Project purchase risk intelligent early warning method and system based on knowledge graph

Abstract

The invention provides a project purchase risk intelligent early warning method and system based on a knowledge graph, which are suitable for purchase activities of various organizations such as enterprises, universities and public institutions and cover various purchase types such as equipment, medical consumables, software and information services, logistics materials, engineering construction and professional services. The method comprises the steps of collecting multi-source data in a project purchasing process and constructing a dynamic knowledge graph; identifying a hidden association relationship between suppliers by fusing the unstructured text data and the structured service data; detecting tendentiousness terms in the bid inviting file by adopting a semantic vector space mapping technology; a risk accumulation evaluation model is constructed based on historical performance data, and full-process risk dynamic early warning from supplier admission to performance is realized; wherein three links of hidden association relation identification, tendency term detection and risk accumulation evaluation are included; according to the method, the implicit association relationship between suppliers can be accurately identified, and dynamic risk early warning of the whole purchasing process is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of procurement risk early warning, in particular to a project procurement risk intelligent early warning method and system based on a knowledge graph, which is suitable for procurement activities of various types of organizations such as enterprises, colleges and universities, and institutions, and can be widely applied to risk identification and early warning of projects such as equipment, medical consumables, software and information services, logistics materials, engineering construction and professional services. BACKGROUND

[0002] In the field of project procurement risk supervision, the existing risk early warning system is limited to identifying the association relationship between suppliers, and the judgment is made through the equity proportion registered in the business or the holding information in the company charter. However, in actual procurement activities, there are often implicit control relationships between suppliers in the form of non-equity, such as family members cross-acting as senior managers, core employees flowing between multiple suppliers leading to business synergy, and capital association through proxy agreement to evade supervision. These implicit associations are difficult to be captured by traditional systems due to the lack of publicly registered equity ties, but they can become important carriers of bid rigging, interest delivery. The existing technology can only make subjective judgments by manually checking the similarity of the bid documents, the frequency of historical cooperation and other single-dimensional information, and lacks systematic fusion analysis of multi-source data such as personnel flow, fund flow and social security overlap, resulting in low accuracy and high omission rate in identifying implicit associations, and failing to timely warn potential collaborative irregularities, which seriously affects the fairness and transparency of the procurement process.

[0003] Based on the above problems, there is an urgent need for a technical solution that can accurately identify the implicit association relationship between suppliers and realize dynamic risk early warning in the whole procurement process. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art, and a project procurement risk intelligent early warning method and system based on a knowledge graph are proposed, which comprises: Collecting multi-source data in the project procurement process and constructing a dynamic knowledge graph; Identifying the implicit association relationship between suppliers by fusing unstructured text data and structured business data; Detecting the bias clauses in the bidding documents using semantic vector space mapping technology; Building a risk accumulation evaluation model based on historical performance data to realize dynamic risk early warning in the whole process from supplier access to performance, and the implicit association relationship identification, bias clause detection and risk accumulation evaluation realize real-time updating of risk value through dynamic weight transmission; The dynamic weight transmission dynamically adjusts the weight coefficients of each evaluation link based on the entity association strength in the knowledge graph.

[0005] Preferably, the multi-source data collected in the procurement process of the project procurement includes business registration information, social security participation records, senior management positions, patent feature data, judgment documents, bid documents and performance evaluation data, the multi-source data is processed by entity linking technology, and is converted into node and edge attributes of a knowledge graph through triple extraction, the triple includes a parameter item, a relationship and a characteristic value.

[0006] Further preferably, identifying the implicit association relationship between the suppliers includes: extracting the senior management cross-position frequency between the suppliers, the historical bid document text similarity, the geodesic distance of the registered address, the frequency of participating in third-party transactions together, the time series correlation of fund transactions, the stock holding risk coefficient and the social security participation personnel overlap degree, and the features are weighted calculated by an improved LabelPropagation algorithm, and the weight value of the weighted calculation is adjusted in real time by a knowledge graph community discovery algorithm.

[0007] Further preferably, detecting the biased clauses in the bid document includes: decomposing the technical parameter paragraph of the bid document into a feature vector, performing cosine distance calculation on the feature vector and a supplier patent feature vector, combining the weight proportion of the parameter in bid evaluation and the expression ambiguity index, and generating a semantic bias risk quantitative value, the expression ambiguity index is obtained by subject distribution analysis of the parameter text by an LDA topic model.

[0008] Further preferably, the quantitative calculation formula of the implicit association relationship is: ; Wherein: represents the monthly average frequency of cross-position of the senior management of the suppliers i and j; represents the cosine similarity of the historical bid documents of the suppliers i and j based on the BERT vector; represents the geodesic distance of the registered address of the suppliers i and j, and the unit is km; represents the frequency of participating in non-associated third-party transactions together by the suppliers i and j; represents the time series Pearson correlation coefficient of the fund transactions of the suppliers i and j; represents the stock holding risk coefficient of the suppliers i and j, which is obtained by standardizing the mentioning frequency in the judgment documents; represents the social security participation personnel overlap degree of the suppliers i and j; represents a dynamic weight coefficient, which satisfies , and is adjusted by a knowledge graph community discovery algorithm according to a real-time network structure.

[0009] Further preferably, the quantitative calculation formula of the semantic bias risk is: ; Wherein: represents the matching degree of the kth technical parameter with the specific supplier patent feature, with a value range of 0-1; represents the weight proportion of the kth technical parameter in bid evaluation; represents the fuzziness index of the kth technical parameter expression, which is obtained by calculating the topic distribution entropy value through the LDA topic model; represents the conventional value range of the kth technical parameter in the industry standard; represents the degree of deviation of the kth technical parameter from the industry conventional range; represents the historical cooperation anomaly frequency of the supplier and the purchaser.

