A machine learning-based financial anomaly identification method

Abstract

The application relates to the technical field of financial risk control, in particular to a financial abnormality identification method based on machine learning, which comprises the following steps: collecting multi-source financial transaction data and transaction association constraint relations respectively, sequentially performing cleaning, missing mark and feature extraction on the collected multi-source financial transaction data to obtain financial features; inputting the financial features, the transaction constraint feature vectors and the abnormality propagation features into a machine learning model for training to obtain a transaction monitoring model; analyzing real-time financial transactions based on the transaction monitoring model to obtain an abnormality identification result; obtaining the recall rate of the abnormality identification result to determine whether the accuracy of the financial transaction abnormality behavior identification meets the requirements; determining whether the transaction abnormality propagation decay gradient needs to be reduced according to the false alarm rate of the abnormal transaction account; and determining the balance net compression coefficient of the transaction constraint execution based on the false transaction identification rate. The application improves the accuracy of the financial transaction abnormality behavior identification.

Description

Technical Field

[0001] This invention relates to the field of financial risk control technology, and in particular to a method for identifying financial anomalies based on machine learning. Background Technology

[0002] With the deepening of enterprise financial digital transformation, financial transaction data is showing a trend of becoming massive, complex, and concealed. Traditional anomaly identification methods based on rule engines or statistical models often rely on expert experience to set fixed thresholds and simple association rules, making it difficult to capture complex abnormal transaction patterns across accounts, periods, and levels. Although some methods have introduced machine learning technology, they often only analyze the characteristics of a single transaction in isolation, lacking the ability to model the overall transaction relationship graph, offsetting behavior, and deep fund flow paths. When faced with concealed anomalies deliberately constructed through methods such as nested fictitious transactions and offsetting positive and negative amounts, the recall rate is low and the false negative rate is high, making it difficult to meet the business needs for accurate identification of financial anomalies in high-risk scenarios. Therefore, there is an urgent need for a machine learning-based financial anomaly identification method that can integrate multi-source financial data, constraint knowledge bases, and anomaly propagation characteristics, which is of great significance for improving the accuracy and sensitivity of financial anomaly identification.

[0003] Chinese Patent Publication No. CN120235448A discloses a method and system for large-scale dynamic data anomaly identification and early warning in the regulatory field. The method includes: S1, integrating and preprocessing data through a large-scale data integration platform; S2, extracting features from the preprocessed data model; S3, identifying abnormal data through an anomaly detection model; S4, setting early warning rules and thresholds in the anomaly detection model to identify and issue early warnings for abnormal data in real time; and S5, handling the abnormal early warning data according to a response strategy. It is evident that the method and system for large-scale dynamic data anomaly identification and early warning in the regulatory field suffer from insufficient accuracy in identifying abnormal financial transaction behavior because the anomaly detection model relies on static early warning rules and thresholds and lacks adaptive updating capabilities for concealed abnormal transaction behavior. Summary of the Invention

[0004] To address this, the present invention provides a machine learning-based method for identifying financial anomalies, thereby overcoming the problem in existing technologies where the anomaly detection models rely on static warning rules and thresholds and lack adaptive updating capabilities for concealed abnormal transaction behaviors, resulting in insufficient accuracy in identifying financial transaction anomalies.

[0005] To achieve the above objectives, the present invention provides a machine learning-based method for identifying financial anomalies, comprising: Multi-source financial transaction data and transaction-related constraint relationships are collected separately. The collected multi-source financial transaction data are then cleaned, missing data is marked, time-series aligned, fused, and feature extracted to obtain financial features. A constraint knowledge base is then constructed based on the transaction-related constraint relationships. Based on the constraint knowledge base, apply the applicable constraints to each financial transaction to obtain the transaction constraint feature vector, obtain the financial transaction relationship to construct the abnormal relationship graph, and generate abnormal propagation features based on the abnormal relationship graph. The financial features, the transaction constraint feature vector, and the anomaly propagation features are input into a machine learning model for training to obtain a transaction monitoring model. Based on the transaction monitoring model, real-time financial transactions are analyzed to obtain anomaly identification results. Obtain the recall rate of the anomaly identification results to determine whether the accuracy of the identification of abnormal financial transaction behavior meets the requirements; If the accuracy of the identification of abnormal financial transaction behavior does not meet the requirements, the underreporting rate of abnormal transaction accounts is obtained to determine whether the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. If the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts does not meet the requirements, then it is determined whether it is necessary to reduce the transaction anomaly propagation attenuation gradient. If it is not necessary to reduce the decay gradient of abnormal transaction propagation, then the net amount compression factor for transaction constraint execution is determined based on the identification rate of fraudulent transactions.

