Credit card risk prevention and control method and system based on Stacking ensemble learning

Abstract

The invention belongs to the field of social risk prevention and control, and discloses a credit card risk prevention and control method and system based on Stacking ensemble learning, and the method comprises the steps: preprocessing a data set, extracting key features, and employing a resampling technology to balance positive and negative examples; constructing a shallow feature rule by means of a diversified machine learning method, and establishing a deep neural rule by using deep learning technologies such as a multi-layer perceptron; based on a Stacking ensemble learning method, a feature rule and a neural rule are fused. According to the invention, the risk can be identified and predicted more accurately, so that the accuracy and timeliness of early warning are improved; through fusion optimization of multiple rules, the problems of ambiguity, conflict, non-uniformity and the like in current social governance and risk prevention and control rules are solved.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the technical field of social risk prevention and control, and in particular relates to a credit card risk prevention and control method and system based on Stacking ensemble learning. Background Art

[0002] In modern society, intelligent technology plays a vital role in preventing and controlling social risks. It can identify and predict potential risks more comprehensively and accurately, and provide support so that effective preventive measures can be taken in a timely manner, thereby maintaining social stability and security. The application of intelligent rule generation makes risk management more flexible and adaptable, and can effectively respond to the ever-changing social environment. In addition, it improves the scientificity and accuracy of the decision-making process, provides more reliable decision-making support for government departments and enterprises, and promotes the sustainable development of society. Rule-based expert system is one of the core methods for early credit card fraud detection. This system uses a series of predefined rules to determine the legitimacy of transactions. Typical rules are usually simple "if-then" or "if-else-then" structures. For example, the system can set trigger conditions based on parameters such as transaction amount, geographic location, time period, etc. When a transaction meets certain conditions, it can be marked as an anomaly. This rule-based engine has clear logic, and experts can directly design and adjust rules. However, this relies on expert experience and is difficult to cover all fraud methods. As fraud methods evolve and become more complex, it is difficult to adapt to new fraud behaviors by relying solely on rules. Furthermore, when the number of rules increases, not only are they difficult to manage, but conflicts arise between the rules, which complicates the rule system and increases maintenance costs.

[0003] The challenges in the field of risk prevention and control are mainly reflected in the difficulty in identifying and variability of risk patterns, the huge size of data, and the high requirements for real-time response. In addition, this field also faces the problem of dealing with sample imbalance: the natural distribution difference between normal transactions and fraudulent transactions makes the sample imbalance problem in the data set particularly serious. This imbalance leads to a variety of problems in model training, including model bias, poor generalization ability, and unstable performance. The bias problem stems from the relative scarcity of positive samples, which makes the model tend to classify samples into the majority class; the poor generalization ability is because the model fails to learn the characteristics of fraudulent transactions on a balanced data distribution; and the unstable performance is due to the sensitivity to changes in data distribution.

[0004] In this context, research on risk prevention and control data has received widespread attention. However, a single model or rule has limitations when processing this data. Therefore, it is particularly important to study multi-rule construction and fusion methods to improve the accuracy and efficiency of data analysis. These studies will help expand the application areas of tabular data processing, enhance decision-making support capabilities, and provide important support for the development and practical application of related fields. Therefore, the present invention hopes to improve the efficiency and accuracy of data analysis by developing a multi-rule construction and fusion process for risk prevention and control data, so as to more effectively process and analyze large-scale and complex social risk data, thereby bringing innovative solutions to social risk prevention and control and other related fields, and promoting the application and development of intelligent technology in these fields.

[0005] In the field of risk prevention and control data rule construction, traditional methods mainly focus on the application of knowledge rules, which are generally implemented through expert systems or automatic rule machines. However, the present invention expands the definition of rules and includes all methods that can reason about data into the category of rules. This broad definition covers reasoning methods from simple feature rules to complex neural rules. Feature rules are reasoning rules based on shallow machine learning models that identify patterns and anomalies by analyzing significant features in the data. Although these rules have a certain degree of interpretability, their ability to capture deep information is limited. Relatively speaking, neural rules are reasoning rules based on deep neural network models that can deeply mine deep information in the data through complex algorithmic models. However, such rules usually require a large amount of data and computing resources, and the interpretability of the model is also a challenge.

[0006] In the field of traditional rule extraction and expression, the construction of single rules has made some progress, effectively alleviating the problem of difficulty in extracting and expressing rules. However, this method still needs to be improved in improving prediction accuracy. By combining feature rules with neural rules, complementary advantages can be achieved, thereby significantly improving the performance of the prediction model. This fusion strategy can not only optimize the performance of existing rules, but also provide a more accurate strategy for solving complex problems.

[0007] Technical problems existing in the application of existing technologies in the industry:

[0008] 1. Simplified rules lead to insufficient prediction accuracy and adaptability

[0009] In the field of risk prevention and control, systems based on a single rule have significant performance limitations, which are manifested in:

[0010] Limited coverage of a single rule: Traditional rules based on “if-then” or specific algorithms can only process historical data or known patterns, and are unable to cope with complex and dynamically changing risk patterns.

[0011] Poor generalization ability: Single-rule models cannot learn deep risk features from limited training data and have difficulty capturing new or complex fraud behaviors.

[0012] Delayed rule updates: Since a single rule relies on expert presets, when fraud patterns evolve rapidly, the rule cannot be updated in time to adapt to new risk scenarios.

[0013] Insufficient ability to cope with complex scenarios: Single rules perform poorly in complex scenarios with multi-dimensional data and multi-variable interactions, and are prone to misjudgments and missed judgments.

[0014] Industry Impact:

[0015] The insufficient performance of the prediction model will lead to a decrease in the overall reliability of the risk prevention and control system, and it will be unable to provide accurate risk warnings for enterprises and government agencies.

[0016] Lack of adaptability makes it difficult to expand the system to new areas, weakening its market competitiveness and application value.