[0010] Further preferably, the calculation formula of the comprehensive risk early warning value is: ; wherein: represents the implicit association strength index; represents the semantic bias risk degree; represents the default degree of the tth historical performance, with a value range of 0-1; represents the concealment score of the tth default, which is obtained by sentiment analysis of the words in the audit report; represents the time decay coefficient, with a value of 0.1 in the last 1 year, and increasing by 0.2 for each additional year; represents the number of performances in the last 3 years; represents the community tightness of the associated network in which the supplier is located, which is calculated by the Modularity algorithm.

[0011] A project procurement risk intelligent early warning system based on a knowledge graph, comprising a multi-modal data fusion module and a dynamic knowledge graph storage module, further comprising an implicit association mining module, a semantic bias detection module, a risk accumulation evaluation module, and an intelligent interception execution module, the output ends of the implicit association mining module are respectively connected to the input ends of the semantic bias detection module and the risk accumulation evaluation module, the output end of the semantic bias detection module is connected to the input end of the risk accumulation evaluation module, the output end of the risk accumulation evaluation module is connected to the input end of the intelligent interception execution module, and each module realizes real-time data interaction through a data bus.

[0012] Further preferably, the implicit association mining module comprises a feature extraction submodule, a weight adjustment submodule, and a community division submodule, the feature extraction submodule is used to extract multi-dimensional features such as the frequency of cross-appointment of senior executives between suppliers and the similarity of bid documents, the weight adjustment submodule dynamically generates weight coefficients using a knowledge graph community discovery algorithm, and the community division submodule divides the associated network by taking the overlap degree of social insurance participants as a constraint condition through an improved LabelPropagation algorithm.

[0013] Further preferably, the risk accumulation evaluation module comprises a historical data storage unit, a decay coefficient calculation unit and a linkage calculation unit, the historical data storage unit is used for storing the default degree and the concealment score of the performance in the past three years, the decay coefficient calculation unit dynamically generates a time decay coefficient according to the performance time, the linkage calculation unit receives the implicit association strength index output by the implicit association mining module and the semantic bias risk degree output by the semantic bias detection module, generates a comprehensive risk warning value through a preset algorithm and sends the comprehensive risk warning value to the intelligent interception execution module.

[0014] Technical effects: The present application breaks through the limitation of traditional dependence on equity information by fusing unstructured text and structured business data to identify the implicit association of suppliers, accurately captures the non-explicit control relationship such as cross-employment of senior managers and overlapping of social security, and realizes real-time updating of the whole process risk through dynamic weight transmission by combining semantic vector space mapping and risk accumulation evaluation, effectively solves the problems of insufficient identification of implicit association and high omission rate in the prior art, improves the accuracy and timeliness of the procurement risk early warning, and guarantees the fairness and transparency of the procurement process. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The project procurement risk intelligent early warning method based on the knowledge graph of the present application is a flow chart; Figure 2 The connection block diagram of the project procurement risk intelligent early warning system based on the knowledge graph of the present application is a connection block diagram; Figure 3 The overall architecture block diagram of the implicit association risk early warning and interception system of the present application is an overall architecture block diagram; Figure 4 The submodule composition block diagram of the implicit association mining module and the risk accumulation evaluation module of the present application is a submodule composition block diagram. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0017] In the description of the present application, it should be understood that the orientations or positional relationships indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0018] Traditional procurement risk early warning systems are limited to explicit equity association in association relationship identification and are difficult to capture non-controlling but actual control of implicit association; the detection of tender document bias clauses relies on keyword matching, and the identification of semantic ambiguity type is insufficient; risk assessment does not consider dynamic cumulative effect, resulting in low accuracy of whole-process risk early warning.

[0019] Based on this, please refer to Figure 1 The embodiment provides a project procurement risk intelligent early warning method based on a knowledge graph, comprising the following steps: Collecting multi-source data in the project procurement process and constructing a dynamic knowledge graph; Fusing unstructured text data and structured business data to identify the implicit association relationship between suppliers; Detecting the bias clauses in the tender document by using a semantic vector space mapping technology; Based on historical performance data, a risk accumulation evaluation model is constructed to realize whole-process risk dynamic early warning from supplier access to performance, and the identification of implicit association, the detection of bias clauses and the risk accumulation evaluation are realized through dynamic weight transmission to realize real-time updating of the risk value; The dynamic weight transmission is based on the dynamic adjustment of the weight coefficients of each evaluation link according to the entity association strength in the knowledge graph.

[0020] The core of the technical scheme is to construct a multi-dimensional linked risk early warning mechanism, which can be widely applied to procurement activities of enterprises, colleges and universities, and various types of organizations such as institutions, covering equipment, medical consumables, software and information services, logistics materials, engineering construction and professional services and other types of procurement. By collecting multi-source data such as business registration information and social security records, a dynamic knowledge graph is constructed to provide a data basis for implicit association identification. In the implicit association identification link, unstructured text data and structured business data are fused to break through the limitations of traditional equity association and capture implicit features such as cross-employment of senior managers and overlapping of social security personnel. The semantic vector space mapping technology is used to process the tender document, the technical parameter paragraph is disassembled into a feature vector, which is matched with the supplier patent feature vector, and the parameter weight and the ambiguity index are combined to realize accurate detection of semantic ambiguity type bias clauses.

