Method and system for constructing financial fraud prediction model

Abstract

The invention discloses a method and system for constructing a financial counterfeiting prediction model, and belongs to the technical field of feature interaction relation mining, and the method comprises the steps: obtaining financial data and a counterfeiting announcement as an initial data set; missing values and abnormal values in the initial data set are preprocessed; counterfeited event screening: quantifying the association degree between the financial counterfeited events and the financial indexes, and screening out counterfeited events with high association degree; mining financial indexes with interaction by using the large model; and optimizing the feature expression capability of the financial counterfeiting prediction model based on the financial indexes with the interaction points. The prediction model constructed by the invention can effectively improve the accuracy and robustness of financial counterfeiting identification.

Description

Technical Field

[0001] This invention relates to the field of feature interaction relationship mining technology, and more specifically to a method and system for constructing a financial fraud prediction model based on large model reasoning and ensemble tree model. Background Technology

[0002] In the financial field, with the diversification of business operations and the increasing complexity of data recording, the complexity and diversity of financial data are growing exponentially. Traditional methods for predicting financial fraud, such as manual auditing and simple statistical analysis, rely excessively on the experience and subjective judgment of auditors. This not only makes them inefficient but also makes it difficult to accurately identify hidden fraud patterns when faced with massive amounts of complex financial data, easily leading to misjudgments and omissions.

[0003] In recent years, the rise of big data and machine learning technologies has brought new opportunities to financial data processing, and data-driven methods for predicting financial fraud have become a research focus. However, existing machine learning methods still have many limitations when processing financial data, due to the specialized and highly structured nature of the data, as well as the concealment and variability of financial fraud patterns, making it difficult to effectively uncover the potential patterns behind complex financial data.

[0004] Meanwhile, large-scale models, exemplified by the Transformer architecture, have achieved remarkable results in the field of natural language processing, offering new directions for solving complex problems with their powerful semantic understanding and reasoning capabilities. However, directly applying large-scale models to the processing of structured financial data presents challenges such as the high degree of data specialization and structure, high computational costs, and the high risk of overfitting due to limited data scale. Ensemble tree models (such as XGBoost and LightGBM), widely used in financial forecasting tasks, while efficient and possessing certain feature learning capabilities in processing structured financial data, struggle to fully capture the complex underlying patterns in financial data when faced with complex financial fraud pattern recognition problems, relying solely on their feature interaction capabilities. This makes them unable to meet the increasingly stringent demands for financial fraud prediction.

[0005] Therefore, how to provide a financial fraud prediction model that can effectively uncover complex interactions in financial data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for constructing a financial fraud prediction model. It utilizes large-scale model inference technology to perform in-depth analysis of raw data, identifying key indicators highly correlated with the target event and their anomaly thresholds. Furthermore, by introducing interactive features into the XGBoost and LightGBM models, the feature representation capability of the model is optimized, significantly improving the model's accuracy in recognizing complex patterns and its predictive performance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] On the one hand, this invention provides a method for constructing a financial fraud prediction model, comprising:

[0009] Collect historical financial data and build an initial dataset;

[0010] The historical financial data in the initial dataset is preprocessed to obtain a preprocessed dataset;

[0011] The preprocessed dataset is randomly sampled and divided into a training set and a test set;

[0012] Based on financial fraud announcements, fraud events were identified, and a fraud event dataset was constructed.

[0013] Based on the large model, financial indicators related to fraudulent behavior and the corresponding abnormal thresholds of the financial indicators are extracted from the fraudulent events to construct a dataset of financial indicators and abnormal thresholds.

[0014] The company's financial indicators are serialized into a serialized dataset of company financial indicators.

[0015] The financial indicators and abnormal threshold datasets are used as the external knowledge base of the large model, and the serialized corporate financial indicator datasets are input into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive functions based on the financial indicators and the interaction constraints between them.

[0016] Using the preprocessed dataset, the interaction constraints between the financial indicators are embedded into the ensemble tree model. By introducing interaction features, the feature expression capability of the financial fraud prediction model is optimized.

