Risk state classification model construction method for supply chain finance

Abstract

The application relates to the technical field of supply chain finance, in particular to a risk state classification model construction method for supply chain finance, which comprises the following steps: obtaining supply chain financial data and risk state labels of a plurality of merchants, and calculating interval proportion coefficients between adjacent risk state labels; in a feature space mapped by a feature extraction network, the geometric interval of each category center is constrained according to the interval proportion coefficients to construct a non-equal interval loss; the feature extraction network is iteratively trained by using the non-equal interval loss, and the interval proportion coefficients are adaptively adjusted according to the local discrimination ratio of adjacent boundaries in the iterative training until the feature extraction network converges; and the risk grading result of a to-be-classified merchant is obtained according to the converged feature extraction network. The technical scheme can make the separation resources be differentially allocated to different difficulty category boundaries in the iterative training, and the classification accuracy at the weak boundary between adjacent categories is improved.

Description

Technical Field

[0001] This application relates to the field of supply chain finance technology, and in particular to a method for constructing a risk status classification model for supply chain finance. Background Technology

[0002] In supply chain finance, companies need to conduct credit risk assessments and classifications of upstream and downstream merchants in order to implement differentiated credit granting strategies and risk control measures. As the scale of supply chain networks expands and business complexity increases, the number of merchants continues to grow. Relying on manual experience for risk classification is no longer sufficient to meet the requirements of timeliness and consistency. There is an urgent need to use classification models to automatically and standardizedly determine the risk status of merchants.

[0003] Currently, Chinese patent application document with publication number CN115147208A discloses a method and system for assessing credit risk in supply chain finance based on artificial intelligence. This method determines the set of evaluation features, collects target risk assessment data, performs dimensionality reduction preprocessing on the data, optimizes the support vector machine parameters using a dynamic mutation particle swarm optimization algorithm, and then trains a strong classifier model through an ensemble learning algorithm. Standard sample data is then input into the strong classifier model to construct an intelligent credit assessment model for assessing credit risk in supply chain finance.

[0004] However, the above methods do not consider the differences in the degree of statistical difference between different adjacent categories when constructing the classification model. In actual supply chain finance scenarios, the degree of data difference between adjacent risk levels is often different. When the data difference between a pair of adjacent categories is small, the equidistant separation strategy cannot allocate enough separation resources to this weak boundary, resulting in insufficient classification accuracy of merchants in the risk transition zone. This can easily lead to early-risk merchants being misclassified as normal merchants, causing missed risk control judgments. Summary of the Invention

[0005] To address the technical problem of insufficient classification accuracy, this application provides a method for constructing a risk status classification model for supply chain finance. This method enables differentiated allocation of separated resources to category boundaries of varying difficulty during iterative training, improving classification accuracy at weak boundaries between adjacent categories. It also effectively reduces the geometric complexity of the classification decision surface, enhancing the model's generalization ability and computational efficiency in high-dimensional feature spaces.

[0006] This application provides a method for constructing a risk status classification model for supply chain finance. The method includes: S101, acquiring supply chain finance data and corresponding risk status labels of multiple merchants, calculating the statistical difference between adjacent risk status labels of the supply chain finance data, and obtaining a spacing ratio coefficient; S102, constraining the geometric spacing of each category center in the feature space mapped by the feature extraction network according to the spacing ratio coefficient to construct a non-equidistant loss; S103, iteratively training the feature extraction network using the non-equidistant loss, and adaptively adjusting the spacing ratio coefficient according to the local discriminant ratio of adjacent boundaries in the feature space during iterative training, until the feature extraction network converges and the final category center is obtained; S104, inputting the supply chain finance data of the merchants to be classified into the converged feature extraction network to obtain new merchant features, calculating the Euclidean distance between the new merchant features and each final category center, and taking the risk status label corresponding to the smallest Euclidean distance as the risk classification result.

[0007] By using the ratio of statistical dissimilarity as the spacing ratio coefficient to constrain the geometric spacing of class centers, the geometric layout of the feature space is matched with the objective distribution of the degree of dissimilarity between classes in the original data. This overcomes the limitation of traditional equidistant class center classification methods that apply the same separation strength to all boundaries. In iterative training, the local discriminant ratio is introduced as a feedback signal to adaptively adjust the spacing ratio coefficient, so that the spacing ratio coefficient gradually approaches the optimal value in the feature space from the initial estimate based on the original data, thereby improving the model's adaptability to boundary conditions.

[0008] Preferably, after acquiring supply chain finance data from multiple merchants, the construction method further includes: acquiring the minimum and maximum values ​​of all supply chain finance data in the historical set; subtracting the minimum value from the supply chain finance data and dividing by the difference between the maximum value and the minimum value to obtain a normalized index; and using the normalized index as the supply chain finance data for subsequent calculations.

[0009] Preferably, the risk status labels include a first category, a second category, and a third category; obtaining the spacing ratio coefficient includes: for any single indicator in the supply chain finance data, obtaining the first standard deviation and the second standard deviation corresponding to two adjacent risk status labels; using the sample size of the two risk status labels to perform a weighted sum of the first and second standard deviations to obtain the combined standard deviation; calculating the absolute value of the difference between the means of the single indicator under the two risk status labels, and dividing the absolute value by the combined standard deviation to obtain the effect size index; calculating the average value of the effect size index of each single indicator to obtain the statistical difference between the adjacent boundaries of the two risk status labels; dividing the statistical difference between the adjacent boundaries of the first category and the second category by the statistical difference between the adjacent boundaries of the second category and the third category to obtain the spacing ratio coefficient.