[0021] Based on historical performance data, a risk accumulation evaluation model is constructed, and the implicit association identification result, the bias clause detection result and the historical performance risk are calculated through a dynamic weight transmission mechanism, wherein the dynamic weight is adjusted in real time according to the entity association strength in the knowledge graph, so that the risk value can be optimized in real time with the updating of each link data, forming a whole-process dynamic early warning closed loop from access to performance. The scheme solves the problems of insufficient identification of implicit association and semantic bias in traditional systems and static risk assessment, realizes whole-process risk dynamic early warning, and improves the accuracy and timeliness of early warning.

[0022] In the prior art, the data collection of project procurement risk early warning is limited to single type or structured data, the data processing method is simple, and it is difficult to effectively integrate multi-source heterogeneous data into a knowledge graph, resulting in incomplete node and edge attributes of the knowledge graph, affecting the comprehensiveness of subsequent risk analysis.

[0023] Based on this, multi-source data in the project procurement process is collected, including business registration information, social security participation records, senior management positions, patent feature data, judicial documents, bidding documents and performance evaluation data. After processing by entity linking technology, the multi-source data is converted into node and edge attributes of the knowledge graph through triple extraction, and the triple includes parameter items, relationships and feature values.

[0024] The technical scheme focuses on the fusion processing of multi-source data, providing high-quality data support for knowledge graph construction. The collected data covers structured data such as business, social security, and senior management positions, as well as unstructured text data such as judicial documents and bidding documents, achieving comprehensive coverage of supplier information. Through entity linking technology, the same entity from different sources is associated and matched, eliminating data redundancy and ambiguity. The triple extraction technology is used to convert the processed data into a structured form of parameter items, relationships and feature values, where the parameter items can correspond to supplier attributes, the relationships can reflect the association between suppliers, and the feature values are specific numerical values or descriptions. These triples are directly converted into node and edge attributes of the knowledge graph, enabling the knowledge graph to fully present the multi-dimensional information and associated network of suppliers, providing a solid data foundation for subsequent implicit association identification and risk assessment.

[0025] The scheme realizes the effective integration of multi-source heterogeneous data, enriches the entity and relationship attributes of the knowledge graph, and provides comprehensive data support for subsequent risk analysis.

[0026] Traditional supplier association relationship identification methods only rely on equity information, and cannot identify non-equity implicit associations such as senior management cross-appointment and social security personnel overlap, making it difficult to discover risks such as bid rigging and bid fraud in a timely manner, affecting the fairness of the procurement process.

[0027] Based on this, the identification of implicit association relationships between suppliers includes: extracting the frequency of senior management cross-appointment between suppliers, the text similarity of historical bidding documents, the geodesic distance of registered addresses, the frequency of participating in third-party transactions together, the time series correlation of fund transactions, the risk coefficient of equity holding, and the overlap degree of social security participants. The improved LabelPropagation algorithm is used to calculate the weight of the features, and the weight value of the weighted calculation is adjusted in real time by the knowledge graph community discovery algorithm.

[0028] The technical scheme realizes accurate identification of the implicit association of the supplier by multi-dimensional feature extraction and intelligent algorithm fusion. Seven types of key features are extracted, covering personnel association, frequency of cross-appointment of senior executives, overlap of social security participants, business association, frequency of joint participation in third-party transactions, historical bid document text similarity, spatial association, registered address geodesic distance, capital association, time series correlation of capital flow, and potential risk association, stock ownership holding risk coefficient, which comprehensively covers the scenarios that may exist implicit control relationship. The improved LabelPropagation algorithm is used to calculate the weight of these features. Compared with traditional algorithms, the improved algorithm takes the overlap of social security participants as an implicit constraint for community division, enhancing the sensitivity to personnel implicit association. The weight value of the weighted calculation is adjusted in real time by the knowledge graph community discovery algorithm, and the weight proportion of each feature is dynamically optimized according to the real-time structure of the supplier association network, ensuring the adaptability of the identification of implicit association in different scenarios, and finally accurately outputting the implicit association strength between suppliers.

[0029] It is worth mentioning that the multi-source data collected in the process of bidding and procurement project includes business registration information, social security record, senior management position information, patent feature data, judgment documents, bid documents, performance evaluation data and company contact personal information, the company contact personal information includes name, ID number, mobile phone number, the multi-source data is processed by entity linking technology, and is converted into the node and edge attributes of the knowledge graph through triple extraction, the triple includes parameter item, relationship and characteristic value.

[0030] The scheme breaks through the limitations of traditional equity association identification, accurately captures multi-dimensional implicit association, effectively identifies the risk of bid rigging, and improves procurement supervision.

[0031] The existing technology detects the tendentious clauses in the bidding documents mainly by keyword matching, and cannot identify the semantic ambiguity type of partiality such as high consistency between technical parameters and specific supplier patent features, resulting in high missed detection rate of tendentious clauses and affecting the fairness of procurement.

[0032] Therefore, detecting the tendentious clauses in the bidding documents includes: decomposing the technical parameter paragraph of the bidding document into a feature vector, calculating the cosine distance with the supplier patent feature vector, combining the weight proportion of the parameter in bid evaluation and the expression ambiguity index to generate a semantic bias risk quantization value, and the expression ambiguity index is obtained by topic distribution analysis of the parameter text by LDA topic model.

