A dynamic weight-based electronic bidding risk early warning system and method

Abstract

This invention discloses a dynamic weight-based electronic bidding risk early warning system and method, specifically relating to the field of bidding supervision. It addresses the shortcomings of existing electronic bidding processes, such as the lack of dynamic correlation analysis of review behavior, difficulty in timely identification of clause-induced review deviation risks, and insufficient accuracy of risk warnings. By analyzing the clause nodes of the bidding documents, the project complexity level, and the supplier's proposed text to generate basic risk weights, and combining this with the reviewer's clause access behavior trajectory during the review process to generate clause traction strength, a normal traction fluctuation envelope is constructed based on the scheme text difference, historical questioning frequency, and historical rejection frequency. This allows for dynamic risk identification of continuously increasing clause-induced behavior. Simultaneously, risk matching is performed using historically identified cases of bid rigging, abnormally low-price bids, and qualification abnormalities, achieving real-time early warning and dynamic supervision of abnormal review behavior during electronic bidding.

Description

Technical Field

[0001] This invention relates to the field of bidding and tendering supervision technology, and more specifically, to an electronic bidding and tendering risk early warning system and method based on dynamic weights. Background Technology

[0002] With the continuous promotion of public resource trading platforms, centralized enterprise procurement platforms, and electronic bidding systems, bidding activities are gradually shifting from traditional offline review methods to a fully electronic review model. In large-scale group procurement, regional centralized procurement, and cross-industry joint procurement scenarios, the review process typically involves a large number of supplier proposal documents, complex technical terms, and multiple rounds of online review operations. As the complexity of bidding business continues to increase, risk behaviors such as bid rigging, abnormally low-price bids, abnormal qualification materials, and targeted screening of terms are gradually showing characteristics of process concealment, behavior decentralization, and dynamic correlation.

[0003] Existing electronic bidding risk supervision methods mostly focus on the final scoring results, price deviations, or analysis of corporate relationships. They lack dynamic correlation analysis of clause access behavior, scoring modification behavior, and continuous monitoring behavior during the review process, making it difficult to identify the continuous influence of some abnormal clauses on the review process.

[0004] In actual review scenarios, certain clauses may continuously attract review attention due to frequent historical disputes, strong supplier selection, or special technical limitations. This can lead to reviewers repeatedly reviewing specific clauses, continuously modifying scores, or even creating review deviation risks. However, existing systems lack a dynamic risk assessment mechanism that combines project complexity, historical risk characteristics, and changes in behavioral patterns, making it difficult to achieve real-time identification and accurate early warning of risky behaviors in the electronic bidding process. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an electronic bidding risk early warning system and method based on dynamic weights to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for risk early warning in electronic bidding based on dynamic weights includes the following steps: S1. During the review stage, analyze the clauses and nodes of the tender documents, the project complexity level and the text of each supplier's proposal, and count the historical rejection frequency, questioning frequency and supplier pass rate reduction ratio of each clause, and generate basic risk weights by weighting. S2. Real-time capture of the supplier's proposal during the bidding review process, including the focus switching and score modification operations of the review body on each clause node, and generate a clause access behavior trajectory in chronological order. S3. Obtain the access behavior trajectory of all supplier solutions for the terms, trace back the behavior trajectory to identify related terms nodes, count the number of repeated visits of each related node in the entire behavior trajectory, and synthesize the term traction strength with the basic risk weight. S4. Calculate the degree of difference between the texts of each supplier's proposal, integrate the project complexity level and the frequency of historical questioning of the clauses to form a behavior correction coefficient, and make a weighted correction to the traction strength of the clauses. S5. Set a review time window. When the window ends, calculate the window change of the traction strength of each clause after the revision. When the strength of the same clause continues to increase in multiple consecutive windows, mark it as a continuous traction node and record the growth rate sequence. S6. Check whether the increase in traction intensity of the continuous traction node exceeds the upper limit of traction intensity fluctuation based on the historical veto frequency. If it exceeds, it is determined that there is a risk of clause-induced review deviation. S7. Extract the clause node content, supplier scheme identifier and associated review entity identifier corresponding to the clause-induced review deviation risk, and generate an electronic bidding risk warning to push to the supervision terminal.

[0008] As a further aspect of the present invention, in step S1, generating the basic risk weight specifically includes: The project complexity level is a project complexity identifier generated based on the size of the supplier's proposal text and the number of review clauses. Each level corresponds to a different complexity quantification value. Extract the records of rejection events, objection events, and bidder approvals for each clause node in the historical review archives. Use the number of rejection event records as the historical rejection frequency and the number of objection event records as the historical objection frequency. Based on the total number of bidding suppliers under the constraints of the clause nodes, the decrease in the number of qualified suppliers after the review is calculated as the supplier pass rate reduction ratio. The frequency of historical rejections, the frequency of historical challenges, and the supplier pass rate reduction ratio are scaled down relative to their respective maximum values ​​across all clause nodes to obtain the rejection weight, challenge weight, and reduction weight. The normalized weighted sum of the rejection weight, challenge weight, and reduction weight is then calculated and used as the basic risk weight for the corresponding clause node.

[0009] As a further aspect of the present invention, in step S2, generating the terms access behavior trajectory specifically includes: For each supplier's solution text review process, monitor the focus operation corresponding to each clause node in the review interface. When the review subject focuses on the clause node, record the corresponding clause node identifier, focus time, and the previous focused clause node identifier as a focus switching event. Monitor the score values ​​of each clause node, and when a change occurs, record the time of the change, the changed value, and the current clause node identifier as a score modification event; The focus switching events and score modification events generated by the same review subject are arranged in ascending order according to their recorded times, generating a trace of the clause access behavior of each review subject, which is grouped by the supplier solution identifier.

[0010] As a further aspect of the present invention, in S3, the traction strength of the composite clause specifically includes: In the clause access behavior trajectory of each review subject corresponding to the supplier solution identifier, when a review subject triggers a score modification event, the time segment between the previous score modification event and the current score modification event is extracted from the clause access behavior trajectory of the review subject. Extract all clause node identifiers whose number of focus counts exceeds a preset normal number threshold within a time segment, and form a set of related clause nodes after deduplication; Traverse the identifier of each clause node in the associated clause node set, retrieve the clause access behavior trajectory of all review subjects, and count the cumulative number of repeated visits to the corresponding clause node. The cumulative number of repeated visits to each associated clause node is correlated with the basic risk weight of the corresponding clause node to generate the clause traction strength of the corresponding clause node.

