Dynamic association network-based surrounding string prediction method, system and terminal device

Abstract

This invention belongs to the field of bidding risk identification technology, specifically disclosing a method, system, and terminal device for predicting bid rigging based on dynamic association networks. The method includes: constructing a dynamic graph structure model; acquiring bid document text data, business registration information data of bidding companies, and historical bidding behavior data; extracting semantic features of the bid document text, company attribute features, and behavioral temporal features to form feature vectors for each node; simultaneously calculating the weight of each edge in the graph structure model; updating the graph structure model; using the updated graph structure model as input, utilizing a graph neural network model to obtain node embedding vectors, and calculating the association degree matrix between companies based on the node embedding vectors, thereby predicting whether bid rigging exists between companies. This invention comprehensively considers multi-source data features such as text semantics, company attributes, and behavioral temporal features, fully capturing the implicit relationships between companies and improving the accuracy of identifying bid rigging.

Description

Technical Field

[0001] This invention relates to the field of bidding risk identification technology, and in particular to a method, system and terminal device for predicting bid rigging based on dynamic correlation networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Identifying bid-rigging and collusion during the bidding process has always been a hot topic in the field of bidding research. During the bidding process, some suppliers may engage in unfair competition such as bid-rigging or collusion to increase their chances of winning the bid and exclude other competitors. Bid-rigging or collusion (hereinafter referred to as bid-rigging) refers to the act of multiple bidders conspiring privately to manipulate the bidding results through pre-negotiated pricing or other means; this behavior is usually intended to ensure that a specific bidder wins the bid, undermines fair competition, and constitutes illegal and irregular conduct.

[0004] In traditional auditing processes, the analysis of bidding information often relies on the subjective experience of the auditors, resulting in outdated auditing methods and low identification efficiency.

[0005] Existing technologies also disclose some methods for automatically identifying bid-rigging behavior, such as:

[0006] By constructing a weighted network through the co-occurrence relationship between "enterprises and bidding projects" and using the number of joint bids as the weight, the risk of bid rigging can be identified by identifying closely related groups; however, this relationship network structure is fixed and cannot dynamically reflect dynamic information such as changes in enterprise equity and qualifications.

[0007] Alternatively, the correlation index between enterprises can be determined by calculating the binary vector correlation coefficient of enterprise attribute information groups, and the related enterprises can be judged by manually preset attribute thresholds, thereby identifying the risk of collusion. However, this method cannot explore potential collaboration at the semantic level and is difficult to capture non-explicit correlations such as "similar schemes but different expressions", resulting in a high risk of missing implicit collusion. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a method, system, and terminal device for predicting bid rigging based on dynamic association networks. It introduces graph neural networks (GNNs) to dynamically update the weights of the association edges between enterprises and combines multi-source data (textual semantics, enterprise attributes, and behavioral time series) to capture potential relationship changes among bidding entities in real time, thereby improving the accuracy of identifying bid rigging behavior.

[0009] In some implementations, the following technical solutions are adopted:

[0010] A method for predicting scrambling based on dynamic association networks includes:

[0011] A dynamic graph structure model is constructed, in which enterprises are used as nodes, and edges are constructed based on the equity relationship, historical cooperation relationship, behavioral collaboration relationship and semantic collaboration relationship of the tender document text between nodes. Each edge is accompanied by an edge weight that is updated in real time.

[0012] Obtain tender document text data, business registration information data of bidding companies, and historical bidding behavior data;

[0013] Based on the acquired data, semantic features of the tender document text, enterprise attribute features, and behavioral temporal features are extracted to form feature vectors for each node; at the same time, the weight of each edge in the graph structure model is calculated; and the graph structure model is updated.

[0014] Using the updated graph structure model as input, a graph neural network model is used to obtain node embedding vectors. Based on the node embedding vectors, the correlation matrix between enterprises is calculated to predict whether there is collusion or bid-rigging among enterprises.

[0015] As a further solution, equity association edges are constructed based on the equity association relationships between nodes. The edge weight of the equity association edge is determined by multiplying the equity hierarchy weight between enterprises by the similarity of equity structure.

[0016] The method for determining the similarity of the equity structure is as follows:

[0017] Construct feature dimensions that can reflect the essence of equity structure, generate binary vectors corresponding to enterprises based on feature dimensions, and calculate the similarity of binary vectors between enterprises as equity structure similarity.

[0018] As a further solution, historical cooperation edges are constructed based on the historical cooperation relationships between nodes. The method for determining the edge weights of these historical cooperation edges is as follows:

[0019] ;

[0020] in, λ represents the number of times the companies jointly bid, λ is the decay coefficient, and t is the time since the last joint bid.

[0021] As a further solution, behavioral collaboration edges are constructed based on the behavioral collaboration relationships between nodes. The method for determining the edge weight of the behavioral collaboration edge is as follows: determine the time series of bid volatility of enterprises in the most recent set number of bids; calculate the Pearson correlation coefficient of the bid volatility of two enterprises as the edge weight of the behavioral collaboration edge.

[0022] As a further solution, semantic collaborative edges are constructed based on the semantic collaborative relationships of the tender text between nodes. The method for determining the edge weights of the semantic collaborative edges is as follows:

[0023] The tender documents of the companies corresponding to node A and node B are split into sentence sequences respectively;

[0024] Each sentence is converted into a continuous numerical vector using a pre-trained model in the bidding and tendering field;

[0025] Calculate the semantic correlation between different sentences in the enterprise tender text corresponding to two nodes, and generate a sentence-level alignment matrix S;

[0026] Based on the sentence-level alignment matrix S, the global implicit collaboration score of the enterprise tender texts corresponding to node A and node B is calculated, which serves as the edge weight of the semantic collaboration edge between node A and node B.

