Fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set

Abstract

The invention relates to the technical field of risk management and data analysis, in particular to a fraud detection method and system for an unbalanced data set, and more particularly relates to a fraud detection method based on an SMOTE-CatBoost-SHAP unbalanced data set. Comprising the following steps: (1) a data preprocessing step; (2) a model training step; (3) a model evaluation step; (4) explaining the model; according to the method, a data balancing technology, an advanced machine learning algorithm and an interpretability framework are comprehensively applied to form a set of complete fraud detection solution; through the comprehensive technical advantages, when the model processes complex and unbalanced trade data, high-precision fraud recognition and interpretable decision support can be achieved at the same time, and a new technical path and a new solution are provided for the field of risk fraud detection.

Description

Technical Field

[0001] The present invention relates to the technical field of risk management and data analysis, and in particular to a fraud detection method and system for an unbalanced data set, and more specifically to a fraud detection method based on a SMOTE-CatBoost-SHAP unbalanced data set. Background Art

[0002] In the field of risk fraud detection, since fraudulent behaviors are often hidden and complex, and fraudulent samples usually account for only a very small number of the overall data set, this leads to serious data imbalance problems; this data imbalance poses a challenge to the training of machine learning models, because traditional classification algorithms are prone to overfitting or underfitting when faced with unbalanced data sets, thus affecting the accuracy and generalization ability of the model; especially in the application scenario of import and export trade fraud detection, with the increase in global trade volume and the diversification of fraud methods, how to effectively identify and prevent fraud has become a key issue that customs and regulatory agencies need to solve urgently.

[0003] Traditional fraud detection methods, such as rule-based expert systems and statistical methods, can identify fraudulent behavior to a certain extent, but have many limitations. Expert systems rely on manually set rules and have difficulty coping with the diversity and dynamic changes of fraudulent behavior. Statistical methods often assume that data follows a specific distribution, which is not always true in complex trade data. Therefore, these methods have limited effectiveness in practical applications.

[0004] With the rapid development of machine learning technology, fraud detection methods based on machine learning have gradually emerged; machine learning models can learn and identify complex fraud patterns from historical transaction data through big data analysis, thereby improving the efficiency and accuracy of fraud detection; however, machine learning models also face challenges on unbalanced data sets; in order to solve this problem, scholars have made many attempts at the data level, algorithm level, and the combination of data and algorithms, such as using undersampling, oversampling, and comprehensive sampling methods to balance the data set, or using cost-sensitive learning and Boosting algorithms to optimize model performance.

[0005] Among them, SMOTE (Synthetic Minority Over-sampling Technique), as an effective oversampling method, balances the data set by synthesizing minority class samples and is widely used in the classification problems of unbalanced data sets; and CatBoost, as a model based on gradient boosting decision trees, demonstrates superior performance in classification problems by optimizing category feature processing and avoiding overfitting.

[0006] Despite this, fraud detection models based on machine learning still face the problem of poor interpretability. These models are often viewed as black boxes, and it is difficult to explain their decision-making process, which leads to regulators’ distrust and low acceptance of the fraud identification conclusions output by the models. In order to improve the interpretability of the model, the SHAP (SHapley Additive ex Planations) method was born.

[0007] SHAP is a model interpretation method based on the Shapley value in game theory. It can assign a contribution value to each feature to indicate the impact of the feature on model predictions. By calculating the SHAP value of each sample, we can intuitively understand how the model makes decisions, thereby improving the interpretability and transparency of the model. Combining the SHAP method with machine learning models can reveal key features that affect the identification of fraudulent transactions and provide regulators with a more reliable basis for fraud detection. Summary of the invention

[0008] The present invention aims to provide a fraud detection solution based on SMOTE-CatBoost-SHAP, aiming to solve the problem of model performance degradation caused by severe data imbalance in import and export trade fraud detection, and enhance the interpretability of the model by introducing the SHAP (SHapley Additive ex Planations) method, thereby improving the accuracy and credibility of fraud detection. The technical solution integrates data balancing technology (SMOTE), advanced machine learning algorithm (CatBoost) and model interpretation tool (SHAP), and is specially designed for processing complex and highly unbalanced trade fraud data. It is of great significance to promote international trade security, improve customs supervision efficiency and promote the application of machine learning technology in risk management.