[0011] As a further aspect of the present invention, the weighted correction of the clause traction strength in S4 specifically includes: Technical parameter description paragraphs and requirement response paragraphs are extracted from the solution texts of each supplier. Each paragraph text is segmented into words, and the words are converted into word vectors in a pre-set word vector dictionary. The average vector of word vectors in the paragraph is calculated as the paragraph embedding vector. Calculate the cosine distance between the embedding vectors of the same type of paragraphs from any two supplier proposals, and take the arithmetic mean of the cosine distances of all supplier pairs as the text distance of the proposals. Read the project complexity level, convert it into a complexity quantification value according to the preset mapping relationship, and read the historical questioning frequency of the corresponding clause node; By using the text distance and complexity quantification of the scheme as behavioral mitigation factors, the current clause traction strength is attenuated and corrected. The frequency of historical questioning is used as a risk enhancement factor to enhance and correct the attenuated clause traction strength, thus obtaining the corrected traction strength of the corresponding clause node.

[0012] As a further aspect of the present invention, in step S5, recording the growth rate sequence specifically includes: At the end of the review time window, each clause node is traversed one by one, and the latest generated value is extracted from all the revised traction strength values ​​in the window as the traction strength of the current window of the corresponding clause node. For each clause node, compare the current window traction strength with the traction strength of the immediately preceding window. If the current value is greater than the previous value, set the window change flag of the corresponding clause node to increase. Continuously monitor multiple review time windows. When the window change indicator of a clause node shows continuous growth and the number of consecutive growths reaches the preset number of consecutive growths, mark the corresponding clause node as a continuous traction node. Extract the traction intensity of each window in the window time sequence to form a growth rate sequence.

[0013] As a further aspect of the present invention, in step S6, determining the risk of clause-induced review deviation specifically includes: Extract the historical veto frequency records and corresponding historical traction intensity records of the continuous traction nodes in historical projects; Based on the historical veto frequency, the historical traction intensity records are filtered by interval, and the historical traction intensity change records corresponding to the historical veto frequency of the current continuous traction node are extracted to form the normal traction intensity change set of the corresponding clause node. Perform fluctuation boundary fitting on the set of normal traction intensity changes to generate the normal traction degree fluctuation envelope for the corresponding clause node; Calculate the current growth rate corresponding to the continuous traction node based on the growth rate sequence, and check whether the current growth rate exceeds the normal traction fluctuation envelope of the corresponding clause node; When the current growth rate exceeds the normal traction fluctuation envelope, it is determined that there is a risk of clause-induced review deviation.

[0014] On the other hand, the present invention provides an electronic bidding risk early warning system based on dynamic weights, comprising: The data collection module is used to collect bidding documents, bidding activities, historical review cases, and market supervision information in a unified manner. The weight adjustment module is used to dynamically adjust the basic risk weight and clause traction strength calculation weight of the corresponding clause nodes based on the project complexity level, changes in the review stage, and the behavior of the review subjects. The traction analysis module is used to count the number of repeated visits to clause nodes based on the clause access behavior trajectory of each review subject, and generate the clause traction strength of the corresponding clause node. The risk identification module is used to identify risk patterns based on the deviation between the growth rate corresponding to the continuous traction node and the normal traction intensity fluctuation envelope, combined with historically identified cases of bid rigging, abnormally low-price bidding, and qualification fraud, and generate corresponding risk warning information. The early warning push module is used to generate corresponding graded early warning messages based on the degree of deviation in the risk warning information, and push them to the procurement supervision terminal and the bidding management terminal.

[0015] The technical effects and advantages of the electronic bidding risk early warning system and method based on dynamic weights of the present invention are as follows: This invention constructs a dynamic risk early warning mechanism based on the behavioral trajectory of clause nodes. It performs unified correlation analysis on clause access behavior, scoring modification behavior, and continuous monitoring behavior during the electronic bidding review process. This enables the identification of abnormal clause influence on the review process before the review results are formed, improving the ability to identify review deviation risks in advance. By introducing basic risk weights, project complexity levels, and scheme text differences, it dynamically corrects normal review monitoring behaviors in different project scenarios, reducing false alarms in complex projects and highly differentiated scheme scenarios. By constructing a normal traction fluctuation envelope based on historical rejection frequencies, it achieves differentiated risk assessment for clauses of different risk levels, improving the scenario adaptability of risk identification. Simultaneously, by combining historically identified cases of bid rigging, abnormally low-price bidding, and qualification abnormalities for risk pattern matching, it enhances the comprehensiveness and accuracy of electronic bidding risk identification, contributing to improved transparency, traceability, and risk supervision efficiency in the bidding process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an electronic bidding risk early warning method based on dynamic weights according to the present invention; Figure 2 This is a schematic diagram of the structure of an electronic bidding risk early warning system based on dynamic weights according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Figure 1 This invention presents a method for risk early warning in electronic bidding based on dynamic weights, which includes the following steps: S1. During the review stage, analyze the clauses and nodes of the tender documents, the project complexity level and the text of each supplier's proposal, and count the historical rejection frequency, questioning frequency and supplier pass rate reduction ratio of each clause, and generate basic risk weights by weighting. S2. Real-time capture of the supplier's proposal during the bidding review process, including the focus switching and score modification operations of the review body on each clause node, and generate a clause access behavior trajectory in chronological order. S3. Obtain the access behavior trajectory of all supplier solutions for the terms, trace back the behavior trajectory to identify related terms nodes, count the number of repeated visits of each related node in the entire behavior trajectory, and synthesize the term traction strength with the basic risk weight. S4. Calculate the degree of difference between the texts of each supplier's proposal, integrate the project complexity level and the frequency of historical questioning of the clauses to form a behavior correction coefficient, and make a weighted correction to the traction strength of the clauses. S5. Set a review time window. When the window ends, calculate the window change of the traction strength of each clause after the revision. When the strength of the same clause continues to increase in multiple consecutive windows, mark it as a continuous traction node and record the growth rate sequence. S6. Check whether the increase in traction intensity of the continuous traction node exceeds the upper limit of traction intensity fluctuation based on the historical veto frequency. If it exceeds, it is determined that there is a risk of clause-induced review deviation. S7. Extract the clause node content, supplier scheme identifier and associated review entity identifier corresponding to the clause-induced review deviation risk, and generate an electronic bidding risk warning to push to the supervision terminal.