[0027] As a further embodiment, the semantic features of the tender document text include a semantic vector of the tender document text; the enterprise attribute features include a vector composed of registered capital, qualification level, and equity penetration ratio; and the behavioral time series features include a time series vector of price volatility.

[0028] As a further solution, the association matrix S between enterprises is calculated based on node embedding vectors, specifically as follows:

[0029] ;

[0030] in, The elements in the correlation matrix represent the correlation strength between company i and company j. , These are the node embedding vectors for the nodes corresponding to enterprise i and enterprise j, respectively; Let be the average weight of all edges between firm i and firm j.

[0031] As a further embodiment, the graph neural network model includes a two-layer graph neural network algorithm, wherein the first layer graph neural network algorithm randomly samples a fixed number of nodes from its neighboring nodes for each node. The neighbors are used to calculate a weighted average of the neighbor features using the mean aggregation function. After concatenating the node's own features with the aggregated features of its neighbors, the weighted average is then applied using a weight matrix. Perform a linear transformation and activation to obtain the first-layer node embedding vector. ;

[0032] The second-layer graph neural network algorithm uses the embedding vectors of the first-layer nodes. As an initial feature of each node, a fixed number of random samples are taken from its neighboring nodes. Neighbors < The weighted average of neighbor features is calculated using the mean aggregation function. The node's own features are then concatenated with the aggregated features of its neighbors, and the result is weighted using a weight matrix. A linear transformation is performed, and a modularity optimization term is introduced to guide the model to identify closely related community structures; finally, the activation yields the first-layer node embedding vector. .

[0033] In other embodiments, the following technical solutions are adopted:

[0034] A scrambling prediction system based on dynamic association networks includes:

[0035] The graph structure construction module is configured to: construct a dynamic graph structure model, wherein the graph structure model uses enterprises as nodes, and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and semantic collaboration relationships of tender documents between nodes, with each edge having an edge weight that is updated in real time;

[0036] The data acquisition module is configured to acquire tender document text data, business registration information data of bidding companies, and historical bidding behavior data.

[0037] The graph structure update module is configured to: extract semantic features of the tender document text, enterprise attribute features, and behavioral temporal features based on the acquired data to form feature vectors for each node; calculate the weight of each edge in the graph structure model; and update the graph structure model.

[0038] The behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain node embedding vectors, calculate the correlation matrix between enterprises based on the node embedding vectors, and thus predict whether there is collusion between enterprises.

[0039] In other embodiments, the following technical solutions are adopted:

[0040] A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to perform the aforementioned method for predicting scrambling based on a dynamic association network.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] (1) This invention constructs equity association edges, historical cooperation edges, semantic collaboration edges and behavioral collaboration edges between enterprise nodes respectively, and dynamically updates the edge weights of each edge based on real-time acquired data, mining explicit and implicit associations between enterprises from different dimensions, enriching the modeling dimensions of the graph structure model; comprehensively considering the multi-source data features such as text semantics, enterprise attributes and behavioral time sequence, fully capturing implicit associations between enterprises, and improving the accuracy of identifying collusion behavior.

[0043] (2) This invention can identify implicit collaborations such as sentence rewriting and synonym replacement through the relationship between semantics of the tender document text; through dynamic information of enterprise attributes such as equity and qualification changes, it can capture the potential relationship changes between bidding entities in real time and realize the dynamic update of the graph structure model; through behavioral time series data such as price volatility, it can transform implicit enterprise collaborative behaviors into explicit graph relationships, thereby fully exploring the non-explicit associations between enterprises to achieve in-depth exploration of potential collusion groups.

[0044] (3) The present invention designs a two-layer graph neural network algorithm. The first layer aggregates direct neighbor nodes to capture the direct relationship between enterprises; the second layer realizes global relationship fusion to mine the deep implicit relationship between enterprises. It can identify the hidden bidding behavior of "main enterprise and multi-layer shell enterprises" and avoid the risk of missed judgment.

[0045] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0046] Figure 1 This is a flowchart of the tag-scratching prediction method based on dynamic association network in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of a bid-rigging prediction system based on a dynamic association network in an embodiment of the present invention. Detailed Implementation

[0048] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] Example 1

[0051] In one or more embodiments, a method for predicting flag-grabbing schemes based on dynamic association networks is disclosed, combining... Figure 1 Specifically, it includes the following process:

[0052] S101: Construct a dynamic graph structure model. The graph structure model uses enterprises as nodes and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and semantic collaboration relationships of tender documents between nodes. Each edge is accompanied by an edge weight that is updated in real time.

[0053] In this embodiment, enterprises are used as nodes, and equity-related edges, historical cooperation edges, behavioral collaboration edges, and semantic collaboration edges are constructed based on the equity relationships, historical cooperation relationships, behavioral collaboration relationships, and semantic collaboration relationships between nodes. By constructing multiple collaboration edges, both explicit (equity, historical cooperation) and implicit (behavioral, semantic collaboration) relationships can be covered simultaneously, reducing missed detections; multi-source data (business registration, bidding records, behavioral time series, text) are integrated to comprehensively characterize enterprise relationships; dynamic weight updates are supported to adapt to relationship changes and enhance robustness; and cross-validation of association signals accurately identifies complex bid-rigging patterns.

