Financial fraud detection model training method and device and medium

Abstract

The invention relates to a financial fraud detection model training method and device and a medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring a financial fraud data set; preprocessing the financial fraud data set to obtain standardized financial data; using the standardized financial data to train a TabTransform model, and obtaining a first prediction result; using the standardized financial data to train an XGBoot model to obtain a second prediction result; training a LightGBM model by using the standardized financial data to obtain a third prediction result; and performing weighted voting on the first prediction result, the second prediction result and the third prediction result through a weighted integration method to obtain a final prediction result. The model obtained by using the scheme can accurately identify financial fraud data.

Description

Technical Field

[0001] The present disclosure belongs to the field of artificial intelligence technology, and in particular relates to a financial fraud detection model training method, device and medium. Background Art

[0002] Current financial fraud detection models are generally based on a single model, which results in an inability to accurately identify whether financial data contains fraud. Specifically, these models include TabTransformer, XGBoost, and LightGBM.

[0003] TabTransformer is a deep learning model based on the Transformer architecture, specifically designed for tabular data such as financial transaction data and customer behavior data. Its core concept is to use a self-attention mechanism to process the complex relationships between features in tabular data. This approach aims to address the shortcomings of traditional machine learning algorithms in processing tabular data with diverse and heterogeneous features. However, this single model suffers from high computational complexity, poor interpretability, and demanding preprocessing.

[0004] XGBoost is an efficient implementation method based on gradient boosting and is widely used in various classification tasks including financial fraud detection. Its key advantage is that it combines multiple weak classifiers into a strong classifier and gradually adjusts the weight of each tree to improve the accuracy of the model. However, this single model is sensitive to parameters, has low computational efficiency, and suffers from overfitting problems.

[0005] LightGBM can effectively handle missing values ​​and categorical features. Its histogram algorithm enables better memory and computational efficiency when processing large-scale data. However, this single model is highly complex, handles missing values ​​poorly, and requires complex parameter tuning.

[0006] As mentioned above, the above-mentioned model is prone to errors in identifying fraud in financial data, which may cause fraudulent data to escape legal scrutiny and cause unbearable losses. Summary of the Invention

[0007] This disclosure proposes a financial fraud detection model training method, device, and medium to address the problem of inaccurate identification of whether financial data contains fraud by existing models.

[0008] According to a first aspect of the present disclosure, a financial fraud detection model training method is provided, the method comprising: obtaining a financial fraud dataset; preprocessing the financial fraud dataset to obtain standardized financial data; using the standardized financial data to train a TabTransformer model to obtain a first prediction result; using the standardized financial data to train an XGBoot model to obtain a second prediction result; using the standardized financial data to train a LightGBM model to obtain a third prediction result; and performing weighted voting on the first prediction result, the second prediction result, and the third prediction result through a weighted integration method to obtain a final prediction result.

[0009] In some embodiments, preprocessing the financial fraud dataset to obtain standardized financial data includes: deleting erroneous records from the financial fraud dataset to obtain initial data; and performing mean filling on the initial data to obtain standardized financial data.

[0010] In some embodiments, the standardized financial data includes category characteristic data and numerical characteristic data, the category characteristic data includes transaction type data, source account data, and destination account data, and the numerical characteristic data includes transaction time data, transaction amount data, pre-transaction source account amount data, and post-transaction source account amount data.

[0011] In some embodiments, the use of the standardized financial data to train the TabTransformer model to obtain a first prediction result includes: using LabelEncoder to perform integer conversion on the category feature data to obtain integer data; embedding the integer data to obtain an embedding vector; processing the embedding vector to learn the contextual relationship between each feature in the embedding vector to obtain first data to be relied upon; performing a linear scaling operation on the numerical feature data to obtain second data to be spliced; performing feature vector splicing and pooling processing on the first data to be spliced ​​and the second data to be spliced ​​to obtain financial data to be input; inputting the financial data to be input into the Inception-ResNet structure for residual processing to obtain financial data after residual processing; inputting the financial data after residual processing into a multi-layer perceptron to obtain a first prediction result.

[0012] According to a second aspect of the present disclosure, a financial fraud detection model training device is provided, including: a financial fraud data set acquisition module, used to acquire a financial fraud data set; a financial fraud data preprocessing module, used to preprocess the financial fraud data set to obtain standardized financial data; a TabTransformer model training module, used to train the TabTransformer model using the standardized financial data to obtain a first prediction result; an XGBoot model training module, used to train the XGBoot model using the standardized financial data to obtain a second prediction result; a LightGBM model training module, used to train the LightGBM model using the standardized financial data to obtain a third prediction result; a prediction result weighted voting module, used to perform weighted voting on the first prediction result, the second prediction result, and the third prediction result through a weighted integration method to obtain a final prediction result.

