Dynamic association network-based constellation mark prediction method and system, and terminal equipment

Abstract

The invention belongs to the technical field of bidding and tendering risk identification, and specifically discloses a dynamic association network-based bidding and tendering risk prediction method and system, and a terminal device. The method comprises the following steps: constructing a dynamic graph structure model; obtaining bidding document text data, industry and commerce information data of a bidding enterprise and historical bidding behavior data; respectively extracting bidding document text semantic features, enterprise attribute features and behavior time sequence features to form a feature vector of each node; calculating the weight of each edge in the graph structure model; updating the graph structure model; and taking the updated graph structure model as input, obtaining a node embedding vector by using a graph neural network model, and calculating an association degree matrix between the enterprises based on the node embedding vector so as to predict whether a surrounding and bidding behavior exists between the enterprises. According to the method, multi-source data features such as text semantics, enterprise attributes and behavior time sequences are comprehensively considered, the implicit incidence relation between enterprises is fully captured, and the recognition accuracy of the bidding behavior is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bidding risk identification, and in particular to a method, system and terminal device for predicting bid rigging based on a dynamic association network. Background Art

[0002] The statements in this section merely provide 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 long been a hot topic in the field. To increase their chances of winning bids and exclude other competitors, some suppliers may engage in unfair competition practices such as bid rigging and collusion. Bid rigging and collusion (hereinafter referred to as bid rigging) refers to the private collusion between multiple bidders to manipulate the bidding results through pre-negotiated bids or other means. This practice, often intended to favor a specific bidder, undermines fair competition and constitutes an illegal and irregular act.

[0004] In the traditional audit process, the analysis of bidding information often relies on the subjective experience of the reviewer. The review method is relatively outdated and the identification efficiency is low.

[0005] The prior art also discloses some methods for automatically identifying bid-rigging behaviors, such as: A weighted network is constructed through the co-occurrence relationship of "enterprise-bidding project", with the number of joint bids as the weight. The risk of bid-rigging behavior is 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.

[0006] Alternatively, the correlation index between enterprises can be determined by calculating the binary vector correlation coefficient of the enterprise attribute information group, and the related enterprises can be judged based on manually preset attribute thresholds, thereby identifying the risk of bid-rigging behavior. However, this method cannot tap into the potential synergy at the semantic level, and it is difficult to capture non-explicit correlation relationships such as "similar plans but different expressions", resulting in a high risk of missing implicit bid-rigging behavior. Summary of the Invention

[0007] In order to solve the above problems, the present invention proposes a bid-rigging prediction method, system and terminal device based on a dynamic association network. Graph neural network (GNN) is introduced to dynamically update the weights of the association edges between enterprises. Multi-source data (text semantics, enterprise attributes, and behavior time series) are combined to capture the potential relationship changes of bidding entities in real time, thereby improving the accuracy of identifying bid-rigging behaviors.

[0008] In some embodiments, the following technical solutions are adopted: A method for predicting bid rigging based on a dynamic association network, comprising: Build a dynamic graph structure model that uses enterprises as nodes and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and bid document text semantic collaboration relationships between nodes. Each edge is accompanied by a real-time updated edge weight. Obtain bid document data, bidding company business information data, and historical bidding behavior data; Based on the acquired data, the semantic features of the bid document text, the enterprise attribute features, and the behavioral temporal features are extracted to form the feature vector of 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. Taking the updated graph structure model as input, the graph neural network model is used to obtain the node embedding vector. Based on the node embedding vector, the correlation matrix between enterprises is calculated to predict whether there is bid rigging between enterprises.

[0009] As a further solution, the equity-related edge is constructed based on the equity-related relationship between the nodes. The edge weight of the equity-related edge is determined by the product of the equity level weight and the equity structure similarity between the enterprises. The method for determining the equity structure similarity is: Construct a characteristic dimension that can reflect the essence of the equity structure, generate a binary vector corresponding to the enterprise based on the characteristic dimension, and calculate the similarity of the binary vectors between enterprises as the equity structure similarity.

[0010] As a further solution, historical cooperation edges are constructed based on the historical cooperation relationships between nodes. The edge weight of the historical cooperation edge is determined as follows: ; in, is the number of joint biddings by enterprises, λ is the decay coefficient, and t is the time since the last joint bidding.

[0011] As a further solution, a behavioral collaborative edge is constructed based on the behavioral collaborative relationship between nodes. The method for determining the edge weight of the behavioral collaborative edge is: determining the time series of the enterprise's quotation volatility in the most recent set number of bids; calculating the Pearson correlation coefficient of the quotation volatility of the two enterprises as the edge weight of the behavioral collaborative edge.

