Risk control AUC optimization adaptive enhancement tree model establishment method

Abstract

The invention discloses a risk control AUC optimization adaptive enhancement tree model establishment method. The model establishment method comprises the following steps: establishing a decision tree model; original data are collected, data preprocessing is carried out, and preprocessed data are obtained; and training the decision tree model by adopting a small sample learning technology and a sample weighting mechanism, and obtaining a risk control AUC optimization adaptive enhancement tree model by adopting an attention mechanism enhancement model. And through small sample learning and a dynamic sample weighting mechanism, the AUC value of the model is optimized, and the prediction performance is improved especially under the conditions of class imbalance and sample scarcity.

Description

Technical Field

[0001] The present invention relates to the field of risk control, and in particular to a method for establishing an AUC-optimized adaptive boosting tree model for risk control. Background Art

[0002] The technical problems existing in the existing technology include:

[0003] Small sample learning and overfitting problems: Traditional decision trees are prone to overfitting when processing small sample data, resulting in insufficient model generalization ability and inability to effectively cope with sparse data environments, especially in areas such as financial credit.

[0004] Data imbalance and AUC optimization challenges: When data categories are severely imbalanced, traditional models find it difficult to effectively improve the AUC value, resulting in a decrease in the model's ability to distinguish minority class samples and an excessive bias towards the majority class.

[0005] Computational complexity and training efficiency: Models that combine deep learning attention mechanisms usually introduce a lot of computational overhead, resulting in a significant increase in training time, and performance cannot be taken into account in application scenarios with large-scale datasets or small sample data.

[0006] Problem of dynamic weighting of feature importance: Existing models are unable to dynamically adjust weights based on the risk distribution or feature weights of the sample, resulting in insufficient attention paid to key features and reducing the model's adaptability to different sample types.

[0007] Existing technologies use ensemble learning: using methods such as bagging, boosting, and stacking to improve the generalization ability of the model, but there is still a bias problem for small sample data, and the model is difficult to effectively deal with sparse samples.

[0008] Weighted optimization: AUC is optimized by adjusting sample weights. However, traditional weighting strategies perform poorly in response to dynamic changes in risk characteristics, making it difficult for the model to focus on key features.

[0009] Attention-based boosted tree model: Some studies have attempted to introduce the attention mechanism into the ensemble learning model. However, the dynamic weighting strategy of the attention mechanism has not been fully combined with the AUC-optimized tree, and the model still faces a trade-off between generalization ability and computational complexity.

[0010] shortcomings of existing technologies;

[0011] Existing decision tree-based models perform well in big data scenarios. However, in small-sample learning scenarios, due to the limited number of samples and the high dimensionality of the feature space, the models are prone to overfitting and fail to fully capture the true characteristics of the data distribution. Small sample data is often noisy, resulting in reduced model generalization ability. This is particularly true in the fields of financial credit and credit risk assessment, where existing models perform poorly when the number of samples is insufficient.

[0012] While some studies have adopted sample weighting strategies to handle imbalanced data, existing weighting strategies are typically static and difficult to dynamically adjust based on sample characteristics, resulting in insufficient model recognition for minority samples. In a small-sample learning environment, sample imbalance exacerbates model bias and further reduces the model's AUC performance.

[0013] While models incorporating attention mechanisms improve accuracy and generalization, they significantly increase computational complexity, resulting in excessively long training times. Especially with small sample sizes, the high computational cost makes the model difficult to apply in real-world business scenarios, limiting its practical feasibility.

[0014] Current AUC optimization mainly adjusts data weights through static weighting or simple sampling methods, but these methods cannot adjust weights in real time according to the dynamic characteristics of samples, resulting in the model being unable to maintain stable AUC performance under different data distributions. Summary of the Invention

[0015] In view of the above problems, the present invention is proposed to provide a method for establishing an AUC optimized adaptive boosting tree model for risk control that overcomes the above problems or at least partially solves the above problems.

[0016] According to one aspect of the present invention, a method for establishing an AUC-optimized adaptive boosting tree model for risk control is provided, the model establishment method comprising:

[0017] Build a decision tree model;

[0018] Collect raw data and perform data preprocessing to obtain preprocessed data;

[0019] The small sample learning technology and sample weighting mechanism are used to train the decision tree model, and the attention mechanism is used to enhance the model to obtain the AUC optimized adaptive enhanced tree model for risk control.

[0020] Optionally, the collecting of raw data and performing data preprocessing to obtain preprocessed data specifically includes: cleaning, feature selection and engineering of the raw data set.

[0021] Optionally, the use of small sample learning technology and sample weighting mechanism to train the decision tree model specifically includes:

[0022] Initialize attention model and decision tree parameters;

[0023] Training the attention model;

[0024] Build a decision tree;

[0025] Make predictions for each decision tree and obtain probability values;

[0026] Average the predictions of all trees to get the final probability;

[0027] Output the category label according to the probability threshold.

[0028] Optionally, constructing the decision tree specifically includes:

[0029] Perform weighted sampling according to the sample weight to obtain the data set after sample resampling;

[0030] A plurality of decision trees are trained on the resampled data set.

[0031] Optionally, the training attention model specifically includes:

[0032] Calculate category weights;

[0033] Calculate sample weights;

[0034] Train the attention model using the AUC loss function:

[0035] Forward propagation: calculate model output; backpropagation: update model parameters.

[0036] Optionally, the training attention model specifically includes:

[0037] The multi-head attention mechanism captures the relationship between input features and then maps the attention output to the target dimension through a fully connected layer;

[0038] Among them, the attention model output formula is: output = AttentionModel(X);

[0039] Among them, AttentionModel is a neural network model based on a multi-head attention mechanism, which is used to calculate the weight of the sample. The input is the feature matrix X, and the output is the weight of each sample.