[0009] The technical solution adopted by the present invention to solve the technical problem is: a fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set, comprising the following steps:

[0010] (1) Data preprocessing step: obtain the original data set, standardize it, and then divide the data set into a training data set and a test data set;

[0011] (2) Model training steps: Using the training data set, combined with the SMOTE algorithm, data balancing is performed, and then the CatBoost fraud detection model is trained;

[0012] (3) Model evaluation step: Use the test dataset to verify the recognition performance of the SMOTE-CatBoost-based fraud detection model;

[0013] (4) Model interpretation step: The SHAP method is introduced to calculate the contribution value of each feature to the model prediction results and identify the key features that have important contributions to the fraud detection results.

[0014] The data preprocessing step also includes the operations of performing outlier detection and missing value processing on the original data set.

[0015] In the data balancing process, the SMOTE algorithm balances the data set by randomly selecting minority class samples and synthesizing new minority class samples.

[0016] During the training process, the CatBoost fraud detection model adopts a ranking boosting strategy to ensure that only the part before the current data is used to calculate the residual when building each tree, thereby avoiding data leakage problems.

[0017] The CatBoost fraud detection model adopts target smoothing coding technology for category features, encoding the category features as statistical information of the category features and the target variable, avoiding the dimensionality expansion problem caused by one-hot coding.

[0018] In the model evaluation step, indicators such as accuracy, recall, and F1 score are used to evaluate the recognition performance of the fraud detection model based on SMOTE-CatBoost.

[0019] In the model interpretation step, the SHAP method quantifies the marginal contribution of each feature to the prediction result by calculating the Shapley value of each feature, thereby providing transparency of the model decision-making process.

[0020] The SHAP method also includes visualizing Shapley values ​​to intuitively show which variables have higher weights in judging fraud risk.

[0021] A fraud detection method based on a SMOTE-CatBoost-SHAP unbalanced data set also includes a comparison step: comparing the performance of a fraud detection model based on SMOTE-CatBoost with other machine learning models, such as logistic regression, decision tree classifier, gradient boosting classifier, K nearest neighbor classifier, naive Bayes classifier, multilayer perceptron, stochastic gradient descent classifier, etc., to verify the superiority of the proposed model.

[0022] The fraud detection method based on the SMOTE-CatBoost-SHAP unbalanced data set can be applied to import and export trade fraud detection scenarios, and is verified using a real customs declaration commodity data set to prove its effectiveness and practicality.

[0023] Beneficial effects of this technical solution: The present invention comprehensively applies data balancing technology, advanced machine learning algorithms and interpretability frameworks to form a complete fraud detection solution. This comprehensive technical advantage enables the model to simultaneously achieve high-precision fraud identification and interpretable decision support when processing complex and unbalanced trade data, providing a new technical path and solution for the field of risk fraud detection, which is specifically reflected as follows:

[0024] In the face of ever-changing fraud methods, the model proposed in this invention demonstrates good adaptability and flexibility; through SHAP analysis, new fraud patterns or new variants of known fraud methods can be discovered and explained in a timely manner, providing powerful decision-making support for customs and regulatory agencies; this ability to dynamically adjust and optimize fraud detection strategies is unmatched by traditional fraud detection systems based on rules or statistical methods.

[0025] The embodiment of the present invention takes import and export trade fraud detection as the application scenario, and verifies the classification accuracy and interpretability of the model through a real customs declaration commodity data set; the experimental results show that the model can not only effectively identify fraudulent behavior, but also provide clear explanations and visual analysis, which greatly enhances the value of the model in practical applications; this is of great significance to improving the efficiency of customs supervision, combating trade fraud, and maintaining a fair trade order. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the execution logic principle in the embodiment.

[0027] Figure 2 Schematic diagram of feature selection results based on decision tree in the embodiment.

[0028] Figure 3 Schematic diagram of SHAP feature analysis in the embodiment.