[0017] 2. The rule fusion method lacks systematization and efficiency

[0018] Current rule fusion methods mainly focus on simple rule superposition or linear combination, which has the following problems:

[0019] Lack of optimization strategy: The weight distribution of rule fusion is usually set manually or simply weighted average, which makes it difficult to scientifically reflect the relative importance of each rule.

[0020] Conflicts between rules are not effectively resolved: When different rules produce contradictory inferences on the same data, it is difficult for existing systems to make clear judgments, resulting in instability in prediction results.

[0021] Inefficient fusion: When processing large-scale high-dimensional data, the parallel processing and computational optimization of multiple rules are insufficient, resulting in unsatisfactory operating efficiency and real-time performance of the fusion algorithm.

[0022] Insufficient utilization of data correlation: Existing fusion methods rarely fully exploit the contextual information of the data and the nonlinear relationship between features, which limits the improvement of prediction performance.

[0023] Industry impact: Rule conflicts and inefficient fusion strategies will lead to a decline in the decision-making quality of the forecasting system, making it difficult to meet the high requirements of real-time risk prevention and control.

[0024] Efficiency issues restrict the application of the system in large-scale data scenarios, increase operating costs, and weaken user experience and system credibility. Summary of the invention

[0025] In view of the problems existing in the prior art, the present invention provides a credit card risk prevention and control method and system based on Stacking ensemble learning.

[0026] The present invention is implemented as follows: a credit card risk prevention and control method based on Stacking ensemble learning, the method specifically comprising:

[0027] S1: Dataset preprocessing, extracting key features, and using resampling technology to balance positive and negative samples;

[0028] S2: Use a variety of machine learning methods to build shallow feature rules, and use deep learning technologies such as multi-layer perceptrons to build deep neural rules;

[0029] S3: Based on the Stacking ensemble learning method, it integrates feature rules and neural rules.

[0030] Furthermore, in S1, some resampling techniques are used to balance the positive and negative samples. The preprocessed data set is defined as D, where the positive and negative data sets are M and N (generally, M<<N), respectively. The number of data in D is expressed as |D|=M+N. Let the resampled data set be D new :

[0031] (1) Upsampling

[0032] Random upsampling: By randomly copying minority class samples, the number of positive and negative samples is balanced, and NM minority class samples are copied. new |=2*N;

[0033] SMOTE: Synthetic Minority Oversampling Technology, which increases the number of minority samples by inserting newly generated samples between them. Assuming k new samples are generated, each minority sample participates in generating new samples, D new |=M*k+N;

[0034] ADASYN: Adaptive synthetic sampling method, which adjusts the number of new samples generated according to the number of minority class nearest neighbors of each minority class sample. For each minority class sample x i , generate k i new samples, where k i is x i The number of nearest neighbors of the minority class.

[0035] (2) Downsampling

[0036] Random downsampling: By randomly removing majority class samples, reduce their number to match the number of minority class samples, remove NM majority class samples, |D new |=2*M;

[0037] TomekLinks: Reduce noise by identifying and removing minority class samples that are too close to majority class samples. Suppose T majority class noise samples are removed, |D new |=M+NT;

[0038] (3) Combination of upsampling and downsampling

[0039] SMOTEENN: Combining SMOTE and Edited Nearest Neighbors (ENN), SMOTE is used to generate new samples first, and then ENN is used to remove new samples that are too close to the majority class samples. Suppose k new samples are generated by SMOTE, and then T samples are removed by ENN. D new |=M+N+kT;

[0040] SMOTETomek: Combining SMOTE and Tomek Links, first remove minority class samples that are too close to majority class samples, and then use SMOTE to generate new samples. Suppose k new samples are generated by SMOTE, and then T samples are removed by TomekLinks. new |=M+N+kT.

[0041] Further, the S2 defines the input data D new ={(x1,y1),(x2,y2),…,(x n ,y n )}, where x i is the eigenvector, y i is the corresponding label (0 or 1), for each feature x of the data i , a single rule R(·) is constructed through machine learning methods to obtain the corresponding prediction output

[0042] (1) Feature rules

[0043] Classification decision tree: Build decision rules by learning the features of samples. For each node, select the best features and thresholds to split the data:

[0044]

[0045] Among them, x j is the selected feature. t is the segmentation threshold;

[0046] Logistic regression: Logistic regression is used for binary classification problems, and its prediction formula is:

[0047]

[0048] Among them, β0, β1, β2…β n are the parameters of logistic regression;

[0049] K-nearest neighbor method: K-NN predicts the category of a new sample based on the categories of the nearest k neighbors:

[0050]

[0051] Among them, {y1,y2,…,y k} is the category of the k nearest neighbors of x, mode(·) is the mode operation;

[0052] Support Vector Machine: SVM distinguishes samples of different categories by finding the hyperplane that maximizes the margin:

[0053]

[0054] Where: W is the weight vector, b is the bias term;

[0055] (2) Neural rules

[0056] The neural rules are mainly constructed by multi-layer perceptron, which is a feedforward neural network consisting of multiple layers, including input layer, hidden layer and output layer. Its output can be expressed as:

[0057]

[0058] Where x is the input feature vector, W i and b i are the weight matrix and bias vector of the i-th layer respectively. σ is the activation function. Two neural rules are constructed according to the different loss functions. Their loss functions are different:

[0059] For the binary classification problem, a cross entropy loss function is used, which is expressed as:

[0060]

[0061] Where N is the number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability that the model predicts that the i-th sample is a positive class;

[0062] Another method is to learn feature representation by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs through contrastive learning. The contrastive loss function is defined as:

[0063]

[0064] Among them, y i is a sample pair (x i ,xj ) label, y i =0 means x i and x j Belong to the same class of samples, y i =1 means x i and x j For samples belonging to different categories, for MLP, we can define the feature representation function f as the output of the network, that is, f(x) = MLP(x). Then, we use the above contrast loss function to train MLP so that the feature representation learned by the network can distinguish samples of different categories.

