A Public Resource Transaction Risk Early Warning Method Based on Big Data Analysis and Engineering Consulting

By combining graph neural networks and support vector machines with the results of engineering consulting verification, the problem of identifying hidden related behaviors of trading entities in public resource transactions has been solved, achieving stability and consistency in risk assessment and improving the timeliness and accuracy of supervision.

CN123118253APending Publication Date: 2026-11-17ZHENGZHOU SEARUN INFO TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611235056.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing public resource transaction risk warning technologies are insufficient to identify hidden related behaviors of transaction entities in multi-section continuous bidding scenarios. Risk assessment results are unstable and fail to be effectively combined with engineering consulting verification conclusions, leading to delays in regulatory intervention.

Method used

By analyzing the relationships between transaction entities in different batches of announcements using graph neural networks, determining risk levels using support vector machines, and combining this with engineering consulting verification results, a stable risk warning mechanism is formed, enabling a structured expression of the related behaviors and process states of transaction entities.

Benefits of technology

It improves the ability to identify repeated participation by trading entities, reduces the impact of fluctuations in single indicators for risk level determination, enhances the stability and consistency of risk warning results, and provides clear verification basis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118253A_ABST
    Figure CN123118253A_ABST
Patent Text Reader

Abstract

This invention relates to the field of transaction risk early warning technology, specifically a public resource transaction risk early warning method based on big data analysis and engineering consulting. It uses a graph neural network to jointly characterize the frequency and co-occurrence relationships of subject pairs across different announcement batches, enabling repeated participation behaviors among transaction subjects to form stable associations across batches. This integrates the originally dispersed subject behaviors across multiple transaction units into a continuous association chain, enhancing the overall ability to identify hidden association behaviors. Support vector machines are used to determine the level boundaries based on the proportion of covered bidding subjects, the time interval for bid evaluation and public announcement, and the number of rejected bids, reducing the impact of fluctuations in a single indicator on risk level determination. By structurally expressing the set of transaction subject associations, the stability and consistency of risk early warning results are improved, providing clear screening criteria for subsequent verification actions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of transaction risk early warning technology, and in particular to a method for early warning of public resource transaction risks based on big data analysis and engineering consulting. Background Technology

[0002] The technology field of transaction risk early warning specifically refers to the technical direction of identifying, classifying, and triggering disposal procedures for risks in the behavior of transaction entities, the status of the transaction process, and the transaction results in public resource transactions. The technology field is based on the constraints of public resource transaction business rules, regulatory requirements, and engineering consulting implementation specifications. By constructing risk identification models, risk assessment rule sets, and risk response mechanisms, it conducts joint analysis of transaction plans, bidding document terms, and the status of the transaction process to achieve systematic identification and control of bid rigging, abnormal pricing, deviations from contract performance, violations, and transaction risks. The core of the technology field lies in quantifying transaction risks through rule logic, model calculations, and engineering consulting verification results, and establishing a definite correlation between the risk assessment results and subsequent regulatory actions and engineering consulting disposal.

[0003] The public resource transaction risk early warning method based on big data analysis and engineering consulting is a technical solution for constructing risk identification, risk early warning, and risk disposal throughout the entire public resource transaction process. Its purpose is to identify potential violations during the transaction implementation stage and before completion, and to trigger intervention measures before the risks spread. The method uses logical correlation analysis of the behavioral characteristics, bidding behavior patterns, and performance history of transaction entities, combined with the compliance verification results of bidding document terms by engineering consulting, to form a transaction risk level identifier. Based on the level boundaries, it outputs risk classification results, thereby achieving pre-event early warning and in-event control of transaction activities. The method achieves the technical effects of reducing the probability of illegal transactions, shortening the time for risk identification, and improving the consistency of regulatory enforcement.

[0004] Existing public resource transaction risk early warning technologies, in practice, often focus on single transaction projects and individual transaction results. The risk identification process emphasizes static judgment of current transaction behavior, lacking a systematic collection of instances where transaction entities repeatedly participate in different announcement batches. This allows some entities to circumvent centralized supervision by participating in multiple bidding sections and multiple announcement batches. In the risk assessment process, existing technologies typically use entity behavior characteristics, pricing models, and performance records as the main analytical basis, failing to incorporate the results of engineering consulting's verification of the compliance of bidding document terms into the risk assessment process. Furthermore, the lack of stable correlations between various types of information makes risk assessment results susceptible to bias from a single source. Fluctuations in indicators make it difficult to reflect the cumulative risks of persistent abnormal behavior. Risk scoring results are often only used to mark high-risk projects, failing to establish a linkage with engineering consulting verification conclusions. Subsequent verification actions still rely on experience-based judgment, increasing execution uncertainty. Furthermore, the lack of continuous recording of the entity's association status and consulting verification results after risk identification makes it difficult to effectively track abnormal behavior in subsequent transactions. In multi-section continuous bidding scenarios, when the same entity combination appears repeatedly but the risk score of a single transaction does not exceed the threshold, existing technology cannot promptly combine engineering consulting to discover potential risks at the clause level, thereby delaying regulatory intervention and amplifying the impact of potential violations. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a public resource transaction risk early warning method based on big data analysis and engineering consulting.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a public resource transaction risk early warning method based on big data analysis and engineering consulting, comprising the following steps:

[0007] Step 1: Based on the announcement page of the public resource trading platform, read the announcement number and the bid section number, extract the bid entity identifier and associate it with the tenderer identifier and the agency identifier, combine them to form the cumulative number of occurrences of the same bid section entity pair, record the source of the announcement batch, and establish a set of transaction entity association relationships;

[0008] Step 2: Based on the set of relationships between the transaction entities, a graph neural network is used to characterize the frequency of occurrence and co-occurrence of entity pairs in different batches of announcements, mark repeated entity pairs in consecutive batches, adjust the sorting position of entity pairs according to the degree of repetition, and form a sequence of suspicious association links;

[0009] Step 3: Based on the suspicious association link sequence, read the qualification review results of the corresponding transaction unit and read the time interval for the bid evaluation announcement. After calculating the proportion of the link covering the bidding entities, record the number of rejected bids. The support vector machine will perform level boundary judgment on each recorded value to generate a transaction risk level identifier.