[0006] Furthermore, the recall rate based on the anomaly identification results determines whether the accuracy of identifying abnormal financial transaction behaviors meets the requirements, including: The recall rate of the anomaly identification results is compared with the preset recall rate; If the recall rate of the anomaly identification results is greater than the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to meet the requirements. If the recall rate of the anomaly identification result is less than or equal to the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to be unsatisfactory.

[0007] Furthermore, given that the accuracy of identifying abnormal financial transaction behavior does not meet the requirements, the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts is determined based on the underreporting rate of abnormal transaction accounts.

[0008] Furthermore, based on the underreporting rate of abnormal transaction accounts, determine whether the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements, including: Compare the underreporting rate of abnormal transaction accounts with the preset first underreporting rate; If the underreporting rate of the abnormal transaction account is less than or equal to the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. If the underreporting rate of the abnormal transaction accounts is greater than the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts does not meet the requirements.

[0009] Further, determine whether it is necessary to reduce the decay gradient of transaction anomaly propagation, including: The underreporting rate of the abnormal transaction accounts is compared with the preset first underreporting rate and the preset second underreporting rate, respectively; If the underreporting rate of the abnormal transaction account is greater than the preset first underreporting rate and less than or equal to the preset second underreporting rate, it is determined to reduce the attenuation gradient of abnormal transaction propagation. If the underreporting rate of the abnormal transaction account is greater than the preset second underreporting rate, it is determined that there is no need to reduce the transaction anomaly propagation attenuation gradient.

[0010] Furthermore, the reduction in the transaction anomaly propagation attenuation gradient is determined by the difference between the underreporting rate of the abnormal transaction account and the preset first underreporting rate.

[0011] Furthermore, based on the condition that the underreporting rate of the abnormal transaction accounts is greater than the preset second underreporting rate, it is initially determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior is determined based on the fraudulent transaction identification rate.

[0012] Furthermore, the comprehensiveness of the financial transaction behavior representation by the transaction constraint feature vector is determined based on the fraudulent transaction detection rate, including: Compare the fraudulent transaction detection rate with the preset detection rate; If the fraudulent transaction identification rate is greater than the preset identification rate, then the transaction constraint feature vector is determined to meet the requirements for comprehensive representation of financial transaction behavior. If the fraudulent transaction identification rate is less than or equal to the preset identification rate, it is determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the net amount compression coefficient for the execution of transaction constraints is increased.

[0013] Furthermore, the fraudulent transaction identification rate is the ratio of the number of fraudulent transactions identified in the anomaly identification results to the total number of actual fraudulent transactions.

[0014] Furthermore, the increase in the net amount compression coefficient for the execution of the transaction constraint is determined by the difference between the preset identification rate and the fraudulent transaction identification rate.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: The method of this invention determines whether the accuracy of identifying abnormal financial transaction behavior meets the requirements by using the recall rate of the anomaly identification results. Since financial transaction data typically exhibits multi-account association and cross-period flow characteristics, abnormal behavior may be scattered across multiple weakly associated transaction nodes. When the model fails to identify some latent abnormal samples, the recall rate decreases. By determining the accuracy of identifying abnormal financial transaction behavior, the overall ability of the model to capture real abnormal transaction samples can be quantified. The attenuation gradient of transaction anomaly propagation can be adjusted by using the underreporting rate of abnormal transaction accounts. Because abnormal transactions are usually concealed, the weak manifestation of abnormal features at local transaction nodes leads to the truncation of deep anomaly propagation chains, reducing the model's sensitivity to identifying potential abnormal behavior. By reducing the attenuation gradient of transaction anomaly propagation, the attenuation of abnormal features can be reduced. The attenuation rate during cross-account and multi-level transaction propagation allows abnormal behavior features in weakly correlated nodes to maintain high correlation strength in deep propagation links, thereby enhancing the model's continuous perception of long-link hidden abnormal transactions. The net amount compression coefficient of the offsetting amount in transaction constraint execution is adjusted by the false transaction identification rate. Since offsetting behavior reconstructs the original transaction sequence through positive and negative amount offsetting or netting, the original transaction distribution in the feature space undergoes nonlinear compression and structural distortion, which weakens or masks the statistical features of abnormal transaction patterns. By increasing the net amount compression coefficient of the offsetting amount in transaction constraint execution, the smoothed cash flow differences during the offsetting process can be reverse-enhanced and mapped, making the originally compressed abnormal fluctuations explicit again in the feature space, repairing the feature distribution offset caused by offsetting, and improving the accuracy of identifying abnormal financial transaction behavior.