[0065] Furthermore, in the training phase, S3 first independently trains multiple single rules R i , each single rule R i Generate predictions on the training set Where x is the original input feature, the output of these base learners is concatenated with the original input feature and used as the input of the new learning task:

[0066]

[0067] The second layer of learning is performed through the meta-learner, which also uses a multi-layer perceptron and a contrastive learning method to extract the hidden layer embedding results that have been trained contrastively, connect these prediction results with the original features of the sample, and input them into the neural-feature rule fusion module to predict the final classification results. The fusion module uses a three-layer multi-layer perceptron structure to achieve more accurate fraud detection. The basic structure of MLP:

[0068]

[0069] Among them, x new is the input feature vector, W i and b i are the weight matrix and bias vector of the i-th layer, σ is the activation function, and ReLU is used here. The loss function uses formula (7). Finally, our model can be based on the output of the meta-learner Provide accurate detection results, which is a confidence level between 0 and 1. Set the fraud risk prevention and control threshold θ according to the prevention and control needs. When the transaction is judged as risky, It is considered as a normal transaction.

[0070] Another object of the present invention is to provide a credit card risk prevention and control system based on Stacking ensemble learning, which specifically includes:

[0071] Data preprocessing module, used to refine features and balance positive and negative samples;

[0072] Rule building module, used to build feature rules and neural rules;

[0073] Rule fusion module, used to fuse feature rules and neural rules.

[0074] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0075] First, the present invention can improve the accuracy and timeliness of early warning: The present invention can more accurately identify and predict risks, thereby improving the accuracy and timeliness of early warning. Through the intelligent construction technology of multiple rules, the automatic extraction and expression of feature rules and neural rules are realized, solving the problem that complex risk data rules are difficult to summarize and learn. The combination of feature rules and neural rules enhances the adaptability and flexibility of the rules, enabling social governance and risk prevention and control rules to better cope with the ever-changing social environment.

[0076] The present invention can enhance the adaptability and flexibility of rules: intelligent construction technology improves the adaptability and flexibility of rules, enabling social governance and risk prevention and control rules to more effectively cope with the ever-changing social environment. The present invention solves the problems of ambiguity, conflict and inconsistency in current social governance and risk prevention and control rules through the fusion and optimization of multiple rules. It realizes intelligent rule extraction and expression, improves the intelligence level of the system platform, the completeness of early warning rules and processing efficiency, and provides a more accurate and efficient solution for social governance and risk prevention and control.

[0077] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are as follows: the technical solution of the present invention has significant economic benefits in the field of fraud detection and risk management of financial institutions and credit card issuers. Through the fusion of multiple rules and intelligent construction methods, the present invention can more accurately identify and intercept fraudulent transactions, thereby effectively reducing fraud losses. This more efficient risk identification mechanism directly improves the security of transactions and provides customers with a more reliable transaction environment. At the commercial level, by improving the accuracy of fraud detection and reducing the false alarm rate, the present invention helps to reduce customers' unnecessary worries and inconveniences, enhance customers' trust in financial institutions, and thus improve customer loyalty. At the same time, more accurate fraud detection can reduce business interruptions caused by false alarms, thereby optimizing user experience and improving customer satisfaction. More importantly, since the present invention can provide accurate risk identification and prevention and control measures, financial institutions can further increase their market share and enhance their market influence in a highly competitive market. By reducing potential fraud losses and reducing operating costs, the overall income and profitability of financial institutions will be significantly improved. The commercial value of this technical solution is not only reflected in direct economic benefits, but also provides important strategic support for the long-term stable operation and customer relationship management of financial institutions.

[0078] (2) The technical solution of the present invention overcomes technical prejudice:

[0079] The present invention effectively overcomes the technical bias problem of traditional models through integrated learning methods and multi-rule fusion. In fraud detection, a single model usually shows bias towards certain specific types of data, resulting in excessive bias. A single model performs better for certain common types of fraud samples, but has weak recognition capabilities for uncommon or new fraud samples and cannot fully cover diverse fraud patterns.

[0080] The present invention effectively alleviates this problem by integrating multiple rules. Specifically, by using the combination of feature rules and neural rules and the stacking integration method, the system can show consistent excellent results on different types of data, reducing the dependence on a single data model. The application of ensemble learning methods not only improves the accuracy and robustness of the model, but also reduces the bias of the model in identifying certain specific types of fraudulent data by combining diverse data features.

[0081] In addition, the present invention optimizes the contrast loss so that each rule can have a balanced output during the fusion process, which not only improves the effect of rule fusion, but also avoids the dominance of a single rule in a specific scenario. This diversified rule integration method improves the fairness and objectivity of the model, ensures the comprehensiveness and accuracy of risk identification, effectively overcomes the technical bias of a single model, and provides more fair and objective decision support for fraud detection.

[0082] Third, the existing technical problems solved by the technical solution of the present invention in industrial applications

[0083] 1. Limitations of single model performance

[0084] Existing problem: When faced with complex credit card fraud scenarios, a single rule or model is unable to fully capture the potential relationships between multidimensional data due to its limitations, resulting in low prediction accuracy and insufficient generalization ability.

[0085] The solution of the present invention: Based on the Stacking ensemble learning method, through the fusion of feature rules and neural rules, combined with the advantages of shallow machine learning models (such as decision trees, logistic regression, support vector machines, etc.) and deep learning models (such as multi-layer perceptrons), the model's ability to identify complex fraudulent behaviors is significantly improved.

[0086] 2. Sample imbalance problem

[0087] Existing problem: Fraudulent transactions account for a very low proportion of credit card transactions, resulting in an extremely unbalanced data set sample, making the model prone to the majority class (normal transactions), thereby reducing the ability to identify fraudulent behavior.

[0088] The solution of the present invention is as follows: in the data preprocessing stage, by combining upsampling (such as SMOTE, ADASYN) and downsampling (such as Tomek Links), the distribution of positive and negative samples in the data set is effectively balanced, providing balanced data for subsequent model training and improving the classification performance of the model.

[0089] 3. Shortcomings of rule fusion methods

[0090] Existing problems: Traditional rule fusion methods often simply superimpose different rules, fail to fully utilize the complementarity between rules, and find it difficult to achieve efficient rule integration.

