Multi-source data fusion enterprise associated transaction audit intelligent checking system

Abstract

The invention relates to the technical field of transaction audit checking, in particular to a multi-source data fusion enterprise associated transaction audit intelligent checking system which comprises a data acquisition module, a network construction module, a chiral feature generation module, a feature injection module, a test distribution module, a distribution calculation module and a statistical test module. The chiral feature generation module is used for generating a plurality of chiral feature pairs based on the transaction data, each chiral feature pair comprises a left-hand feature and a right-hand feature, and the left-hand feature and the right-hand feature are equivalent in three dimensions of transaction amount, transaction category and transaction quantity; by setting a chiral feature pair and injecting a test transaction which is completely equivalent in commercial essence and is only in mirror symmetry in microscopic features into a normal transaction stream, a ghost agency can be effectively induced to expose systematic preference of the ghost agency in a coordination behavior, so that the ghost agency is identified from massive noise, and the accuracy of the ghost agency is improved. And effective detection of the ghost medium with extremely high concealment is realized.

Description

Technical Field

[0001] This invention relates to the technical field of transaction auditing and verification, and in particular to an intelligent verification system for enterprise related-party transaction auditing that integrates multi-source data. Background Technology

[0002] In the field of auditing related-party transactions, identifying and verifying concealed and unfair related-party transactions has always been a core challenge for regulation and internal control. Traditional auditing methods mainly rely on verifying information about directly related parties, such as equity relationships, overlapping senior management, and registered addresses, and then analyzing the fairness of transaction pricing and terms based on this. However, as companies' methods of deliberately circumventing regulation become increasingly sophisticated, a more covert form of related-party transaction—the ghost intermediary model—has begun to emerge. In this model, related-party transactions are not conducted through entities with direct equity or personnel connections, but rather use an intermediary entity that is difficult to trace within the deep network of the supply chain as a cover. This intermediary has no connection with the transacting parties at the public information level, but substantially influences or even manipulates transaction terms and pricing by providing key but difficult-to-value technical services, channel access, quality certification, or logistics coordination, thereby achieving the transfer of benefits.

[0003] Currently, the technical means to address such issues have significant limitations. On the one hand, existing auditing systems mostly rely on rule engines to match known related parties or use graph algorithms to uncover first- and second-degree direct connections. They lack effective detection capabilities for implicit collusion carried out through long supply chains and multi-layered nested relationships, where there is no direct connection. On the other hand, even if data analysis reveals statistical anomalies in certain transactions, it is difficult to attribute these anomalies to the systematic manipulation of a specific, hidden intermediary entity. Instead, they are easily interpreted as random market fluctuations or business coincidences. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes an intelligent verification system for auditing related-party transactions of enterprises through multi-source data fusion, comprising: The data acquisition module is used to collect transaction data from the enterprise resource planning system, supply chain management system, and financial system. The transaction data includes contract number, transaction timestamp, transaction amount, transaction category, and counterparty information. The network construction module is used to construct the supply chain network topology based on the counterparty information in the transaction data; A chiral feature generation module is used to generate multiple chiral feature pairs based on the transaction data; The feature injection module is used to inject the chiral feature pairs into the normal transaction process of the supply chain network topology to form a test transaction set that includes left-hand test transactions and right-hand test transactions; The test distribution module is used to execute the test transaction distribution process, distributing the test transaction set to each counterparty in the supply chain network topology. The data collection module is used to collect counterparty response data to the left-hand test transaction and the right-hand test transaction based on counterparty information in the transaction data; The distribution calculation module is used to calculate the chiral feature selection distribution based on the selection response data; The statistical testing module establishes a random distribution expectation, performs a statistical comparison test between the chiral feature selection distribution and the random distribution expectation, and generates a ghost mediator identification signal when the test result shows a statistically significant deviation.

[0005] Furthermore, multiple chiral feature pairs are generated based on the transaction data, as follows: each chiral feature pair includes a left-handed feature and a right-handed feature. The left-handed feature and the right-handed feature are equivalent in the three dimensions of transaction amount, transaction category and transaction quantity, but have mirror symmetry in the two dimensions of contract number generation algorithm and timestamp perturbation algorithm. The chiral feature generation module includes a contract number generation unit and a time perturbation unit; The contract number generation unit is used to generate left-hand contract numbers and right-hand contract numbers based on the contract number. The left-hand contract number is generated using a first prime number modular operation and a first hash salt value, and the right-hand contract number is generated using a second prime number modular operation and a second hash salt value. The first prime number modular operation and the second prime number modular operation are mirror images of each other, and the first hash salt value and the second hash salt value are inverse operations of each other. The time perturbation unit is used to apply time perturbation to the transaction timestamp, wherein the left-hand time perturbation is calculated based on the fractional part of pi, and the right-hand time perturbation is calculated based on the fractional part of the natural constant. The amplitude range of the left-hand and right-hand time perturbations is controlled within a preset time window.

