes of regulatory violations. Therefore, a top-down approach is more suitable for constructing the supply chain knowledge graph.
The transformed data is stored in the Neo4j graph database to create a knowledge graph of company associations and regulatory violation information. Given the frequent changes in inter-enterprise interest associations within the supply chain and the non-specific timing of related enterprises' regulatory violations, a year-based storage method is adopted, storing supply chain knowledge graphs of different years in separate graph database backups. Each year's knowledge graph contains varying numbers of entities and relationships, covering over 10,000 entities and 20,000 relationships.
Figure 2 is part of the supply chain knowledge graph we have constructed.
Fig. 2
Part of constructed supply chain knowledge graph
4.3 Supply Chain Disruption Risk Prediction Model
A method for predicting supply chain disruption risk based on knowledge graphs is proposed. It leverages the supply chain risk knowledge graph to integrate various association information aspects of supply chain enterprises and employs a hypergraph neural network model to apply risk factors related to the same industry, region, etc., of enterprises to the task of disruption risk prediction. With the dynamic attention mechanism, the attention coefficients of each enterprise node can be adaptively adjusted based on the information from its neighbors and the current layer, thereby better capturing the features of enterprises and the importance of supply and sales relationships.
4.3.1 Construction of related enterprise hypergraph
Hypergraphs play a crucial role in predicting supply chain disruption risks because hyperedges reflect common factors that enterprises in the supply chain face. Therefore, it is natural to use hypergraphs to capture shared risk information, such as industry downturns, regional economic policy changes, or guarantee risks caused by the same stakeholders. This paper constructs a Hypergraph Neural Network (HGNN) to achieve vectorized representations of cluster risk features for associated enterprises.
Hyperedges in a hypergraph can connect multiple nodes simultaneously [29]. Based on the constructed knowledge graph
, we categorize companies involved in the same region, industry, or with the same investors. The representation of any subset
is as follows:
the subset
includes
enterprise nodes
and hyperedge types belong to one of the three types: Industry relations
, region relations
, and shareholder relations
By integrating all subsets
. we can construct the enterprise association hypergraph. According to the definition of hypergraph in section 3, the hypergraph
has the node set
, the hyperedge set
, the type set
of each hyperedge, and the diagonal matrix
of hyperedge weights. The enterprise association hypergraph
can be represented by an association matrix
with a scale of
, specifically as follows:
Based on the adjacency matrix of each generated hypergraph
, we need to weight and sum the hyperedges
connected to the nodes
, as well as the weights
of each hyperedge, to calculate the node degree
of each node in
.Additionally, we need to sum up the number of nodes connected by each hyperedge to obtain the marginal degree of the hyperedge, denoted as
. We then place the
and
of each hypergraph into matrices and perform a diagonalization operation to obtain the diagonal matrices
and
, respectively.
We integrate hypergraph neural network HGNN[30] to complete the aggregation of enterprise associated cluster features. the aggregation of enterprise associated cluster characteristics involves two stages of neural network computations: In the first stage, based on the enterprise node features in each hypergraph
and the adjacency matrix
, we perform the operation
to sum up the feature vectors of the points connected by the hyperedge, and obtain the hyperedge features of each hypergraph. In the second stage, we aggregate the hyperedge features by using the operation
to update the features of the enterprise nodes in each hypergraph. The specific iterative update formula for the node feature vector in each hypergraph is as follows:
Where
represents the number of network iterations,
and
are the feature vector representations of each node in the hypergraph after the
and
iterations, respectively.
and
, are the diagonal matrix of node degree and edge degree.
is the diagonal matrix of hyperedge weights, which needs to be obtained through model training hyperparameters.
is the vector dimension conversion matrix for the
layer.
Finally, the node feature vector
output after the last iteration is taken as the feature representation of the enterprise associated cluster.
4.3.2 Dynamic Relation Attention aggregation
This module tackles the challenge of modeling diverse relationship types in real-world graphs by independently computing node representations for each relationship type within the supply chain risk knowledge graph.
We propose learning multiple node representations for each node by computing a representation per relation and a self-transformation. The neighborhood set of
for
is defined by
,The node representation of
for
at the
-th layer is denoted by
,
represents the dimension of node representations at the current stage. Let
be the node representation of
at the
-th layer. We compute
for
,where MLP is a multi-layer perceptron. Using the dynamic multi-head attention with
heads, we compute the node representation for
at each head as follows:
Which
,
is calculated using the following formula
where
denotes a vertical concatenation,
is a non-linear function.
. To aggregate the outputs from
heads, we concatenate the resulting representations from different heads. By computing a node representation per relation using (6), we have
representations for each node.
