Enterprise credit score intelligent evaluation method based on traffic travel consumption data

Abstract

The application discloses an enterprise credit score intelligent evaluation method based on traffic travel consumption data, relates to the technical field of enterprise credit intelligent evaluation, and comprises the following steps: extracting multi-dimensional credit features from original traffic consumption records and constructing an enhanced behavior portrait, constructing a dynamic enterprise credit evaluation graph network, performing relationship reasoning and credit state diffusion calculation by injecting a feature vector, quantifying and conducting the risk of associated enterprises, performing time sequence evolution evaluation of the credit level based on the portrait, introducing an external risk label for constraint correction after generating a preliminary trajectory, starting backtracking verification by using historical score data flow, and finally outputting the credit score after consistency test. The method overcomes the limitations of isolated evaluation, can systematically identify associated risks, and generates a stable and logically self-consistent credit evolution result in the time dimension.

Description

Technical Field

[0001] This invention belongs to the field of intelligent enterprise credit assessment technology, specifically an intelligent assessment method for enterprise credit scoring based on transportation and travel consumption data. Background Technology

[0002] Current corporate credit assessment technologies primarily rely on traditional static data such as financial statements, business registration records, and legal disputes, combined with pre-defined scoring cards or statistical models for calculation. These methods typically treat each company as an independent assessment unit, analyzing it in isolation using its own characteristic indicators. Existing technologies struggle to effectively capture the complex networks of relationships between companies formed through supply chains, regional clusters, and business collaborations, and cannot quantify the potential impact of related-party risk transmission on corporate creditworthiness, resulting in insufficient ability to identify systemic risks.

[0003] Existing credit scoring models mostly output static scores at specific points in time or use simple moving average methods to process time series. These methods lack the ability to model the continuous and smooth evolution of credit status, and the scoring results are easily affected by fluctuations in single-period data, leading to abrupt changes. Furthermore, existing solutions lack a closed-loop mechanism for automatically and logically verifying current assessment results against historical scoring sequences, making it difficult to guarantee consistency of scores over time. This may result in assessment results that contradict historical trends or defy common business sense, affecting the reliability and interpretability of the model. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] To this end, this invention proposes an intelligent assessment method for enterprise credit scoring based on transportation and travel consumption data, including:

[0006] Obtain the original collection of traffic consumption records of the enterprise entity;

[0007] The original set of traffic consumption records is parsed at multiple levels to generate a set of structured consumption behavior fragments;

[0008] Based on the structured set of consumer behavior fragments, cross-data source enterprise attribute association mapping is performed to form an enhanced enterprise behavior profile;

[0009] Based on the enhanced corporate behavior profile, credit-related features are deeply extracted and fused to generate a multi-dimensional credit feature vector.

[0010] A dynamic enterprise credit assessment graph network is constructed, and the multidimensional credit feature vector is injected into the dynamic enterprise credit assessment graph network to perform relational reasoning and credit status diffusion calculation.

[0011] Based on the results of credit status diffusion calculation, a time-series evolution assessment of corporate credit level is performed to generate a preliminary credit trajectory.

[0012] External compliance and risk labels are introduced to constrain and correct the preliminary credit trajectory, and a benchmark credit score is output.

[0013] A scoring backtracking verification mechanism is initiated, using historical scoring data streams to perform consistency checks on the benchmark credit score, and generating the final corporate credit score accordingly.

[0014] Furthermore, the multi-level parsing of the original traffic consumption record set includes:

[0015] The first level of analysis is performed to separate the consumption time series, consumption geographical location series, and consumption amount series from the original set of traffic consumption records; the second level of analysis is performed to identify periodic consumption patterns and sudden consumption clusters for the consumption time series, analyze the permanent residence area and flow path map for the consumption geographical location series, and divide the consumption amount series into stable intervals and abnormal fluctuation intervals.

[0016] The third level of analysis is performed to align and associate the periodic consumption patterns, sudden consumption clusters, resident areas, flow path maps, stable intervals and abnormal fluctuation intervals in time and space, and synthesize the set of structured consumption behavior fragments.

[0017] Furthermore, the execution of cross-data source enterprise attribute association mapping includes:

[0018] The enterprise identification information in the structured consumer behavior fragment set is matched with the business registration database to obtain the enterprise's legal industry classification and registered capital information;

[0019] At the same time, the consumption geographic location sequence in the structured consumption behavior fragment set is associated with the regional economic indicator database to map the economic activity parameters of the enterprise's operating area;

[0020] Semantically correlate the aforementioned cyclical consumption patterns with publicly available news data on the company's business activities to infer the company's business activity cycle;

[0021] By integrating the aforementioned legal industry classifications, registered capital information, economic activity parameters, and business activity cycles, the enhanced corporate behavior profile is formed.

[0022] Furthermore, the deep extraction and fusion of credit-related features includes:

[0023] Consumer stability features are extracted from the enhanced corporate behavior profile, and these features are quantified by the variance of the consumption amount sequence and the regularity of the consumption time interval.

[0024] Extract consumption scale adaptability characteristics, which reflect the degree of matching between consumption level and enterprise registered capital and the average level of the industry to which it belongs;

[0025] Geographic risk association features are extracted, which are calculated based on the regional economic activity parameters mapped by the consumption geographic location sequence and the frequency of known credit risk events in the enterprise's operating area.

