A Novel Risk-Perception Model Based on Blockchain for Supply Chain Finance of China Real Estate

More than 220 enterprises in China's real estate industry have gone bankrupt


Introduction
In 2023, China's real estate industry attracted a wave of bankruptcies among real estate enterprises as several leading enterprises defaulted on their debts.For instance, Evergrande's debt reached 2.43 trillion CNY according to its 2022 financial report, while Yango Group accumulated a total debt of 274.6 billion CNY.On August 6, 2023, Yango Group was delisted from the Shenzhen Stock Exchange.The National Bureau of Statistics of China reported that the country's investment in property development fell by 8.5% yearon-year, while domestic lending dropped by 11.5% and the use of foreign capital decreased by 43%.According to the People's Court Announcement Network, as of October 15, 2023, more than 220 real estate-related enterprises in China had issued bankruptcy announcements.
The role of real estate finance becomes particularly critical to the healthy development of China's real estate industry, and risk regulation is a top priority.Real estate finance involves a variety of financial activities and services in the process of purchasing, developing, constructing, and investing in property, ensuring the availability of funds for real estate projects from planning to completion.Real Estate Supply Chain Finance (RE-SCF) is an important part of real estate finance, focusing on optimizing cash flow and financial transactions in the supply chain of real estate projects, and providing short-term financing to suppliers and contractors.Globalization has led to a more complex division of labor, resulting in a growing number of levels in the real estate supply chain and more intricate relationships between participants.At the same time, the high degree of correlation between Supply Chain Finance (SCF) information flow, cash flow, business flow, and logistics significantly facilitates the rapid propagation of SCF risk [11,22].Consequently, the scope and impact of losses caused by SCF risk have dramatically expanded.Investments in real estate development contributed 7,517.297 billion CNY in total value added to all industries.Real estate development investments also drove the employment of more than 115 million people.The wave of real estate debt defaults has undoubtedly hampered economic development and caused mass unemployment among the population.According to the Wind database, from the beginning of 2022, the growth rate of China's real estate investment declined rapidly, and the unemployment rate of migrant workers once exceeded 6%.What's more, according to cement ren.com,China's largest database of cement professionals, more than 19 regions had already reduced the price of cement by more than CNY 100 per tonne.The bankruptcy of real estate enterprises has seriously affected the government's land transfer fees and the revenue of the financial sector.Additionally, in the upstream, construction industry accounts payable have become difficult to cash, resulting in a significant number of layoffs or closures among small and medium-sized enterprises (SMEs).Since the first half of 2023, over 1,300 construction enterprises have declared bankruptcy.Meanwhile, downstream real estate operating industries such as hotels, resorts, and retail operations have been experiencing revenue pressures, leading to salary cuts and layoffs.
To enhance capital liquidity in both the upstream and downstream sectors of the real estate industry and to accelerate the progress of projects, SCF offers a flexible and effective financing solution for real estate enterprises.This approach can optimize fund flows, reduce risks, and improve the stability and efficiency of the industry.SCF risk is the biggest obstacle to the development of supply chain finance, and the core part of risk control is risk assessment [17].However, the current SCF risk assessment still has defects such as asymmetric information, inaccurate risk classification, and delayed risk warnings.Therefore, this study proposes a blockchain-based intelligent perception model for RE-SCF.This model enables comprehensive risk monitoring across the entire system, with risk levels further refined using a Graph Convolutional Neural Network (GCNN).This model can utilize the open and transparent nature of blockchain to achieve multi-level information flow, reducing the probability of risk arising from information silos.Furthermore, integrating blockchain with GCNN enables rapid and accurate risk levels assessment.This integration also quantifies the probability of enterprise risk, thereby providing regulatory authorities with a quantitative standard for controlling supply chain financial risks.

Review of SCF Risk Control
The purpose of SCF risk assessment in real estate is to identify and quantify the risk status of enterprises in the supply chain of real estate projects, thereby devising management strategies to safeguard the project's stability and profitability.In response to the real estate financial crisis, the United States, Greece, and Japan have notably adopted measures.These measures feature economic stimulation, government intervention, and financial regulatory actions.Artificial intelligence (AI) can conduct in-depth analysis of extensive historical transaction data, financial statements, and market dynamics, enabling more accurate predictions of potential risks.Consequently, AI, alongside expert assessment methods, has become a principal approach for banks and other financial institutions.Zhang et al. [25] used back propagation neural network to classify the credit risk of the automobile industry into three levels, which provides theoretical support for improving the profitability of banks.Sang [18] compared the back propagation neural network with a support vector machine (SVM) and verified that SVM is more suitable for the credit assessment of banks and other financial institutions in the case of small samples.To compensate for the defects of a single model, Zhu et al. [27] proposed a new integrated machine-learning method to construct an integrated model of random subspace-real AdaBoost to predict the credit risk of of multisource data.Under the framework of an integrated learning model, Lei et al. [8] constructed the chaotic grasshopper optimization algorithm to extract the financial features of enterprises through the complex data-preprocessing process, then used SVM to classify the data, and finally optimized it using the sticky mushroom algorithm to construct the SCF risk precautionary system.Table 1 presents a comparison of the models, accuracy, and sample sizes proposed by some scholars.

