Supplier illegal behavior intelligent identification method and system based on Monte Carlo method

Abstract

The invention discloses a supplier illegal behavior intelligent identification method and system based on a Monte Carlo method, and relates to the technical field of data processing, and the method comprises the steps: carrying out the preprocessing of an original bidding record data set, and obtaining an analysis bidding record data set; extracting multi-dimensional features by a multi-dimensional analysis feature mechanism, and carrying out supplier relation network feature quantitative processing to obtain an initial feature engineering data set; fusing the initial feature engineering data set into an enhanced multi-dimensional feature engineering data set by using multi-granularity time analysis mechanism time sequence mode analysis; constructing an abnormal behavior portrait by using a Monte Carlo simulation model, constructing a multi-level data view, simulating the multi-level data view, and obtaining a quantitative comprehensive risk score; and performing assessment and identification by using a multi-dimensional risk assessment and intelligent identification terminal to obtain an integrity comprehensive risk score, dividing suppliers based on a preset threshold, and generating an illegal behavior adjustment report. According to the invention, differentiated schemes are provided for suppliers with different risks, and the scientificity and pertinence of management and control are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for intelligent identification of supplier violations based on the Monte Carlo method. Background Technology

[0002] In procurement bidding, collusion among suppliers can occur, such as bid rigging and collusion. A common tactic is manipulating bids to ensure a specific supplier wins the contract. A one-yuan difference in bids could be a deliberate strategy; for example, one supplier might bid 100 yuan while another bids 101 yuan, creating the illusion of intense competition when in reality it's collusion. Because the price difference is extremely small, it usually doesn't affect the bidding outcome—for example, the supplier with the lower bid wins. However, it could be used to conceal collusion, such as rotating bids or protecting specific suppliers. The Monte Carlo method is a statistical simulation method based on random sampling used to assess the probability of an event occurring. In this scenario, by simulating a large number of random bidding scenarios, the frequency of a one-yuan bid difference under random conditions is observed, and the actual observed frequency is compared with the simulated frequency. If the actual frequency is significantly higher than the simulated frequency, it indicates that the one-yuan bid difference may not be randomly generated, suggesting potential irregularities.

[0003] The existing approximation scheme uses the traditional hypothesis testing method, the chi-square goodness-of-fit test, to test whether the observed price difference distribution conforms to a uniform distribution. The steps are as follows: Establish hypotheses: Null hypothesis (H0): The distribution of price differences is not significantly different from a uniform distribution. Alternative hypothesis (H1): The distribution of price differences is significantly different from a uniform distribution. Prepare data: Collect price difference data, i.e., the price difference between any two suppliers in all bids. Determine the range of price differences and divide the data into intervals. Calculate the theoretical frequency: Calculate the expected frequency for each interval based on the uniform distribution. Expected frequency for each interval under a uniform distribution = total frequency / number of intervals. Calculate the chi-square statistic for each interval: For each interval, calculate (observed frequency - expected frequency)^2 / expected frequency. Calculate the total chi-square statistic: Sum the chi-square statistics of each interval to obtain the total chi-square statistic. Determine the degrees of freedom: Under a uniform distribution, the degrees of freedom are the number of intervals minus 1. Calculate the p-value: p_value = 1 - chi2.cdf(chi_square_statistic, degrees_of_freedom). If the p-value is less than the significance level, the null hypothesis is rejected, and it is concluded that the price difference distribution is significantly different from the theoretical distribution, that is, the price difference distribution is not uniform and may be manipulated.

[0004] Existing approximation schemes use the traditional hypothesis testing method, the chi-square goodness-of-fit test, which has the following drawbacks: Sensitivity to distribution assumptions: The chi-square test relies on a pre-defined theoretical distribution. If this distribution does not match the actual distribution under random conditions, the test results may be unreliable. For example, in bidding, the price difference may not naturally be uniformly distributed; using a uniform distribution as the theoretical distribution would lead to erroneous conclusions. Requires sufficient sample size: The chi-square test requires that the expected frequency of each interval cannot be too small. If the sample size is insufficient, interval merging may be necessary, which may mask important details. For example, if the focus is on a 1-yuan price difference, but the expected frequency of the 1-yuan price difference is less than 5, it must be merged with other intervals, making it impossible to test whether the 1-yuan price difference is anomaly individually. May fail to detect complex collusion patterns: The chi-square test examines the entire distribution and may be insensitive to certain specific patterns, especially when such patterns are relatively small in the overall distribution. For example, if collusion occurs only in a few bids, and most price differences are random, the chi-square test may fail to detect these few but important anomalies. Low sensitivity to extreme values: After grouping the data, the chi-square test ignores the specific numerical differences within each group and only considers the frequency. It is not sensitive to extreme values ​​because extreme values ​​are often placed in the last or first group, which may contain a wide range of values, thus diluting the impact of extreme values.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes an intelligent identification method and system for supplier violations based on the Monte Carlo method, thereby overcoming the aforementioned technical problems existing in the prior art.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention proposes an intelligent identification method for supplier violations based on the Monte Carlo method, comprising:

[0009] S1. Obtain the original bid record dataset and preprocess it to obtain the analytical bid record dataset;

[0010] S2. Multidimensional feature extraction is performed on the bidding record dataset using a multidimensional analysis feature mechanism, and supplier feature quantification is performed on the extracted multidimensional feature results using a supplier relationship network to obtain the initial feature engineering dataset.

[0011] S3. Utilize a multi-granularity time analysis mechanism to perform time series pattern analysis on the bidding record dataset, and fuse the time series distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset.

[0012] S4. Construct a Monte Carlo simulation model, and use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile;

[0013] S5. Identify and filter the abnormal behavior profiles and construct a multi-level data view. Use the Monte Carlo simulation model to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. Evaluate and classify the randomization scheme to obtain a quantitative comprehensive risk score.

[0014] S6. Utilize multi-dimensional risk assessment and intelligent identification terminals to evaluate and identify quantitative comprehensive risk scores, integrate and process the evaluation and identification results to obtain a comprehensive integrity risk score, classify suppliers based on preset thresholds and comprehensive integrity risk scores, and generate violation adjustment reports to achieve supplier violation identification.

[0015] Furthermore, the original bid record dataset is obtained and preprocessed to obtain the bid record dataset for analysis, which includes:

[0016] S11. Generate the original bid record dataset based on the bid identifier, supplier identifier, bid amount, and bid time;

[0017] S12. Use a hierarchical cleaning strategy to process the missing values ​​in the original bidding record dataset, mark invalid data with missing values ​​in any field, and record the reasons and number of deletions to obtain a preliminary cleaned dataset.

[0018] S13. Use a multi-level anomaly detection method to detect the quotation amount field in the preliminary cleaned dataset. Based on the seasonal fluctuation characteristics of coal prices and the improved box plot method, identify the detection results and dynamically adjust the threshold. Use the median instead of the mean and the median absolute deviation instead of the standard deviation for the adjusted Z score. Mark the observations with an absolute value greater than the preset threshold as outliers.

[0019] S14. Using the edit distance and fuzzy matching algorithm of the business knowledge base, entity normalization is performed on the supplier name inconsistency problem to obtain intermediate data of bidding records with unified supplier identification.

[0020] S15. Integrate outlier data, data quality reports, and intermediate data from bidding records into a star schema data warehouse to form an analytical bidding record dataset.

[0021] Furthermore, a multidimensional feature extraction mechanism is used to extract multidimensional features from the bid record dataset, and a supplier relationship network is used to quantify the extracted multidimensional features, resulting in an initial feature engineering dataset including:

[0022] S21. Construct a multi-dimensional analysis feature mechanism based on price spread characteristics, relationship characteristics, time series characteristics, and derived characteristics;

[0023] S22. Using a multidimensional analysis feature mechanism, feature extraction is performed on the analysis bidding record dataset, and the feature extraction results are transformed into information carriers of collusion behavior. Based on the information carriers, a multidimensional feature set of individual supplier behaviors is generated.

[0024] S23. Construct a time-weighted supplier relationship network based on joint bidding records, and use a community detection algorithm to identify potential related groups in the supplier relationship network to obtain community structure characteristics;

[0025] S24. Utilize supplier relationship networks and community structure features to perform network correlation quantization and enhancement fusion on the multidimensional feature set to obtain the initial feature engineering dataset.

[0026] Furthermore, a time-weighted supplier relationship network is constructed based on joint bidding records, and a community detection algorithm is used to identify potential related groups in the supplier relationship network, yielding community structure characteristics including:

[0027] S231. Use the community discovery algorithm to run the consensus strategy multiple times, and run it independently a preset number of times with different random seeds.

[0028] S232. Based on the community partitioning results of the preset number of independent runs, the frequency of common communities among supplier nodes is counted, a consensus matrix of the supplier relationship network is constructed, and a consensus matrix dataset is obtained.

[0029] S233. Based on the consensus matrix dataset, input it again into the community structure detection to eliminate the randomness bias of a single algorithm run and obtain the supplier community division result;

[0030] S234. Based on the supplier community segmentation results, extract supplier nodes with high intermediary centrality to obtain the competitive health of the supplier network.

[0031] Furthermore, a multi-granularity time analysis mechanism is used to perform time-series pattern analysis on the bid record dataset, and the time-series distribution analysis results are fused with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset, including:

[0032] S31. Construct a multi-granularity time analysis mechanism based on macro-periodicity analysis, meso-aggregation analysis and micro-sequence dependence analysis;

[0033] S32. Utilize a multi-granularity time analysis mechanism to perform macro-periodic analysis on the bidding record dataset, identify abnormal patterns in bidding behavior on seasonal and cyclical time scales, and obtain macro-periodic characteristics.

[0034] S33. Use a multi-granularity time analysis mechanism to perform meso-level clustering analysis on the bidding record dataset, and test the distribution pattern on the time axis of the bidding events based on the point process model to obtain the time clustering characteristics.