[0033] The technical scheme realizes accurate detection of the semantic fuzzy type of the tendentious clause through semantic vector matching and multi-factor fusion. The technical parameter paragraph of the bidding document is structured and processed, and is decomposed into a feature vector containing technical indicators, value range and other information. At the same time, the supplier patent features are converted into corresponding feature vectors, and the matching degree of the two is calculated by cosine distance, and the correlation degree of the technical parameters and the specific supplier patent is quantified. The weight proportion of the parameter in the bid evaluation is introduced, the influence of the high weight parameter on the procurement result is highlighted, and the misjudgment caused by the matching of the low weight parameter is avoided. The parameter text is analyzed by the LDA topic model, the fuzzy expression index is calculated, and the fuzzy degree of the parameter expression is quantified, such as the tailor-made for a specific supplier. The matching degree, weight proportion and fuzzy index are fused and calculated to generate a semantic bias risk quantization value, realizing accurate identification and quantitative evaluation of the semantic fuzzy type of the tendentious clause.

[0034] The identification of the implicit association relationship between suppliers includes: extracting the frequency of cross-appointment of senior managers between suppliers, the similarity of historical bidding file texts, the geodesic distance of registered addresses, the frequency of participating in third-party transactions, the time series correlation of fund transactions, the stockholding risk coefficient, the overlap degree of social security participants, the repetition frequency of contact names between suppliers, the association frequency of contact ID numbers, and the association frequency of contact mobile phone numbers. The features are weighted and calculated by an improved project procurement LabelPropagation algorithm, and the weight value of the weighted calculation is adjusted in real time by a knowledge graph community discovery algorithm.

[0035] The scheme improves the identification ability of the semantic fuzzy type of the tendentious clause, reduces the missed detection rate, and ensures the fairness of the procurement process.

[0036] The traditional method lacks quantitative evaluation means for the implicit association relationship between suppliers, and cannot accurately measure the strength of non-equity association, resulting in strong subjectivity in judging the implicit control relationship and affecting the objectivity of risk early warning.

[0037] Therefore, the quantitative calculation formula of the implicit association relationship is: ; Wherein: represents the monthly average frequency of cross-appointment of senior managers between suppliers i and j; represents the cosine similarity of the historical bidding file based on the BERT vector between suppliers i and j; represents the geodesic distance between the registered addresses of suppliers i and j, in km; represents the frequency of non-associated third-party transactions participated in by suppliers i and j; represents the time series Pearson correlation coefficient of fund transactions between suppliers i and j; The risk coefficient of equity holding on behalf of suppliers i and j is obtained by standardization calculation based on the frequency of mention in court documents; This indicates the degree of overlap in social security participants between suppliers i and j; Represents dynamic weighting coefficients, satisfying The algorithm is adjusted by the knowledge graph community based on the real-time network structure.

[0038] This formula quantifies the supplier through weighted fusion of multi-dimensional features. and The formula identifies the strength of implicit connections between individuals, overcoming the limitations of traditional methods that rely solely on equity relationships. Its core design logic involves transforming intangible characteristics such as personnel, business, and financial connections into calculable values, and then dynamically weighting these characteristics to balance their contribution across different scenarios.

[0039] First item Focusing on the synergistic relationship between personnel, business operations, and space. Among these, Frequency of cross-appointment of senior executives and The product of the similarity of the tender documents reflects the possibility of business collaboration between the two parties through personnel overlap; divided by This introduces the spatial decay effect—the closer the physical distance, the greater the ease of implicit association, and the larger this value becomes. Dynamic weights The knowledge graph community discovers that algorithms are adjusted in real time, for example, in industries with frequent executive turnover. It will automatically increase the weight of personnel associations.

[0040] Second item Regarding the potential connection between business and funds. Frequency of joint participation in third-party transactions and The product of the correlation of fund transfers reflects the fund linkage formed between the two parties in their business cooperation; the denominator The risk coefficient for nominee shareholding mitigates the risk of misjudging related parties when nominee shareholding exists by introducing the negative impact of suspected nominee shareholding; if A higher value, such as frequent mentions of nominee shareholding in court documents, will lower the overall value of this item, preventing the misjudgment of false associations caused by nominee shareholding as genuine cooperation. Dynamic weighting In capital-intensive procurement scenarios, such as engineering project procurement, the weight of financial connections will be significantly increased and strengthened.

[0041] Third item Directly quantify the implicit connections among grassroots personnel by measuring the overlap of social security participants. Capture non-executive-level personnel mobility relationships, such as the same group of employees rotating between two suppliers. Dynamic weighting. It carries a higher weight in labor-intensive industries because the mobility of grassroots personnel in these industries has a more significant impact on relationships.

[0042] Three items passed The constraints form a dynamic balance, ensuring that the formula can adaptively adjust the priority of each feature under different industries and procurement scenarios, and the final output is... The value can be directly used to determine whether there is a hidden relationship between suppliers that is not controlled by a controlling shareholder but is actually controlled, providing a quantitative basis for identifying suspected bid rigging and collusion.

[0043] This technical solution constructs a quantitative evaluation formula for the strength of implicit relationships through multi-dimensional feature weighted fusion. The formula includes three core terms, quantifying the implicit relationships between suppliers from three dimensions: personnel-business relationship, capital-potential risk relationship, and implicit personnel relationship. In the first term... Frequency of cross-appointment of senior executives and Multiplying the similarity of tender documents reflects the collaborative relationship between personnel and business; dividing by... The distance between registered addresses, in kilometers, reflects the attenuation effect of spatial distance on the strength of the association; in the second item... Frequency of joint transactions and Multiplying the correlation between funds reflects the linkage between business and funds; the denominator... The risk coefficient for nominee shareholding weakens the misjudgment of related parties when there is suspicion of nominee shareholding; the third item Social security overlap directly quantifies the implicit connections among grassroots personnel.