[0010] Preferably, the geometric spacing of each category center is constrained according to the spacing ratio coefficient to construct a non-equidistant loss, including: obtaining an intra-class clustering term that clusters similar samples toward their respective category centers; constructing a spacing constraint term that determines the deviation of the actual category center spacing from the target spacing according to the spacing ratio coefficient; constructing a collinearity regularization term based on the degree to which the category centers deviate from a straight line in the feature space; and performing a weighted summation of the intra-class clustering term, the spacing constraint term, and the collinearity regularization term to obtain the non-equidistant loss.

[0011] The non-equidistant loss is composed of a weighted sum of intra-class clustering, spacing constraint, and collinearity regularization. The intra-class clustering term drives similar samples to cluster towards the class center to reduce intra-class divergence. The spacing constraint term drives the class centers to be arranged in a non-equidistant proportion by penalizing the deviation between the actual spacing and the target spacing. The collinearity regularization term constrains the three class centers to be arranged on a straight line, simplifying the classification decision surface into a hyperplane perpendicular to this line. The synergistic effect of the three terms enables the feature space to simultaneously possess a compact intra-class distribution, a reasonable inter-class spacing ratio, and a simple geometric structure.

[0012] Preferably, the collinearity regularization term is obtained as follows: a first vector pointing from the first category center to the second category center, and a second vector pointing from the first category center to the third category center are obtained; the magnitude product between the first vector and the second vector is calculated; the absolute value of the inner product between the first vector and the second vector is calculated; the difference between the magnitude product and the absolute value of the inner product is used as the collinearity deviation metric; the collinearity deviation metric is divided by the maximum distance between each pair of the three category centers to obtain the collinearity regularization term.

[0013] Preferably, before iteratively training the feature extraction network using the non-equidistant loss, the construction method further includes: setting the initial center of the first category at the origin of the feature space; setting the initial center of the second category at a position along the positive direction of a specific coordinate axis at a distance of a first target spacing from the origin; and setting the initial center of the third category at a position along the positive direction of the specific coordinate axis at a distance of a total target spacing from the origin, wherein the first target spacing and the total target spacing are determined according to the spacing ratio coefficient.

[0014] Preferably, the method for obtaining the local discriminant ratio is as follows: calculate the average Euclidean distance from the feature vectors of all samples in the first category on both sides of any adjacent boundary to the corresponding category center, as the first divergence; calculate the average Euclidean distance from the feature vectors of all samples in the second category on both sides of the adjacent boundary to the corresponding category center, as the second divergence; calculate the Euclidean distance between the category centers of the first category and the second category, as the inter-class separation degree; divide the mean of the first divergence and the second divergence by the inter-class separation degree to obtain the local discriminant ratio.

[0015] The local discriminant ratio extends the standard Fisher discriminant ratio from one-dimensional data space to high-dimensional feature space. It uses the average Euclidean distance from the sample to the class center to represent intra-class divergence and the Euclidean distance between class centers to represent inter-class separation. It can quantify the local weakness of each adjacent boundary and provide targeted feedback signals for the dynamic adjustment direction of the spacing ratio coefficient, so that the adjustment process has a clear basis rather than blind search.

[0016] Preferably, the adjacent boundaries include a first boundary between the first category and the second category, and a second boundary between the second category and the third category; adaptively adjusting the spacing ratio coefficient based on the local discriminant ratio of the adjacent boundaries in the feature space includes: using the local discriminant ratio of the first boundary as a first local ratio, and using the local discriminant ratio of the second boundary as a second local ratio; in response to the first local ratio being greater than the second local ratio, using the learning rate coefficient multiplied by the difference between the two to obtain an increment, and superimposing the increment on the spacing ratio coefficient; In response to the second local ratio being greater than the first local ratio, the reduction is obtained by multiplying the learning rate coefficient by the difference between the two, and the reduction is subtracted from the spacing ratio coefficient.

[0017] By comparing the local discriminant ratios of the two boundaries, the adjustment direction is determined, and the adjustment range is determined by multiplying the learning rate coefficient by the difference. This achieves bidirectional adaptive adjustment of the spacing ratio coefficient. When the weak boundary changes, the adjustment direction switches accordingly, enabling the spacing ratio coefficient to continuously track the dynamic evolution of boundary separation in the feature space. This avoids the problem of the fixed spacing ratio coefficient deviating from the actual optimal ratio in the later stages of training.

[0018] Preferably, the convergence determination condition of the feature extraction network is to satisfy any one of the following: the first condition is that the classification accuracy of the validation set has not improved in multiple consecutive training rounds; the second condition is that in multiple consecutive training rounds, the maximum center displacement among the adjacent round displacements of all the category centers is less than a preset convergence threshold.

[0019] Preferably, after obtaining the new merchant features, the construction method further includes: obtaining the second smallest Euclidean distance from the new merchant features to all final category centers; dividing the smallest Euclidean distance by the second smallest Euclidean distance to obtain a distance ratio; using the difference between 1 and the distance ratio as the classification confidence; and generating and outputting a verification label for the merchant to be classified in response to the classification confidence being lower than a confidence threshold.