[0035] S34. Using a multi-granularity time analysis mechanism, perform micro-sequence dependency analysis on the bidding record dataset to analyze the time interval of consecutive bidding events, the synchronicity of the bid sequence, and the winning bid rotation pattern to obtain behavioral sequence characteristics;

[0036] S35. Integrate and quantify the macro-periodic characteristics, temporal clustering characteristics, and behavioral sequence characteristics to obtain the time series distribution analysis results;

[0037] S36. The temporal distribution analysis results are fused with the initial feature engineering dataset to obtain an enhanced multidimensional feature engineering dataset.

[0038] Furthermore, a Monte Carlo simulation model is constructed, and this model is used to identify and construct an enhanced multidimensional feature engineering dataset, resulting in an abnormal behavior profile including:

[0039] S41. Construct a Monte Carlo simulation model based on overall frequency test, supplier pair analysis, conditional pattern test, and time series pattern analysis, combined with randomized counterfactual scenarios.

[0040] S42. A large-scale simulation was conducted using a randomization strategy that preserves the segment structure to construct a zero distribution of the univariate frequency of bid difference. The percentile ranking of the observed frequency, the effect size index, and the nonparametric confidence interval were calculated for statistical comparison and evaluation to obtain macroscopic test evidence.

[0041] S43. Screen supplier pairs with statistical significance and perform conditional randomization tests for pairing. Based on multiple comparison correction and risk assessment of fusion statistical significance, effect size and behavioral persistence, obtain combined-level test evidence.

[0042] S44. Conduct stratified Monte Carlo simulations and statistical significance assessments on basic business conditions, complex interaction conditions, and dynamic market conditions to obtain contextualized test evidence.

[0043] S45. Using Monte Carlo simulation, a stochastic benchmark of periodic, clustered, and sequence-dependent time patterns is constructed and statistically tested to obtain time series test evidence.

[0044] S46. Based on macroscopic verification evidence, combined verification evidence, contextualized verification evidence, and time-series verification evidence, cross-validation and evidence fusion are performed on the enhanced multidimensional feature engineering dataset to construct a complete profile of abnormal behavior.

[0045] Furthermore, Monte Carlo simulations were used to construct stochastic benchmarks for periodic, clustered, and sequence-dependent time patterns, and statistical tests were performed. The time-series test evidence obtained includes:

[0046] S451. Construct a time series model based on dynamic condition testing, and use the time series model to analyze the correlation between the univariate frequency of price difference and the market index;

[0047] S452. Using time series decomposition technology, the univariate event sequence of price difference is decomposed into trend component, periodic component and residual component. The periodic component is subjected to spectral analysis to identify significant periodicity. Monte Carlo simulation is used to generate event sequences with random time distribution. The difference in periodic intensity between the actual sequence and the random sequence is compared to evaluate the statistical significance of periodicity.

[0048] S453. Using point process analysis, the price difference univariate event is regarded as an event point on the time axis. The interval time distribution of the actual event sequence is calculated and compared with the random sequence generated by the homogeneous Poisson process.

[0049] S454. Use the L-function to analyze the degree of clustering across multiple time scales, and use Monte Carlo simulation to construct the confidence envelope of the K-function. Determine whether the actual K-function value exceeds the envelope to test for significant temporal clustering.

[0050] Furthermore, the abnormal behavior profiles are identified and filtered to construct a multi-level data view. The Monte Carlo simulation model is used to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. The randomization scheme is then evaluated and divided to obtain a quantitative comprehensive risk score, including:

[0051] S51. Use an adaptive threshold algorithm to analyze the global distribution of the number of joint bids in the abnormal behavior profile of suppliers to determine the quantile threshold. Combine the joint bid intensity, time coverage, price correlation, bid complementarity and multi-dimensional screening criteria to obtain the set of supplier pairs that meet the conditions.

[0052] S52. Based on the supplier's basic information on the jointly participated bid section, details of both parties' quotations, information on the competitive environment and business background, and the joint bidding event sequence, multi-dimensional features of each event point and time interval sequence arranged in chronological order, the information is input into the context data of the performance of both parties in other bid sections individually, to obtain a multi-level data view;

[0053] S53. Using the constrained randomization in the Monte Carlo simulation model, randomization tests are performed on the multi-level data view and a zero distribution is constructed to obtain the randomization scheme for the supplier conditions.

[0054] S54. Compare the actual observed price difference univariate frequency with the simulated zero distribution, calculate the p-value, effect size, relative risk ratio and perform confidence interval analysis, use the stratified false discovery rate control method to perform multiple comparison correction, and combine sensitivity analysis to test the impact of randomization scheme on the results to obtain the statistical significance assessment results.

[0055] S55. Based on the statistical significance assessment results, integrate multi-dimensional indicators such as behavioral pattern risk, time consistency, and network relationship risk, use a weighted formula to calculate a quantitative comprehensive risk score, and classify supplier pairs into high, medium, and low risk levels according to preset thresholds to obtain a quantitative comprehensive risk score.

[0056] Furthermore, a multi-dimensional risk assessment and intelligent identification terminal is used to evaluate and identify the quantitative comprehensive risk score. The evaluation and identification results are then integrated and processed to obtain an integrity comprehensive risk score. Based on preset thresholds and the integrity comprehensive risk score, suppliers are classified into different levels, and a violation adjustment report is generated to achieve supplier violation identification, including:

[0057] S61. Using a multi-dimensional risk assessment and intelligent identification terminal, the quantitative comprehensive risk score and the corresponding supplier-related data are assessed and identified to obtain the assessment input dataset.

[0058] S62. Based on the terminal, a multi-dimensional evaluation framework is used to evaluate the statistical significance dimension. Combined with indicators such as p-value, effect size, and confidence interval width obtained from the Monte Carlo test, the p-value is mapped to a statistical risk score through a transformation function to obtain the statistical risk score result.

[0059] S63. Use unsupervised anomaly detection algorithms to identify unconventional patterns in price synchronicity, price amount last digit regularity, and price adjustment consistency, quantify them as behavioral anomaly indices, and obtain behavioral pattern risk scores.

[0060] S64. Use time series analysis and change point detection techniques to identify the starting point, duration and evolution trend of anomalous patterns, assess the stability and avoidance behavior of anomalous patterns, and obtain time consistency risk scores.

[0061] S65. Calculate node centrality and community structure indicators based on supplier relationship network, and evaluate supplier association strength and connection relationship of high-risk suppliers to obtain network relationship risk score results;

[0062] S66. Use a weighted fusion model to integrate statistical risk scoring results, behavioral pattern risk scoring results, time consistency risk scoring results, and network relationship risk scoring results, and set dynamic weights according to business importance to obtain a comprehensive integrity risk score.

[0063] S67. Based on preset thresholds, suppliers or supplier pairs are divided into three risk levels: high, medium, and low, and adjustment reports are generated. Based on the violation adjustment reports, the results of supplier violation identification are achieved.

[0064] Secondly, this invention also provides an intelligent identification system for supplier violations based on the Monte Carlo method, comprising:

[0065] The bid data preprocessing module is used to acquire the original bid record dataset and preprocess it to obtain the analytical bid record dataset;

[0066] The supplier feature quantification and extraction module is used to extract multidimensional features from the analysis bidding record dataset using a multidimensional analysis feature mechanism, and to quantify the extracted multidimensional feature results using a supplier relationship network to obtain the initial feature engineering dataset.

[0067] The multi-granularity temporal feature fusion module is used to perform temporal pattern analysis on the bidding record dataset using a multi-granularity time analysis mechanism, and to fuse the temporal distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset.

[0068] The Monte Carlo simulation model analysis module is used to construct a Monte Carlo simulation model, and then use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile.

[0069] The supplier risk simulation assessment module is used to identify and screen abnormal behavior profiles and construct multi-level data views. The Monte Carlo simulation model is used to test and simulate the multi-level data views to obtain randomization schemes for supplier conditions. The randomization schemes are then evaluated and divided to obtain a quantitative comprehensive risk score.

[0070] The multidimensional risk level determination module is used to evaluate and identify the quantitative comprehensive risk score using a multidimensional risk assessment and intelligent identification terminal, and to integrate and process the evaluation and identification results to obtain the integrity comprehensive risk score. Based on the preset threshold and the integrity comprehensive risk score, the module classifies suppliers into different levels and generates a violation adjustment report to achieve the identification of supplier violations.

[0071] The beneficial effects of this invention are as follows:

[0072] 1) This invention can intelligently identify whether the phenomenon occurs randomly, in order to determine whether there is a high-risk violation, and formulate corresponding investigation and monitoring recommendations for each of the three risk levels: high, medium and low.

[0073] 2) This invention uses Monte Carlo simulation testing that combines multiple dimensions and conditions to achieve greater accuracy. It abandons the limitations of traditional mark-scraping detection that relies on simple statistical thresholds or rules. It uses the Monte Carlo method to conduct simulation testing from multiple dimensions such as overall frequency, supplier pairs, conditional patterns, and time-series patterns. At the same time, it conducts targeted testing for specific conditions, which can more comprehensively and accurately identify abnormal patterns and discover more targeted collusion behaviors.

[0074] 3) This invention enables data-driven zero-distribution construction to adapt to complex business scenarios. By constructing a zero-distribution through randomized raw data, the significance of observed values ​​can be tested without relying on any prior distribution assumptions. This avoids the dependence of traditional statistical tests on theoretical distributions and is more suitable for the complex distribution characteristics of actual business data.

[0075] 4) This invention can detect covert collusion by analyzing supplier relationship networks. It constructs a network of suppliers’ joint bidding relationships, identifies potential colluding groups with the help of community detection algorithms, and conducts risk assessment by combining network characteristics. This can break through the limitations of traditional detection methods and accurately detect more covert collusion behaviors.

[0076] 5) This invention utilizes a dynamic risk assessment system to achieve differentiated management and control. It constructs a dynamic risk assessment system that integrates multiple factors such as statistical significance, behavioral patterns, temporal consistency, and network relationships, outputs quantitative risk scores and levels, and provides differentiated treatment solutions for suppliers with different risk levels, thereby improving the scientific nature and pertinence of management and control.