[0029] Figure 4 is a waterfall chart of SHAP values ​​in the embodiment. DETAILED DESCRIPTION

[0030] The specific implementation of this technical solution is described in detail below in conjunction with specific embodiments:

[0031] Example 1: Application to fraud detection scenarios for import and export declaration companies

[0032] With the continuous growth of global trade, the number and types of imported and exported goods are increasing, and import and export declarations have become an important link in international trade. However, the problem of trade fraud in import and export declarations has also become prominent, becoming a risk area that governments and regulators around the world focus on. According to a report by the World Customs Organization (WCO), trade fraud not only seriously affects tariff revenue, but may also lead to unfair competition and market chaos. In recent years, with the complexity of cross-border trade and the rapid growth of transaction volume, fraud methods have become increasingly diverse, including false reporting of commodity values, false commodity classifications, and falsification of origin. Import and export declaration fraud refers to the behavior of enterprises or individuals deliberately falsely declaring commodity information in order to obtain illegal economic benefits, thereby circumventing tariffs or obtaining improper preferences. These fraudulent behaviors have become more difficult to detect with the increase in global trade activities. Traditional fraud detection methods face many challenges, such as the large scale of transaction data, complex data dimensions, and dynamic changes in fraud patterns. Therefore, how to use advanced technology to effectively identify fraud in import and export declarations has become a key issue that customs and regulators need to solve urgently.

[0033] This embodiment will be based on SMOTE-CatBoost-SHAP for import and export declaration fraud detection of severely unbalanced data sets; Figure 1 As shown, the implementation of this embodiment is divided into the following four parts:

[0034] (1) Data collection and preprocessing

[0035] First, the original data is obtained and the data set is standardized; then, the constructed data set is divided into a training data set and a test data set in proportion.

[0036] (2) Model training and optimization

[0037] This embodiment trains the SMOTE-CatBoost model based on the training data set, and compares SMOTE-LR, SMOTE-DTC, SMOTE-GBC, SMOTE-KNNC, SMOTE-BNB, SMOTE-GNB, SMOTE-MLPC and SMOTE-SGDC.

[0038] (3) Model testing and performance evaluation

[0039] This example uses a test data set to verify the recognition performance of the import and export declaration enterprise fraud detection model based on SMOTE-CatBoost.

[0040] (4) Model interpretability analysis

[0041] This embodiment uses the SHAP method to identify key features that contribute significantly to the results of fraud detection and identification of import and export declaration companies.

[0042] In the execution experiment of this embodiment, the import and export declaration enterprise fraud detection model is mainly constructed based on CatBoost; CatBoost is a model based on gradient boosting decision trees (GradientBoostingDecisionTrees, GBDT), which specifically optimizes category feature processing and avoids overfitting; the main innovation of CatBoost lies in ordered boosting (OrderedBoosting) and native category feature processing; ordered boosting divides the data into sequential blocks to ensure that when building each tree, only the part before the current data is used to calculate the residual, thereby avoiding the common data leakage problem in gradient boosting; especially when processing time series data or small sample data sets, ordered boosting can significantly improve the generalization performance of the model.

[0043] For categorical features, CatBoost does not require traditional one-hot encoding or target encoding, but uses a special target smoothing encoding technology; it is based on the correlation between categorical features and target variables, by encoding categorical features as statistical information about them and target variables, and combining ranking boosting to prevent information leakage, which not only maintains the information integrity of categorical features, but also avoids the dimensionality expansion problem caused by one-hot encoding; CatBoost's function form is:

[0044]

[0045] Among them, Tk(x) represents the prediction of the kth decision tree, αk is the weight of the kth tree, and K is the total number of trees in the model; the final prediction value f(x) is obtained by the weighted sum of all decision trees.