[0019] In S1, basic risk weights are generated.

[0020] When generating a project complexity level based on the size of the supplier proposal documents and the number of review clauses, all supplier proposal documents corresponding to the current electronic bidding project are read. The length of the effective technical description content in each proposal document is counted, and table of contents pages, duplicate attachment pages, and fixed template pages are filtered out, retaining only the technical description content actually used for review. Subsequently, the number of review clauses used for scoring in the bidding documents is counted, and the clauses are classified and marked. Clauses involving cross-system interfaces, compatibility verification, technical parameter adaptation, and multi-module linkage requirements are marked as highly relevant clauses, while clauses involving only single function confirmation or basic qualification verification are marked as ordinary clauses. During the statistical process, highly relevant clauses are counted in the clause size statistics at twice the amount of ordinary clauses to avoid underestimating the complexity of large and complex projects due to the high degree of clause relevance. Finally, a project complexity label is generated based on the average size of the supplier proposal documents and the weighted number of review clauses. In practice, for example, projects with an average document length of less than 100 pages and fewer than 40 review clauses are classified as low complexity; projects with an average document length between 100 and 300 pages and between 40 and 100 review clauses are classified as medium complexity; and projects with an average document length of more than 300 pages or more than 100 review clauses are classified as high complexity. Each complexity level is expressed by a quantitative value between zero and one.

[0021] When extracting the records of rejection events, objections, and successful bidders for each clause node in the historical review archives, the process begins by searching for historical review clauses in historical projects that are identical or similar to the current clause, based on the clause node identifier. Synonymous technical terms in the clause descriptions are then grouped together; for example, "domestic database compatibility," "database adaptation capability," and "database interface compatibility requirements" are grouped into the same category of compatibility clauses to avoid discrepancies in historical statistical results due to differences in clause naming. Subsequently, the review result records in the corresponding historical projects are read. Review events where the clause directly causes the supplier to lose more than 70% of the full score for that clause are identified as rejection events, and these are accumulated to form the historical rejection frequency. For suppliers during the review process... If a written objection, review and reconsideration application, or regulatory complaint is submitted after the review, and the corresponding clause node is clearly mentioned in the review record, it will be identified as a challenge event and accumulated into historical challenge frequency. When calculating the supplier pass rate reduction ratio, first read the total number of suppliers participating in the bidding in the historical project, and then read the number of qualified suppliers that finally passed the technical review. The difference between the two is taken as the supplier pass rate reduction ratio as the proportion of the total number of suppliers participating in the bidding. For example, if ten suppliers participated in the bidding in a historical project, and only four of them finally passed the technical review, the corresponding supplier pass rate reduction ratio is recorded as 60%. Non-review elimination cases caused by suppliers voluntarily withdrawing from the bid, missing pre-qualification, or abnormal document uploads are not included in the supplier pass rate reduction ratio statistics.

[0022] When scaling the historical rejection frequency, historical challenge frequency, and supplier pass rate reduction ratio, the maximum value of the corresponding indicator is extracted from all historical clause nodes. Then, the historical rejection frequency, historical challenge frequency, and supplier pass rate reduction ratio for each clause node are divided by the maximum value of the corresponding indicator to generate the rejection weight, challenge weight, and reduction weight. In actual implementation, if the maximum historical rejection frequency for a certain type of clause is 20 times, and the current clause has a historical rejection frequency of 10 times, the corresponding rejection weight is recorded as 0.5. Then, normalized weighted fusion processing is performed on the three weights. In actual projects, the weights of the rejection weight, challenge weight, and reduction weight are set based on the weight ratio preset before project bidding, with the rejection weight weight higher than the other two to highlight the direct impact of the clause on the review results. After weighted fusion, the basic risk weight of the corresponding clause node is generated, and the basic risk weight is subjected to range restriction processing to constrain the final result to between zero and one.

[0023] In S2, the access behavior trajectory of the terms is generated.

[0024] For each supplier's solution text review process, when monitoring the focus operation of each clause node in the review interface, an independent focus status identifier is established for each review clause in the electronic review interface. When the reviewer enters the review area corresponding to a clause through mouse click, keyboard switching, scroll positioning, or page jump, and stays there for more than the set focus confirmation time, it is determined that the reviewer has formed a valid focus on that clause node. In actual implementation, the focus confirmation time is set to two seconds. Quick swipes or accidental touches less than two seconds are not counted as valid focus events to avoid distortion of behavior trajectory due to accidental stops during interface browsing. For clause nodes that include attachment viewing windows, parameter comparison windows, and technical response comparison windows, when the reviewer opens the corresponding associated window and the window content is bound to the current clause node, the focus status of the current clause node is maintained synchronously until the window is closed or the reviewer switches to another clause node, ending the current focus record. After verification, the system records the current supplier solution identifier, the current clause node identifier, the focus start time, and the identifier of the previous clause node in the focus state, and generates a corresponding focus switching event. If the reviewer continuously browses different technical parameter tabs within the same clause node without leaving the current clause area, a new focus switching event is not generated; instead, the focus duration of the current clause node is continued. If the reviewer frequently moves between two clause nodes within a short period, all focus switching records are retained, and a switching interval field is added to the records. When the interval between two adjacent switching events is less than five seconds, the corresponding behavior is marked as a high-frequency revisit behavior for subsequent identification of the clause node's continuous influence on the review process. In multi-window review scenarios, when the reviewer simultaneously opens multiple supplier solutions for comparison, the main interface window currently in the input focus state is used as the actual focus source, and browsing behavior in non-input focus windows is excluded.