[0077] 6) This invention optimizes and ensures the efficiency of large-scale data processing through parallel computing. In response to the problem of large computational load in Monte Carlo simulation, parallel computing technology is used to optimize algorithm efficiency, ensuring that the system can efficiently process large-scale bidding data and meet the business needs of real-time or near real-time monitoring.

[0078] 7) The interpretability and visualization of this invention support scientific decision-making. It not only outputs risk identification results, but also provides detailed statistical evidence and visualization reports at the same time, helping business personnel to intuitively understand the test results and providing strong support for decision-making.

[0079] 8) The adaptive learning of this invention enables continuous optimization of the system. The system can dynamically adjust model parameters and risk thresholds based on historical data feedback, forming an adaptive closed loop of data input, result feedback and parameter optimization, ensuring that the detection capability continues to improve with business development. Attached Figure Description

[0080] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in 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.

[0081] Figure 1 This is a flowchart of a supplier violation intelligent identification method based on the Monte Carlo method according to an embodiment of the present invention.

[0082] Figure 2 This is a basic flowchart of a supplier violation intelligent identification method based on the Monte Carlo method according to an embodiment of the present invention.

[0083] Figure 3 This is a detailed flowchart of a supplier violation intelligent identification method based on the Monte Carlo method according to an embodiment of the present invention.

[0084] Figure 4 This is a flowchart of the Monte Carlo simulation model in the intelligent identification method for supplier violations based on the Monte Carlo method according to an embodiment of the present invention.

[0085] Figure 5 This is a schematic diagram of a supplier violation intelligent identification system based on the Monte Carlo method according to an embodiment of the present invention. Detailed Implementation

[0086] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0087] According to embodiments of the present invention, a method and system for intelligent identification of supplier violations based on the Monte Carlo method are proposed.

[0088] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-3 As shown, the intelligent identification method for supplier violations based on the Monte Carlo method according to an embodiment of the present invention includes:

[0089] Step S1: Obtain the original bid record dataset and preprocess it to obtain the analysis bid record dataset;

[0090] Step S2: Use the multidimensional analysis feature mechanism to extract multidimensional features from the analysis bidding record dataset, and use the supplier relationship network to quantify the extracted multidimensional feature results to obtain the initial feature engineering dataset.

[0091] Step S3: Perform time series pattern analysis on the bidding record dataset using a multi-granularity time analysis mechanism, and fuse the time series distribution analysis results with the initial feature engineering dataset to obtain an enhanced multidimensional feature engineering dataset;

[0092] Step S4: Construct a Monte Carlo simulation model and use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile;

[0093] Step S5: Identify and filter the abnormal behavior profiles and construct a multi-level data view. Use the Monte Carlo simulation model to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. Evaluate and classify the randomization scheme to obtain a quantitative comprehensive risk score.

[0094] Step S6: Use a multi-dimensional risk assessment and intelligent identification terminal to assess and identify the quantitative comprehensive risk score, and integrate and process the assessment and identification results to obtain the integrity comprehensive risk score. Based on the preset threshold and the integrity comprehensive risk score, classify suppliers into different levels and generate a violation adjustment report to achieve supplier violation identification.

[0095] In this optional embodiment, the original bid record dataset is obtained and preprocessed to obtain the bid record dataset for analysis, including:

[0096] S11. Generate the original bid record dataset based on the bid identifier, supplier identifier, bid amount, and bid time;

[0097] S12. Use a hierarchical cleaning strategy to process the missing values ​​in the original bidding record dataset, mark invalid data with missing values ​​in any field, and record the reasons and number of deletions to obtain a preliminary cleaned dataset.

[0098] S13. Use a multi-level anomaly detection method to detect the quotation amount field in the preliminary cleaned dataset. Based on the seasonal fluctuation characteristics of coal prices and the improved box plot method, identify the detection results and dynamically adjust the threshold. Use the median instead of the mean and the median absolute deviation instead of the standard deviation for the adjusted Z score. Mark the observations with an absolute value greater than the preset threshold as outliers.

[0099] S14. Using the edit distance and fuzzy matching algorithm of the business knowledge base, entity normalization is performed on the supplier name inconsistency problem to obtain intermediate data of bidding records with unified supplier identification.

[0100] S15. Integrate outlier data, data quality reports, and intermediate data from bidding records into a star schema data warehouse to form an analytical bidding record dataset.

[0101] Specifically, data preparation and preprocessing are the cornerstones of the entire analysis process, and their quality directly determines the reliability and effectiveness of subsequent Monte Carlo analysis and risk identification. This stage, through a systematic data processing workflow, transforms raw, messy bidding records into a structured, high-quality analytical dataset, while extracting multi-dimensional features to support deep pattern recognition. The primary task of data cleaning is to establish a complete and consistent foundational dataset. In practice, a layered cleaning strategy is adopted: the first layer handles missing values, targeting four core fields: bid identifier, supplier identifier, bid amount, and bid time. Any record missing even one of these fields is marked as invalid data. However, unlike simple deletion, the system records the reason for and the number of deletions, generating a data quality report for subsequent traceability. For the bid amount field, a multi-level anomaly detection mechanism is implemented: based on business rules, obviously unreasonable values ​​such as zero, negative values, or extreme values ​​exceeding 10 times the historical high price are filtered out; statistical methods are used to identify outliers; and an improved box plot method (i.e., IQR method) is used in conjunction with the seasonal fluctuation characteristics of coal prices to dynamically adjust the threshold. An improved box plot method is used, but instead of simply applying a fixed threshold of 1.5 times the IQR, it is dynamically adjusted based on the actual distribution characteristics of coal procurement price data. The skewness and kurtosis of historical bid amounts are analyzed. For right-skewed distributions, indicating a small number of extremely high bids, asymmetric boundary settings are used: a lower limit of Q1 - 1.0 * IQR and an upper limit of Q3 + 2.0 * IQR. This allows for more accurate identification of truly abnormally high bids without mistakenly deleting reasonable high-price data. Simultaneously, IQR is calculated separately for data of different procurement types and coal quality specifications to avoid misjudgments due to category differences. The distribution-based method is Z-score standardization, but considering that bid data often does not follow a normal distribution, robust statistics are used for standardization. The median is used instead of the mean, and MAD (median absolute deviation) is used instead of the standard deviation. The adjusted Z-score is calculated, and observations with an absolute value greater than 3.5 are marked as outliers. In particular, for the issue of inconsistent supplier names, such as a certain coal industry group and a certain coal industry group company being regarded as different suppliers, a fuzzy matching algorithm based on edit distance and business knowledge base is used for entity normalization to ensure the accuracy of supplier identification.

[0102] All processed data is integrated into a unified analytical data warehouse, employing a star schema design. Centered on the bidding fact table, it connects supplier, time, and bid segment dimensions. Data quality verification is integrated throughout the entire process: in the cleaning phase, data loss rates are recorded and corrections are made; in the feature engineering phase, the rationality and stability of feature distributions are checked; in the network construction phase, network connectivity and scale-free properties are verified; and in the time analysis phase, the stationarity of the time series is checked. The final output dataset not only includes the cleaned raw fields but also four types of feature matrices, network topology, and time series objects, providing a comprehensive and high-quality data foundation for subsequent Monte Carlo simulations and machine learning analysis.

[0103] In this optional embodiment, a multidimensional feature extraction mechanism is used to extract multidimensional features from the bid record dataset, and a supplier relationship network is used to quantify the extracted multidimensional features to obtain an initial feature engineering dataset including:

[0104] S21. Construct a multi-dimensional analysis feature mechanism based on price spread characteristics, relationship characteristics, time series characteristics, and derived characteristics;

[0105] S22. Using a multidimensional analysis feature mechanism, feature extraction is performed on the analysis bidding record dataset, and the feature extraction results are transformed into information carriers of collusion behavior. Based on the information carriers, a multidimensional feature set of individual supplier behaviors is generated.

[0106] S23. Construct a time-weighted supplier relationship network based on joint bidding records, and use a community detection algorithm to identify potential related groups in the supplier relationship network to obtain community structure characteristics;

[0107] S24. Utilize supplier relationship networks and community structure features to perform network correlation quantization and enhancement fusion on the multidimensional feature set to obtain the initial feature engineering dataset.

[0108] In this optional embodiment, a time-weighted supplier relationship network is constructed based on joint bidding records, and a community detection algorithm is used to identify potential related groups in the supplier relationship network, resulting in community structure characteristics including:

[0109] S231. Use the community discovery algorithm to run the consensus strategy multiple times, and run it independently a preset number of times with different random seeds.

[0110] S232. Based on the community partitioning results of the preset number of independent runs, the frequency of common communities among supplier nodes is counted, a consensus matrix of the supplier relationship network is constructed, and a consensus matrix dataset is obtained.

[0111] S233. Based on the consensus matrix dataset, input it again into the community structure detection to eliminate the randomness bias of a single algorithm run and obtain the supplier community division result;

[0112] S234. Based on the supplier community segmentation results, extract supplier nodes with high intermediary centrality to obtain the competitive health of the supplier network.

[0113] Specifically, the core objective of feature engineering is to transform raw data into an information carrier capable of revealing potential collusion, i.e., a feature engineering dataset. The construction of a multi-dimensional analysis feature mechanism includes four main feature families: price difference features, relationship features, time series features, and derived features. Price difference features not only calculate absolute differences but also introduce relative difference indicators, such as standardized price difference (in other words, the price difference divided by the bid budget) and ranking difference (i.e., the difference in the ranking and quantile of the bid in the current bidding). Relationship features extract the frequency of joint bidding, the joint bidding rate, and the intensity of competition for each pair of suppliers by constructing a supplier-bid participation matrix. The joint bidding rate is the proportion of joint bidding times to the total number of bids for each supplier. Regarding time series features, the distribution of bidding time intervals, bid response times, and autocorrelation of historical bid sequences are calculated for each supplier. The bid response time is the time difference from the issuance of the tender document to the bid. Derived features combine business knowledge, such as calculating the price difference patterns of supplier pairs in historical bidding situations, identifying whether there is a recurring pattern where the price difference between supplier A and supplier B is consistently 1 yuan when supplier A wins the bid, and vice versa.