[0054] Each node contains basic attributes (such as registered capital, qualification level, equity penetration ratio, etc.), text attributes (semantic vector of tender document), and behavioral attributes (such as price volatility, etc.).

[0055] As a specific implementation method, equity association edges are used to analyze the equity relationships between enterprises and determine whether there are bid-rigging groups formed through equity ties. Equity associations (such as direct control, indirect shareholding, and cross-shareholding) reflect the capital ties and alignment of interests between enterprises, which are an important basis for the formation of bid-rigging groups. Through equity association edges, potential colluding entities such as "parent company-subsidiary" and "related enterprises" can be quickly identified. These enterprises are more likely to collude to manipulate bids due to their aligned interests.

[0056] The weight of equity-related edges can accurately measure the strength of the association. For example, companies with direct shareholding and highly similar equity structures are significantly more likely to collude in targeting each other, providing a quantitative basis for identifying core gang members.

[0057] The method for determining the edge weight of equity-related edges is: the product of the weight of equity hierarchy among enterprises and the similarity of equity structure, i.e.:

[0058] ;

[0059] in, Weighting based on equity levels, such as in the case of direct control. =0.8, when indirectly holding shares =0.5; represents the equity structure similarity, and represents the binary vector correlation coefficient.

[0060] Similarity of equity structure The calculation process is as follows:

[0061] (1) Determine the characteristic dimensions of the equity structure:

[0062] Select a set number of characteristic dimensions that can reflect the essence of the equity structure (e.g., preset 8 dimensions):

[0063] Dimension 1: Are there any individual shareholders holding more than 30% of the shares?

[0064] Dimension 2: Are there any corporate shareholders (enterprises) holding more than 50% of the shares?

[0065] Dimension 3: Whether there are multiple layers of nested shareholding (≥3 layers);

[0066] Dimension 4: Whether the sum of the shareholding ratios of the top three shareholders is greater than 60% (concentrated shareholding);

[0067] Dimension 5: Whether there are foreign shareholders;

[0068] Dimension 6: Whether there is state-owned capital holding shares;

[0069] Dimension 7: Whether the number of shareholders is greater than 10 (dispersed shareholding);

[0070] Dimension 8: Does cross-shareholding exist (e.g., A holds B, and B also holds A)?

[0071] (2) Based on the actual situation of the enterprise, for each feature dimension, use "1" to indicate that the feature is present and "0" to indicate that the feature is not present, and generate the binary vector of the enterprise.

[0072] For example: the binary vector of company A is ,in (Describe the characteristics of Company A in specific dimension 1). (Explain the specific features of company A in dimension 2), and so on; the binary vector of company B is... The same logic is used for binary encoding.

[0073] (3) Calculate the similarity between the binary vectors of the two companies using the following formula:

[0074] ;

[0075] The numerator represents the number of feature dimensions shared by the two companies (i.e., the intersection of the binary vectors of company A and company B); the denominator represents the number of feature dimensions shared by at least one of the two companies (i.e., the union of the binary vectors of company A and company B).

[0076] Calculated similarity This can then be used as the edge weight of the equity relationship between company A and company B.

[0077] As a specific implementation method, historical cooperation edges are used to reflect the relationship between two companies bidding together. The more times they bid together, the greater the risk of collusion. The method for determining the edge weight of historical cooperation edges is as follows:

[0078] ;

[0079] in, This indicates the number of times the two companies jointly bid. λ is the attenuation factor, and λ is the attenuation coefficient. The time since the last joint bid is in days.

[0080] Bidding syndicates often manipulate bids through "short-term, high-frequency cooperation," while the relevance of long-term cooperation may weaken due to corporate strategic adjustments, market changes, etc. Therefore, the attenuation factor makes recent joint bids carry higher weight (the smaller t is, the higher the weight). The closer the model is to 1, the better it reflects the current potential collaborative risks, avoids outdated information interfering with the current risk assessment, and improves the model's sensitivity to recent behavior.

[0081] Structural similarity analysis can uncover hidden connections that appear independent but are actually controlled (such as circumventing regulations by distributing shares through multiple intermediary companies), filling the gaps in traditional identification methods that rely solely on surface-level business registration information. Changes in equity structure (such as equity transfers or the addition of new shareholders) will update the edge weights in real time, enabling the model to promptly capture changes in corporate relationships and avoid missing detections of bid-rigging activities caused by equity adjustments.

[0082] As a specific implementation method, behavioral collaboration edges are used to reflect the collaborative relationship between two companies, such as whether their price fluctuation trends are the same. These collaborative relationships enhance the sensitivity of the graph structure model to collaborative behavior, which is beneficial for identifying potential bid-rigging behavior. The method for determining the edge weights of behavioral collaboration edges is as follows:

[0083] (1) Obtain the target company's bid price data in at least N historical projects and form a time series. , The bid amounts (in ten thousand yuan) for each of the N projects are listed in chronological order by bid date.

[0084] Remove abnormal quotes (such as quotes that are significantly lower than the cost price or higher than the market price), standardize the quotes (such as subtracting the mean and then dividing by the standard deviation) to eliminate the influence of dimensions.

[0085] (2) Set the time window length as (like (This indicates that the trend is calculated based on the most recent 5 items), then the trend value... For time window The average of all quotes.

[0086] (3) Calculate the time window The relative deviation between the actual price and the trend value in each project The price volatility is obtained as follows: , Representing a time window The actual quote for each item within.