[0025] When monitoring the scoring values ​​of each clause node, the scoring input area corresponding to each clause node in the electronic review interface is read, or automatic scoring is performed based on the clause achievement level. A scoring status mapping relationship is established according to the supplier solution identifier, clause node identifier, and review subject identifier. During the actual review process, when the review subject performs add, modify, or overwrite operations on the scoring result of a supplier solution under the current clause node, the modified scoring value is read in real time, and the time of scoring change, the changed value, and the current clause node identifier are recorded simultaneously, generating a corresponding scoring modification event. For temporary input values ​​that the review subject has not yet confirmed during the scoring process, they are not directly included in the scoring modification event. Only after the scoring box loses its input focus, the confirmation button is clicked, or the page is saved is the current scoring result recognized as a valid scoring modification result, avoiding a large number of invalid scoring modification records due to accidental deletion, temporary input, or keyboard adjustments during the scoring input process. For multiple consecutive submissions by the same review subject within a short period of time... When adjusting the scoring results of the same clause node multiple times, all scoring change records are retained, and a scoring change direction indicator is added to the records. When the scoring results change continuously in the same direction, it is marked as a continuous adjustment behavior. After collecting focus switching events and scoring modification events, the behaviors are aggregated according to the same review subject and the same supplier solution identifier, and sorted in ascending order according to the record time corresponding to each event to generate the clause access behavior trajectory of the corresponding review subject. During the trajectory generation process, similar events that are repeated continuously with an interval of less than one second are merged. For example, two consecutive identical focus switching events are only recorded once to prevent abnormal expansion of the behavior trajectory due to interface refresh, network jitter, or repeated browser triggers. In the case of abnormal page closure, network reconnection, or review terminal switching during the review process, the first event after the abnormality is restored is marked as the trajectory restoration event, and the last valid event identifier before restoration is retained to maintain the continuity of the entire review behavior trajectory in the time dimension.

[0026] In S3, the synthetic traction strength.

[0027] In the access behavior trajectory of each review subject corresponding to the supplier solution identifier, when a review subject triggers a score modification event, the complete access behavior trajectory of that review subject in the current supplier solution is read, and the record position corresponding to the most recent score modification event is located. The time of the most recent score modification event is taken as the start time of the time segment, and the time of the current score modification event is taken as the end time of the time segment. Then, all focus switching events within the above time range are extracted from the behavior trajectory. In actual implementation, for the case of the first score modification event, the time when the review subject first enters the current supplier solution review interface is taken as the start time of the time segment. For instantaneous focus events with a continuous dwell time of less than two seconds within the time segment, they are not included in the subsequent clause node statistics to avoid invalid focus records caused by accidental touch switching, page shaking, or rapid scrolling. During the extraction process, the number of focus times corresponding to all clause nodes within the time segment is counted, and the historical data of the corresponding clause nodes in the historical projects are read simultaneously. The historical average number of focus counts is used to determine whether a clause node exhibits abnormal focus behavior within the current scoring adjustment period if the number of focus counts for a particular clause node in the current time segment exceeds the corresponding historical average number of focus counts. In actual projects, the historical average number of focus counts is statistically analyzed according to the project complexity level, and independent historical focus baselines are established for projects of different complexity levels to avoid misidentifying normal large-scale clause viewing behavior in complex projects as abnormal focus. Subsequently, clause nodes that meet the abnormal focus criteria are deduplicated to form a set of related clause nodes. For cases where the reviewer repeatedly visits the same clause node within a short period of time, the corresponding number of repeated visits is additionally recorded. When the number of repeated visits exceeds three, the clause node is marked as a potential leading node. For clause nodes that simultaneously involve attachment viewing, parameter comparison, and scoring adjustment behavior, the attachment viewing time and parameter comparison time are cumulatively included in the focus duration of the clause node to ensure that the set of related clause nodes can truly reflect the core clause content that the reviewer focuses on before the scoring is modified.

[0028] When traversing each clause node identifier in the associated clause node set, the system reads the clause access behavior trajectory corresponding to all review subjects in the current project and counts the cumulative number of repeated visits to the corresponding clause node in all behavior trajectories. During the statistics process, only access behaviors with an interval of more than three seconds between two adjacent visits are counted as repeated visits; repeated focus events caused by continuous refresh, page repaint, or browser automatic recovery are not included in the statistics. For cases where the same review subject switches between different supplier solutions and then returns to the current clause node, it is judged as a cross-solution review behavior and counted as twice the number of ordinary repeated visits in the cumulative repeated visits, in order to strengthen the ability to identify continuous comparison behaviors of key clauses. In actual projects, if a clause node shows high-frequency repeated visits among multiple review subjects, the subject coverage ratio of the corresponding clause node is calculated. For example, if there are five review subjects participating in the current project review, four of whom are reviewers... If all entities repeatedly access the same clause node, the corresponding entity coverage ratio is recorded as 80%. Then, the basic risk weight of the corresponding clause node is read, and risk correlation calculation is performed on the cumulative number of repeated accesses. In actual implementation, an access concentration coefficient is calculated based on the cumulative number of repeated accesses, and an entity diffusion coefficient is calculated based on the entity coverage ratio. The access concentration coefficient, entity diffusion coefficient, and basic risk weight are then normalized and merged to generate the clause traction strength of the corresponding clause node. For clause nodes with a large number of repeated accesses generated by only a single review entity, the contribution ratio corresponding to the entity diffusion coefficient is automatically reduced to avoid abnormally high clause traction strength due to individual review entity reading habits. The specific contribution ratio is pre-set based on historical data. Furthermore, for clause nodes with a high historical frequency of questioning but a low current entity coverage ratio, the influence of the basic risk weight is preferentially retained during the fusion calculation process.

[0029] In S4, the traction strength of the clause is weighted and corrected.