[0114] Supplier relationship networks are a key tool for detecting collusive behavior. A multi-layered network with temporal weights is constructed: the base layer uses the number of times joint bidding occurs as edge weights; the behavioral layer uses price similarity as weights, including the frequency of price differences of 1 yuan; and the temporal layer reflects changes in relationship strength over different time periods. After network construction, multi-scale community detection is performed using the Louvain algorithm (i.e., a community detection algorithm) to identify the global community structure. A weighted undirected network is constructed based on joint bidding data, where the edge weights w(i,j) = α*f(i,j) + β*s(i,j) + γ*t(i,j). In this equation, w(i,j) represents the temporally weighted network edge weight between supplier i and supplier j in the supplier relationship network, f(i,j) represents the standardized value of the joint bidding frequency between suppliers i and j, s(i,j) represents the similarity of bidding behavior, including the frequency of price differences of 1 yuan, and t(i,j) represents the temporal coordination index. α, β, and γ represent weight parameters set based on business experience. Typically, α is set to a higher value to emphasize the fundamental role of the frequency of joint bidding. Considering the randomness of the Louvain algorithm, a strategy of running it multiple times to reach consensus is adopted to enhance the stability of the results. Specifically, the algorithm is run independently 50 times, each time using a different random seed, resulting in 50 community partitioning results. A consensus matrix C is constructed based on these partitions, where C(i,j) represents the frequency at which suppliers i and j are assigned to the same community. The Louvain algorithm is then applied again to the consensus matrix to obtain the final stable community partitioning. This method significantly improves the robustness of community discovery and reduces the fluctuations caused by random initialization.

[0115] Simultaneously, local clustering coefficients are used to analyze the formation of small groups. Network indicators include node centrality and edge centrality. Node centrality includes degree centrality, proximity centrality, and eigenvector centrality, with particular attention paid to suppliers with high betweenness centrality, as these suppliers play a key role in information transmission or coordination. Furthermore, the structural entropy and clustering coefficient of the network are analyzed to assess the competitive health of the entire supplier network. For detected close communities, internal pricing difference patterns are analyzed. If a price difference of 1 yuan frequently occurs within a community but rarely between communities, it suggests possible collusion by groups.

[0116] In this optional embodiment, a multi-granularity time analysis mechanism is used to perform time-series pattern analysis on the bid record dataset, and the time-series distribution analysis results are fused with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset, including:

[0117] S31. Construct a multi-granularity time analysis mechanism based on macro-periodicity analysis, meso-aggregation analysis and micro-sequence dependence analysis;

[0118] S32. Utilize a multi-granularity time analysis mechanism to perform macro-periodic analysis on the bidding record dataset, identify abnormal patterns in bidding behavior on seasonal and cyclical time scales, and obtain macro-periodic characteristics.

[0119] S33. Use a multi-granularity time analysis mechanism to perform meso-level clustering analysis on the bidding record dataset, and test the distribution pattern on the time axis of the bidding events based on the point process model to obtain the time clustering characteristics.

[0120] S34. Using a multi-granularity time analysis mechanism, perform micro-sequence dependency analysis on the bidding record dataset to analyze the time interval of consecutive bidding events, the synchronicity of the bid sequence, and the winning bid rotation pattern to obtain behavioral sequence characteristics;

[0121] S35. Integrate and quantify the macro-periodic characteristics, temporal clustering characteristics, and behavioral sequence characteristics to obtain the time series distribution analysis results;

[0122] S36. The temporal distribution analysis results are fused with the initial feature engineering dataset to obtain an enhanced multidimensional feature engineering dataset.

[0123] Specifically, time pattern analysis aims to reveal the regularity and concealment of violations over time. A multi-granularity time analysis approach is employed: at the macro level of macro-cyclicality, it analyzes the seasonality and periodicity of bidding behavior, such as checking for abnormally high 1-yuan price differences in certain months or quarters; at the meso level of meso-clustering, it analyzes the temporal clustering of bidding activities, using point process models to examine whether bidding events are uniformly distributed or clustered along the time axis, the latter potentially indicating coordinated action; at the micro level of micro-sequence dependence, it analyzes the time interval distribution of consecutive bidding events. Particularly important analyses include: detection of bidding time synchronicity, i.e., whether different suppliers submit bids sequentially within a very short period; temporal patterns of winning bids, such as whether supplier A and supplier B take turns winning bids with a 1-yuan price difference in consecutive bid segments; and temporal evolution analysis of behavioral patterns, observing dynamic changes in price difference patterns through sliding window techniques. A time series similarity matrix is ​​constructed to analyze the coordinated changes in bidding behavior of different suppliers over time, using the Dynamic Time Warping (DTW) algorithm to measure sequence similarity and identify suspicious combinations of highly synchronized behaviors.

[0124] In this optional embodiment, a Monte Carlo simulation model is constructed, and the Monte Carlo simulation model is used to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile, including:

[0125] S41. Construct a Monte Carlo simulation model based on overall frequency test, supplier pair analysis, conditional pattern test, and time series pattern analysis, combined with randomized counterfactual scenarios.

[0126] S42. A large-scale simulation was conducted using a randomization strategy that preserves the segment structure to construct a zero distribution of the univariate frequency of bid difference. The percentile ranking of the observed frequency, the effect size index, and the nonparametric confidence interval were calculated for statistical comparison and evaluation to obtain macroscopic test evidence.

[0127] S43. Screen supplier pairs with statistical significance and perform conditional randomization tests for pairing. Based on multiple comparison correction and risk assessment of fusion statistical significance, effect size and behavioral persistence, obtain combined-level test evidence.

[0128] S44. Conduct stratified Monte Carlo simulations and statistical significance assessments on basic business conditions, complex interaction conditions, and dynamic market conditions to obtain contextualized test evidence.

[0129] S45. Using Monte Carlo simulation, a stochastic benchmark of periodic, clustered, and sequence-dependent time patterns is constructed and statistically tested to obtain time series test evidence.

[0130] S46. Based on macroscopic verification evidence, combined verification evidence, contextualized verification evidence, and time-series verification evidence, cross-validation and evidence fusion are performed on the enhanced multidimensional feature engineering dataset to construct a complete profile of abnormal behavior.

[0131] In this optional embodiment, a stochastic benchmark of periodic, clustered, and sequence-dependent time patterns is constructed using Monte Carlo simulation and statistically tested to obtain time-series test evidence, including:

[0132] S451. Construct a time series model based on dynamic condition testing, and use the time series model to analyze the correlation between the univariate frequency of price difference and the market index;

[0133] S452. Using time series decomposition technology, the univariate event sequence of price difference is decomposed into trend component, periodic component and residual component. The periodic component is subjected to spectral analysis to identify significant periodicity. Monte Carlo simulation is used to generate event sequences with random time distribution. The difference in periodic intensity between the actual sequence and the random sequence is compared to evaluate the statistical significance of periodicity.

[0134] S453. Using point process analysis, the price difference univariate event is regarded as an event point on the time axis. The interval time distribution of the actual event sequence is calculated and compared with the random sequence generated by the homogeneous Poisson process.

[0135] S454. Use the L-function to analyze the degree of clustering across multiple time scales, and use Monte Carlo simulation to construct the confidence envelope of the K-function. Determine whether the actual K-function value exceeds the envelope to test for significant temporal clustering.

[0136] Specifically, such as Figure 4As shown, the Monte Carlo test's core analytical methods constitute the main framework of the entire violation detection system. This framework systematically evaluates the statistical significance, pattern specificity, and behavioral anomaly of the one-way price difference phenomenon by constructing randomized counterfactual scenarios. The following is a detailed explanation of the technical solutions for the four-dimensional testing methods: Overall Frequency Test, System-Level Anomaly Screening. The overall frequency test aims to assess the general anomaly degree of the one-way price difference phenomenon from a macroscopic perspective. Its technical implementation includes three key stages: Establishing a baseline randomized model. A randomization strategy that preserves the data structure is adopted. In each simulation, the number of suppliers in each bidding segment remains unchanged, and only the bid amount is randomly redistributed among all suppliers participating in that bidding segment. This intra-segment shuffling method ensures the comparability of the simulated scenario with the real scenario in terms of competitive structure. Conducting large-scale simulations, typically 10,000 times, recording the total number of occurrences of the one-way price difference in each simulation to construct a complete zero distribution. Performing statistical comparison and evaluation. Calculating the percentile ranking of the actual observed frequency in the simulated distribution to obtain the precise p-value. The effect size index is calculated, primarily using two core metrics: Standardized Mean Deviation (SD) calculated using Cohen's d method: d = (M_obs - M_sim) / SD_sim, where M_obs represents the observed one-dimensional frequency of the price difference, M_sim represents the average of the random frequencies obtained from Monte Carlo simulation, and SD_sim represents the standard deviation of the simulated frequency distribution. This index measures the degree to which the actual frequency deviates from the random expectation, expressed in standard deviations. In business interpretation, d = 0.2 indicates a small effect (slight anomaly), d = 0.5 indicates a moderate effect (significant anomaly), and d = 0.8 indicates a large effect (severe anomaly). For example, if d = 1.5 is calculated, it means that the actual one-dimensional frequency of the price difference is 1.5 standard deviations higher than the random expectation, which is an extremely strong anomalous signal.

[0137] The hazard ratio (RR) provides a more intuitive business interpretation; its formula is RR = M_obs / M_sim. This metric directly reflects how many times the actual frequency is greater than the random expectation. For example, RR = 3 means that the actual frequency of a one-dollar price difference is three times that under random conditions. The hazard ratio is particularly suitable for explaining the severity of a problem to non-technical personnel because it can be understood without requiring a statistical background. Calculate the confidence interval for the hazard ratio to assess the accuracy of this multiple estimate. Construct nonparametric confidence intervals and determine the reasonable random fluctuation range of the actual frequency using percentile methods. If the actual value falls outside the 95% confidence interval, especially above the upper limit, it indicates the presence of a systemic anomaly. Perform robustness analysis. Test the effects of different randomization strategies, including fully randomized rearrangement, stratified randomization, and guided resampling. Enhance the reliability of conclusions by examining the convergence of results from multiple methods. While the overall frequency test may not be sensitive enough to detect local anomalies, it can provide important diagnostics of the overall health of the procurement system, identifying widespread problems requiring systemic intervention.