[0017] Preferably, the preprocessing includes: missing value processing and normalization processing;

[0018] The missing value processing includes: calculating the proportion of missing values ​​for each financial data in the initial dataset; if the proportion of missing values ​​for the financial data is greater than a preset proportion, then removing the financial data; if the proportion of missing values ​​for the financial data is less than or equal to the preset proportion, then filling the missing values ​​of the financial data with the average value of the corresponding financial data.

[0019] The standardization process includes: using the Z-Score standardization method to scale the values ​​of each historical financial data in the initial dataset to a distribution with a mean of 0 and a standard deviation of 1. For the j-th feature X of the i-th sample in the initial dataset... ij Replaced with the standardized value X calculated using the Z-Score method.ij ',

[0020]

[0021] Where, μ j Let σ be the mean of the j-th historical financial data in the initial dataset. j Let be the standard deviation of the j-th historical financial data in the initial dataset.

[0022] Preferably, the preprocessed dataset is sampled and divided according to a preset positive sample to negative sample sampling ratio and a preset training set to test set partitioning ratio, wherein the preset positive sample to negative sample sampling ratio is 1:2 and the preset training set to test set partitioning ratio is 80:20.

[0023] Preferably, fraud events are obtained based on financial fraud announcement data, and a fraud event dataset is constructed, including:

[0024] Identifying fraudulent events based on data from financial fraud announcements;

[0025] The correlation between the financial fraud events and financial indicators is quantified based on cosine similarity.

[0026] Based on the average and maximum correlation between the financial fraud events and financial indicators, the financial fraud events are filtered to obtain a filtered dataset of fraud events.

[0027] Preferably, the correlation between the financial fraud event and the financial indicators is quantified based on cosine similarity, including:

[0028] The fraudulent events and financial indicators were encoded using a lightweight language model and converted into sentence embedding vectors.

[0029] The correlation between the financial fraud event and financial indicators is determined by calculating the cosine similarity between the sentence embedding vectors:

[0030]

[0031] Where, embedding vector Sentence embedding vectors representing financial fraud events. Sentence embedding vectors representing financial indicators. Represents the dot product of vectors. The norm of the sentence embedding vector representing a financial fraud event. The norm of the sentence embedding vector representing financial indicators.

[0032] Preferably, using the preprocessed dataset, the interaction constraints between the financial indicators are embedded into the ensemble tree model. By introducing interaction features, the feature representation capability of the financial fraud prediction model is optimized, including:

[0033] By utilizing the interaction constraints between the financial indicators, the interaction constraints are applied to model training in the XGBoost model by setting the interaction_constraints parameter; in the LightGBM model, the interaction constraints are applied to model training by setting the feature_interaction parameter or using a custom feature interaction constraint file. The XGBoost model and the LightGBM model are then trained using the preprocessed dataset to obtain a trained financial fraud prediction model.

[0034] On the other hand, the present invention provides a system for constructing a financial fraud prediction model, comprising:

[0035] The data acquisition module is used to collect historical financial data and build an initial dataset;

[0036] The preprocessing module is used to preprocess the historical financial data in the initial dataset to obtain a preprocessed dataset;

[0037] The data stratification module is used to randomly sample and divide the preprocessed dataset to obtain a training set and a test set.

[0038] The fraud incident extraction module is used to obtain fraud incidents based on financial fraud announcement data and construct a fraud incident dataset;

[0039] An anomaly indicator extraction module is used to extract financial indicators related to fraudulent behavior and the corresponding anomaly thresholds from the fraudulent event based on a large model, and to construct a dataset of financial indicators and anomaly thresholds.

[0040] The serialization module is used to serialize the text of company financial indicators to form a serialized dataset of company financial indicators.

[0041] The interaction module is used to use the financial indicators and abnormal threshold dataset as an external knowledge base of the large model, and input the serialized corporate financial indicator dataset into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive functions based on the financial indicators and the interaction constraints between the financial indicators.

[0042] The optimization module is used to embed the interaction constraints between the financial indicators into the ensemble tree model using the preprocessed dataset, thereby optimizing the feature expression capability of the financial fraud prediction model by introducing interaction features.