[0012] As a further solution, a semantic collaborative edge is constructed based on the semantic collaborative relationship of the bid document text between nodes. The edge weight of the semantic collaborative edge is determined as follows: Split the enterprise bid texts corresponding to nodes A and B into sentence sequences respectively; Use the pre-trained model in the bidding field to convert each sentence into a continuous numerical vector; Calculate the semantic relevance between different sentences in the enterprise bid text corresponding to the two nodes and generate a sentence-level alignment matrix S; Based on the sentence-level alignment matrix S, the global implicit collaboration score of the enterprise bid text corresponding to node A and node B is calculated as the edge weight of the semantic collaboration edge between node A and node B.

[0013] As a further solution, the semantic features of the bid text include a semantic vector of the bid 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 quotation volatility.

[0014] As a further solution, the correlation matrix S between enterprises is calculated based on the node embedding vectors, specifically: ; in, is an element in the association matrix, indicating the strength of association between enterprise i and enterprise j; 、 are the node embedding vectors of the nodes corresponding to enterprise i and enterprise j respectively; is the mean weight of all edges between enterprise i and enterprise j.

[0015] As a further solution, the graph neural network model includes a two-layer graph neural network algorithm. The first layer of the graph neural network algorithm randomly samples a fixed number of nodes from its neighboring nodes for each node. The neighbors of the node are calculated using the mean aggregation function to calculate the weighted average of the neighbor features. After the node's own features are spliced ​​with the neighbor aggregation features, the weight matrix Perform linear transformation and activation to obtain the first layer node embedding vector ; The second-layer graph neural network algorithm uses the first-layer node embedding vector As the initial features of the node itself, a fixed number of random samples are taken from each node’s neighboring nodes. Neighbors, < ; Use the mean aggregation function to calculate the weighted average of neighbor features, concatenate the node's own features with the neighbor aggregation features, and then use the weight matrix Perform linear transformation and introduce modularity optimization to guide the model to identify closely related community structures; finally activate and obtain the first-layer node embedding vector .

[0016] In other embodiments, the following technical solutions are adopted: A bid-rigging prediction system based on a dynamic association network, comprising: 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 bid document text semantic collaboration relationships between nodes, with each edge accompanied by a real-time updated edge weight; The data acquisition module is configured to: acquire bid document text data, business information data of bidding enterprises, and historical bidding behavior data; The graph structure update module is configured to: extract the semantic features of the bid text, the enterprise attribute features, and the behavior temporal features based on the acquired data to form 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 behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain the node embedding vector, calculate the correlation matrix between enterprises based on the node embedding vector, and thus predict whether there is collusion in bidding between enterprises.

[0017] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the above-mentioned bid-rigging prediction method based on a dynamic association network.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention constructs equity-related edges, historical cooperation edges, semantic collaboration edges, and behavioral collaboration edges between enterprise nodes, and dynamically updates the edge weights of each edge based on real-time data. It mines explicit and implicit associations between enterprises from different dimensions, enriching the modeling dimensions of the graph structure model; comprehensively considers multi-source data features such as text semantics, enterprise attributes, and behavioral time series, fully captures the implicit association relationships between enterprises, and improves the accuracy of identifying bid-rigging behaviors.

[0019] (2) The present invention can identify implicit collaboration such as sentence rewriting and synonym replacement through the relationship between the semantics of the bid text; it can capture the potential relationship changes between bidding entities in real time through dynamic information of corporate attributes such as corporate equity and qualification changes, and realize the dynamic update of the graph structure model; through behavioral time series data such as quotation volatility, it can transform implicit corporate collaborative behaviors into explicit graph relationships, thereby fully exploring the non-explicit associations between enterprises and realizing in-depth exploration of potential bid-rigging gangs.

[0020] (3) The present invention designs a two-layer graph neural network algorithm. The first layer aggregates direct neighbor nodes to capture the direct connections between enterprises. The second layer realizes global relationship fusion and mines the deep implicit connections between enterprises. It can identify the hidden bid-rigging behavior of "main enterprises and multi-layer vest enterprises" and avoid the risk of missed judgment.

[0021] Other features and advantages of additional aspects of the present invention will be given in part in the following description and in part will become obvious from the following description or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flowchart of a method for predicting bid rigging based on a dynamic association network in an embodiment of the present invention; Figure 2 Schematic diagram of a bid-rigging prediction system based on a dynamic association network in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0025] Example 1 In one or more embodiments, a method for predicting bid rigging based on a dynamic association network is disclosed. Figure 1 , specifically including the following process: S101: Construct a dynamic graph structure model. The graph structure model uses enterprises as nodes and constructs edges based on the equity relationship, historical cooperation relationship, behavioral collaboration relationship, and bid text semantic collaboration relationship between nodes. Each edge is accompanied by a real-time updated edge weight.

[0026] In this example, enterprises are used as nodes, and equity-related edges, historical collaboration edges, behavioral collaboration edges, and semantic collaboration edges are constructed based on the equity relationships, historical collaboration relationships, behavioral collaboration edges, and bid document semantic collaboration relationships between these nodes. By constructing multiple collaboration edges, we can simultaneously cover both explicit (equity, historical collaboration) and implicit (behavior, semantic collaboration) relationships, reducing missed detections. We also integrate multi-source data (industry and commerce, bidding records, behavioral time series, and text) to comprehensively characterize enterprise relationships. We also support dynamic weight updates to adapt to changing relationships and enhance robustness. We also cross-validate association signals to accurately identify complex bid-rigging patterns.