[0040] The neural network model of the multi-head attention mechanism includes: a multi-head attention layer to capture the relationship between input features; and a fully connected layer to map the output of the attention mechanism to the target dimension.

[0041] Optionally, the multi-head attention layer specifically includes:

[0042] Input: Input feature matrix Where n is the number of samples and d is the feature dimension;

[0043] Linear transformation: Perform linear transformation on the input features to obtain query, key and value:

[0044] Q=XW Q , K=XW K , V=XW V

[0045] in, is the learned weight matrix, d k is the dimension of each head;

[0046] Calculate the attention score matrix A:

[0047]

[0048] Among them, softmax is used to normalize the attention score;

[0049] Weighted summation: Perform weighted summation on the value V to obtain the attention output:

[0050] Attention(Q, K, V)=AV.

[0051] Optionally, the fully connected layer specifically includes:

[0052] The first layer is fully connected:

[0053] h=ReLU(W1·Attention(Q, K, V)+b1)

[0054] in, is the weight matrix, b1 is the bias term, and ReLU is the activation function;

[0055] The second fully connected layer: y = W2 h + b2;

[0056] in, is the weight matrix, b2 is the bias term;

[0057] The sample weight calculation formula is:

[0058] sample_weights=class_weights×attention_weights

[0059] Among them, class_weights is the category weight, balancing the category imbalance problem;

[0060] attention_weights is the weight of the attention model output, focusing on difficult samples;

[0061] The calculation formula of the category weight is:

[0062]

[0063] Among them, N is the total number of samples; N0 is the number of negative samples; N1 is the number of positive samples.

[0064] Optionally, the use of the AUC loss function to train the attention model specifically includes:

[0065] The AUC loss function is used to optimize the model's ability to distinguish between positive and negative samples. This time, logarithmic loss is used to indirectly optimize AUC. The logarithmic loss formula is as follows:

[0066]

[0067] Where: y i is the true label of the sample (0 or 1); p i is the probability predicted by the model; N is the number of samples;

[0068] The AUC loss function formula is:

[0069]

[0070] Among them, N + is the number of positive samples; N - is the number of negative samples; p i is the predicted probability of the positive sample; p j is the predicted probability of negative samples.

[0071] Optionally, the sample resampling specifically includes:

[0072] Weighted sampling is used to enhance the weight of difficult samples. The formula is:

[0073]

[0074] Among them, γ is the weight coefficient of difficult samples; when sampling, weighted random sampling is performed according to sample_weights.

[0075] The present invention provides a method for establishing an AUC-optimized adaptive boosting tree model for risk control. The method comprises: establishing a decision tree model; collecting raw data and performing data preprocessing to obtain preprocessed data; training the decision tree model using small sample learning techniques and a sample weighting mechanism, and enhancing the model using an attention mechanism to obtain an AUC-optimized adaptive boosting tree model for risk control. Through small sample learning and a dynamic sample weighting mechanism, the model's AUC value is optimized, particularly improving predictive performance in situations of class imbalance and sample scarcity.

[0076] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0078] Figure 1 A flowchart of a method for establishing an AUC-optimized adaptive boosting tree model for risk control provided by an embodiment of the present invention;

[0079] Figure 2 Schematic diagram of the test results of the original training data and the traditional DecisionTree generation model provided by the embodiment of the present invention;

[0080] Figure 3 Schematic diagram of test results of enhanced training data and traditional DecisionTree generation model provided by an embodiment of the present invention;

[0081] Figure 4 Schematic diagram of the test results of the original training data and the AUCOptimizedTrees generation model provided in an embodiment of the present invention;

[0082] Figure 5 Schematic diagram of the test results of the enhanced training data and AUCOptimizedTrees generation model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0083] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0084] The terms "comprises" and "comprising" and any variations thereof in the description, embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0085] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0086] The main objectives of the present invention are as follows:

[0087] 1. Propose a decision tree model that combines AUC optimization and attention mechanism.

[0088] 2. Through small sample learning and dynamic sample weighting mechanism, the model's AUC value is optimized, especially to improve prediction performance in cases of class imbalance and sample scarcity.

[0089] 3. Conduct experiments on real datasets to verify the effectiveness of the proposed model in small-sample learning scenarios.

[0090] The following steps are adopted to achieve the target effect of the present invention:

[0091] 1. Model Design: Design and implement a model that combines the AUC optimization tree with the attention mechanism to optimize the decision tree training process and enhance the model's adaptability in small-sample learning.

[0092] 2. Data preprocessing: Cleaning, feature selection, and engineering of the original data set to ensure that the data is suitable for model training, especially in small sample learning, to ensure data quality and representativeness of features.

[0093] 3. Training and optimization: Decision tree models are trained using small sample learning techniques and sample weighting mechanisms, and the attention mechanism is used to enhance the model's ability to focus on important features.

[0094] like Figure 1 As shown, the present invention provides a method for establishing an AUC optimized adaptive boosting tree model for risk control, the model establishment method comprising:

[0095] Build a decision tree model;

[0096] Collect raw data and perform data preprocessing to obtain preprocessed data;

[0097] The small sample learning technology and sample weighting mechanism are used to train the decision tree model, and the attention mechanism is used to enhance the model to obtain the AUC optimized adaptive enhanced tree model for risk control.

[0098] The algorithm flow is as follows:

[0099] 1) Initialization

[0100] Initialize the attention model AttentionModel.

[0101] Initialize decision tree parameters.

[0102] 2) Training the attention model

[0103] Calculate class weights class_weights.