[0020] The classification confidence score is calculated by the ratio of the minimum Euclidean distance to the second minimum Euclidean distance. The classification results are expanded from a single category label to a composite output that includes a quantitative assessment of reliability. When the classification confidence score is lower than the threshold, a verification mark is generated to prompt manual verification. This reduces the risk of misjudgment for merchants in the boundary area while ensuring the overall classification efficiency.

[0021] The technical solution of this application has the following beneficial technical effects: By calculating the statistical difference between supply chain finance data of adjacent risk status labels, a spacing ratio coefficient is obtained. Based on this coefficient, a non-equidistant loss is constructed to constrain the geometric spacing of category centers in the feature space. This allows the feature extraction network to arrange category centers according to a non-equidistant ratio that matches the degree of statistical difference in the data during training. At the same time, during iterative training, the spacing ratio coefficient is adaptively adjusted based on the local discriminant ratio to track the dynamic changes in the boundary separation status in the feature space. This allows for the differentiated allocation of separation resources to boundaries of varying difficulty, improving the classification accuracy at weak boundaries with small statistical differences and reducing the probability of misclassifying early-risk merchants as normal merchants.

[0022] Furthermore, the collinear regularization term in the non-equidistant loss constrains the three class centers to be arranged on a straight line, simplifying the classification decision surface into a hyperplane perpendicular to this line. This effectively reduces the geometric complexity of the classification decision surface and improves the model's generalization ability and computational efficiency in high-dimensional feature spaces. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for constructing a risk status classification model for supply chain finance according to an embodiment of this application.

[0024] Figure 2 This is a schematic diagram showing the non-equidistant, collinear arrangement of three category centers in the feature space according to an embodiment of this application. Detailed Implementation

[0025] This application provides a method for constructing a risk status classification model for supply chain finance, applicable to supply chain finance platforms. These platforms need to classify upstream and downstream merchants into three risk levels: normal, warning, and high-risk. Normal corresponds to the first category, indicating a merchant with good operating conditions and a low probability of default; warning corresponds to the second category, indicating early risk signals requiring enhanced monitoring; and high-risk corresponds to the third category, indicating a merchant's severely deteriorated financial situation and a high probability of default. The supply chain finance platform implements differentiated credit granting strategies and risk control measures based on the credit risk classification. Figure 1 As shown, the method for constructing a risk status classification model for supply chain finance includes steps S101 to S104, which are described in detail below.

[0026] S101, acquire supply chain finance data and corresponding risk status labels of multiple merchants, calculate the statistical difference between the supply chain finance data of adjacent risk status labels, and obtain the spacing ratio coefficient.

[0027] In one embodiment, supply chain finance data for multiple merchants is obtained from the business data of a supply chain finance platform. Each merchant's supply chain finance data includes five indicators: debt-to-equity ratio, current ratio, accounts receivable turnover, net profit growth rate, and operating cash flow ratio. Simultaneously, each merchant is labeled with a corresponding risk status tag based on the platform's historical risk control records. It should be noted that the above five indicators are merely preferred examples of this application. In other embodiments, other dimensions of feature data can be selected according to actual business needs, and this application does not impose any restrictions.

[0028] After acquiring supply chain finance data, it is normalized to obtain the minimum and maximum values ​​of all supply chain finance data in the historical set. Specifically, the minimum and maximum values ​​of each indicator are calculated separately for all merchants. The minimum value is subtracted from the supply chain finance data, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized indicator. This normalized indicator is used as the supply chain finance data for subsequent calculations. The value of the normalized indicator falls within the range of 0 to 1, eliminating differences in units and numerical ranges between different indicators.

[0029] Based on normalized supply chain finance data, the statistical difference between adjacent risk status labels is calculated to obtain the distance ratio coefficient. For any single indicator in the supply chain finance data, the first and second standard deviations corresponding to each of the two adjacent risk status labels are obtained. Taking the first boundary "normal-warning" boundary as an example, the first standard deviation is the standard deviation of all merchants in the normal category for this single indicator, and the second standard deviation is the standard deviation of all merchants in the warning category for this single indicator. The first and second standard deviations are weighted and summed using the sample sizes corresponding to the two risk status labels to obtain the combined standard deviation. The specific expression for the combined standard deviation is:

[0030] In the formula, To combine standard deviations, This represents the sample size corresponding to the first risk status label. This represents the sample size corresponding to the second risk status label. The first standard deviation, The second standard deviation is used as the weight for the two standard deviations. The formula uses the sample size of each category as the weight to calculate the weighted average of the two standard deviations, so that the combined standard deviation can comprehensively reflect the overall dispersion of the two categories on this single indicator.

[0031] Calculate the absolute value of the difference between the means of an individual indicator under the two risk status labels, and divide the absolute value by the pooled standard deviation to obtain the effect size index. The effect size index adopts the standard calculation method of the Cohen effect size, and its specific expression is as follows:

[0032] In the formula, For effect size indicators, This represents the average of this single indicator under the first risk status label. The mean of this single indicator under the second risk status label; the effect size indicator is a dimensionless standardized indicator, and the larger the value, the more significant the difference between the two categories in this single indicator.

[0033] Calculate the average effect size of all individual indicators to obtain the statistical difference of each adjacent boundary. The adjacent boundaries include the first boundary between the first category and the second category, and the second boundary between the second category and the third category. That is, the first boundary is the "normal-warning" boundary, and the second boundary is the "warning-high risk" boundary.