[0043] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for constructing a financial fraud prediction model. By integrating multiple data sources, a comprehensive feature set is constructed. Utilizing the reasoning ability of a large model, key indicators highly correlated with the target event and their anomaly thresholds are extracted, which not only enriches the model's input features but also enhances the model's ability to perceive complex patterns. Furthermore, this invention introduces large model reasoning technology to extract key indicators and mine interaction features, fully exploring the potential interaction patterns and rules between features, improving the model's adaptability to nonlinear and non-stationary scenarios, and possessing stronger generalization ability. Even further, this invention optimizes the model's feature expression ability by embedding the interaction-constrained features mined by large model reasoning technology into the XGBoost and LightGBM models, significantly improving the model's recognition accuracy and prediction performance for complex patterns. This effectively solves the problems of insufficient quantification of the correlation between financial data and financial fraud, insufficient feature depth mining, and underutilization of interactions in the prior art, thereby improving prediction accuracy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 A schematic diagram of the overall process of the modeling method provided by the present invention.

[0046] Figure 2 This is a flowchart of a feature interaction relationship mining method based on large model inference and ensemble tree model.

[0047] Figure 3 A schematic diagram of the modeling system provided by the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] This invention addresses the shortcomings of existing technologies, such as insufficient depth in financial data feature mining, inadequate quantification of correlations with fraudulent events, and underutilization of interactions. By deeply mining key features and interactions within financial data, it aims to improve the efficiency and accuracy of complex pattern recognition. This invention proposes a modeling method for mining financial data feature interactions based on large-model inference and ensemble tree models, which effectively improves the accuracy and stability of complex pattern recognition. This method has been effectively validated in the field of financial fraud prediction.

[0050] This invention discloses a method for constructing a financial fraud prediction model, such as... Figure 1-2 As shown, it includes:

[0051] Step 1: Collect historical financial data and build an initial dataset.

[0052] An initial dataset was constructed by collecting the company's historical financial data, including key financial statement information such as the balance sheet, income statement, and cash flow statement. This initial dataset encompasses multiple sample labels, where samples labeled as indicating financial fraud are defined as positive samples, while samples that have not involved financial fraud are defined as negative samples. Each sample contains multiple feature data related to financial fraud.

[0053] Step 2: Preprocess the historical financial data in the initial dataset to obtain the preprocessed dataset.

[0054] Step 3: Randomly sample and divide the preprocessed dataset to obtain the training set and the test set.

[0055] Step 4: Obtain fraud events based on financial fraud announcement data and construct a fraud event dataset.

[0056] Collect financial fraud announcements published by the China Securities Regulatory Commission (CSRC), extract the specific reasons for the fraud incidents, and construct a fraud incident dataset.

[0057] Step 5: Based on the large model, extract financial indicators related to fraudulent behavior and the corresponding abnormal thresholds of financial indicators from the fraud incident, and construct a dataset of financial indicators and abnormal thresholds.

[0058] Based on the background knowledge of the large model, a set of prompt words engineering is designed to extract the key financial indicators that lead to fraudulent behavior from fraudulent events that are highly correlated with financial characteristics, as well as the abnormal thresholds of these indicators in fraudulent events, forming a dataset of key financial indicators and abnormal thresholds.

[0059] Step 6: Serialize the company's financial indicators into a serialized dataset of company financial indicators.

[0060] Step 7: Use the financial indicators and anomaly threshold dataset as the external knowledge base of the large model, and input the serialized company financial indicator dataset into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive effects based on the interaction constraints between financial indicators.

[0061] Specifically, based on the background knowledge of the large model, another set of prompt words is designed to extract valuable and coherent financial indicators from the company's financial background, forming an interactive financial indicator dataset.

[0062] Step 8: Using the preprocessed dataset, embed the interaction constraints between financial indicators into the ensemble tree model. By introducing interaction features, optimize the feature expression capability of the financial fraud prediction model.