[0016] Furthermore, this invention determines whether the accuracy of identifying abnormal financial transaction behavior meets the requirements by setting a preset recall rate. Since financial transaction data usually exhibits characteristics of multi-account association and cross-period flow, abnormal behavior may be scattered across multiple weakly associated transaction nodes. When the model fails to identify some latent abnormal samples, the recall rate decreases. By determining the accuracy of identifying abnormal financial transaction behavior, the overall ability of the model to capture real abnormal transaction samples can be quantified, thereby further improving the accuracy of identifying abnormal financial transaction behavior.

[0017] Furthermore, this invention adjusts the decay gradient of abnormal transaction propagation by setting a preset first false negative rate and a preset second false negative rate. Since abnormal transactions are usually concealed, the abnormal features are weakly manifested at local transaction nodes, causing the deep abnormal propagation chain to be truncated, reducing the model's sensitivity to identify potential abnormal behavior. By reducing the decay gradient of abnormal transaction propagation, the decay rate of abnormal features in the cross-account and multi-level transaction propagation process can be reduced, so that the abnormal behavior features in weakly correlated nodes can still maintain a high correlation strength in the deep propagation chain, thereby enhancing the model's continuous perception ability of long-chain concealed abnormal transactions and further improving the accuracy of identifying abnormal financial transaction behavior.

[0018] Furthermore, this invention adjusts the net amount compression coefficient of the offsetting transaction constraint execution by setting a preset recognition rate. Since the offsetting behavior reconstructs the original transaction sequence through positive and negative amount offsetting or netting, the original transaction distribution in the feature space undergoes nonlinear compression and structural distortion, which weakens or masks the statistical characteristics of abnormal transaction patterns. By increasing the net amount compression coefficient of the offsetting transaction constraint execution, the smoothed cash flow differences during the offsetting process can be reverse-enhanced and mapped, making the originally compressed abnormal fluctuations explicit again in the feature space, repairing the feature distribution offset caused by the offsetting, and further improving the accuracy of identifying abnormal financial transaction behaviors. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the financial anomaly identification method based on machine learning according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process by which the machine learning-based financial anomaly identification method of this invention determines whether the accuracy of the identification of abnormal financial transaction behaviors meets the requirements. Figure 3 This is a flowchart illustrating the machine learning-based financial anomaly identification method of the present invention for determining whether it is necessary to reduce the attenuation gradient of transaction anomaly propagation; Figure 4 This is a flowchart illustrating the process of determining the net amount compression coefficient for transaction constraint execution using a machine learning-based financial anomaly identification method according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] Please see Figure 1 The diagram shown is an overall flowchart of the financial anomaly identification method based on machine learning according to an embodiment of the present invention.

[0023] An embodiment of the present invention provides a machine learning-based method for identifying financial anomalies, comprising: Step S1: Collect multi-source financial transaction data and transaction association constraint relationships respectively. Clean the collected multi-source financial transaction data, mark missing data, align time sequence, fuse data and extract features to obtain financial features. Construct a constraint knowledge base based on the transaction association constraint relationships. Step S2: Apply applicable constraints to each financial transaction based on the constraint knowledge base to obtain transaction constraint feature vectors, obtain financial transaction relationships to construct an anomaly relationship graph, and generate anomaly propagation features based on the anomaly relationship graph; Step S3: Input the financial features, the transaction constraint feature vector, and the anomaly propagation features into the machine learning model for training to obtain the transaction monitoring model; and analyze real-time financial transactions based on the transaction monitoring model to obtain anomaly identification results. Step S4: Obtain the recall rate of the anomaly identification results to determine whether the accuracy of the identification of abnormal financial transaction behavior meets the requirements; Step S5: If the accuracy of the identification of abnormal financial transaction behavior does not meet the requirements, the underreporting rate of abnormal transaction accounts is obtained to determine whether the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. Step S6: If the sensitivity of the transaction monitoring model to identify potential real abnormal transaction accounts does not meet the requirements, then determine whether it is necessary to reduce the transaction abnormality propagation attenuation gradient based on the false negative rate of abnormal transaction accounts. Step S7: If it is not necessary to reduce the decay gradient of abnormal transaction propagation, then determine the net amount compression coefficient for transaction constraint execution based on the false transaction identification rate.