[0027] Each node contains basic attributes (such as registered capital, qualification level, equity penetration ratio, etc.), text attributes (bid semantic vector) and behavioral attributes (such as quotation volatility, etc.).

[0028] As a specific implementation, equity-linked edges are used to analyze equity relationships between companies and determine whether bid-rigging rings exist through equity ties. Equity-linked relationships (such as direct control, indirect shareholding, and cross-shareholding) reflect the capital ties and alignment of interests between companies, and are a key foundation for bid-rigging rings. Equity-linked edges can be used to quickly identify potential collusive entities, such as parent-subsidiary relationships and affiliated companies. These companies, due to their intertwined interests, are more susceptible to bid manipulation collusion.

[0029] The weight of the equity-related edge can accurately measure the strength of the association. For example, companies with direct holdings and highly similar equity structures have a significantly higher probability of colluding in bidding, providing a quantitative basis for identifying core gang members.

[0030] The edge weight of the equity-related edge is determined by multiplying the equity level weight and the equity structure similarity between enterprises, that is: ; in, The weight of the equity level, for example: when directly holding =0.8, when indirectly holding =0.5; is the equity structure similarity, and is the binary vector correlation coefficient.

[0031] Equity structure similarity The calculation process is as follows: (1) Determine the characteristic dimensions of the equity structure: Select a set number of characteristic dimensions that can reflect the essence of the equity structure (for example, 8 dimensions): Dimension 1: Whether there are natural person shareholders holding more than 30% of the shares; Dimension 2: Whether there is a legal person shareholder (enterprise) holding more than 50% of the shares; Dimension 3: whether there are multiple layers of nested holdings (level ≥ 3); Dimension 4: Whether the combined shareholding ratios of the top three shareholders are greater than 60% (concentrated ownership); Dimension 5: Whether there are foreign shareholders; Dimension 6: Whether there is state-owned capital holding; Dimension 7: Is the number of shareholders greater than 10 (dispersed equity)? Dimension 8: Is there cross-holding (e.g. A holds B, and B also holds A).

[0032] (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 a binary vector for the enterprise.

[0033] For example: The binary vector of enterprise A is ,in (Describe the characteristics of the specific dimension 1 of Enterprise A), (Explain the specific dimension 2 characteristics of enterprise A), and so on; the binary vector of enterprise B is , binary encoding is based on the same logic.

[0034] (3) Calculate the similarity between the binary vectors of two enterprises. The specific formula is: ; Among them, the numerator represents the number of characteristic dimensions shared by the two enterprises (i.e., the intersection of the binary vector of enterprise A and the binary vector of enterprise B); the denominator represents the number of characteristic dimensions shared by at least one of the two enterprises (i.e., the union of the binary vector of enterprise A and the binary vector of enterprise B).

[0035] The calculated similarity It can be used as the edge weight of the equity relationship edge between enterprise A and enterprise B.

[0036] As a specific implementation method, the historical cooperation edge is used to reflect the relationship between two companies in jointly bidding. The more times they jointly bid, the greater the risk of bid rigging. The edge weight of the historical cooperation edge is determined as follows: ; in, represents the number of times two companies bid together, is the attenuation factor, λ is the attenuation coefficient, The time since the last joint bid, in days.

[0037] Bidding riggers often manipulate bids through "short-term, high-frequency collaboration," while the relevance of long-term collaboration may weaken due to corporate strategy adjustments, market changes, etc. Therefore, the attenuation factor gives a higher weight to recent joint bids (the smaller t is, the higher the weight is). The closer it is to 1), the better it can reflect the current potential collaborative risk, avoid outdated information interfering with current risk judgment, and improve the model's sensitivity to recent behavior.

[0038] Structural similarity analysis can uncover hidden connections that appear independent but are actually controlled (e.g., circumventing regulation through dispersed shareholdings across multiple intermediary companies), addressing the blind spots of traditional identification methods that rely solely on superficial business information. Changes in equity structure (e.g., equity transfers, new shareholders) update edge weights in real time, enabling the model to promptly capture changes in corporate relationships and avoid under-detecting bid-rigging behaviors caused by equity adjustments.

[0039] As a specific implementation method, behavioral collaboration edges are used to reflect the behavioral collaboration relationship between two companies, such as whether the trends of bid fluctuations are the same. Through these collaborative relationships, the graph structure model can be more sensitive to collaborative behavior, which is conducive to identifying potential bid-rigging behaviors. The edge weight of the behavioral collaboration edge is determined as follows: (1) Obtain the target company's bidding and quotation data in at least N historical projects to form a time series , The quoted amounts of the projects corresponding to N (unit: ten thousand yuan) are sorted in chronological order by the bidding date.