For each node, we have
) representations at each layer as described above. Then aggregating these representations using a dynamic attention with
attention heads:
Which
,and
is computed by
The attention coefficient
indicates the importance of the
-th relation for computing the representation of
at the
-th layer.
.This attention coefficient can differ depending on nodes and layers.
Then, we aggregate different layers with
heads, and
:
is computed by
Note that the attention coefficient
is learned to imply the importance of the
-th layer’s representation for computing the final representation of node
.
.Finally, the node feature vector
output after the last iteration is taken as the feature representation.
4.3.3 Supply chain Interrupt prediction
We concatenate
with
,and additional information initially collected
and fuse them into a more comprehensive vector representation
of the listed enterprise's own characteristics through linear layer transformation, as shown in the following formula:
and
are the training hyperparameters. Using the Mish[31] activation function, we perform nonlinear feature transformation on the enterprise associated cluster feature, and further use the MLP neural network with two hidden layers and the ReLU[32] activation layer to make the feature dimension of
. We introduce a hyperparameter
to achieve effective concatenation of the mixed features
of the enterprise's own operational and risk information and the embedding features
. The comprehensive feature representation
of violation risk of each enterprise node is obtained, and the specific formula is as follows:
In order to output the final classification of the violation prediction of listed enterprises, we map the
comprehensive features of each enterprise node into the prediction category of violation through a fully connected layer and use the minimum cross-entropy loss function to evaluate the model loss
. The corresponding formula is as follows:
Where
represents the company entries in the training set with labeled violation tags,
represents whether the target enterprise in the supply chain has disruption risk.
is the hyperparameter that needs to be trained in the prediction model, and
represents the bias vector.
5 Experiments and results
To validate the performance of our model for predicting supply chain disruption risks, we first introduce our experimental dataset, evaluation metrics, and experimental setup. Then, we compare our proposed method with existing baseline methods to assess its performance.
5.1 Experimental data
We collected operational information of listed companies in China and data on related companies in the supply chain from 2020 to 2022. The data sources include the CSMAR financial database and public disclosures from the China Securities Regulatory Commission, the Shenzhen Stock Exchange, and the Shanghai Stock Exchange. Given the time-dependent nature of the risk effects of supply chain disruption events and the relatively long time span of the target samples, risk information may differ across different years. Moreover, fluctuations in both the internal and external supply chain environment can have an impact, resulting in significant differences in operational performance indicators across different years. Therefore, it is not suitable to allocate all data into a single dataset for training the prediction model. We constructed three supply chain knowledge graphs at different times by year, which serve as data support for evaluating enterprise risk indices and constructing hypergraphs, but they do not directly participate in model training. The scale of each knowledge graph is shown in Table
KG2020 | KG2021 | KG2022 | |
|---|---|---|---|
Enterprise nodes | 2451 | 2437 | 2372 |
Shareholders nodes | 3392 | 3425 | 3410 |
Industry nodes | 38 | 38 | 38 |
Location nodes | 27 | 26 | 27 |
Amount nodes | 11235 | 10994 | 10326 |
All nodes | 17143 | 16920 | 16173 |
All relationships | 52827 | 53098 | 52735 |
Subsequently, we constructed datasets for three different years to predict supply chain disruption risks. Each dataset used for disruption prediction includes samples of indicator types for companies and their enterprise association hypergraphs. The indicator type samples consist of the company's basic operational indicators for the year and violation category labels.
5.2 Experimental setup
We implemented the model for predicting supply chain disruption risks along with other baseline models using the PyTorch deep learning framework and the PyTorch Geometric (PyG) graph neural network library. Regarding the parameter settings for the neural network models, we used the AdamW optimizer as the optimization strategy and employed a cosine annealing learning rate scheduler to train the neural network models. The initial learning rate was set to 0.0005, the models were trained for 300 epochs, and the dropout rate was set to 0.2. We evaluated the model's performance on the test dataset to assess its classification effectiveness. To comprehensively evaluate the classification performance of the model, we chose accuracy, recall, and F1 score as the evaluation metrics for the classification models. The specific calculation formulas are shown in equations (16) to (18). .Additionally, considering the imbalance between violation samples and compliance samples, we select metrics G-mean (19) and AUC suitable for evaluating the prediction performance of the binary model under the unbalanced data set.
Where TP represents the number of correctly predicted violation samples, TN represents the number of correctly predicted compliance samples, FP represents the number of compliance samples predicted as violations, and FN represents the number of violation samples predicted as compliance.