[0087] By continuously moving the time window, the price volatility of all quotes can be calculated.

[0088] (4) Obtain the price volatility series of Company A respectively. Price volatility series of Company B Pearson correlation coefficient was used. Measuring the correlation of volatility series:

[0089] ;

[0090] You can | |The edge weight as a behavior collaboration edge.

[0091] If | | The volatility exceeds the set threshold, and at least k projects show high volatility for both Company A and Company B simultaneously. If the set value is not met, then there is a suspicion that Company A and Company B are colluding to manipulate prices.

[0092] As a specific implementation method, semantic collaborative edges are used to reflect the relationship between the bid documents of two companies; the edge weights of semantic collaborative edges can be determined by the similarity between the bid documents, specifically as follows:

[0093] (1) The tender documents of Company A and Company B are split into sentence sequences respectively. The sentence sequence of Company A is as follows: Company B's sentence sequence is as follows: ; and Each represents a sentence, i=1,2,…,m, j=1,2,…,n.

[0094] (2) Use a pre-trained model in the bidding field (such as the BERT fine-tuning model) to convert each sentence into a low-dimensional continuous numerical vector.

[0095] (3) Calculate the semantic relevance between sentences in the two companies' tender documents to determine which sentences are highly semantically related:

[0096] ;

[0097] in, Sentence and sentences The semantic relevance between them This is a learnable parameter matrix used to capture cross-textual semantic interactions; It also represents sentences in the sentence sequence of Company B, but with... They can be different. The larger the value, the more accurate the sentence is. and sentences The closer the semantic connection between them.

[0098] (4) Generate according to the above calculation method. The sentence-level alignment matrix S, the elements in matrix S .

[0099] Based on the sentence-level alignment matrix S, calculate the global implicit collaboration score of the tender documents of Company A and Company B:

[0100] ;

[0101] in, , The importance weights of sentence pairs can be automatically generated using a hierarchical attention mechanism.

[0102] Global implicit collaborative score of the tender documents of Company A and Company B It can be used as the edge weight of semantic collaborative edges.

[0103] This embodiment maps deep semantic structures sentence by sentence, which can identify implicit collaborations under sentence rewriting and synonym replacement (such as the semantic consistency between "cost control" and "expense management"), and capture logical connections scattered across different chapters (such as the echo between the technical cooperation statement and the business support statement), thus achieving deep detection from "superficial similarity" to "intent alignment".

[0104] S102: Obtain tender document text data, business registration information data of bidding companies, and historical bidding behavior data.

[0105] In this embodiment, the tender document text data is obtained directly from the bidding platform, including technical solutions, commercial quotations, project response terms, etc. This text contains a large amount of semantic coordination information that may be involved in bid rigging, such as similar technical descriptions and unusually consistent commercial terms. Simultaneously, basic information of the bidding companies, such as company name, unified social credit code, and registered address, is obtained to construct the basic attributes of the company nodes.

[0106] The business registration information of bidding companies is obtained from a corporate business registration information database. This includes data on the companies' equity structure, such as shareholder information, shareholding ratios, and the number of equity penetration layers. This data is used to analyze the equity relationships between companies and determine whether there are any bid-rigging groups formed through equity ties. In addition, the company's qualification information, such as qualification type, level, and validity period, is collected and compared with the qualification requirements of the project tender to check for signs of illegal activities such as qualification theft. As a specific example, a company's qualification type that perfectly matches the project's qualification type scores 1; a partial match is assigned a score based on the degree of matching (e.g., 0.5); and a non-match scores 0.

[0107] Establish a local or cloud-based database of historical bidding behavior, storing data such as bidding time, bid amount, and winning results for companies in past bidding activities. Regularity in bidding time and unusually close or different bids are important indicators for identifying bid-rigging. For example, if multiple companies submit bids for multiple projects at extremely close times, it may indicate coordinated bidding; specific regular fluctuations in bids, such as increasing or decreasing by a fixed difference, also strongly suggest bid-rigging.

[0108] It should be noted that all of the above data can be obtained through legal means.

[0109] S103: Based on the acquired data, extract the semantic features of the tender document text, the enterprise attribute features, and the behavioral temporal features to form the feature vector of each node; at the same time, calculate the weight of each edge in the graph structure model; and update the graph structure model.

[0110] Specifically, the semantic features of the tender document text include the semantic vector of the tender document text, with dimensions of [missing information]. The enterprise attribute characteristics include a vector composed of registered capital (normalized value), qualification level (unique hot coded value), and equity penetration ratio, with dimensions of [missing information]. Behavioral time-series features include a price volatility time-series vector, with dimensions of [missing information]. ;

[0111] For each node, extract the aforementioned feature vectors separately, and combine these feature vectors together to obtain the feature vector of each node, with dimension [dimensionality missing]. .

[0112] Based on the acquired data, the edge weight determination method recorded in S101 is used to determine the edge weights of equity association edges, historical cooperation edges, behavioral collaboration edges, and semantic collaboration edges between each node.

[0113] The feature vector of each node and the calculated edge weights are assigned to the graph structure model to form the updated graph structure model.

[0114] S104: Using the updated graph structure model as input, the graph neural network model is used to obtain node embedding vectors. Based on the node embedding vectors, the correlation matrix between enterprises is calculated to predict whether there is collusion between enterprises.