[0013] According to a third aspect of the present disclosure, a financial fraud detection model training device is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the above-mentioned financial fraud detection model training method based on instructions stored in the memory.

[0014] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the instructions are executed by a processor, the financial fraud detection model training method as described above is implemented.

[0015] By adopting the above technical solution, the embodiments of the present disclosure can achieve the following beneficial technical effects: This solution integrates the combined detection of three models: TabTransformer, XGBoost, and LightGBM, integrates the advantages of each model, and uses weighted integration in the combined detection method. By weighted voting on the prediction results of different models, the final prediction result is obtained, which reduces the misjudgment of a single model in certain special cases and improves the overall prediction accuracy.

[0016] This paper optimizes TabTransformer and introduces the Inception-ResNet structure into TabTransformer. The Inception-ResNet structure introduces residual links on this basis, which not only utilizes the multi-scale feature extraction capability of the Inception module, but also alleviates the gradient vanishing problem in deep networks through skip links. Therefore, due to the training of deep networks and the efficient transmission of feature information, PyTorch's mixed precision training is used. Mixed precision training can reduce the amount of computation and memory usage, thereby accelerating the model training process.

[0017] XGBoost and LightGBM excel at processing structured data, efficiently capturing nonlinear relationships, efficiently handling large datasets, and possessing strong memory optimization capabilities. TabTransformer can capture complex nonlinear relationships between features. Through the Transformer architecture, it can effectively process complex categorical features and learn rich models from tabular data. Its strengths in long-range dependencies and feature interactions can effectively address multidimensional, sparse data, and high-dimensional features, making it particularly suitable for financial fraud detection tasks involving large amounts of transaction records and user information.

[0018] As mentioned above, TabTransformer captures deep feature interactions and long-term dependencies, while XGBoost and LightGBM can efficiently process specific dominant features. The combination of the three can give full play to their respective advantages and enhance the robustness of this solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0020] The present disclosure can be more clearly understood from the following detailed description with reference to the accompanying drawings.

[0021] Figure 1 is a flowchart illustrating a method for training a financial fraud detection model according to some embodiments of the present disclosure.

[0022] Figure 2 is a flow chart illustrating the detection of financial fraud using a combined decision maker according to some embodiments of the present disclosure.

[0023] Figure 3 is a schematic diagram showing an encoder component of a multi-head self-attention layer according to some embodiments of the present disclosure.

[0024] Figure 4 is a flow chart illustrating a method for TabTransformer training according to some embodiments of the present disclosure.

[0025] Figure 5 It is a schematic diagram showing the process of splicing, pooling and residual processing of financial data according to some embodiments of the present disclosure.

[0026] Figure 6 is a block diagram of a financial fraud detection model training device according to some embodiments of the present disclosure.

[0027] Figure 7 It is a block diagram of a financial fraud detection model training device according to some other embodiments of the present disclosure.

[0028] Figure 8is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0029] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure.

[0030] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0031] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0032] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0033] In all examples shown and discussed herein, any specific values ​​should be understood as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0034] It should be noted that like reference numerals and letters refer to like items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0035] Currently, current financial fraud detection models are generally based on a single model, which results in an inability to accurately identify whether financial data contains fraud. Specifically, these models include TabTransformer, XGBoost, and LightGBM.

[0036] TabTransformer is a deep learning model based on the Transformer architecture, specifically designed for tabular data such as financial transaction data and customer behavior data. Its core concept is to use a self-attention mechanism to process the complex relationships between features in tabular data. This approach aims to address the shortcomings of traditional machine learning algorithms in processing tabular data with diverse and heterogeneous features. However, this single model suffers from high computational complexity, poor interpretability, and demanding preprocessing.

[0037] XGBoost is an efficient implementation method based on gradient boosting and is widely used in various classification tasks including financial fraud detection. Its key advantage is that it combines multiple weak classifiers into a strong classifier and gradually adjusts the weight of each tree to improve the accuracy of the model. However, this single model is sensitive to parameters, has low computational efficiency, and suffers from overfitting problems.

[0038] LightGBM can effectively handle missing values ​​and categorical features. Its histogram algorithm enables better memory and computational efficiency when processing large-scale data. However, this single model is highly complex, handles missing values ​​poorly, and requires complex parameter tuning.