5.3 Results of the Analysis
The main methods for quantitatively predicting supply chain disruption risks are based on machine learning models or graph representation learning models. To verify the effectiveness of these methods, we selected traditional machine learning models and graph representation learning models as baseline models in the field of supply chain risk control and compared them with the proposed supply chain disruption risk prediction model. The input data for the machine learning models consists solely of the company's operational indicators. On the other hand, graph-based learning models use the company's operational indicators, triplet information from the knowledge graph, and meta-path information generated from relationships in the knowledge graph as input.
We compare the following specific baseline models:
(1) Logistic Regression (LR)[33]: A statistical model used for binary classification problems.
(2) LightGBM[34]: a gradient boosting framework that uses tree-based learning algorithms, designed for speed and efficiency, especially for large datasets.
(3) RGCN[35]: a type of graph neural network designed to handle multi-relational graph data by incorporating different types of relationships between entities in the graph.
(4) RGAT[36]: an extension of the graph attention network (GAT) that leverages attention mechanisms to model multi-relational graph data, capturing the importance of different relations between entities.
(5) Ie-HGCN [37]: An interpretable and efficient heterogeneous graph convolutional network automatically discovers effective meta-paths, ensuring both high graph representation capacity and improved interpretability.
Models | AUC | G-Mean | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
LR | 0.7833 | 0.6574 | 0.5439 | 0.4794 | 0.5104 |
LightGBM | 0.8106 | 0.7155 | 0.6389 | 0.5633 | 0.5998 |
RGCN | 0.8457 | 0.7558 | 0.6471 | 0.6604 | 0.6328 |
RGAT | 0.8513 | 0.7613 | 0.6303 | 0.6728 | 0.6455 |
Ie-HGCN | 0.8424 | 0.7592 | 0.6458 | 0.6773 | 0.6392 |
HG-DRA | 0.8592 | 0.7724 | 0.6374 | 0.6815 | 0.6503 |
Models | AUC | G-Mean | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
LR | 0.8092 | 0.6938 | 0.5989 | 0.5604 | 0.5814 |
LightGBM | 0.7904 | 0.7235 | 0.6032 | 0.6607 | 0.6353 |
RGCN | 0.8275 | 0.7758 | 0.6564 | 0.6804 | 0.6528 |
RGAT | 0.8317 | 0.7626 | 0.6822 | 0.6693 | 0.6701 |
Ie-HGCN | 0.8339 | 0.7891 | 0.6438 | 0.6740 | 0.6679 |
HG-DRA | 0.8402 | 0.7924 | 0.6603 | 0.6815 | 0.6793 |
Models | AUC | G-Mean | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
LR | 0.7983 | 0.6998 | 0.6059 | 0.6004 | 0.5794 |
LightGBM | 0.7992 | 0.7539 | 0.6689 | 0.6323 | 0.6041 |
RGCN | 0.8075 | 0.7885 | 0.6854 | 0.6792 | 0.6328 |
RGAT | 0.8007 | 0.7901 | 0.6930 | 0.6799 | 0.6049 |
Ie-HGCN | 0.7829 | 0,7742 | 0.6827 | 0.6603 | 0.6383 |
HG-DRA | 0.8102 | 0.7732 | 0.7072 | 0.6824 | 0.6503 |
We conducted multiple rounds of experiments on three datasets and calculated the average values to reflect the performance of the models, testing their effectiveness and generalization ability. The experimental results are shown in Table 3, Table 4, and Table 5. The results indicate that the proposed model exhibits good predictive performance.
It can be observed that heterogeneous graph neural network models (such as RGCN) outperform machine learning models in terms of predictive performance. This is because heterogeneous graph networks not only learn the basic information of each enterprise but also take into account the risk association relationships between enterprises. This demonstrates that the diverse relationships among enterprises in the supply chain are an important means for predicting supply chain disruption risks. Our model effectively captures the relationships between supply chain enterprises, making full use of the macro-level relationships and multi-layer enterprise relationships related to risk.
5.4 Ablation Experiments
In addition to the comparative experiments mentioned above, we conducted ablation experiments from three perspectives to analyze the importance of different modules in our model for the task of predicting supply chain disruption risks. The specifics are as follows:
(1) HG-DRA w/o HGNN: Removing the hypergraph neural network (HGNN) to eliminate the acquisition of supply chain-related enterprise association clustering features.
(2) HG-DRA w/o DRA: Removing the dynamic relational attention network to eliminate the acquisition of multi-layer risk relationship features among enterprises.
(3) HG-DRA w/o basic: Removing the basic information features of supply chain enterprises.