[0006] Furthermore, the feature injection module includes: A transaction filtering unit is used to filter normal transactions that meet certain conditions from the transaction data; The test transaction creation unit is used to create corresponding left-hand and right-hand test transactions for each qualified normal transaction; The deployment unit is used to deploy the left-hand test transaction and the right-hand test transaction to the supply chain network through a randomized allocation mechanism.

[0007] Furthermore, the chiral feature selection distribution is calculated based on the selected response data, as follows: the chiral feature selection distribution includes a node-level chiral preference distribution and a community-level chiral preference distribution; The distributed computing module includes: The statistics unit is used to count the number of times each counterparty selected a left-hand test trade and the number of times it selected a right-hand test trade; A node-level computing unit is used to calculate the node-level chirality preference index of each trading counterparty based on the number of times; The community identification unit is used to identify the community to which a trading counterparty belongs based on the supply chain network topology. Community-level computing units are used to calculate community-level chirality preference indicators for counterparties within each community.

[0008] Furthermore, the statistical testing module includes: The hypothesis testing unit is used to establish the expected random distribution. The expected random distribution is established as follows: under the null hypothesis, each counterparty has an equal probability of choosing a left-handed test trade and a right-handed test trade, that is, they follow a binomial distribution with p=0.5. The chiral feature selection follows a binomial distribution, and the initial abnormal signal is output. An effect calculation unit is used to calculate the effect size index, which includes the selection probability difference value and the standardized mean difference; The network effect analysis unit is used to identify community divisions in the network based on the supply chain network topology using a community detection algorithm, calculate the consistency index of chiral preferences within the community, and output the network aggregation signal. The signal synthesis unit is used to receive the preliminary abnormal signal and the network aggregation signal, and when both signals are received simultaneously, it generates the ghost intermediary identification signal.

[0009] Furthermore, it also includes a causal inference module, which comprises: A matching unit is used to construct a propensity score matching model, wherein the propensity score matching model uses transaction features as covariates and chiral feature selection as a processing variable. The treatment effect unit is used to estimate the average treatment effect of chiral selection by controlling for confounding factors using the propensity score matching model. The mediation effect unit is used to perform causal mediation effect analysis to quantify the influence of phantom mediators in transaction choices. The causal mediation effect analysis includes calculating the average causal mediation effect and the average direct effect.

[0010] Furthermore, it also includes an evidence evaluation module, which comprises: The first scoring unit is used to assign the first evidence score based on the results of the statistical significance test; The second scoring unit is used to assign second evidence scores based on the effect size index; The third scoring unit is used to assign third evidence scores based on the results of network aggregation effect analysis. The comprehensive evaluation unit is used to weight and sum the scores of the first piece of evidence, the second piece of evidence, and the third piece of evidence to obtain the total strength of evidence index. The grade determination unit is used to set the threshold for the overall evidence strength grade.

[0011] Furthermore, it also includes a false alarm control module, which is used to apply an error detection rate control method to correct the preliminary abnormal signals and network aggregation signals output by the statistical test module, and to estimate the confidence intervals of the chiral feature selection distribution and effect size index through the bootstrap resampling method.

[0012] Compared with the prior art, the present invention can achieve at least the following beneficial effects: 1. This invention sets up chiral feature pairs and injects test transactions that are commercially equivalent but mirror-symmetric only in microscopic features into the normal transaction flow. This can effectively induce phantom intermediaries to expose their systematic preferences in coordination behavior, thereby identifying them from massive noise and achieving effective detection of highly concealed phantom intermediaries.