[0063] Furthermore, the characteristics of certain financial statement data in Step 1 include, but are not limited to, stock code, total assets, fixed assets, cash and cash equivalents, cash and cash equivalents ratio, accounts receivable, accounts receivable ratio, inventory, inventory ratio, total liabilities, accounts payable, accounts payable ratio, advances from customers, total equity, equity ratio, total asset ratio, total debt ratio, current ratio, debt-to-equity ratio, net cash flow from operating activities, net cash flow from operating activities ratio, cash received from sales of goods and services, cash received from sales of goods and services ratio, cash paid to and on behalf of employees, cash paid to and on behalf of employees ratio, net cash flow from investing activities, net cash flow from investing activities ratio, and acquisition of fixed assets. Cash paid for intangible assets and other long-term assets; cash paid for the purchase and construction of fixed assets as a percentage of intangible assets and other long-term assets; net cash flow from financing activities; net cash flow from financing activities as a percentage of total cash flow; increase in cash and cash equivalents; increase in cash and cash equivalents as a percentage of total cash flow; net profit attributable to owners of the parent company; total operating revenue; total operating costs; ratio of total operating revenue to total operating costs; operating costs; operating expenses; operating expenses as a percentage of total operating costs; selling expenses; administrative expenses; financial expenses; operating profit; total profit; income tax expense; taxes and surcharges; total profit as a percentage of total profit; operating profit as a percentage of total profit; net profit attributable to owners of the parent company as a percentage of total profit; net profit attributable to owners of the parent company after deducting non-recurring gains and losses; percentage of accounts receivable; percentage of net profit after deducting non-recurring gains and losses.

[0064] Furthermore, preprocessing includes: missing value handling and standardization;

[0065] Missing value handling includes: calculating the proportion of missing values ​​for each financial data in the initial dataset; if the proportion of missing values ​​for a financial data is greater than a preset proportion, then the financial data is removed; if the proportion of missing values ​​for a financial data is less than or equal to the preset proportion, then the missing values ​​of the financial data are filled with the average value of the corresponding financial data.

[0066] The standardization process includes: using the Z-Score standardization method to scale the values ​​of each historical financial data point in the initial dataset to a distribution with a mean of 0 and a standard deviation of 1. For the j-th feature X of the i-th sample in the initial dataset... ij Replaced with the standardized value X calculated using the Z-Score method. ij ',

[0067]

[0068] Where, μ j Let σ be the mean of the j-th historical financial data in the initial dataset. j Let be the standard deviation of the j-th historical financial data in the initial dataset.

[0069] Preferably, the preprocessed dataset is sampled and divided according to a preset positive sample to negative sample sampling ratio and a preset training set to test set splitting ratio. The preset positive sample to negative sample sampling ratio is 1:2, and the preset training set to test set splitting ratio is 80:20.

[0070] Furthermore, based on financial fraud announcement data, fraud events are identified, and a fraud event dataset is constructed, including:

[0071] The specific implementation of the screening for fraudulent events is as follows:

[0072] First, collect the financial fraud announcements published by the China Securities Regulatory Commission (CSRC) and extract the specific reasons for the fraud incidents. fraud ={d1,d2,…,d N}, where N is the number of fraudulent events, d i This is the descriptive text for the i-th fraudulent event. Simultaneously, the set of financial indicator names F = {f1, f2, ..., f...} is extracted from the financial indicator feature mapping file. M}, where M is the number of financial indicators, f j Let j be the name of the j-th financial indicator.

[0073] Based on cosine similarity, the correlation between financial fraud events and financial indicators is quantified.

[0074] Based on the average and maximum correlation between financial fraud events and financial indicators, financial fraud events are filtered to obtain a filtered dataset of fraud events.

[0075] Preferably, the correlation between financial fraud events and financial indicators is quantified based on cosine similarity, including:

[0076] A lightweight language model is used to encode fraudulent events and financial indicators separately, converting them into sentence embedding vectors. Specifically, the description of fraudulent events and the name of financial indicators are encoded separately using a lightweight language model and converted into sentence embedding vectors. The implementation involves encoding each fraudulent event description d... i and the name of each financial indicator f j The model outputs the corresponding embedding vectors respectively. and in D is the dimension of the embedding vector.