Blockchain-Based SCF Risk
The primary causes of SCF risk, such as repeated pledges and false warehouse receipts, are rooted in information asymmetry [6,8].Recognizing this, researchers have turned to blockchain technology.Its core architecture, which is based on Bitcoin, offers several key features.These include openness, transparency, immutability, decentralization, traceability, and shared maintenance.Together, these characteristics have been effectively utilized to mitigate SCF (Supply Chain Finance) risk [1,3,12].Dong et al. [5] contend that blockchain technology facilitates the sharing of information in supply chain transactions, thereby enabling wider adoption of diverse financing tools within the supply chain.Natanelov et al. [14] demonstrated that blockchain's smart contract can shorten the cash flow cycle in the beef chain between Australia and China, while also reducing operational risk in blockchain-based SCF.Caniato et al. [2] com- In light of the above shortcomings, based on both domestic and foreign research foundations, this study established a blockchain-based SCF perception model.Firstly, to improve the risk quantification model and subdivide the risk into more categories.Secondly, to expand the sample size of the study and improve the representativeness and accuracy of the model.Thirdly, to utilize blockchain technology to enable the SCF risk to be monitored, assessed, warned and to form a comprehensive risk management framework for RE-SCF.

Intelligent Perception Model Framework for RE-SCF Risk
Smart contract is a set of digitally defined promises, automatically executed by the system when predefined contract terms are met [16].The blockchain intelligent perception framework for RE-SCF risk established by integrating GCNN into blockchain smart contract is shown in Figure 1.
The framework of the intelligent perception model is explained below: the smart contract in the blockchain is used as the programming environment to establish an intelligent perception model integrating risk monitoring, assessment, and categorized early warnings.Risk monitoring determines the presence of risks in real estate supply chain enterprises.Risk assessment goes a step further by quantifying these risks.
The framework of the intelligent perception model for RE-SCF risk communication among SCF participants and diminishing the risk associated with SCF.Soni et al. [19] proposed a decision-making framework to help SMEs develop sustainable SCF using Industry 4.0 technology.In a similar vein, Ahram et al. [1] and O'Leary [15] asserted that blockchain-enabled supply chain networks have shown improvements in transparency and accountability.Kshetri [7] employed multiple case studies to showcase blockchain's impacts on reducing costs, enhancing speed, reliability, risk mitigation, and flexibility in the supply chain.
Scholars have quantified SCF risk through AI methods and used blockchain to address information asymmetry and reduce risk, but the following shortcomings still exist: 1 Risk must be classified into more categories to accurately reflect the real risk status of enterprises [26]. 2 Research on SMEs' credit risk is concentrated, and there are gaps in the research regarding other risk.3 The number of research samples is insufficient, and the data can be easily tampered with, leading to low credibility of the model.4 Most of the machine learning uses centralized contract terms are met [16].The blockchain intelligent perception framework for RE-SCF risk established by integrating GCNN into blockchain smart contract is shown in Figure 1.

Figure 1
The framework of the intelligent perception model for The framework of the intelligent perception model is explained below: the smart contract in the blockchain is used as the programming environment to establish an intelligent perception model integrating risk monitoring, assessment, and categorized early warnings.Risk monitoring determines the presence of risks in real estate supply chain enterprises.Risk assessment goes a Furthermore, categorized early warnings merge the risk levels of enterprises with their scale to issue tailored warnings.This approach not only strengthens the enterprises' implementation of risk management measures but also reduces the likelihood of SCF risk occurrences.
Smart contract in blockchain can reduce operational risk in RE-SCF [2,4], but they have problems such as a lack of intelligence and flexibility, while AI has defects such as high computational cost, low resource utilization, and code vulnerability.By integrating blockchain with GCNN, the integrity, security, and validity of data can be effectively enhanced by the blockchain's consensus mechanism, asymmetric encryption, and hash algorithm [1,20,22].This integration provides high-quality data sources and more distributed arithmetic power for AI, while AI can also add intelligent effects to smart contract, expanding and diversifying their functions and improving the blockchain's ability to process data [10].Details are shown in Table 2.
As depicted in Figure 2, the blockchain architecture is stratified into six distinct layers: data layer, network layer, contract layer, incentive layer, consensus layer, and application layer.Within the contract layer, we have embedded functionalities for risk monitoring, risk assessment, and categorized early warnings, allowing for the full utilization of the advantages presented by the blockchain-enabled neural network.Under the premise of ensuring security, the distributed structure of blockchain provides more distributed arithmetic power for AI, which reduces redundant costs.