[0044] The three items are weighted by dynamic coefficients. The fusion and weighting of features are adjusted by the knowledge graph community discovery algorithm based on the real-time network structure to ensure that the contribution of each dimension of features is reasonable in different scenarios, and the final output is... The value can accurately quantify the strength of implicit associations.

[0045] This formula enables the objective quantification of the strength of implicit associations, providing accurate data support for association risk assessment and improving the objectivity of risk judgment.

[0046] Traditional technical solutions lack a systematic approach to quantifying the risk of semantic bias in bidding documents. They rely solely on keyword matching or subjective judgment, failing to comprehensively assess multiple factors such as the matching degree of technical parameters, evaluation weights, and ambiguity of expression, resulting in inaccurate risk quantification results.

[0047] Based on this, the quantitative calculation formula for semantic bias risk is as follows: ; in: This represents the degree of matching between the k-th technical parameter and the patent features of a specific supplier, with a value ranging from 0 to 1. This indicates the weight percentage of the k-th technical parameter in the bid evaluation; The fuzziness index of the kth technical parameter expression is obtained by calculating the topic distribution entropy value through the LDA topic model. The kth technical parameter represents the conventional value range in the industry standard. The kth technical parameter represents the degree of deviation from the industry conventional range. The kth technical parameter represents the historical cooperation abnormal frequency between the supplier and the purchaser.

[0048] The formula is aimed at the problem of semantic fuzzy type bias in the tender document, and through the multi-factor fusion of technical matching degree, evaluation weight and expression fuzziness, the precise quantification of the inclination clause is realized. The core of the formula design is to convert the text semantic features into mathematical parameters to avoid subjective judgment bias.

[0049] Molecular part The weighted semantic matching sum reflects the association degree of the technical parameter and the specific supplier. Among them, The technical parameter and the matching degree of the supplier's patent are quantified as the consistency of the feature, The parameter weight in the evaluation highlights the influence of key parameters, The expression fuzziness index amplifies the wind of fuzzy expression; for example, when the parameter description is to adopt the industry leading technology, The value increases, and the overall value increases.

[0050] The accumulation of the product of the three ensures that the parameter combination with high matching degree, high weight and high fuzziness is identified. The denominator part The conventional deviation normalization term is used to correct the influence of the parameter deviation from the industry convention. The product of the conventional range of the industry standard and The degree of deviation quantifies the abnormal degree of the parameter; the square root processing avoids the interference of individual extreme values on the whole, ensuring that the denominator will not be amplified too much due to the large deviation of a single parameter, thereby balancing the semantic matching risk of the numerator.

[0051] The formula is finally multiplied by , wherein The historical cooperation abnormal frequency introduces the influence of historical behavior - if the supplier has multiple abnormal cooperation records with the purchaser, this item will amplify the semantic bias risk, realizing the linkage evaluation of historical behavior and current clauses.

[0052] For example, When, , the risk value is significantly higher than that without abnormal records.

[0053] Overall, the formula converts the semantic bias risk into a comparable numerical value by the logic of associating the numerator, correcting the denominator, and supplementing the history, solves the problem of insufficient identification of fuzzy expressions by traditional keyword matching, and provides an objective basis for tender document compliance review.

[0054] The technical solution realizes precise evaluation of the inclined clause by constructing a semantic bias risk quantification formula through multi-dimensional parameter fusion. The formula molecule part is a weighted summation term, which multiplies and accumulates the matching degree of each technical parameter with the supplier's patent , the weight proportion of the parameter in the bid evaluation , and the expression fuzziness index , considering both the matching degree of technical features and the influence of high-weight parameters, while incorporating the factor of expression fuzziness.

[0055] The denominator part is a normalization term, which is the square root of the product of the industry conventional value range and the deviation degree , which weakens the interference of parameters with small deviation from the conventional range on the overall risk, and avoids misjudgment due to normal parameter fluctuations. Finally, multiply by , which introduces the influence of the historical cooperation abnormal frequency , and amplifies the risk of related clauses for suppliers with bad cooperation records. Each parameter is a dimensionless proportion or index, and the semantic bias risk can be converted into a quantifiable value through the formula to realize objective evaluation.

[0056] The scheme realizes precise quantification of semantic bias risk, avoids subjective judgment bias, and provides an objective basis for tender document compliance review.

[0057] In the traditional technical solution, the supplier's implicit association, semantic bias of the tender document, and historical performance risk are not analyzed in conjunction, and the time accumulation effect of the risk is not considered, resulting in incomplete comprehensive risk assessment and inability to reflect the dynamic risk of the whole process.

[0058] Therefore, the calculation formula of the comprehensive risk early warning value is: ; Wherein: represents the implicit association strength index; represents the semantic bias risk degree; represents the degree of breach of the tth historical performance, with a value range of 0-1; represents the concealment score of the tth breach, obtained by word sentiment analysis of the audit report; represents the time decay coefficient, with a value of 0.1 in the last 1 year, increasing by 0.2 for each additional year; represents the number of performances in the last 3 years; ​​​​​Community tightness, which represents the community tightness of the associated network where the supplier is located, is calculated by the Modularity algorithm.

[0059] This formula is the progression and integration of the previous two formulas. Through the whole-process risk linkage of implicit association, semantic bias, and historical performance, it realizes dynamic comprehensive evaluation of procurement risk, and solves the problem of risk fragmentation in traditional evaluation and the lack of time dimension consideration.

[0060] First term of the formula For the current stage collaborative risk, the square root form is used to balance the implicit association strength index and the influence of semantic bias risk degree .