[0010] Step 4: Based on the suspicious association link sequence and the transaction risk level identifier, filter high-level transaction units, select the corresponding main links before sorting, match the corresponding verification item types of main co-occurrence, proxy binding and review intersection, arrange them in the order of transaction links, and output the verification action arrangement sequence;

[0011] Step 5: Based on the sequence of verification actions, write the execution conclusion marker for each verification item, accumulate the number of abnormal markers and divide the pointing intervals, write the pointing intervals back to the main body related entries and update the attention markers synchronously to obtain the main body related risk status set;

[0012] Step Six: Based on the aforementioned set of risk statuses associated with the main entities, filter the main entity-related entries marked as valid, associate them with engineering consulting business types, perform compliance checks on the terms of the tender documents and generate a set of consulting handling results, write the set of consulting handling results into the transaction unit, and obtain the set of engineering consulting handling statuses.

[0013] As a further embodiment of the present invention, the transaction entity association set includes entity pairs, announcement batch source, occurrence frequency, and bid section number; the suspicious association link sequence includes entity pairs, announcement batches, consecutive batch duplicate markers, and sorting position; the transaction risk level identifier includes transaction unit, coverage ratio of bidding entities, number of rejected bid markers, and risk level identifier; the verification action sequence includes entity link, verification item type, transaction process order, and verification action sequence; the entity association risk status set includes entity association entries, number of abnormal markers, pointing range, and attention markers; and the engineering consulting handling status set includes transaction unit, risk level identifier, and coverage ratio of bidding entities.

[0014] As a further aspect of the present invention, the specific steps for establishing the set of relationships between the transaction entities are as follows:

[0015] Based on the public resource trading platform announcement page, read the announcement number and bid section number, locate the bid entity identifier, associate the tenderer identifier and agency identifier, form entity pairs according to the announcement number and bid section number, record the batch sequence number and the number of occurrences, and generate a set of co-occurrence entries of bid section entities;

[0016] Based on the set of co-occurring entries of the subject in the tender section, merge entries of the same subject and accumulate the number of times they appear in different batches, retain the corresponding tenderer identifier and agent identifier, solidify the record of subject pair and batch relationship, and establish a set of transaction subject relationship.

[0017] As a further aspect of the present invention, the specific steps for forming the suspected association link sequence are as follows:

[0018] Based on the set of relationships between the transaction entities, the batch number of the corresponding announcement of the entity pair is read, the number is arranged in the order of batch time, the appearance position of each entity pair in different batches is recorded, the appearance trajectory of the entity pair is formed in time sorting, and the appearance trajectory is constructed into the entity pair relationship structure. The graph neural network is used to jointly characterize the relationship strength and stability of the entity pair in the co-occurrence relationship of multiple batches, and the batch distribution record set of the entity pair is obtained.

[0019] Based on the batch distribution record set of the subject pairs, the adjacent batch numbers of the occurrence trajectory of the subject pairs are compared, the number of consecutive number segments and the position of the number break are recorded, the continuous occurrence segment record is formed and associated with the corresponding subject pairs, and a continuous marker set of subject pairs is generated.

[0020] Based on the aforementioned continuous marker set of subject pairs, the subject pair records are arranged according to the range of consecutive occurrence segments and the range of cumulative occurrence batches. A limited number of subject pair entries are extracted before the arrangement, and the subject identifier and batch trajectory are retained to form a suspicious association link sequence.

[0021] As a further aspect of the present invention, the graph neural network, through the subject pair association structure formed by the set of transaction subject association relationships, loads the subject pair identifier as a node identifier item and loads the co-occurrence order and occurrence frequency of the subject pair in the announcement batch as a connection relationship value item, inputs the node identifier item and the connection relationship value item into the network, transmits the values ​​between nodes, and updates the association representation value corresponding to the connection relationship in the multi-round calculation process, thereby completing the update of the association representation value in the subject pair association structure.

[0022] As a further aspect of the present invention, the specific steps for generating the transaction risk level identifier are as follows:

[0023] Based on the suspicious association link sequence, the corresponding transaction unit number is read, the qualification review conclusion field is located, and the start and end time field of the bid evaluation announcement is read. The review status and announcement duration values ​​are collected according to the transaction unit to generate a transaction process status record set.

[0024] Based on the transaction status record set, the number of bidding entities covered by suspicious related links in each transaction unit is counted, the corresponding invalidation mark field is read and the number of marks is accumulated, and the number of coverage and the number of marks are written into the transaction unit record to generate a risk element statistical record set.

[0025] Based on the aforementioned risk factor statistical record set, the support vector machine performs classification boundary determination on the risk factor combination consisting of the number of subjects covered, the value of the public announcement duration, and the number of times the bid was rejected, and writes the interval code combination into the record corresponding to the transaction unit number to generate a transaction risk level identifier.

[0026] As a further aspect of the present invention, the support vector machine performs pairwise comparison operations on the input multidimensional numerical vector set, selects the vector at the boundary position as the constraint vector, iteratively calculates the interval relationship between the constraint vectors, updates the decision function used to distinguish different numerical distribution regions during the calculation process, and solidifies the decision function after the decision function stabilizes.