[0024] Specifically, multi-source financial transaction data includes bank statements, corporate transfer records, and bill settlement statements.

[0025] Specifically, transaction-related constraints include constraints on the transaction entities, constraints on monetary amounts, and constraints on temporal relationships.

[0026] Specifically, financial characteristics include the dispersion of transaction amounts, the ratio of fund inflows to outflows, and the proportion of nighttime transactions.

[0027] Specifically, the constraint knowledge base is a set of executable logical expressions and threshold conditions that transform transaction-related constraints into transaction-related constraints.

[0028] Specifically, the transaction monitoring model can be a deep neural network, a graph neural network, or an ensemble learning model, with a preferred embodiment being a deep neural network.

[0029] Specifically, the process of inputting financial features, transaction constraint feature vectors, and anomaly propagation features into a machine learning model to train a transaction monitoring model involves performing feature-level fusion processing on structured financial features, transaction constraint feature vectors, and anomaly propagation features to form a unified multidimensional feature representation vector. After standardizing the multidimensional feature representation vector, it is input into the machine learning model. Through supervised learning training on historically labeled normal transaction samples and abnormal transaction samples, the model learns the joint mapping relationship between financial feature patterns, transaction constraints, and anomaly propagation structures, and outputs a transaction monitoring model for financial transaction analysis.

[0030] Specifically, the process of generating anomaly propagation features based on anomaly association graphs involves using nodes in the anomaly association graphs to represent financial transaction entities, edges to represent fund flow relationships or dependency relationships between transactions, performing graph propagation calculations based on the anomaly association graphs, multi-hop diffusion propagation of the initial anomaly scores of nodes, generating anomaly intensity representations for each node at different propagation orders, aggregating the anomaly intensity at each propagation order, and obtaining anomaly propagation features that characterize the propagation intensity of anomalies in the transaction network.

[0031] Specifically, the process of analyzing real-time financial transactions based on a transaction monitoring model to obtain anomaly identification results involves obtaining the corresponding financial features, transaction constraint feature vectors, and anomaly propagation features extracted from real-time financial transaction data, inputting the real-time transaction features into a trained transaction monitoring model for forward inference calculation, outputting anomaly scores for the corresponding transactions, and judging the anomaly scores or anomaly probability values ​​based on preset anomaly judgment thresholds. When the preset threshold is exceeded, the transaction is judged as an anomaly to obtain anomaly identification results.

[0032] Specifically, the anomaly identification results include anomaly identification results for monetary amounts, anomaly identification results for the relationships between transaction entities, and anomaly identification results for the timing of transactions.

[0033] Specifically, the transaction constraint feature vector is a vectorized representation of whether each transaction triggers a specific rule and the strength of the trigger, based on a constraint knowledge base.

[0034] Specifically, the characteristics of abnormal propagation include node centrality, path depth, and community division attributes.

[0035] Specifically, the transaction anomaly propagation attenuation gradient is used to characterize the attenuation rate of the abnormal signal as it propagates along the association path in the anomaly association graph with increasing propagation depth.

[0036] Specifically, the net offset compression factor for transaction constraint enforcement is an adjustment parameter that measures the degree of nonlinear amplification of the residual difference after amount offset when netting offset transaction pairs.

[0037] In implementation, the method of this invention determines whether the accuracy of identifying abnormal financial transaction behavior meets the requirements by using the recall rate of the anomaly identification results. Since financial transaction data typically exhibits multi-account association and cross-period flow characteristics, abnormal behavior may be scattered across multiple weakly associated transaction nodes. When the model fails to identify some latent abnormal samples, the recall rate decreases. By determining the accuracy of identifying abnormal financial transaction behavior, the overall ability of the model to capture real abnormal transaction samples can be quantified. The attenuation gradient of transaction anomaly propagation is adjusted by the underreporting rate of abnormal transaction accounts. Because abnormal transactions are often concealed, the weak manifestation of abnormal features at local transaction nodes leads to the truncation of deep anomaly propagation chains, reducing the model's sensitivity to identifying potential abnormal behavior. By reducing the attenuation gradient of transaction anomaly propagation, the attenuation of abnormal features across accounts and multiple levels can be reduced. The attenuation rate during transaction propagation allows abnormal behavior features in weakly correlated nodes to maintain high correlation strength in deep propagation links, thereby enhancing the model's continuous perception of hidden abnormal transactions in long links. The net amount compression coefficient of the offsetting amount in transaction constraint execution is adjusted by the false transaction identification rate. Since offsetting behavior reconstructs the original transaction sequence through positive and negative amount offsetting or netting, the original transaction distribution in the feature space undergoes nonlinear compression and structural distortion, which weakens or masks the statistical features of abnormal transaction patterns. By increasing the net amount compression coefficient of the offsetting amount in transaction constraint execution, the smoothed cash flow differences during the offsetting process can be reverse-enhanced and mapped, making the originally compressed abnormal fluctuations explicit again in the feature space, repairing the feature distribution offset caused by offsetting, and improving the accuracy of identifying abnormal financial transaction behavior.