[0091] The solution of the present invention is to adopt the Stacking integrated learning framework, conduct secondary learning on the prediction results of multiple single rules through a meta-learner, and introduce a multi-layer perceptron (MLP) as a fusion module to further optimize the feature representation and improve the prediction performance.

[0092] 4. Challenges of Dynamic Changes in Fraud Patterns

[0093] Existing problems: The rapid changes in fraud behaviors and patterns make it difficult for static rules and models to adapt to new types of fraud, resulting in high system maintenance costs and lack of real-time performance.

[0094] The solution of the present invention: through the synergistic fusion of feature rules and neural rules and the dynamic learning ability of the Stacking framework, the adaptability and robustness of the model are enhanced, and new fraud patterns can be quickly captured.

[0095] Significant technical advancement of the present invention

[0096] 1. Significantly improve prediction accuracy and robustness

[0097] This invention combines feature rules with neural rules and stacking ensemble learning methods to more comprehensively mine data features and improve the ability to capture complex fraud patterns. Experiments show that the fraud detection accuracy, recall rate and F1 score of this method are significantly higher than those of traditional single models and simple rule stacking methods.

[0098] 2. Solve the problem of sample imbalance

[0099] By combining sampling techniques such as SMOTEENN and SMOTETomek, the problem of uneven distribution of positive and negative samples in the dataset is solved, the model's recognition ability for minority classes (fraudulent transactions) is improved, and the missed detection rate and misjudgment rate are significantly reduced.

[0100] 3. Achieve efficient rule integration

[0101] The Stacking framework is introduced to perform secondary learning of multi-rule prediction results through a meta-learner, which effectively integrates the advantages of shallow models and deep learning models and optimizes the synergy between rules.

[0102] The use of the MLP structure further improves the nonlinear expression capability of the fusion process, enabling the model to maintain efficient decision-making performance in complex scenarios.

[0103] 4. Enhance the ability to adapt to dynamic risks

[0104] The present invention is based on the neural rules of the contrastive learning method, and by continuously optimizing the deep-level feature expression, it improves the adaptability to the dynamic changes of fraudulent behavior and realizes the rapid detection of new fraud patterns.

[0105] 5. Provide flexible risk prevention and control threshold settings

[0106] In the final prediction stage, the model outputs a confidence score from 0 to 1, and users can set the fraud risk prevention and control threshold (θ) according to prevention and control needs. This flexibility allows the system to adjust the detection strategy according to different scenarios (such as high-risk transactions or large transactions), thereby reducing risks while reducing interference with normal transactions.

[0107] 6. Reduce operation and maintenance costs

[0108] Based on the dynamic characteristics of ensemble learning, the present invention reduces the reliance on frequent rule updates and reduces system maintenance costs. At the same time, combined with the high efficiency of multi-rule fusion, it shortens the model training and reasoning time and improves the operating efficiency of the system.

[0109] The technical solution of the present invention uses an innovative Stacking ensemble learning method, combined with the advantages of feature rules and neural rules, to effectively solve the problems of low prediction accuracy, sample imbalance, insufficient rule fusion, and fraud pattern changes in the prior art. It significantly improves the performance and efficiency of fraud detection in industrial applications, provides financial institutions with more powerful risk prevention and control capabilities, reduces operating costs, and provides important technical support for the development of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 It is a flow chart of a credit card risk prevention and control method based on Stacking ensemble learning provided by an embodiment of the present invention;

[0111] Figure 2 It is a schematic diagram of risk prevention and control of Stacking integrated learning provided by an embodiment of the present invention;

[0112] Figure 3 It is a credit card risk prevention and control module diagram based on Stacking ensemble learning provided by an embodiment of the present invention;

[0113] Figure 4 It is the experimental result of constructing a single rule on two types of risk data sets provided by the embodiment of the present invention;

[0114] Figure 5 These are the experimental results of comparing the Stacking ensemble learning method with a single rule on two types of risk data sets provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0115] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0116] Industrial Application Example 1: Bank Credit Card Transaction Fraud Detection System

[0117] In bank credit card transactions, fraudulent activities (such as stolen credit cards, forged identity transactions, etc.) pose a serious threat to user assets and bank reputation. Existing rules and models are difficult to meet the needs of complex transaction scenarios due to their singleness and static nature, and are prone to false positives and false negatives.

[0118] The bank adopts the credit card risk prevention and control method based on the Stacking ensemble learning of the present invention to analyze and assess the risks of real-time transaction data, thereby detecting and preventing suspicious fraudulent transactions.

[0119] 1. Data collection and preprocessing

[0120] Obtain transaction data from the bank's credit card system, including transaction amount, time, geographic location, transaction device and other characteristics.

[0121] Use SMOTEENN to balance the ratio of positive and negative samples to ensure that the model can identify fraudulent transactions.

[0122] 2. Model Construction

[0123] Feature rules: Use traditional machine learning models such as classification decision trees and logistic regression to analyze transaction features (such as anomalies in time and location) and generate preliminary predictions.

[0124] Neural rules: Based on multi-layer perceptron (MLP), it deeply analyzes complex patterns and potential correlations in trading behaviors.

[0125] 3. Stacking ensemble learning

[0126] The outputs of feature rules and neural rules are combined with the original transaction data as input to train the MLP-based meta-learner, fuse the prediction results, and generate the final risk score.

[0127] 4. Real-time risk detection

[0128] In actual transactions, the system determines the risk category of a transaction based on the risk score (confidence between 0 and 1) and the preset threshold θ:

[0129] Above the threshold: Marked as a fraudulent transaction, triggering the bank's risk control mechanism (such as freezing the account, notifying the user by SMS).

[0130] Below the threshold: normal transaction, automatically released.

[0131] The detection accuracy rate increased to over 95%, the false alarm rate decreased by 30%, and the missed alarm rate decreased by 25%.

[0132] The system achieves a response to real-time transactions within seconds, ensuring user experience while reducing the bank’s economic losses.