[0027] As a further aspect of the present invention, the specific steps for outputting the verification action sequence are as follows:

[0028] Based on the suspicious association link sequence and the transaction risk level identifier, read the transaction unit number and filter the corresponding high-level records, locate the association link sorting sequence number within the transaction unit, extract the link entries before sorting and write them into the subject relationship identifier field, and generate the verification link selection set;

[0029] Based on the selected set of verification links, the subject relationship identifier field is read and matched with the corresponding verification item number one by one. The qualification review, bid evaluation and bid award order fields in the transaction unit are read, the verification item numbers are rearranged in order, and the verification action sequence is output.

[0030] As a further aspect of the present invention, the specific steps for obtaining the subject-related risk status set are as follows:

[0031] Based on the sequence of verification actions, the transaction unit number is read and the corresponding verification item execution conclusion mark is written one by one. The conclusion marks are collected by transaction unit and the number of abnormal marks is accumulated. The number value is assigned to the preset interval code to obtain the interval record set.

[0032] Based on the aforementioned set of records for the specified interval, the interval code is written into the corresponding entity-related entry, the value of the attention marker within the entry is updated, and the reference relationship of the transaction unit number is retained. The status of the related entries is summarized according to the entity identifier to obtain the entity-related risk status set.

[0033] As a further aspect of the present invention, the specific steps for obtaining the transaction risk level identifier are as follows:

[0034] Based on the aforementioned set of risk statuses associated with the main entities, select the main entity-related entries marked as valid, associate them with engineering consulting business types, perform compliance checks on the tender document terms of the corresponding transaction units, and generate a set of engineering consulting disposal results.

[0035] Based on the aforementioned deviation marker record set, the deviation markers are associated with the corresponding transaction units of the transaction risk level identifier. The risk level identifier value is corrected and fixed according to the deviation marker quantity range to obtain the updated transaction risk level identifier.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0037] 1. In this invention, a graph neural network is used to jointly characterize the frequency of occurrence and co-occurrence relationship of subject pairs in different announcement batches, so that the repeated participation behavior between trading subjects forms a stable association representation across batches, thereby integrating the subject behaviors that were originally scattered in multiple trading units into a continuous association link, enhancing the overall ability to identify hidden association behaviors;

[0038] 2. In this invention, the risk assessment process is shifted from judgment based on empirical thresholds to boundary differentiation based on the distribution of multidimensional results by using a support vector machine to determine the risk level based on the proportion of bidding entities covered, the time interval for bid evaluation and public announcement, and the number of rejected bids. This reduces the impact of fluctuations in a single indicator on the risk level determination.

[0039] 3. In this invention, by structurally expressing the set of relationships between transaction entities, the relationship characteristics of the entities and the status of the transaction process are made to correspond within the same risk judgment framework, thereby improving the stability and consistency of risk warning results and providing clear screening criteria for subsequent verification actions. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] Example 1

[0043] Please see Figure 1 This invention provides a technical solution: a public resource transaction risk early warning method based on big data analysis and engineering consulting, comprising the following steps:

[0044] Step 1: Based on the announcement page of the public resource trading platform, read the announcement number and the bid section number, extract the bid entity identifier and associate it with the tenderer identifier and the agency identifier, combine them to form the cumulative number of occurrences of the same bid section entity pair, record the source of the announcement batch, and establish a set of transaction entity association relationships;

[0045] Step 2: Based on the set of relationships between transaction entities, a graph neural network is used to characterize the frequency of occurrence and co-occurrence of entity pairs in different batches of announcements, mark repeated entity pairs in consecutive batches, adjust the order of entity pairs according to the degree of repetition, and form a sequence of suspicious association links;

[0046] Step 3: Based on the suspicious link sequence, read the qualification review results of the corresponding transaction unit and the time interval for the bid evaluation announcement. After calculating the proportion of the link covering the bidding entities, record the number of rejected bids. The support vector machine will perform level boundary judgment on each recorded value to generate a transaction risk level identifier.

[0047] Step 4: Based on the suspicious association link sequence and transaction risk level identifier, screen high-level transaction units, select the corresponding main links before sorting, match the main co-occurrence, proxy binding and review intersection corresponding to the verification item types, arrange them in the order of transaction links, and output the verification action arrangement sequence;

[0048] Step 5: Based on the sequence of verification actions, write the execution conclusion marker for each verification item, accumulate the number of abnormal markers and divide the pointing intervals, write the pointing intervals back to the main entity-related entries and update the attention markers synchronously to obtain the main entity-related risk status set;

[0049] Step Six: Based on the set of risk status of related entities, filter the related entity entries marked as valid, associate them with the engineering consulting business type, perform compliance verification on the terms of the bidding documents and generate a set of consulting disposal results, write the set of consulting disposal results into the transaction unit, and obtain the set of engineering consulting disposal status.

[0050] The transaction entity relationship set includes entity pairs, announcement batch source, frequency of occurrence, and bid section number; the suspicious relationship link sequence includes entity pairs, announcement batches, repeated markers for consecutive batches, and sorting position; the transaction risk level identifier includes transaction unit, coverage ratio of bidding entities, number of rejected bid markers, and risk level identifier; the verification action sequence includes entity link, verification item type, transaction process order, and verification action sequence; the entity relationship risk status set includes entity relationship items, number of abnormal markers, pointing range, and attention markers; and the engineering consulting handling status set includes transaction unit, risk level identifier, and coverage ratio of bidding entities.

[0051] The specific steps for establishing a set of relationships between transaction entities are as follows:

[0052] Based on the public resource trading platform announcement page, read the announcement number and bid section number, locate the bid entity identifier, associate the tenderer identifier and agency identifier, form entity pairs according to the announcement number and bid section number, record the batch sequence number and the number of occurrences, and generate a set of co-occurrence entries of bid section entities;

[0053] Based on the set of co-occurring items of the main entities in the tender section, merge the items of the same main entity and accumulate the number of times they appear in different batches, retain the corresponding tenderer's identifier and agency's identifier, solidify the record of the relationship between the main entity and the batch, and establish a set of the relationship between the transaction entities.