[0061] For example, if , , this term is ; if one of them is 0.2 and the other is 0.8, this term is , avoiding the dominance of single-stage risk over the overall evaluation, and reflecting the synergistic effect of association and clauses.

[0062] Middle term is the historical performance risk accumulation term, which introduces a time decay mechanism to handle the impact of historical default. The product of the degree of default and the hidden score quantifies the actual risk of a single default; The timeliness of risk is reflected in the default in the past year , the decay coefficient is , and the default three years ago , the decay coefficient is

[0063] , ensuring that recent defaults contribute more to risk. After accumulation, add 1 to avoid the term being 0 when there is no historical default, ensuring the reasonableness of the base risk value. Finally multiply community tightness, introducing the structural influence of the associated network where the supplier is located. Through the Modularity algorithm, it reflects the depth of the supplier's embedding in the associated network; the tighter the network, the higher the value, and the overall risk value is amplified.

[0064] For example, a supplier in a tightly associated network , its comprehensive risk is significantly higher than that of an isolated supplier , reflecting the transmission of individual risk and network risk.

[0065] Overall, the formula realizes the whole-process risk coverage from access to bidding to performance through the three-step progressive logic of current collaborative risk, historical cumulative risk and network structure risk, and each parameter is dynamically adjusted in real time with data updates, such as refreshing every hour 、 and , ensuring that the output value can accurately reflect the real-time risk status of the supplier, providing a quantitative basis for intelligent interception and dynamic management.

[0066] The technical solution constructs a comprehensive risk early warning formula through multi-dimensional risk linkage and time decay mechanism. The first term of the formula is the collaborative risk of implicit association and semantic bias, which adopts square root form to balance the influence of the two and avoid excessive weight of a single factor.

[0067] The middle term is the historical performance risk accumulation term, which multiplies the degree of default and the concealment score , and then multiplies the time decay coefficient , where is the time from now, in years, so that the weight of recent default is higher, in line with the risk timeliness law. After adding 1 to the cumulative value, the basic risk value is ensured to be non-zero.

[0068] Finally, multiply by the community tightness to reflect the tightness of the associated network where the supplier is located. The tighter the network, the greater the risk value. Among the parameters, the unit is years, and the rest are dimensionless indices or proportions. Through the formula, multi-link risk can be integrated, and a dynamically updated comprehensive risk value can be output.

[0069] The scheme realizes comprehensive assessment of whole-process risk, reflects dynamic changes in risk, and improves the comprehensiveness and accuracy of early warning.

[0070] The module settings of traditional procurement risk early warning systems are scattered, and there is a lack of effective data interaction and linkage mechanism between functional modules, which cannot realize the collaborative work of implicit association recognition, semantic bias detection and risk assessment, resulting in low overall early warning efficiency of the system.

[0071] Based on this, please refer to Figure 2The embodiment provides a project procurement risk intelligent early warning system based on a knowledge graph, which comprises a multi-modal data fusion module and a dynamic knowledge graph storage module, further comprises an implicit association mining module, a semantic bias detection module, a risk accumulation evaluation module and an intelligent interception execution module, the output ends of the implicit association mining module are connected with the input ends of the semantic bias detection module and the risk accumulation evaluation module respectively, the output end of the semantic bias detection module is connected with the input end of the risk accumulation evaluation module, the output end of the risk accumulation evaluation module is connected with the input end of the intelligent interception execution module, and each module realizes real-time data interaction through a data bus.

[0072] The technical scheme realizes intelligent early warning of the whole process of procurement risks by constructing a system architecture with multiple modules in cooperation.

[0073] The implicit association mining module extracts supplier association features from the knowledge graph, calculates an implicit association strength index, and sends the index to the semantic bias detection module and the risk accumulation evaluation module respectively.

[0074] The risk accumulation evaluation module sends the comprehensive risk value to the intelligent interception execution module to trigger early warning or interception actions.

[0075] The system realizes the cooperative work of each module, improves the integrity and efficiency of risk early warning, and ensures the controllability of risks in the whole process of procurement.

[0076] In the traditional system, the function of the implicit association mining module is single, only limited association features can be extracted, and the weight is fixed, which cannot be dynamically adjusted according to the network structure of the supplier, so that the precision of the implicit association recognition is insufficient.

[0077] Therefore, please refer to Figure 3The implicit association mining module comprises a feature extraction submodule, a weight adjustment submodule, and a community division submodule. The feature extraction submodule is used to extract multi-dimensional features such as the frequency of cross-appointment of senior managers between suppliers and the similarity of bid documents. The weight adjustment submodule dynamically generates weight coefficients by using a knowledge graph community discovery algorithm. The community division submodule divides the association network by taking the overlap degree of social insurance participants as a constraint condition by using an improved LabelPropagation algorithm.

[0078] The technical solution improves the depth and accuracy of implicit association mining through the cooperative work of the three submodules. The feature extraction submodule comprehensively covers multi-dimensional features such as personnel, business, and space. In addition to the frequency of cross-appointment of senior managers and the similarity of bid documents, it also includes the distance of registered addresses, the correlation of fund flows, and the like, to ensure that no potential association clues are missed. The weight adjustment submodule dynamically generates weight coefficients for each feature by analyzing the topological structure of the supplier association network in real time based on the knowledge graph community discovery algorithm. For example, in a network with a high degree of personnel association, the weight of the overlap degree of social insurance participants is increased, and in a scenario with prominent business association, the influence of the frequency of common transactions is enhanced, thereby avoiding the lack of scene adaptability caused by fixed weights. The community division submodule uses an improved LabelPropagation algorithm to take the overlap degree of social insurance participants as a hard constraint condition, and when the community is divided, it preferentially classifies suppliers with a high degree of social insurance overlap into the same community, thereby solving the problem that traditional algorithms are not sensitive to the implicit association of grassroots personnel, and finally outputting a precise supplier association network.