[0133] Industrial application example 2: Third-party payment platform anti-fraud system

[0134] Third-party payment platforms need to deal with a variety of fraudulent behaviors (such as account theft, malicious refunds, and false payments). Traditional rule-based prevention and control methods are difficult to adapt to complex and changing fraud patterns, and are prone to high false alarm rates, affecting normal user experience.

[0135] The payment platform introduces the Stacking integrated learning method of the present invention to perform real-time analysis and risk control of payment behaviors, thereby reducing fraud losses and optimizing user experience.

[0136] 1. Data collection and preprocessing

[0137] Collect platform payment data, including transaction amount, payment method, account history, IP address, device type, etc.

[0138] The SMOTETomek method is used to process unbalanced data and optimize the ratio of positive and negative samples.

[0139] 2. Model construction and fusion

[0140] Feature rules: Use support vector machine (SVM) and K-nearest neighbor algorithm (KNN) to extract abnormal patterns of payment behavior from simple features.

[0141] Neural rules: Analyze complex payment behavior characteristics and discover potential fraud associations through multi-layer perceptron (MLP).

[0142] Stacking integration: Use an MLP-based meta-learner to fuse feature rules and neural rules to generate the final payment risk score.

[0143] 3. Risk Management Process

[0144] The system dynamically adjusts the prevention and control strategy based on the risk score:

[0145] High-risk transactions: Restrict payment operations and send a verification SMS to the user for secondary confirmation.

[0146] Medium-risk transactions: Marked as suspicious and subject to manual backend review.

[0147] Low-risk transactions: payments are completed automatically.

[0148] 4. Real-time update and optimization

[0149] The system regularly updates the sample library and optimizes model performance through online learning.

[0150] Introduce a user feedback mechanism to further improve the adaptability and accuracy of the model.

[0151] The payment fraud detection accuracy rate reached 97%, and the complaint rate of platform users decreased by 40%.

[0152] The system's false alarm rate dropped by 35%, saving approximately 15% of daily manual review costs, greatly improving operational efficiency.

[0153] like Figure 1 As shown, an embodiment of the present invention provides a credit card risk prevention and control method based on Stacking ensemble learning, and the method specifically includes:

[0154] S1: Dataset preprocessing, extracting key features, and using resampling technology to balance positive and negative samples;

[0155] S2: Use a variety of machine learning methods to build shallow feature rules, and use deep learning technologies such as multi-layer perceptrons to build deep neural rules;

[0156] S3: Based on the Stacking ensemble learning method, it integrates feature rules and neural rules.

[0157] The embodiment of the present invention provides a credit card risk prevention and control method based on Stacking ensemble learning, which aims to improve the accuracy of risk identification through multi-level feature fusion and model optimization. First, in the data set preprocessing stage, the system will clean and organize the credit card data, and select key features that can effectively reflect the user's credit risk, such as transaction amount, repayment status, etc. In order to solve the problem of imbalance between positive and negative examples in sample data, resampling techniques, such as undersampling or oversampling, are used to achieve sample balance, thereby improving the robustness of the model.

[0158] Next, shallow feature rules are constructed through a variety of machine learning algorithms. For example, traditional machine learning methods such as decision trees and logistic regression are used to extract the credit performance characteristics of customers in different dimensions and form a variety of shallow feature rules. These shallow rules can obtain basic feature patterns from the data, which can provide a basis for deep feature extraction in the subsequent stage.

[0159] In the process of deep feature extraction, deep learning algorithms such as multilayer perceptron (MLP) are used to build a deep neural network model. By analyzing and modeling the nonlinear relationship of key features through neural networks, deep neural networks can dig out more complex credit risk patterns, which provides a more accurate basis for credit card risk identification and makes up for the shortcomings of shallow feature rules.

[0160] Finally, based on the Stacking ensemble learning method, shallow feature rules and deep neural rules are fused. By setting up multiple primary models (such as shallow feature rules and neural rule models) and using a secondary learner for final prediction, the advantages of the models can be complemented. Stacking ensemble learning can effectively utilize the characteristics of different models and ultimately generate a stable and efficient credit card risk prevention and control strategy, providing banks with accurate risk assessment support.

[0161] S1, in actual real life, risk prevention and control data sets are difficult to collect, and naturally have the problem of imbalance between positive and negative samples. That is, the number of fraudulent transactions in real life is far less than the number of normal transactions. Therefore, if you want to obtain a powerful model through the risk prevention and control data set, it is very necessary to have a balancing strategy for it. After preprocessing the data set to extract key features, some resampling techniques are used to balance the positive and negative samples.

[0162] Define the preprocessed data set as D, where the number of positive and negative data samples is M and N respectively (generally, M<<N), then the number of data in D is expressed as |D|=M+N, and the resampled data set is D new (1) Upsampling

[0163] Random upsampling: By randomly copying minority class samples, the number of positive and negative samples is balanced.

[0164] Copy NM minority class samples. |D new |=2*N

[0165] SMOTE: Synthetic Minority Oversampling Technique, which increases the number of minority class samples by inserting newly generated samples between them.

[0166] Assume that k new samples are generated, and each minority class sample participates in generating new samples. new |=M*k+N

[0167] ADASYN: An adaptive synthetic sampling method that adjusts the number of new samples generated according to the number of minority class nearest neighbors of each minority class sample.

[0168] For each minority class sample x i , generate k i new samples, where k i is x i The number of nearest neighbors of the minority class.

[0169] (2) Downsampling

[0170] Random downsampling: Reduce the number of majority class samples to match the number of minority class samples by randomly removing them.

[0171] Remove NM majority class samples. |D new |=2*M

[0172] TomekLinks: Reduces noise by identifying and removing minority class samples that are too close to majority class samples.

[0173] Assume that T majority class noise samples are removed. |D new |=M+NT

[0174] (3) Combination of upsampling and downsampling

[0175] SMOTEENN: Combining the methods of SMOTE and Edited Nearest Neighbors (ENN), first generate new samples through SMOTE, and then use ENN to remove new samples that are too close to the majority class samples.