[0040] Eliminate abnormal quotations (such as quotations that are significantly lower than the cost price or higher than the market price), standardize the quotations (such as subtracting the mean and dividing by the standard deviation) to eliminate dimensional effects.

[0041] (2) Set the time window length to (like , indicating that the trend is calculated based on the last five items), then the trend value Time window The average value of all quotes within.

[0042] (3) Calculation time window The relative deviation between the actual quotation and the trend value of each project , and get the quote volatility: , Represents a time window Actual quote for each item within.

[0043] By continuously moving the time window, the volatility of all quotes can be calculated.

[0044] (4) Obtain the quotation volatility series of enterprise A respectively and the quote volatility series of firm B ; Using Pearson correlation coefficient Measuring the correlation of volatility series: ; You can put | |Edge weight as the behavior coordination edge.

[0045] If | | is greater than the set threshold, and at least k projects have high volatility at the same time for both enterprise A and enterprise B, that is, >set value; then there is suspicion of price manipulation between Enterprise A and Enterprise B.

[0046] As a specific implementation method, the semantic collaborative edge is used to reflect the association relationship between the bid documents of two enterprises. The edge weight of the semantic collaborative edge can be determined by the similarity between the bid documents. The specific method is as follows: (1) Split the bid documents of Enterprise A and Enterprise B into sentence sequences respectively. The sentence sequence of Enterprise A is , the sentence sequence of enterprise B is ; and They all represent sentences, i=1,2,…,m, j=1,2,…,n.

[0047] (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.

[0048] (3) Calculate the semantic relevance between sentences in the two corporate bid documents to determine which sentences are semantically highly related: ; in, Expressing sentences and sentences The semantic relevance between is a learnable parameter matrix used to capture cross-text semantic interactions; It also represents the sentences in the sentence sequence of enterprise B, but it is different from It can be different. The larger the value of and sentences The closer the semantic connection between them.

[0049] (4) According to the above calculation method, generate The sentence-level alignment matrix S, the elements in the matrix S .

[0050] Based on the sentence-level alignment matrix S, the global implicit collaboration score of the bid texts of Enterprise A and Enterprise B is calculated: ; in, , Represents the importance weight of a sentence pair, which can be automatically generated through a hierarchical attention mechanism.

[0051] Global implicit collaboration scores of the bid documents of Enterprise A and Enterprise B It can be used as the edge weight of the semantic collaborative edge.

[0052] This embodiment maps the deep semantic structure sentence by sentence, which can not only identify implicit collaboration under sentence rewriting and synonym replacement (such as the semantic consistency between "cost control" and "expense control"), but also capture logical connections scattered in different chapters across paragraphs (such as the echo between the cooperation expression in the technical part and the support expression in the business part), realizing deep detection from "superficial similarity" to "intention alignment".

[0053] S102: Acquire bid document text data, business information data of bidding companies, and historical bidding behavior data.

[0054] In this embodiment, bid document text data is obtained directly from the bidding platform, including technical proposals, commercial quotations, project response terms, and other content. These texts contain a large amount of semantically coordinated information that may be involved in bid rigging, such as similar technical descriptions and unusually consistent commercial terms. Basic information about the bidding companies is also obtained, such as company name, unified social credit code, and registered address, to construct the basic attributes of the company node.

[0055] The bidding companies' business information is obtained from the company's business information database. This data, including shareholder information, shareholding ratios, and the number of layers of equity penetration, is used to analyze equity relationships between companies and determine whether there are bid-rigging rings formed through equity ties. Furthermore, information on the companies' qualifications, such as qualification type, level, and validity period, is collected and compared with the qualifications required for the project bidding to identify any violations such as reliance on other qualifications. For example, if the company's qualifications completely match the project's requirements, a score of 1 is assigned. A partial match is assigned a score based on the degree of match (e.g., 0.5). A mismatch is assigned a score of 0.

[0056] Establish a local or cloud-based database of historical bidding behavior, storing data such as bid times, bid amounts, and winning bids from past bidding activities. Patterns such as regularity in bid times and unusually close or divergent bids are important indicators for identifying bid-rigging. For example, if multiple companies bid very close together on multiple projects, this could indicate coordinated bidding. Regular fluctuations in bids, such as increases or decreases by a fixed margin, can also be a strong indicator of bid-rigging.

[0057] It should be noted that all of the above data can be obtained through legal means.

[0058] S103: Based on the acquired data, the semantic features of the bid text, the enterprise attribute features, and the behavior temporal features are extracted to form a feature vector 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.

[0059] Specifically, the semantic features of the bid text include the bid text semantic vector, with the dimension ; Enterprise attribute features include registered capital (normalized value), qualification level (unique hot encoding value) and equity penetration ratio of the vector, the dimension is ; The behavioral time series features include the quote volatility time series vector, with the dimension ; For each node, extract the above feature vectors separately, combine these feature vectors together to obtain the feature vector of each node, the dimension is .