[0138] Supplier pair analysis precisely identifies suspicious combinations. Supplier pair analysis focuses on identifying abnormal interaction patterns between specific supplier pairs. Its technical implementation includes four core steps: constructing a supplier pair network and screening. Extracting all supplier pairs that have jointly bid from the data, calculating the number of joint bids for each pair, the joint bid strength index (the geometric mean of the number of joint bids to their respective total bids), and the background success rate. Setting reasonable thresholds, such as ≥5 joint bids, ensures the analysis is statistically significant. Performing a pair-specific Monte Carlo test. For each eligible supplier pair, performing conditionally randomized simulations: randomly reassigning the bidding order of the two suppliers only for the bids they jointly participate in, while maintaining their bidding relationships with other suppliers. Through numerous simulations, typically 5000, constructing a univariate frequency zero distribution of the bid difference for that specific supplier pair. Multiple comparison correction and risk assessment. Simultaneously testing a large number of supplier pairs, with rigorous multiple comparison correction to control false positives. The False Discovery Rate (FDR) control method, using the classic Benjamini-Hochberg method, follows this process: Perform all tests and obtain p-values: For the N selected supplier pairs, perform Monte Carlo tests to obtain N independent p-values: p1, p2, ..., p no Sort the p-values: Sort these p-values ​​from smallest to largest: p （1） ≤p （2） ≤...≤p （n）oSet the FDR control level q: Determine an acceptable upper limit for the false detection rate, for example, q = 0.05. We are willing to accept that up to 5% of the cases in the final suspect list are false positives. Find the critical threshold k: Starting with the largest p-value, compare and find the largest index k that satisfies $p_{(k)}\leq\frac{k}{N}\timesq$. Make a decision: Reject the null hypothesis (i.e., classify as anomalous) for all supplier pairs whose corresponding p-values ​​satisfy $p_{(i)}\leqp_{(k)}$. Simultaneously, calculate the risk score for each supplier pair by weighted fusion of the following three dimensions: statistical significance (p-value), effect size (abnormality strength), and behavioral persistence (time span of the anomalous pattern). Based on the relative importance of each dimension in business judgment, the designed weight allocation is: statistical significance 40%, effect size 40%, and behavioral persistence 20%. The comprehensive scoring formula is: Risk Score = w1 × Significance Score + w2 × Effect Size Score + w3 × Persistence Score. The score was eventually normalized to a range of 0-1, with higher scores indicating higher risk.

[0139] Pattern Recognition and Classification. For high-risk supplier pairs identified in the risk assessment phase, further analysis of their behavioral patterns is conducted: is it a rigid cooperation maintaining a 1 yuan price difference, or a conditional cooperation only occurring within a specific price range? Is it a one-way master-slave relationship, such as supplier A always charging 1 yuan more than supplier B, or a two-way equal relationship, or sometimes supplier A charging 1 yuan more, sometimes supplier B charging 1 yuan more? These pattern classifications provide specific directions for subsequent investigations. Conditional Pattern Testing: Revealing hidden coordination logic, conditional pattern testing aims to discover anomalous behaviors activated under specific circumstances, identifying more covert and strategic forms of collusion. This testing unfolds through three levels: Basic condition testing focuses on simple but important business scenarios as hierarchical basic business conditions. Testing whether the 1 yuan price difference phenomenon is abnormally concentrated under specific conditions: for example, only occurring when the budget of a large bid exceeds a specific threshold, only occurring in bids managed by a specific procurement officer, or only occurring in bids in a certain region. Performing stratified Monte Carlo simulation: Selecting a subset of bids that meet the conditions from the data, randomizing the prices within each bid, and constructing a condition-specific zero distribution. Compare the differences in the degree of anomalousness between conditional and unconditional subsets. Complex interaction condition testing analyzes multi-factor combination scenarios as complex interaction conditions. Use conditional decision trees or rule discovery algorithms to identify the set of conditions satisfied by supplier combinations with abnormally frequent 1-yuan price differences. For example, it was found that when suppliers A, B, and C bid simultaneously, the bid type is a long-term agreement, and supplier D is absent, the probability of suppliers A and B having a 1-yuan price difference is abnormally high. Multi-condition combinations often reveal carefully designed avoidance strategies. Statistical significance is assessed by comparing the differences between scenarios satisfying conditional combinations and random combinations. Dynamic condition testing focuses on conditional behavior based on market conditions, i.e., dynamic market conditions. Analyze whether 1-yuan price differences are more frequent during periods of market tension, such as peak winter coal consumption or periods of high price volatility. Build a time series model to correlate the frequency of 1-yuan price differences with market indices, such as coal price indices and inventory levels, and test whether this correlation exceeds random expectations using Monte Carlo simulation. Dynamic condition testing can detect coordinated behavior by suppliers using market conditions as cover. The frequency of 1-yuan price differences is denoted as... A simplified example of a time series model is shown below:

[0140] ;

[0141] in, This represents the autoregressive term, used to capture the inherent temporal inertia of the phenomenon. β and γ represent weighting parameters set based on business experience. Indicates a price index. Indicates inventory. Indicates control variables, This represents the random error term. Time-series pattern analysis: Capturing the coordinated temporal characteristics, time-series pattern analysis reveals the non-random characteristics of the univariate price difference phenomenon from a time dimension, focusing on three types of time anomaly patterns: Periodicity pattern detection analyzes whether the univariate price difference occurs regularly within a specific time period. Using time-series decomposition techniques, the univariate price difference event sequence is decomposed into trend components, periodic components, and residual components. Spectral analysis is performed on the periodic components, using Fourier transform or wavelet analysis to identify significant periods, such as monthly or quarterly periods. Event sequences with random time distributions are generated through Monte Carlo simulation, and the difference in periodic intensity between the actual sequence and the random sequence is compared to assess whether the detected periodicity is statistically significant. Clustering pattern detection analyzes whether the univariate price difference is abnormally concentrated in time. Using point process analysis, each univariate price difference is considered an event point on the time axis. The interval time distribution of the actual event sequence is calculated and compared with a sequence generated completely randomly by a homogeneous Poisson process. L-function analysis is used to analyze the degree of clustering of events across multiple time scales. This second-order analysis method, based on point process theory, is used to quantitatively examine whether the univariate event of price difference is uniformly random distributed along the time axis or exhibits a significant clustering or dispersion pattern within specific time intervals. Monte Carlo simulation is used to construct the confidence envelope of the K-function. If the actual K-function value exceeds the envelope, it indicates the existence of significant temporal clustering. Such clustering may suggest that suppliers are acting in concentrated bursts during specific periods.

[0142] Sequence dependency detection analyzes the temporal correlation and sequential patterns among events with a one-yuan bid difference. It examines whether consecutive bid segments with a one-yuan bid difference occur too close in time and whether the winning bids exhibit a regular rotation pattern. For example, it analyzes whether the winning bid rotation sequence from supplier A to supplier B to supplier A and then to supplier B is significantly more frequent than expected randomly. By constructing a Markov chain model, defining the states as {supplier A wins, supplier B wins, others}, it focuses on the following three patterns: whether the transition probabilities from supplier A to supplier B and from supplier B to supplier A are significantly higher than random; whether the three-step transitions from supplier A to supplier B to supplier A, or from supplier B to supplier A and then to supplier B, are abnormally frequent; and whether the transition pattern changes after a one-yuan bid difference occurs. It compares the difference between the actual state transition probabilities and the random transition probabilities and uses permutation tests to assess significance. Sequence dependency detection can uncover subtle temporal coordination mechanisms among suppliers. These four dimensions of testing are not conducted in isolation but rather are mutually corroborated and work synergistically through an integrated framework. Overall frequency testing identifies systemic risks, supplier analysis pinpoints responsible parties, conditional pattern testing reveals behavioral triggering mechanisms, and time-series pattern analysis captures the temporal characteristics of coordination. By integrating these four pieces of evidence, a complete profile of abnormal behavior can be constructed, providing multi-level and multi-dimensional scientific basis for subsequent risk assessment and investigation decisions. This comprehensive testing system significantly enhances the ability to detect complex and covert collusion, while cross-validation through multiple pieces of evidence strengthens the credibility of the results.

[0143] In this optional embodiment, the abnormal behavior profile is identified and filtered to construct a multi-level data view. The Monte Carlo simulation model is used to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. The randomization scheme is then evaluated and divided to obtain a quantitative comprehensive risk score, including:

[0144] S51. Use an adaptive threshold algorithm to analyze the global distribution of the number of joint bids in the abnormal behavior profile of suppliers to determine the quantile threshold. Combine the joint bid intensity, time coverage, price correlation, bid complementarity and multi-dimensional screening criteria to obtain the set of supplier pairs that meet the conditions.

[0145] S52. Based on the supplier's basic information on the jointly participated bid section, details of both parties' quotations, information on the competitive environment and business background, and the joint bidding event sequence, multi-dimensional features of each event point and time interval sequence arranged in chronological order, the information is input into the context data of the performance of both parties in other bid sections individually, to obtain a multi-level data view;

[0146] S53. Using the constrained randomization in the Monte Carlo simulation model, randomization tests are performed on the multi-level data view and a zero distribution is constructed to obtain the randomization scheme for the supplier conditions.

[0147] S54. Compare the actual observed price difference univariate frequency with the simulated zero distribution, calculate the p-value, effect size, relative risk ratio and perform confidence interval analysis, use the stratified false discovery rate control method to perform multiple comparison correction, and combine sensitivity analysis to test the impact of randomization scheme on the results to obtain the statistical significance assessment results.