[0104] Calculate sample weights sample_weights.

[0105] The attention model is trained using the AUC loss function: forward propagation: calculating the model output; backpropagation: updating the model parameters.

[0106] 3) Build a decision tree

[0107] Weighted sampling is performed according to the sample weights to obtain a resampled data set, solving the data scarcity problem in small sample learning.

[0108] Train multiple decision trees on the resampled dataset.

[0109] 4) Prediction

[0110] Make predictions for each decision tree and get the probability value.

[0111] The predictions of all trees are averaged to get the final probability.

[0112] Output class labels based on probability thresholds.

[0113] 4. Evaluate model performance using AUC and KS metrics and compare it with the traditional decision tree model, especially in small sample learning environments, focusing on verifying the robustness and accuracy of the model.

[0114] Effects that can be achieved by the invention:

[0115] 1. Propose and implement a model that combines AUC optimization tree with attention mechanism, with special attention paid to the optimization effect in small-sample learning.

[0116] 2. Verify the superiority of the model in terms of indicators such as AUC value and KS value on the dataset, especially demonstrating its advantages in small sample scenarios.

[0117] 3. Provide a sample weighting mechanism to optimize AUC, which can effectively solve the problems of sample imbalance and small sample learning.

[0118] The AUC-Optimized Adaptive Augmented Ensemble Trees (AUC-AAET) model combines decision trees and an attention mechanism. First, the model weights the input features using the attention mechanism, dynamically adjusting the importance of each sample in training. This allows the model to focus on a small number of key samples, particularly in small-sample learning. Next, the model is trained using decision trees, using dynamically weighted samples to optimize the AUC when constructing each tree. During training, the model's performance in binary classification problems is gradually improved by optimizing sample weights and AUC loss, ensuring both stability and accuracy, especially when the sample size is small.

[0119] The attention mechanism is used to dynamically adjust sample weights, focusing on samples that are difficult to classify (i.e., difficult samples), thereby improving classification performance. The model captures the relationship between input features through a multi-head attention mechanism, and then maps the attention output to the target dimension through a fully connected layer. The output formula of its attention model is: output = AttentionModel(X). Among them, AttentionModel is a neural network model based on a multi-head attention mechanism (MultiheadAttention), which is used to calculate the weight of the sample. The input is the feature matrix X, and the output is the weight of each sample. Its structure includes the following parts: MultiheadAttention layer (MultiheadAttention): used to capture the relationship between input features; Fully Connected Layers (Fully Connected Layers): used to map the output of the attention mechanism to the target dimension.

[0120] 1. Multi-head attention mechanism layer

[0121] 1) Input:

[0122] Input feature matrix Where n is the number of samples and d is the feature dimension.

[0123] 2) Linear transformation:

[0124] Perform a linear transformation on the input features to obtain the query, key, and value:

[0125] Q=XW Q , K=XW K , V=XW V

[0126] in, is the learnable weight matrix, d k is the dimension of each head.

[0127] 3) Attention score:

[0128] Calculate the attention score matrix A:

[0129]

[0130] Among them, softmax is used to normalize the attention score.

[0131] 4) Weighted summation:

[0132] Perform a weighted summation on the value V to get the attention output:

[0133] Attention(Q, K, V) = AV

[0134] 2. Fully connected layer

[0135] 1) First layer full connection:

[0136] h=ReLU(W1·Attention(Q, K, V)+b1)

[0137] in, is the weight matrix, b1 is the bias term, and ReLU is the activation function.

[0138] 2) Second layer full connection:

[0139] y=W2·h+b2

[0140] in, is the weight matrix and b2 is the bias term.

[0141] 3. The sample weight calculation formula is:

[0142] sample_weights=class_weights×attention_weights

[0143] Among them: class_weights is the class weight, which is used to balance the class imbalance problem; attention_weights is the weight output by the attention model, which is used to focus on difficult samples.

[0144] The AUC loss function is used to optimize the model's ability to distinguish between positive and negative samples. This time, logarithmic loss is used to indirectly optimize AUC. The logarithmic loss formula is as follows:

[0145]

[0146] Where: y i is the true label of the sample (0 or 1); p i is the probability predicted by the model; N is the number of samples.

[0147] The AUC loss function formula is:

[0148]

[0149] Where: N + is the number of positive samples; N - is the number of negative samples; p i is the predicted probability of the positive sample; p j is the predicted probability of negative samples.

[0150] Class weight is used to balance the class imbalance problem. The formula is:

[0151]

[0152] Where: N is the total number of samples; N0 is the number of negative samples; N1 is the number of positive samples.

[0153] Sample resampling introduces a weighting mechanism for difficult samples. Difficult samples are those that are located near the decision boundary and are difficult to classify. By assigning higher training weights to difficult samples, the algorithm can pay more attention to these boundary samples during training, improving classification performance. The weight of difficult samples is increased through weighted sampling. The formula is:

[0154]

[0155] Where: γ is the weight coefficient of difficult samples; weighted random sampling is performed according to sample_weights during sampling.

[0156] Sample limitations

[0157] 1) The sales invoice amount in the past 12 months is greater than RMB 100,000;

[0158] 2) The most recent invoicing record is no more than 3 months from the time of data collection;

[0159] 3) Data collection is from January 2024 or earlier. Data collection is done as close to the present time as possible, while meeting the overall sample requirements (industry, region, etc.).

[0160] Data preprocessing and feature processing include: for the original data, first perform data cleaning, including missing value processing, outlier detection, etc.; then perform feature selection and engineering, select important features and perform data enhancement processing to ensure the effectiveness of model training.