Figure 2 Layer structure diagram of blockchain combined with
As depicted in Figure 2, the blockchain architecture is stratified into six distinct layers: data layer, network layer, contract layer, incentive layer, consensus layer, and application layer.Within the contract layer, we have embedded functionalities for risk monitoring, risk assessment, and categorized early warnings, allowing for the full utilization of the advantages presented by the blockchain-enabled neural network.

Establishment of Intelligent Perception Model for RE-SCF Risk
Based on the advantages of combining blockchain with AI as well as the perception framework proposed in   Based on the advantages of combining blockchain with AI as well as the perception framework proposed in Figure 1, the RE-SCF risk intelligent perception model was established as shown in Figure 3. Firstly, a risk assessment system tailored to the specific characteristics of China's real estate industry was established, integrating the industry's evaluation criteria into the enterprise's performance evaluation standards.Secondly, correlation analysis was used to remove the multicollinearity among the indicators.
Then, the multi-dimensional indicators were fused using principal component analysis.Finally, a blockchain and a graph convolutional neural network were employed to establish an intelligent perception model, as shown in Figure 3.The specific content of the established risk intelligent perception model is outlined in Sections 4.1, 4.2, and 4.3.

Figure 3
Flow chart of RE-SCF risk intelligent perception

. Construction of Blockchain
This study takes government regulators and real estate enterprises as blockchain nodes, establishing the consortium chain.The data from the risk assessment system were merged on a per-enterprise basis and subsequently uploaded into the consortium chain.
Uploading RE-SCF data can leverage the data depository, cross-validation, and chronological relationship among blockchain technology timestamps.These features ensure the authenticity and reliability of the data in terms of completeness, reasonableness, and cause-and-effect logic.This process provides a real source of data for the subsequent risk monitoring, assessment, and categorized early warnings [11].The addition of government nodes can provide real-time

Construction of Blockchain
This study takes government regulators and real estate enterprises as blockchain nodes, establishing the consortium chain.The data from the risk assessment system were merged on a per-enterprise basis and subsequently uploaded into the consortium chain.
Uploading RE-SCF data can leverage the data depository, cross-validation, and chronological relationship among blockchain technology timestamps.These features ensure the authenticity and reliability of the data in terms of completeness, reasonableness, and cause-and-effect logic.This process provides a real source of data for the subsequent risk monitoring, assessment, and categorized early warnings [11].The addition of government nodes can provide real-time supervision, timely grasp of industry dynamics, and macro-control.The smart contract of the blockchain provides a programming environment for subsequent risk monitoring, assessment, and categorized early warnings while avoiding operational risk.
1 Refer to the financial institution risk management framework and the real estate industry evaluation system in the enterprise performance evaluation standard values, as published by the State Administration for Market Regulation and the Standardization Administration of China.
2 Referring to the risk management factors identified by Ying et al. [23], the current status of the enterprise was measured using four secondary indicators: profitability, solvency, operating capacity, and growth capacity.The current status of the supply chain was measured using the supply chain operation status and financing status.
3 The real estate industry is characterized by high leverage and high turnover.Additionally, before its collapse, Evergrande aggressively issued commercial papers and repeatedly delayed payments to downstream agents.To address these issues, additional metrics have been incorporated into the RE-SCF risk assessment system.These include the gearing ratio, cash ratio, days payable outstanding and days sales outstanding.These additions aim to more accurately reflect the capital status of enterprises and their position within the supply chain.The inclusion of the business cycle also indirectly reflects whether the enterprise has problems such as unfinished buildings and slow construction.Combining the three factors mentioned above, the established risk assessment system is shown in Table 3.
The preprocessing of data in Table 3 is performed by deleting most of the missing indicators.For example, 'Year of transaction' indicator, which is missing data for most of the companies, was deleted.For a small number of indicators with missing values, the industry average was used instead.when the 'gross margin' indicator is missing in parts of the manufacturing industry, the industry's average gross margin value can be used to fill in the missing data.The processing of the remaining indicators was consistent with the above principle.papers and repeatedly delayed payments to downstream agents.To address these issues, additional metrics have been incorporated into the RE-SCF risk assessment system.These include the gearing ratio, cash ratio, days payable outstanding and days sales outstanding.These additions aim to more accurately reflect the capital status of enterprises and their position within the supply chain.The inclusion of the business cycle also indirectly reflects whether the enterprise has problems such as unfinished buildings and slow construction.
Combining the three factors mentioned above, the established risk assessment system is shown in Table 3. (1)

Third
Where 'ρ' denotes the Spearman correlation coefficient, 'd' represents the rank difference between x and y, and 'n' indicates the sample capacity.

Multidimensional Indicators Fusion for RE-SCF Risk
Multidimensional indicators fusion plays a key role in providing more comprehensive and accurate analysis, enhancing decision support.Especially in the field of RE-SCF with complex and changing environments, it can effectively identify, capture, and assess potential risk factors.Based on the results of the correlation analysis, the risk assessment indicators were deleted; the RE-SCF risk assessment indicators system was established.The indicators were fused using principal component analysis (PCA), and the fused indicator was used as the criteria for RE-SCF risk assessment.The specific steps are as follows: 1 To eliminate differences in dimensions and numerical ranges among various datasets while Where 'ρ' denotes the Spearman correlation coefficient, 'd' represents the rank difference between x and y, and 'n' indicates the sample capacity.