[0026] Extract behavioral consistency features, which assess the degree of agreement between the inferred business activity cycle and the actual consumption behavior cycle;

[0027] The feature interleaving and fusion algorithm is used to nonlinearly combine and reduce the consumption stability feature, consumption scale adaptability feature, geographical risk correlation feature and behavioral consistency feature to generate the multidimensional credit feature vector.

[0028] Furthermore, the construction of the dynamic enterprise credit assessment graph network includes:

[0029] Initialize a static enterprise relationship graph by using enterprises as nodes in the graph network and supply chain relationships, equity relationships, and geographical proximity relationships between enterprises as edges in the graph network.

[0030] The multidimensional credit feature vector is injected into the static enterprise association graph as the initial attribute of the corresponding enterprise node;

[0031] Design a credit state propagation rule, wherein the credit state propagation rule defines the credit influence weight between nodes based on the type and strength of the edge, and the similarity of the target node to the credit features of the source node;

[0032] The credit status propagation rule is iteratively run on the static enterprise association graph to realize the diffusion and rebalancing of credit features along the graph network edges, and to complete the relationship reasoning and credit status diffusion calculation, thereby transforming the static enterprise association graph into an enterprise credit assessment graph network carrying dynamic credit status.

[0033] Furthermore, the temporal evolution assessment of the enterprise's credit level includes:

[0034] Load the credit status snapshots of the enterprise credit assessment graph network from multiple time slices in the past from historical storage;

[0035] The credit status of the current time slice and historical credit status snapshots are arranged in chronological order to construct a sequence of corporate credit status evolution.

[0036] Trend analysis is performed on the evolution sequence of the enterprise's credit status to identify long-term improvement, deterioration, or cyclical fluctuation patterns in credit status.

[0037] Volatility analysis was performed on the evolution sequence of the enterprise's credit status to quantify the severity and suddenness of changes in credit status.

[0038] By combining the results of the trend analysis and the volatility analysis, a preliminary credit trajectory describing the company's credit history and potential future trends is generated.

[0039] Furthermore, the introduction of external compliance and risk labels to constrain the initial credit trajectory includes:

[0040] Access the enterprise public credit information platform to query and obtain the target enterprise's administrative penalty label, serious dishonesty label, and honor and award label;

[0041] Establish a label influence mapping table, which defines the specific adjustment direction and magnitude of different types of labels on credit scores;

[0042] The administrative penalty label and the serious dishonesty label are used as negative constraints, and the honor and reward label is used as a positive constraint. The credit level indicated by the preliminary credit trajectory is quantitatively corrected according to the label influence mapping table.

[0043] The quantified and corrected results are mapped to a preset credit scoring range to generate the benchmark credit score.

[0044] Furthermore, the initiation scoring backtracking verification mechanism includes:

[0045] Extract the historical score sequence of the enterprise entity over a continuous period of time from the historical credit scoring database;

[0046] Calculate the continuity index between the benchmark credit score and the historical score sequence in terms of recent trends;

[0047] Calculate the relative position index of the benchmark credit score and the score distribution of a group of companies of the same industry and size at the current point in time;

[0048] A consistency threshold is set. If both the continuity indicator and the relative position indicator meet the consistency threshold, the benchmark credit score is directly used as the final corporate credit score.

[0049] If any indicator is not met, a scoring review process is triggered to adjust the benchmark credit score until the consistency threshold is met, and the adjusted score is used as the final corporate credit score.

[0050] Furthermore, the step of associating the consumption geographic location sequence in the structured consumer behavior fragment set with the regional economic indicator database includes:

[0051] Analyze the aforementioned consumption geographic location sequence and statistically analyze the different administrative divisions involved in the enterprise's consumption behavior and their frequencies;

[0052] Obtain the quarterly GDP growth rate, corporate tax density, and consumer complaint rate for the corresponding administrative division from the regional economic indicators database;

[0053] A comprehensive regional economic health index is constructed for each administrative division unit. The comprehensive regional economic health index is calculated by weighting the quarterly GDP growth rate, corporate tax density, and consumer complaint rate.

[0054] The economic activity parameter of the region where the enterprise operates is calculated by weighting the comprehensive index of regional economic health based on consumption frequency.

[0055] Furthermore, the feature interleaving and fusion algorithm includes:

[0056] The aforementioned consumption stability characteristics, consumption scale adaptability characteristics, geographical risk correlation characteristics, and behavioral consistency characteristics are normalized to the same numerical range;

[0057] A fully connected feature interaction network is constructed, and the normalized features are input into the feature interaction network. The feature interaction network learns the high-order combination relationship between features through multi-layer nonlinear transformation.

[0058] Principal component analysis is performed on the output of the last layer of the feature interaction network, and principal components with contributions exceeding a set threshold are retained.