[0115] In this embodiment, the graph neural network model adopts a two-layer GraphSAGE (SAmple and aggreGatE) algorithm;

[0116] The first layer of the GraphSAGE algorithm mainly performs feature aggregation of neighboring nodes. The specific operation process is as follows:

[0117] For each node v, a fixed number of random samples are taken from its neighbors N(v). Neighbors (such as) =20), to avoid excessive computational complexity;

[0118] Calculate a weighted average of neighbor features using the mean aggregation function:

[0119] ;

[0120] in, Let be the initial feature vector of the neighbor node u.

[0121] After concatenating the node's own features with the aggregated features of its neighbors, the weight matrix is ​​used... Perform a linear transformation and activate: || Where σ is the ReLU activation function, This indicates feature splicing.

[0122] Finally, the first-layer node embedding vector is obtained. , dimension (generally (to achieve feature dimensionality reduction).

[0123] The second-layer GraphSAGE algorithm mainly achieves global relation fusion, and the specific operation process is as follows:

[0124] Update sample count ( Less than (Focusing on core neighbors)

[0125] Recalculate the mean aggregation of the neighbor features: ; This represents the first-level node embedding vector after the neighbor node u has been aggregated by the first-level GraphSAGE algorithm.

[0126] Similar to the first layer, but the weight matrix is ​​updated to... And introduce modularity optimization terms: ,in Modularity function gradient, The regularization coefficient (e.g., 0.01) guides the model to identify closely related community structures. The sum of edge weights within community i is the proportion of the total edge weights. It is a key indicator for measuring the tightness of connections within a community and is used to calculate the "difference between community structure and random network" (modularity is essentially a comparison of community connectivity between real networks and random networks). The sum of the edge weights connected to community i is the proportion of the total edge weights.

[0127] In this context, a community structure refers to a tightly knit subgroup formed by nodes (enterprises). Nodes within a community are densely connected, while connections to external nodes are sparse. In bid-rigging scenarios, a community typically corresponds to a potential bid-rigging gang. Its internal enterprises form a highly interconnected network through various means such as equity control, plagiarism of bid documents, and coordinated bidding behavior, while their connections to enterprises outside the gang are relatively weak.

[0128] This embodiment will... As a gradient term, it is added to the GNN training ( The core of the modularity optimization term ▽Q(G) is to measure the difference between the "sum of edge weights within the community" and the "expected weights when randomly assigned". Maximizing Q will cause the model to map clusters of nodes with close internal connections and sparse external connections to similar positions in a low-dimensional space, forming a cluster structure. Therefore, the modularity optimization term can directly guide the model to identify such closely related subgroups (i.e., community structures), thereby enabling the identification of hidden groups.

[0129] Finally, the second-layer node embedding vector is obtained. , dimension (like This vector integrates the direct associations and indirect transmission relationships of nodes (such as the indirect associations of ABC).

[0130] In this embodiment, the first-layer GraphSAGE algorithm samples... direct neighbors ( (Larger), aggregating the direct association features of nodes (such as direct equity or cooperative relationships between enterprises), the output node embedding vector It already includes local direct relationship information. The second-layer GraphSAGE algorithm uses... For input, sampling One core neighbor ( < Focusing on more critical relationships, the aggregated "neighbors" not only include the node's direct neighbors, but also indirectly include the "neighbors' neighbors" (i.e., indirect relationships) through the embedding vectors of the first layer. This multi-level aggregation allows features to convey relationships over longer distances (such as the indirect relationship of ABC), thereby covering the global relationship network.

[0131] This embodiment overcomes the limitations of traditional pairwise association analysis through two-layer aggregation, uncovering complex relationships of "direct association + indirect transmission" (e.g., companies A and C have no direct cooperation, but form a potential association through their common neighbor B). Through end-to-end training, it automatically learns the weight allocation of different types of edges (equity-related edges, semantic collaboration edges, behavioral collaboration edges, and historical cooperation edges), avoiding the bias of manually preset weights. The second layer, aggregation of indirect neighbors, can identify hidden groups of "mastermind companies + multiple shell companies" (e.g., company D indirectly controls G through E and F, forming a bid-rigging network), which traditional methods cannot identify due to a lack of deep relationship mining.

[0132] After two layers of GraphSAGE computation, the output dimension of each node is... Embedded vector This vector integrates the node's own features with the association features of its neighboring nodes, and is used to characterize the potential association attributes of an enterprise in a dynamic relationship network.

[0133] Based on node embedding vectors, calculate the inter-enterprise association matrix S, where the elements... Represents the association strength between companies i and j; node embedding vector (Model output) integrates the enterprise's attribute features, textual semantic features, behavioral temporal features, and the association information of neighboring nodes (such as indirect equity and historical cooperation transmission), and is a low-dimensional mapping of the enterprise's "comprehensive association attributes" in a dynamic relationship network.

[0134] therefore, The essence is the similarity of the embedding vectors of two nodes. The higher the similarity, the stronger the consistency of the two companies in multi-dimensional related features, and the higher the possibility of potential collusion and bid-rigging.

[0135] In this embodiment, The specific calculation method is as follows:

[0136] ;

[0137] in, The elements in the correlation matrix represent the correlation strength between company i and company j. , These are the node embedding vectors for the nodes corresponding to enterprise i and enterprise j, respectively; It is the average edge weight of all edges (equity-related edges, semantic collaboration edges, behavioral collaboration edges, and historical cooperation edges) between enterprise i and enterprise j.

[0138] It should be noted that if the two companies have no direct relationship (e.g., this is their first contact), then... (Default neutral weights to avoid completely ignoring potential implicit associations).