Multidimensional Indicators Fusion for RE-SCF Risk
Multidimensional indicators fusion plays a key role in providing more comprehensive and accurate analysis, enhancing decision support.Especially in the field of RE-SCF with complex and changing environments, it can effectively identify, capture, and assess potential risk factors.Based on the results of the correlation analysis, the risk assessment indicators were deleted; the RE-SCF risk assessment indicators system was established.The indicators were fused using principal component analysis (PCA), and the fused indicator was used as the criteria for RE-SCF risk assessment.The specific steps are as follows: 1 To eliminate differences in dimensions and numerical ranges among various datasets while mitigating issues like gradient vanishing and explosion, normalization was applied.The normalization formula is shown below: mitigating issues like gradient vanishing and explosion, normalization was applied.The normalization formula is shown below: where 'xi'represents the ith x and 'xi ' ' the value of the ith x after normalization. 2 PCA was performed on the data, according to the empirical principle given by Kaiser [6].When the KMO value is > 0.5 and the sig value is < 0.01, it is suitable to utilize PCA to fuse the multidimensional data and extract the key components.

RE-SCF Risk Assessm
GCNN is used to uncover data pa indicators system.Then a risk a built, which rates the enterprise data is input into the model, G assessment on the enterprise an different labels to represent var enterprise.The process of build model is shown below: (1) Parameter Settings The parameters to be used in the risk assessment section and their meaning are shown in Tabl where 'x i 'represents the ith x and 'x i ' ' the value of the ith x after normalization.

RE-SCF Risk Monitoring
RE-SCF risk monitoring used the smart contract in the blockchain as the technical support and determines risk thresholds for the fusion of the indicators obtained.Establishing these risk thresholds helps identify potential risks within enterprises.Should any risks be identified, a risk assessment is then conducted to further quantify the risk levels faced by real estate enterprises.

RE-SCF Risk Assessment
GCNN is used to uncover data patterns based on the risk indicators system.Then a risk assessment model was built, which rates the enterprise risk level.When the data is input into the model, GCNN performs a risk assessment on the enterprise and uses the output of different labels to represent various risk levels of the enterprise.The process of building a risk assessment model is shown below: 1 Parameter Settings The parameters to be used in the risk assessment section and their meaning are shown in Table 4.
Adjacency matrix: matrix with edge links model is shown below: (1) Parameter Settings The parameters to be used in the risk assessment section and their meaning are shown in Table 4.
Enterprises were represented as nodes, enterprisetrading relationships as edges, and indicators within the RE-SCF risk assessment system as feature data.The ratio of the training set was selected based on when the model achieved the highest accuracy rate. (

3) Selection of activation function
The common activation functions were averaged three times, and the best activation function was selected based on the accuracy of the activation function.

Degree matrix:
diagonal matrix formed by node self-loop Data Segmentation Enterprises were represented as nodes, enterprise-trading relationships as edges, and indicators within the RE-SCF risk assessment system as feature data.The ratio of the training set was selected based on when the model achieved the highest accuracy rate.
3 Selection of activation function The common activation functions were averaged three times, and the best activation function was selected based on the accuracy of the activation function.

Convolution operation
Based on the algorithm and training time considerations, this study uses three convolutional layers.Through the nodes and edges to establish graph data, as well as node feature data and graph data into the convolution layers for convolution operations.The general operations process is shown in Figure 4.  (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
To prevent the m is necessary to i convolution ope of the number o the number of i were updated accuracy of the After the above obtained, and t were compared error (MSE), r mean absolute verify the perf relevant explana as follows: The the true value, is the amount of

RE-SCF ings
In SCF, enterp "bonding" effec spread.SCF r especially the ri likely to be trans chain, and losse of the core ente overall risk of micro-and sma Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.
Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.
of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.
(7) Comparative study of AI models After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency. (3)

Cross-entropy formula
5 Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.(3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using  (3) Cross-entropy formula The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency. ( Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.
is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using  (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.

Table 5
Learning rate setting

Learning rate Advantages Disadvantages
Large It can speed up the convergence of weights, which is conducive to improving the accuracy.
Accuracy is not stable.

Small
The speed of weight convergence is more stable, and the accuracy fluctuates less.

The accuracy is not High enough.
A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.
is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency.Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.

Table 5
Learning rate setting

Learning rate Advantages Disadvantages
Large It can speed up the convergence of weights, which is conducive to improving the accuracy.
Accuracy is not stable.

Small
The speed of weight convergence is more stable, and the accuracy fluctuates less.

The accuracy is not High enough.
A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.
To prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using  (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.

Table 5
Learning rate setting

Learning rate Advantages Disadvantages
Large It can speed up the convergence of weights, which is conducive to improving the accuracy.
Accuracy is not stable.