[0046] Then the model interpretability of SHAP is integrated; SHAP is an algorithm for explaining machine learning models. It assigns a contribution value to each feature based on the Shapley value in game theory to indicate the impact of the feature on the model prediction; SHAP provides a consistent and unified method to explain the predictions of any machine learning model, and can quantify the marginal contribution of each feature to the prediction results; the theoretical basis of the SHAP algorithm comes from the Shapley value in game theory; the Shapley value is a fair method for allocating the benefits of each participant in a cooperative game. In machine learning, features can be regarded as participants in the game, and the predicted value of the model can be regarded as benefits; the Shapley value determines the importance of each feature in the current sample by calculating the marginal contribution of the feature in all possible feature combinations; specifically, for each sample x, its model prediction f(x) is decomposed into the sum of the contributions of each feature:

[0047]

[0048] in, is the baseline value (usually the average prediction of the model on the scene data); is the Shapley value of feature i, which indicates the marginal contribution of feature i to the prediction result.

[0049] The calculation of the Shapley value involves traversing all possible subsets of the features and calculating the marginal contribution; for a feature i, the Shapley value The formula is:

[0050]

[0051] Where N is the set of all features, S is a subset of feature i, excluding i itself, f(s) represents the prediction value of the model when only feature subset S is used; f(s∪{i}) represents the prediction value after adding feature i to feature subset S. is the weight assigned to each subset in the Shapley value, ensuring that the marginal contribution of each feature can be calculated fairly.

[0052] This embodiment effectively solves the problem of overfitting or underfitting of fraud detection models in unbalanced data sets by combining SMOTE data balancing technology with CatBoost machine learning algorithm. SMOTE technology makes the data set more balanced by synthesizing minority class samples, while CatBoost algorithm further improves the fraud identification ability of the model in complex trade data through its unique ranking boosting and native category feature processing technology. Compared with existing models such as logistic regression, decision tree classifier, and gradient boosting classifier, the SMOTE-CatBoost model proposed in this embodiment shows higher classification accuracy and generalization performance in fraud detection tasks. At the same time, the present invention innovatively introduces the SHAP interpretation framework to provide strong interpretability support for the SMOTE-CatBoost model. SHAP calculates the contribution value (SHAP value) of each feature to the model prediction result, so that researchers and business personnel can intuitively understand which variables have higher weights in judging fraud risks. This transparent interpretation method not only improves the credibility of the model, but also helps business experts to more accurately identify important influencing factors of fraudulent behavior, thereby optimizing fraud detection strategies.

[0053] Example 2: Import and export trade fraud detection model based on SMOTE-CatBoost-SHAP and interpretability improvement method

[0054] This embodiment aims to combine data balancing technology with machine learning to weaken the above phenomenon and improve model performance. In addition, when conducting risk assessment and disposal based on model detection results, business experts pay more attention to the interpretability of black box models. Based on the above model architecture, this embodiment introduces SHAP (SHapley Additive exPlanations) to identify important influencing factors of fraudulent behavior, thereby improving the interpretability and scenario usability of the model. The study takes import and export trade fraud detection as an application scenario, and uses a real customs declaration form data set to verify the classification accuracy and interpretability of the SMOTE-CatBoost-SHAP risk fraud detection model proposed in this embodiment.

[0055] This embodiment is based on SMOTE-CatBoost-SHAP for import and export declaration fraud detection for severely unbalanced data sets; the research framework is as follows Figure 1 As shown, it is divided into the following four parts:

[0056] (1) Data collection and preprocessing: First, obtain the original data and standardize the dataset; then, divide the constructed dataset into training dataset and test dataset in proportion.

[0057] (2) Model training and optimization: In this embodiment, the SMOTE-CatBoost model is trained based on the training data set, and SMOTE-LR, SMOTE-DTC, SMOTE-GBC, SMOTE-KNNC, SMOTE-BNB, SMOTE-GNB, SMOTE-MLPC and SMOTE-SGDC are compared.

[0058] (3) Model testing and performance evaluation: Using the test dataset, verify the recognition effect of the import and export declaration enterprise fraud detection model based on SMOTE-CatBoost.

[0059] (4) Model interpretability analysis: This embodiment uses the SHAP method to identify key features that contribute significantly to the results of fraud detection for import and export declaration companies.