Models | AUC | G-Mean | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
HG-DRA w/o HGNN: | 0.8195 | 0.7463 | 0.6619 | 0.6104 | 0.6394 |
HG-DRA w/o DRA | 0.8197 | 0.7238 | 0.6489 | 0.5793 | 0.6048 |
HG-DRA w/o basic | 0.7853 | 0.6985 | 0.5664 | 0.4692 | 0.5128 |
HG-DRA | 0.8230 | 0.7793 | 0.6592 | 0.6024 | 0.6273 |
The results show that removing any module has a certain impact on the experimental performance. In particular, after removing the basic information of the enterprise, the recall rate and F1 value have a significant decrease.
6 Conclusion
This study leverages knowledge graph technology, hypergraph neural networks, and graph attention techniques to comprehensively capture and analyze the complex relationship network among upstream and downstream enterprises in the supply chain. This encompasses the associations between suppliers, manufacturers, distributors, retailers, and other stakeholders, as well as their dependencies and multi-layer risk association relationships. We have verified that dynamically adjusting the attention coefficients of enterprise nodes on heterogeneous graphs is effective for risk detection. Our model effectively addresses the challenges of predicting supply chain disruption risks inherent in traditional supply chain management methods.
Acknowledgements
This research was supported by the Shandong Province Science and Technology Project (2023TSGC0509, 2022TSGC2234), Qingdao Science and Technology Plan Project (23-1-5-yqpy-2-qy).
A
A
Author Contribution
Wang: Conceptualization, Methodology ; Zhao: Investigation, Software,original draft, Methodology ; Liu: Data Collection, Writing–review;Li: Methodology, Writing–review ;Zhang: Writing–review and editing;Xiong: Writing–review and editing.
A
Data Availability
Data are available on request to the authors.
Declarations
Conflict of interest
The authors declare that they have no conflict of interest.
References:
1.
Agarwal N, Seth N, Agarwal A (2022) Evaluation of supply chain resilience index: a graph theory based approach. BIJ 29:735–766. https://doi.org/10.1108/BIJ-09-2020-0507
2.
Gao Q, Xu H, Li A (2022) The analysis of commodity demand predication in supply chain network based on particle swarm optimization algorithm. J Comput Appl Math 400:113760. https://doi.org/10.1016/j.cam.2021.113760
3.
Liu J-B, Peng X-B, Zhao J (2023) Analyzing the spatial association of household consumption carbon emission structure based on social network. J Comb Optim 45:79. https://doi.org/10.1007/s10878-023-01004-x
4.
Brintrup A, Wichmann P, Woodall P et al (2018) Predicting Hidden Links in Supply Networks. https://doi.org/10.1155/2018/9104387. Complexity 2018:9104387
5.
Kosasih EE, Brintrup A (2022) A machine learning approach for predicting hidden links in supply chain with graph neural networks. Int J Prod Res 60:5380–5393. https://doi.org/10.1080/00207543.2021.1956697
6.
Wu F, Xiao X, Zhou X et al (2022) Complex risk contagions among large international energy firms: A multi-layer network analysis. Energy Econ 114:106271. https://doi.org/10.1016/j.eneco.2022.106271
7.
Zhang Z, Ji Y, Shen J et al (2022) Heterogeneous Information Network based Default Analysis. on Banking Micro and Small Enterprise Users
8.
Kui Z, Yu X, Wei WP, Gang YC (2022) A method of Evaluation for Small and Medium-sized Enterprises. In: Proceedings of the 2022 14th International Conference on Machine Learning and Computing. Association for Computing Machinery, New York, NY, USA, pp 358–367
9.
Li J, Chang Y, Wang Y, Zhu X (2023) Tracking down financial statement fraud by analyzing the supplier-customer relationship network. Comput Ind Eng 178:109118. https://doi.org/10.1016/j.cie.2023.109118
10.
Brody S, Alon U, Yahav E (2022) How Attentive are Graph Attention Networks?
11.
Shekarian M, Mellat Parast M (2021) An Integrative approach to supply chain disruption risk and resilience management: a literature review. Int J Logistics Res Appl 24:427–455. https://doi.org/10.1080/13675567.2020.1763935
12.
Hou L, Bi G, Guo Q (2025) An improved sparrow search algorithm optimized LightGBM approach for credit risk prediction of SMEs in supply chain finance. J Comput Appl Math 454:116197. https://doi.org/10.1016/j.cam.2024.116197
13.
Wang L, Jia F, Chen L, Xu Q (2023) Forecasting SMEs’ credit risk in supply chain finance with a sampling strategy based on machine learning techniques. Ann Oper Res 331:1–33. https://doi.org/10.1007/s10479-022-04518-5
14.
Kipf TN, Welling M (2017) Semi-Supervised Classification with Graph Convolutional Networks
15.