Small
The speed of weight convergence is more stable, and the accuracy fluctuates less.

The accuracy is not High enough.
A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.
To prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using  Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.
A learning rate that is either too large or too small can hinder to the improvement of the model's accuracy.Therefore, the learning rate at the highest accuracy and the lowest loss rate of the model was selected for the Adam optimizer.

Small
The speed of weight convergence is more stable, and the accuracy fluctuates less.
The accuracy is not High enough.

Perform iterations
To prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.(3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.  10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, Calling up the training set, Equation ( 3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.A learning rate that is either too large or too small can (6) Perform iterations To prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across Calling up the training set, Equation ( 3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 (3) Cross-entropy formula (5) Back propagation The model was initialized with a gradient before the next iteration.To prevent the previous gradient from affecting the current gradient, the gradient operation must be performed.The weights were updated by back propagation using the Adam optimizer so that the weight matrix in the convolutional network was updated in real time to further improve the accuracy.
Adam's process formula The different learning rate of the optimizer determined the magnitude of updating each weight in the gradient direction, as shown in Table 5.

Table 5
Learning rate setting

Learning rate Advantages Disadvantages
Large It can speed up the convergence of weights, which is conducive to improving the accuracy.
Accuracy is not stable.

Small
The speed of weight convergence is more stable, and the accuracy fluctuates less.
The accuracy is not High enough.

(6) Perform iterations
To prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.

(7) Comparative study of AI models
After the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The  � in Equations ( 10)-( 12) represents the true value,  � � represents the predicted value, and n is the amount of data.

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and microenterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across

RE-SCF Risk Categorized Early Warnings
In SCF, enterprises are closely connected, and the "bonding" effect of finance makes it easier for risk to spread.SCF risk will be magnified exponentially, especially the risk of core enterprises, which are more likely to be transmitted to other enterprises in the supply chain, and losses will increase exponentially.The risk of the core enterprises has a significant impact on the overall risk of SCF, while the risk of the marginal micro-and small enterprises has little impact on it.
After predicting the enterprise risk using a GCNN risk assessment model, the model acted accordingly.If the core enterprise risk result is standard (with an output label of 1) or high-risk (with an output label of 0), it will automatically return the risk assessment result, provide an early warning and remind the enterprise to take measures in time.Similarly, if the risk results of medium-sized enterprises, small enterprises, and micro enterprises are high-risk, they will also receive warnings.

RE-SCF Case Study
This section analyzes the real estate supply chain, using data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood process-ing, financial industry, and warehousing and transportation agency.

Correlation Analysis
The data were preprocessed, and since it does not satisfy normal distribution, the Spearman correlation test was performed on the third-level indicators within the same second-level indicator.The results are presented in Table 6.Profit size, net interest rate, gross margin, and operating income are strongly correlated or have a highly significant correlation.To prevent pseudo-correlation, a partial correlation analysis was conducted, the results are shown in Table 7.
Table 7 reveals that the net interest rate is weakly correlated with operating income by ignoring the effect of profit scale and gross margin.This suggests net interest rate is pseudo-correlated with operating income.Consequently, net interest rate and operating income were retained, and profit scale and gross margin were deleted.The assessment indicators system obtained after the correlation analysis is shown in Table 8.

Principal Component Analysis Fusion Indicators
1 Using formula (2), the data for the aforementioned indicators were normalized.The data were subjected to a PCA.From Table 9, it can be seen that KMO value = 0.6 > 0.5, and sig value = 0.000 < 0.01.According to the empirical principle of Kaiser [6], it is appropriate to apply PCA for the fusion of multidimensional indicators, facilitating the extraction of the primary component.
3 As can be seen in Figure 5, the cumulative contribution of the first principal component extracted by PCA has the highest eigenvalue with 93.6% information retention.Using the multi-dimensional data in the risk assessment indicator system was fused, and the first principal component  � obtained is given by the following step.

Figure 5
Gravel diagram 4) From table 10, it can be concluded that:

r RE-
Using the above data as a basis for merging by enterprise, the enterprise and the government were used as nodes to establish a consortium chain.Part of the block data of the established consortium chain are  mensional data in the risk assessment indicator system was fused, and the first principal component obtained is given by the following step.
4 From table 10, it can be concluded that: = 0.000 < 0.01.According to the empirical principle of Kaiser [6], it is appropriate to apply PCA for the fusion of multidimensional indicators, facilitating the extraction of the primary component.3) As can be seen in Figure 5, the cumulative contribution of the first principal component extracted by PCA has the highest eigenvalue with 93.6% information retention.Using Table 10, the multi-4) From table 10, it can be concluded that:

Construction of Blockchain
Using the above data as a basis for merging by enterprise, the enterprise and the government were used as nodes to establish a consortium chain.Part of the block data of the established consortium chain are

Construction of Blockchain
Using the above data as a basis for merging by enterprise, the enterprise and the government were used as nodes to establish a consortium chain.Part of the block data of the established consortium chain are shown below.Each block generally comprises two parts, the header and the body.The header includes the version, nonce, Merkle root, timestamp, hash value of the previous block, hash value of the current block, among other details.Some of the block data are shown in Table 11.The 'index' identifies the block's position in the blockchain; 'timestamp' records the time when the block data was written; 'nonce' is a random number, in mining to search for a nonce value that satisfies the condition; and 'hash' is a fixed-length output gen- erated using the SHA-256 function.'Previous Hash' denotes the hash value of the previous block.