[0139] Due to node embedding vector and It is a high-dimensional dense vector (such as) Cosine similarity is used to measure directional consistency (reflecting the degree of overlap in feature patterns), and the formula is: The value range is [-1, 1]. The closer it is to 1, the more consistent the embedding vectors of the two companies are (the more similar the associated features).

[0140] To reinforce the impact of known explicit relationships (such as equity ownership and historical cooperation), the mean of the edge weights between the two firms is introduced. As a correction factor.

[0141] After obtaining the correlation matrix S between enterprises, all correlation strengths less than the set threshold are set to 0, and all correlation strengths are normalized to [0,1]. Enterprise nodes with correlation strengths greater than the set threshold are judged to have collusion behavior.

[0142] In this embodiment, during the training process of the graph neural network, a dynamic masking mechanism is used to enhance the model's ability to capture sensitive semantic associations, targeting implicit collusion and bid-rigging features in bidding texts (such as reused technical solutions, unusually consistent response logic, and veiled interest-related rhetoric).

[0143] Dynamic mask generation is divided into three levels:

[0144] Dynamic masking of the text feature space: During the training phase, the feature vector of the input tender text is masked with a certain probability, and the mask position is dynamically adjusted. Unlike BERT's fixed mask, this mechanism dynamically generates the mask based on the following rules: limited semantic density, collaborative feature enhancement, and time-dependent awareness.

[0145] Adversarial training of multimodal feature fusion: Perturb the weights of semantic collaborative edges to simulate possible word rewriting behaviors (synonym replacement, parameter shift, word order reordering) of companies that collude in bidding.

[0146] Temporal dynamic mask update: The mask strategy evolves over time. Every 10 epochs (rounds) of training, the mask strategy is dynamically adjusted based on the model's current accuracy in recognizing crosstalk samples. A time decay factor is introduced to reduce the mask probability of historical high-frequency mask positions, thus avoiding model overfitting to specific perturbation patterns.

[0147] Specifically, the calculation process for the dynamic mask rate is as follows:

[0148] Define dynamic mask rate N is the length of the text sequence, and T is the number of training rounds; , This represents the total number of times the i-th position was masked during historical training. Let be the accuracy of identifying cross-label samples after the t-th round of training; To control the attenuation rate of the historical mask frequency; The initial mask probability before adjustment; This is a function that dynamically adjusts the mask strength based on accuracy.

[0149] (1) Historical mask frequency attenuation: , The function is an indicator function. It takes the value 1 when the i-th position is masked in round t-1, and 0 otherwise. This formula indicates that the historical frequency decays exponentially with time, while the masking record of the current round is accumulated.

[0150] (2) Accuracy adjustment coefficient: , To adjust the amplitude (e.g.) =0.2), For target accuracy (e.g.) =0.9), when the accuracy is lower than the target, increase the mask probability to strengthen training; when it is higher than the target, decrease the mask probability to avoid overfitting.

[0151] (3) Dynamic mask rate calculation: , For frequency decay exponent (e.g.) =0.1), which controls the probability decay rate of historical high-frequency mask positions. This formula combines accuracy adjustment and historical frequency decay to ensure that the mask position changes dynamically with training.

[0152] The policy update mechanism every 10 epochs (rounds) is as follows: when t mod 10 = 0, the historical frequency matrix is ​​reset. The mask records from 10 epochs ago are weakened through exponential decay; the base mask probability is dynamically adjusted. , The adaptive coefficients are adjusted based on the accuracy changes over the last 10 rounds. If the accuracy improves, the initial mask strength is reduced. To reduce mask perturbation; if it decreases, increase it. To increase the difficulty of training.

[0153] The specific process of modularity optimization is as follows:

[0154] Modularity function As a regularization term in graph neural networks, it guides the model to identify closely related community structures and reduces the false negative rate.

[0155] in, This represents the ratio of the sum of edge weights within community i to the total edge weights. For example, if the sum of the weights of equity-related edges and semantic collaboration edges among enterprises within community i is 0.3, and the total edge weights are 1.0, then... =0.3.

[0156] This represents the proportion of the sum of edge weights connected to community i to the total edge weights. For example, if the sum of edge weights between community i and external firms is 0.2, then... =0.2.

[0157] In this embodiment, whenever a new bidding record is generated or the company's business registration information changes, the following update is triggered:

[0158] Calculate the edge weights (equity, cooperation, semantics, behavior) between the new enterprise and existing nodes. Only update the affected edge weights and the embedding vectors of the corresponding nodes to avoid full retraining.

[0159] Adversarial examples simulating novel collusion relationships (such as fictitious corporate relationships involving "dispersed equity + cross-regional bidding") are periodically generated and injected into the relationship network for robustness training, improving the ability to identify unknown relationship patterns. (Adversarial examples refer to artificially constructed data samples that simulate novel and covert collusion relationships in reality. These samples appear to be normal bidding data on the surface, but actually contain carefully designed relationship information with collusion-like characteristics. Injecting them into the relationship network for training aims to enable the model to learn to identify these hidden abnormal patterns, enhance its resistance to various potential collusion behaviors, and thus improve the model's robustness and generalization ability.)