[0060] In this embodiment, CatBoost is a model based on Gradient Boosting Decision Trees (GBDT), which specifically optimizes category feature processing and avoids overfitting; the main innovation of CatBoost lies in ordered boosting (Ordered Boosting) and native category feature processing; ordered boosting divides the data into sequential blocks to ensure that when building each tree, only the part before the current data is used to calculate the residual, thereby avoiding the common data leakage problem in gradient boosting; especially when processing time series data or small sample data sets, ordered boosting can significantly improve the generalization performance of the model.

[0061] For categorical features, CatBoost does not require traditional one-hot encoding or target encoding, but uses a special target smoothing encoding technology; it is based on the correlation between categorical features and target variables, by encoding categorical features as statistical information about them and target variables, and combining ranking boosting to prevent information leakage, which not only maintains the information integrity of categorical features, but also avoids the dimensionality expansion problem caused by one-hot encoding; CatBoost's function form is:

[0062]

[0063] Among them, T k (x) represents the prediction of the kth decision tree, α k is the weight of the kth tree, K is the total number of trees in the model; the final prediction value f(x) is obtained by the weighted sum of all decision trees.

[0064] SHAP is an algorithm for explaining machine learning models. It is based on the Shapley value in game theory and assigns a contribution value to each feature to indicate the impact of the feature on model predictions. It can quantify the marginal contribution of each feature to the prediction results.

[0065] The theoretical basis of the SHAP algorithm comes from the Shapley value in game theory; the Shapley value is a fair method for allocating benefits to each participant in a cooperative game. In machine learning, features can be regarded as participants in the game, and the predicted value of the model can be regarded as benefits; the Shapley value determines the importance of each feature in the current sample by calculating the marginal contribution of the feature in all possible feature combinations; specifically, for each sample x, its model prediction f(x) is decomposed into the sum of the contributions of each feature:

[0066]

[0067] in, is the baseline value (usually the average prediction of the model on the background data); is the Shapley value of feature i, which indicates the marginal contribution of feature i to the prediction result.

[0068] The calculation of the Shapley value involves traversing all possible subsets of the features and calculating the marginal contribution; for a feature i, the Shapley value The formula is:

[0069]

[0070] Where N is the set of all features, S is a subset of feature i, excluding i itself,

[0071] f(S) represents the prediction value of the model when only the feature subset S is used; f(S∪{i}) represents the prediction value after adding feature i to the feature subset S. is the weight assigned to each subset in the Shapley value, ensuring that the marginal contribution of each feature can be calculated fairly.

[0072] Based on this, the following will be further demonstrated and explained through specific experiments:

[0073] By mining and exploring the big data resources at the disposal of the customs, we collected 1 million registered import and export trading companies, and took the four major dimensions of basic information of the company, transaction information, operating conditions, and violation records as the basic directions. We selected basic information including companies, personnel, transactions, and previous penalty records; there were 33,406 high-risk, 46,327 medium-risk, and 920,267 low-risk samples.

[0074] In order to ensure the performance and stability of the model, this embodiment systematically preprocesses the original data; first, the missing values ​​are filled, the outliers are detected and processed, the samples with mixed data types are removed, and the categorical variables are one-hot encoded; in order to facilitate the model to recognize and utilize time information, the time series data is timestamp converted; secondly, in order to solve the problem of class imbalance in the data set, the SMOTE (Synthetic Minority Over-sampling Technique) method is used to process the data. SMOTE effectively increases the number of minority class samples by generating new synthetic samples between minority class samples, thereby balancing the class distribution; this embodiment selects 500,000 low-risk enterprises, 46,327 medium-risk enterprises, and 33,406 high-risk enterprises. The SMOTE technology is used to balance the data sets of high-, medium- and low-risk enterprises, and 500,000 enterprise data are obtained respectively, and then the training set and test set are divided according to the ratio of 7:3.

[0075] In order to improve the performance of the model, this embodiment eliminates irrelevant or redundant features through feature selection engineering, thereby reducing the number of features, improving the accuracy of the fraud detection model for import and export declaration companies, and reducing the running time; the tree model selects features that can take into account the interaction between features, process nonlinear features and high-dimensional data, and is suitable for data sets with complex relationships between features. Therefore, this embodiment selects a decision tree as an evaluation model for feature selection.