[0054] Based on the announcement page of the public resource trading platform, the text content of the announcement page is obtained through network requests. The returned text is decoded according to UTF-8 encoding rules. The announcement number field and the bid section number field in the page are located item by item. The location process is performed based on the fixed tag hierarchy order in the page. The located announcement number and bid section number content are concatenated into strings with the announcement number first and the bid section number last, connected by a fixed separator. A unique entity identifier value is assigned to the concatenation result and written to a memory mapping table. The names of the bidding entities listed in the announcement page are read one by one, and the names of the bidding entities are matched and extracted according to the character continuity rule. The system performs deduplication on the names of bidders that appear repeatedly on the same announcement page and numbers them according to their order of appearance. It extracts the fields of the corresponding names of the tendering party and the agency on the page and stores them in string form. It establishes a multi-field association record between the bidder number and the names of the tendering party and the agency. It performs a pairwise combination operation on the bidder numbers corresponding to the same announcement number and the bid section number. The combination rule is limited to not generating duplicate sequential combinations. It assigns a batch number to the generated subject pair record. The batch number is numbered sequentially after being sorted according to the announcement release time. The starting value of the number is 1. The initial value of the occurrence count of each subject pair record is set to 1. It generates a set of co-occurring entries of the bid section subjects.

[0055] Based on the set of co-occurring entries of the subject pair, the subject pair identifier is used as the grouping basis. All records in the entry set are traversed one by one and grouped. The records in each group are sorted according to the batch number. For multiple records corresponding to the same subject pair, the occurrence count is accumulated. The accumulation rule is to sum the occurrence counts of each record in the group as integers. The names of the tendering party and the agency corresponding to the group are retained. The retention rule is to select the field content of the record with the smallest batch number as the association field. The correspondence between the subject pair identifier and the batch number is fixed. The fixing method is to write the subject pair identifier and the corresponding batch number list into a key-value storage structure. In the key-value storage structure, the key is the subject pair identifier and the value is an array of batch numbers. The subject pair records that have completed the merging process are uniformly summarized and written into the result set to establish a set of transaction subject association relationships.

[0056] The specific steps to form a suspicious link sequence are as follows:

[0057] Based on the set of relationships between transaction entities, the batch number of the corresponding announcement of the entity pair is read, the number is arranged in the order of batch time, the position of each entity pair in different batches is recorded, the occurrence trajectory of the entity pair in time order is formed, and the occurrence trajectory is constructed into the entity pair relationship structure. The graph neural network is used to jointly characterize the relationship strength and stability of the entity pair in the co-occurrence relationship of multiple batches, and the batch distribution record set of entity pair is obtained.

[0058] Based on the batch distribution record set of subject pairs, the occurrence trajectory of subject pairs is compared with the adjacent batch numbers, the number of consecutive number segments and the position of number break are recorded, the continuous occurrence segment record is formed and associated with the corresponding subject pair, and a continuous marker set of subject pairs is generated.

[0059] Based on the continuous tag set of subject pairs, the subject pair records are arranged according to the range of consecutive occurrence segments and the range of cumulative occurrence batches. A limited number of subject pair entries are extracted before the arrangement, and the subject identifier and batch trajectory are retained to form a suspicious association link sequence.

[0060] Based on the set of relationships between transaction entities, a sequential traversal sorting method is used to read the announcement batch number corresponding to each entity pair identifier. The announcement batch numbers are then arranged chronologically according to the announcement publication time field, which is stored in year-month-day-hour-minute-second format and sorted from earliest to latest. The position index of each entity pair in the sorted batch number sequence is recorded, represented as an integer and numbered consecutively starting from 1. The position indices of the entity pairs in different batches are combined sequentially to form a time series record. This time series record is written into the entity pair trajectory field, and an adjacency list construction method is used to construct the entity pair trajectory. The structure is transformed into a subject-pair association structure, where each subject pair is recorded as a node. The association between nodes is based on the co-occurrence order of the same batch. Then, a graph neural network algorithm is used to process the subject-pair association structure. The graph neural network contains three propagation layers, with 128 nodes in each layer. The input data includes the occurrence count of the subject pair and the batch position index, with the values ​​limited to between 0 and 100. The propagation process performs three rounds of information updates in layer order. In each round of updates, the associated value record is written to each node. After all rounds of processing are completed, the node records are summarized to generate a subject-pair batch distribution record set.

[0061] Based on the batch distribution record set of subject pairs, a sequential comparison method is adopted to read the batch number sequence corresponding to the subject pair one by one. The difference calculation operation is performed on adjacent numbers in the batch number sequence. The difference calculation method is to subtract the preceding number from the subsequent number. The difference result is compared with integers. When the difference value is equal to 1, it is recorded as a continuous numbering state. When the difference value is greater than 1, it is recorded as a numbering break state. Segment counting is performed on the continuous numbering state. The segment counting rule is to increment the number by 1 for each continuous numbering state. The break position index is recorded for the numbering break state. A continuous numbering segment count record and a break position record are formed for each subject pair. The continuous numbering segment count and the break position index are written into the record field corresponding to the subject pair identifier. Only one continuous numbering segment statistical result is retained for the same subject pair. After the statistics of all subject pairs are completed, a continuous marker set of subject pairs is generated.

[0062] Based on the continuous tag set of subject pairs, an interval sorting and filtering method is adopted. The number of consecutive numbered segments and the cumulative number of batches corresponding to the subject pairs are read one by one. The number of consecutive numbered segments is classified according to a preset interval. The interval division rule is that 1 to 2 is the first interval, 3 to 5 is the second interval, and 6 and above is the third interval. The cumulative number of batches is marked in the same way. The two types of interval tags are combined to form a sorting key value. All subject pair records are arranged from high to low according to the sorting key value. A limited number threshold is set for the arrangement result. The limited number threshold is set to 50 records. The first 50 subject pair records are extracted from the arrangement result. The subject identifier field and batch trajectory field are retained for the extracted subject pair records. The extracted results are uniformly written into the result set to form a suspicious association link sequence.