[0059] The retained principal components are concatenated to form a multidimensional credit feature vector with a lower dimension than the total number of original features but a higher information density.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] By constructing a dynamic corporate credit assessment graph network, corporate entities and their multidimensional credit feature vectors are mapped to nodes and node attributes in the graph. Relationships between nodes are inferred and constructed based on data such as consumer behavior. Credit state diffusion calculations are performed on this network, utilizing the message passing mechanism of graph neural networks to allow credit signals to iteratively propagate and aggregate along the relationships between multi-hop neighbor nodes. This technology can automatically uncover and quantify the implicit interconnected risks between enterprises, placing the credit assessment of a single enterprise within its own business ecosystem network. Credit scoring no longer relies solely on the isolated attributes of the enterprise itself but incorporates the state of its interconnected network, thus enabling earlier and more accurate warnings of cascading credit deterioration caused by common risks from upstream and downstream enterprises, related groups, or regions. This represents a leap from point-based assessment to systemic risk assessment.

[0062] Based on the temporal features extracted from the enhanced behavioral profile, a continuous evolution assessment of credit level is conducted to generate a preliminary credit trajectory. Building upon this, external hard compliance and risk labels are introduced as constraints to perform boundary correction on the credit trajectory, ensuring it conforms to the logic of known risk events. A scoring backtracking verification mechanism is initiated, utilizing historical scoring data streams to construct a consistency verification model. By comparing the predictive continuity and the reasonableness of fluctuations between the current benchmark score and the historical trajectory, abnormal deviations are detected and smoothed. The effect of this technique is to generate a final credit score with temporal consistency and logical robustness. The scoring results not only reflect the current state but also form a smooth, interpretable, and self-consistent credit evolution curve with historical information, avoiding abrupt score changes and improving the model's stability and decision credibility over time. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the steps of the intelligent assessment method for enterprise credit scoring based on transportation and travel consumption data described in this invention.

[0064] Figure 2 A flowchart for performing cross-data source enterprise attribute association mapping;

[0065] Figure 3 A heatmap showing the correlation between credit features;

[0066] Figure 4 A line chart showing the time-series evolution trend of credit.

[0067] Figure 5 A correlation analysis diagram of factors influencing credit scores. Detailed Implementation

[0068] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] See Figure 1The system retrieves a set of original transportation consumption records for the enterprise entity from authorized data interfaces or databases. This set contains consumption details for all vehicles under the enterprise's name within a certain period. The original transportation consumption record set undergoes multi-level parsing to generate a structured set of consumption behavior fragments. Based on this structured set of consumption behavior fragments, cross-data source enterprise attribute association mapping is performed, forming an enhanced enterprise behavior profile by associating with external business registration, regional economic, and news data sources. Based on this enhanced enterprise behavior profile, deep extraction and fusion of credit-related features are performed, extracting key features from multiple dimensions and generating multi-dimensional credit feature vectors through fusion algorithms. A dynamic enterprise credit assessment graph network is constructed, with enterprises as nodes and various inter-enterprise relationships as edges. Multi-dimensional credit feature vectors are injected into network nodes as initial attributes. The credit status of nodes is updated through relational reasoning and credit status diffusion calculation within the network. Based on the results of the credit status diffusion calculation, a temporal evolution assessment of the enterprise's credit level is performed, generating a preliminary credit trajectory by analyzing the historical sequence of credit status. External compliance and risk labels from the enterprise public credit information platform are introduced to perform constrained corrections on the preliminary credit trajectory, outputting a benchmark credit score. A scoring backtesting and verification mechanism is initiated, which uses the company's own historical scoring data stream and the scoring distribution of the same industry group to conduct a consistency test on the benchmark credit score. After necessary review and adjustment, the final corporate credit score is generated.

[0070] In one embodiment of the present invention, see [reference] Figure 2The system retrieves a set of original transportation consumption records for a business entity over the past six months. This set includes details of taxi, ride-hailing, public transportation, and shared bicycle transactions. Each record contains a consumption timestamp, the geographic coordinates of the transaction, the consumption amount, and a unique identifier for the business. First-level parsing is performed, sequentially reading each record from the original transportation consumption record set. The consumption timestamps are extracted and sorted to form a consumption time series, the geographic coordinates are extracted to form a consumption geographic series, and the consumption amounts are extracted to form a consumption amount series. These three series maintain a one-to-one correspondence in the index. The second level of analysis involves using Fourier transform to detect periodic signals and identify periodic consumption patterns, such as peak commuting times on weekdays, for the consumption time series. Density-based clustering algorithms are also used to identify clusters of sudden consumption events occurring within a short period. For the consumption geographic location series, the DBSCAN spatial clustering algorithm identifies frequently occurring geographic point clusters as permanent areas, and a trajectory reconstruction algorithm connects discrete geographic points in chronological order to form a flow path map. For the consumption amount series, the mean and standard deviation within a sliding window are calculated. Intervals within ±2 standard deviations of the mean are classified as stable intervals, while intervals outside this range are marked as abnormal fluctuation intervals. The third level of analysis, based on the timestamps and coordinates of the consumption time and geographic location series, aligns and associates periodic consumption patterns, sudden consumption clusters, permanent areas, flow path maps, stable intervals, and abnormal fluctuation intervals in time and space. For example, a sudden consumption cluster is associated with its corresponding flow path segment and the abnormally high consumption amount interval at that time, synthesizing a final structured set of consumption behavior segments.