[0076] The goal of feature selection is to maximize the accuracy of the model calculated in the five-fold cross validation by optimizing the feature combination. The model uses an iterative cycle method to gradually adjust the selected feature ratio to find the best feature subset; the final result is as follows Figure 2 As shown, the horizontal axis represents the percentage of the selected features in the total features, and the vertical axis represents the accuracy of the model under this feature ratio; the study found that as the number of selected features increases, the accuracy of the model also increases; in order to make full use of the information in the data and improve the prediction performance of the model, this embodiment selects a total of 25 features through feature engineering to ensure that the model can capture the important patterns and laws in the data to the greatest extent.

[0077] In order to objectively evaluate the effectiveness of the algorithm, three evaluation indicators, namely accuracy, recall rate and F1 value, are used to measure the model performance; the evaluation indicators are all based on the confusion matrix, as shown in Table 1 below:

[0078]

[0079]

[0080] Table 1 Confusion matrix

[0081]

[0082] Among them, True Positive (TP) represents the number of correctly identified samples; False Positive (FP) represents the number of incorrectly identified samples; True Negative (TN) represents the number of correctly identified samples as incorrect; False Negative (FN) represents the number of incorrectly identified samples as correct.

[0083] In order to verify the effectiveness of the fraud identification model based on SMOTE-CatBoost proposed in this embodiment, a comparative analysis is performed with the SMOTE-LR, SMOTE-DTC, SMOTE-GBC, SMOTE-KNNC, SMOTE-BNB, SMOTE-GNB, SMOTE-MLPC and SMOTE-SGDC models; the comparison results of various models in terms of the three indicators of accuracy, recall rate and F1 value are shown in Table 2.

[0084] From the perspective of the accuracy of fraud identification of each model after balancing the data set, SMOTE-CatBoost has the highest fraud identification accuracy of 84.8%, and SMOTE-BNB has the lowest fraud identification accuracy of 8.1%; specifically, the recognition accuracy of the SMOTE-CatBoost model is 66.9%, 3.8%, 4.0%, 15.4%, 7.5%, 946.9%, 115.2% and 108.4% higher than the eight fraud identification models SMOTE-LR, SMOTE-DTC, SMOTE-GBC, SMOTE-KNNC, SMOTE-BNB, SMOTE-GNB, SMOTE-MLPC and SMOTE-SGDC, respectively.

[0085]

[0086]

[0087] Table 2 Evaluation indicators of different models

[0088] The import and export declaration fraud detection and identification based on SMOTE-CatBoost has a high prediction accuracy, but because it is a "black box" model, it is difficult for the risk management department to understand which variables play a key role. To address this problem, the SHAP framework is introduced to explain the model results. While providing feature importance ranking, it focuses on explaining the specific impact of different feature values ​​on the prediction results. The SHAP value analysis chart of all variables on all samples is drawn as follows: Figure 3 As shown in the figure, each point represents a sample, and the sample point changes from blue to red, indicating that the SHAP value of the sample changes from small to large; the image of each indicator is composed of the SHAP values ​​of all sample points, and the SHAP value equals zero as the middle dividing line; the sample point on the left has a negative SHAP value for this indicator, that is, this indicator has a negative contribution to the identification of fraud in import and export declaration companies; the sample point on the right has a positive SHAP value for this indicator, that is, this indicator has a positive contribution to the detection and identification of fraud in import and export declaration companies; it can be seen that the features based on the SHAP value can accurately analyze the impact of each sample and each feature on the detection and identification of fraud in import and export declaration companies.

[0089] Depend on Figure 3It can be seen that the SHAP values ​​of features such as validity date, inspection rate, and annual review date fluctuate greatly, indicating that these features have a significant impact on the judgment of the model; among them, the annual review date and validity date have a negative impact on the enterprise fraud risk assessment, that is, the later the annual review date and validity date, the lower the enterprise fraud risk, which is in line with the logic in reality. The later the annual review time and validity time usually means that the enterprise complies with laws and regulations and management norms, and the fraud risk is relatively low; the inspection rate also shows a negative impact. The higher the inspection rate, the lower the enterprise fraud risk. This may be because by increasing supervision and constraints and reducing corporate risk expectations, companies are prompted to comply with the rules more and reduce the probability of fraud.