[0077] The correlation between financial fraud events and financial indicators is determined by calculating the cosine similarity between sentence embedding vectors:

[0078]

[0079] Where, embedding vector Sentence embedding vectors representing financial fraud events. Sentence embedding vectors representing financial indicators. Represents the dot product of vectors. The norm of the sentence embedding vector representing a financial fraud event. The norm of the sentence embedding vector representing financial indicators.

[0080] For each fraudulent event description di, calculate its cosine similarity matrix S with all financial indicators. i ∈R 1×M S i (j)=sim(d i ,f j ).

[0081] Based on the cosine similarity matrix S i Calculate the description d for each fraudulent event. i Average cosine similarity and maximum cosine similarity

[0082]

[0083] According to the preset threshold θ avg and θ max Screening out fraudulent events that are highly correlated with financial indicators. and

[0084] Set θ avg =0.4 and θ max=0.6. If a certain fraudulent event description d i If the above conditions are met, the fraudulent event is considered to have a high correlation with the financial indicators and is retained; otherwise, the fraudulent event is excluded. Ultimately, the filtered fraudulent event dataset D is formed. filtered .

[0085] Furthermore, in step 5, the specific implementation of extracting key financial indicators and anomaly thresholds is as follows:

[0086] Based on the background knowledge of the large model, a prompting word project was designed: "Please read the provided textfile carefully, which contains the reasons why the China Securities Regulatory Commission (CSRC) determines that the company has committed financial fraud. This document outlines various financial and operational violations that have been identified as signs of the company's fraudulent activities. Your task is to analyze this article and provide fraud markers and thresholds for the financial indicators mentioned. This method should help us understand which financial indicators are the most important in detecting fraudulent activities and may help enhance the fraud detection model." This guides the large model to extract key financial indicators and their abnormal thresholds that lead to fraudulent behavior from fraudulent events that are highly correlated with financial characteristics.

[0087] The filtered dataset of fraudulent events descriptions D filtered By combining the above-mentioned prompts with the large model, text output containing key financial indicators and their abnormal thresholds is generated.

[0088] In step 6, the specific implementation of knowledge base construction is as follows:

[0089] The key financial indicators and their anomaly thresholds extracted in step 5 are integrated and constructed into an external knowledge base for the large model. This knowledge base stores the key financial indicators and their corresponding anomaly thresholds in a structured form, providing the large model with important reference information for detecting financial fraud. For example, the extracted key financial indicators and their anomaly thresholds are as follows:

[0090] 1.TotalAssets Ratio(TOTAL_ASSETS_RATIO)

[0091] Fraud Sign:Dramatic fluctuations in total assets may indicate assetmanipulation or misrepresentation.

[0092] Threshold: Less than 20% or more than 80%.

[0093] Using the above method, key financial indicators and their abnormal thresholds are extracted from the description of fraudulent events, forming a dataset D of key financial indicators and abnormal thresholds. key_indicators .

[0094] The specific implementation of financial indicator serialization is as follows:

[0095] The financial indicator data from the training set, segmented in step 3, is used as the object for serialization. For each financial indicator in the company's financial statements, it is converted one by one into a textual description in natural language, according to predefined format specifications. Taking a single line of data as an example, the serialized text would be: "The Debt-to-Asset Ratio is 75%. The Accounts Receivable is $120,000. The Net Cash Flow from Operations Ratio is 3%." Following this conversion logic, all financial indicators in the entire dataset are comprehensively processed, ultimately generating a serialized text dataset.

[0096] Furthermore, the serialized corporate financial indicator dataset from step 6 and the external knowledge base are input into the large model. The external knowledge base contains key financial indicators and their anomaly threshold information, providing the large model with professional knowledge background on financial fraud detection. The serialized corporate financial indicator dataset presents the company's financial situation in the form of natural language descriptions. A special cue word engineering is designed to guide the large model in reasoning about financial fraud.

[0097] Furthermore, by utilizing the preprocessed dataset, the interaction constraints between financial indicators are embedded into the ensemble tree model. By introducing interaction features, the feature representation capability of the financial fraud prediction model is optimized, including:

[0098] By leveraging the interaction constraints between financial indicators, the interaction_constraints parameter in the XGBoost model is used to apply these constraints to model training. Similarly, in the LightGBM model, the feature_interaction parameter or a custom feature interaction constraint file is used to apply these constraints to model training. Both the XGBoost and LightGBM models are then trained using a preprocessed dataset to obtain a well-trained financial fraud prediction model.