[0013] 2. This invention, through modular design, automates the entire process from multi-source data acquisition, supply chain network construction, test feature generation and injection, response data collection, distributed computation, to final statistical verification. This significantly improves audit efficiency, freeing auditors from tedious data collection and preliminary analysis, allowing them to focus on in-depth analysis of high-risk signals. It also avoids errors and subjective biases that may arise from manual operations. Beyond analyzing the behavioral preferences of individual counterparties, and more importantly, by constructing a supply chain network topology and conducting community discovery, it can identify groups of companies exhibiting highly consistent behavior. This network-level correlation analysis reveals the coordinated group behavior patterns hidden behind individuals, providing an unprecedented perspective and evidence for discovering complex collusion and fraud. It makes audit conclusions more objective, rigorous, and quantifiable, greatly enhancing the persuasiveness and reliability of audit findings, and providing a solid basis for subsequent in-depth investigations and regulatory actions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a system principle block diagram of the present invention. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] Please refer to Figure 1 This invention provides an intelligent verification system for auditing related-party transactions of enterprises through multi-source data fusion, including a data acquisition module, a network construction module, a chiral feature generation module, a feature injection module, a test distribution module, a distributed computing module, and a statistical testing module.

[0017] The data acquisition module is used to collect transaction data from enterprise resource planning systems, supply chain management systems, and financial systems. The transaction data includes contract number, transaction timestamp, transaction amount, transaction category, and counterparty information. It should be noted that the Enterprise Resource Planning (ERP) system, Supply Chain Management (SRM) system, and Financial System are internal information systems used by the enterprise to manage core business processes. The data acquisition module automatically extracts the required transaction data by establishing connections with the databases of these systems or by calling their provided application programming interfaces (APIs). This multi-source data fusion approach ensures the comprehensiveness and accuracy of the data upon which subsequent analysis is based.

[0018] The network construction module is used to construct a supply chain network topology based on the counterparty information in the transaction data. The supply chain network topology has enterprises as nodes and historical transaction relationships between enterprises as edges. It should be noted that the core function of the network construction module is to transform and abstract the aforementioned tabular transaction data into a graph structure composed of nodes and edges, i.e., a supply chain network topology. Specifically, taking enterprises as nodes means that each independent legal entity appearing in the counterparty information (including the company itself and all counterparties) is regarded as a point (node) in the network.

[0019] Specifically, using historical transaction relationships between enterprises as edges means that if two enterprises have at least one transaction record in the collected historical transaction data, a connection (edge) is established between the nodes representing these two enterprises. This edge can be weighted, with the weight determined by factors such as transaction frequency and total transaction amount; or it can be unweighted, simply indicating the existence of a transaction relationship. In this way, the originally isolated, list-style transaction data is organized into a holistic network that intuitively reflects the business connections between enterprises. This network topology serves as the foundational data carrier for subsequent graph algorithm operations such as community discovery and influence analysis. This module can run periodically to update the network structure and reflect the latest transaction status.

[0020] The chiral feature generation module is used to generate multiple chiral feature pairs based on the transaction data. Each chiral feature pair includes a left-handed feature and a right-handed feature. The left-handed feature and the right-handed feature are equivalent in the three dimensions of transaction amount, transaction category and transaction quantity, but have mirror symmetry in the two dimensions of contract number generation algorithm and timestamp perturbation algorithm. It should be noted that the "chiral feature pair" is the core innovative concept of this system, inspired by the concept of chiral molecules in chemistry. Just as the left and right hands are structurally mirror-symmetric but not completely superimposed, the left-handed and right-handed features in this system exhibit a mirror-symmetric relationship in key identification dimensions. Specifically, "maintaining equivalence in the three dimensions of transaction amount, transaction category, and transaction quantity" means that for the same basic transaction, the left-handed and right-handed test transactions are completely identical in these three core business elements. For example, if the basic transaction is "purchasing 100 tons of raw material A at a unit price of 5,000 yuan," then the generated left-handed and right-handed test transactions will maintain the exact same product name "raw material A," quantity "100 tons," and total amount "500,000 yuan," ensuring complete commercial equivalence between the two test transactions.

[0021] It should be noted that the mirror symmetry in the two dimensions of contract number generation algorithm and timestamp perturbation algorithm is specifically as follows: the contract number generation unit generates seemingly normal contract numbers that contain identification markers through specific algorithm rules. Specifically, the "first prime number modular operation" and "second prime number modular operation" are mirror images of each other, meaning that two different prime numbers are selected for modular operation. For example, the first prime number can be 1000003, and the second prime number can be 1000033. These two prime numbers are numerically close but have different mathematical properties. The "first hash salt value" and "second hash salt value" are inverse operations of each other, meaning that a hash function based on the Fibonacci sequence and its inverse function are used. For example, for a transaction with transaction ID 123, the first hash salt value may be calculated as fib_hash(123)=55, while the second hash salt value is calculated as 1000-55=945. With this design, a left-rotating contract number may generate a format such as "CONTRACT-L0555", while a right-rotating contract number will generate a format such as "CONTRACT-R9445". Both are visually valid contract numbers, but they contain different algorithmic features.