[0071] In some embodiments, the enterprise identification information (i.e., the unified social credit code) in the structured consumer behavior fragment set is matched against the business registration database. Upon successful matching, the enterprise's legal industry classification and registered capital information are obtained. Simultaneously, the consumption geographic location sequence in the structured consumer behavior fragment set is analyzed to statistically analyze the different administrative divisions involved in the enterprise's consumption behavior and their frequency of occurrence. For example, the statistical results show that the enterprise's consumption occurred in regions A, B, and C, with frequencies of 50%, 30%, and 20%, respectively. The latest quarterly GDP growth rate, enterprise tax density, and consumer complaint rate for regions A, B, and C are obtained from the regional economic indicator database. A comprehensive regional economic health index is constructed for each administrative division. The comprehensive regional economic health index is calculated by weighting the quarterly GDP growth rate, enterprise tax density, and consumer complaint rate. The calculation formula is as follows:

[0072]

[0073] in: Represents the comprehensive index of regional economic health. This represents the normalized quarterly GDP growth rate. Represents the normalized corporate tax density. This represents the normalized consumer complaint rate. , , For pre-set weighting coefficients and satisfying The weighted average of the calculated regional economic health index for each region is calculated based on consumption frequency. Specifically, the regional economic health index for region A is multiplied by 0.5, for region B by 0.3, and for region C by 0.2, and then summed to calculate the economic activity parameter for the region where the company operates. Furthermore, the analyzed periodic consumption patterns, such as peak consumption times on Monday mornings and Friday afternoons, are semantically correlated with publicly available news data on the company's business activities. Natural language processing techniques are used to extract time keywords and business verbs from news headlines and summaries to infer that the company's business activity cycle may be related to weekly customer visits and project meetings. Integrating and matching the obtained legal industry classification, registered capital information, calculated economic activity parameters, and inferred business activity cycles creates an enhanced corporate behavior profile.

[0074] In one embodiment of the present invention, credit-related features are deeply extracted and fused based on an existing enhanced corporate behavior profile. The enhanced corporate behavior profile includes information about a company with a registered capital of 5 million yuan belonging to the software and information technology services industry, its consumption amount sequence, consumption time interval sequence, inferred business activity cycle, mapped regional economic activity parameters, and known risk event frequencies over the past six months. Consumption stability features are extracted from the enhanced corporate behavior profile. These features are quantified by the variance of the consumption amount sequence and the regularity of the consumption time intervals. Specifically, the variance of the consumption amount sequence is calculated, and the standard deviation of the intervals between adjacent consumption time points is calculated. The variance and the reciprocal of the standard deviation are weighted and summed to obtain a scalar value, which serves as the consumption stability feature value. The consumption scale adaptability feature is extracted, reflecting the degree of matching between consumption level and the company's registered capital and the average level of its industry. Specifically, it calculates the ratio of the company's average monthly consumption to its registered capital. Simultaneously, it queries the industry database for the average monthly transportation consumption benchmark value for the software and information technology services industry, calculates the ratio of the company's average monthly consumption to this industry benchmark value, and combines these two ratios after normalization into a single consumption scale adaptability feature value. The geographic risk correlation feature is extracted, calculated based on the regional economic activity parameter mapped by the consumption geographic location sequence and the frequency of known credit risk events in the company's operating area. Specifically, it performs a reverse transformation on the regional economic activity parameter to obtain a basic risk coefficient, then superimposes the total number of risk events such as industrial and commercial administrative penalties and contract disputes in the company's resident area over the past year, retrieved from a public risk event database. After logarithmic scaling, this is multiplied by the basic risk coefficient to generate the geographic risk correlation feature value. Extract behavioral consistency features to assess the degree of agreement between the inferred business activity cycle and the actual consumption behavior cycle. Specifically, the inferred business activity cycle, such as "the middle of the week is the business peak", is cross-compared with the high-frequency consumption time periods identified by clustering in the actual consumption time series. The proportion of overlapping time windows is calculated and used as the behavioral consistency feature value.

[0075] In some embodiments, a feature interleaving and fusion algorithm is used to nonlinearly combine and reduce the dimensionality of consumption stability features, consumption scale adaptability features, geographical risk correlation features, and behavioral consistency features. The four values ​​of consumption stability features, consumption scale adaptability features, geographical risk correlation features, and behavioral consistency features are transformed to the same numerical range of 0 to 1 using a maximum-minimum normalization method. A fully connected feature interaction network with two hidden layers is constructed, with the first hidden layer containing eight neurons and the second hidden layer containing four neurons. The four normalized feature values ​​are input as input vectors to the feature interaction network. The feature interaction network learns higher-order combination relationships between features through nonlinear transformations between layers. The output of the first hidden layer is given by the formula:

[0076]

[0077] in: It is the input feature vector. It is the weight matrix of the first layer. It is the bias vector of the first layer. It is the ReLU activation function. This is the output vector of the first hidden layer. Principal component analysis is performed on the output vector of the last layer of the feature interaction network to calculate the covariance matrix, its eigenvalues, and eigenvectors, retaining principal components with a cumulative contribution exceeding 85%. These retained principal components are then sequentially concatenated to form a multidimensional credit feature vector with a lower dimension than the original total number of features but higher information density. In essence, after processing by the feature interleaving and fusion algorithm, the original four independent features are merged into a single multidimensional credit feature vector containing three elements.