[0063] Graph neural networks, through the subject pair association structure formed by the set of transaction subject association relationships, load the subject pair identifier as node identifier item and load the co-occurrence order and occurrence frequency of subject pairs in the announcement batch as connection relationship value item. Input the node identifier item and connection relationship value item into the network, pass the values ​​between nodes, and update the association representation value corresponding to the connection relationship in multiple rounds of calculation, thus completing the update of the association representation value in the subject pair association structure.

[0064] Graph neural networks, according to the formula:

[0065]

[0066] in: Indicates the first The feature matrix of the subject to the node in a wheel graph neural network. Indicates completion of the first The result obtained after round of calculation Wheel body node feature matrix, This serves as an index for the rounds in the propagation process of a graph neural network. This indicates that the subject is related to the first element in the association structure. Each subject to a node Indicates the relationship with the first The first entity has a relationship with the node. Each subject to a node Indicates the subject's... With the subject The cumulative number of times the event occurred across different announcement batches. Indicates the subject's... With the subject The co-occurrence sequence number in the announcement batch Indicates the subject's... With the subject The interval value between the corresponding batch numbers of the announcement. This represents the smoothing parameter used to calculate stability by balancing co-occurrence counts. This represents the order decay parameter used to adjust the degree of influence of co-occurrence order. This represents the interval decay parameter used to adjust the degree of impact of the announcement batch interval. This represents the self-join weight parameter used to indicate the association relationship between the subject and its own nodes. This represents the degree matrix calculated based on the weighted association relationships between the subjects. Indicates the first The weight matrix used for feature transformation in a wheel graph neural network. This represents an activation operator that is applied element-by-element to the computation result. Represents the natural constant;

[0067] Execution process: First, the set of relationships between transaction entities is constructed into a subject-pair relationship structure, and each subject pair is written as a node into the initial feature matrix. Subsequently in In the round calculation, based on the subject's... With the subject Cumulative co-occurrence counts in the announcement batch Co-occurrence sequence number and the batch interval values ​​of the announcement By introducing a smoothing parameter Sequential decay parameters Interval attenuation parameters With self-connection weight parameters The relationships between subject pairs are expressed using a unified weighted average and combined with a degree matrix. Normalization is performed on the weighted association relationships, and the normalization result is compared with the feature matrix of the previous subject pair. and weight matrix After multiplying, we get the first... Wheel body with feature matrix The numerical update is completed by activating the ReLU operator, and the process is indexed by propagation round. The process is executed round by round, thereby continuously updating the association representation values ​​of the subject in the associated structure in multiple rounds of calculation, providing basic data support for subsequent transaction risk level determination and verification action ranking.

[0068] The specific steps for generating a transaction risk level identifier are as follows:

[0069] Based on the suspicious related link sequence, read the corresponding transaction unit number, locate the qualification review conclusion field and read the bid evaluation announcement start and end time field, collect the review status and announcement duration values ​​according to the transaction unit, and generate a transaction process status record set.

[0070] Based on the transaction status record set, the number of bidding entities covered by suspicious related links within each transaction unit is counted, the corresponding invalidation mark field is read and the number of marks is accumulated, and the number of coverage and the number of marks are written into the transaction unit record to generate a risk element statistical record set.

[0071] Based on the statistical record set of risk factors, the support vector machine performs classification boundary determination on the risk factor combination consisting of the number of subjects covered, the value of the public announcement duration, and the number of times the bid was rejected. The interval code combination is written into the record corresponding to the transaction unit number to generate a transaction risk level identifier.

[0072] Based on the suspicious link sequence, a sequential field positioning method is adopted. The transaction unit number recorded in the link sequence is read one by one. The transaction unit number is used as the search key to access the transaction data storage table. The fields in the storage table are scanned one by one in the order of field name. For the record whose field name is equal to the qualification review conclusion identifier, the value reading operation is performed. The reading content is limited to the pass and fail status flags and stored in integer form. Then, the bidding announcement information field under the same transaction unit number is located. The positioning method is to perform reading operations on the records whose field name is equal to the bidding announcement start time identifier and the bidding announcement end time identifier respectively. The start time and end time are parsed into time values ​​in the format of year, month, day, hour, minute and second. The time difference operation is performed on the two. The calculation method is to subtract the start time from the end time and convert it into an hour value. The hour value is recorded in integer form. The qualification review status value and announcement duration value of each transaction unit are collected and written into the transaction unit record structure. After all transaction units are processed, a transaction process status record set is generated.

[0073] Based on the transaction status record set, a counting and summarizing method is used to read the suspicious related link records corresponding to each transaction unit in the record set one by one. A uniqueness comparison operation is performed on the bid entity identifier associated in each link record. The comparison rule is that the same bid entity identifier is counted only once. The number of bid entity identifiers after the comparison is completed is counted as an integer and recorded as the number of covered bid entities. Then, the bid rejection mark field in the data record corresponding to the same transaction unit number is read. The bid rejection mark field uses 0 and 1 to represent the non-rejected and rejected bid status. A sequential traversal operation is performed on the bid rejection mark field. An accumulation count operation is performed on the record with a mark value of 1. The accumulation result is recorded as the number of bid rejection marks. The value of the number of covered bid entities and the value of the number of bid rejection marks are written into the transaction unit record structure. After the statistical writing of all transaction units is completed, a risk element statistical record set is generated.