[0148] S55. Based on the statistical significance assessment results, integrate multi-dimensional indicators such as behavioral pattern risk, time consistency, and network relationship risk, use a weighted formula to calculate a quantitative comprehensive risk score, and classify supplier pairs into high, medium, and low risk levels according to preset thresholds to obtain a quantitative comprehensive risk score.

[0149] Specifically, the analysis involves in-depth supplier pair analysis: Supplier pair identification and screening mechanisms utilize adaptive threshold algorithms instead of fixed rules. The global distribution of joint bidding frequency across all supplier pairs is analyzed to determine reasonable quantile thresholds. For example, supplier pairs ranking in the top 15% for joint bidding frequency are included in the analysis, rather than using a fixed minimum number. Multi-dimensional screening criteria ensure analysis quality: Joint Bidding Strength: The proportion of joint bidding frequency to each supplier's total bidding frequency is calculated to identify closely related pairs. Time Coverage: The joint bidding time span must cover at least one complete procurement cycle, such as 6 months, excluding temporary pairs. Price Correlation: The correlation coefficient of historical bids from two suppliers is calculated to identify pairs with highly correlated pricing behavior. Winning Bid Complementarity: The analysis examines whether a pattern emerges where one supplier's winning bid inevitably leads to the other's participation. Supplier pair data extraction and feature engineering construct multi-level data views for each supplier pair. Bid Segment Level Data Extraction includes: Basic information of the jointly participated bid segments: Bid Segment ID, Release Time, Deadline, Budget Amount, and Procurement Type. Bid Details: Precise bid amount, bid time, and whether the bid was successful. Competitive Environment Information: Number of other participating suppliers and their bids for this tender section. Business Background Information: Project to which this tender section belongs, procurement department, and evaluation method.

[0150] The time-series data is organized into a structured format, including the following aspects: a sequence of joint bidding events arranged chronologically; multidimensional features for each event point: price difference, relative ranking, and bidding result; a time interval sequence: the distribution of time differences in consecutive joint bids; and contextual data, including: the performance of each party in other bidding sections individually, a comparative analysis of their historical bidding success rates, and their bidding behavior patterns in non-joint bidding sections. Enhanced feature engineering is performed. Based on the extracted raw data, a deep feature set is constructed. Price difference feature family: absolute difference sequence: the price difference in each joint bid; relative difference: the ratio of price difference to the section budget; ranking difference: the difference in ranking between the two parties' bids in the current bid; standardized difference: the difference adjusted to consider the intensity of competition in the section; behavioral pattern feature family: bidding order pattern: the time regularity of who bids first and who bids last; price adjustment pattern: the adjustment range of one party's bid relative to historical bids; response pattern: the response strategy to the other party's bid, such as always 1 yuan higher or always 1 yuan lower; fluctuation synchronicity: the time correlation of the fluctuations in the bids of both parties. Relationship dynamics feature family, relationship strength change: time trend of joint bidding frequency. Behavioral convergence: whether the bidding strategies of both parties converge over time. Dependency asymmetry: analyzing whether there is a pattern where one party always follows the other. Supplier pair level Monte Carlo test, the core test adopts a customized Monte Carlo simulation method. Randomization scheme for each supplier pair maintaining design conditions, including bid segment structure randomization, hierarchical randomization, and constraint randomization. Bid segment structure maintenance randomization: in each simulation, the set of all suppliers for each bid segment remains unchanged, and the bid allocation is only randomly determined within that supplier pair. Specifically, for each bid segment with joint bidding, the bid amounts of both parties are exchanged with a 50% probability, and the bids of other suppliers remain unchanged. This randomization maintains the competitive structure and bid distribution of the bid segment, only destroying the specific association that may exist between the supplier pairs. Considering that different bid segment types may have different bidding strategies, hierarchical randomization is implemented: hierarchical by bid segment budget size: large bid segments and small bid segments are randomized separately. Hierarchical by procurement type: different types such as open tenders and invited tenders are handled separately. Stratified by time period: Bidding segments for different years or quarters are randomized separately. Constraints on randomization to address business specificities: Maintaining historical success rates: The randomized bid allocation should ensure that the success rates for both parties are close to historical levels. Maintaining bid distribution characteristics: The mean and variance of the bid distributions for both parties should remain essentially unchanged after randomization. Maintaining time pattern constraints: Avoiding the generation of bid time series that are practically impossible. Next, execute an efficient parallel simulation process: Dynamically determine the required number of simulations based on preliminary analysis results. Conduct 1000 rapid simulations to evaluate the stability of the results.If the p-value is near the boundary, such as 0.04-0.06, increase the number of simulations to 10,000 to obtain a more accurate estimate; if the p-value is significantly small, less than 0.01 or large, greater than 0.1, then 1,000 simulations are sufficient.

[0151] After constructing a complete null distribution, not only are basic statistics calculated, but its distribution characteristics are also analyzed, including distribution shape, extreme values, and distribution fit. Sensitivity analysis tests the impact of randomization scheme selection on the results: comparing the consistency of p-values ​​under different randomization strategies. The sensitivity of the results to changes in the random seed is analyzed. The stability of the results is tested after the number of simulations increases to a certain threshold. Through thousands of such simulations, the system can depict the normal fluctuation range of the number of times a bid difference occurs between suppliers in a purely random situation, i.e., the statistical null distribution. The statistical significance evaluation system performs a rigorous statistical significance evaluation by comparing the actual observed frequency of bid differences with the simulated null distribution. This not only calculates precise p-values ​​to measure the extremes of the observed results but also integrates multiple indicators such as effect size and relative risk ratio, and quantifies the uncertainty of the estimate through confidence interval analysis. Considering the multiple comparison problems that may arise from simultaneously testing a large number of supplier pairs, the system adopts a stratified false discovery rate control method to ensure the reliability of the overall conclusions. All this statistical evidence is ultimately integrated into a quantitative comprehensive risk score: Risk Score = w1 * Statistical Significance Transformation + w2 * Behavioral Pattern Risk Transformation + w3 * Temporal Consistency + w4 * Network Relationship Risk Transformation. Risk classification and prioritization: High risk: Score > 0.7, immediate investigation recommended. Medium risk: Score 0.4-0.7, enhanced monitoring and regular review recommended. Low risk: Score < 0.4, routine monitoring sufficient. The system categorizes supplier pairs into different risk levels based on the risk score, providing clear and scientific action guidelines for subsequent priority investigations and precise monitoring.

[0152] In this optional embodiment, a multi-dimensional risk assessment and intelligent identification terminal is used to assess and identify the quantitative comprehensive risk score, and the assessment and identification results are integrated and processed to obtain an integrity comprehensive risk score. Based on a preset threshold and the integrity comprehensive risk score, suppliers are classified into different levels, and a violation adjustment report is generated to achieve supplier violation identification, including:

[0153] S61. Using a multi-dimensional risk assessment and intelligent identification terminal, the quantitative comprehensive risk score and the corresponding supplier-related data are assessed and identified to obtain the assessment input dataset.

[0154] S62. Based on the terminal, a multi-dimensional evaluation framework is used to evaluate the statistical significance dimension. Combined with indicators such as p-value, effect size, and confidence interval width obtained from the Monte Carlo test, the p-value is mapped to a statistical risk score through a transformation function to obtain the statistical risk score result.

[0155] S63. Use unsupervised anomaly detection algorithms to identify unconventional patterns in price synchronicity, price amount last digit regularity, and price adjustment consistency, quantify them as behavioral anomaly indices, and obtain behavioral pattern risk scores.

[0156] S64. Use time series analysis and change point detection techniques to identify the starting point, duration and evolution trend of anomalous patterns, assess the stability and avoidance behavior of anomalous patterns, and obtain time consistency risk scores.

[0157] S65. Calculate node centrality and community structure indicators based on supplier relationship network, and evaluate supplier association strength and connection relationship of high-risk suppliers to obtain network relationship risk score results;

[0158] S66. Use a weighted fusion model to integrate statistical risk scoring results, behavioral pattern risk scoring results, time consistency risk scoring results, and network relationship risk scoring results, and set dynamic weights according to business importance to obtain a comprehensive integrity risk score.

[0159] S67. Based on preset thresholds, suppliers or supplier pairs are divided into three risk levels: high, medium, and low, and adjustment reports are generated. Based on the violation adjustment reports, the results of supplier violation identification are achieved.

[0160] Specifically, the multi-dimensional risk assessment and intelligent identification terminal: The risk assessment and intelligent identification system is a multi-dimensional, multi-level analytical framework that transforms statistical findings into actionable business insights. The system assesses from a statistical significance perspective, considering not only the p-value obtained from the Monte Carlo test but also multiple statistical indicators such as effect size and confidence interval width to comprehensively measure the credibility and strength of the deviation of the univariate phenomenon of price difference from randomness. The statistical risk score does not simply rely on a p-value threshold but uses a transformation function to map the p-value to a risk score. Extremely low p-values ​​(less than 0.001) correspond to high risk, moderate p-values ​​to medium risk, and p-values ​​above the significance level are considered low risk. This transformation process also considers test power and sample size to avoid misjudgments due to insufficient data. At the behavioral pattern risk assessment level, the system deeply analyzes other abnormal characteristics in pricing behavior besides price difference, including the synchronicity of pricing time, the pattern of the last digit of the pricing amount, and the consistency of the magnitude and direction of pricing adjustments. For example, if two suppliers not only frequently have price differences of one yuan, but also always submit their quotes at almost the same time, or if the last digit of their quotes shows a fixed pattern, these collaborative behaviors will significantly increase their behavioral risk scores. The system identifies these unusual patterns through an unsupervised anomaly detection algorithm and quantifies them into a behavioral anomaly index, which is corroborated with the statistical risk score. The time consistency risk assessment focuses on examining the stability and evolution trend of the anomaly pattern over time.