[0030] When extracting technical parameter description paragraphs and requirement response paragraphs from each supplier's proposal text, the process reads all proposal text files submitted by suppliers in the current project and locates the chapters according to the clause directory structure in the tender documents. For the technical parameter description paragraphs, priority is given to extracting the main text containing equipment parameters, interface indicators, compatibility descriptions, performance indicators, and system architecture descriptions. For the requirement response paragraphs, the response texts from suppliers addressing each requirement in the tender are extracted. During the extraction process, the table of contents, cover page, legal statement page, and duplicate attachment pages are filtered, retaining only the valid content actually used in the technical review. Subsequently, word segmentation is performed on each paragraph. In actual implementation, a pre-built word vector dictionary based on an industry dictionary and historical bidding corpus is used. The industry dictionary includes high-frequency technical terms in electronic bidding scenarios such as databases, middleware, servers, interface protocols, network security, and domestic adaptation. Synonymous technical expressions, such as "interface compatibility," "protocol adaptation," and "system compatibility support," are uniformly mapped to... To avoid semantic distance distortion caused by differences in word expression habits among different suppliers, similar word vector regions are used. After word segmentation, each word is converted into a word vector in the corresponding word vector dictionary, and all word vectors within the same paragraph are averaged and aggregated to generate the paragraph embedding vector. In actual projects, the word vector dimension is set to 128 dimensions, with different initial weights for industry technical terms and general descriptive terms. Technical parameter terms have a higher weight than general descriptive terms to highlight the differences in technical solutions. For templated descriptive statements that appear repeatedly in a paragraph, such as "fully meets the requirements" or "strictly follows the tender documents," a weight reduction process is performed before calculating the paragraph embedding vector to avoid incorrect compression of the solution text distance due to multiple suppliers copying a large number of templated expressions. For supplier solutions containing a large number of charts, parameter tables, or attached images, the key parameter fields in the corresponding charts are parsed into text before participating in the paragraph embedding vector generation to ensure that the semantic distance of the solution can truly reflect the degree of difference between the supplier's technical solutions.

[0031] When calculating the cosine distance between the embedding vectors of similar paragraphs from any two supplier proposals, the process first establishes a correspondence between paragraphs of the same type according to the technical parameter description paragraph and the requirement response paragraph. Semantic distance calculation is only allowed between paragraphs of the same type to avoid distortion of the distance results caused by mixing technical parameter content and requirement response content. Then, the cosine distance is calculated for the corresponding paragraph embedding vectors of any two supplier proposals, and the arithmetic mean of the cosine distances between all supplier pairs is taken to generate the text distance of the proposal corresponding to the current project. In actual implementation, if there are five suppliers in the project, the cosine distances between ten pairs of suppliers are calculated separately, and the ten results are taken. The average of the results is used as the final solution text distance. A small solution text distance indicates high semantic similarity among multiple supplier solutions, while a large solution text distance indicates significant differences in technical routes, parameter configurations, and demand response methods among supplier solutions. After generating the solution text distance, the current project complexity level is read and converted into a complexity metric value according to a preset mapping relationship. In actual projects, low-complexity, medium-complexity, and high-complexity projects are respectively assigned to complexity metric values. Subsequently, the historical challenge frequency of the corresponding clause nodes is read, and behavioral mitigation factors and risk enhancement factors are generated respectively. Among them, the solution text distance and complexity metric value together... When significant differences exist between proposed solutions and the project is highly complex, it indicates that the review body's repeated visits and score adjustments around clause nodes are normal review phenomena. Therefore, the behavior mitigation factor is used to attenuate the current clause's traction strength. In practice, the distance and complexity metric values ​​of the proposed solution texts are first normalized, and then their weighted average is calculated as the behavior mitigation factor. The current clause's traction strength is then divided by the behavior mitigation factor to generate the attenuated clause's traction strength. Subsequently, the frequency of historical objections is converted into a risk enhancement factor. For clause nodes with a high frequency of historical objections, the corresponding clause node's risk is increased. The risk enhancement ratio is determined, and the weakened clause traction strength is enhanced by using the risk enhancement factor. In actual implementation, the frequency of historical objections is normalized according to the highest objection frequency among all clause nodes to generate a risk enhancement ratio between zero and one. The weakened clause traction strength is then weighted and fused with the risk enhancement ratio to generate the final adjusted traction strength. For clause nodes with small text distances but consistently high historical objection frequencies, the contribution weight of the risk enhancement ratio is increased during the enhancement adjustment process to strengthen the risk identification capability of potential targeted screening clauses. All of the above weights are pre-set based on historical bidding records before the review begins.

[0032] In S5, the growth rate sequence is recorded.

[0033] When reviewing each clause node one by one at the end of the review time window, a fixed-duration review time window is divided according to the current project review progress. In actual implementation, each review time window is set to five minutes, and multiple window intervals are continuously divided according to the review start time. Then, the revised traction strength records corresponding to all clause nodes in the current time window are read and sorted according to the generation time. The latest generated revised traction strength value is extracted as the window traction strength of the corresponding clause node in the current window. During the extraction process, for clause nodes in the window that have not generated a new revised traction strength record, the window traction strength corresponding to the previous window is automatically inherited to avoid the interruption of window traction strength due to the review subject not accessing a clause for a short period of time. Then, the current window traction strength is compared with the window traction strength of the immediately preceding window. When the current window traction strength is greater than the previous window traction strength, the window traction strength is determined. When determining the traction strength, the window change flag for the corresponding clause node is set to increase. If the difference between the current window traction strength and the previous window traction strength is lower than the set minimum fluctuation value, the increase flag is not executed. In actual projects, the minimum fluctuation value is set to 5% of the historical average traction strength to filter out small fluctuations caused by normal reading behavior or short-term repeated access during the review process. If a clause node shows alternating increases and decreases between multiple adjacent windows, the number of consecutive increases is reset to zero, and only continuous and stable increases are retained. In large-scale centralized procurement projects, if the review subject frequently switches between multiple supplier proposals, causing rapid changes in traction strength within a local window in a short period of time, the traction strength of windows within two consecutive windows is smoothed, and the average value of adjacent windows is used as the final window traction strength to reduce the impact of short-term behavioral fluctuations on subsequent continuous traction judgments.