[0161] The prediction label is defined as whether the company has ever had a record of overdue payments exceeding 30 days in history. If the same customer has two credit application records, the record with the bad prediction label is used. If both are bad or both are good, the record with the later prediction label is used.

[0162] Data cleaning and processing were performed to remove features with null value rates greater than 70%.

[0163] Dataset partitioning and data augmentation

[0164] The data is divided into a training set and a test set. The training set and test set each contain 500 records, with a good-to-bad ratio of 4:1. The number of real samples that meet the project requirements does not exceed 500.

[0165] Data augmentation: The ADA-TABDDPM algorithm was used to augment the training set. The augmented training set consisted of 10,390 records, with a good-to-bad ratio of 0.51:0.49, meeting the project requirements.

[0166] The test set is usually used to evaluate the real performance of the model without data augmentation and retaining its real distribution.

[0167] The DecisionTree algorithm was used to sort the features by importance and select the modeling features. The feature index names and descriptions are as follows:

[0168]

[0169]

[0170]

[0171] 1. Category Description

[0172] The characteristic variables in the business performance model are divided into the following seven categories, each of which reflects a different aspect of the business activities:

[0173] 1) Abnormal signal

[0174] What it means: Focuses on indicators that may be associated with unusual behavior, often used for risk identification.

[0175] Number of features: 1

[0176] 2) Income Capacity

[0177] Meaning: Evaluates a company's revenue scale, operating activity and stability, and reflects its financial status.

[0178] Number of features: 6

[0179] 3) Downstream customers

[0180] Meaning: Analyze the relationship between a company and its counterparties (customers), including customer concentration, stability, and transaction size.

[0181] Number of features: 17

[0182] 4) Income Fluctuations

[0183] Meaning: Measures the seasonality, growth trend, and stability of a company's revenue.

[0184] Number of features: 6

[0185] 5) Business activity

[0186] Meaning: Evaluates a company's operating history and frequency of activities to reflect its business continuity.

[0187] Number of features: 2

[0188] 6) Operating radius

[0189] Meaning: Measures a company's market coverage and operational diversity.

[0190] Number of features: 3

[0191] 7) Cancel the red check

[0192] Meaning: Pay attention to the cancellation and red cancellation of invoices to reflect the transaction quality and potential risks.

[0193] Number of features: 2

[0194] 2. Relationships between major categories

[0195] 1) Earning capacity and income fluctuation

[0196] Both are related to a company's financial performance. Earnings capacity focuses on the static size of revenue, while revenue volatility focuses on dynamic trends and stability.

[0197] 2) Downstream customers and operating radius

[0198] Downstream customers focus on the specific characteristics of counterparties, while operating radius assesses market coverage from a geographical and industry perspective. Together, they reflect a company's customer base and market size.

[0199] 3) Abnormal signal and invalid red rush

[0200] Both are related to risk. Abnormal signals detect potential abnormal behavior, while voided red strikes reflect transaction quality issues.

[0201] 4) Business activity and other categories

[0202] As a basic indicator, business activity is indirectly related to major categories such as revenue capacity and downstream customers, and reflects the company's ability to continue operating.

[0203] The relationship shows that the model comprehensively evaluates the company's performance capability through multi-dimensional characteristics, focusing on both financial performance and customer relationships and potential risks.

[0204] 3. Detailed description of features

[0205] 1) Abnormal signal

[0206] Chinese name: L1-L6 nighttime ticket issuance ratio

[0207] Business Implications: This value indicates the ratio of invoices issued at night to the total number of invoices issued during the six months prior to the observation time point and the month prior to the observation time point. Nighttime invoicing may signal unusual behavior, such as tax evasion or fraud.

[0208] 2) Income Capacity

[0209] Chinese name: Average number of valid tickets per month for L1-L3 (based on the actual number of ticket issuance months)

[0210] Business significance: Calculate the average number of valid invoices per month from the previous three months to the previous month to reflect short-term business activity and revenue capacity.

[0211] Chinese name: L1-L3 monthly average effective ticket amount - natural logarithm (based on the actual number of ticket issuance months)

[0212] Business Implications: Calculate the natural logarithm of the average monthly valid ticket amount from the previous three months to the previous month to smooth the data distribution and reflect revenue stability.

[0213] Chinese name: L19 total amount of valid tickets - natural logarithm

[0214] Business significance: It represents the natural logarithm of the total amount of valid tickets in the previous 19 months, reflecting the historical revenue scale.

[0215] Chinese name: Total number of L4 valid tickets

[0216] Business significance: Indicates the number of valid invoices in the previous four months, reflecting the business activity during that period.

[0217] Chinese name: L14 total amount of valid tickets

[0218] Business meaning: Indicates the total amount of valid tickets in the previous 14 months, reflecting the revenue scale during that period.

[0219] Chinese name: Average monthly valid ticket amount for L1-L24 (based on the actual number of ticket issuance months)

[0220] Business significance: Calculate the average monthly valid ticket amount from the previous 24 months to the previous month to reflect the long-term income stability.

[0221] 3) Downstream customers

[0222] Chinese name: L13-L15 The proportion of the total amount of valid tickets of the top ten counterparties

[0223] Business significance: It reflects the proportion of the top ten counterparties in terms of transaction amount from the previous 15 months to the previous 13 months, indicating customer concentration.

[0224] Chinese name: Average establishment months of the top ten counterparties (enterprises) from L1 to L12

[0225] Business Implications: Calculate the average establishment time of the top ten trading counterparties from the previous 12 months to the previous month to assess customer maturity.