[0176] Assume that k new samples are generated by SMOTE, and then T samples are removed by ENN.

[0177] D new |=M+N+kT

[0178] SMOTETomek: Combines SMOTE and Tomek Links methods to first remove minority class samples that are too close to majority class samples, and then use SMOTE to generate new samples.

[0179] Assume that k new samples are generated by SMOTE, and then T samples are removed by Tomek Links. new |=M+N+kT

[0180] In S2, with the help of diverse machine learning methods, we successfully constructed shallow feature rules and thoroughly verified them through extensive experiments. Furthermore, the remarkable feature extraction ability of deep learning technology and its powerful ability to handle complex nonlinear relationships provide a breakthrough method for credit card fraud detection. Deep learning technologies such as multi-layer perceptrons are used to establish deep neural rules.

[0181] Define input data D new ={(x1,y1),(x2,y2),…,(x n ,y n )}, where x i is the eigenvector, y i is the corresponding label (0 or 1). For each feature x of a data i , we construct a single rule R(·) through machine learning methods to obtain the corresponding prediction output

[0182] (1) Feature rules

[0183] Classification decision tree: Build decision rules by learning the features of samples. For each node, we select the best features and thresholds to split the data:

[0184]

[0185] Among them, x j is the selected feature. t is the segmentation threshold.

[0186] Logistic regression: Logistic regression is used for binary classification problems, and its prediction formula is:

[0187]

[0188] Among them, β0, β1, β2…β n are the parameters of logistic regression.

[0189] K-nearest neighbor method: K-NN predicts the category of a new sample based on the categories of the nearest k neighbors:

[0190]

[0191] Among them, {y1,y2,…,y k} is the category of the k nearest neighbors of x, and mode(·) is the mode operation.

[0192] Support Vector Machine: SVM distinguishes samples of different categories by finding the hyperplane that maximizes the margin:

[0193]

[0194] Where: W is the weight vector. b is the bias term.

[0195] (2) Neural rules

[0196] Neural rules are mainly constructed by multi-layer perceptrons. Multi-layer perceptrons are a feedforward neural network consisting of multiple layers, including input layer, hidden layer and output layer. Its output can be expressed as:

[0197]

[0198] Where x is the input feature vector. i and b i are the weight matrix and bias vector of the i-th layer respectively. σ is the activation function.

[0199] Two neural rules are constructed based on different loss functions, and their loss functions are different:

[0200] For the binary classification problem, a cross entropy loss function is used, which is expressed as:

[0201]

[0202] Where N is the number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability that the model predicts that the i-th sample is a positive class.

[0203] Another method is to learn feature representation by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs through contrastive learning. The contrastive loss function is defined as:

[0204]

[0205] Among them, y i is a sample pair (x i ,x j ) label, y i =0 means x i and x j Belong to the same class of samples. i =1 means x i and x jBelong to different classes of samples. For MLP, we can define the feature representation function f as the output of the network, that is, f(x) = MLP(x). Then, we use the above contrast loss function to train MLP so that the feature representation learned by the network can distinguish samples of different categories.

[0206] In the field of traditional rule extraction and expression, although progress has been made in the construction of single rules and the difficulties of rule extraction and expression have been solved, there is still room for improvement in improving prediction accuracy. The S3, by combining feature rules with neural rules, achieves complementary advantages and significantly improves the performance of the prediction model. This fusion strategy not only optimizes the performance of existing rules, but also provides a more accurate strategy for solving complex problems.

[0207] The core of this technology is the Stacking ensemble learning method, which improves prediction performance by combining multiple different machine learning and deep learning models. The basic idea of ​​Stacking is to use the prediction results of multiple base learners (basic models) as the input of the meta learner (combined model), thereby improving the prediction ability of the overall model.

[0208] In the training phase, we first independently train multiple single rules R i . Each single rule R i Generate predictions on the training set Where x is the original input feature. The output of these base learners is concatenated with the original input feature and used as the input of the new learning task:

[0209]

[0210] The second layer of learning is performed through a meta-learner. The meta-learner can be a suitable model such as logistic regression, gradient boosting tree, or deep neural network, and its goal is to learn how to best combine the outputs of the base learners. Here we also use a multi-layer perceptron as the meta-learner. In addition to the contrastive learning method, we extract the hidden layer embedding results that have been contrastively trained. These prediction results are connected with the original features of the sample and input into the neural-feature rule fusion module to predict the final classification result. This fusion module uses a three-layer multi-layer perceptron structure to achieve more accurate fraud detection. The basic structure of MLP:

[0211]

[0212] Among them, x new is the input feature vector. W i and b i are the weight matrix and bias vector of the i-th layer respectively. σ is the activation function, and ReLU is used here. The loss function uses formula (7). Finally, our model can be based on the output of the meta-learner Provide accurate detection results. It is a confidence level between 0 and 1. According to the prevention and control needs, set the fraud risk prevention and control threshold θ. When the transaction is judged as risky, It is considered as a normal transaction.

[0213] Figure 2 This is a schematic diagram of Stacking ensemble learning risk prevention and control.

[0214] like Figure 3 As shown, a credit card risk prevention and control system based on Stacking ensemble learning provided by an embodiment of the present invention specifically includes:

[0215] Data preprocessing module, used to refine features and balance positive and negative samples;

[0216] Rule building module, used to build feature rules and neural rules;

[0217] Rule fusion module, used to fuse feature rules and neural rules.

[0218] 1. Specific application fields or related products of the present invention.

[0219] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0220] 1. Data preparation

[0221] The European credit card fraud dataset is derived from 284,807 transaction records of European credit card users within two days in September 2013, including 492 fraudulent transactions, accounting for 0.172% of the total number of transactions. Each transaction data contains 30-dimensional features, of which 28-dimensional features are preprocessed into numerical attributes through PCA transformation mapping, and the two-dimensional features of transaction time and amount are not PCA transformed. It is expected to predict transaction risks through 30-dimensional features, and the transaction type is represented by a binary variable, 0 for normal transactions and 1 for fraudulent transactions. This dataset is a public dataset. The original dataset has been preprocessed and directly divided into 80% training set and 20% test set. Through random sampling without replacement, 227,452 negative examples (normal transaction samples) and 392 positive examples (fraudulent transaction samples) of the original training set are obtained, and 56,863 negative examples and 100 positive examples of the test set are obtained.