RE-SCF Risk Monitoring
This study focuses on the RE-SCF risk of enterprises as the research objective using the blockchain's smart contract to delineate the risk threshold for F 1 .This approach facilitates the dynamic monitoring of enterprise risk.Based on the "Standard Values for Enterprise Performance Evaluation" issued by the State-owned Assets Supervision and Administration Commission of the State Council, the specific delineation is shown in Figure 6.

RE-SCF Risk Assessment
Taking the real estate supply chain as a sample, GCNN was employed to establish a risk assessment model.When the data is input, GCNN conducts a risk assessment of the enterprise.The risk assessment model-building processes are illustrated below.
1 Data Segmentation As depicted in Figure 7, the model has the highest accuracy when the training set share is 75%, so the number of training and test sets are selected as in Table 12.

RE-SCF Risk Monitoring
This study focuses on the RE-SCF risk of enterprises as the research objective using the blockchain's smart contract to delineate the risk threshold for  � .This approach facilitates the dynamic monitoring of enterprise risk.Based on the "Standard Values for Enterprise Performance Evaluation" issued by the Stateowned Assets Supervision and Administration Commission of the State Council, the specific delineation is shown in Figure 6.

RE-SCF Risk Assessment
Taking the real estate supply chain as a sample, GCNN was employed to establish a risk assessment model.When the data is input, GCNN conducts a risk assessment of the enterprise.The risk assessment model-building processes are illustrated below. (

1) Data Segmentation
As depicted in Figure 7, the model has the highest accuracy when the training set share is 75%, so the number of training and test sets are selected as in Table 12.

Figure 7
Accuracy with d

RE-SCF Risk Monitoring
This study focuses on the RE-SCF risk of enterprises as the research objective using the blockchain's smart contract to delineate the risk threshold for  � .This approach facilitates the dynamic monitoring of enterprise risk.Based on the "Standard Values for Enterprise Performance Evaluation" issued by the Stateowned Assets Supervision and Administration Commission of the State Council, the specific delineation is shown in Figure 6.

RE-SCF Risk Assessment
Accuracy with different percentages of the training set  The common activation functions were averaged three times, as shown in Figure 8. Elu has the highest accuracy, so it was chosen as the activation function of the model.(3) Back Propagation After each output result was obtained, the weights were updated through backpropagation with Adam's optimizer.As shown in Figure 9, the accuracy is plotted on the left vertical axis and the loss rate is plotted on the right vertical axis.The learning rate of 0.01 yields the highest accuracy and the smallest loss rate, so the learning rate of Adam was taken as 0.01.

Figure 9
Accuracy and loss rate for different learning rates (4) Perform Iterations As shown in Figure 10, after 1500 iterations, the accuracy of the model is stable at about 94%, which also represent the shortest time to achieve this accuracy.Therefore, the number of iterations of the model was selected as 1500.In comparison with MLP and SVM, and the comparison results are shown in Figure 11.Here, the others, and the error is lower, which is a better performance.

Figure 10
Accuracy at different numbers of iterations Comparison results of different models

Categorized Early Warnings of RE-SCF Risk
The results of the partial categorization of 398 pieces of data are shown in    (3) Back Propagation After each output result was obtained, the weights were updated through backpropagation with Adam's optimizer.As shown in Figure 9, the accuracy is plotted on the left vertical axis and the loss rate is plotted on the right vertical axis.The learning rate of 0.01 yields the highest accuracy and the smallest loss rate, so the learning rate of Adam was taken as 0.01.

Figure 9
Accuracy and loss rate for different learning rates (4) Perform Iterations As shown in Figure 10, after 1500 iterations, the accuracy of the model is stable at about 94%, which also represent the shortest time to achieve this accuracy.Therefore, the number of iterations of the model was selected as 1500.In comparison with MLP and SVM, the others, and the error is lower, which is a better performance.

Figure 10
Accuracy at different numbers of iterations Comparison results of different models

Categorized Early Warnings of RE-SCF Risk
The results of the partial categorization of 398 pieces of data are shown in Table 13, from which it can be seen that Pudong Construction, Feiliks, Yongan Forestry, and core high-risk, core standard, and other high-risk enterprises were warned.A combination of classified early warnings and smart contract solves the problem of untimely risk warnings.Timely warnings to high-risk   After each output result was obtained, the weights were updated through backpropagation with Adam's optimizer.As shown in Figure 9, the accuracy is plotted on the left vertical axis and the loss rate is plotted on the right vertical axis.The learning rate of 0.01 yields the highest accuracy and the smallest loss rate, so the learning rate of Adam was taken as 0.01.

