Transform-based export enterprise risk prediction method and system

Abstract

The invention discloses an export enterprise risk prediction method and system based on Transform, relates to the field of data processing, is applied to an enterprise risk control system, and comprises the following steps: constructing a multi-dimensional index system, and dynamically distributing index weights by adopting a subjective and objective fused multi-target combination weighting model; designing a Transform risk prediction model, capturing dynamic association of risk indexes through a space-time attention mechanism, and training the risk prediction model in combination with historical risk evaluation index data; establishing a weight comparison correction mechanism, comparing a combined weighting result with a model prediction value, and optimizing weight distribution through error feedback; and establishing a risk early warning linkage mechanism of a regional level by combining a corrected weighting result, and triggering graded early warning when an enterprise risk value exceeds a threshold value, thereby realizing accurate prediction and collaborative prevention and control of export enterprise risk indexes.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a Transformer-based export enterprise risk prediction method and system.

[0002] summary

[0003] The present invention discloses a Transformer-based export enterprise risk prediction method and system, which relates to the field of data processing and is applied to enterprise risk control systems. The method comprises: constructing a multi-dimensional indicator system, and dynamically allocating indicator weights using a subjective-objective integrated multi-objective combined weighting model; designing a Transformer risk prediction model, capturing the dynamic correlation of risk indicators through a spatiotemporal attention mechanism, and training the risk prediction model in combination with historical risk evaluation indicator data; establishing a weight comparison and correction mechanism, comparing the combined weighting result with the model prediction value, and optimizing the weight distribution through error feedback; establishing a regional-level risk warning linkage mechanism in combination with the corrected weighted result, and triggering a graded warning when the enterprise risk value exceeds a threshold, thereby realizing accurate prediction and coordinated prevention and control of export enterprise risk indicators. Background Art

[0004] Against the backdrop of a complex and volatile global trade environment, the risks faced by exporting companies are multi-dimensional, dynamic, and non-linear. Traditional risk assessment methods often rely on single-dimensional financial analysis or static indicator systems, making it difficult to fully capture the complex risks faced by exporting companies that are highly time-varying. With the acceleration of global supply chain restructuring and the evolution of digital trade rules, there is an urgent need to build an intelligent risk assessment and prediction system that integrates multi-source heterogeneous data and has dynamic learning capabilities.

[0005] However, existing models (such as Logistic regression, decision trees, etc.) lack the ability to integrate time series correlation and cross-modal data, resulting in delayed risk warnings and high misjudgment rates. Summary of the Invention

[0006] The purpose of the present invention is to provide a Transformer-based export enterprise risk prediction method and system to solve the above-mentioned problems existing in traditional risk prediction methods. The method includes a multi-dimensional indicator construction module, a multi-objective combination weighting module, a Transformer risk prediction module, a weight comparison and correction module, and a risk warning module. The specific scheme is as follows:

[0007] Obtain risk data from multiple export companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system;

[0008] Construct a multi-objective combined weighted evaluation model, integrate subjective and objective weighting methods, and assign weights to each indicator of the multi-dimensional indicator system; based on the multi-objective combined weighted evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to output a comprehensive risk score S1;

[0009] Construct a Transformer risk prediction model, train the model based on weighted historical risk assessment indicator data, and output a risk probability value P1; input the target enterprise's real-time data into the model to obtain a real-time risk prediction value P2;

[0010] The comprehensive risk score S1 generated by the multi-objective combination weighted evaluation model is standardized and aligned with the Transformer risk prediction value P1 and then compared. The weights of each indicator of the multi-objective combination are adjusted through error feedback, and the adjusted weights are used to correct the comprehensive risk score S2. Combined with the weighted results of S2 and P2, a regional risk map is constructed, the transmission path is simulated, and a graded warning is triggered.