[0222] The IBM anti-money laundering transaction dataset contains attributes such as time, initiating account, receiving account, and amount. In the preprocessing, the payment method attributes are manually encoded according to the complexity of the payment. Then the timestamp is converted to the number of seconds since the Unix epoch (1970-01-01). Label Encoding is used to quantify the remaining object type columns, and finally the data is normalized. It is expected to predict transaction risks through 10-dimensional features, and the transaction type is represented by a binary variable, 0 for normal transactions and 1 for fraudulent transactions. This dataset is a public dataset. The original dataset is divided into 80% training set and 20% test set. Through random sampling without replacement, the original training set has 811,705 negative examples (normal transaction samples) and 828 positive examples (fraudulent transaction samples), and the test set has 202,927 negative examples and 209 positive examples.

[0223] 2. Indicator design

[0224] In response to the above model evaluation indicator selection problem, this study proposed a method to control the first type error rate (FNR), that is, to minimize the missed positive examples, and reduce the second type error rate (FPR) under this premise. In this way, the second type error rate (FPR) is reduced under the premise of controlling the first type error rate (FNR). This strategy aims to ensure that the model reduces the false positive examples while also reducing the false negative examples.

[0225] By controlling FNR, the model can more accurately identify fraudulent behavior and reduce the risk of missed reports. This is crucial for financial institutions, e-commerce platforms, etc., because missed reports of fraudulent behavior lead to significant economic losses. Lowering FPR while controlling FNR means that the model can also maintain a high level of accuracy when identifying non-fraudulent behavior and reduce unnecessary false positives. In addition, paying attention to both FNR and FPR can better balance the model's performance in identifying positive and negative examples, making the model more robust. At the same time, it can more intuitively understand the model's performance in the fraud detection process, which is helpful for subsequent tuning and improvement of the model.

[0226] Table 1 Confusion matrix

[0227]

[0228] The table shows the four types of prediction results of the confusion matrix. The correct prediction is TP, which means the actual situation is risky and the predicted situation is risky; the first type of error is FN, which means the actual situation is risky and the predicted situation is risk-free; the second type of error is FP, which means the actual situation is risk-free but the model misjudges it as risky; and the correct prediction is TN, which means the actual situation is risk-free and the predicted situation is risk-free. The false positive rate (FPR) and false negative rate (FNR) expressed in formula (1) and formula (2) are used to evaluate the model. The lower the two indicators, the better. FNR = FN / (TP+FN)#(10)

[0229] FPR=FP / (FP+TN)#(11)

[0230] 3. Build a single rule

[0231] Table 2 shows the single rule with the best performance in the European credit card fraud dataset and the IBM anti-money laundering transaction dataset test set experiments. By controlling the risk prevention and control threshold θ = 0.5, the European credit card fraud dataset FNR is controlled to 3%, and the IBM anti-money laundering transaction dataset FNR is controlled to 9%.

[0232] Table 2 Single rule experimental results

[0233]

[0234]

[0235] 4. Perform multi-rule fusion

[0236] The results obtained by the Stacking fusion method are shown in Table 3. By controlling the risk prevention and control threshold θ = 0.5, the European credit card fraud dataset FNR = 3%, and the IBM anti-money laundering transaction dataset FNR = 9%

[0237] Table 3 Stacking fusion experiment results

[0238] Rule Name TP TN FP FN FNR FPR Stacking 97 55794 1069 3 0.03 0.0188 Stacking 190 192915 10012 19 0.09 0.0431

[0239] Extensive experiments conducted on the European credit card fraud risk dataset show that, compared with the single rule method, the FPR is reduced by (0.0310-0.0188) / 0.0310=39.4% while maintaining the FNR at 0.03.

[0240] A large number of experiments conducted on the IBM anti-money laundering transaction dataset show that, compared with the single rule method, the FPR is reduced by (0.0510-0.0431) / 0.0510=15.5% while maintaining the FNR at 0.09.

[0241] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0242] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A credit card risk prevention and control method based on Stacking ensemble learning, characterized in that: The method specifically includes: S1: Dataset preprocessing, extracting key features, and using resampling technology to balance positive and negative samples; S2: Use a variety of machine learning methods to build shallow feature rules, and use deep learning technologies such as multi-layer perceptrons to build deep neural rules; S3: Based on the Stacking ensemble learning method, it integrates feature rules and neural rules.

2. The credit card risk prevention and control method based on Stacking ensemble learning as claimed in claim 1, characterized in that: In S1, some resampling techniques are used to balance the positive and negative samples. The preprocessed data set is defined as D, where the positive and negative data sets are M and N respectively (generally, M<<N). The number of data in D is expressed as |D|=M+N. Let the resampled data set be D new : (1) Upsampling Random upsampling: By randomly copying minority class samples, the number of positive and negative samples is balanced, and NM minority class samples are copied. new |=2*N; SMOTE: Synthetic Minority Oversampling Technology, which increases the number of minority samples by inserting newly generated samples between them. Assuming k new samples are generated, each minority sample participates in generating new samples, D new |=M*k+N; ADASYN: Adaptive synthetic sampling method, which adjusts the number of new samples generated according to the number of minority class nearest neighbors of each minority class sample. For each minority class sample x i , generate k i new samples, where k i is x i The number of nearest neighbors of the minority class. (2) Downsampling Random downsampling: By randomly removing majority class samples, reduce their number to match the number of minority class samples, remove NM majority class samples, |D new |=2*M; TomekLinks: Reduce noise by identifying and removing minority class samples that are too close to majority class samples. Suppose T majority class noise samples are removed, |D new |=M+NT; (3) Combination of upsampling and downsampling SMOTEENN: Combining SMOTE and Edited Nearest Neighbors (ENN), first generate new samples through SMOTE, and then use ENN to remove new samples that are too close to the majority class samples. Suppose k new samples are generated through SMOTE, and then T samples are removed through ENN. new |=M+N+kT; SMOTETomek: Combining SMOTE and Tomek Links, first remove minority class samples that are too close to majority class samples, and then use SMOTE to generate new samples. Suppose k new samples are generated by SMOTE, and then T samples are removed by Tomek Links. new |=M+N+kT.