[0060] Based on the acquired data, the edge weight determination method recorded in S101 is used to determine the edge weights of the equity-related edges, historical cooperation edges, behavioral collaboration edges, and semantic collaboration edges between each node.

[0061] The feature vector of each node and the calculated edge weight are assigned to the graph structure model to form an updated graph structure model.

[0062] S104: Using the updated graph structure model as input, the graph neural network model is used to obtain the node embedding vector, and the correlation matrix between enterprises is calculated based on the node embedding vector, so as to predict whether there is bid rigging between enterprises.

[0063] In this embodiment, the graph neural network model uses a two-layer GraphSAGE (SAmple and aggreGatE) algorithm; The first layer of the GraphSAGE algorithm mainly implements feature aggregation of neighbor nodes. The specific operation process is as follows: 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; Use the mean aggregation function to calculate the weighted average of neighbor features: ; in, is the initial eigenvector of neighbor node u.

[0064] After the node's own features are spliced ​​with the neighbor's aggregated features, the weight matrix Perform linear transformation and activate: || , where σ is the ReLU activation function, Represents feature splicing.

[0065] Finally, the first layer node embedding vector is obtained , the dimension is (generally , achieving feature dimensionality reduction).

[0066] The second-layer GraphSAGE algorithm mainly realizes global relationship fusion. The specific operation process is as follows: Update the number of samples ( Less than , focusing on core neighbors); Calculate the mean aggregate of neighbor features again: ; Represents the first-layer node embedding vector of neighbor node u after aggregation by the first-layer GraphSAGE algorithm.

[0067] Similar to the first layer, but the weight matrix is ​​updated to , and introduce modularity optimization items: ,in is the modularity function The gradient, is the regularization coefficient (e.g. 0.01), guiding the model to identify closely related community structures. The ratio of the sum of the edge weights within community i to the total edge weights is a key indicator for measuring the tightness of the internal connections within the community and is used to calculate the "difference between the community structure and the random network" (modularity is essentially a comparison of the community connectivity between the real network and the random network); is the ratio of the sum of edge weights connected to community i to the total edge weights.

[0068] A community structure refers to a tightly knit subgroup of nodes (companies). Nodes within a community are densely connected, while connections to external nodes are sparse. In bid-rigging scenarios, a community often represents a potential bid-rigging ring. Companies within the group form a highly interconnected network through various means, such as equity control, bid plagiarism, and coordinated bidding behavior, while maintaining weak connections with companies outside the ring.

[0069] In this embodiment, Added as a gradient term to GNN training ( The core of the modularity optimization term, ▽Q (G), measures the difference between the sum of edge weights within a community and the expected weights under random assignment. Maximizing Q causes the model to map groups of nodes with strong internal connections and sparse external connections to similar locations in low-dimensional space, forming a clustering structure. Therefore, the modularity optimization term can directly guide the model to identify these closely connected subgroups (i.e., community structures), thereby enabling the identification of hidden groups.

[0070] Finally, the second layer node embedding vector is obtained , the dimension is (like ), which has integrated the direct association and indirect conduction relationship of the nodes (such as the indirect association of ABC).

[0071] In this embodiment, the first layer of GraphSAGE algorithm is to sample direct neighbors ( Large), aggregates the direct correlation features of the nodes (such as direct equity and cooperative relationships between enterprises), and outputs the node embedding vector The local direct relationship information is already included. The second layer of GraphSAGE algorithm is based on For input, sampling Core neighbors ( < , focusing on more critical connections). In this case, the aggregated "neighbors" include not only the node's direct neighbors but also indirectly include the "neighbors' neighbors" (i.e., indirect connections) through the first-layer embedding vector. This multi-level aggregation allows features to convey longer-range connections (such as the indirect relationship between ABC), thus covering the global relationship network.

[0072] This implementation overcomes the limitations of traditional pairwise association analysis through two-layer aggregation, exploring complex relationships involving "direct associations + indirect transmissions" (e.g., companies A and C have no direct collaboration but form a potential connection through their shared neighbor B). Through end-to-end training, it automatically learns the weights for different edge types (equity-related edges, semantic collaboration edges, behavioral collaboration edges, and historical collaboration edges), avoiding the bias of manually pre-set weights. The second-layer aggregation of indirect neighbors can identify hidden groups involving "main culprit companies + multiple layers of front companies" (e.g., company D indirectly controls company G through companies E and F, forming a bid-rigging network), which traditional methods cannot identify due to their lack of deep relationship mining.

[0073] After two layers of GraphSAGE calculation, the output dimension of each node is Embedding vector of ,This vector combines the node’s own characteristics with the association characteristics of neighboring nodes, and is used to characterize the potential association attributes of enterprises in dynamic relationship networks.