[0161] The system divides the entire observation period into multiple time windows, analyzing the frequency and pattern of a one-yuan bid difference within each window to determine if they persist. Consistent, long-term discrepancies are more risky than occasional anomalies. Through time series analysis and change point detection techniques, the system can identify the starting point, duration, and potential termination point of abnormal patterns, thus determining whether the behavior is a long-term coordination strategy or a short-term experiment. Furthermore, the system assesses whether abnormal patterns temporarily disappear during specific periods and then reappear later, such as during audits; such avoidance behavior itself is a high-risk signal. Network relationship risk assessment places suppliers within a broader interactive network for analysis. Using a supplier relationship network constructed from bidding data, the system calculates network centrality indicators for each supplier, such as degree centrality, betweenness centrality, and eigenvector centrality. Suppliers at the network center tend to have stronger coordination capabilities and influence, thus posing a higher risk. Simultaneously, a community detection algorithm is used to identify tightly connected supplier groups within the network. If a one-yuan bid difference frequently occurs within a group but rarely between groups, it suggests potential collusion. Network risk assessment also evaluates the strength of historical connections between suppliers, such as the frequency and stability of their joint bidding, and whether they have strong connections with other known high-risk suppliers. The system integrates the risk scores from these four dimensions into a single comprehensive risk score using a weighted fusion model. The weighting is based on business importance and can be dynamically adjusted according to different procurement scenarios. After comprehensive scoring, the system classifies suppliers or supplier pairs into high, medium, and low risk levels based on preset thresholds. High-risk levels trigger immediate alerts, recommending an immediate in-depth investigation and potentially implementing temporary measures such as suspending bidding eligibility; medium-risk levels indicate the need for enhanced monitoring and regular review; and low-risk levels are included in routine monitoring.

[0162] Specifically, this invention uses the Monte Carlo method to construct counterfactual scenarios by simulating random processes, thereby assessing whether observed patterns are random. Specifically, by randomly rearranging bid data while maintaining the bid segment structure and bid distribution, the number of times a 1-yuan price difference occurs under random conditions is calculated, thus obtaining the p-value of the observed value. It does not require assuming a theoretical distribution; it uses actual data to construct an empirical distribution through resampling. Therefore, the Monte Carlo method is insensitive to distribution assumptions. It has relatively flexible requirements for sample size; simulations can be performed even with a small sample size, but an excessively small sample size can lead to excessive randomness in the simulation and unstable results. The Monte Carlo method does not require merging intervals, thus avoiding information loss. Simulation experiments can be designed to maintain the bid segment structure, supplier participation patterns, etc., thereby enabling the detection of more complex patterns. For example, by keeping the number of suppliers in each bid segment constant in the simulation, the probability of a 1-yuan price difference occurring under randomly assigned bids can be tested. The Monte Carlo method can also be further extended, for example, considering specific relationships between supplier pairs, time series patterns, etc., by designing more complex simulation schemes to detect these patterns. Using observational data directly, extreme values ​​are used as a baseline for random rearrangement in the simulation. Therefore, the probability of observed extreme values ​​appearing in the simulation directly affects the p-value. If extreme values ​​occur frequently, they are less likely to appear in the random simulation, resulting in a smaller p-value. Thus, the Monte Carlo method is more sensitive to extreme values.

[0163] Specifically, the core value of this invention lies in providing a scientific and practical detection framework. Compared with traditional statistical methods, this technology analyzes data entirely based on its inherent characteristics, without requiring prior assumptions about the distribution patterns of price differences. This fundamentally avoids the risk of misjudgment due to incorrect assumptions. In the specific scenario of coal procurement, the price formation mechanism is complex, influenced by multiple factors such as coal quality differences, transportation costs, and regional markets. Price data often exhibits non-standard distribution characteristics. The Monte Carlo method, by learning the true distribution pattern from historical data, can naturally adapt to these business characteristics, ensuring the reliability of the detection results. Its unique feature lies in its high sensitivity to complex violation patterns. It can detect simple frequency anomalies and identify hidden violations that are difficult to detect using traditional methods, such as conditional collusion between specific supplier combinations, time-based rotating bidding patterns, or coordinated manipulation by multiple supplier groups. By constructing randomized counterfactual scenarios, this method accurately answers the crucial question of "how likely is this phenomenon to occur by chance," providing intuitive and convincing evidence for regulatory decisions. When a pair of suppliers differs by one yuan in multiple joint bids, this invention demonstrates that the probability of such a situation occurring in thousands of randomized experiments is extremely low, proving that this pattern is not random. In practice, this invention can flexibly handle common data quality issues in real-world business, such as insufficient sample size and incomplete data. Even if some suppliers have a limited number of joint bids, it can still provide meaningful risk assessments. This method supports multi-dimensional comprehensive analysis, organically integrating factors such as price differences, supplier relationship networks, and time series patterns to generate quantitative risk scores. This allows regulatory resources to be prioritized for the highest-risk cases, improving investigation efficiency. This quantitative assessment provides a scientific basis for management decisions and helps establish a precise risk classification and control mechanism.

[0164] Furthermore, this invention boasts excellent operability and sustainability. Utilizing parallel computing technology, it can handle the computational demands of large-scale data simulations, ensuring analytical efficiency. The system possesses self-optimization capabilities; with data accumulation and investigation feedback, it can continuously adjust and improve the detection model, forming a virtuous cycle of continuous improvement. From an organizational value perspective, the establishment of this scientific detection system enhances the accuracy of violation detection and creates a strong deterrent against potential violations, contributing to a fair and competitive procurement environment and promoting the healthy development of the supply chain ecosystem. This invention provides a complete, scientific, and efficient technical solution for coal procurement supervision, ensuring rigor at the statistical level and demonstrating practicality at the business level, making it an ideal choice for achieving intelligent risk identification.

[0165] like Figure 5 As shown, according to another embodiment of the present invention, a supplier violation intelligent identification system based on the Monte Carlo method is also provided, comprising:

[0166] Bid data preprocessing module 1 is used to acquire the original bid record dataset and preprocess it to obtain the analytical bid record dataset.

[0167] Supplier feature quantification and extraction module 2 is used to extract multidimensional features from the analysis bidding record dataset using a multidimensional analysis feature mechanism, and to quantify the extracted multidimensional feature results using a supplier relationship network to obtain the initial feature engineering dataset.

[0168] The multi-granularity temporal feature fusion module 3 is used to perform temporal pattern analysis on the bidding record dataset using a multi-granularity time analysis mechanism, and to fuse the temporal distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset.

[0169] Monte Carlo simulation model analysis module 4 is used to construct a Monte Carlo simulation model, and to use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile.

[0170] The supplier risk simulation assessment module 5 is used to identify and screen abnormal behavior profiles and construct multi-level data views. The Monte Carlo simulation model is used to test and simulate the multi-level data views to obtain randomization schemes for supplier conditions. The randomization schemes are then evaluated and divided to obtain a quantitative comprehensive risk score.

[0171] The multidimensional risk level determination module 6 is used to evaluate and identify the quantitative comprehensive risk score using a multidimensional risk assessment and intelligent identification terminal, and to integrate and process the evaluation and identification results to obtain a comprehensive integrity risk score. Based on the preset threshold and the comprehensive integrity risk score, the module classifies suppliers into different levels and generates a violation adjustment report to achieve the identification of supplier violations.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of supplier violations based on the Monte Carlo method, characterized in that, include: S1. Obtain the original bid record dataset and preprocess it to obtain the analytical bid record dataset; S2. Multidimensional feature extraction is performed on the bidding record dataset using a multidimensional analysis feature mechanism, and supplier feature quantification is performed on the extracted multidimensional feature results using a supplier relationship network to obtain the initial feature engineering dataset. S3. Utilize a multi-granularity time analysis mechanism to perform time series pattern analysis on the bidding record dataset, and fuse the time series distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset. S4. Construct a Monte Carlo simulation model, and use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile; S5. Identify and filter the abnormal behavior profiles and construct a multi-level data view. Use the Monte Carlo simulation model to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. The randomization schemes are evaluated and classified to obtain a quantitative comprehensive risk score; S6. Utilize multi-dimensional risk assessment and intelligent identification terminals to evaluate and identify quantitative comprehensive risk scores, integrate and process the evaluation and identification results to obtain a comprehensive integrity risk score, classify suppliers based on preset thresholds and comprehensive integrity risk scores, and generate violation adjustment reports to achieve supplier violation identification.

2. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The process of obtaining the original bid record dataset and preprocessing it to obtain the bid record dataset for analysis includes: S11. Generate the original bid record dataset based on the bid identifier, supplier identifier, bid amount, and bid time; S12. Use a hierarchical cleaning strategy to process the missing values ​​in the original bidding record dataset, mark invalid data with missing values ​​in any field, and record the reasons and number of deletions to obtain a preliminary cleaned dataset. S13. Use a multi-level anomaly detection method to detect the quotation amount field in the preliminary cleaned dataset. Based on the seasonal fluctuation characteristics of coal prices and the improved box plot method, identify the detection results and dynamically adjust the threshold. Use the median instead of the mean and the median absolute deviation instead of the standard deviation for the adjusted Z score. Mark the observations with an absolute value greater than the preset threshold as outliers. S14. Using the edit distance and fuzzy matching algorithm of the business knowledge base, entity normalization is performed on the supplier name inconsistency problem to obtain intermediate data of bidding records with unified supplier identification. S15. Integrate outlier data, data quality reports, and intermediate data from bidding records into a star schema data warehouse to form an analytical bidding record dataset.

3. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The method involves using a multidimensional analysis feature mechanism to extract multidimensional features from the bid record dataset, and then using a supplier relationship network to quantify the extracted multidimensional features, resulting in an initial feature engineering dataset including: S21. Construct a multi-dimensional analysis feature mechanism based on price spread characteristics, relationship characteristics, time series characteristics, and derived characteristics; S22. Using a multidimensional analysis feature mechanism, feature extraction is performed on the analysis bidding record dataset, and the feature extraction results are transformed into information carriers of collusion behavior. Based on the information carriers, a multidimensional feature set of individual supplier behaviors is generated. S23. Construct a time-weighted supplier relationship network based on joint bidding records, and use a community detection algorithm to identify potential related groups in the supplier relationship network to obtain community structure characteristics; S24. Utilize supplier relationship networks and community structure features to perform network correlation quantization and enhancement fusion on the multidimensional feature set to obtain the initial feature engineering dataset.

4. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 3, characterized in that, The supplier relationship network, constructed based on time-weighted joint bidding records, and the potential associated groups within the supplier relationship network using a community detection algorithm, yields the following community structure characteristics: S231. Use the community discovery algorithm to run the consensus strategy multiple times, and run it independently a preset number of times with different random seeds. S232. Based on the community partitioning results of the preset number of independent runs, the frequency of common communities among supplier nodes is counted, a consensus matrix of the supplier relationship network is constructed, and a consensus matrix dataset is obtained. S233. Based on the consensus matrix dataset, input it again into the community structure detection to eliminate the randomness bias of a single algorithm run and obtain the supplier community division result; S234. Based on the supplier community segmentation results, extract supplier nodes with high intermediary centrality to obtain the competitive health of the supplier network.

5. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The process involves using a multi-granularity time analysis mechanism to perform time-series pattern analysis on the bid record dataset, and then fusing the time-series distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset, which includes: S31. Construct a multi-granularity time analysis mechanism based on macro-periodicity analysis, meso-aggregation analysis and micro-sequence dependence analysis; S32. Utilize a multi-granularity time analysis mechanism to perform macro-periodic analysis on the bidding record dataset, identify abnormal patterns in bidding behavior on seasonal and cyclical time scales, and obtain macro-periodic characteristics. S33. Use a multi-granularity time analysis mechanism to perform meso-level clustering analysis on the bidding record dataset, and test the distribution pattern on the time axis of the bidding events based on the point process model to obtain the time clustering characteristics. S34. Using a multi-granularity time analysis mechanism, perform micro-sequence dependency analysis on the bidding record dataset to analyze the time interval of consecutive bidding events, the synchronicity of the bid sequence, and the winning bid rotation pattern to obtain behavioral sequence characteristics; S35. Integrate and quantify the macro-periodic characteristics, temporal clustering characteristics, and behavioral sequence characteristics to obtain the time series distribution analysis results; S36. The temporal distribution analysis results are fused with the initial feature engineering dataset to obtain an enhanced multidimensional feature engineering dataset.

6. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The construction of the Monte Carlo simulation model, and the use of the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset, yields an abnormal behavior profile, including: S41. Construct a Monte Carlo simulation model based on overall frequency test, supplier pair analysis, conditional pattern test, and time series pattern analysis, combined with randomized counterfactual scenarios. S42. A large-scale simulation was conducted using a randomization strategy that preserves the segment structure to construct a zero distribution of the univariate frequency of bid difference. The percentile ranking of the observed frequency, the effect size index, and the nonparametric confidence interval were calculated for statistical comparison and evaluation to obtain macroscopic test evidence. S43. Screen supplier pairs with statistical significance and perform conditional randomization tests for pairing. Based on multiple comparison correction and risk assessment of fusion statistical significance, effect size and behavioral persistence, obtain combined-level test evidence. S44. Conduct stratified Monte Carlo simulations and statistical significance assessments on basic business conditions, complex interaction conditions, and dynamic market conditions to obtain contextualized test evidence. S45. Using Monte Carlo simulation, a stochastic benchmark of periodic, clustered, and sequence-dependent time patterns is constructed and statistically tested to obtain time series test evidence. S46. Based on macroscopic verification evidence, combined verification evidence, contextualized verification evidence, and time-series verification evidence, cross-validation and evidence fusion are performed on the enhanced multidimensional feature engineering dataset to construct a complete profile of abnormal behavior.

7. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 6, characterized in that, The method of constructing a stochastic benchmark with periodic, clustered, and sequence-dependent time patterns using Monte Carlo simulation and performing statistical tests yields time-series test evidence including: S451. Construct a time series model based on dynamic condition testing, and use the time series model to analyze the correlation between the univariate frequency of price difference and the market index; S452. Using time series decomposition technology, the univariate event sequence of price difference is decomposed into trend component, periodic component and residual component. The periodic component is subjected to spectral analysis to identify significant periodicity. Monte Carlo simulation is used to generate event sequences with random time distribution. The difference in periodic intensity between the actual sequence and the random sequence is compared to evaluate the statistical significance of periodicity. S453. Using point process analysis, the price difference univariate event is regarded as an event point on the time axis. The interval time distribution of the actual event sequence is calculated and compared with the random sequence generated by the homogeneous Poisson process. S454. Use the L-function to analyze the degree of clustering across multiple time scales, and use Monte Carlo simulation to construct the confidence envelope of the K-function. Determine whether the actual K-function value exceeds the envelope to test for significant temporal clustering.

8. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The process involves identifying and filtering abnormal behavior profiles and constructing a multi-level data view. The Monte Carlo simulation model is then used to test and simulate the multi-level data view to obtain a randomization scheme for supplier conditions. The randomization schemes are evaluated and categorized to obtain a quantitative comprehensive risk score, including: S51. Use an adaptive threshold algorithm to analyze the global distribution of the number of joint bids in the abnormal behavior profile of suppliers to determine the quantile threshold. Combine the joint bid intensity, time coverage, price correlation, bid complementarity and multi-dimensional screening criteria to obtain the set of supplier pairs that meet the conditions. S52. Based on the supplier's basic information on the jointly participated bid section, details of both parties' quotations, information on the competitive environment and business background, and the joint bidding event sequence, multi-dimensional features of each event point and time interval sequence arranged in chronological order, the information is input into the context data of the performance of both parties in other bid sections individually, to obtain a multi-level data view; S53. Using the constrained randomization in the Monte Carlo simulation model, randomization tests are performed on the multi-level data view and a zero distribution is constructed to obtain the randomization scheme for the supplier conditions. S54. Compare the actual observed price difference univariate frequency with the simulated zero distribution, calculate the p-value, effect size, relative risk ratio and perform confidence interval analysis, use the stratified false discovery rate control method to perform multiple comparison correction, and combine sensitivity analysis to test the impact of randomization scheme on the results to obtain the statistical significance assessment results. S55. Based on the statistical significance assessment results, integrate multi-dimensional indicators such as behavioral pattern risk, time consistency, and network relationship risk, use a weighted formula to calculate a quantitative comprehensive risk score, and classify supplier pairs into high, medium, and low risk levels according to preset thresholds to obtain a quantitative comprehensive risk score.

9. The intelligent identification method for supplier violations based on the Monte Carlo method according to claim 1, characterized in that, The process involves using a multi-dimensional risk assessment and intelligent identification terminal to evaluate and identify a quantitative comprehensive risk score, integrating the evaluation and identification results to obtain a comprehensive integrity risk score, classifying suppliers based on preset thresholds and the comprehensive integrity risk score, and generating a violation adjustment report to achieve supplier violation identification. S61. Using a multi-dimensional risk assessment and intelligent identification terminal, the quantitative comprehensive risk score and the corresponding supplier-related data are assessed and identified to obtain the assessment input dataset. S62. Based on the terminal, a multi-dimensional evaluation framework is used to evaluate the statistical significance dimension. Combined with indicators such as p-value, effect size, and confidence interval width obtained from the Monte Carlo test, the p-value is mapped to a statistical risk score through a transformation function to obtain the statistical risk score result. S63. Use unsupervised anomaly detection algorithms to identify unconventional patterns in price synchronicity, price amount last digit regularity, and price adjustment consistency, quantify them as behavioral anomaly indices, and obtain behavioral pattern risk scores. S64. Use time series analysis and change point detection techniques to identify the starting point, duration and evolution trend of anomalous patterns, assess the stability and avoidance behavior of anomalous patterns, and obtain time consistency risk scores. S65. Calculate node centrality and community structure indicators based on supplier relationship network, and evaluate supplier association strength and connection relationship of high-risk suppliers to obtain network relationship risk score results; S66. Use a weighted fusion model to integrate statistical risk scoring results, behavioral pattern risk scoring results, time consistency risk scoring results, and network relationship risk scoring results, and set dynamic weights according to business importance to obtain a comprehensive integrity risk score. S67. Based on preset thresholds, suppliers or supplier pairs are divided into three risk levels: high, medium, and low, and adjustment reports are generated. Based on the violation adjustment reports, the results of supplier violation identification are achieved.

10. A supplier violation intelligent identification system based on the Monte Carlo method, used to implement the supplier violation intelligent identification method based on the Monte Carlo method as described in any one of claims 1-9, characterized in that, include: The bid data preprocessing module is used to acquire the original bid record dataset and preprocess it to obtain the analytical bid record dataset; The supplier feature quantification and extraction module is used to extract multidimensional features from the analysis bidding record dataset using a multidimensional analysis feature mechanism, and to quantify the extracted multidimensional feature results using a supplier relationship network to obtain the initial feature engineering dataset. The multi-granularity temporal feature fusion module is used to perform temporal pattern analysis on the bidding record dataset using a multi-granularity time analysis mechanism, and to fuse the temporal distribution analysis results with the initial feature engineering dataset to obtain an enhanced multi-dimensional feature engineering dataset. The Monte Carlo simulation model analysis module is used to construct a Monte Carlo simulation model, and then use the Monte Carlo simulation model to identify and construct an enhanced multidimensional feature engineering dataset to obtain an abnormal behavior profile. The supplier risk simulation assessment module is used to identify and screen abnormal behavior profiles and construct multi-level data views. The Monte Carlo simulation model is used to test and simulate the multi-level data views to obtain a randomization scheme for supplier conditions. The randomization schemes are evaluated and classified to obtain a quantitative comprehensive risk score; The multidimensional risk level determination module is used to evaluate and identify the quantitative comprehensive risk score using a multidimensional risk assessment and intelligent identification terminal, and to integrate and process the evaluation and identification results to obtain the integrity comprehensive risk score. Based on the preset threshold and the integrity comprehensive risk score, the module classifies suppliers into different levels and generates a violation adjustment report to achieve the identification of supplier violations.

Images