[0226] Chinese name: Month-on-month growth in the number of L7-L9 counterparties (enterprises)

[0227] Business Implications: Compares the month-over-month changes in the number of trading counterparties from the first nine months to the first seven months to reflect the dynamics of the customer base.

[0228] Chinese name: Number of overlapping L7-L12 top three trading counterparties on a month-on-month basis

[0229] Business significance: Calculates the number of overlaps between the top three trading counterparties from the previous 12 months to the previous 7 months to reflect the stability of major customers.

[0230] Chinese name: Logarithm of the total amount of valid tickets of the top five trading counterparties (L13-L24) month-on-month (different counterparties)

[0231] Business significance: Calculate the logarithm of the month-on-month growth in the transaction amounts of the top five counterparties from the previous 24 months to the previous 13 months to reflect transaction dynamics.

[0232] Chinese name: L22-L24 Total amount of valid tickets of the top ten counterparties

[0233] Business significance: Indicates the bill amounts of the top ten trading counterparties in the previous 24 to 22 months, reflecting the transaction scale.

[0234] Chinese name: Total effective bill amount of the top three L4-L6 counterparties year-on-year (different counterparties)

[0235] Business Implications: Compares the year-on-year growth in the transaction amounts of the top three counterparties from the first six months to the first four months with the same period last year, reflecting annual changes.

[0236] Chinese name: L13-L18 The total amount of valid tickets of the largest counterparty (locked molecular counterparty)

[0237] Business significance: Calculates the month-on-month change in the transaction amount with the largest counterparty from the previous 18 months to the previous 13 months to reflect the dynamics of the largest customer.

[0238] Chinese name: Number of overlapping L7-L9 top three trading counterparties year-on-year

[0239] Business significance: Calculates the number of overlaps between the top three trading counterparties in the first nine months to the first seven months and the same period last year, reflecting annual stability.

[0240] Chinese name: Year-on-year growth in the number of L1-L6 counterparties (individuals)

[0241] Business Implications: Comparing the year-on-year growth in the number of individual trading counterparties from the first six months to the first month compared with the same period last year, reflecting changes in individual customers.

[0242] Chinese name: Number of overlapping L1-L3 top five trading counterparties year-on-year

[0243] Business significance: Calculates the number of overlaps between the top five trading counterparties in the previous three months and the previous month and the same period last year, reflecting short-term stability.

[0244] Chinese name: Total amount of valid tickets of the top five counterparties L7-L9

[0245] Business significance: Indicates the bill amounts of the top five trading counterparties from the previous 9 to 7 months, reflecting the transaction scale.

[0246] Chinese name: L13-L15 The proportion of the total amount of valid bills of the top three trading counterparties

[0247] Business significance: It reflects the proportion of the transaction amount of the top three counterparties in the first 15 to 13 months, indicating the dependence on major customers.

[0248] Chinese name: L28-L30 HHI coefficient of the effective bill amount of the top ten trading counterparties

[0249] Business Implications: Calculates the HHI index of the top ten counterparties by transaction amount from the previous 30 months to the previous 28 months to reflect customer concentration.

[0250] Chinese name: L13-L24 HHI coefficient of the effective bill amount of all counterparties

[0251] Business Implications: Calculates the HHI index of all counterparty amounts from the previous 24 months to the previous 13 months to reflect transaction diversity.

[0252] Chinese name: L10-L12 Month-on-month loss of all trading counterparties

[0253] Business significance: Calculates the number of counterparty churn from the previous 12 months to the previous 10 months to reflect customer stability.

[0254] Chinese name: L28-L30 number of counterparties (enterprises)

[0255] Business Implications: Compares the month-over-month changes in the number of counterparties from the previous 30 months to the previous 28 months to reflect long-term customer dynamics.

[0256] 4) Income Fluctuations

[0257] Chinese name: The total number of valid tickets in the first half of the year compared with the same period last year

[0258] Business Implications: Compares the year-on-year growth in the number of valid tickets in the first half of the year with that in the same period last year, reflecting seasonal fluctuations.

[0259] Chinese name: Average monthly valid ticket amount for L1-L12 (based on the number of natural months) / Average monthly valid ticket amount for L1-L36 (based on the number of natural months)

[0260] Business Implications: Compares the ratio of the average monthly ticket amount in the previous 12 months with that in the previous 36 months to reflect recent revenue fluctuations.

[0261] Chinese name: Month-on-month average number of valid tickets for L31-L33 (based on the actual number of ticket issuance months)

[0262] Business significance: Calculate the month-on-month change in the average monthly number of tickets from the previous 33 months to the previous 31 months to reflect the long-term trend.

[0263] Chinese name: L1-L6 valid ticket number variation coefficient month-on-month

[0264] Business significance: Calculate the month-on-month change in the coefficient of variation of the number of tickets from the previous 6 months to the previous month to reflect revenue stability.

[0265] Chinese name: Average number of valid tickets per month in the third quarter of the year (based on the number of natural months) compared with the same period last year

[0266] Business significance: Compares the year-on-year growth in the average monthly ticket volume in the third quarter of the current year with the same period last year, reflecting quarterly fluctuations.

[0267] Chinese name: The total number of valid tickets in the first quarter of the year

[0268] Business significance: Compare the year-on-year growth in the number of valid tickets in the first quarter of the current year with that in the same period last year, reflecting quarterly changes.

[0269] 5) Business activity

[0270] Chinese name: The number of invoices from the earliest date to today (as of the end of the month)

[0271] Business significance: Calculates the number of days from the earliest invoice to the observation time point, reflecting the business operation history of the enterprise.

[0272] Chinese name: Coefficient of variation of L1-L6 invoicing days

[0273] Business significance: Calculates the coefficient of variation of the number of invoicing days from the previous six months to the previous month to reflect the stability of operating frequency.