[0090] In this embodiment, the waterfall chart used is an effective visualization tool, which aims to intuitively show the contribution of each feature to the model prediction result; it comprehensively presents the prediction value of a specific sample, influencing factors, the average prediction value of all features, and the average marginal contribution of each feature; in this figure, the model output for a specific sample shows the positive or negative impact of each feature on the final prediction result; according to the analysis results of the sample waterfall chart, it can be seen that "total tax amount" and "enterprise level" are the features that have the greatest positive effect on the prediction result, which significantly improves the prediction value of the model; in contrast, "enterprise credit mark" shows the greatest negative effect, which inhibits the improvement of the prediction result; the effects of all features ultimately converge to form the prediction result of the model, which fully reflects the interaction and importance of features in the decision-making process.

[0091] This example uses public data to conduct a modeling analysis on fraud detection of import and export declaration companies, and clearly draws the following conclusions:

[0092] (1) The embodiment constructs an import and export declaration enterprise fraud detection model based on SMOTE-CatBoost. To verify the performance of the model, a comparative experiment is conducted with seven classification models. The results show that the import and export declaration enterprise fraud detection based on SMOTE-CatBoost has the highest recognition accuracy, while the import and export declaration enterprise fraud detection based on SMOTE-BNB has the lowest recognition accuracy.

[0093] (2) To verify the importance of different variables to the results. The SHAP method was used to analyze the importance of the overall model features and found that the features such as the validity date, inspection rate, and annual review date had a greater impact; the SHAP model interpretation results showed that the annual review date, validity date, and inspection rate had a negative and significant impact, providing a reference for the key monitoring indicators of fraud detection for import and export declaration companies.

Claims

1. A fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set, characterized by: The following steps are involved: (1) Data preprocessing step: obtain the original data set, standardize it, and then divide the data set into a training data set and a test data set; (2) Model training steps: Using the training data set, combined with the SMOTE algorithm, data balancing is performed, and then the CatBoost fraud detection model is trained; (3) Model evaluation step: Use the test dataset to verify the recognition performance of the SMOTE-CatBoost-based fraud detection model; (4) Model interpretation step: The SHAP method is introduced to calculate the contribution value of each feature to the model prediction results and identify the key features that have important contributions to the fraud detection results.

2. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: The data preprocessing step also includes the operations of performing outlier detection and missing value processing on the original data set.

3. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: In the data balancing process, the SMOTE algorithm balances the data set by randomly selecting minority class samples and synthesizing new minority class samples.

4. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: During the training process, the CatBoost fraud detection model adopts a ranking boosting strategy to ensure that only the part before the current data is used to calculate the residual when building each tree, thereby avoiding data leakage problems.

5. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: The CatBoost fraud detection model adopts target smoothing coding technology for category features, encoding the category features as statistical information of the category features and the target variable, avoiding the dimensionality expansion problem caused by one-hot coding.

6. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: In the model evaluation step, indicators such as accuracy, recall, and F1 score are used to evaluate the recognition performance of the fraud detection model based on SMOTE-CatBoost.

7. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: In the model interpretation step, the SHAP method quantifies the marginal contribution of each feature to the prediction result by calculating the Shapley value of each feature, thereby providing transparency of the model decision-making process.

8. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 7, characterized in that: The SHAP method also includes visualizing Shapley values ​​to intuitively show which variables have higher weights in judging fraud risk.

9. The fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to claim 1, characterized in that: It also includes a comparison step: comparing the performance of the SMOTE-CatBoost-based fraud detection model with other machine learning models, such as logistic regression, decision tree classifier, gradient boosting classifier, K-nearest neighbor classifier, naive Bayes classifier, multi-layer perceptron, stochastic gradient descent classifier, etc., to verify the superiority of the proposed model.

10. A fraud detection method based on SMOTE-CatBoost-SHAP unbalanced data set according to any one of claims 1 to 9, characterized in that: This method is applied to the import and export trade fraud detection scenario and is verified using a real customs declaration commodity dataset to prove its effectiveness and practicality.

Images