[0038] Please continue reading. Figure 2 The diagram shown is a logical flowchart illustrating the process by which the machine learning-based financial anomaly identification method of this invention determines whether the accuracy of the identification of abnormal financial transaction behaviors meets the requirements.

[0039] Specifically, the recall rate based on the anomaly identification results determines whether the accuracy of identifying abnormal financial transaction behaviors meets the requirements, including: The recall rate of the anomaly identification results is compared with the preset recall rate; If the recall rate of the anomaly identification results is greater than the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to meet the requirements. If the recall rate of the anomaly identification result is less than or equal to the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to be unsatisfactory.

[0040] Understandably, in the process of identifying financial anomalies based on machine learning, the core logic of using a preset recall rate to characterize the accuracy of identifying abnormal financial transaction behavior is to transform the model's ability to identify abnormal financial transaction behavior into a quantifiable recall rate range. By comparing the actual recall rate with the preset threshold, it is determined whether the accuracy of anomaly behavior identification meets the requirements. The preset recall rate can be set according to actual working conditions. The setting of the preset recall rate aims to ensure the accuracy and practicality of identifying abnormal financial transaction behavior. Optionally, the preset recall rate is determined through a limited number of trials by evaluating the effect of different identification result recall rates on the identification of abnormal financial transaction behavior. The determined preset recall rate should meet the requirement of being neither too small nor causing excessive interference to the identification process of abnormal financial transaction behavior. For example, the preset recall rate is generally selected in the range of [94%, 96%].

[0041] Preferably, the preset recall rate is 95%.

[0042] Specifically, the recall rate of anomaly identification results is the ratio of the number of abnormal transactions correctly identified by the model to the total number of actual abnormal transactions.

[0043] In practice, this invention determines whether the accuracy of identifying abnormal financial transaction behavior meets the requirements by setting a preset recall rate. Since financial transaction data usually exhibits characteristics of multi-account association and cross-period flow, abnormal behavior may be scattered in multiple weakly associated transaction nodes. When the model fails to identify some latent abnormal samples, the recall rate decreases. By determining the accuracy of identifying abnormal financial transaction behavior, the overall ability of the model to capture real abnormal transaction samples can be quantified, thereby further improving the accuracy of identifying abnormal financial transaction behavior.

[0044] Please continue reading. Figure 3 The flowchart shown is a process for determining whether to reduce the decay gradient of transaction anomaly propagation using a machine learning-based financial anomaly identification method according to an embodiment of the present invention.

[0045] Specifically, given that the accuracy of identifying abnormal financial transaction behavior does not meet the requirements, the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts is determined based on the underreporting rate of abnormal transaction accounts.

[0046] Specifically, the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts is determined based on the underreporting rate of abnormal transaction accounts, including: Compare the underreporting rate of abnormal transaction accounts with the preset first underreporting rate; If the underreporting rate of the abnormal transaction account is less than or equal to the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. If the underreporting rate of the abnormal transaction accounts is greater than the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts does not meet the requirements.

[0047] Specifically, determining whether it is necessary to reduce the decay gradient of transaction anomaly propagation includes: The underreporting rate of the abnormal transaction accounts is compared with the preset first underreporting rate and the preset second underreporting rate, respectively; If the underreporting rate of the abnormal transaction account is greater than the preset first underreporting rate and less than or equal to the preset second underreporting rate, then the transaction anomaly propagation attenuation gradient is reduced. If the underreporting rate of the abnormal transaction account is greater than the preset second underreporting rate, it is determined that there is no need to reduce the transaction anomaly propagation attenuation gradient.