[0022] The chiral feature generation module includes a contract number generation unit and a time perturbation unit; The contract number generation unit is used to generate left-hand contract numbers and right-hand contract numbers based on the contract number. The left-hand contract number is generated using a first prime number modular operation and a first hash salt value, and the right-hand contract number is generated using a second prime number modular operation and a second hash salt value. The first prime number modular operation and the second prime number modular operation are mirror images of each other, and the first hash salt value and the second hash salt value are inverse operations of each other. The time perturbation unit is used to apply time perturbation to the transaction timestamp, wherein the left-hand time perturbation is calculated based on the fractional part of pi, and the right-hand time perturbation is calculated based on the fractional part of the natural constant. The amplitude range of the left-hand and right-hand time perturbations is controlled within a preset time window.

[0023] It should be noted that the design purpose of the time perturbation unit is to embed a recognizable pattern while maintaining the reasonableness of transaction time. The specific implementation of the left-handed time perturbation based on the fractional part of pi is as follows: take the second value of the base timestamp, multiply it by pi π, calculate its sine value, and multiply it by a preset time window value. For example, for the base timestamp "2024-05-21 14:30:00", the calculation sin(30π)300≈0*300=0 seconds, that is, the left-handed timestamp remains unchanged. The right-handed time perturbation based on the fractional part of the natural constant uses the natural constant e for a similar calculation: cos(30*e)300≈cos(81.6)300≈0.15*300=45 seconds, resulting in the right-handed timestamp "2024-05-21 14:30:45". The "preset time window" is usually set to 300 seconds (5 minutes). This time window can both ensure the natural fluctuation range of trading time and provide sufficient signal differentiation for subsequent statistical analysis.

[0024] It should be noted that this design of chiral feature pairs ensures that test transactions can be naturally integrated into the normal business flow. To transaction participants, left-handed and right-handed test transactions appear to be completely normal business opportunities, and the algorithmic differences are imperceptible. However, when a "phantom intermediary" is involved in systematic coordination, it will statistically show a systematic preference for a certain chiral feature, thereby exposing the hidden coordination behavior.

[0025] The feature injection module is used to inject the chiral feature pairs into the normal transaction process of the supply chain network topology to form a test transaction set that includes left-hand test transactions and right-hand test transactions; It's important to note that the core function of the feature injection module is to subtly integrate the generated chiral feature pairs into the company's normal supply chain transaction process. This process must ensure the naturalness and concealment of the test transaction. Specifically, the normal transaction process injected into the supply chain network is as follows: the entire process from initiation to confirmation of the test transaction is completely consistent with the company's daily real transaction process, including purchase requests, price inquiries, contract approvals, and order generation, making it impossible for transaction participants to distinguish between the test transaction and a real transaction.

[0026] The feature injection module includes a transaction screening unit, a test transaction creation unit, and a deployment unit.

[0027] The transaction screening unit is used to screen normal transactions that meet the criteria from the transaction data. The screening criteria include excluding transactions involving known related parties, ensuring that the transactions meet the commercial substance requirements, and controlling the number of test transactions to be within a preset proportion of 5% of the total transaction volume of the supply chain network. It should be noted that the transaction screening unit employs multi-level screening criteria to ensure the effectiveness and security of test transactions. The phrase "excluding transactions involving known related parties" refers to filtering out counterparties already identified as related parties using the company's existing related party list, avoiding redundant testing of known relationships and preventing alerting potential collateral. Ensuring transactions meet commercial substance requirements specifically means that the selected basic transactions must have a genuine commercial background and reasonable transaction terms, such as prices conforming to market conditions and the transaction categories falling within the company's normal business scope, avoiding any impact on the detection effect due to the irrationality of the test transactions themselves. The phrase "controlling the number of test transactions within a preset proportion of 5% of the total transaction volume of the supply chain network" is to minimize the impact on the company's normal operations while ensuring statistical significance. For example, if the company's average monthly transaction volume is 1000 transactions, the total number of test transactions injected each month should be controlled to around 50.