[0074] Based on the node embedding vector, calculate the correlation matrix S between enterprises, where the elements Represents the strength of association between enterprise i and j; node embedding vector The model output integrates the enterprise's attribute characteristics, textual semantic characteristics, behavioral temporal characteristics, 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 the dynamic relationship network.

[0075] therefore, The essence of is the similarity of the embedding vectors of the two nodes - the higher the similarity, the stronger the consistency of the two companies in multi-dimensional correlation characteristics, and the higher the possibility of potential bid-rigging collaboration.

[0076] In this embodiment, The specific calculation method is as follows: ; in, is an element in the association matrix, indicating the strength of association between enterprise i and enterprise j; 、 are the node embedding vectors of the nodes corresponding to enterprise i and enterprise j respectively; is the average edge weight of all edges (equity-related edges, semantic collaboration edges, behavioral collaboration edges, and historical collaboration edges) between enterprise i and enterprise j.

[0077] It should be noted that if there is no direct boundary between the two companies (such as the first contact), (The default is neutral weighting to avoid completely ignoring potential hidden associations).

[0078] Since the node embedding vector and is a high-dimensional dense vector (such as , cosine similarity is used to measure directional consistency (reflecting the overlap of feature patterns), the formula is: , the value range is: [-1,1]. The closer it is to 1, the more consistent the directions of the embedding vectors of the two companies are (the more similar the associated features are).

[0079] In order to strengthen the influence of known explicit relationships (such as equity and historical cooperation), the edge weight mean between the two companies is introduced. as a correction factor.

[0080] 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 strength greater than the set threshold are judged to have engaged in bid rigging.

[0081] In this embodiment, during the training process of the graph neural network, a dynamic masking mechanism is used to improve the model's ability to capture sensitive semantic associations, targeting the implicit collusion and collaboration features of the bidding text (such as reused technical solutions, unusually consistent response logic, and obscure interest-related language).

[0082] Dynamic mask generation is divided into three levels: Dynamic masking of the text feature space: During the training phase, the input proposal text feature vector is masked according to a certain probability, and the mask position is dynamically adjusted. Unlike BERT's fixed masking, this mechanism dynamically generates masks based on the following rules: limited semantic density, collaborative feature enhancement, and temporal dependency awareness.

[0083] Adversarial training with multimodal feature fusion: Perturb the weights of semantically collaborative edges to simulate possible bid-rigging behavior (synonym replacement, parameter shift, and word order reorganization).

[0084] Time-series dynamic mask update: The masking strategy evolves over time. Every 10 epochs (rounds) of training, the masking strategy is dynamically adjusted based on the model's current recognition accuracy for bid-rigging samples. A time decay factor is introduced to reduce the masking probability for historically high-frequency mask positions to prevent the model from overfitting specific perturbation patterns.

[0085] Specifically, the calculation process of the dynamic mask rate is as follows: Defining dynamic masking rates , N is the length of the text sequence, T is the number of training rounds; , Indicates the total number of times the i-th position has been masked in historical training; is the recognition accuracy of the colluded bidding samples after the tth round of training; To control the decay rate of the historical mask frequency; is the initial mask probability without adjustment; is a function that dynamically adjusts the mask strength according to the accuracy.

[0086] (1) Historical mask frequency attenuation: , is an indicator function, which takes 1 when the i-th position is masked in round t-1, and takes 0 otherwise; this formula indicates that the historical frequency decays exponentially over time, while accumulating the mask records of the current round.

[0087] (2) Accuracy adjustment coefficient: , To adjust the amplitude (such as =0.2), is the target accuracy (e.g. =0.9), when the accuracy is lower than the target, the mask probability is increased to strengthen the training; when it is higher than the target, the mask probability is reduced to avoid overfitting.

[0088] (3) Dynamic mask rate calculation: , is the frequency attenuation index (e.g. = 0.1), which controls the probability decay rate of the historical high-frequency mask position. This formula combines accuracy adjustment and historical frequency decay to ensure that the mask position changes dynamically with training.

[0089] The strategy update mechanism for every 10 epochs (rounds): When t mod 10 = 0, the historical frequency matrix is ​​reset. , weaken the mask record 10 epochs ago through exponential decay; dynamically adjust the basic mask probability: , is the adaptive coefficient, which adjusts the initial mask strength according to the accuracy changes in the past 10 rounds. If the accuracy improves, reduce To reduce mask disturbance; if it decreases, increase To increase the difficulty of training.

[0090] The specific process of modularity optimization is as follows: Modularity function , as a regularization term of the graph neural network, it guides the model to identify closely related community structures and reduce the missed detection rate.

[0091] in, is the ratio of the sum of the 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 between enterprises in community i is 0.3, and the total edge weight is 1.0, then =0.3.