[0074] Based on a statistical record set of risk factors, a support vector machine (SVM) algorithm is used to construct risk factor combinations for each trading unit, including the number of entities covered, the duration of public disclosure, and the number of times a bid is rejected. These risk factor combinations are then standardized by subtracting the mean of the corresponding element set from each value and dividing by the standard deviation. The mean and standard deviation are pre-calculated and stored from historical trading records. The standardized risk factor combinations are then input into a SVM classification model. The SVM uses a radial basis function with kernel parameters (0.5) and a penalty parameter of 1. Classification is performed on the input data sequentially by trading unit. The classification output is then mapped to a preset level interval code, defined as three integer interval values. A correspondence is established between the mapped interval codes and the trading unit numbers, and this correspondence is written into a storage record. After writing to all trading units, a trading risk level identifier is generated.

[0075] Support Vector Machine (SVM) performs pairwise comparison operations on the input multidimensional numerical vector set, selects the vector at the boundary position as the constraint vector, iteratively calculates the interval relationship between the constraint vectors, updates the decision function used to distinguish different numerical distribution regions during the calculation process, and solidifies the decision function after it stabilizes.

[0076] Support Vector Machine, according to the formula:

[0077]

[0078] in: This represents the weight vector of the decision hyperplane in the support vector machine. Represents the weight vector The Middle The weight components corresponding to each feature dimension This represents the feature dimension index, with values ​​ranging from 1 to 3. This represents the bias parameters in the decision function of the support vector machine. Represents the set of slack variables. Indicates the first The relaxation variables corresponding to each training sample Indicates the sample index. This represents the total number of training samples. Indicates the first The feature scaling factor corresponding to each feature dimension This represents the error penalty coefficient parameter. Indicates the relationship with the first Individual sample category identifier The corresponding category weight coefficients, Indicates the first The risk level category identifier of each sample Indicates the first The confidence weight coefficients corresponding to each sample This represents the quadratic penalty coefficient for the squared term of the slack variable. This represents the parameter set operator used to find the minimum value of the objective expression;

[0079] Execution process: First, construct a sample index by considering the number of entities covered, the duration of public announcement, and the number of times the bid was rejected for each trading unit. The three-dimensional feature vectors are summarized to form a total number of samples. The training set, then indexed according to feature dimensions. Initialize weight vector Each component in And configure the corresponding feature scaling coefficients for each feature dimension. Simultaneously, slack variables were configured for each sample. And identified according to sample category Select the corresponding category weight coefficient Then, combined with the sample confidence weight coefficient With penalty coefficient parameter and secondary penalty coefficient Substitute all parameters into the above formula, and then perform a calculation on the parameter set. Perform the minimization operation and continuously update the weight vector during the iteration process. Bias parameters and slack variables The parameter set is output after the target expression value stabilizes, and is used for subsequent calculation of transaction risk level.

[0080] The specific steps for outputting the sequence of verification actions are as follows:

[0081] Based on the suspicious related link sequence and the transaction risk level identifier, the transaction unit number is read and the corresponding high-level record is filtered. The sorting sequence number of the related links within the transaction unit is located, the link entries before sorting are extracted and written into the subject relationship identifier field to generate the verification link selection set.

[0082] Based on the verification link selection set, read the subject relationship identifier field and match the corresponding verification item number one by one, read the qualification review, bid evaluation and bid award sequence field in the transaction unit, rearrange the verification item numbers in order, and output the verification action sequence.

[0083] Based on suspicious related link sequences and transaction risk level identifiers, a conditional filtering and sorting interception method is adopted. The transaction unit number of each record in the suspicious related link sequence is read one by one. The transaction unit number is used as the search key to access the transaction risk level storage table. The transaction risk level field in the storage table is subjected to numerical comparison operation. Records with risk level values ​​equal to a preset high-level mark are filtered. The high-level mark is manually set and fixed in advance by regulatory rules and has a value of 3. The filtered set of transaction unit numbers is processed one by one. The corresponding related link sorting number field of each transaction unit number is read. The sorting number field is an integer and arranged in ascending order. The sorting number field is traversed sequentially. The interception threshold is set to 20. Link records with sorting numbers less than or equal to 20 are retained. The retained link records are written into the subject relationship identifier field. The subject relationship identifier field consists of subject pair identifier strings. The filtered and intercepted link records are summarized and written into the result set to generate the verification link selection set.

[0084] Based on the selected set of verification links, a sequential matching and rearrangement method is adopted. The subject relationship identifier field of each record in the selected set is read one by one. The subject relationship identifier field is used as the index key to access the verification item mapping table. The correspondence between the subject relationship identifier and the verification item number in the mapping table is matched one by one. The verification item number is represented in integer form and is predefined. The verification item number that has been matched is written into a temporary sequence structure. The qualification review order field, bid evaluation order field and bid award order field corresponding to the same transaction unit number are read respectively. The order field value is an integer and is pre-set by the transaction process rules. The verification item number in the temporary sequence structure is sorted by multiple keys according to the rule that the qualification review order field value is the first priority, the bid evaluation order field value is the second priority, and the bid award order field value is the last priority. The sorted verification item number sequence is processed by sequential numbering, with the starting value of the number being 1. The sorted and numbered verification item number sequence is written into the result set. After all transaction units are processed, the verification action arrangement sequence is output.

[0085] The specific steps to obtain the set of risk states associated with the subject are as follows:

[0086] Based on the sequence of verification actions, the transaction unit number is read and the corresponding verification item execution conclusion mark is written one by one. The conclusion marks are collected by transaction unit and the number of abnormal marks is accumulated. The quantity value is assigned to the preset interval code to obtain the interval record set.

[0087] Based on the set of records pointing to the interval, the interval code is written into the corresponding subject-related entries, the value of the attention mark in the entry is updated and the reference relationship of the transaction unit number is retained. The status of the related entries is summarized according to the subject identifier to obtain the subject-related risk status set.