[0028] The test transaction creation unit is used to create corresponding left-hand test transactions and right-hand test transactions for each qualified normal transaction. The left-hand test transaction uses the left-hand contract number and left-hand timestamp, and the right-hand test transaction uses the right-hand contract number and right-hand timestamp. It should be noted that the specific workflow of the test transaction creation unit is as follows: After a qualified basic transaction is selected, the system will create two identical copies of the transaction based on its commercial terms (such as product specifications, quantity, price, etc.). One copy is given a left-handed characteristic, i.e., a left-handed contract number and a left-handed timestamp; the other copy is given a right-handed characteristic, i.e., a right-handed contract number and a right-handed timestamp. For example, based on a real "steel purchase" transaction, the system will generate two test transactions that are identical in terms of commercial terms. One uses a left-handed contract number "PO-L-20240521-001" and a left-handed timestamp "2024-05-21 10:30:00", and the other uses a right-handed contract number "PO-R-20240521-001" and a right-handed timestamp "2024-05-21 10:30:45".

[0029] The deployment unit is used to deploy the left-hand test transaction and the right-hand test transaction to the supply chain network through a randomized allocation mechanism, ensuring that the left-hand test transaction and the right-hand test transaction maintain a balance in the three dimensions of transaction time, transaction object and transaction environment.

[0030] It should be noted that the randomized allocation mechanism adopted by the deployment unit ensures the scientific nature of the testing and the reliability of the results. The specific implementation of "maintaining balance across the three dimensions of transaction time, transaction object, and transaction environment" includes: in the transaction time dimension, the issuance times of left-hand and right-hand test transactions should be evenly distributed across different working days and hours; in the transaction object dimension, the same supplier may receive left-hand or right-hand test transactions at different times, avoiding a fixed pattern; in the transaction environment dimension, ensuring that the distribution ratio of the two types of test transactions is basically consistent under various market conditions. This balance design effectively eliminates the interference of external factors on the test results, and when systematic biases occur, they can be attributed with greater certainty to the coordinating effect of "phantom intermediaries."

[0031] The test distribution module is used to execute the test transaction distribution process, distributing the test transaction set to each counterparty in the supply chain network topology.

[0032] It's important to note that the core function of the test distribution module is to simulate the transaction initiation process in real-world business scenarios. The test transaction distribution process specifically involves sending the test transaction to the target counterparty through existing business system channels, following the company's normal procurement or sales processes. Specifically, for procurement-related test transactions, the system issues an inquiry or purchase order to the supplier through the company's procurement management system; for sales-related test transactions, it issues a quotation or sales contract to the customer through the sales management system. The entire distribution process relies entirely on the company's existing information systems and standard business processes, ensuring that the test transaction is identical in form to a real transaction.

[0033] The data collection module is used to collect counterparty selection response data for the left-hand test transaction and the right-hand test transaction based on counterparty information in the transaction data. The selection response data includes accepting the transaction, rejecting the transaction, and modifying the transaction terms. It should be noted that the data collection module is responsible for comprehensively capturing the counterparty's feedback behavior regarding the test transaction. The selected response data includes three types: when the counterparty fully accepts all terms of the test transaction, it is recorded as "Acceptance of Transaction"; when the counterparty explicitly rejects the transaction or fails to respond within a certain period, it is recorded as "Rejection of Transaction"; when the counterparty requests modifications to key terms such as price, quantity, and delivery date, it is recorded as "Modification of Transaction Terms". The system will record the specific details of each response type. For example, in the case of modifying transaction terms, it will accurately record the modified terms, their original content, the modified content, and the modification time, among other detailed information.

[0034] The distribution calculation module is used to calculate the chiral feature selection distribution based on the selected response data, wherein the chiral feature selection distribution includes a node-level chiral preference distribution and a community-level chiral preference distribution; It should be noted that the core task of the distributed computing module is to transform the collected discrete response data into systematic distribution characteristics, thereby revealing potential patterns. The chiral feature selection distribution is a multi-level analytical framework, where the node-level chiral preference distribution focuses on the selection behavior of individual counterparties, while the "community-level chiral preference distribution" analyzes selection patterns from a more macro-level group perspective. This two-layer analytical structure can simultaneously capture individual characteristics and group effects, providing richer evidence for identifying "phantom intermediaries."

[0035] The distributed computing module includes: a statistical unit, a node-level computing unit, a community identification unit, and a community-level computing unit.