[0092] is the ratio of the sum of the edge weights connected to community i to the total edge weights. For example, if the sum of the edge weights between community i and external enterprises is 0.2, then =0.2. In this embodiment, whenever a new bidding record is generated or the business information of an enterprise is changed, the following updates are triggered: Calculate various edge weights (equity, collaboration, semantics, behavior) between new companies and existing nodes. Update only the affected edge weights and corresponding node embeddings to avoid full retraining.

[0093] Adversarial examples simulating novel bid-rigging relationships (such as fictitious corporate connections involving "dispersed equity and cross-regional bidding") are regularly generated and injected into the relationship network for robustness training, improving the ability to identify unknown relationship patterns. (Adversarial examples are artificially constructed data samples that simulate novel, hidden bid-rigging relationships in reality. These samples appear to be normal bidding data on the surface, but in fact contain carefully designed relationship information that resembles bid-rigging characteristics. Injecting them into the relationship network for training aims to enable the model to learn to identify these hidden abnormal patterns, enhancing its ability to resist various potential bid-rigging behaviors, thereby improving the model's robustness and generalization ability.)

[0094] As a specific example, the process of constructing adversarial samples based on data features is as follows: Regarding equity structure: Public information from industrial and commercial data is used to fabricate complex structures of dispersed equity among companies. For example, multiple fictitious intermediate holding companies are created, each holding a small stake in the target company, creating a seemingly dispersed ownership structure that is actually controlled by a single entity. Assuming a real-world bid-rigging ring holds equity in a core bidding company through multiple layers of nested companies, this structure is simulated to generate adversarial examples. For example, Company A indirectly holds 5%, 3%, and 4% of Company E's equity through three seemingly unrelated companies, B, C, and D, respectively. This creates the illusion of dispersed ownership, but the actual controller is the same person. This equity structure data is used as part of the adversarial example.

[0095] Bidding behavior: We simulate unusual cross-regional bidding patterns. Based on the normal regional distribution and frequency of bidding in historical bidding data, we create unusual bidding behavior. For example, in a certain region, typical companies often bid on local projects but less frequently on remote projects. When constructing adversarial examples, we assume that some companies frequently bid on cross-regional projects, with regular intervals between bids that differ significantly from typical corporate bidding behavior. For example, Company F frequently bids on multiple projects far from its registered location within a short period of time, with each bid occurring within a week, which is inconsistent with typical bidding habits of local companies.

[0096] Semantic collaboration: Using natural language processing technology, we rewrite bid documents. We extract common expressions from a large number of real bids and, through methods like synonym replacement and sentence structure transformation, construct bid documents with similar semantics but distinct wording. For example, we rewrite "Our company possesses advanced technical solutions" to "Our company possesses cutting-edge technical response strategies." This creates semantic collaboration across multiple bids, yet is difficult to detect from the surface text. These processed bid documents serve as semantic features of adversarial examples and are injected into the relational network, training the model to identify these implicit semantic collaborations.

[0097] As a specific example, the process of constructing adversarial samples combined with attack algorithms is as follows: We apply the principles of adversarial example generation algorithms based on images to relational network data. Taking a gradient-based approach as an example, we calculate the gradient information of the relational network model when processing current, normal bidding data. Then, we apply small perturbations to the equity correlation, bidding behavior characteristics, and semantic similarity in the data in a direction that causes the model to misidentify. Assuming the model's judgment of normal equity correlation is relatively stable, we use an algorithm to add small perturbations to the equity correlation data, causing the model's judgment to deviate from this data. This simulates new types of bid-rigging relationships that are difficult for the model to identify, thus constructing adversarial examples.

[0098] This embodiment introduces a graph neural network (GNN) to dynamically update the weights of the associated edges between enterprises, and combines multi-source data (text 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 collusion in bids.

[0099] Example 2 In one or more embodiments, a bid-rigging prediction system based on a dynamic association network is disclosed. Figure 2 , specifically including: 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 bid document text semantic collaboration relationships between nodes, with each edge accompanied by a real-time updated edge weight; The data acquisition module is configured to: acquire bid document text data, business information data of bidding enterprises, and historical bidding behavior data; The graph structure update module is configured to: extract the semantic features of the bid text, the enterprise attribute features, and the behavior temporal features based on the acquired data to form 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 behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain the node embedding vector, calculate the correlation matrix between enterprises based on the node embedding vector, and thus predict whether there is collusion in bidding between enterprises.

[0100] It should be noted that the specific implementation of the above modules is exactly the same as that in Example 1 and will not be described in detail.

[0101] Example 3 In one or more embodiments, a terminal device is disclosed, which includes a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, and the instructions are suitable for the processor to load and execute the bid-rigging prediction method based on a dynamic association network described in Example 1.

[0102] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0103] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0104] During implementation, each step of the above method may be completed by an integrated logic circuit of hardware in a processor or by instructions in the form of software.