[0034] When continuously monitoring multiple review time windows, the window change indicators corresponding to each clause node are read in chronological order, and the number of consecutive increases is accumulated. In actual implementation, three consecutive increases are used as the criterion for a continuous traction node. That is, when the same clause node maintains an increasing window traction intensity in three consecutive review time windows, the clause node is marked as a continuous traction node. If there is a single instance of flat growth without a decrease during continuous growth, the continuous growth status is maintained to avoid interruption of continuous traction judgment due to short-term pauses in review operations within a local window. For different project complexity levels, the corresponding criteria for judging the number of consecutive increases are set independently. For example, low-complexity projects use two consecutive increases as the criterion, medium-complexity projects use three consecutive increases, and high-complexity projects use four consecutive increases to adapt to the differences in the duration of normal review behavior in different projects. After marking the continuous traction nodes, extract the window traction strength in the corresponding window according to the review time window order, and calculate the change in traction strength between adjacent windows to form a growth rate sequence. In actual projects, if the window traction strength corresponding to a certain clause node is 10, 15, 22, and 30 respectively, the corresponding growth rate sequences are recorded as 5, 7, and 8 respectively. For clause nodes with continuously expanding growth rates, an additional growth acceleration indicator is added to identify the continuous absorption trend of the clause node on the review process. For clause nodes with continuous growth rates but gradually decreasing rates, the growth contribution ratio in subsequent risk assessment is reduced to avoid some clause nodes that naturally cool down in the later stages of review from maintaining a high-risk state for a long time. At the same time, the continuous traction behavior of different supplier solutions sharing the same clause node is uniformly associated. When multiple supplier solutions simultaneously form a continuous traction phenomenon under the same clause node, the priority of the corresponding growth rate sequence in subsequent risk identification is increased.

[0035] In step S6, the risk of clause-induced review deviation is determined.

[0036] When extracting historical veto frequency records and corresponding historical traction strength records for continuous traction nodes in historical projects, the process begins by retrieving historical clause nodes in the historical review archive that are consistent with the current clause content or have similar technical semantics, based on the clause node identifier corresponding to the current continuous traction node. Clauses with different names but consistent actual technical constraints are grouped according to a unified clause category. For example, "Domestic Database Adaptation Requirements," "Database Compatibility Requirements," and "Heterogeneous Database Interface Support Requirements" are grouped into database compatibility clauses to avoid discrepancies in historical statistical results due to differences in clause naming across historical projects. Subsequently, the historical veto frequency records and historical traction strength records for the corresponding historical projects are read. The historical traction strength records are regenerated offline using the same clause traction strength calculation method as the current project to ensure consistency in statistical scope between historical and current project traction strength. After extracting historical records, the historical traction strength records are filtered based on historical veto frequency. In actual implementation, the historical veto frequency corresponding to the current continuous traction node is used as the basis for selection. Centered on this, a corresponding frequency filtering range is established. For example, if the historical rejection frequency of the current clause node is ten times, then historical project traction strength change records with a historical rejection frequency between eight and twelve times will be filtered. For clause categories with a small number of historical projects, the filtering range is automatically expanded to ensure that the normal traction strength change set contains at least twenty sets of valid historical records. For obviously abnormal historical projects, such as large-scale abnormal review events caused by regulatory notices, review suspensions, or concentrated complaints, the corresponding historical traction strength records are not included in the normal traction strength change set. During the filtering process, the project complexity level of the corresponding historical project is additionally read. When the difference between the historical project complexity level and the current project complexity level exceeds a set level, the corresponding historical traction strength change record participates in the subsequent set construction in a downweighted manner to avoid excessive differences in normal traction behavior between low-complexity projects and high-complexity projects, which would lead to distortion of the fluctuation range. Finally, all the filtered historical traction strength change records are arranged in the order of the review time window to form the normal traction strength change set of the corresponding clause node.

[0037] When performing fluctuation boundary fitting on the normal traction intensity change set, all historical growth rate records in the normal traction intensity change set are read, and corresponding historical growth trajectories are generated according to the review time window order. In actual implementation, a sliding window boundary fitting method is used to calculate the fluctuation boundary of the historical growth trajectory. The sliding window length is set to three consecutive review time windows. The maximum and minimum growth rates within the corresponding window range are statistically analyzed segment by segment, and the boundary results within all window ranges are smoothly connected to generate the normal traction fluctuation envelope of the corresponding clause node. For local abnormal growth behaviors in the historical growth trajectory that have a single sudden increase followed by a rapid decline, a neighborhood averaging method is used for smoothing correction during the fitting process to avoid local extreme values ​​causing abnormal expansion of the normal traction fluctuation envelope. After completing the generation of the normal traction fluctuation envelope, the current growth rate is calculated based on the growth rate sequence corresponding to the current continuous traction node, and the current growth rate is... The magnitude is mapped to the corresponding review time window position, and then it is checked whether the current growth magnitude exceeds the normal traction fluctuation envelope of the corresponding window position. In actual implementation, when the current growth magnitude exceeds the normal traction fluctuation envelope for two consecutive review time windows, it is determined that the corresponding clause node has experienced abnormal traction growth. For cases where only a single window shows a short-term excess, the observation status is maintained, and the risk is not immediately determined to avoid misidentifying normal growth behavior caused by short-term concentrated viewing, concentrated discussion, or technical clarification during the review process as review deviation risk. For cases where multiple review subjects simultaneously form continuous growth around the same clause node and the growth magnitude synchronously exceeds the normal traction fluctuation envelope, the corresponding risk level is increased. When the current growth magnitude exceeds the normal traction fluctuation envelope, it is determined that the corresponding clause node has experienced clause-induced review deviation risk, and the corresponding growth window position, excess magnitude, and the identification of the review subjects involved in forming the continuous traction behavior are recorded simultaneously.

[0038] In step S7, the content of the clause node corresponding to the clause-induced review deviation risk, the supplier's scheme identifier, and the associated review subject identifier are extracted, and an electronic bidding risk warning is generated and pushed to the supervision terminal.