Li Y, Vinyals O, Dyer C et al (2018) Learning Deep Generative Models of Graphs
16.
Veličković P, Cucurull G, Casanova A et al (2018) Graph Attention Networks
17.
Yu B, Yin H, Zhu Z (2018) Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting. In: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. pp 3634–3640
18.
Guo Q, Zhuang F, Qin C et al (2022) A Survey on Knowledge Graph-Based Recommender Systems. IEEE Trans Knowl Data Eng 34:3549–3568. https://doi.org/10.1109/TKDE.2020.3028705
19.
Ji S, Pan S, Cambria E et al (2022) A Survey on Knowledge Graphs: Representation, Acquisition, and Applications. IEEE Trans Neural Networks Learn Syst 33:494–514. https://doi.org/10.1109/TNNLS.2021.3070843
20.
Kosasih EE, Margaroli F, Gelli S et al (2024) Towards knowledge graph reasoning for supply chain risk management using graph neural networks. Int J Prod Res 62:5596–5612. https://doi.org/10.1080/00207543.2022.2100841
21.
Geng S, Fu Z, Tan J et al (2022) Path Language Modeling over Knowledge Graphsfor Explainable Recommendation. In: Proceedings of the ACM Web Conference 2022. Association for Computing Machinery, New York, NY, USA, pp 946–955
22.
Tominac PA, Zhang W, Zavala VM (2022) Spatio-temporal economic properties of multi-product supply chains. Comput Chem Eng 159:107666. https://doi.org/10.1016/j.compchemeng.2022.107666
23.
Zhang C, Chen J, Shu T, Tan J (2022) Enterprise Event Risk Detection Based on Supply Chain Contagion. In: 2022 IEEE 9th International Conference on Data Science and Advanced Analytics (DSAA). pp 1–10
24.
Kim H, Choi J, Whang JJ (2023) Dynamic Relation-Attentive Graph Neural Networks for Fraud Detection. In: 2023 IEEE International Conference on Data Mining Workshops (ICDMW). pp 1092–1096
25.
Gong G, Huang X, Wu S et al (2021) Punishment by Securities Regulators, Corporate Social Responsibility and the Cost of Debt. J Bus Ethics 171:337–356. https://doi.org/10.1007/s10551-020-04438-z
26.
Chen Y, Wu Z (2023) Financial Fraud Detection of Listed Companies in China: A Machine Learning Approach. Sustainability 15:105. https://doi.org/10.3390/su15010105
27.
Xu H, Fan G, Song Y (2022) Novel Key Indicators Selection Method of Financial Fraud Prediction Model Based on Machine Learning Hybrid Mode. Mobile Information Systems 2022:6542652. https://doi.org/10.1155/2022/6542652
28.
Adler J, Parmryd I (2010) Quantifying colocalization by correlation: The Pearson correlation coefficient is superior to the Mander’s overlap coefficient. Cytometry Part A 77A:733–742. https://doi.org/10.1002/cyto.a.20896
29.
Gao Y, Zhang Z, Lin H et al (2022) Hypergraph Learning: Methods and Practices. IEEE Trans Pattern Anal Mach Intell 44:2548–2566. https://doi.org/10.1109/TPAMI.2020.3039374
30.
Feng Y, You H, Zhang Z et al (2019) Hypergraph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence 33:3558–3565. https://doi.org/10.1609/aaai.v33i01.33013558
31.
Misra D (2020) Mish: A Self Regularized. Non-Monotonic Activation Function
32.
Shang W, Sohn K, Almeida D, Lee H (2016) Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units. In: Proceedings of The 33rd International Conference on Machine Learning. PMLR, pp 2217–2225
33.
Applied Logistic Regression
34.
Ke G, Meng Q, Finley T et al (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems. Curran Associates, Inc
35.
Chen J, Hou H, Gao J et al (2019) RGCN: Recurrent Graph Convolutional Networks for Target-Dependent Sentiment Analysis. In: Douligeris C, Karagiannis D, Apostolou D (eds) Knowledge Science, Engineering and Management. Springer International Publishing, Cham, pp 667–675
36.
Ishiwatari T, Yasuda Y, Miyazaki T, Goto J (2020) Relation-aware Graph Attention Networks with Relational Position Encodings for Emotion Recognition in Conversations. In: Webber B, Cohn T, He Y, Liu Y (eds) Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online, pp 7360–7370
37.
Yang Y, Guan Z, Li J et al (2023) Interpretable and Efficient Heterogeneous Graph Convolutional Network. IEEE Trans Knowl Data Eng 35:1637–1650. https://doi.org/10.1109/TKDE.2021.3101356