[0034] The spacing ratio coefficient is obtained by dividing the statistical difference between the adjacent boundaries of the first and second categories by the statistical difference between the adjacent boundaries of the second and third categories. The spacing ratio coefficient represents the ratio of the statistical differences between two adjacent boundaries. Since the second and third categories are two different categories, the statistical difference between the adjacent boundaries of the second and third categories is non-zero, so the denominator will not be zero in the calculation of the spacing ratio coefficient.

[0035] In one example, the effect sizes of the five indicators at the first boundary, "normal-warning," are 0.31, 0.28, 0.35, 0.22, and 0.29, respectively. What is the statistical difference at the first boundary? The effect sizes of the five indicators at the second boundary, "early warning - high risk," are 0.72, 0.65, 0.81, 0.58, and 0.69, respectively. Therefore, the statistical difference at the second boundary is 0.29. If it equals 0.69, then the spacing ratio factor is... equal Divide by The result was 0.42. When When the value is less than 1, it indicates that the statistical difference between the "normal-warning" boundary and the "warning-high-risk" boundary is less than that between the "warning" and "high-risk" boundaries. In the subsequent feature space, the target distance between the center of the normal category and the center of the warning category should be proportionally less than the target distance between the center of the warning category and the center of the high-risk category.

[0036] Among them, when the statistical difference of the second boundary When it is less than 0.05, The cutoff value is set to 0.05 to avoid numerical anomalies in division operations. The cutoff threshold of 0.05 corresponds to a level of almost no difference in the empirical classification of the Cohen effect.

[0037] In this way, the spacing ratio coefficient transforms the degree of difference between the two boundaries into a dimensionless ratio, providing an objective and accurate data basis for the subsequent non-equidistant layout of the category centers.

[0038] S102, based on the geometric spacing of the category centers in the feature space of the feature extraction network mapping constrained by the spacing ratio coefficient, construct a non-equidistant loss.

[0039] In one embodiment, the feature extraction network is a three-layer fully connected network. The input layer has a dimension of 5, corresponding to 5 supply chain finance data indicators. The hidden layers have dimensions of 32 and 64, respectively. Batch normalization and linear rectified activation functions are used between the layers. The output layer has a dimension of 64, meaning the feature extraction network maps 5-dimensional supply chain finance data into feature vectors in a 64-dimensional feature space. The number of hidden layers and the dimension of the output layer in the fully connected network can be adjusted as needed, and this application does not impose any restrictions.

[0040] Before iteratively training the feature extraction network using non-equidistant loss, the initial positions of the center of the first, second, and third categories are set in the feature space. The initial center of the first category is set at the origin of the feature space, i.e., the center of the normal category is initialized as a 64-dimensional zero vector; the initial center of the second category is set at a position along the positive direction of a specific coordinate axis, with a distance of the first target distance from the origin, i.e., the center of the warning category is initialized along the positive direction of the first coordinate axis, with a distance of the first target distance from the origin. The initial center of the third category is set at a position along the positive direction of a specific coordinate axis, with a distance from the origin equal to the total target distance. In other words, the high-risk category center is initialized with a distance along the positive direction of the first coordinate axis, with a distance from the origin equal to the total target distance. The location.

[0041] The first target spacing and the total target spacing are determined based on the spacing ratio coefficient: setting the baseline spacing. Equal to the square root of the feature dimension, in this embodiment The first target spacing is equal to the square root of 64, which is 8. equal Total target spacing equal In the example of S101, the spacing ratio factor If it is 0.42, then equal , equal This initialization method ensures that the three class centers satisfy the constraints of collinear alignment and target spacing ratio from the start of training, reducing invalid searches in the early stages of training. For example... Figure 2 As shown, the three category centers are arranged along the same straight line in the feature space, with normal category centers, warning category centers, and high-risk category centers distributed sequentially. The first target spacing is the distance between the normal category center and the warning category center, and the total target spacing is the distance between the normal category center and the high-risk category center. The difference between the total target spacing and the first target spacing is the second target spacing, which is the distance between the warning category center and the high-risk category center. Smaller than the second target spacing This reflects a non-equidistant layout based on the spacing ratio coefficient; similar samples cluster towards their respective category centers in the feature space, forming a compact cluster distribution.

[0042] After determining the initial position of the category center, a non-equidistant loss is constructed, which includes intra-class clustering terms, spacing constraints, and collinearity regularization terms.

[0043] Obtain the intra-class clustering term, which represents the aggregation of similar samples towards their respective class centers. The construction of the intra-class clustering term is based on the fundamental form of the standard centrality loss function, proposed by Wen et al. in 2016, which constrains similar samples to cluster towards their class centers in the feature space. The specific expression for the intra-class clustering term is:

[0044] In the formula, This represents the number of samples in the training batch. For the first The 64-dimensional feature vector obtained by mapping each sample through a feature extraction network For the first The category center of the category to which each sample belongs. This is the Euclidean distance. For each sample in the training batch, this term calculates the squared Euclidean distance between its feature vector and its class center, multiplied by the mean of all samples in the batch. This drives similar samples to cluster toward their respective class centers in the feature space, reducing intra-class divergence.

[0045] Based on the spacing ratio coefficient, a spacing constraint term is constructed to determine the deviation of the actual category center spacing from the target spacing; the specific expression of the spacing constraint term is:

[0046] In the formula, , , The centers are categorized into three types: normal, warning, and high-risk. The Euclidean distance between the centers of the normal category and the warning category is... The Euclidean distance between the center of the warning category and the center of the high-risk category is... The first target spacing is... , The target spacing between early warning and high-risk areas is... ; for and The absolute value of the difference. Among them, and The ratio is the spacing ratio coefficient. ,when When less than 1, Less than That is, the distance between the "normal" and "early warning" boundary targets is smaller than the distance between the "early warning" and "high-risk" boundary targets. The network is forced to distinguish between the two categories of normal and early warning in a smaller space, so as to learn more refined discriminative features to distinguish between these two statistically small categories.