[0011] Optionally, the subjective weighting method and the objective weighting method specifically include:

[0012] The subjective weighting method adopts the analytic hierarchy process (AHP), in which experts compare the importance of indicators in pairs and generate a subjective weight vector W s , the subjective weight is obtained using the following formula:

[0013] Construct a judgment matrix, the formula is as follows:

[0014] A=(α ij ) n×n ;

[0015] Among them, α ij It represents the importance scale value of indicator i relative to j (1-9 scale method), and n means that n indicators are included;

[0016] Normalize the judgment matrix A by column, the formula is as follows:

[0017]

[0018] The weight vector is obtained by summing and normalizing the rows. The formula is as follows:

[0019]

[0020] After obtaining the initial weights of each indicator, the consistency index and consistency ratio are obtained using the following formula:

[0021]

[0022] Wherein, CI is a consistency index, CR is a consistency ratio, if CR < 0.1, it is passed, and λ max is the maximum eigenvalue of the judgment matrix, and RI is an average random consistency index.

[0023] The objective weighting method adopts a CRITIC entropy weight method to generate an objective weight vector W o , and the objective weight is obtained by using the following formula:

[0024] The index is standardized, the original data matrix X=(x ij ) m×n (m is a positive sample), and is normalized according to positive / negative indexes, and the formula is as follows:

[0025]

[0026] The information amount and conflict of the index are calculated, the data information amount of the index j is measured by a standard deviation, the conflict degree of the index j and other indexes is measured by a Pearson correlation coefficient, the comprehensive information amount C j of the index j is calculated by combining the information amount and the conflict, and the formula is as follows:

[0027]

[0028] Wherein, σ j is a standard deviation of the jth index, and r jk is a Pearson correlation coefficient.

[0029] The weight is distributed according to the proportion of the comprehensive information amount C j , and the formula is as follows:

[0030]

[0031] Optionally, the multi-objective combination weighting evaluation model determines the initial weight of each index, and specifically includes the following steps.

[0032] A multi-objective programming model is established, a linear weighting method is used to convert into a single objective optimization, and the initial weight W init is solved by using a Lagrange multiplier method, and the formula is as follows:

[0033]

[0034] Optionally, the most weight is generated by using a dynamic adjustment mechanism, and the comprehensive risk score S1 is output, and specifically includes the following steps.

[0035] The dynamic adjustment mechanism generates the most weight by using the initial weight and the historical weight weighted fusion, and the formula is as follows:

[0036] W final=λ(t)·W init +(1-λ(t))·W histroy ;

[0037] Where λ(t) is the time decay factor, which is dynamically adjusted with the rate of change of the external environment. Based on the final weight and the standardized value of the indicator, the enterprise risk score is calculated as follows:

[0038]

[0039] Optionally, the constructing of the Transformer risk prediction model specifically includes:

[0040] The multimodal encoding layer, spatiotemporal attention module and risk prediction layer take as input the historical risk assessment index data processed by multi-objective combination weighting, and output the risk probability value, where:

[0041] The multimodal encoding layer uses linear embedding for numerical indicators and BERT model pre-trained word vectors for textual policy data;

[0042] The spatiotemporal attention module includes a spatial attention layer to calculate the association weights between different indicators; a temporal attention layer to capture the long-term dependencies of historical sequences;

[0043] The risk prediction layer outputs the risk probability value through a fully connected network;

[0044] Optionally, a model is trained based on weighted historical risk assessment indicator data to output a risk probability value P1. The real-time data of the target enterprise is input into the model to obtain a real-time risk prediction value P2, specifically including:

[0045] The predicted value is output through average pooling and Sigmoid function. The formula is as follows:

[0046] P1=σ(W p AvgPool(MultiHead(X+P))+b p );

[0047] P2=σ(W p AvgPool(MultiHead(X input +P))+b p );

[0048] Where X is the weighted historical risk assessment indicator data, X input is the real-time data input after empowerment, W p is the weight matrix, P is the temporal position encoding matrix, MultiHead(·) represents the calculation of h = 8 attention heads, and Avgpool(·) reduces the dimension of features along the time dimension to extract the overall trend;