[0078] In one embodiment of the present invention, when constructing a dynamic enterprise credit assessment graph network, a set of related enterprises are used as nodes in the graph network. The set of nodes includes the target enterprise and its upstream and downstream partners, holding subsidiaries, and other enterprises with adjacent registered addresses. Supply chain relationships, equity relationships, and geographical proximity relationships between enterprises are used as edges in the graph network. Supply chain relationships are extracted from publicly available bidding and procurement data, equity relationships are extracted from industrial and commercial equity penetration data, and geographical proximity relationships are calculated based on the latitude and longitude of the enterprise's registered address, where the distance is less than a set threshold. A static enterprise association graph containing nodes and edges is initialized. Multidimensional credit feature vectors are injected into the static enterprise association graph as initial attributes of the corresponding enterprise nodes. The multidimensional credit feature vector of each enterprise node is a numerical vector, which is assigned to the corresponding node as its initial credit status attribute.

[0079] Design a credit state propagation rule. The credit state propagation rule defines the credit influence weights between nodes based on the edge type and strength, and the similarity of the target node's credit features to the source node. For any connection between the source node and the target node... With the target node The edge, its credit influence weight From the formula:

[0080]

[0081] in: These are basic coefficients set based on edge type (supply chain, equity, geographical proximity). It is a specific numerical value that characterizes the strength of the edge. and These are nodes and nodes The multidimensional credit feature vector This represents the square of the Euclidean distance between two vectors. It is a positive parameter that adjusts the degree of influence of similarity. The credit status propagation rule is iteratively run on a static enterprise association graph. In each iteration, each node updates its own credit status vector through a weighted aggregation function based on the credit status of all its neighboring nodes and their corresponding credit influence weights. This achieves the diffusion and rebalancing of credit features along the graph network edges, completing relational reasoning and credit status diffusion calculation. After a preset number of iterations, the node credit status tends to stabilize, thus transforming the static enterprise association graph into an enterprise credit evaluation graph network carrying dynamic credit status.

[0082] In some embodiments, the edge strength in the credit state propagation rule For supply chain relationships, this could be annual transaction volume; for equity relationships, it could be shareholding ratio; and for geographical proximity relationships, it could be the square of the inverse of the distance. It can be understood that the iterative process of the credit state propagation rule simulates the transmission effect of credit risk or credit reputation in a business network. Optionally, the weighted aggregation function can be a weighted sum of the credit state vectors of neighboring nodes, combined with the decay of the node's own state in the previous round. Optionally, the termination condition for the iteration can be set to the average change in the credit state vectors of all nodes in two consecutive iterations being less than a very small threshold.

[0083] See Figure 3The graph presents the linear correlations among four core credit characteristics—consumption stability, scale adaptability, geographic risk correlation, and behavioral consistency—during the credit feature integration stage. The colors and values ​​in the graph jointly represent the magnitude of the correlation coefficients: red to yellow represent positive correlations, and blue to dark blue represent negative correlations; the darker the color, the closer the absolute value of the correlation coefficient is to 1. The graph clearly shows: Strong negative correlation pairs: the correlation coefficient between geographic risk correlation and behavioral consistency is -0.97, and the correlation coefficient between geographic risk correlation and consumption stability is -0.84. This indicates that the closer a company's consumption behavior matches its business activity cycle (high behavioral consistency), the lower the regional credit risk mapped by its consumption location (low geographic risk correlation); companies with higher consumption stability also have lower geographic risk in their operating areas. Strong positive correlation pairs: the correlation coefficient between consumption stability and behavioral consistency is 0.80, and the correlation coefficient between scale adaptability and behavioral consistency is 0.64. This indicates that the more stable a company's consumption behavior, the higher the degree of alignment between its actual consumption cycle and its business activity cycle; the closer the consumption scale matches the company's capital and industry level (high scale adaptability), the stronger its behavioral consistency. The correlation coefficients for moderately correlated pairs are: 0.44 for consumption stability and 0.56 for scale adaptability and -0.56 for geographical risk. This indicates that companies with higher consumption stability generally have better scale adaptability, while companies with better scale adaptability tend to have lower geographical risk in their operating areas. These correlations provide crucial linear priors for subsequent feature interleaving and fusion algorithms, helping to retain more discriminative credit feature information during nonlinear combination and dimensionality reduction, thereby improving the information density and predictive power of multidimensional credit feature vectors.

[0084] In one embodiment of the present invention, when performing a time-series evolution assessment of corporate credit levels, snapshots of the credit status of the corporate credit assessment graph network over multiple time slices are loaded from historical storage. These time slices are on a monthly basis; for example, snapshots of the corporate credit assessment graph network credit status of the target company on the last day of each of the past twelve months are loaded. The corporate credit assessment graph network credit status of the current time slice is arranged chronologically with the historical credit status snapshots to construct a corporate credit status evolution sequence. Each element in the corporate credit status evolution sequence represents a comprehensive value of the corporate credit assessment graph network credit status at a given point in time. Trend analysis is performed on the corporate credit status evolution sequence, using the least squares method to fit a straight line. The slope of the line is calculated to identify long-term improvement, deterioration, or cyclical fluctuation patterns in the corporate credit assessment graph network credit status. A positive slope indicates long-term improvement, and a negative slope indicates long-term deterioration. Volatility analysis is performed on the corporate credit status evolution sequence, calculating the ratio of the standard deviation to the mean of the corporate credit status evolution sequence to quantify the severity and suddenness of changes in the corporate credit assessment graph network credit status. By combining the results of trend analysis and volatility analysis, a preliminary credit trajectory describing the company's credit history and potential future trends is generated. The preliminary credit trajectory is represented by a data structure that includes trend direction, volatility amplitude, and prediction confidence interval.