[0039] As mentioned above, the above-mentioned model is prone to errors in identifying fraud in financial data, which may cause fraudulent data to escape legal scrutiny and cause unbearable losses.

[0040] TabTransformer: A deep learning model based on the Transformer architecture, designed for supervised and semi-supervised learning tasks of structured tabular data. It dynamically captures complex dependencies between features through a self-attention mechanism, significantly improving the modeling capabilities of tabular data.

[0041] XGBoost: It is an efficient ensemble learning algorithm based on gradient boosting decision tree optimization. The core idea is to iteratively combine multiple weak learners, gradually correct the prediction error, and finally form a strong prediction model.

[0042] LightGBM: An efficient distributed gradient boosting framework developed by Microsoft, designed for large-scale data and high-dimensional feature scenarios.

[0043] In view of this, the present disclosure proposes a financial fraud detection model training method, device and medium. This solution integrates the combined detection of three models: TabTransformer, XGBoost, and LightGBM, integrating the advantages of each model. The combined detection method uses weighted integration to obtain the final prediction result by weighted voting on the prediction results of different models, reducing the misjudgment of a single model in certain special cases and improving the overall prediction accuracy.

[0044] This paper optimizes TabTransformer and introduces the Inception-ResNet structure into TabTransformer. The Inception-ResNet structure introduces residual links on this basis, which not only utilizes the multi-scale feature extraction capability of the Inception module, but also alleviates the gradient vanishing problem in deep networks through skip links. Therefore, due to the training of deep networks and the efficient transmission of feature information, PyTorch's mixed precision training is used. Mixed precision training can reduce the amount of computation and memory usage, thereby accelerating the model training process.

[0045] XGBoost and LightGBM excel at processing structured data, efficiently capturing nonlinear relationships, efficiently handling large datasets, and possessing strong memory optimization capabilities. TabTransformer can capture complex nonlinear relationships between features. Through the Transformer architecture, it can effectively process complex categorical features and learn rich models from tabular data. Its strengths in long-range dependencies and feature interactions can effectively address multidimensional, sparse data, and high-dimensional features, making it particularly suitable for financial fraud detection tasks involving large amounts of transaction records and user information.

[0046] As mentioned above, TabTransformer captures deep feature interactions and long-term dependencies, while XGBoost and LightGBM can efficiently process specific dominant features. The combination of the three can give full play to their respective advantages and enhance the robustness of this solution.

[0047] Figure 1 FIG. 1 is a flow chart illustrating a method for detecting financial fraud according to some embodiments of the present disclosure. Figure 1 As shown, the financial fraud detection method includes steps S110 to S160.

[0048] In step S110 , a financial fraud dataset is obtained.

[0049] In step S120, the financial fraud dataset is preprocessed to obtain standardized financial data.

[0050] In some embodiments, an operation of deleting erroneous records is performed on the financial fraud dataset to obtain initial data; and mean filling is performed on the initial data to obtain standardized financial data.

[0051] In some embodiments, the standardized financial data includes category characteristic data and numerical characteristic data, the category characteristic data includes transaction type data, source account data, and destination account data, and the numerical characteristic data includes transaction time data, transaction amount data, pre-transaction source account amount data, and post-transaction source account amount data.

[0052] In step S130, the TabTransformer model is trained using the standardized financial data to obtain a first prediction result.

[0053] In some embodiments, the use of the standardized financial data to train the TabTransformer model to obtain a first prediction result includes: using LabelEncoder to perform integer conversion on the category feature data to obtain integer data; embedding the integer data to obtain an embedding vector; processing the embedding vector to learn the contextual relationship between each feature in the embedding vector to obtain first data to be relied upon; performing a linear scaling operation on the numerical feature data to obtain second data to be spliced; performing feature vector splicing and pooling processing on the first data to be spliced ​​and the second data to be spliced ​​to obtain financial data to be input; inputting the financial data to be input into the Inception-ResNet structure for residual processing to obtain financial data after residual processing; inputting the financial data after residual processing into a multi-layer perceptron to obtain a first prediction result.

[0054] In step S140, the XGBoot model is trained using the standardized financial data to obtain a second prediction result.

[0055] In step S150, the LightGBM model is trained using the standardized financial data to obtain a third prediction result.

[0056] In step S160, weighted voting is performed on the first prediction result, the second prediction result, and the third prediction result using a weighted integration method to obtain a final prediction result.

[0057] like Figure 2 As shown in Figure 3, the combined decision maker based on this method is divided into three modules.