[0047] A collinearity regularization term is constructed based on the degree to which the category centers deviate from a straight line in the feature space. Specifically, this includes: obtaining the collinearity regularization term derived from the first category center. Pointing to the center of the second category The first vector, the first vector for ; and by the first category center Pointing to the center of the third category The second vector, the second vector for Calculate the product of the magnitudes of the first vector and the second vector, and the absolute value of the inner product between the first vector and the second vector; use the difference between the absolute value of the product of magnitudes and the inner product as a collinearity deviation metric.

[0048] The collinearity deviation metric is used to approximate the degree to which the first and second vectors deviate from their collinear arrangement, that is, the degree to which the class centers deviate from their collinear arrangement. When the three class centers are completely collinear, the first and second vectors have the same or opposite directions, their inner product equals the product of their magnitudes, and the collinearity deviation metric is zero. The greater the degree of collinearity deviation, the larger the angle between the first and second vectors, and the larger the collinearity deviation metric. Dividing the collinearity deviation metric by the maximum distance between any two of the three class centers yields the collinearity regularization term. Let... for and , and , and The maximum of the three distances, and the specific expression for the collinearity regularization term, is:

[0049] In the collinearity regularization term, divide by Normalization is performed to make the collinear regularization term a dimensionless quantity that is not affected by the absolute distance between the class centers, and only reflects the relative degree to which the three class centers deviate from the collinear arrangement.

[0050] The non-uniformity loss is obtained by weighted summing of the intra-class clustering term, the spacing constraint term, and the collinearity regularization term. The specific expression for the non-uniformity loss is:

[0051] In the formula, This is the weighting coefficient for the spacing constraint term, with a value of 0.5. The weight coefficient for the collinearity regularization term is set to 0.1. The collinearity regularization term is squared and then multiplied by the weight coefficient, so that the penalty for small deviations is weak, while the penalty for large deviations is greatly enhanced, effectively suppressing large deviations.

[0052] in, The determination was made by performing a grid search on the validation set in the range of 0.1 to 1.0 with a step size of 0.1. The value was determined by searching the validation set in the range of 0.01 to 0.5 with a step size of 0.01, and the value corresponding to the highest classification accuracy on the validation set was selected. and The value of , and in order to quickly determine and The maximum number of searches is set to 100.

[0053] During training, the total loss function is a weighted sum of the cross-entropy classification loss and the non-equidistant loss; that is, the total loss equals the cross-entropy classification loss plus... Multiply by a non-equidistant loss, where the cross-entropy classification loss is the standard multi-class cross-entropy loss function, used to drive the network to learn the correct classification boundary. The overall weight for the non-equidistant loss is set to 0.01. In the early stages of training, the typical value for the cross-entropy classification loss is approximately 1.1, while the typical value for the intra-class clustering term in the non-equidistant loss is approximately 50 to 100. By setting the overall weight to 0.01, and multiplying the non-equidistant loss by 0.01, the magnitudes of the two losses are matched, ensuring a balanced contribution from both parts during gradient updates. This overall weight is adjusted based on the initial magnitude ratio of the two loss parts observed in preliminary experiments.

[0054] Non-equidistant loss embeds the spacing ratio coefficient into the loss function through a spacing constraint term, ensuring that the class center spacing matches the statistical difference ratio of the data. This allocates more feature space separation resources to the boundaries with larger statistical differences. The collinearity regularization term constrains the three class centers to be aligned on a straight line, simplifying the classification decision surface to a hyperplane perpendicular to this line and reducing the geometric complexity of the classification decision.

[0055] S103, the feature extraction network is iteratively trained using the non-equidistant loss, and the spacing ratio coefficient is adaptively adjusted according to the local discriminant ratio of adjacent boundaries in the feature space during the iterative training until the feature extraction network converges to obtain the final category center.

[0056] In one embodiment, based on non-equidistant loss and an initialized feature extraction network, the feature extraction network is iteratively trained using normalized supply chain finance data. The training process employs an adaptive learning rate optimizer with a learning rate of 0.001, a batch size of 256, and a maximum training epoch limit of 200.

[0057] The execution process for each training batch is as follows: 256 samples are randomly selected from the training set, and their normalized supply chain finance data is input into the feature extraction network to obtain 256 64-dimensional feature vectors. Based on these feature vectors and the current three class centers, the total loss function is calculated, and the gradient is calculated and the network parameters are updated using the backpropagation algorithm. Simultaneously, for each class appearing in the current batch, the arithmetic mean of the feature vectors of all samples belonging to that class in the batch is calculated as the batch class mean. The class center of that class is updated using an exponential moving average method, i.e., the new class center for any class. Satisfying the relation: ; As the old category center of this category, This is the batch mean of the category; when a batch does not contain samples of a certain category, the category center of that category remains unchanged.