[0085] In a specific implementation, when introducing external compliance and risk labels to constrain and calibrate the initial credit trajectory, the system accesses the enterprise's public credit information platform. Using the enterprise's unified social credit code as the query condition, it retrieves and obtains the target enterprise's administrative penalty labels, serious dishonesty labels, and honorary award labels. A label impact mapping table is established, defining the specific adjustment direction and magnitude of different types of labels on the credit score. An example of a label impact mapping table is shown in Table 1 below.

[0086] Table 1: Tag Influence Mapping Table

[0087]

[0088] Administrative penalty labels and serious dishonesty labels are used as negative constraints, while honor and reward labels are used as positive constraints. The credit level indicated by the initial credit trajectory is quantitatively corrected based on the label impact mapping table. The quantitative correction process uses the following formula:

[0089]

[0090] in: This represents the revised credit score. This represents the baseline predicted score extracted from the initial credit trajectory. This is the total number of related tags found. It is the first The effectiveness decay coefficient of each label, This refers to the adjustment range corresponding to the label influence mapping table. The quantified correction result is mapped to a preset credit score range of 0 to 100 to generate a baseline credit score.

[0091] In some embodiments, the method for extracting the comprehensive value of a company's credit status from a snapshot of its credit status in a corporate credit assessment graph network can be to take the first principal component score of the multidimensional credit feature vector of the company node in the corporate credit assessment graph network after credit status diffusion calculation as the comprehensive value. Optionally, in addition to linear trend fitting, seasonal decomposition algorithms can be introduced in trend analysis to identify periodic fluctuation patterns. Optionally, absolute mean deviation can be used as an alternative indicator in volatility analysis to reduce the impact of outliers. It is understood that time-series evolution assessment aims to capture the dynamic changes in corporate credit status. External label correction provides a rigid adjustment mechanism based on factual regulatory information.

[0092] See Figure 4 This chart illustrates the monthly credit composite score trend of a target company over the past 12 months. Specifically, the monthly credit composite score is obtained by extracting the first principal component score from the multidimensional credit feature vectors of the corresponding nodes in the company's credit assessment graph network through principal component analysis. In the chart, blue dots represent the actual credit composite score for each month, and the pink dashed line is the trend line obtained through least squares fitting, with a slope of 0.35, indicating a long-term improving trend in the company's credit status during the observation period. The yellow horizontal line represents the average credit composite score over the 12 months (76.9), serving as a baseline for evaluating monthly performance. In terms of fluctuations, the credit composite scores show significant peaks in July and November, especially in July reaching a high of nearly 79.9, while February and June are relatively low, reflecting some short-term fluctuations in the company's credit status. The positive slope of the trend line, combined with the fluctuations of the monthly values ​​around the mean, constitutes the preliminary trajectory of the company's credit history, providing foundational data for subsequent binding corrections using external compliance and risk labels.

[0093] In one embodiment of the present invention, when the scoring backtracking verification mechanism is initiated, the historical scoring sequence of the enterprise over a continuous period of time is extracted from the historical credit scoring database, using the enterprise's unified social credit code as an index. For example, the final enterprise credit scores recorded at the end of each month for the past twelve months are extracted to form a historical scoring sequence [70,72,71,73,75,74,76,77,78,76,77,78]. The benchmark credit score generated in this assessment is 78 points. The continuity index of the benchmark credit score and the historical scoring sequence in terms of recent trends is calculated. The continuity index is quantified by comparing the degree of agreement between the benchmark credit score and the predicted value based on the recent historical scoring sequence. The last three values ​​of the historical scoring sequence [76,77,78] are selected for linear fitting to obtain a prediction line. The reciprocal of the absolute difference between the benchmark credit score 78 and the predicted value of the prediction line in the current period is calculated as part of the continuity index. At the same time, the standard deviation of the values ​​of the last six months of the historical scoring sequence is calculated to measure recent fluctuations. The reciprocal of the absolute difference and the recent fluctuation values ​​are combined to generate the final continuity index value. The relative position index of the benchmark credit score and the group of companies of the same industry and size in the score distribution at the current time point is calculated. The relative position index is obtained by determining the percentile ranking of the benchmark credit score in the reference group. The latest credit scores of all companies belonging to the "software and information technology services industry" and with registered capital between 5 million and 10 million are retrieved from the historical credit score database to form a reference group score set. The percentile ranking of the benchmark credit score 78 in this set is calculated and this percentile value is used as the relative position index.