[0058] First, a module that performs data preprocessing.

[0059] The original input data has characteristics such as missing data, errors, and financial invisibility of merchant accounts. In addition, due to the extreme imbalance of the financial fraud dataset, it is necessary to retain transaction records and perform data preprocessing operations such as mean filling and deleting erroneous records. The judgment formulas for some data errors are as follows:

[0060] oldbalanceOrig CASH_IN -newbalanCeOrig CASH_IN ≠amount;newbalanceDest CASH_IN -OldbalanceDestCASH_IN ≠amount;|newbalanceDest TRANSFER -oldbalanceDest TRANSFER |≠amount; After data cleaning, accounts are categorized into personal and corporate accounts based on the first character of their names. Therefore, two new features, is_merchant_orig and is_merchant_dest, are added to divide all features into two parts. Categorical features, such as type (transaction type), nameOrig (source account), and nameDest (destination account), represent transaction type and account information. These discrete values ​​cannot be directly used as decision maker inputs. Numerical features, such as step (transaction time), amount (transaction amount), oldbalanceOrig (source account balance before transaction), and newbalanceOrig (source account balance after transaction), describe the specific amount and account status of each transaction. These numerical features are directly input to the decision maker after normalization and linear scaling. The data splitting step splits the dataset into training and test sets, with a fixed 20% fraudulent record ratio, to ensure accurate model performance evaluation during training: Data_loader,test_dataloader = train_test_split (test_size = 0.2).

[0061] Second, the module that performs model training.

[0062] During the training process, the same training set data was used to train the TabTransformer, XGBoost, and LightGBM models respectively, and each model learned different types of data patterns: TabTransformer is a deep learning model based on the Transformer architecture, which learns the complex relationship between categorical features through the self-attention mechanism and handles the embedding and interaction of numerical features.

[0063] XGBoost is an ensemble learning method based on gradient boosting. It performs classification through a combination of decision trees and is good at capturing nonlinear relationships between features.

[0064] LightGBM uses a gradient-boosted decision tree method and histogram optimization technology to improve training speed and memory efficiency.

[0065] The characteristics of the above models are shown in the following table:

[0066]

[0067] XGBoost and LightGBM models can directly call the module structure.

[0068] Regarding the structure of the TabTransformer model: After data cleaning, the original input data can be considered to consist of two main parts: categorical features include type (transaction type), nameOrig (source account), nameDest (destination account), etc. These features represent transaction type and account information. They are all discrete values ​​and cannot be directly used as input to the decision maker; numerical features include step (transaction time), amount (transaction amount), oldbalanceOrig (source account amount before the transaction), newbalanceOrig (source account amount after the transaction), etc. These features describe the specific amount and account status of each transaction. They are all numerical features that can be directly input into the decision maker after normalization and linear scaling.

[0069] like Figure 3 As shown, these two types of features need to be processed separately. The preprocessing of the input data primarily involves the following: For each categorical feature (type, nameOrig, and nameDest), LabelEncoder is used to encode them into numeric values, converting each category into a unique integer. For each categorical feature, an EmbeddingLayer is defined to map the category number into a fixed-dimensional vector space. Embedding vectors help capture similarities between categories during subsequent model training. For example, each distinct value of nameOrig and nameDest is mapped to a fixed-dimensional embedding vector, helping the model learn the relationships between different categories. However, the embedding vectors cannot be directly input into the decision maker. Categorical features processed by the embedding layer are fed into the Encoder to further learn the contextual relationships between each feature. The Self-Attention Mechanism in the Transformer model helps the model focus on the relative importance of different features, thereby capturing more complex patterns. Each layer attempts to refine and optimize the relationships between input features, helping the model better understand the underlying structure of the categorical features. Finally, the features processed by Transformer Encoder will be fused by pooling operations (maximum pooling and mean pooling) for classification in the subsequent MLP layer.

[0070] like Figure 4 As shown in the figure, for numerical features, the data standardization step is first performed. StandardScaler is used to standardize the numerical features to ensure that the features have the same scale and avoid the influence of scale differences on the prediction performance of some features during model training. Then, the standardized numerical features are input into the fully connected layer for linear transformation. The data flow processing is as follows: Figure 5shown.

[0071] Third, a module that performs model fusion.