[0079] The implicit association mining module comprises a feature extraction submodule, a weight adjustment submodule, and a community division submodule. The feature extraction submodule is used to extract multi-dimensional features such as the frequency of cross-appointment of senior managers between suppliers, the similarity of bid documents, the repetition frequency of contact names between suppliers, the association frequency of contact ID numbers, and the association frequency of contact mobile phone numbers. The weight adjustment submodule dynamically generates weight coefficients by using a knowledge graph community discovery algorithm. The community division submodule divides the association network by taking the overlap degree of social insurance participants as a constraint condition by using an improved project procurement LabelPropagation algorithm.

[0080] The module improves the comprehensiveness and scene adaptability of implicit association recognition and provides precise association data for risk assessment.

[0081] The traditional risk accumulation assessment module lacks systematic processing of historical performance data and does not consider the time decay effect of default and the linkage with real-time risks, resulting in a large deviation between the risk accumulation assessment result and the actual risk.

[0082] Based on this, please refer to Figure 4The risk accumulation evaluation module comprises a historical data storage unit, a decay coefficient calculation unit and a linkage calculation unit. The historical data storage unit is used to store the default degree and concealment score of performance in the past three years. The decay coefficient calculation unit dynamically generates a time decay coefficient according to the performance time. The linkage calculation unit receives the implicit association strength index output by the implicit association mining module and the semantic bias risk degree output by the semantic bias detection module, generates a comprehensive risk warning value through a preset algorithm and sends it to the intelligent interception execution module.

[0083] The technical scheme realizes dynamic accumulation evaluation of risks through the cooperation of three units. The historical data storage unit specially stores the performance data in the past three years, including the default degree and concealment score of each performance, to provide a data basis for risk accumulation. The decay coefficient calculation unit dynamically generates a decay coefficient according to the performance time. The default in the past one year is given a higher coefficient, and the coefficient decreases over time, which is consistent with the rule that the influence of risks decreases over time, avoiding treating the long-term slight default and the recent serious default equally.

[0084] The linkage calculation unit, as the core, receives the implicit association strength index and the semantic bias risk degree, combines the default data of the historical data storage unit and the output of the decay coefficient calculation unit, performs fusion calculation through a preset comprehensive risk formula, generates a final comprehensive risk warning value, and sends it to the intelligent interception execution module in real time, to ensure that the risk evaluation result can trigger a response action in time. The module realizes linkage evaluation of historical and real-time risks, and improves the accuracy and timeliness of risk accumulation evaluation.

[0085] An AI project procurement auxiliary supervision submodule is added between the risk accumulation assessment module and the intelligent interception execution module, and the submodule includes a tender document intelligent review unit, an automatic evaluation unit, a blockchain storage unit and a fund penetration monitoring unit. The tender document intelligent review unit automatically screens the restrictive clauses of the qualification requirements and business conditions in the tender document by constructing a standard clause database covering various industries, calculates the deviation degree of the clauses from the industry's conventional requirements by combining semantic vector space mapping technology, generates a compliance detection report and marks high-risk clauses; the automatic evaluation unit connects the supplier's qualification, performance and credit data in the dynamic knowledge graph to realize the AI project procurement automatic evaluation and scoring of the tenderer's qualification verification, performance award matching, financial index calculation and credit record inquiry, and directly output the preliminary bid recommendation for small-scale and general-purpose project procurement projects, greatly shortening the evaluation period; the blockchain storage unit uploads key data such as the bid specification, bid process score and voting result to the government blockchain platform in real time to ensure that the bid information is tamper-proof and traceable throughout the process, and supports the supervision channel connection with the discipline inspection and supervision and judicial departments; the fund penetration monitoring unit connects the financial fund supervision system to track the account fund flow direction of the bid winner, subcontracting unit and material supplier in real time, identify fund abnormal transfer and interception and trigger early warning, and strengthen the government investment project fund supervision.

[0086] At the same time, an anonymous bid evaluation support function is added in the implicit correlation mining module, the identity information of the bidder is hidden by desensitizing the feature vector of the bid document, and the correlation network divided by the knowledge graph community discovery algorithm is combined to realize the random extraction of bid evaluation experts and the random allocation of bid evaluation seats, avoiding the interference of experts and bidders due to implicit correlation. In the semantic bias detection module, the AI project procurement auxiliary bid cleaning function is supplemented, which automatically identifies missing items in the bid and serious unbalanced bidding, generates a bid cleaning report by combining the cost indicators in the historical performance data, provides an objective reference for the bid committee, and further improves the scientificity and rationality of the bid decision.