[0048] It is understandable that the preset first false negative rate is less than the preset second false negative rate, and the three intervals divided by the preset first false negative rate and the preset second false negative rate correspond to three different scenarios: The first interval is when the false alarm rate is less than or equal to the preset first false alarm rate, which corresponds to the situation where the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. The second interval is when the false negative rate is greater than the preset first false negative rate and less than or equal to the preset second false negative rate. The corresponding situation is: because abnormal transactions are usually hidden, the abnormal features are weakly manifested at local transaction nodes, which leads to the interruption of the deep abnormal propagation chain and reduces the sensitivity of the model to identify potential abnormal behavior. The third interval is when the underreporting rate is greater than the preset second underreporting rate. The corresponding situation is that the offsetting behavior will reconstruct the original transaction sequence through the offsetting of positive and negative amounts or netting, which causes the original transaction distribution in the feature space to undergo nonlinear compression and structural distortion, resulting in the weakening or masking of the statistical characteristics of abnormal transaction patterns. At this time, it is necessary to further determine whether the transaction constraint feature vector meets the requirements for the comprehensiveness of the representation of financial transaction behavior.

[0049] Understandably, in the process of identifying financial anomalies based on machine learning, a preset first and second false negative rate are used to characterize the sensitivity of the transaction monitoring model in identifying potentially real abnormal transaction accounts. The core logic is to transform the model's sensitivity in identifying potentially real abnormal transaction accounts into a quantifiable false negative rate range. By comparing the actual false negative rate with the preset false negative rate threshold, it is determined whether the accuracy of anomaly identification meets the requirements. The preset first and second false negative rates can be set according to actual working conditions. The setting of the preset first and second false negative rates aims to ensure the accuracy and practicality of identifying financial transaction anomalies. Optionally, the preset first and second false negative rates are determined through a limited number of trials by evaluating the effect of different false negative rates on the identification of financial transaction anomalies. The determined preset first and second false negative rates should be neither too small nor too disruptive to the identification process of financial transaction anomalies. For example, the preset first false negative rate is generally selected in the range of [1%, 3%], and the preset second false negative rate is generally selected in the range of [4%, 6%].

[0050] Preferably, the first false negative rate is 2% in a preferred embodiment, and the second false negative rate is 5% in a preferred embodiment.

[0051] Specifically, the underreporting rate of abnormal transaction accounts is the ratio of the number of accounts that are actually abnormal but not identified by the model to the total number of actual abnormal transaction accounts.

[0052] Specifically, the reduction in the transaction anomaly propagation attenuation gradient is determined by the difference between the underreporting rate of abnormal transaction accounts and a preset first underreporting rate.

[0053] Specifically, when the difference between the underreporting rate of abnormal transactions and the preset first underreporting rate is within 2%, the transaction anomaly propagation attenuation gradient is reduced to 0.9 times the original value. When the difference between the underreporting rate of abnormal transactions and the preset first underreporting rate exceeds 2%, the transaction anomaly propagation attenuation gradient is reduced by 0.04 for every 1% increase beyond the original value, in addition to being reduced to 0.9 times the original value. For example, when the difference between the underreporting rate of abnormal transactions and the preset first underreporting rate is 3%, the current transaction anomaly propagation attenuation gradient is 0.7, and the reduced transaction anomaly propagation attenuation gradient is 0.7×0.9-0.04×1=0.59.

[0054] In implementation, this invention adjusts the decay gradient of abnormal transaction propagation by setting a preset first false negative rate and a preset second false negative rate. Since abnormal transactions are usually concealed, the abnormal features are weakly manifested at local transaction nodes, which leads to the truncation of deep abnormal propagation chains and reduces the model's sensitivity to identify potential abnormal behavior. By reducing the decay gradient of abnormal transaction propagation, the decay rate of abnormal features in cross-account and multi-level transaction propagation can be reduced, so that the abnormal behavior features in weakly correlated nodes can still maintain a high correlation strength in deep propagation links. This enhances the model's ability to continuously perceive long-link concealed abnormal transactions and further improves the accuracy of identifying abnormal financial transaction behavior.

[0055] Please continue reading. Figure 4 The diagram shown is a logical flowchart of the process of determining the net amount compression coefficient for transaction constraint execution using the financial anomaly identification method based on machine learning in an embodiment of the present invention.

[0056] Specifically, based on the condition that the underreporting rate of the abnormal transaction accounts is greater than the preset second underreporting rate, it is initially determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior is determined based on the fraudulent transaction identification rate.