[0072] The trained TabTransformer, XGBoost, and LightGBM models are each used to predict on the test data to obtain the prediction results of the three models, which are then fused using a voting-based aggregate decision maker. This voting mechanism can be based on the following method: each model makes a prediction, and the final decision is made by the majority of models voting. If the majority of models predict fraud, the final prediction is fraud, which is a hard voting method, that is,

[0073]

[0074] On the Kaggle public dataset Fraudulent Transactions Dat (financial fraud transaction dataset), the above model made a combined decision, and the final Precision and Recall rates reached 98.82% and 87.53%, respectively. These are both better than the results of using XGBoost, LightGBM, and TabTransformer alone, achieving better classification results and accurate financial fraud detection.

[0075] Kaggle: A global data science competition platform and developer community operated by Google.

[0076] This solution proposes a combined decision maker that integrates TabTransformer, XGBoost and LightGBM, combining the advantages of deep learning and traditional machine learning algorithms to optimize the financial fraud detection method.

[0077] The TabTransformer model can capture complex relationships in category features through the self-attention mechanism of the encoder in the Transformer, and adds a residual module or Inception-ResNet module to the original TabTransformer. At the same time, it combines the efficiency and accuracy of XGBoost and LightGBM in processing high-dimensional features to optimize the financial fraud detection method.

[0078] Experiments were conducted on a public financial fraud dataset on Kaggle to verify the performance differences between single models and combined models. The results showed that the combined decision maker's approach to financial fraud detection significantly improved performance in metrics such as AUC-ROC and F1-score compared to single models.

[0079] Figure 6 is a block diagram illustrating a financial fraud detection model training apparatus according to some embodiments of the present disclosure. Figure 6As shown, the financial fraud detection model training device 600 includes a financial fraud data set acquisition module 610, a financial fraud data preprocessing module 620, a TabTransformer model training module 630, an XGBoot model training module 640, a LightGBM model training module 650, and a prediction result weighted voting module 660.

[0080] The financial fraud dataset acquisition module 610 is configured to acquire a financial fraud dataset;

[0081] a financial fraud data preprocessing module 620 configured to preprocess the financial fraud data set to obtain standardized financial data;

[0082] a TabTransformer model training module 630 configured to train a TabTransformer model using the standardized financial data to obtain a first prediction result;

[0083] An XGBoot model training module 640 is configured to train the XGBoot model using the standardized financial data to obtain a second prediction result;

[0084] A LightGBM model training module 650 is configured to train the LightGBM model using the standardized financial data to obtain a third prediction result;

[0085] The prediction result weighted voting module 660 is configured to perform weighted voting on the first prediction result, the second prediction result, and the third prediction result through a weighted integration method to obtain a final prediction result.

[0086] This device can accurately identify financial fraud data.

[0087] Figure 7 is a block diagram illustrating a financial fraud detection model training apparatus according to some other embodiments of the present disclosure. Figure 7 As shown, the financial fraud detection model training apparatus 700 includes a memory 710 and a processor 720 coupled to the memory 710. The memory 710 is configured to store instructions for executing the corresponding embodiments of the financial fraud detection model training method. The processor 720 is configured to execute the financial fraud detection model training method according to any of the embodiments of the present disclosure based on the instructions stored in the memory 710.

[0088] Figure 7 is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure. Figure 7As shown, computer system 700 may be implemented as a general-purpose computing device. Computer system 700 includes memory 710, processor 720, and bus 730 that connects various system components.

[0089] Memory 710 may include, for example, system memory, non-volatile storage media, and the like. System memory, for example, stores an operating system, application programs, a boot loader, and other programs. System memory may include volatile storage media, such as random access memory (RAM) and / or cache memory. Non-volatile storage media, for example, stores instructions for executing at least one corresponding embodiment of the financial fraud detection model training method. Non-volatile storage media include, but are not limited to, disk storage, optical storage, and flash memory.

[0090] The processor 720 can be implemented using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gates or transistors, or other discrete hardware components. Accordingly, each module, such as the financial fraud dataset acquisition module, the financial fraud data preprocessing module, the TabTransformer model training module, the XGBoot model training module, the LightGBM model training module, and the prediction result weighted voting module, can be implemented by a central processing unit (CPU) executing instructions in a memory that execute corresponding steps, or by dedicated circuits that execute corresponding steps.

[0091] The bus 730 may use any of a variety of bus architectures, including, but not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, and a Peripheral Component Interconnect (PCI) bus.

[0092] Computer system 700 may also include input / output interfaces 740, a network interface 750, a storage interface 760, and the like. These interfaces 740, 750, and 760, as well as memory 710 and processor 720, may be connected via bus 730. Input / output interfaces 740 provide connection interfaces for input / output devices such as a display, mouse, and keyboard. Network interface 750 provides a connection interface for various networked devices. Storage interface 760 provides a connection interface for external storage devices such as floppy disks, USB flash drives, and SD cards.