[0088] Based on the sequence of verification actions, a sequential writing and interval mapping method is used to read the transaction unit numbers recorded in the sequence one by one. The transaction unit number is used as an index key to access the verification item execution record table. The verification item numbers that match the transaction unit numbers in the table are located one by one. An execution conclusion mark is written to each verification item number. The execution conclusion mark value is limited to 0 and 1 and stored in integer form. Then, a collection operation is performed on all execution conclusion marks under the same transaction unit number. The collection method is to aggregate records by transaction unit number. An anomaly mark counting operation is performed on the collection result. The counting rule is to accumulate the records with execution conclusion mark values ​​equal to 1 to obtain the anomaly mark count value. An interval mapping process is performed on the anomaly mark count value. The interval code is pre-set manually and stored. The interval division rule is that 0 to 1 corresponds to code 1, 2 to 3 corresponds to code 2, and 4 and above corresponds to code 3. A correspondence is established between the mapped interval code and the transaction unit number and written into the storage structure. After all transaction units are processed, a set of records pointing to intervals is generated.

[0089] Based on a set of records pointing to intervals, an associated writing and summarizing method is used. The correspondence between transaction unit numbers and interval codes in the record set is read record by record. The transaction unit number is used as an index key to access the subject-related entry table. Each subject-related entry in the table is located, and the interval code is written to the corresponding subject-related entry field. An update operation is performed on the attention mark field in the subject-related entry. The attention mark value is limited to 0 and 1 and stored as an integer. The update rule is to write 1 if the interval code is greater than 1, otherwise write 0. A write operation is performed on the transaction unit number reference relationship field in the subject-related entry. This reference relationship field is stored as an array of transaction unit numbers. Subsequently, a grouping operation is performed on all subject-related entries according to the subject identifier. A status summarizing operation is performed on multiple related entries corresponding to the same subject identifier. The summarizing content includes a set of interval codes and a set of attention marks. The summarizing result is written to a result set. After processing all subjects, a subject-related risk status set is obtained.

[0090] The specific steps to obtain a transaction risk level label are as follows:

[0091] Based on the set of risk status of related entities, we screen the related entity entries marked as valid, associate them with engineering consulting business types, perform compliance verification on the tender document terms of the corresponding transaction units, and generate a set of engineering consulting disposal results.

[0092] Based on the record set of deviation markers, the deviation markers are associated with the corresponding transaction units of the transaction risk level identifier. The risk level identifier value is corrected and fixed according to the range of deviation markers, resulting in the updated transaction risk level identifier.

[0093] Based on the risk status set of the main entity association, the method of consulting object positioning and clause marking is adopted. The transaction unit number referenced in the main entity association entries is read one by one. The transaction unit number is used as the index key to access the tender document clause record table. The clause number matching the transaction unit number in the table is located one by one. Each clause number is written with a compliance verification mark. The value of the verification mark is limited to 0 and 1 and stored in integer form. Then, a collection operation is performed on all verification marks under the same transaction unit number. The collection method is to aggregate records by transaction unit number. A deviation mark counting operation is performed on the collection result. The counting rule is to accumulate the records with a verification mark value of 1 to obtain the deviation mark quantity value. The deviation mark quantity value is processed by interval mapping. The interval code is pre-set and fixedly stored. The interval code is established with the transaction unit number and written into the storage structure to generate the engineering consulting disposal result set.

[0094] Based on the engineering consulting outcome set, a risk level correction and result solidification method is adopted. The transaction unit number and its corresponding interval code of the record in the outcome set are read one by one. The transaction unit number is used as the index key to access the transaction risk level identifier storage table. The risk level identifier records that match the transaction unit number are located one by one. The original risk level identifier value is read and temporarily stored. The correspondence between the interval code and the original risk level identifier value is determined. According to the determination result, the risk level identifier value is corrected and written. The corrected risk level identifier value is stored in integer form. Then, the corrected risk level identifier records under the same transaction unit number are overwritten and written. The integrity of the correction result is checked. The records that pass the check are solidified and stored. After all transaction units are processed, the updated transaction risk level identifier is obtained.

[0095] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A public resource transaction risk early warning method based on big data analysis and engineering consulting, characterized in that, Includes the following steps: Step 1: Based on the announcement page of the public resource trading platform, read the announcement number and the bid section number, extract the bid entity identifier and associate it with the tenderer identifier and the agency identifier, combine them to form the cumulative number of occurrences of the same bid section entity pair, record the source of the announcement batch, and establish a set of transaction entity association relationships; Step 2: Based on the set of relationships between the transaction entities, a graph neural network is used to characterize the frequency of occurrence and co-occurrence of entity pairs in different batches of announcements, mark repeated entity pairs in consecutive batches, adjust the order of entity pairs according to the degree of repetition, and form a sequence of suspicious association links; Step 3: Based on the suspicious association link sequence, read the qualification review results of the corresponding transaction unit and read the time interval for the bid evaluation announcement. After calculating the proportion of the link covering the bidding entities, record the number of rejected bids. The support vector machine will perform level boundary judgment on each recorded value to generate a transaction risk level identifier. Step 4: Based on the suspicious association link sequence and the transaction risk level identifier, filter high-level transaction units, select the corresponding main links before sorting, match the corresponding verification item types of main co-occurrence, proxy binding and review intersection, arrange them in the order of transaction links, and output the verification action arrangement sequence; Step 5: Based on the sequence of verification actions, write the execution conclusion marker for each verification item, accumulate the number of abnormal markers and divide the pointing intervals, write the pointing intervals back to the main body related entries and update the attention markers synchronously to obtain the main body related risk status set; Step Six: Based on the aforementioned set of risk statuses associated with the main entities, filter the main entity-related entries marked as valid, associate them with engineering consulting business types, perform compliance checks on the terms of the tender documents and generate a set of consulting handling results, write the set of consulting handling results into the transaction unit, and obtain the set of engineering consulting handling statuses.

2. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The transaction entity association set includes entity pairs, announcement batch source, frequency of occurrence, and bid section number; the suspicious association link sequence includes entity pairs, announcement batches, consecutive batch duplicate markers, and sorting position; the transaction risk level identifier includes transaction unit, coverage ratio of bidding entities, number of rejected bid markers, and risk level identifier; the verification action sequence includes entity link, verification item type, transaction process order, and verification action sequence; the entity association risk status set includes entity association entries, number of abnormal markers, pointing range, and attention markers; and the engineering consulting handling status set includes transaction unit, risk level identifier, and coverage ratio of bidding entities.

3. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps for establishing the set of relationships between the transaction entities are as follows: Based on the public resource trading platform announcement page, read the announcement number and bid section number, locate the bid entity identifier, associate the tenderer identifier and agency identifier, form entity pairs according to the announcement number and bid section number, record the batch sequence number and the number of occurrences, and generate a set of co-occurrence entries of bid section entities; Based on the set of co-occurring entries of the subject in the tender section, merge entries of the same subject and accumulate the number of times they appear in different batches, retain the corresponding tenderer identifier and agent identifier, solidify the record of subject pair and batch relationship, and establish a set of transaction subject relationship.

4. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps for forming the suspected association link sequence are as follows: Based on the set of relationships between the transaction entities, the batch number of the corresponding announcement of the entity pair is read, the number is arranged in the order of batch time, the appearance position of each entity pair in different batches is recorded, the appearance trajectory of the entity pair is formed in time sorting, and the appearance trajectory is constructed into the entity pair relationship structure. The graph neural network is used to jointly characterize the relationship strength and stability of the entity pair in the co-occurrence relationship of multiple batches, and the batch distribution record set of the entity pair is obtained. Based on the batch distribution record set of the subject pairs, the adjacent batch numbers of the occurrence trajectory of the subject pairs are compared, the number of consecutive number segments and the position of the number break are recorded, the continuous occurrence segment record is formed and associated with the corresponding subject pairs, and a continuous marker set of subject pairs is generated. Based on the aforementioned continuous marker set of subject pairs, the subject pair records are arranged according to the range of consecutive occurrence segments and the range of cumulative occurrence batches. A limited number of subject pair entries are extracted before the arrangement, and the subject identifier and batch trajectory are retained to form a suspicious association link sequence.

5. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The graph neural network, through the subject pair association structure formed by the set of transaction subject association relationships, loads the subject pair identifier as a node identifier item and loads the co-occurrence order and occurrence frequency of the subject pair in the announcement batch as a connection relationship value item. It inputs the node identifier item and the connection relationship value item into the network, transmits the values ​​between nodes, and updates the association representation value corresponding to the connection relationship in multiple rounds of calculation, thus completing the update of the association representation value in the subject pair association structure.

6. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps for generating the transaction risk level identifier are as follows: Based on the suspicious association link sequence, the corresponding transaction unit number is read, the qualification review conclusion field is located, and the start and end time field of the bid evaluation announcement is read. The review status and announcement duration values ​​are collected according to the transaction unit to generate a transaction process status record set. Based on the transaction status record set, the number of bidding entities covered by suspicious related links in each transaction unit is counted, the corresponding invalidation mark field is read and the number of marks is accumulated, and the number of coverage and the number of marks are written into the transaction unit record to generate a risk element statistical record set. Based on the aforementioned risk factor statistical record set, the support vector machine performs classification boundary determination on the risk factor combination consisting of the number of subjects covered, the value of the public announcement duration, and the number of times the bid was rejected, and writes the interval code combination into the record corresponding to the transaction unit number to generate a transaction risk level identifier.

7. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The support vector machine performs pairwise comparison operations on the input multidimensional numerical vector set, selects the vector at the boundary position as the constraint vector, iteratively calculates the interval relationship between the constraint vectors, updates the decision function used to distinguish different numerical distribution regions during the calculation process, and solidifies the decision function after it stabilizes.

8. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps for outputting the sequence of verification actions are as follows: Based on the suspicious association link sequence and the transaction risk level identifier, read the transaction unit number and filter the corresponding high-level records, locate the association link sorting sequence number within the transaction unit, extract the link entries before sorting and write them into the subject relationship identifier field, and generate the verification link selection set; Based on the selected set of verification links, the subject relationship identifier field is read and matched with the corresponding verification item number one by one. The qualification review, bid evaluation and bid award order fields in the transaction unit are read, the verification item numbers are rearranged in order, and the verification action sequence is output.

9. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps to obtain the set of associated risk states of the subject are as follows: Based on the sequence of verification actions, the transaction unit number is read and the corresponding verification item execution conclusion mark is written one by one. The conclusion marks are collected by transaction unit and the number of abnormal marks is accumulated. The number value is assigned to the preset interval code to obtain the interval record set. Based on the aforementioned set of records for the specified interval, the interval code is written into the corresponding entity-related entry, the value of the attention marker within the entry is updated, and the reference relationship of the transaction unit number is retained. The status of the related entries is summarized according to the entity identifier to obtain the entity-related risk status set.

10. The public resource transaction risk early warning method based on big data analysis and engineering consulting according to claim 1, characterized in that, The specific steps to obtain the transaction risk level identifier are as follows: Based on the aforementioned set of risk statuses associated with the main entities, select the main entity-related entries marked as valid, associate them with engineering consulting business types, perform compliance checks on the tender document terms of the corresponding transaction units, and generate a set of engineering consulting disposal results. Based on the aforementioned deviation marker record set, the deviation markers are associated with the corresponding transaction units of the transaction risk level identifier. The risk level identifier value is corrected and fixed according to the deviation marker quantity range to obtain the updated transaction risk level identifier.