[0057] Specifically, the comprehensiveness of the financial transaction behavior representation by the transaction constraint feature vector is determined based on the fraudulent transaction detection rate, including: Compare the fraudulent transaction detection rate with the preset detection rate; If the fraudulent transaction identification rate is greater than the preset identification rate, then the transaction constraint feature vector is determined to meet the requirements for comprehensive representation of financial transaction behavior. If the fraudulent transaction identification rate is less than or equal to the preset identification rate, it is determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the net amount compression coefficient for the execution of transaction constraints is increased.

[0058] It is understandable that the two preset recognition rate intervals correspond to two different scenarios: The first interval is when the fraudulent transaction identification rate is less than or equal to the preset identification rate. The corresponding situation is: because the offsetting behavior will reconstruct the original transaction sequence through the offsetting of positive and negative amounts or netting, the original transaction distribution in the feature space will undergo nonlinear compression and structural distortion, which will weaken or cover up the statistical characteristics of the abnormal transaction pattern. At this time, it is necessary to increase the offsetting net amount compression coefficient of the transaction constraint execution. The second interval is when the fraudulent transaction identification rate is greater than the preset identification rate, which corresponds to the situation where the determination of the transaction constraint feature vector meets the requirements for the comprehensiveness of the representation of financial transaction behavior.

[0059] Understandably, in the process of identifying financial anomalies based on machine learning, a preset recognition rate is used to characterize the comprehensiveness of the transaction constraint feature vector's representation of financial transaction behavior. The core logic is to transform the comprehensiveness of the transaction constraint feature vector's representation of financial transaction behavior into a quantifiable recognition rate range. By comparing the actual recognition rate with the preset recognition rate threshold, it is determined whether the comprehensiveness of the transaction constraint feature vector's representation meets the requirements. The preset recognition rate can be set according to actual working conditions. The setting of the preset recognition rate aims to ensure the accuracy and practicality of identifying financial transaction anomalies. Optionally, the preset recognition rate is determined through a limited number of trials by evaluating the identification effect of different false transaction recognition rates on financial transaction anomalies. The determined preset recognition rate should satisfy the condition that it is neither too small nor too disruptive to the identification process of financial transaction anomalies. For example, the preset recognition rate is generally selected in the range of [97%, 99%].

[0060] Preferably, the preset recognition rate is 98% in this preferred embodiment.

[0061] Specifically, the fraudulent transaction identification rate is the ratio of the number of fraudulent transactions correctly identified by the model to the total number of actual fraudulent transactions.

[0062] Specifically, the increase in the net amount compression coefficient for the execution of the transaction constraint is determined by the difference between the preset identification rate and the fraudulent transaction identification rate.

[0063] Specifically, when the difference between the preset identification rate and the fraudulent transaction identification rate is within 3%, the net amount compression coefficient for transaction constraint execution increases to 1.1 times the original value. When the difference between the preset identification rate and the fraudulent transaction identification rate exceeds 3%, in addition to increasing to 1.1 times the original value, for every 1% exceeding 3%, the net amount compression coefficient for transaction constraint execution increases by 0.04. For example, when the difference between the preset identification rate and the fraudulent transaction identification rate is 5%, the current net amount compression coefficient for transaction constraint execution is 0.6, and the increased net amount compression coefficient for transaction constraint execution is 0.6×1.1+0.04×2=0.74.

[0064] In practice, this invention adjusts the net amount compression coefficient of the offsetting transaction constraint execution by setting a preset recognition rate. Since the offsetting behavior reconstructs the original transaction sequence through positive and negative amount offsetting or netting, the original transaction distribution in the feature space undergoes nonlinear compression and structural distortion, which weakens or masks the statistical characteristics of abnormal transaction patterns. By increasing the net amount compression coefficient of the offsetting transaction constraint execution, the smoothed cash flow differences during the offsetting process can be reverse-enhanced and mapped, making the originally compressed abnormal fluctuations explicit again in the feature space, repairing the feature distribution offset caused by the offsetting, and further improving the accuracy of identifying abnormal financial transaction behaviors.