[0058] After each training epoch, the local discriminant ratio of adjacent boundaries in the feature space is calculated. This local discriminant ratio is constructed based on the standard Fisher discriminant ratio. The standard Fisher discriminant ratio is a statistical indicator that measures the separability of two classes of data, and its expression is:

[0059] In the formula, and These are the means of the two types of data, and The variances of the two classes of data are respectively used to measure the ratio of inter-class distance to intra-class divergence in a one-dimensional data space. In this application, it is necessary to evaluate the ease of separation of the "normal-warning" boundary and the "warning-high risk" boundary in a 64-dimensional feature space. The Fisher discriminant ratio is extended from one dimension to a higher dimension space. The variance is replaced with the average Euclidean distance from the sample to the class center to characterize the intra-class divergence in the higher dimension space, and the mean difference is replaced with the Euclidean distance between the class centers to characterize the inter-class distance in the higher dimension space. This is then transformed into a local discriminant ratio for a single boundary.

[0060] Specifically, adjacent boundaries include a first boundary and a second boundary; the local discriminant ratio is obtained as follows: calculate the average Euclidean distance from the feature vectors of all samples in the first class on both sides of any adjacent boundary to the corresponding class center, as the first divergence; calculate the average Euclidean distance from the feature vectors of all samples in the second class on both sides of the adjacent boundary to the corresponding class center, as the second divergence; calculate the Euclidean distance between the class centers of the first class and the second class, as the inter-class separation; divide the mean of the first divergence and the second divergence by the inter-class separation to obtain the local discriminant ratio. The expression for the local discriminant ratio is:

[0061] In the formula, Choosing 1 or 2 corresponds to the "normal-warning" boundary and the "warning-high risk" boundary, respectively. The number of samples in the first category at this boundary. The number of samples in the second category at this boundary. For the first category The feature vector of each sample For the second category The feature vector of each sample As the category center of the first category, As the category center of the second category, for The Euclidean distance is used to characterize the inter-class separation, which is a positive number greater than 0. In the local discriminant ratio, the numerator is the arithmetic mean of the intra-class average Euclidean distances of the two classes on both sides of the boundary, representing the overall intra-class divergence at the boundary; the denominator is the Euclidean distance between the centers of the two classes on both sides of the boundary, representing the degree of inter-class separation at the boundary. The larger the local discriminant ratio, the greater the intra-class divergence relative to the inter-class distance at the boundary, that is, the weaker the boundary and the more difficult it is to classify correctly.

[0062] In one example, suppose that after the 20th training epoch, the local discriminative ratio of the "normal-warning" boundary is... The local discrimination ratio of the "early warning - high risk" boundary is 0.38. It is 0.25. Greater than This indicates that the boundary between "normal" and "warning" is weaker.

[0063] The spacing ratio coefficient is adaptively adjusted based on the local discriminant ratio of adjacent boundaries in the feature space. Specifically, this includes: first, using the local discriminant ratio of the first boundary as the first local ratio, and the local discriminant ratio of the second boundary as the second local ratio; in response to the first local ratio being greater than the second local ratio, it indicates that the "normal-warning" boundary is weaker, due to the first target spacing... ; This is a spacing ratio factor; when the "normal-warning" boundary is weaker, the spacing of that boundary should be increased. To improve classification ability, the learning rate coefficient is multiplied by the difference between the two values ​​to obtain an increment, which is then added to the distance scaling factor. Since the second local ratio is greater than the first local ratio, indicating a weaker "early warning - high risk" boundary, the learning rate coefficient is multiplied by the difference between the two values ​​to obtain a decrease, which is then subtracted from the distance scaling factor.

[0064] Among them, the learning rate coefficient The learning rate coefficient was set to 0.05. The preliminary experiment was conducted as follows: Five candidate values ​​(0.01, 0.03, 0.05, 0.08, and 0.1) were tested on the training set, with five iterations performed for each. The convergence speed and stability of the classification accuracy on the validation set were used as evaluation metrics. The convergence speed of classification accuracy was defined as the increase in classification accuracy before and after five iterations, and stability was defined as the negative of the variance of the classification accuracy over the five iterations. 0.05 corresponds to the fastest convergence and the result without oscillations; therefore, the learning rate coefficient was chosen accordingly. The value is set to 0.05. Meanwhile, the adjusted spacing ratio coefficient is limited to the range of 0.2 to 0.8. The lower limit of 0.2 ensures that the spacing between the first targets is not too small, causing the centers of the normal and warning categories to become too close and unable to separate. The upper limit of 0.8 ensures that the layout maintains its non-equidistant characteristics.

[0065] For example, in round 20, the first local ratio is 0.38, the second local ratio is 0.25, the first local ratio is greater than the second local ratio, and the increase is... If the current spacing ratio is 0.42, then the adjusted spacing ratio will be... The adjusted spacing ratio coefficient is substituted into the spacing constraint term to update the first target spacing. This dynamic adjustment mechanism takes effect in the next training cycle, enabling the spacing scaling factor to adaptively track changes in boundary separation in the feature space.

[0066] After each training round, the current positions of the three class centers are recorded simultaneously. The displacement of each class center between adjacent rounds is calculated, which is the Euclidean distance between the current round class center position and the previous round class center position. The maximum value of the three class center displacements is taken as the maximum center displacement of the current round.

[0067] The convergence criterion for a feature extraction network is that it satisfies any one of the following: The first condition is that the classification accuracy of the validation set does not improve for multiple consecutive training rounds. In this embodiment, it is set to 10 consecutive rounds. The classification accuracy usually recovers its upward trend within 5 to 8 rounds. 10 rounds are sufficient to determine whether the performance has saturated.