[0036] The statistics unit is used to count the number of times each counterparty selected a left-hand test trade and the number of times it selected a right-hand test trade; It should be noted that the statistical unit uses a cumulative counting method to build the basic dataset. Specifically, for each counterparty, the system will count the total number of times they selected left-handed test transactions and the total number of times they selected right-handed test transactions across all test scenarios. For example, supplier A received 15 test transaction opportunities during a three-month testing period, selecting left-handed test transactions 10 times and right-handed test transactions 5 times. This basic count data provides the numerator and denominator for subsequent probability calculations. The system ensures that the statistical time window is long enough to eliminate the influence of random factors; it is generally recommended to include at least 20 test transaction opportunities to produce reliable statistical results.

[0037] The node-level computing unit is used to calculate the node-level chirality preference index of each counterparty based on the number of times. The node-level chirality preference index includes the left-hand selection probability and the right-hand selection probability. It should be noted that the node-level computing unit transforms the basic counting data into more interpretable probability indicators. The formula for calculating the "left-hand selection probability" is: the number of times a left-hand test transaction is selected divided by the total number of test transaction opportunities. Using the previous example, supplier A's left-hand selection probability is 10 / 15 ≈ 0.67, and its right-hand selection probability is 5 / 15 ≈ 0.33. Theoretically, these two probabilities should be close to 0.5 in the absence of any external influences. When a counterparty's left-hand or right-hand selection probability deviates significantly from 0.5, it indicates a possible systematic bias. The system will establish an independent selection probability profile for each counterparty and track the trends of these probabilities over time.

[0038] The community identification unit is used to identify the community to which a trading counterparty belongs based on the supply chain network topology; It should be noted that the community identification unit employs complex network analysis algorithms to discover implicit group structures within the supply chain. Specifically, the "identification of the communities to which trading counterparties belong based on the supply chain network topology" is implemented through modularly optimized community discovery algorithms, such as the Louvain algorithm or the Infomap algorithm. These algorithms can automatically divide the network into several communities with tightly connected internal connections and sparse external connections based on the connection density and patterns between nodes. For example, in a supply chain network with 200 nodes, the algorithm may identify five main communities, each containing 30-50 enterprises. These communities often correspond to different industry segments or geographical clusters.

[0039] The community-level computing unit is used to calculate the community-level chirality preference index of counterparties within each community. The community-level chirality preference index includes the community average selection probability and the community selection consistency coefficient.

[0040] It should be noted that the community-level computing unit performs group behavior analysis based on community division. The "community average selection probability" refers to the arithmetic mean of the selection probabilities of all member nodes within a community, reflecting the overall chiral preference direction of the community. The "community selection consistency coefficient," calculated using the reciprocal of the standard deviation or coefficient of variation, measures the degree of consistency in the selection behavior of members within the community. For example, the average left-handed selection probability of community C1 is 0.72, and the consistency coefficient is 0.85; the average left-handed selection probability of community C2 is 0.48, and the consistency coefficient is 0.92. A high consistency coefficient combined with an average selection probability significantly deviating from 0.5 is a strong indicator of the existence of "phantom intermediaries," as it suggests that firms within the community exhibit highly coordinated selection behavior.

[0041] The statistical testing module establishes a random distribution expectation, performs a statistical comparison test between the chiral feature selection distribution and the random distribution expectation, and generates a ghost mediator identification signal when the test result shows a statistically significant deviation. The statistical testing module includes: a hypothesis testing unit, an effect calculation unit, a network effect analysis unit, and a signal synthesis unit.

[0042] The hypothesis testing unit is used to establish a random distribution expectation, which is established as follows: Under the null hypothesis, each counterparty has an equal probability of selecting a left-handed test trade and a right-handed test trade, i.e., they follow a binomial distribution with p=0.5. Chiral feature selection also follows a binomial distribution, and a preliminary anomaly signal is output, indicating that the left-handed test trade and the right-handed test trade are selected with equal probability. The exact binomial test is used to calculate the probability of deviation between the observed selection distribution and the expected distribution. A significance level threshold of 0.05 is set; when the probability of deviation is less than the significance level threshold of 0.05, a preliminary anomaly signal is output. The effect calculation unit is used to calculate the effect size index, which includes the selection probability difference value and the standardized mean difference; The network effect analysis unit is used to identify community divisions in the network based on the supply chain network topology using a community detection algorithm, calculate the consistency index of chiral preferences within a community (including intra-community correlation coefficient and Kendall's concordance coefficient), calculate the correlation index of chiral preferences between communities (including Pearson correlation coefficient and Spearman's rank correlation coefficient), and output a network aggregation signal when the intra-community chiral preference consistency index exceeds a first preset threshold of 0.8 and the inter-community chiral preference correlation index exceeds a second preset threshold of 0.6. The signal synthesis unit is used to receive the preliminary abnormal signal and the network aggregation signal, and when both signals are received simultaneously, it generates the ghost intermediary identification signal.