[0105] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it 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 on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting bid rigging based on a dynamic association network, characterized in that: include: Build a dynamic graph structure model that uses enterprises as nodes and constructs edges based on equity relationships, historical cooperation relationships, behavioral collaboration relationships, and bid document text semantic collaboration relationships between nodes. Each edge is accompanied by a real-time updated edge weight. Obtain bid document data, bidding company business information data, and historical bidding behavior data; Based on the acquired data, the semantic features of the bid document text, the enterprise attribute features, and the behavioral temporal features are extracted to form the feature vector of 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. Taking the updated graph structure model as input, the graph neural network model is used to obtain the node embedding vector. Based on the node embedding vector, the correlation matrix between enterprises is calculated to predict whether there is bid rigging between enterprises.

2. The method for predicting bid rigging based on a dynamic association network according to claim 1, characterized in that: The equity-related edge is constructed based on the equity-related relationship between the nodes. The edge weight of the equity-related edge is determined by multiplying the equity level weight and the equity structure similarity between the enterprises. The method for determining the equity structure similarity is: Construct a characteristic dimension that can reflect the essence of the equity structure, generate a binary vector corresponding to the enterprise based on the characteristic dimension, and calculate the similarity of the binary vectors between enterprises as the equity structure similarity.

3. The method for predicting bid rigging based on a dynamic association network according to claim 1, characterized in that: The historical cooperation relationship between nodes is used to construct a historical cooperation edge. The edge weight of the historical cooperation edge is determined as follows: ； in, is the number of joint biddings by enterprises, λ is the decay coefficient, and t is the time since the last joint bidding.

4. The method for predicting bid rigging based on a dynamic association network according to claim 1, wherein: A behavioral collaborative edge is constructed based on the behavioral collaborative relationship between nodes. The method for determining the edge weight of the behavioral collaborative edge is as follows: determining the time series of the bid volatility of the enterprise in the most recent set number of bids; calculating the Pearson correlation coefficient of the bid volatility of the two enterprises as the edge weight of the behavioral collaborative edge.

5. The method for predicting bid rigging based on a dynamic association network according to claim 1, wherein: The semantic collaborative relationship between the bid documents is used to construct a semantic collaborative edge. The edge weight of the semantic collaborative edge is determined as follows: Split the enterprise bid texts corresponding to nodes A and B into sentence sequences respectively; Use the pre-trained model in the bidding field to convert each sentence into a continuous numerical vector; Calculate the semantic relevance between different sentences in the enterprise bid text corresponding to the two nodes and generate a sentence-level alignment matrix S; Based on the sentence-level alignment matrix S, the global implicit collaboration score of the enterprise bid text corresponding to node A and node B is calculated as the edge weight of the semantic collaboration edge between node A and node B.

6. The method for predicting bid rigging based on a dynamic association network according to claim 1, wherein: The bid document text semantic features include a bid document text semantic vector; the enterprise attribute features include a vector composed of registered capital, qualification level and equity penetration ratio; and the behavior time series features include a quote volatility time series vector.

7. The method for predicting bid rigging based on a dynamic association network according to claim 1, wherein: The correlation matrix S between enterprises is calculated based on the node embedding vector, specifically: ； in, is an element in the association matrix, indicating the strength of association between enterprise i and enterprise j; 、 are the node embedding vectors of the nodes corresponding to enterprise i and enterprise j respectively; is the mean weight of all edges between enterprise i and enterprise j.

8. The method for predicting bid rigging based on a dynamic association network according to claim 1, wherein: The graph neural network model includes a two-layer graph neural network algorithm. The first layer of the graph neural network algorithm randomly samples a fixed number of nodes from its neighboring nodes for each node. The neighbors of the node are calculated using the mean aggregation function to calculate the weighted average of the neighbor features. After the node's own features are spliced ​​with the neighbor aggregation features, the weight matrix Perform linear transformation and activation to obtain the first layer node embedding vector ; The second-layer graph neural network algorithm uses the first-layer node embedding vector As the initial features of the node itself, a fixed number of random samples are taken from each node’s neighboring nodes. Neighbors, < ; Use the mean aggregation function to calculate the weighted average of neighbor features, concatenate the node's own features with the neighbor aggregation features, and then use the weight matrix Perform linear transformation and introduce modularity optimization to guide the model to identify closely related community structures; finally activate and obtain the first-layer node embedding vector .

9. A bid-rigging prediction system based on a dynamic association network, characterized by: 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 bid document text semantic collaboration relationships between nodes, with each edge accompanied by a real-time updated edge weight; The data acquisition module is configured to: acquire bid document text data, business information data of bidding enterprises, and historical bidding behavior data; The graph structure update module is configured to: extract the semantic features of the bid text, the enterprise attribute features, and the behavior temporal features based on the acquired data to form 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 behavior prediction module is configured to: take the updated graph structure model as input, use the graph neural network model to obtain the node embedding vector, calculate the correlation matrix between enterprises based on the node embedding vector, and thus predict whether there is collusion in bidding between enterprises.

10. A terminal device comprising a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the bid-rigging prediction method based on a dynamic association network as described in any one of claims 1-8.

Images