[0039] When identifying risks based on the deviation between the growth rate corresponding to a continuous traction node and the normal traction intensity fluctuation envelope, the following steps are taken: First, the growth rate sequence corresponding to the current continuous traction node is read, and the excess ratio of the current growth rate exceeding the normal traction intensity fluctuation envelope in each review time window is calculated. In actual implementation, when the current growth rate is higher than the upper boundary of the normal traction intensity fluctuation envelope at the corresponding window position, the ratio of the difference between the excess part and the corresponding boundary value to the boundary value is taken as the degree of deviation. Then, historically identified cases of bid rigging, abnormally low-price bidding, and qualification fraud are read, and the clause-driven growth trajectory, the trajectory of changes in the focus of the review subject, and the similarity records of supplier proposal texts in the corresponding cases are extracted. Among them, in bid rigging cases, the focus is on extracting the continuous traction behavior of multiple supplier proposals forming synchronous growth under the same clause node; in abnormally low-price bidding cases, the focus is on extracting the clause nodes of low-priced suppliers that show rapid growth in the later stages of the review. The study focuses on identifying behavioral characteristics of rapid growth, particularly in cases of qualification fraud, by extracting key behavioral features of concentrated repeated visits and continuous score adjustments in qualification review clauses within a short period. After extracting historical cases, the growth trajectory corresponding to the current continuous growth node is matched with the risk behavior trajectory execution patterns in historical cases. In actual projects, when the current growth trajectory meets similar conditions in terms of continuous growth direction, growth rate, and growth duration as in historical cases, the current project is deemed to have a corresponding risk pattern. For cases where there is only partial similarity within a single time window but a significant difference in the overall growth trend, the risk is not directly determined to exist. When the current continuous growth node matches multiple types of historical risk case characteristics, the corresponding risk level is increased according to the number of risk overlaps. After completing the risk pattern matching, corresponding risk warning information is generated, including clause node identifiers, corresponding supplier solution identifiers, deviation degree, risk type, and the identifier of the review entity that formed the continuous growth behavior.

[0040] When generating tiered early warning messages based on the degree of deviation in risk warning information, the risk level is first divided according to the magnitude of the deviation. In actual implementation, risk behaviors with a deviation of less than 10% are marked as Level 1 warnings, those with a deviation between 10% and 30% are marked as Level 2 warnings, and those with a deviation of more than 30% are marked as Level 3 warnings. For composite risk behaviors that match both bid-rigging and fraudulent qualification cases, the risk level is automatically upgraded by one level. Subsequently, corresponding tiered early warning messages are generated based on the risk level. Level 1 early warning messages only include clause node identifiers, supplier solution identifiers, and deviation information. Level 2 early warning messages additionally include historical case matching results and corresponding growth rate sequences. Level 3 early warning messages further... The system adds identifiers for review entities involved in continuous risk management activities, risk formation time windows, and summaries of corresponding risk behavior trajectories. During the push process, Level 1 warning messages are pushed to the bidding management end, while Level 2 and Level 3 warning messages are simultaneously pushed to the procurement supervision end. For Level 3 warning messages, the manual review status is additionally recorded after push, and subsequent risk trajectory information is automatically added when the corresponding clause node continues to show continuous growth. If the same clause node triggers multiple warnings of the same level repeatedly within a short period of time, the warnings are merged, retaining only the latest risk behavior trajectory and cumulative deviation, thus avoiding a large number of duplicate warning messages on the procurement supervision end. At the same time, risk warning records that have been reviewed and confirmed to be normal review behaviors are marked, and the matching priority of the corresponding risk patterns is reduced during subsequent historical case updates.

[0041] Example 2 The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces an electronic bidding risk warning system based on dynamic weights.

[0042] Figure 2 A schematic diagram of a dynamic weight-based electronic bidding risk early warning system according to the present invention is provided. The dynamic weight-based electronic bidding risk early warning system includes: The data collection module is used to collect bidding documents, bidding activities, historical review cases, and market supervision information in a unified manner. The weight adjustment module is used to dynamically adjust the basic risk weight and clause traction strength calculation weight of the corresponding clause nodes based on the project complexity level, changes in the review stage, and the behavior of the review subjects. The traction analysis module is used to count the number of repeated visits to clause nodes based on the clause access behavior trajectory of each review subject, and generate the clause traction strength of the corresponding clause node. The risk identification module is used to identify risk patterns based on the deviation between the growth rate corresponding to the continuous traction node and the normal traction intensity fluctuation envelope, combined with historically identified cases of bid rigging, abnormally low-price bidding, and qualification fraud, and generate corresponding risk warning information. The early warning push module is used to generate corresponding graded early warning messages based on the degree of deviation in the risk warning information, and push them to the procurement supervision terminal and the bidding management terminal.

[0043] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0044] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0045] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0047] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0048] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0049] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0051] In conclusion, 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 dynamic weight-based electronic bidding risk early warning method, characterized in that, Includes the following steps: S1. During the review stage, analyze the clauses and nodes of the tender documents, the project complexity level and the text of each supplier's proposal, and count the historical rejection frequency, questioning frequency and supplier pass rate reduction ratio of each clause, and generate basic risk weights by weighting. S2. Real-time capture of the supplier's proposal during the bidding review process, including the focus switching and score modification operations of the review body on each clause node, and generate a clause access behavior trajectory in chronological order. S3. Obtain the access behavior trajectory of all supplier solutions for the terms, trace back the behavior trajectory to identify related terms nodes, count the number of repeated visits of each related node in the entire behavior trajectory, and synthesize the term traction strength with the basic risk weight. S4. Calculate the degree of difference between the texts of each supplier's proposal, integrate the project complexity level and the frequency of historical questioning of the clauses to form a behavior correction coefficient, and make a weighted correction to the traction strength of the clauses. S5. Set a review time window. When the window ends, calculate the window change of the traction strength of each clause after the revision. When the strength of the same clause continues to increase in multiple consecutive windows, mark it as a continuous traction node and record the growth rate sequence. S6. Check whether the increase in traction intensity of the continuous traction node exceeds the upper limit of traction intensity fluctuation based on the historical veto frequency. If it exceeds, it is determined that there is a risk of clause-induced review deviation. S7. Extract the clause node content, supplier scheme identifier and associated review entity identifier corresponding to the clause-induced review deviation risk, and generate an electronic bidding risk warning to push to the supervision terminal.