[0160] As a concrete example, the process of constructing adversarial examples based on data features is as follows:

[0161] Regarding equity structure: Utilizing publicly available business registration data, a complex, fictitious equity structure is created among companies. For example, multiple fictitious intermediate holding companies are established, each holding a small percentage of the target company's equity, constructing an equity structure that appears dispersed but is actually controlled by the same entity. Assuming a real bid-rigging group holds equity in the core bidding company through multiple nested companies, this structure is simulated to generate adversarial examples. For instance, company A indirectly holds 5%, 3%, and 4% of company E's equity through three seemingly unrelated companies B, C, and D, respectively, creating the illusion of dispersed equity, but the actual controller is the same person. This equity structure data is used as part of the adversarial examples.

[0162] Regarding bidding behavior: The adversarial examples simulate abnormal patterns in cross-regional bidding. Based on the normal regional distribution and frequency of bidding in historical bidding data, unusual bidding behaviors are generated. For example, in a certain region, normal companies participate in local projects more frequently and participate in distant projects less frequently. When constructing adversarial examples, some companies are designed to frequently participate in cross-regional bidding projects, with bidding time intervals showing a regularity that differs significantly from normal corporate bidding behavior. For example, Company F frequently participates in multiple projects far from its registered location within a short period, with each bidding interval within one week, which is inconsistent with the normal bidding habits of local companies.

[0163] In terms of semantic collaboration: Natural language processing techniques are used to rewrite the content of tender documents. Common expressions are extracted from a large number of real tender documents, and synonym replacement, sentence transformation, and other methods are used to construct tender documents with similar semantics but different expressions. For example, "Our company has advanced technical solutions" is rewritten as "Our company has cutting-edge technical response strategies," making multiple tender documents semantically collaborative, but not noticeable on the surface. These processed tender documents are used as semantic feature parts of adversarial examples, injected into a relational network, and the model is trained to recognize this implicit semantic collaboration.

[0164] As a specific example, the process of constructing adversarial examples in conjunction with attack algorithms is as follows:

[0165] Inspired by adversarial example generation algorithms in the image domain, this approach is applied to relational network data. Taking a gradient-based method as an example, the gradient information of the relational network model when processing current normal bidding data is calculated. Then, along the direction that causes the model to misidentify, small perturbations are made to the equity correlation, bidding behavior features, and semantic similarity in the data. Assuming that the model's judgment of normal equity correlation is relatively stable, the algorithm adds small perturbations to the equity correlation data, causing the model to deviate from these judgments, simulating new types of collusion and bid-rigging relationships that are difficult for the model to identify, thereby constructing adversarial examples.

[0166] This embodiment introduces a graph neural network (GNN) to dynamically update the weights of the edges connecting enterprises, and combines multi-source data (textual semantics, enterprise attributes, and behavioral time series) to capture the potential relationship changes of bidding entities in real time, thereby improving the accuracy of identifying collusive bidding behavior.

[0167] Example 2

[0168] In one or more embodiments, a flag-grabbing prediction system based on dynamic association networks is disclosed, combined with... Figure 2 Specifically, it includes:

[0169] The graph structure construction module is configured to: construct a dynamic graph structure model, wherein the graph structure model uses enterprises as nodes, and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and semantic collaboration relationships of tender documents between nodes, with each edge having an edge weight that is updated in real time;

[0170] The data acquisition module is configured to acquire tender document text data, business registration information data of bidding companies, and historical bidding behavior data.

[0171] The graph structure update module is configured to: extract semantic features of the tender document text, enterprise attribute features, and behavioral temporal features based on the acquired data to form feature vectors for each node; calculate the weight of each edge in the graph structure model; and update the graph structure model.

[0172] The behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain node embedding vectors, calculate the correlation matrix between enterprises based on the node embedding vectors, and thus predict whether there is collusion between enterprises.

[0173] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0174] Example 3

[0175] In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the processor to perform the tag-blocking prediction method based on dynamic association network described in Embodiment 1.

[0176] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0177] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0178] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0179] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting scrambling based on dynamic association networks, characterized in that, include: A dynamic graph structure model is constructed, in which enterprises are used as nodes, and edges are constructed based on the equity relationship, historical cooperation relationship, behavioral collaboration relationship and semantic collaboration relationship of the tender document text between nodes. Each edge is accompanied by an edge weight that is updated in real time. Behavioral collaboration edges are constructed based on the behavioral collaboration relationships between nodes. The method for determining the edge weight of the behavioral collaboration edge is as follows: determine the time series of bid volatility of enterprises in the most recent set number of bids, set a time window, continuously move the time window, and calculate the bid volatility of all bids; calculate the Pearson correlation coefficient of the bid volatility of two enterprises as the edge weight of the behavioral collaboration edge. Semantic collaborative edges are constructed based on the semantic collaborative relationships between the tender documents of nodes. The method for determining the edge weight of the semantic collaborative edges is as follows: the tender documents of the enterprises corresponding to nodes A and B are split into sentence sequences respectively; each sentence is converted into a continuous numerical vector using a pre-trained model in the bidding domain; the semantic correlation between different sentences in the tender documents of the enterprises corresponding to the two nodes is calculated, and a sentence-level alignment matrix S is generated. Based on the sentence-level alignment matrix S, the global implicit collaboration score of the enterprise tender documents corresponding to node A and node B is calculated, which is used as the edge weight of the semantic collaboration edge between node A and node B. Obtain tender document text data, business registration information data of bidding companies, and historical bidding behavior data; Based on the acquired data, semantic features of the tender document text, enterprise attribute features, and behavioral temporal features are extracted to form feature vectors for each node; at the same time, the weight of each edge in the graph structure model is calculated. The graph structure model is updated; the behavioral time-series features include the time-series vector of price volatility. Using the updated graph structure model as input, a graph neural network model is used to obtain node embedding vectors. Based on the node embedding vectors, the correlation matrix between enterprises is calculated to predict whether there is collusion between enterprises. The graph neural network model includes a two-layer graph neural network algorithm. The first layer graph neural network algorithm randomly samples a fixed number of neighboring nodes from each node. The neighbors are used to calculate a weighted average of the neighbor features using the mean aggregation function. After concatenating the node's own features with the aggregated features of its neighbors, the weighted average is then applied using a weight matrix. Perform a linear transformation and activation to obtain the first-layer node embedding vector. ; The second-layer graph neural network algorithm uses the embedding vectors of the first-layer nodes. As an initial feature of each node, a fixed number of random samples are taken from its neighboring nodes. Neighbors < The weighted average of neighbor features is calculated using the mean aggregation function. The node's own features are then concatenated with the aggregated features of its neighbors, and the result is weighted using a weight matrix. A linear transformation is performed, and a modularity optimization term is introduced to guide the model to identify closely related community structures; finally, the activation yields the first-layer node embedding vector. .