[0068] The second condition is that the maximum center displacement among the adjacent displacements of all category centers in multiple consecutive training rounds is less than a preset convergence threshold. In this embodiment, the maximum center displacement is set to be less than 0.01 for 5 consecutive rounds. The convergence threshold of 0.01 is based on a baseline interval of 8, and 0.01 corresponds to 0.125% of the baseline interval. At this point, the change in the category center position can be ignored, the geometric structure of the feature space is stable, and the requirement of 5 consecutive rounds eliminates the interference of random fluctuations in a single round.

[0069] The third condition is reaching the maximum training epoch limit. The maximum training epoch limit is a well-known parameter in the deep learning model training process. In this embodiment, it is set to 200, that is, training is terminated when the training epoch reaches 200.

[0070] When training terminates due to the second condition, it indicates that the geometric layout of the feature space has reached a steady state and the relative positional relationship of the category centers has stabilized. The first condition monitors whether the classification performance is saturated, and the second condition monitors whether the spatial structure is stable. The two conditions determine the best time to terminate training from different perspectives. After training is completed, the parameters of the feature extraction network and the final positions of the three category centers are saved, which are the final category centers.

[0071] The dynamic adjustment mechanism enables the spacing ratio coefficient to be adaptively adjusted according to the actual separation status of the feature space during training. The local discriminant ratio quantifies the weakness of each boundary from the perspective of high-dimensional feature space, providing targeted feedback signals for the adjustment direction and ensuring that the feature extraction network is fully trained.

[0072] S104, input the supply chain finance data of the merchants to be classified into the converged feature extraction network to obtain new merchant features, calculate the Euclidean distance between the new merchant features and the final category center, and output the risk status label corresponding to the minimum Euclidean distance as the risk classification result.

[0073] In one embodiment, after the feature extraction network converges, it is used to determine the risk level of the merchants to be classified. Supply chain finance data of the merchants to be classified is obtained, including five indicators: debt-to-equity ratio, current ratio, accounts receivable turnover, net profit growth rate, and operating cash flow ratio. Normalization is performed on the supply chain finance data of the merchants to be classified to obtain normalized five-dimensional indicator data.

[0074] Normalized supply chain finance data is input into a converged feature extraction network. After layer-by-layer transformation through the input layer and two hidden layers, a new 64-dimensional merchant feature is obtained. The Euclidean distance between the new merchant feature and the three final class centers is calculated, and the distance from the new merchant feature to the normal class center is obtained. Distance to the center of the warning category and the distance to high-risk category centers The risk status label corresponding to the minimum value among the three Euclidean distances is taken as the risk classification result and output.

[0075] After obtaining the new merchant features, the classification confidence score is further calculated to assess the reliability of the risk grading results. The second-smallest Euclidean distance from the new merchant features to all final category centers is obtained (the second smallest of the three Euclidean distances). The smallest Euclidean distance is divided by the second-smallest Euclidean distance to obtain the distance ratio. The difference between 1 and the distance ratio is calculated to obtain the classification confidence score. The specific expression is:

[0076] In the formula, The minimum Euclidean distance. The distance is the second smallest Euclidean distance, and the classification confidence score ranges from 0 to 1. When the distance between the new merchant feature and the nearest class center is much smaller than the distance between the new merchant feature and the second nearest class center, the distance ratio approaches 0, and the classification confidence score approaches 1, indicating that the classification result is highly reliable. When the new merchant feature is located in the middle of the two nearest class centers, the distance ratio approaches 1, and the classification confidence score approaches 0, indicating that the merchant is in the boundary area between two risk levels, and the classification result is sensitive to small perturbations.

[0077] In one example, suppose the Euclidean distances from a merchant's new merchant features to the three final category centers are respectively , , The minimum Euclidean distance is The second smallest Euclidean distance to the center of the corresponding warning category is The distance ratio is The classification confidence level is Although the merchant was classified as being at the warning level, its new merchant characteristics were almost equidistant from the center of the warning category and the center of the high-risk category, with a classification confidence level of only 0.02.

[0078] In response to a classification confidence level falling below a confidence threshold, a verification tag is generated and output for the merchant to be classified. The confidence threshold is set to 0.1. This threshold is set by statistically analyzing the classification confidence distribution of misclassified samples in the validation set, and taking the 90th percentile of the classification confidence level of the misclassified samples as the confidence threshold, ensuring that 90% of the misclassified samples can be captured by the verification tag and submitted for manual review.

[0079] In the example above, the classification confidence score of 0.02 is lower than the confidence threshold of 0.1, generating a review flag and prompting risk control personnel to conduct a manual review of the merchant.

[0080] For each merchant to be classified, the risk classification result, classification confidence score, and whether a review mark exists are output for risk control personnel to make subsequent credit decisions. Risk control personnel can adopt differentiated processing strategies for different merchants based on the classification confidence score. For example, for merchants with high classification confidence scores, the corresponding credit strategy can be directly implemented based on the risk classification result, while merchants with review marks require manual intervention for secondary review.