2. The dynamic weight-based electronic bidding risk early warning method according to claim 1, characterized in that, In step S1, generating the basic risk weights specifically includes: The project complexity level is a project complexity identifier generated based on the size of the supplier's proposal text and the number of review clauses. Each level corresponds to a different complexity quantification value. Extract the records of rejection events, objection events, and bidder approvals for each clause node in the historical review archives. Use the number of rejection event records as the historical rejection frequency and the number of objection event records as the historical objection frequency. Based on the total number of bidding suppliers under the constraints of the clause nodes, the decrease in the number of qualified suppliers after the review is calculated as the supplier pass rate reduction ratio. The frequency of historical rejections, the frequency of historical challenges, and the supplier pass rate reduction ratio are scaled down relative to their respective maximum values ​​across all clause nodes to obtain the rejection weight, challenge weight, and reduction weight. The normalized weighted sum of the rejection weight, challenge weight, and reduction weight is then calculated and used as the basic risk weight for the corresponding clause node.

3. The method for risk early warning in electronic bidding based on dynamic weights according to claim 1, characterized in that, In S2, the specific steps for generating the terms access behavior trajectory include: For each supplier's solution text review process, monitor the focus operation corresponding to each clause node in the review interface. When the review subject focuses on the clause node, record the corresponding clause node identifier, focus time, and the previous focused clause node identifier as a focus switching event. Monitor the score values ​​of each clause node, and when a change occurs, record the time of the change, the changed value, and the current clause node identifier as a score modification event; The focus switching events and score modification events generated by the same review subject are arranged in ascending order according to their recorded times, generating a trace of the clause access behavior of each review subject, which is grouped by the supplier solution identifier.

4. The method for risk early warning in electronic bidding based on dynamic weights according to claim 1, characterized in that, In S3, the traction strength of the synthetic clause specifically includes: In the clause access behavior trajectory of each review subject corresponding to the supplier solution identifier, when a review subject triggers a score modification event, the time segment between the previous score modification event and the current score modification event is extracted from the clause access behavior trajectory of the review subject. Extract all clause node identifiers whose number of focus counts exceeds a preset normal number threshold within a time segment, and form a set of related clause nodes after deduplication; Traverse the identifier of each clause node in the associated clause node set, retrieve the clause access behavior trajectory of all review subjects, and count the cumulative number of repeated visits to the corresponding clause node. The cumulative number of repeated visits to each associated clause node is correlated with the basic risk weight of the corresponding clause node to generate the clause traction strength of the corresponding clause node.

5. The method for risk early warning in electronic bidding based on dynamic weights according to claim 1, characterized in that, In S4, the weighted adjustment of the clause's traction strength specifically includes: Technical parameter description paragraphs and requirement response paragraphs are extracted from the solution texts of each supplier. Each paragraph text is segmented into words, and the words are converted into word vectors in a pre-set word vector dictionary. The average vector of word vectors in the paragraph is calculated as the paragraph embedding vector. Calculate the cosine distance between the embedding vectors of the same type of paragraphs from any two supplier proposals, and take the arithmetic mean of the cosine distances of all supplier pairs as the text distance of the proposals. Read the project complexity level, convert it into a complexity quantification value according to the preset mapping relationship, and read the historical questioning frequency of the corresponding clause node; By using the text distance and complexity quantification of the scheme as behavioral mitigation factors, the current clause traction strength is attenuated and corrected. The frequency of historical questioning is used as a risk enhancement factor to enhance and correct the attenuated clause traction strength, thus obtaining the corrected traction strength of the corresponding clause node.

6. The method for risk early warning in electronic bidding based on dynamic weights according to claim 1, characterized in that, In step S5, recording the growth rate sequence specifically includes: At the end of the review time window, each clause node is traversed one by one, and the latest generated value is extracted from all the revised traction strength values ​​in the window as the traction strength of the current window of the corresponding clause node. For each clause node, compare the current window traction strength with the traction strength of the immediately preceding window. If the current value is greater than the previous value, set the window change flag of the corresponding clause node to increase. Continuously monitor multiple review time windows. When the window change indicator of a clause node shows continuous growth and the number of consecutive growths reaches the preset number of consecutive growths, mark the corresponding clause node as a continuous traction node. Extract the traction intensity of each window in the window time sequence to form a growth rate sequence.

7. The method for risk early warning in electronic bidding based on dynamic weights according to claim 1, characterized in that, In S6, determining the risk of clause-induced review deviation specifically includes: Extract the historical veto frequency records and corresponding historical traction intensity records of the continuous traction nodes in historical projects; Based on the historical veto frequency, the historical traction intensity records are filtered by interval, and the historical traction intensity change records corresponding to the historical veto frequency of the current continuous traction node are extracted to form the normal traction intensity change set of the corresponding clause node. Perform fluctuation boundary fitting on the set of normal traction intensity changes to generate the normal traction degree fluctuation envelope for the corresponding clause node; Calculate the current growth rate corresponding to the continuous traction node based on the growth rate sequence, and check whether the current growth rate exceeds the normal traction fluctuation envelope of the corresponding clause node; When the current growth rate exceeds the normal traction fluctuation envelope, it is determined that there is a risk of clause-induced review deviation.

8. A dynamic weight-based electronic bidding risk early warning system, used to implement the dynamic weight-based electronic bidding risk early warning method according to any one of claims 1-7, characterized in that, include: The data collection module is used to collect bidding documents, bidding activities, historical review cases, and market supervision information in a unified manner. The weight adjustment module is used to dynamically adjust the basic risk weight and clause traction strength calculation weight of the corresponding clause nodes based on the project complexity level, changes in the review stage, and the behavior of the review subjects. The traction analysis module is used to count the number of repeated visits to clause nodes based on the clause access behavior trajectory of each review subject, and generate the clause traction strength of the corresponding clause node. The risk identification module is used to identify risk patterns based on the deviation between the growth rate corresponding to the continuous traction node and the normal traction intensity fluctuation envelope, combined with historically identified cases of bid rigging, abnormally low-price bidding, and qualification fraud, and generate corresponding risk warning information. The early warning push module is used to generate corresponding graded early warning messages based on the degree of deviation in the risk warning information, and push them to the procurement supervision terminal and the bidding management terminal.

Images