[0099] To verify the effectiveness of the method of this invention, we conducted an empirical experiment using financial fraud prediction as an application scenario. In the experiment, we evaluated the method under three scenarios: without considering interaction features, with random interaction features, and with considering mined interaction features. The method of this invention utilizes interaction features mined by four major models—GPT, DeepSeek, Gemini, and Ernie—and embeds them into the prediction model for comprehensive comparative analysis.

[0100] The experimental results are listed in Table 1:

[0101] Table 1: Performance comparison of the proposed method on different large models with and without considering interaction features, random interaction features, and considering interaction features (training set:test set = 80:20)

[0102]

[0103] As can be seen from the classification results in Table 1, the method proposed in this invention shows a significant advantage over randomly selected feature interactions and methods that do not consider feature interactions.

[0104] In summary, this invention utilizes large-scale model inference technology for in-depth data analysis, accurately extracting key indicators and their anomaly thresholds, and further mining interactive features. By constructing an external knowledge base and a serialized dataset, this invention embeds these key constraint information into the XGBoost and LightGBM models, optimizing the models' feature representation capabilities. This invention effectively solves the problems of insufficient feature mining depth, insufficient correlation quantification, and underutilization of interactions in existing technologies, significantly improving the model's accuracy and predictive performance in identifying complex patterns. The validation results in the field of financial fraud prediction further demonstrate its effectiveness, showcasing its powerful capabilities in complex data processing and pattern recognition.

[0105] On the other hand, this invention provides a system for constructing a financial fraud prediction model, such as... Figure 3 As shown, it includes:

[0106] The data acquisition module is used to collect historical financial data and build an initial dataset;

[0107] The preprocessing module is used to preprocess the historical financial data in the initial dataset to obtain a preprocessed dataset;

[0108] The data stratification module is used to randomly sample and divide the preprocessed dataset to obtain training and test sets.

[0109] The fraud incident extraction module is used to obtain fraud incidents based on financial fraud announcement data and construct a fraud incident dataset;

[0110] The abnormal indicator extraction module is used to extract financial indicators related to fraudulent behavior and corresponding abnormal thresholds from fraudulent events based on a large model, and to construct a dataset of financial indicators and abnormal thresholds.

[0111] The serialization module is used to serialize the text of company financial indicators to form a serialized dataset of company financial indicators.

[0112] The interaction module is used to use the financial indicators and abnormal threshold dataset as an external knowledge base of the large model, and input the serialized corporate financial indicator dataset into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive functions based on the interaction constraints between financial indicators.

[0113] The optimization module is used to embed the interaction constraints between financial indicators into the ensemble tree model using the preprocessed dataset. By introducing interaction features, it optimizes the feature expression capability of the financial fraud prediction model.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0115] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a financial fraud prediction model, characterized in that, include: Collect historical financial data and build an initial dataset; The historical financial data in the initial dataset is preprocessed to obtain a preprocessed dataset; The preprocessed dataset is randomly sampled and divided into a training set and a test set; Based on financial fraud announcements, fraud events were identified, and a fraud event dataset was constructed. Based on the large model, financial indicators related to fraudulent behavior and the corresponding abnormal thresholds of the financial indicators are extracted from the fraudulent events to construct a dataset of financial indicators and abnormal thresholds. The company's financial indicators are serialized into a serialized dataset of company financial indicators. The financial indicators and abnormal threshold datasets are used as the external knowledge base of the large model, and the serialized corporate financial indicator datasets are input into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive functions based on the financial indicators and the interaction constraints between them. Using the preprocessed dataset, the interaction constraints between the financial indicators are embedded into the ensemble tree model. By introducing interaction features, the feature expression capability of the financial fraud prediction model is optimized.