3. The credit card risk prevention and control method based on Stacking ensemble learning as claimed in claim 1, characterized in that: S2, defines the input data D new ={(x1,y1),(x2,y2),…,(x n ,y n )}, where x i is the eigenvector, y i is the corresponding label (0 or 1), for each feature x of the data i , a single rule R(·) is constructed through machine learning methods to obtain the corresponding prediction output (1) Feature rules Classification decision tree: Build decision rules by learning the features of samples. For each node, select the best features and thresholds to split the data: Among them, x j is the selected feature. t is the segmentation threshold; Logistic regression: Logistic regression is used for binary classification problems, and its prediction formula is: Among them, β0, β1, β2…β n are the parameters of logistic regression; K-nearest neighbor method: K-NN predicts the category of a new sample based on the categories of the nearest k neighbors: Among them, {y1,y2,…,y k } is the category of the k nearest neighbors of x, mode(·) is the mode operation; Support Vector Machine: SVM distinguishes samples of different categories by finding the hyperplane that maximizes the margin: Where: W is the weight vector, b is the bias term; (2) Neural rules The neural rules are mainly constructed by multi-layer perceptron, which is a feedforward neural network consisting of multiple layers, including input layer, hidden layer and output layer. Its output can be expressed as: Where x is the input feature vector, W i and b i are the weight matrix and bias vector of the i-th layer respectively. σ is the activation function. Two neural rules are constructed according to the different loss functions. Their loss functions are different: For the binary classification problem, a cross entropy loss function is used, which is expressed as: Where N is the number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability that the model predicts that the i-th sample is a positive class; Another method is to learn feature representation by contrastive learning to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs. The contrastive loss function is defined as: Among them, y i is a sample pair (x i ,x j ) label, y i =0 means x i and x j Belong to the same class of samples, y i =1 means x i and x j For samples belonging to different categories, for MLP, we can define the feature representation function f as the output of the network, that is, f(x) = MLP(x). Then, we use the above contrast loss function to train MLP so that the feature representation learned by the network can distinguish samples of different categories.

4. The credit card risk prevention and control method based on Stacking ensemble learning as claimed in claim 1, characterized in that: In the training phase, S3 first independently trains multiple single rules R i , each single rule R i Generate predictions on the training set Where x is the original input feature, the output of these base learners is concatenated with the original input feature and used as the input of the new learning task: The second layer of learning is performed through the meta-learner, which also uses a multi-layer perceptron and a contrastive learning method to extract the hidden layer embedding results that have been trained contrastively, connect these prediction results with the original features of the sample, and input them into the neural-feature rule fusion module to predict the final classification results. The fusion module uses a three-layer multi-layer perceptron structure to achieve more accurate fraud detection. The basic structure of MLP: Among them, x new is the input feature vector, W i and b i are the weight matrix and bias vector of the i-th layer, σ is the activation function, and ReLU is used here. The loss function uses formula (7). Finally, our model can be based on the output of the meta-learner Provide accurate detection results, which is a confidence level between 0 and 1. Set the fraud risk prevention and control threshold θ according to the prevention and control needs. When the transaction is judged as risky, It is considered as a normal transaction.

5. A credit card risk prevention and control system based on Stacking ensemble learning as described in claims 1-4, characterized in that: The system specifically includes: The dataset processing module is used to preprocess the original credit card transaction dataset, extract key features, and balance positive and negative samples using resampling technology; A shallow rule generation module, connected to the data set processing module, for generating shallow feature rules based on a variety of machine learning algorithms; The neural rule generation module is connected to the data set processing module and uses a multi-layer perceptron model to extract deep neural feature rules; The integrated learning module is connected to the shallow rule generation module and the neural rule generation module, and uses the Stacking integrated learning method to fuse the shallow feature rules and neural rules to output the final risk judgment results.

6. The system according to claim 1, wherein: The data set processing module further comprises: Upsampling units, including random upsampling units, SMOTE units, and ADASYN units, are used to replicate or generate minority class samples to balance sample data; Downsampling units, including random downsampling units and TomekLinks units, are used to remove majority class samples or noise data to reduce imbalance; The up- and down-sampling combination unit combines the SMOTEENN and SMOTETomek methods to generate and optimize the balance of sample sets through specific rules.

7. The system according to claim 1, characterized in that The shallow rule generation module further includes: Decision tree module, used to construct rule trees for credit risk based on feature thresholds; Logistic regression module, used to build risk classification models based on linear combination features; K-nearest neighbor module, which infers the risk of transactions based on the categories of the nearest neighbors; The support vector machine module separates different transaction categories by maximizing the hyperplane of the margin.

8. The system of claim 1, wherein: The neural rule generation module further comprises: Multi-layer perceptron building block, extracting deep features based on multiple hidden layers; The first loss function optimization unit uses the cross entropy loss function to generate a classification model for 0 / 1 class labels; The contrastive learning optimization unit adopts the contrastive loss function and the embedding vector obtained through contrastive learning training is used to improve the model's ability to distinguish between positive and negative samples.

9. The system according to claim 1, characterized in that The integrated learning module further comprises: A multi-model fusion module is used to receive and combine the prediction results output by the shallow rule generation module and the neural rule generation module; The meta-learner module further learns the multi-model fusion results through a multi-layer perceptron structure to generate the final risk prediction confidence.

10. The system according to claim 1, wherein: It also includes a risk judgment module, which is connected to the integrated learning module and generates a risk judgment based on the final predicted risk confidence. When the predicted confidence exceeds the set threshold, the system judges it as a high-risk transaction, otherwise it is judged as a normal transaction.

Images