[0274] 6) Operating radius

[0275] Chinese name: L31-L33 Number of intra-provincial trading counterparties / Number of extra-provincial trading counterparties

[0276] Business significance: Calculate the ratio of the number of trading counterparties within the province to the number of trading counterparties outside the province from the previous 33 months to the previous 31 months to reflect the market scope.

[0277] Chinese name: L13-L18 Number of industries involved by counterparties (Level 3 industries)

[0278] Business implications: Calculate the number of third-level industries involved by counterparties in the previous 18 to 13 months to reflect business diversity.

[0279] Chinese name: L34-L36 Number of intra-provincial trading counterparties / Number of extra-provincial trading counterparties

[0280] Business significance: Calculate the ratio of the number of intra-provincial and extra-provincial counterparties from the previous 36 months to the previous 34 months to reflect long-term market coverage.

[0281] 7) Cancel the red check

[0282] Chinese name: Top three trading counterparties for L1-L12, number of trading counterparties issuing red invoices for L1-L6

[0283] Business significance: Calculates the number of red cancellation invoices issued in the first six months among the top three trading counterparties in the first 12 months to reflect transaction risk.

[0284] Chinese name: Coefficient of variation of the number of voided invoices from L1 to L12

[0285] Business significance: Calculates the coefficient of variation of the number of voided invoices from the previous 12 months to the previous month to reflect the stability of voiding.

[0286] The Herfindahl-Hirschman Index (HHI) measures market or customer concentration. Higher values indicate greater market or customer concentration and greater risk concentration; lower values indicate more diversified markets or customers and lower risk. In enterprise risk management, the HHI can be used to analyze a company's reliance on a small number of counterparties.

[0287] 1) HHI coefficient calculation formula

[0288]

[0289] n: number of trading counterparties (e.g., top ten trading counterparties, n=10); s i : The proportion of the amount of the i-th counterparty, calculated as follows:

[0290]

[0291] Among them, A i is the effective ticket amount of the i-th counterparty, Atotal It is the sum of the valid ticket amounts of the top ten trading counterparties.

[0292] 2) Numerical interpretation

[0293] HHI value range: 0≤HHI≤1.

[0294] The closer the HHI is to 1, the more concentrated the transaction amount is among a few counterparties, indicating a high concentration risk.

[0295] The closer the HHI is to 0, the more evenly the transaction amount is distributed and the lower the concentration risk.

[0296] 4. Introduction to Coefficient of Variation

[0297] The coefficient of variation (CV) measures the relative dispersion of data, representing the ratio of the standard deviation to the mean. A larger CV indicates greater data volatility and less stability; a smaller CV indicates more stable data. In revenue or ticket count analysis, the CV can be used to assess the volatility of revenue or transaction volume.

[0298] 1) Calculation formula for coefficient of variation

[0299]

[0300] σ: sample standard deviation, reflecting the degree of dispersion of the data; μ: sample mean, reflecting the average level of the data.

[0301] 2) Standard deviation formula

[0302]

[0303] x i : The value of the i-th month (such as the number of valid tickets); n: the number of samples; μ: the mean.

[0304] 3) Numerical interpretation

[0305] CV>0: The coefficient of variation increases and the volatility increases.

[0306] CV<0: The coefficient of variation decreases and the volatility decreases.

[0307] CV=0: No significant change in volatility.

[0308] During the training process, the sample weights are first initialized and dynamically weighted using the attention mechanism. The model is then trained using the decision tree algorithm and optimized using the AUC loss function.

[0309] Libraries and model tools used for modeling, including data processing, modeling, evaluation, and visualization modules, support a variety of machine learning algorithms and deep learning frameworks.

[0310] 1) Data processing tools

[0311] pandas: data reading and processing.

[0312] numpy: numerical calculations and array operations.

[0313] 2) Machine Learning Model Tools

[0314] RandomForestClassifier: Random forest classification model.

[0315] BaggingClassifier: Bagging integration method.

[0316] AdaBoostClassifier: Adaptive boosting model.

[0317] DecisionTreeClassifier: Decision tree classification model.

[0318] XGBClassifier: XGBoost gradient boosting model.

[0319] LGBMClassifier: LightGBM gradient boosting model.

[0320] LogisticRegression: Logistic regression.

[0321] 3) Model evaluation tools

[0322] accuracy_score: Accuracy.

[0323] recall_score: recall rate.

[0324] precision_score: precision rate.

[0325] f1_score:F1 score.

[0326] roc_auc_score: AUC score.

[0327] classification_report: comprehensive classification report.

[0328] GridSearchCV: Grid Search Cross Validation.

[0329] roc_curve: draw ROC curve.

[0330] 4) Visualization Tools

[0331] matplotlib.pyplot: for plotting and visualizing model results.

[0332] The SimHei font supports Chinese display, and axes.unicode_minus ensures that the minus sign is displayed normally.

[0333] 5) Deep Learning Tools

[0334] torch: The main deep learning framework.

[0335] torch.nn: used to build neural networks.

[0336] torch.optim: Optimizer module for model parameter optimization.

[0337] The present invention first uses the ADA-TABDDPM algorithm to generate synthetic data, enhances the original data set, and then inputs the enhanced data set into the AUCOptimizedTrees algorithm model for training. Finally, it generates a credit risk commercial performance model with better performance for SMEs in the manufacturing industry.

[0338] In this modeling process, the original training data and the enhanced training data were used to evaluate the performance of the model on the test dataset.

[0339] The model trained using augmented data performed significantly better on the test set than the model trained using original data; the model trained using the AUCOptimizedTrees algorithm performed significantly better on the test set than the traditional DecisionTree algorithm.