[0049] Optionally, the comprehensive risk score S1 generated by the multi-objective combined weighted evaluation model is standardized and aligned with the Transformer risk prediction value P1, specifically including:

[0050] The PlattScaling method is used to calibrate the probability output through logistic regression. The formula is as follows:

[0051]

[0052] P1 norm =P1∈[0,1];

[0053] Where a and b are calibration parameters, which are optimized by the validation set. (default ε=0.05), the scores are considered consistent, otherwise weight adjustment is triggered;

[0054] Optionally, adjusting the weights through error feedback to generate a revised comprehensive risk score S2 specifically includes:

[0055] Dynamically adjust the multi-objective combination weight W based on error feedback final , using the constrained gradient descent method:

[0056]

[0057] Where η is the learning rate (the default is η = 0.01), X is the normalized original indicator value vector, and the adjusted weight is used to calculate the modified score S2. The formula is as follows:

[0058]

[0059] Optionally, the weighted results of combining S2 and P2 are used to construct a regional risk map, simulate the transmission path, and trigger a graded warning, specifically including:

[0060] The enterprise-level revised score S2 is weighted and integrated with the real-time prediction value P2 to generate the enterprise comprehensive risk value R;

[0061]

[0062] Where a∈[0,1] is the weight coefficient (the default is a=0.6), which reflects the balance between expert experience and model prediction;

[0063] Based on the adjusted weighted results, a regional risk early warning linkage mechanism is established, with graded early warning triggers:

[0064] Yellow warning (0.7<R<0.8): notify the enterprise to conduct self-inspection;

[0065] Orange alert (0.8≤R≤0.9: notify regional regulatory authorities;

[0066] Red alert (R≥0.9): Initiate a cross-regional emergency response and freeze the export qualifications of high-risk enterprises.

[0067] The present invention provides a Transformer-based export enterprise risk prediction system, comprising:

[0068] Multi-dimensional indicator construction module. Obtain risk data from multiple export companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system;

[0069] Multi-objective combination weighting module. Build a multi-objective combination weighting evaluation model, integrating subjective and objective weighting methods to assign weights to each indicator in the multi-dimensional indicator system. Based on the multi-objective combination weighting evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to output a comprehensive risk score.

[0070] Transformer risk prediction module. Construct a Transformer risk prediction model, which includes a multimodal encoding layer, a spatiotemporal attention module, and a risk prediction layer. The model is trained based on weighted historical risk assessment indicator data to obtain a trained risk prediction model and output a risk prediction value. Furthermore, the target enterprise's real-time data is input into the risk prediction model to obtain a real-time risk prediction value.

[0071] Weight comparison and correction module: The comprehensive risk score generated by the multi-objective combination weighted evaluation model is compared with the Transformer risk prediction value after standardization and alignment. The weights of each indicator in the multi-objective combination are adjusted through error feedback to obtain a corrected comprehensive risk score.

[0072] Risk warning module: Combining the weighted results of the revised comprehensive risk score and the real-time risk prediction value, a regional risk warning linkage mechanism is established. When the risk level of an enterprise exceeds the warning threshold, an early warning message is sent to the relevant regulatory authorities and surrounding enterprises.

[0073] The beneficial effects of the present invention are:

[0074] The present invention provides a Transformer-based export enterprise risk prediction method and system. The method cleans raw data and extracts key enterprise features through preset rules to construct a multi-dimensional risk evaluation index system, solving the problems of single indicators and missing dimensions in traditional methods. The risk prediction variables designed based on the characteristics of export enterprises can improve the industry adaptability of the evaluation system; innovatively integrates the hierarchical analysis method and the entropy weight method, and realizes continuous optimization of weight distribution through a dynamic adjustment mechanism; utilizes the self-attention mechanism of the Transformer model to break through the long-range dependency limitations of the traditional model, and can effectively capture the temporal correlation characteristics of the risk evolution of export enterprises; through standardized alignment and error feedback mechanisms, realizes two-way verification of the empowered score and the model prediction value, and automatically generates a weight correction coefficient matrix, so that the system has the ability of continuous self-optimization, and provides an intelligent decision support tool for coping with the complex and changing international trade environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0076] Figure 1 This is a flowchart of a Transformer-based export enterprise risk prediction method;