Figure 9
Accuracy and loss rate for different learning rates (4) Perform Iterations As shown in Figure 10, after 1500 iterations, the accuracy of the model is stable at about 94%, which also the others, and the error is lower, which is a better performance.

Figure 10
Accuracy at different numbers of iterations

Figure 11
Comparison results of different models

Categorized Early Warnings of RE-SCF Risk
The results of the partial categorization of 398 pieces of data are shown in Table 13, from which it can be seen that Pudong Construction, Feiliks, Yongan Forestry, and core high-risk, core standard, and other high-risk 3 Back Propagation After each output result was obtained, the weights were updated through backpropagation with Adam's optimizer.As shown in Figure 9, the accuracy is plotted on the left vertical axis and the loss rate is plotted on the right vertical axis.The learning rate of 0.01 yields the highest accuracy and the smallest loss rate, so the learning rate of Adam was taken as 0.01.

Perform Iterations
As shown in Figure 10, after 1500 iterations, the accuracy of the model is stable at about 94%, which also represent the shortest time to achieve this accuracy.Therefore, the number of iterations of the model was selected as 1500.In comparison with MLP and SVM, and the comparison results are shown in Figure 11.Here, the accuracy is plotted on the left vertical axis, and MAE, MSE, and RMSE are plotted on the right vertical axis.The accuracy of the GCNN model is higher than the others, and the error is lower, which is a better performance.After each output result was obtained, the weights were updated through backpropagation with Adam's optimizer.As shown in Figure 9, the accuracy is plotted on the left vertical axis and the loss rate is plotted on the right vertical axis.The learning rate of 0.01 yields the highest accuracy and the smallest loss rate, so the learning rate of Adam was taken as 0.01.

Figure 9
Accuracy and loss rate for different learning rates (4) Perform Iterations As shown in Figure 10, after 1500 iterations, the accuracy of the model is stable at about 94%, which also the others, and the error is lower, which is a better performance.

Figure 10
Accuracy at different numbers of iterations

Figure 11
Comparison results of different models

Categorized Early Warnings of RE-SCF Risk
The results of the partial categorization of 398 pieces of data are shown in Table 13, from which it can be seen that Pudong Construction, Feiliks, Yongan Forestry, and core high-risk, core standard, and other high-risk

Categorized Early Warnings of RE-SCF Risk
The results of the partial categorization of 398 pieces of data are shown in Table 13, from which it can be seen that Pudong Construction, Feiliks, Yongan Forestry, and core high-risk, core standard, and other high-risk enterprises were warned.A combination of classified early warnings and smart contract solves the problem of untimely risk warnings.Timely warnings to high-risk enterprises significantly reduce the risk in RE-SCF, thereby enhancing the real estate finance ecosystem's operation.

Conclusion
A significant number of bankruptcies have emerged in China's real estate industry.Traditional risk assessment exhibits flaws such as imprecise risk categorization, slow assessment speed, delayed warnings, and data vulnerability to tampering.Against this backdrop, this study utilizes smart contract in blockchain and GCNN to propose an intelligent risk-perception model that integrates risk monitoring, assessment, and categorized warnings.A comparison of GCNN with MLP and SVM reveals that GCNN's accuracy reaches 94%, while MLP is at 77%, and SVM at 67%, indicating GCNN's superiority over other models.
The model established in this paper has the following advantages: 1 Utilizing the blockchain's transparent nature, it facilitates multi-level information flow, thus reducing the probability of information silo.Additionally, the immutable characteristic of blockchain overcomes the traditional risk assessment model's vulnerability to data tampering and low credibility.
2 By expanding the research sample size and subdi-viding risk into more categories, the model's representativeness and accuracy are enhanced.
3 The GCNN technology in blockchain smart contract provides faster assessments than expert-judgment evaluations.It also offers precise risk levels, quantifying the probability of enterprises being classified as high-risk, thus providing a quantitative basis for financial regulation.
4 Smart contract automatically issues alerts based on risk assessment results, providing a quicker warning mechanism compared to traditional methods that rely on monitoring repayment behaviors and market fluctuations.This feature enables a swifter response to potential risk.

Figure 2 Layer
Figure 2 Layer structure diagram of blockchain combined with mobile phones laptops Mobile phones, computers, websites Financial service Application layer

Figure 1
Figure 1, the RE-SCF was established as s assessment system characteristics of C established, integratin into the enterprise's p Secondly, correlation multicollinearity amo dimensional indicato component analysis.convolutional neura establish an intellige Figure 3.The specif intelligent perception 4.2, and 4.3.