[0340] The model performance was evaluated using AUC and KS indicators, and compared with the traditional decision tree model.

[0341] The experimental verification results are as follows:

[0342] like Figure 2 As shown, the test results of the original training data and the traditional DecisionTree generation model.

[0343] like Figure 3 As shown in Figure 2, the test results of the enhanced training data and the traditional DecisionTree generation model.

[0344] like Figure 4 As shown, the test results of the original training data and the AUCOptimizedTrees generated model.

[0345] like Figure 5 As shown, the enhanced training data and the test results of the AUCOptimizedTrees generation model.

[0346] Model evaluation and results analysis

[0347] The performance of different models on the test set was compared, focusing on two key indicators: AUC and KS.

[0348] Table xx Model evaluation results

[0349]

[0350] Experimental results show that the model, after data augmentation and AUCOptimizedTrees learning, performs well in the credit risk commercial performance prediction task. Key evaluation results are as follows: AUC improvement: The AUC value on the test dataset reached 0.8013, an increase of approximately 10.97% compared to the traditional decision tree model, significantly enhancing the model's ability to distinguish between companies with different credit performances; KS improvement: On the imbalanced test dataset, the KS value reached 0.5875, a 29.35% improvement compared to the traditional decision tree model, demonstrating increased robustness when predicting negative examples.

[0351] This paper addresses the credit risk assessment challenges faced by small and medium-sized enterprises (SMEs) in the manufacturing industry and proposes a commercial performance prediction model that integrates an AUC optimization strategy, an attention mechanism, and an adaptive weighting algorithm. This prediction model demonstrates significant advancements and innovations in small-sample learning and credit risk management scenarios.

[0352] Adapt to data environments with small samples and imbalanced categories

[0353] The performance data of small and medium-sized manufacturing enterprises often have the characteristics of small data volume and uneven distribution of positive and negative samples. Traditional credit risk models are prone to overfitting when data is insufficient, or cause classification bias due to category imbalance.

[0354] The improvements of the present invention include:

[0355] AUC-Optimized Adaptive Boosting Tree (AUC-AAET): To address the class imbalance problem, the model optimization objective directly improves the AUC rather than traditional loss function optimization. This allows for more reasonable weight distribution of minority class samples and improves the ability to identify defaulting companies.

[0356] Dynamic sample weighting mechanism: Dynamically adjust sample weights based on sample difficulty and model prediction error, allowing the model to learn more effectively under small sample conditions, improve data utilization efficiency, and reduce dependence on large samples.

[0357] Adaptive feature selection strategy: Adaptively adjust feature weights based on data distribution to improve the generalization ability of the model when data is limited, ensure that key features are fully utilized, and improve stability in small sample environments.

[0358] Attention mechanism improves feature extraction capabilities

[0359] Credit information for small and medium-sized manufacturing enterprises often contains significant noise, and the contribution of features varies dynamically. Traditional models struggle to effectively capture the most influential credit risk factors. This study proposes:

[0360] Decision tree model integrating attention mechanism: Based on the decision tree model, an attention mechanism is added to enable the model to automatically focus on the features that are most critical to the company's performance, thereby improving the accuracy of feature learning.

[0361] Dynamic weighting strategy for risk characteristics: Dynamically adjust the importance of features so that the model can adapt to changes in the company's credit status and improve the stability and interpretability of the prediction.

[0362] Improve the model's ability to predict credit risk

[0363] In application scenarios such as manufacturing supply chain finance, credit rating, and loan approval, it is crucial to improve the accuracy and stability of credit risk predictions.

[0364] The present invention has made improvements at the modeling level:

[0365] The AUC optimization-enhanced decision tree combined with the attention mechanism enables the model to more accurately extract effective information when data is scarce, thereby improving prediction accuracy.

[0366] The adaptive weighting mechanism enables the model to dynamically adjust sample weights, improve the stability of credit risk scores, and make high-risk enterprises easier to identify.

[0367] The proposed commercial performance model for credit risk in manufacturing SMEs integrates an AUC optimization strategy, an attention mechanism, and an adaptive weighting approach, overcoming technical bottlenecks in small sample environments and credit risk identification. By optimizing the AUC objective function, attention feature extraction, and adaptive weighting, the model effectively identifies high-risk enterprises in situations with limited data and imbalanced categories, achieving higher prediction accuracy and applicability in credit risk management scenarios within the manufacturing industry.

[0368] Beneficial effects:

[0369] 1. AUC optimization and dynamic sample weighting mechanism

[0370] Existing models primarily rely on fixed weighting strategies when processing imbalanced data. This static adjustment of sample weights makes it difficult to adapt to changing data characteristics across scenarios. This paper uses an adaptive AUC weighting mechanism to dynamically adjust sample weights. By introducing an AUC loss function for target optimization, the model adaptively adjusts its focus on minority class samples, improving the model's AUC performance on imbalanced data. Through dynamic weight adjustment, this paper significantly improves the model's classification performance on small, imbalanced data.

[0371] 2. Multi-head attention mechanism enhances feature capture capabilities

[0372] Traditional decision tree models rely on fixed splitting rules to capture features, failing to fully tap into the deep connections between features. This paper innovatively incorporates a multi-head attention mechanism to weight input features, enabling the model to more effectively capture key features and improving the robustness and discriminative power of feature representation. The multi-head attention mechanism helps the model capture multi-dimensional information in the feature space, enhancing the model's ability to learn complex features and further improving its generalization performance.