[0094] A consistency threshold is set, which includes a lower limit for the continuity indicator and a threshold range for the relative position indicator. For example, the lower limit for the continuity indicator is set to 0.7, and the threshold range for the relative position indicator is set to [0.3, 0.7] (i.e., a percentile ranking between 30% and 70% is considered normal). If both the continuity indicator and the relative position indicator meet the consistency threshold, and the calculated value of the continuity indicator is 0.82, which is greater than 0.7, and the calculated value of the relative position indicator is 0.65, which is within the range [0.3, 0.7], then the baseline credit score of 78 is directly used as the final corporate credit score output. If either the continuity indicator or the relative position indicator does not meet the consistency threshold, a scoring review process is triggered. The scoring review process includes automatically re-verifying the generation logic of the multidimensional credit feature vector, reviewing the accuracy of the external compliance and risk label mapping, checking the propagation parameters of the dynamic corporate credit assessment graph network, generating an adjusted score based on the review results until both the new continuity indicator and the new relative position indicator corresponding to the adjusted score meet the consistency threshold, and using the adjusted score as the final corporate credit score.

[0095] In some embodiments, the continuity index can be calculated by measuring the similarity between the predicted value of an autoregression model based on historical scoring sequences and the benchmark credit score, and the similarity calculation formula is as follows:

[0096]

[0097] in: Represents a continuous index value. This represents the predicted score for the current period using a time series model built from historical score sequences. This represents the baseline credit score generated in this assessment. Representing the range of historical rating sequences, this formula ensures that the continuity index value falls within the range of 0 to 1. In some embodiments, the relative position index can be calculated directly by dividing the ranking number of the benchmark credit score in the reference group rating set by the total number of people in the group.

[0098] Optionally, the length of the recent historical scoring sequence can be defined as the last three to six months. Optionally, in addition to industry and size, the definition of the reference group can also include dimensions such as years of operation to improve comparability. It can be understood that the scoring backtesting and verification mechanism enhances the longitudinal rationality and horizontal comparability of the scoring results by introducing two constraints: the time dimension and the group dimension. The scoring review process is a feedback and adjustment link to ensure the consistency of the scoring logic and the robustness of the results.

[0099] See Figure 5 In the enterprise credit scoring model based on transportation and travel consumption data, the core of the expansion phase is to quantify the correlation strength and direction between each key feature and the final credit score, thereby providing a basis for feature fusion and model optimization. The figure shows strongly positively correlated features: consumption stability (0.85), behavioral consistency (0.82), scale adaptability (0.78), and regional economic health (0.75) all show significant positive correlations, indicating that improvements in these features directly drive improvements in credit scores. Among them, consumption stability has the strongest correlation, reflecting that the regularity of enterprise consumption behavior is the core basis for credit assessment. Strongly negatively correlated features: historical volatility (-0.70) and geographical risk correlation (-0.65) show significant negative correlations, meaning that the greater the historical volatility of enterprise consumption behavior and the higher the risk of the operating area, the lower its credit score. The correspondence between the correlation direction and business logic: positively correlated features correspond to the stability and compliance of enterprise operations, while negatively correlated features correspond to potential risks and uncertainties. This result is completely consistent with the technical path of "consumption stability → multi-dimensional credit features → dynamic graph network assessment" in the project background, and also verifies the rationality of the feature extraction and fusion process.

[0100] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A smart assessment method for enterprise credit scoring based on transportation and travel consumption data, characterized in that: The method includes: Obtain the original collection of traffic consumption records of the enterprise entity; The original set of traffic consumption records is parsed at multiple levels to generate a set of structured consumption behavior fragments; Based on the structured set of consumer behavior fragments, cross-data source enterprise attribute association mapping is performed to form an enhanced enterprise behavior profile; Based on the enhanced corporate behavior profile, credit-related features are deeply extracted and fused to generate a multi-dimensional credit feature vector. A dynamic enterprise credit assessment graph network is constructed, and the multidimensional credit feature vector is injected into the dynamic enterprise credit assessment graph network to perform relational reasoning and credit status diffusion calculation. Based on the results of credit status diffusion calculation, a time-series evolution assessment of corporate credit level is performed to generate a preliminary credit trajectory. External compliance and risk labels are introduced to constrain and correct the preliminary credit trajectory, and a benchmark credit score is output. A scoring backtracking verification mechanism is initiated, using historical scoring data streams to perform consistency checks on the benchmark credit score, and generating the final corporate credit score accordingly.

2. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 1, characterized in that, The multi-level parsing of the original traffic consumption record set includes: The first level of analysis is performed to separate the consumption time series, consumption geographical location series, and consumption amount series from the original set of traffic consumption records; the second level of analysis is performed to identify periodic consumption patterns and sudden consumption clusters for the consumption time series, analyze the permanent residence area and flow path map for the consumption geographical location series, and divide the consumption amount series into stable intervals and abnormal fluctuation intervals. The third level of analysis is performed to align and associate the periodic consumption patterns, sudden consumption clusters, resident areas, flow path maps, stable intervals and abnormal fluctuation intervals in time and space, and synthesize the set of structured consumption behavior fragments.

3. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 2, characterized in that, The execution of cross-data source enterprise attribute association mapping includes: The enterprise identification information in the structured consumer behavior fragment set is matched with the business registration database to obtain the enterprise's legal industry classification and registered capital information; At the same time, the consumption geographic location sequence in the structured consumption behavior fragment set is associated with the regional economic indicator database to map the economic activity parameters of the enterprise's operating area; Semantically correlate the aforementioned cyclical consumption patterns with publicly available news data on the company's business activities to infer the company's business activity cycle; By integrating the aforementioned legal industry classifications, registered capital information, economic activity parameters, and business activity cycles, the enhanced corporate behavior profile is formed.

4. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 3, characterized in that, The deep extraction and fusion of credit-related features includes: Consumer stability features are extracted from the enhanced corporate behavior profile, and these features are quantified by the variance of the consumption amount sequence and the regularity of the consumption time interval. Extract consumption scale adaptability characteristics, which reflect the degree of matching between consumption level and enterprise registered capital and the average level of the industry to which it belongs; Geographic risk association features are extracted, which are calculated based on the regional economic activity parameters mapped by the consumption geographic location sequence and the frequency of known credit risk events in the enterprise's operating area. Extract behavioral consistency features, which assess the degree of agreement between the inferred business activity cycle and the actual consumption behavior cycle; The feature interleaving and fusion algorithm is used to nonlinearly combine and reduce the consumption stability feature, consumption scale adaptability feature, geographical risk correlation feature and behavioral consistency feature to generate the multidimensional credit feature vector.

5. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 4, characterized in that, The construction of the dynamic enterprise credit assessment graph network includes: Initialize a static enterprise relationship graph by using enterprises as nodes in the graph network and supply chain relationships, equity relationships, and geographical proximity relationships between enterprises as edges in the graph network. The multidimensional credit feature vector is injected into the static enterprise association graph as the initial attribute of the corresponding enterprise node; Design a credit state propagation rule, wherein the credit state propagation rule defines the credit influence weight between nodes based on the type and strength of the edge, and the similarity of the target node to the credit features of the source node; The credit status propagation rule is iteratively run on the static enterprise association graph to realize the diffusion and rebalancing of credit features along the graph network edges, and to complete the relationship reasoning and credit status diffusion calculation, thereby transforming the static enterprise association graph into an enterprise credit assessment graph network carrying dynamic credit status.

6. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 5, characterized in that, The time-series evolution assessment of the credit level of the implementing enterprise includes: Load the credit status snapshots of the enterprise credit assessment graph network from multiple time slices in the past from historical storage; The credit status of the current time slice and historical credit status snapshots are arranged in chronological order to construct a sequence of corporate credit status evolution. Trend analysis is performed on the evolution sequence of the enterprise's credit status to identify long-term improvement, deterioration, or cyclical fluctuation patterns in credit status. Volatility analysis was performed on the evolution sequence of the enterprise's credit status to quantify the severity and suddenness of changes in credit status. By combining the results of the trend analysis and the volatility analysis, a preliminary credit trajectory describing the company's credit history and potential future trends is generated.

7. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 6, characterized in that, The introduction of external compliance and risk labels to constrain and correct the initial credit trajectory includes: Access the enterprise public credit information platform to query and obtain the target enterprise's administrative penalty label, serious dishonesty label, and honor and award label; Establish a label influence mapping table, which defines the specific adjustment direction and magnitude of different types of labels on credit scores; The administrative penalty label and the serious dishonesty label are used as negative constraints, and the honor and reward label is used as a positive constraint. The credit level indicated by the preliminary credit trajectory is quantitatively corrected according to the label influence mapping table. The quantified and corrected results are mapped to a preset credit scoring range to generate the benchmark credit score.

8. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 1, characterized in that, The initiation scoring backtracking verification mechanism includes: Extract the historical score sequence of the enterprise entity over a continuous period of time from the historical credit scoring database; Calculate the continuity index between the benchmark credit score and the historical score sequence in terms of recent trends; Calculate the relative position index of the benchmark credit score and the score distribution of a group of companies of the same industry and size at the current point in time; A consistency threshold is set. If both the continuity indicator and the relative position indicator meet the consistency threshold, the benchmark credit score is directly used as the final corporate credit score. If any indicator is not met, a scoring review process is triggered to adjust the benchmark credit score until the consistency threshold is met, and the adjusted score is used as the final corporate credit score.

9. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 3, characterized in that, The step of linking the consumption geographic location sequence in the structured consumption behavior fragment set with the regional economic indicator database includes: Analyze the aforementioned consumption geographic location sequence and statistically analyze the different administrative divisions involved in the enterprise's consumption behavior and their frequencies; Obtain the quarterly GDP growth rate, corporate tax density, and consumer complaint rate for the corresponding administrative division from the regional economic indicators database; A comprehensive regional economic health index is constructed for each administrative division unit. The comprehensive regional economic health index is calculated by weighting the quarterly GDP growth rate, corporate tax density, and consumer complaint rate. The economic activity parameter of the region where the enterprise operates is calculated by weighting the comprehensive index of regional economic health based on consumption frequency.

10. The intelligent assessment method for enterprise credit scoring based on transportation consumption data as described in claim 4, characterized in that, The feature interleaving and fusion algorithm includes: The aforementioned consumption stability characteristics, consumption scale adaptability characteristics, geographical risk correlation characteristics, and behavioral consistency characteristics are normalized to the same numerical range; A fully connected feature interaction network is constructed, and the normalized features are input into the feature interaction network. The feature interaction network learns the high-order combination relationship between features through multi-layer nonlinear transformation. Principal component analysis is performed on the output of the last layer of the feature interaction network, and principal components with contributions exceeding a set threshold are retained. The retained principal components are concatenated to form a multidimensional credit feature vector with a lower dimension than the total number of original features but a higher information density.

Images