[0065] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for identifying financial anomalies based on machine learning, characterized in that, include: Multi-source financial transaction data and transaction-related constraint relationships are collected separately. The collected multi-source financial transaction data are then cleaned, missing data is marked, time-series aligned, fused, and feature extracted to obtain financial features. A constraint knowledge base is then constructed based on the transaction-related constraint relationships. Based on the constraint knowledge base, apply the applicable constraints to each financial transaction to obtain the transaction constraint feature vector, obtain the financial transaction relationship to construct the abnormal relationship graph, and generate abnormal propagation features based on the abnormal relationship graph. The financial features, the transaction constraint feature vector, and the anomaly propagation features are input into a machine learning model for training to obtain a transaction monitoring model. Based on the transaction monitoring model, real-time financial transactions are analyzed to obtain anomaly identification results. Obtain the recall rate of the anomaly identification results to determine whether the accuracy of the identification of abnormal financial transaction behavior meets the requirements; If the accuracy of the identification of abnormal financial transaction behavior does not meet the requirements, the underreporting rate of abnormal transaction accounts is obtained to determine whether the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. If the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts does not meet the requirements, then it is determined whether it is necessary to reduce the transaction anomaly propagation attenuation gradient. If it is not necessary to reduce the decay gradient of abnormal transaction propagation, then the net amount compression factor for transaction constraint execution is determined based on the identification rate of fraudulent transactions.

2. The financial anomaly identification method based on machine learning according to claim 1, characterized in that, The recall rate based on the anomaly identification results determines whether the accuracy of identifying abnormal financial transaction behavior meets the requirements, including: The recall rate of the anomaly identification results is compared with the preset recall rate; If the recall rate of the anomaly identification results is greater than the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to meet the requirements. If the recall rate of the anomaly identification result is less than or equal to the preset recall rate, then the accuracy of the identification of abnormal financial transaction behavior is determined to be unsatisfactory.

3. The financial anomaly identification method based on machine learning according to claim 2, characterized in that, If the accuracy of identifying abnormal financial transaction behavior does not meet the requirements, the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts is determined based on the underreporting rate of abnormal transaction accounts.

4. The financial anomaly identification method based on machine learning according to claim 3, characterized in that, The sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts is determined based on the underreporting rate of abnormal transaction accounts, including: Compare the underreporting rate of abnormal transaction accounts with the preset first underreporting rate; If the underreporting rate of the abnormal transaction account is less than or equal to the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts meets the requirements. If the underreporting rate of the abnormal transaction accounts is greater than the preset first underreporting rate, then it is determined that the sensitivity of the transaction monitoring model in identifying potential real abnormal transaction accounts does not meet the requirements.

5. The financial anomaly identification method based on machine learning according to claim 4, characterized in that, Determine whether it is necessary to reduce the decay gradient of transaction anomaly propagation, including: The underreporting rate of the abnormal transaction accounts is compared with the preset first underreporting rate and the preset second underreporting rate, respectively; If the underreporting rate of the abnormal transaction account is greater than the preset first underreporting rate and less than or equal to the preset second underreporting rate, it is determined to reduce the attenuation gradient of abnormal transaction propagation. If the underreporting rate of the abnormal transaction account is greater than the preset second underreporting rate, it is determined that there is no need to reduce the transaction anomaly propagation attenuation gradient.

6. The financial anomaly identification method based on machine learning according to claim 5, characterized in that, The reduction in the attenuation gradient of abnormal transaction propagation is determined by the difference between the underreporting rate of abnormal transaction accounts and a preset first underreporting rate.

7. The financial anomaly identification method based on machine learning according to claim 6, characterized in that, Given that the underreporting rate of the abnormal transaction accounts is greater than the preset second underreporting rate, it is initially determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior is determined based on the fraudulent transaction identification rate.

8. The financial anomaly identification method based on machine learning according to claim 7, characterized in that, The comprehensiveness of the financial transaction behavior representation by the transaction constraint feature vector is determined based on the fraudulent transaction detection rate, including: Compare the fraudulent transaction detection rate with the preset detection rate; If the fraudulent transaction identification rate is greater than the preset identification rate, then the transaction constraint feature vector is determined to meet the requirements for comprehensive representation of financial transaction behavior. If the fraudulent transaction identification rate is less than or equal to the preset identification rate, it is determined that the comprehensiveness of the transaction constraint feature vector in representing financial transaction behavior does not meet the requirements, and the net amount compression coefficient for the execution of transaction constraints is increased.

9. The financial anomaly identification method based on machine learning according to claim 8, characterized in that, The fraudulent transaction identification rate is the ratio of the number of fraudulent transactions identified in the anomaly identification results to the total number of actual fraudulent transactions.

10. The financial anomaly identification method based on machine learning according to claim 9, characterized in that, The increase in the net amount compression coefficient for the execution of the transaction constraints is determined by the difference between the preset identification rate and the fraudulent transaction identification rate.

Images