[0077] Figure 2 This is a schematic diagram of the iterative process of a multi-objective combined weighting model for export enterprise risk prediction based on Transformer; DETAILED DESCRIPTION

[0078] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0079] S1: Obtain risk data from multiple exporting companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system;

[0080] For example, through the API interface, it can connect to the General Administration of Customs' trade data platform, corporate credit system, and international logistics tracking platform in real time to obtain original data such as financial statements, customs declarations, supply chain orders, exchange rate fluctuation records, and policy and regulatory changes of export companies; and use a distributed database to store enterprise-level time series data.

[0081] For example, in the missing value processing, the KNN interpolation method is used to fill the continuous indicators and the mode is used to fill the categorical indicators; dynamic thresholds are defined to automatically trigger data elimination or manual review processes;

[0082] Optionally, risk assessment indicators include financial dimension indicators, supply chain dimension indicators, and policy dimension indicators;

[0083] For example, financial dimension indicators may include debt-to-asset ratio and quick ratio; supply chain dimension indicators may include supplier concentration and logistics punctuality rate; and policy dimension indicators may include HS code tax refund change index and trade impact coefficient.

[0084] S2: Construct a multi-objective combined weighted evaluation model, integrating subjective and objective weighting methods to assign weights to each indicator in the multi-dimensional indicator system; based on the multi-objective combined weighted evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to generate a comprehensive risk score;

[0085] Optionally, experts can be invited to compare the indicators pairwise using the analytic hierarchy process (AHP) to construct a judgment matrix. After passing the consistency test (CR<0.1), the weights are calculated using the following formula:

[0086]

[0087] Optionally, the entropy weight method is used to calculate the information entropy of each indicator to obtain the objective weight vector. The calculation formula is as follows:

[0088]

[0089] Optionally, a dynamic adjustment mechanism is used to generate the final weights and output a comprehensive risk score. The calculation formula is as follows:

[0090] W final =λ(t)·W init +(1-λ(t))·W histroy ;

[0091]

[0092] For example, the initial combination weights α=0.6 (subjective) and β=0.4 (objective) are set, and the fusion coefficient is automatically adjusted through a sliding window mechanism. When it is detected that the exchange rate volatility exceeds the threshold of 5%, the objective weight ratio is triggered to increase to 55%; finally, a dynamic weight matrix is ​​generated, such as the supplier concentration weight is adjusted from 0.21 to 0.23.

[0093] S3: Construct a Transformer risk prediction model, train it based on the weighted historical risk assessment indicator data, obtain a trained risk prediction model, and output a risk prediction value; input the target enterprise's real-time data into the risk prediction model, and output a real-time risk prediction value;

[0094] For example, the numerical data in the multimodal encoding layer is mapped to the high-dimensional space through the fully connected layer, the text data uses BERT to extract semantic features, and the graph data generates node embeddings through GraphSAGE; the spatiotemporal attention module calculates the time dimension self-attention of the time series data to capture long-term dependencies, and constructs a spatial attention matrix based on the correlation between the enterprise's geographic coordinates and the supply chain; the risk prediction layer maps the attention output to the risk probability value through a two-layer fully connected network (the activation function is LeakyReLU).