[0093] Here, various aspects of the present disclosure are described with reference to flowcharts and / or block diagrams of methods, devices, and computer program products according to embodiments of the present disclosure. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks, can be implemented by computer-readable program instructions.

[0094] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable device to produce a machine, so that the processor executes the instructions to produce means for implementing the functions specified in one or more blocks in the flowcharts and / or block diagrams.

[0095] These computer-readable program instructions may also be stored in a computer-readable memory, which cause the computer to operate in a specific manner to produce an article of manufacture, including instructions for implementing the functions specified in one or more blocks in the flowcharts and / or block diagrams.

[0096] The present disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

[0097] This paper analyzes business opportunity information from multiple perspectives, ultimately achieving accurate allocation of business opportunities and providing strong support for enterprises to seize market opportunities.

[0098] The financial fraud detection model training method, apparatus, and medium according to the present disclosure have been described in detail. To avoid obscuring the underlying principles of the present disclosure, some details known in the art have been omitted. Based on the above description, those skilled in the art will fully understand how to implement the technical solutions disclosed herein.

[0099] Although some specific embodiments of the present disclosure have been described in detail through examples, those skilled in the art will understand that the above examples are for illustration only and are not intended to limit the scope of the present disclosure. Those skilled in the art will understand that the above embodiments may be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A financial fraud detection model training method, characterized in that: The method comprises: Obtain financial fraud datasets; Preprocessing the financial fraud dataset to obtain standardized financial data; Training a TabTransformer model using the standardized financial data to obtain a first prediction result; Using the standardized financial data to train the XGBoot model to obtain a second prediction result; Using the standardized financial data to train the LightGBM model, obtaining a third prediction result; The first prediction result, the second prediction result, and the third prediction result are weightedly voted using a weighted integration method to obtain a final prediction result.

2. The financial fraud detection model training method according to claim 1, characterized in that: The preprocessing of the financial fraud dataset to obtain standardized financial data includes: Deleting erroneous records from the financial fraud dataset to obtain initial data; Mean filling is performed on the initial data to obtain standardized financial data.

3. The financial fraud detection model training method according to claim 1 or 2, characterized in that: The standardized financial data includes category characteristic data and numerical characteristic data. The category characteristic data includes transaction type data, source account data, and destination account data. The numerical characteristic data includes transaction time data, transaction amount data, source account amount data before the transaction, and source account amount data after the transaction.

4. The financial fraud detection model training method according to claim 3, characterized in that: The step of training the TabTransformer model using the standardized financial data to obtain a first prediction result includes: Using LabelEncoder to convert the category feature data into integer to obtain integer data; Performing embedding processing on the integer data to obtain an embedding vector; Processing the embedding vector to learn a contextual relationship between each feature in the embedding vector to obtain first data to be relied upon; Performing a linear scaling operation on the numerical feature data to obtain second data to be spliced; Performing feature vector splicing and pooling processing on the first data to be spliced ​​and the second data to be spliced ​​to obtain financial data to be input; Inputting the financial data to be input into the Inception-ResNet structure to perform residual processing to obtain financial data after residual processing; The financial data after the residual processing is input into a multi-layer perceptron to obtain a first prediction result.

5. A financial fraud detection model training device, characterized in that: include: A financial fraud dataset acquisition module is used to acquire financial fraud datasets; A financial fraud data preprocessing module, configured to preprocess the financial fraud data set to obtain standardized financial data; a TabTransformer model training module, configured to train a TabTransformer model using the standardized financial data to obtain a first prediction result; An XGBoot model training module, configured to train the XGBoot model using the standardized financial data to obtain a second prediction result; A LightGBM model training module, configured to train the LightGBM model using the standardized financial data to obtain a third prediction result; The prediction result weighted voting module is used to perform weighted voting on the first prediction result, the second prediction result, and the third prediction result through a weighted integration method to obtain a final prediction result.

6. A financial fraud detection model training device, characterized in that: include: Memory; as well as A processor coupled to the memory, wherein the processor is configured to execute the financial fraud detection model training method according to any one of claims 1 to 4 based on instructions stored in the memory.

7. A computer-readable storage medium, characterized in that Computer program instructions are stored thereon, and when the instructions are executed by a processor, the financial fraud detection model training method according to any one of claims 1 to 4 is implemented.

Images