[0373] 3. Adaptive multi-objective optimization strategy

[0374] Traditional tree models often use a single-objective optimization approach, ignoring the trade-offs between different performance metrics (such as AUC, precision, and recall). This paper introduces an adaptive multi-objective optimization strategy that dynamically adjusts the weights between objectives based on AUC performance, achieving comprehensive optimization of different evaluation metrics. Compared with traditional models, this strategy improves AUC while also taking into account the model's overall predictive capabilities, providing a more optimal solution for model application in complex business scenarios.

[0375] 4. Efficiency and explainability

[0376] In terms of balancing model computational complexity and performance, this paper optimizes the interaction between the attention mechanism and the boosted tree model, enabling the model to improve performance while maintaining high computational efficiency. Compared to complex deep learning-based models (such as Attention-based XGBoost and NAF), this paper provides higher interpretability through a lightweight attention-enhanced tree structure, while reducing the time complexity and computational cost of model training, making it more suitable for rapid deployment and large-scale application in industrial scenarios.

[0377] 5. Wide applicability of application scenarios

[0378] This method is particularly effective in commercial contract risk modeling for small and medium-sized manufacturing enterprises, particularly in credit risk assessment, credit default prediction, and supply chain contract performance analysis. By dynamically adjusting sample weights and enhancing attention, the model maintains stable predictive performance across diverse business scenarios, expanding the application boundaries of the AUC-AAET model.

[0379] In summary:

[0380] The present invention organically combines small sample enhancement, attention mechanism, dynamic sample weighting, AUC optimization and multi-objective optimization strategy through the AUC-AAET model, significantly improving the generalization ability and prediction accuracy of the model, solving the performance bottleneck of traditional methods in small sample data and unbalanced environments, and providing a more accurate and efficient modeling tool for commercial performance risk assessment of small and medium-sized enterprises in the manufacturing industry.

[0381] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for establishing an AUC-optimized adaptive boosting tree model for risk control, characterized in that: The model building method comprises: Build a decision tree model; Collect raw data and perform data preprocessing to obtain preprocessed data; The small sample learning technology and sample weighting mechanism are used to train the decision tree model, and the attention mechanism is used to enhance the model to obtain the AUC optimized adaptive enhanced tree model for risk control.

2. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 1, characterized in that: The collecting of raw data and performing data preprocessing to obtain preprocessed data specifically includes: cleaning the raw data set, feature selection and engineering.

3. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 1, characterized in that: The use of small sample learning technology and sample weighting mechanism to train the decision tree model specifically includes: Initialize attention model and decision tree parameters; Training the attention model; Build a decision tree; Make predictions for each decision tree and obtain probability values; Average the predictions of all trees to get the final probability; Output the category label according to the probability threshold.

4. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 3, characterized in that: The construction of the decision tree specifically includes: Perform weighted sampling according to the sample weight to obtain the data set after sample resampling; A plurality of decision trees are trained on the resampled data set.

5. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 3, characterized in that: The training attention model specifically includes: Calculate category weights; Calculate sample weights; Train the attention model using the AUC loss function: Forward propagation: calculate model output; backpropagation: update model parameters.

6. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 5, characterized in that: The training attention model specifically includes: The multi-head attention mechanism captures the relationship between input features and then maps the attention output to the target dimension through a fully connected layer; Among them, the attention model output formula is: output = AttentionModel(X); Among them, AttentionModel is a neural network model based on a multi-head attention mechanism, which is used to calculate the weight of the sample. The input is the feature matrix X, and the output is the weight of each sample. The neural network model of the multi-head attention mechanism includes: a multi-head attention layer to capture the relationship between input features; and a fully connected layer to map the output of the attention mechanism to the target dimension.

7. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 6, characterized in that: The multi-head attention layer specifically includes: Input: Input feature matrix Where n is the number of samples and d is the feature dimension. Linear transformation: Perform linear transformation on the input features to obtain query, key and value: Q=XW Q ,K=XW K ,V=XW V Among them, W Q , W K , is the learned weight matrix, d k is the dimension of each head; Calculate the attention score matrix A: Among them, softmax is used to normalize the attention score; Weighted summation: Perform weighted summation on the value V to obtain the attention output: Attention(Q, K, V)=AV.

8. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 6, characterized in that: The fully connected layer specifically includes: The first layer is fully connected: h=ReLU(W1·Attention(Q,K,V)+b1) in, is the weight matrix, b1 is the bias term, and ReLU is the activation function; The second fully connected layer: y = W2 h + b2; in, is the weight matrix, b2 is the bias term; The sample weight calculation formula is: sample_weights=class_weights×attention-weights Among them, class_weights is the category weight, balancing the category imbalance problem; attention_weights is the weight of the attention model output, focusing on difficult samples; The calculation formula of the category weight is: Among them, N is the total number of samples; N0 is the number of negative samples; N1 is the number of positive samples.

9. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 5, characterized in that: The use of the AUC loss function to train the attention model specifically includes: The AUC loss function is used to optimize the model's ability to distinguish between positive and negative samples. This time, logarithmic loss is used to indirectly optimize AUC. The logarithmic loss formula is as follows: Where: y i is the true label of the sample (0 or 1); p i is the probability predicted by the model; N is the number of samples: The AUC loss function formula is: Among them, N + is the number of positive samples; N - is the number of negative samples; p i is the predicted probability of the positive sample; p j is the predicted probability of negative samples.

10. The method for establishing an AUC-optimized adaptive boosting tree model for risk control according to claim 4, characterized in that: The sample resampling specifically includes: Weighted sampling is used to enhance the weight of difficult samples. The formula is: Among them, γ is the weight coefficient of difficult samples; when sampling, weighted random sampling is performed according to sample_weights.

Images