[0087] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments within the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A knowledge graph-based project procurement risk intelligent early warning method, characterized in that, include: Collect multi-source data during the project procurement process and construct a dynamic knowledge graph; Identify implicit relationships between suppliers by integrating unstructured text data with structured business data; Semantic vector space mapping technology is used to detect biased clauses in bidding documents; A risk accumulation assessment model is built based on historical performance data to achieve dynamic early warning of risks throughout the entire process from supplier access to performance. The identification of implicit relationships, detection of biased clauses and risk accumulation assessment are achieved through dynamic weight transmission to update risk values ​​in real time. The dynamic weight transmission is based on dynamically adjusting the weight coefficients of each evaluation stage according to the entity association strength in the knowledge graph. 2.The knowledge graph-based project procurement risk intelligent early warning method according to claim 1, characterized in that, The multi-source data collected during the project procurement process includes business registration information, social security participation records, senior management appointment information, patent feature data, court documents, tender documents and performance evaluation data. After being processed by entity linking technology, the multi-source data is extracted into node and edge attributes of a knowledge graph through triple extraction. The triple includes parameter items, relations and feature values. 3.The knowledge graph-based project procurement risk intelligent early warning method according to claim 1, characterized in that, Identifying implicit relationships between suppliers includes: extracting the frequency of cross-appointment of senior executives among suppliers, text similarity of historical tender documents, geodesic distance of registered addresses, frequency of joint participation in third-party transactions, time series correlation of fund transfers, risk coefficient of nominee shareholding, and overlap of social security participants. The improved LabelPropagation algorithm is used to perform weighted calculations on the features, and the weight values ​​of the weighted calculations are adjusted in real time by the knowledge graph community discovery algorithm. 4.The knowledge graph-based project procurement risk intelligent early warning method according to claim 1, characterized in that, The detection of biased clauses in bidding documents includes: decomposing the technical parameter paragraphs of the bidding documents into feature vectors, calculating the cosine distance between these vectors and the feature vectors of the supplier's patents, and combining the weight of the parameters in the bid evaluation with the ambiguity index of the expression to generate a semantic bias risk quantification value. The ambiguity index of the expression is obtained by performing topic distribution analysis on the parameter text using the LDA topic model. 5.The knowledge graph-based project procurement risk intelligent early warning method according to claim 3, characterized in that, The quantitative calculation formula of the implicit correlation relationship is: ; wherein: represents the monthly average frequency of executive cross-appointment between suppliers i and j; represents the cosine similarity of the BERT vector-based historical bidding files of suppliers i and j; represents the geodesic distance between the registered addresses of suppliers i and j, in km; represents the frequency of non-associated third-party transactions in which suppliers i and j participate together; represents the Pearson correlation coefficient of the time series of the financial transactions between suppliers i and j; represents the stockholding risk coefficient of suppliers i and j, calculated by normalizing the frequency of mentions in judicial documents; represents the overlap of social security participants between suppliers i and j; represents the dynamic weight coefficient, satisfying adjusted by the knowledge graph community discovery algorithm according to the real-time network structure. 6.The knowledge graph-based project procurement risk intelligent early warning method according to claim 4, characterized in that, The quantitative calculation formula of semantic bias risk is: ; wherein: represents the matching degree of the kth technical parameter with the specific supplier patent feature, and the value range is 0-1; represents the weight proportion of the kth technical parameter in the bid evaluation; represents the fuzziness index of the kth technical parameter expression, which is obtained by calculating the topic distribution entropy value through the LDA topic model; represents the conventional value range of the kth technical parameter in the industry standard; represents the degree of deviation of the kth technical parameter from the industry conventional range; represents the historical cooperation anomaly frequency of the supplier and the purchaser. 7.The knowledge graph-based project procurement risk intelligent early warning method according to claim 1, characterized in that, The formula for calculating the comprehensive risk early warning value is: ; Wherein: represents the implicit association strength index; represents the semantic bias risk degree; represents the default degree of the tth historical performance, the value range is 0-1; represents the concealment score of the tth default, obtained by analyzing the sentiment of the audit report; represents the time decay coefficient, the value is 0.1 in the last 1 year, and increases by 0.2 every 1 year; represents the number of performance in the last 3 years; represents the community tightness of the association network where the supplier is located, which is calculated by the Modularity algorithm.

8. A knowledge graph-based project procurement risk intelligent early warning system applied to the knowledge graph-based project procurement risk intelligent early warning method of any one of claims 1-7. include: The multimodal data fusion module and dynamic knowledge graph storage module are characterized by further including an implicit association mining module, a semantic bias detection module, a risk accumulation assessment module, and an intelligent interception execution module. The output of the implicit association mining module is connected to the input of the semantic bias detection module and the risk accumulation assessment module, respectively. The output of the semantic bias detection module is connected to the input of the risk accumulation assessment module, and the output of the risk accumulation assessment module is connected to the input of the intelligent interception execution module. The modules achieve real-time data interaction through a data bus. 9.The knowledge graph-based project procurement risk intelligent early warning system according to claim 8, characterized in that, The latent association mining module includes a feature extraction submodule, a weight adjustment submodule, and a community partitioning submodule. The feature extraction submodule is used to extract multi-dimensional features such as the frequency of cross-appointment of senior executives among suppliers and the similarity of tender documents. The weight adjustment submodule uses a knowledge graph community discovery algorithm to dynamically generate weight coefficients. The community partitioning submodule uses an improved LabelPropagation algorithm to partition the association network by taking the overlap of social security participants as a constraint. 10.The knowledge graph-based project procurement risk intelligent early warning system according to claim 8, characterized in that, The risk accumulation evaluation module comprises a historical data storage unit, a decay coefficient calculation unit and a linkage calculation unit, the historical data storage unit is used for storing the default degree and the concealment score of the performance in the past three years, the decay coefficient calculation unit dynamically generates a time decay coefficient according to the performance time, the linkage calculation unit receives the implicit association strength index output by the implicit association mining module and the semantic bias risk degree output by the semantic bias detection module, generates a comprehensive risk early warning value through a preset algorithm and sends the comprehensive risk early warning value to the intelligent interception execution module.

Images