2. The method for constructing a financial fraud prediction model according to claim 1, characterized in that, The preprocessing includes: missing value handling and normalization; The missing value processing includes: calculating the proportion of missing values ​​for each financial data in the initial dataset; if the proportion of missing values ​​for the financial data is greater than a preset proportion, then removing the financial data; if the proportion of missing values ​​for the financial data is less than or equal to the preset proportion, then filling the missing values ​​of the financial data with the average value of the corresponding financial data. The standardization process includes: using the Z-Score standardization method to scale the values ​​of each historical financial data in the initial dataset to a distribution with a mean of 0 and a standard deviation of 1. For the j-th feature X of the i-th sample in the initial dataset... ij Replaced with the standardized value X calculated using the Z-Score method. ij ', Where, μ j Let σ be the mean of the j-th historical financial data in the initial dataset. j Let be the standard deviation of the j-th historical financial data in the initial dataset.

3. The method for constructing a financial fraud prediction model according to claim 1, characterized in that, The preprocessed dataset is sampled and divided according to a preset positive sample to negative sample sampling ratio and a preset training set to test set splitting ratio. The preset positive sample to negative sample sampling ratio is 1:2, and the preset training set to test set splitting ratio is 80:

20.

4. The method for constructing a financial fraud prediction model according to claim 1, characterized in that, Based on financial fraud announcements, fraud events were identified, and a fraud event dataset was constructed, including: Identifying fraudulent events based on data from financial fraud announcements; The correlation between the financial fraud events and financial indicators is quantified based on cosine similarity. Based on the average and maximum correlation between the financial fraud events and financial indicators, the financial fraud events are filtered to obtain a filtered dataset of fraud events.

5. The method for constructing a financial fraud prediction model according to claim 1, characterized in that, The correlation between the financial fraud events and financial indicators is quantified based on cosine similarity, including: The fraudulent events and financial indicators were encoded using a lightweight language model and converted into sentence embedding vectors. The correlation between the financial fraud event and financial indicators is determined by calculating the cosine similarity between the sentence embedding vectors: Where, embedding vector Sentence embedding vectors representing financial fraud events. Sentence embedding vectors representing financial indicators. Represents the dot product of vectors. The norm of the sentence embedding vector representing a financial fraud event. The norm of the sentence embedding vector representing financial indicators.

6. The method for constructing a financial fraud prediction model according to claim X, characterized in that, Using the preprocessed dataset, the interaction constraints between the financial indicators are embedded into the ensemble tree model. By introducing interaction features, the feature representation capability of the financial fraud prediction model is optimized, including: By utilizing the interaction constraints between the financial indicators, the interaction constraints are applied to model training in the XGBoost model by setting the interaction_constraints parameter; in the LightGBM model, the interaction constraints are applied to model training by setting the feature_interaction parameter or using a custom feature interaction constraint file. The XGBoost model and the LightGBM model are then trained using the preprocessed dataset to obtain a trained financial fraud prediction model.

7. A system for constructing a financial fraud prediction model, characterized in that, include: The data acquisition module is used to collect historical financial data and build an initial dataset; The preprocessing module is used to preprocess the historical financial data in the initial dataset to obtain a preprocessed dataset; The data stratification module is used to randomly sample and divide the preprocessed dataset to obtain a training set and a test set. The fraud incident extraction module is used to obtain fraud incidents based on financial fraud announcement data and construct a fraud incident dataset; An anomaly indicator extraction module is used to extract financial indicators related to fraudulent behavior and the corresponding anomaly thresholds from the fraudulent event based on a large model, and to construct a dataset of financial indicators and anomaly thresholds. The serialization module is used to serialize the text of company financial indicators to form a serialized dataset of company financial indicators. The interaction module is used to use the financial indicators and abnormal threshold dataset as an external knowledge base of the large model, and input the serialized corporate financial indicator dataset into the large model to reason and describe the company's current financial situation, explore the interaction relationship between financial indicators, and construct a financial indicator dataset with interactive functions based on the financial indicators and the interaction constraints between the financial indicators. The optimization module is used to embed the interaction constraints between the financial indicators into the ensemble tree model using the preprocessed dataset, thereby optimizing the feature expression capability of the financial fraud prediction model by introducing interaction features.

Images