Figure 3 Flow
Figure 3 Flow chart of RE-SC

Figure 1 ,
Figure 1, the RE-SCF risk intelligent perception model was established as shown in Figure 3. Firstly, a risk assessment system tailored to the specific characteristics of China's real estate industry was established, integrating the industry's evaluation criteria into the enterprise's performance evaluation standards.Secondly, correlation analysis was used to remove the multicollinearity among the indicators.Then, the multidimensional indicators were fused using principal component analysis.Finally, a blockchain and a graph convolutional neural network were employed to establish an intelligent perception model, as shown in Figure 3.The specific content of the established risk intelligent perception model is outlined in Sections 4.1, 4.2, and 4.3.

Figure 3
Figure 3 Flow chart of RE-SCF risk intelligent perception

( 4 )
Convolution operationBased on the algorithm and training time considerations, this study uses three convolutional layers.Through the nodes and edges to establish graph data, as well as node feature data and graph data into the convolution layers for convolution operations.The general operations process is shown in Figure4.

Figure 4
Figure 4 Convolutional algorithm data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency.

( 4 )
Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency.

( 6 )
Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency.

( 8 )
Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.E= -∑      �1 data from the Wind database spanning the last three years.It covers a total of 1,203 data entries across various sectors, including the real estate industry, construction manufacturing, wood processing, financial industry, and warehousing and transportation agency.

Figure 5
Figure 5 Gravel diagram

Figure 6 threshold Figure 7
Figure 6Risk threshold

Figure 8 Accuracy with different activation functions Figure 8
Figure 8 Accuracy with different activation functions Figure 8 Accuracy with different activation functions

Figure 9
Figure 9Accuracy and loss rate for different learning rates

Figure 8
Figure 8Accuracy with different activation functions

Figure 10
Figure 10Accuracy at different numbers of iterations

Figure 8
Figure 8Accuracy with different activation functions

Figure 11
Figure 11Comparison results of different models

Figure 8
Figure 8Accuracy with different activation functions

Table 2
Blockchain-enabled AI 1 Decentralized distributed structure.2 Shared computing resource environment.1 Traditional Centralized Computing Cost is Too High. 2 Low resource utilization rate, easy-to-invade code vulnerability.

Table 3
RE-SCF risk assessment indicators system

Table 4
Parameter-setting

Table 4
Parameter-setting

Table 4
Parameter-setting t Exponential Moving Average of the Gradient at t Time, with m 0 = 0 β 1 , β 2 The Exponential Decay Rate, A Default Value of 0.9 for β 1 , Default Value of 0.999 for β 2 v t Exponential Moving Average of the Squared Gradient at Time t v 0 = 0 F 1 Results of The Principal Component Analysis

Table 5
Learning rate setting Calling up the training set, Equation (3) was used to calculate the outputs and labels of the training set.These calculated labels were then used to compare the labels of the test set and original data.his comparison aimed to calculate the space occupied by test set labels to obtain the outputs.

Table 5
Learning rate setting

Table 5
Learning rate setting

Table 5
Learning rate setting

Table 5
Learning rate setting

7
Comparative study of AI modelsAfter the above steps, the accuracy of the model was obtained, and the results of GCNN, MLP, and SVM were compared in terms of accuracy, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).This comparison was to verify the performance of the GCNN model.The relevant explanations about MSE, RMSE, and MAE are as follows: The y i in Equations (10)-(12) represents the true value, ŷi represents the predicted value, and n is the amount of data.

Table 5
Learning rate setting Perform iterationsTo prevent the model from overfitting or underfitting, it is necessary to iterate the back propagation step and the convolution operation step.According to the reflection of the number of iterations in the accuracy of the model, the number of iterations was selected, and the weights were updated after each iteration to improve the accuracy of the model.
(6)earning rate that is either too large or too small can(6)

Table 5
Learning rate setting

Table 6
Correlation analysis

Table 7
Biased correlation analysis

Table 8
Supply chain finance risk assessment system

Table 9
Results of KMO and Bartlett's test

Table 10
Matrix of component score coefficients

Table 9
Results of KMO and Bartlett's test

Table 10
Matrix of component score coefficients

Table 11
Part of the block

Table 11
Part of the block

Table 12
Percentage of tra Each block generally comprises two parts, the header and the body.The header includes the version, nonce, Merkle root, timestamp, hash value of the previous block, hash value of the current block, among other details.Some of the block data are shown in Table11.The 'index' identifies the block's position in the blockchain; 'timestamp' records the time when the block data was written; 'nonce' is a random number, in mining to search for a nonce value that satisfies the condition; and 'hash' is a fixed-length output generated using the SHA-256 function.'Previous Hash' denotes the hash value of the previous block.

Table 11
Part of the block

Table 12 Percentage
of training set and test set

Table 12
Percentage of training set and test set 2 Selection of Activation Functions

Table 13
, from which it can be seen that Pudong Construction, Feiliks, Yongan Forestry, and core high-risk, core standard, and other high-risk enterprises were warned.A combination of classified early warnings and smart contract solves the problem of untimely risk warnings.Timely warnings to high-risk enterprises significantly reduce the risk in RE-SCF,