2. The method for predicting smuggling based on dynamic association networks as described in claim 1, characterized in that, Equity association edges are constructed based on the equity association relationships between nodes. The edge weight of the equity association edge is determined by multiplying the equity hierarchy weight between enterprises by the similarity of equity structure. The method for determining the similarity of the equity structure is as follows: Construct feature dimensions that can reflect the essence of equity structure, generate binary vectors corresponding to enterprises based on feature dimensions, and calculate the similarity of binary vectors between enterprises as equity structure similarity.

3. The method for predicting smuggling based on dynamic association networks as described in claim 1, characterized in that, Historical cooperation edges are constructed based on the historical cooperation relationships between nodes. The method for determining the edge weights of these historical cooperation edges is as follows: ； in, λ represents the number of times the companies jointly bid, λ is the decay coefficient, and t is the time since the last joint bid.

4. The method for predicting smuggling based on dynamic association networks as described in claim 1, characterized in that, The semantic features of the tender document text include a semantic vector of the tender document text; the enterprise attribute features include a vector composed of registered capital, qualification level, and equity penetration ratio.

5. The method for predicting smuggling based on dynamic association networks as described in claim 1, characterized in that, The association matrix S between enterprises is calculated based on node embedding vectors, specifically as follows: ； in, The elements in the correlation matrix represent the correlation strength between company i and company j. , These are the node embedding vectors for the nodes corresponding to enterprise i and enterprise j, respectively; Let be the average weight of all edges between firm i and firm j.

6. A collusion prediction system based on dynamic association networks, characterized in that, include: The graph structure construction module is configured to: construct a dynamic graph structure model, wherein the graph structure model uses enterprises as nodes, and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and semantic collaboration relationships of tender documents between nodes, with each edge having an edge weight that is updated in real time; construct behavioral collaboration edges based on the behavioral collaboration relationships between nodes, and the edge weight of the behavioral collaboration edges is determined by: determining the time series of bid volatility of enterprises in the most recent set number of bids, setting a time window, continuously moving the time window, and calculating the bid volatility of all bids; calculating the Pearson correlation coefficient of the bid volatility of two enterprises as the edge weight of the behavioral collaboration edge; Semantic collaborative edges are constructed based on the semantic collaborative relationships between the tender documents of nodes. The method for determining the edge weight of the semantic collaborative edges is as follows: the tender documents of the enterprises corresponding to nodes A and B are split into sentence sequences respectively; each sentence is converted into a continuous numerical vector using a pre-trained model in the bidding domain; the semantic correlation between different sentences in the tender documents of the enterprises corresponding to the two nodes is calculated, and a sentence-level alignment matrix S is generated. Based on the sentence-level alignment matrix S, the global implicit collaboration score of the enterprise tender documents corresponding to node A and node B is calculated, which is used as the edge weight of the semantic collaboration edge between node A and node B. The data acquisition module is configured to acquire tender document text data, business registration information data of bidding companies, and historical bidding behavior data. The graph structure update module is configured to: extract semantic features of the tender document text, enterprise attribute features, and behavioral time-series features based on the acquired data, forming a feature vector for each node; simultaneously calculate the weight of each edge in the graph structure model; and update the graph structure model; the behavioral time-series features include the price volatility time-series vector. The behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain node embedding vectors, calculate the correlation matrix between enterprises based on the node embedding vectors, and thus predict whether there is collusion between enterprises. The graph neural network model includes a two-layer graph neural network algorithm. The first layer graph neural network algorithm randomly samples a fixed number of neighboring nodes from each node. The neighbors are used to calculate a weighted average of the neighbor features using the mean aggregation function. After concatenating the node's own features with the aggregated features of its neighbors, the weighted average is then applied using a weight matrix. Perform a linear transformation and activation to obtain the first-layer node embedding vector. ; The second-layer graph neural network algorithm uses the embedding vectors of the first-layer nodes. As an initial feature of each node, a fixed number of random samples are taken from its neighboring nodes. Neighbors < The weighted average of neighbor features is calculated using the mean aggregation function. The node's own features are then concatenated with the aggregated features of its neighbors, and the result is weighted using a weight matrix. A linear transformation is performed, and a modularity optimization term is introduced to guide the model to identify closely related community structures; finally, the activation yields the first-layer node embedding vector. .

7. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-5, which is a method for predicting scrambling based on dynamic association networks.

Images