[0081] It should be noted that the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for constructing a risk status classification model for supply chain finance, characterized in that, The construction method includes: S101, acquire supply chain finance data and corresponding risk status labels of multiple merchants, calculate the statistical difference between the supply chain finance data of adjacent risk status labels, and obtain the spacing ratio coefficient. S102, In the feature space mapped by the feature extraction network, the geometric spacing of each category center is constrained according to the spacing ratio coefficient to construct a non-equidistant loss. S103, the feature extraction network is iteratively trained using the non-equidistant loss, and the spacing ratio coefficient is adaptively adjusted according to the local discriminant ratio of adjacent boundaries in the feature space during the iterative training until the feature extraction network converges to obtain the final category center; S104, input the supply chain finance data of the merchants to be classified into the converged feature extraction network to obtain new merchant features, calculate the Euclidean distance between the new merchant features and each final category center, and take the risk status label corresponding to the minimum Euclidean distance as the risk classification result.

2. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, After acquiring supply chain finance data from multiple merchants, the construction method further includes: Obtain the minimum and maximum values ​​of all supply chain finance data in the historical dataset; Subtract the minimum value from the supply chain finance data and divide by the difference between the maximum value and the minimum value to obtain the normalized index; The normalized index will be used as the supply chain finance data for subsequent calculations.

3. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, The risk status labels include a first category, a second category, and a third category; the method of obtaining the spacing ratio coefficient includes: for any single indicator in the supply chain finance data, obtaining the first standard deviation and the second standard deviation corresponding to two adjacent risk status labels; and using the sample size of the two risk status labels to perform a weighted summation of the first standard deviation and the second standard deviation to obtain the combined standard deviation. Calculate the absolute value of the difference between the means of the individual indicators under the two risk status labels, divide the absolute value by the pooled standard deviation to obtain the effect size index; calculate the average value of the effect size index of each individual indicator to obtain the statistical difference between the adjacent boundaries of the two risk status labels. The spacing ratio coefficient is obtained by dividing the statistical difference between the adjacent boundaries of the first and second categories by the statistical difference between the adjacent boundaries of the second and third categories.

4. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, The geometric spacing between the centers of each category is constrained according to the spacing ratio coefficient to construct a non-equidistant loss, including: obtaining the intra-class clustering term that causes samples of the same type to cluster toward their respective category centers; Based on the aforementioned spacing ratio coefficient, a spacing constraint term is constructed to determine the deviation of the actual category center spacing from the target spacing; based on the degree to which the category centers deviate from a straight line in the feature space, a collinearity regularization term is constructed. The non-equidistant loss is obtained by weighting and summing the intra-class clustering term, the spacing constraint term, and the collinearity regularization term.

5. The method for constructing a risk status classification model for supply chain finance according to claim 4, characterized in that, The method for obtaining the collinear regularization term is as follows: Obtain a first vector pointing from the center of the first category to the center of the second category, and a second vector pointing from the center of the first category to the center of the third category; Calculate the product of the magnitudes of the first vector and the second vector; calculate the absolute value of the inner product between the first vector and the second vector; use the difference between the product of the magnitudes and the absolute value of the inner product as a collinearity deviation metric. The collinearity deviation metric is divided by the maximum distance between any two of the three category centers to obtain the collinearity regularization term.

6. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, Before iteratively training the feature extraction network using the non-equidistant loss, the construction method further includes: The initial center of the first category is set at the origin of the feature space; The initial center of the second category is set at a position along the positive direction of a specific coordinate axis, at a distance from the origin equal to the first target distance. The initial center of the third category is set at a position along the positive direction of the specific coordinate axis, at a distance from the origin equal to the total target distance. The first target distance and the total target distance are determined based on the distance ratio coefficient.

7. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, The method for obtaining the local discrimination ratio is as follows: Calculate the average Euclidean distance from the feature vectors of all samples in the first class on both sides of any adjacent boundary to the center of the corresponding class, and use it as the first divergence. Calculate the average Euclidean distance from the feature vectors of all samples in the second category on both sides of the adjacent boundary to the center of the corresponding category, and use it as the second divergence; Calculate the Euclidean distance between the class centers of the first category and the second category, and use it as the inter-class separation degree; Divide the mean of the first divergence and the second divergence by the inter-class separation ratio to obtain the local discriminant ratio.

8. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, The adjacent boundaries include a first boundary between the first category and the second category, and a second boundary between the second category and the third category; the spacing ratio coefficient is adaptively adjusted based on the local discriminant ratio of the adjacent boundaries in the feature space, including: The local discrimination ratio of the first boundary is used as the first local ratio, and the local discrimination ratio of the second boundary is used as the second local ratio. In response to the first local ratio being greater than the second local ratio, an increment is obtained by multiplying the learning rate coefficient by the difference between the two, and the increment is superimposed on the spacing ratio coefficient; In response to the second local ratio being greater than the first local ratio, the reduction is obtained by multiplying the learning rate coefficient by the difference between the two, and the reduction is subtracted from the spacing ratio coefficient.

9. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, The convergence criterion for the feature extraction network is that it satisfies any one of the following: The first condition is that the classification accuracy of the validation set does not improve in multiple consecutive training rounds; The second condition is that, within multiple consecutive training rounds, the maximum center displacement among the adjacent round displacements of all the category centers is less than a preset convergence threshold.

10. The method for constructing a risk status classification model for supply chain finance according to claim 1, characterized in that, After obtaining the new merchant characteristics, the construction method further includes: Obtain the second smallest Euclidean distance from the new merchant features to all final category centers; The distance ratio is obtained by dividing the minimum Euclidean distance by the second smallest Euclidean distance. The difference between 1 and the distance ratio is used as the classification confidence score; In response to the classification confidence level being lower than the confidence threshold, a verification tag for the merchant to be classified is generated and output.

Images