[0095] S4: The comprehensive risk assessment result generated by the multi-objective combination weighted assessment model is compared with the Transformer risk prediction value after normalization and alignment, and the weights of each indicator of the multi-objective combination are adjusted through error feedback to obtain a revised comprehensive risk score;

[0096] Optionally, the Platt Scaling method is used to standardize the composite score and Transformer risk prediction value;

[0097] Optionally, error feedback dynamically adjusts the weights of each indicator in the multi-objective combination, and the gradient descent method is used to update the weights. The calculation formula is as follows:

[0098]

[0099] Where η is the learning rate (the default is η = 0.01), and X is the normalized original indicator value vector;

[0100] Optionally, use the adjusted weights to calculate a revised score, using the following formula:

[0101]

[0102] S5: Combine the weighted results of the revised comprehensive risk score and the real-time risk prediction value to establish a regional risk early warning linkage mechanism. When the risk level of a certain enterprise exceeds the early warning threshold, an early warning message will be sent to the relevant regulatory authorities and surrounding enterprises.

[0103] Optionally, the revised comprehensive risk score is weighted with the real-time risk prediction value to obtain the final result, which is calculated as follows:

[0104]

[0105] Where a∈[0,1] is the weight coefficient (the default is a=0.6), which reflects the balance between expert experience and model prediction;

[0106] For example, through the adjusted weighted results, a risk warning linkage mechanism is established at the regional level to trigger a graded warning. When the risk level of an enterprise exceeds the warning threshold, warning information is sent to relevant regulatory departments and surrounding enterprises. For example, a yellow warning (0.7<R<0.8) notifies the enterprise to conduct self-inspection; an orange warning (0.8≤R≤0.9) notifies the regional regulatory department; and a red warning (R≥0.9) initiates a cross-regional emergency response and freezes the export qualifications of high-risk enterprises.

[0107] The above is a neural network-based enterprise risk prediction method provided in one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding Transformer-based export enterprise risk prediction system, including:

[0108] Multi-dimensional indicator construction module. Obtain risk data from multiple export companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system;

[0109] Multi-objective combination weighting module. Build a multi-objective combination weighting evaluation model, integrating subjective and objective weighting methods to assign weights to each indicator in the multi-dimensional indicator system. Based on the multi-objective combination weighting evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to output a comprehensive risk score.

[0110] Transformer risk prediction module. Construct a Transformer risk prediction model, which includes a multimodal encoding layer, a spatiotemporal attention module, and a risk prediction layer. The model is trained based on weighted historical risk assessment indicator data to obtain a trained risk prediction model and output a risk prediction value. Furthermore, the target enterprise's real-time data is input into the risk prediction model to obtain a real-time risk prediction value.

[0111] Weight comparison and correction module: The comprehensive risk score generated by the multi-objective combination weighted evaluation model is compared with the Transformer risk prediction value after standardization and alignment. The weights of each indicator in the multi-objective combination are adjusted through error feedback to obtain a corrected comprehensive risk score.

[0112] Risk warning module: Combining the weighted results of the revised comprehensive risk score and the real-time risk prediction value, a regional risk warning linkage mechanism is established. When the risk level of an enterprise exceeds the warning threshold, an early warning message is sent to the relevant regulatory authorities and surrounding enterprises.

Claims

1. A Transformer-based export enterprise risk prediction method, characterized by: include: Obtain risk data from multiple export companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system; Construct a multi-objective combined weighted evaluation model, integrating subjective and objective weighting methods to assign weights to each indicator in the multi-dimensional indicator system; based on the multi-objective combined weighted evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to generate a comprehensive risk score; Construct a Transformer risk prediction model, train it based on the weighted historical risk assessment indicator data, obtain the trained risk prediction model, and output the risk prediction value; Inputting the real-time data of the target enterprise into the risk prediction model and outputting a real-time risk prediction value; The comprehensive risk assessment result generated by the multi-objective combination weighted evaluation model is compared with the Transformer risk prediction value after standardization and alignment, and the weights of each indicator of the multi-objective combination are adjusted through error feedback to obtain a revised comprehensive risk score; By combining the weighted results of the revised comprehensive risk score and the real-time risk prediction value, a regional risk warning linkage mechanism is established. When the risk level of an enterprise exceeds the warning threshold, warning information will be sent to relevant regulatory departments and surrounding enterprises.