[0043] The intelligent verification system also includes an evidence evaluation module, which includes: The first scoring unit is used to assign the first evidence score based on the statistical significance test results. The bias probability value is assigned 3 points when it is less than 0.001, 2 points when it is less than 0.01, and 1 point when it is less than 0.05. The second scoring unit is used to assign a second evidence score based on the effect size index, where 3 points are assigned when the effect size index value is greater than 0.2, 2 points are assigned when it is greater than 0.1, and 1 point is assigned when it is greater than 0.05; The third scoring unit is used to assign third evidence scores based on the network clustering effect analysis results, with 2 points assigned when the network clustering effect consistency index is greater than 0.8; The comprehensive evaluation unit is used to weight and sum the scores of the first piece of evidence, the second piece of evidence, and the third piece of evidence to obtain the total strength of evidence index. The rating determination unit is used to set the threshold for the overall evidence strength level. When the overall evidence strength index is greater than or equal to 6, it is determined to be a high confidence level; when it is greater than or equal to 4, it is determined to be a medium confidence level; and in other cases, it is determined to be a low confidence level.

[0044] The intelligent verification system also includes a false alarm control module, which is used to apply the error detection rate control method to correct the preliminary abnormal signals and network aggregation signals output by the statistical test module, and to estimate the confidence intervals of the chiral feature selection distribution and effect size index through the bootstrap resampling method.

[0045] It should be noted that the false positive control module employs multiple hypothesis testing correction (FDR) techniques to prevent Type I errors in statistical inference. Specifically, the FDR method uses the Benjamini-Hochberg procedure. This method first sorts all statistical test p-values ​​in ascending order. Then, for the i-th smallest p-value, it compares it with (i / m)α (where m is the total number of tests and α is the significance level). The largest k such that p(k) ≤ (k / m)α is found, and finally, the first k tests are considered significant. For example, when simultaneously testing the consistency of preferences among 20 communities, using a traditional significance level of 0.05 might lead to multiple false positives. However, with FDR control, only those tests with p-values ​​less than (i / 20)*0.05 are considered significant.

[0046] It should be noted that the bootstrap resampling method estimates the sampling distribution of statistics by repeatedly sampling the original data with replacement. The specific implementation process for estimating the confidence intervals of the chiral selection distribution and effect size indices is as follows: 1000 bootstrap samples are drawn with replacement from the original sample. For each bootstrap sample, the chiral selection probability, community consistency coefficient, average treatment effect, and other statistics are recalculated. Then, based on the distribution of these 1000 bootstrap statistics, the 2.5% and 97.5% quantiles are used as the upper and lower limits of the 95% confidence interval. For example, if the point estimate of the average left-handed selection probability of a community is 0.72, the 95% confidence interval calculated using the bootstrap method is [0.68, 0.76]. This interval estimate provides a quantitative basis for the reliability of the statistical conclusions.

[0047] In summary, this patent constructs an automated system that integrates multi-source data fusion, chiral testing and active transaction injection, supply chain network analysis and statistical hypothesis testing. This system enables effective, accurate and quantifiable intelligent identification and verification of phantom intermediary-style hidden related transactions that are difficult to detect by traditional methods and exert influence indirectly through complex supply chain networks.

[0048] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0049] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0050] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0051] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0052] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-source data fusion-based intelligent auditing system for related-party transactions of enterprises, characterized in that: include: The data acquisition module is used to collect transaction data from the enterprise resource planning system, supply chain management system, and financial system. The transaction data includes contract number, transaction timestamp, transaction amount, transaction category, and counterparty information. The network construction module is used to construct the supply chain network topology based on the counterparty information in the transaction data; A chiral feature generation module is used to generate multiple chiral feature pairs based on the transaction data; The feature injection module is used to inject the chiral feature pairs into the normal transaction process of the supply chain network topology to form a test transaction set that includes left-hand test transactions and right-hand test transactions; The test distribution module is used to execute the test transaction distribution process, distributing the test transaction set to each counterparty in the supply chain network topology. The data collection module is used to collect counterparty response data to the left-hand test transaction and the right-hand test transaction based on counterparty information in the transaction data; The distribution calculation module is used to calculate the chiral feature selection distribution based on the selection response data; The statistical testing module establishes a random distribution expectation, performs a statistical comparison test between the chiral feature selection distribution and the random distribution expectation, and generates a ghost mediator identification signal when the test result shows a statistically significant deviation.

2. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 1, characterized in that, Based on the transaction data, multiple chiral feature pairs are generated as follows: Each chiral feature pair includes a left-handed feature and a right-handed feature. The left-handed feature and the right-handed feature are equivalent in the three dimensions of transaction amount, transaction category and transaction quantity, but have mirror symmetry in the two dimensions of contract number generation algorithm and timestamp perturbation algorithm. The chiral feature generation module includes a contract number generation unit and a time perturbation unit; The contract number generation unit is used to generate left-hand contract numbers and right-hand contract numbers based on the contract number. The left-hand contract number is generated using a first prime number modular operation and a first hash salt value, and the right-hand contract number is generated using a second prime number modular operation and a second hash salt value. The first prime number modular operation and the second prime number modular operation are mirror images of each other, and the first hash salt value and the second hash salt value are inverse operations of each other. The time perturbation unit is used to apply time perturbation to the transaction timestamp, wherein the left-hand time perturbation is calculated based on the fractional part of pi, and the right-hand time perturbation is calculated based on the fractional part of the natural constant. The amplitude range of the left-hand and right-hand time perturbations is controlled within a preset time window.

3. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 1, characterized in that, The feature injection module includes: A transaction filtering unit is used to filter normal transactions that meet certain conditions from the transaction data; The test transaction creation unit is used to create corresponding left-hand and right-hand test transactions for each qualified normal transaction; The deployment unit is used to deploy the left-hand test transaction and the right-hand test transaction to the supply chain network through a randomized allocation mechanism.

4. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 1, characterized in that, The chiral feature selection distribution is calculated based on the selected response data, as follows: the chiral feature selection distribution includes a node-level chiral preference distribution and a community-level chiral preference distribution; The distributed computing module includes: The statistics unit is used to count the number of times each counterparty selected a left-hand test trade and the number of times it selected a right-hand test trade; A node-level computing unit is used to calculate the node-level chirality preference index of each trading counterparty based on the number of times; The community identification unit is used to identify the community to which a trading counterparty belongs based on the supply chain network topology. Community-level computing units are used to calculate community-level chirality preference indicators for counterparties within each community.

5. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 1, characterized in that, The statistical testing module includes: The hypothesis testing unit is used to establish the expected random distribution. The expected random distribution is established as follows: under the null hypothesis, each counterparty has an equal probability of choosing a left-handed test trade and a right-handed test trade, that is, they follow a binomial distribution with p=0.

5. The chiral feature selection follows a binomial distribution, and the initial abnormal signal is output. An effect calculation unit is used to calculate the effect size index, which includes the selection probability difference value and the standardized mean difference; The network effect analysis unit is used to identify community divisions in the network based on the supply chain network topology using a community detection algorithm, calculate the consistency index of chiral preferences within the community, and output the network aggregation signal. The signal synthesis unit is used to receive the preliminary abnormal signal and the network aggregation signal, and when both signals are received simultaneously, it generates the ghost intermediary identification signal.

6. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 1, characterized in that, It also includes a causal inference module, which includes: A matching unit is used to construct a propensity score matching model, wherein the propensity score matching model uses transaction features as covariates and chiral feature selection as a processing variable. The treatment effect unit is used to estimate the average treatment effect of chiral selection by controlling for confounding factors using the propensity score matching model. The mediation effect unit is used to perform causal mediation effect analysis to quantify the influence of phantom mediators in transaction choices. The causal mediation effect analysis includes calculating the average causal mediation effect and the average direct effect.

7. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 5, characterized in that, It also includes an evidence evaluation module, which comprises: The first scoring unit is used to assign the first evidence score based on the results of the statistical significance test; The second scoring unit is used to assign second evidence scores based on the effect size index; The third scoring unit is used to assign third evidence scores based on the results of network aggregation effect analysis. The comprehensive evaluation unit is used to weight and sum the scores of the first piece of evidence, the second piece of evidence, and the third piece of evidence to obtain the total strength of evidence index. The grade determination unit is used to set the threshold for the overall evidence strength grade.

8. The intelligent verification system for auditing related-party transactions of enterprises based on multi-source data fusion as described in claim 5, characterized in that, It also includes a false alarm control module, which is used to apply the error detection rate control method to correct the initial abnormal signal and network aggregation signal output by the statistical test module, and to estimate the confidence interval of the chiral feature selection distribution and effect size index by using the bootstrap resampling method.

Images