2. The Transformer-based export enterprise risk prediction method according to claim 1, characterized in that: The subjective weighting method and the objective weighting method specifically include: The subjective weighting method adopts the analytic hierarchy process (AHP), in which experts compare the importance of indicators in pairs and generate subjective weight vectors. The subjective weights are obtained using the following formula: The objective weighting method adopts the CRITIC entropy weight method to generate an objective weight vector, and the objective weight is obtained using the following formula:

3. The Transformer-based export enterprise risk prediction method according to claim 1, characterized in that: The multi-objective combined weighted evaluation model specifically includes: A multi-objective programming model is established and converted into a single-objective optimization using the linear weighted method. The initial weight is solved using the Lagrange multiplier method. The formula is as follows:

4. The method for predicting export enterprise risk based on Transformer according to claim 1, characterized in that: The Transformer risk prediction model includes: The multimodal encoding layer, spatiotemporal attention module and risk prediction layer take as input the historical risk assessment index data processed by multi-objective combination weighting, and output the risk probability value, where: The multimodal encoding layer uses linear embedding for numerical indicators and BERT model pre-trained word vectors for textual policy data; The spatiotemporal attention module includes a spatial attention layer to calculate the association weights between different indicators; a temporal attention layer to capture the long-term dependencies of historical sequences; The risk prediction layer outputs the risk probability value through a fully connected network.

5. The Transformer-based export enterprise risk prediction method according to claim 1, characterized in that: The method of adjusting the weights of each indicator of the multi-objective combination through error feedback and updating the weights using the gradient descent method specifically includes: The following formula is used to adjust the weight of each indicator in the multi-objective combination: Where η is the learning rate (the default is η = 0.01), and X is the normalized original indicator value vector.

6. The Transformer-based export enterprise risk prediction method according to claim 1, characterized in that: The establishment of a regional risk early warning linkage mechanism specifically includes: Based on the adjusted weighted results, a regional risk early warning linkage mechanism is established, with graded early warning triggers: Yellow warning (0.7<R<0.8): notify the enterprise to conduct self-inspection; Orange alert (0.8≤R≤0.9: notify regional regulatory authorities; Red alert (R≥0.9): Initiate a cross-regional emergency response and freeze the export qualifications of high-risk enterprises.

7. A Transformer-based export enterprise risk prediction system, characterized by: include: Multi-dimensional indicator construction module. Obtain risk data from multiple export companies, clean the raw data based on preset rules, and build a multi-dimensional risk assessment indicator system; Multi-objective combination weighting module. Build a multi-objective combination weighting evaluation model, integrating subjective and objective weighting methods to assign weights to each indicator in the multi-dimensional indicator system. Based on the multi-objective combination weighting evaluation model, determine the initial weight of each indicator, and generate the final weight through a dynamic adjustment mechanism to output a comprehensive risk score. Transformer risk prediction module. Construct a Transformer risk prediction model, which includes a multimodal encoding layer, a spatiotemporal attention module, and a risk prediction layer. It is trained based on weighted historical risk assessment indicator data to obtain a trained risk prediction model and output a risk prediction value. In addition, the real-time data of the target enterprise is input into the risk prediction model to obtain a real-time risk prediction value; Weight comparison and correction module: The comprehensive risk score generated by the multi-objective combination weighted evaluation model is compared with the Transformer risk prediction value after standardization and alignment. The weights of each indicator in the multi-objective combination are adjusted through gradient descent feedback to obtain the corrected comprehensive risk score. Risk warning module: Combining the weighted results of the revised comprehensive risk score and the real-time risk prediction value, a regional risk warning linkage mechanism is established. When the risk level of an enterprise exceeds the warning threshold, an early warning message is sent to the relevant regulatory authorities and surrounding enterprises.

Images