Bidding enterprise abnormal behavior detection method based on bidding and tendering data

Abstract

The invention discloses a bidding enterprise abnormal behavior detection method based on bidding and tendering data, and the method comprises the steps: firstly constructing a bidding enterprise relation network based on bidding and tendering records, and carrying out the community detection of the bidding enterprise relation network under different scales, and obtaining a community set under the corresponding scale; secondly, on the basis of community sets under different scales, robust enterprise groups which stably appear under multiple scales are identified; and then constructing a bidding enterprise group hierarchy tree based on the robust enterprise groups, and revealing inclusion relationships and hierarchy associations among different bidding enterprise groups. And finally, traversing the bidding enterprise group hierarchical tree to screen candidate abnormal enterprise groups, and carrying out abnormal bidding behavior collaborative analysis on the candidate abnormal enterprise groups in combination with the bidding flow diagram and the sub-group event sequence diagram. The method effectively breaks through the limitation that a traditional method only depends on bidding content similarity or single-scale community detection, and has remarkable advantages in the aspects of robustness, interpretability and recognition precision.

Description

Technical Field

[0001] This invention belongs to the field of enterprise abnormal behavior detection technology, specifically relating to a method for detecting abnormal behavior of bidding enterprises based on bidding data. Background Technology

[0002] The bidding system is the core mechanism for resource allocation in engineering construction and government procurement, and its fairness and impartiality are directly related to the healthy operation of the market economy. With intensifying market competition, illegal activities such as bid rigging and collusion are becoming increasingly covert, organized, and complex. Unscrupulous companies often manipulate bidding results by forming alliances of interest and setting up dummy bids, seriously disrupting market order. Using big data technology to detect abnormal behavior by bidding companies has become an important technical means for regulatory authorities to identify clues of violations.

[0003] Some methods, such as the patent application number 202410504274.8 (A method and system for identifying and analyzing bid-rigging behavior based on big data), determine whether bid-rigging behavior exists by comparing the similarity of the total bid price, price list of main items, and project plan provided by the bidders. This type of judgment method based on explicit content similarity is easily circumvented by companies through strategies such as price manipulation or text spoofing. Other methods, such as the patent application number 201911151844 (A method for identifying and analyzing bid-rigging behavior based on big data), construct a community relationship network among participating companies based on bidding data and use community detection algorithms to mine potential company groups, thereby determining whether bid-rigging behavior exists. However, in actual bidding scenarios, companies from different industries or business areas often participate in the same type of project, naturally forming small-scale clusters. These industry clusters are not equivalent to bid-rigging groups; failure to distinguish between them will lead to misidentification. Summary of the Invention

[0004] This invention proposes a method for detecting abnormal behavior of bidding companies based on bidding data. The method first constructs a relationship network of bidding companies based on historical bidding records. Then, it performs community detection on this network at different scales to identify robust company groups that consistently appear across multiple scales. Further, by analyzing the inclusion relationships among these robust company groups, a hierarchical tree of bidding company groups is constructed, thereby identifying closely related bidding company groups at different levels. Finally, a multi-view collaborative visualization view is designed to comprehensively present the bidding characteristics of different company groups, thus assisting in the detection of abnormal behavior by bidding companies.

[0005] A method for detecting abnormal behavior of bidding companies based on bidding data includes the following steps:

[0006] Step 1: Build a network of relationships among bidding companies based on bidding records.

[0007] Obtain historical bidding records and build a network of relationships among bidding companies. ,in Let the set of nodes correspond to all participating companies, and the set of edges be... It encodes the relationship between any two companies participating in a bid. Adjacency matrix. elements in The connection weight between nodes i and j is defined as the cumulative number of projects in which the two companies jointly participate in bidding.

[0008] Step 2: Conduct community testing on the relationship network of bidding companies at different scales to obtain the community set at the corresponding scale.

[0009] Within a given resolution range Within, a series of resolution parameters are selected with a fixed step size. For each resolution parameter With a generalized modularity function with a resolution parameter As the optimization objective, the Louvain community detection algorithm is used to divide the bidding companies' relationship network into communities, obtaining the community structure under the corresponding resolution parameters. ,

[0010] In a single resolution parameter The resulting community structure This study describes the closely related bidding groups of firms at the analytical scale induced by this resolution parameter, characterizing the joint bidding relationships among firms at this scale. As the resolution parameter changes, the original bidding firm relationship network is mapped into a series of community sets induced by different resolution parameters. This allows us to characterize the organizational structure of corporate joint bidding relationships under multi-scale conditions.

[0011] Step 3: Based on community sets at different scales, identify robust corporate groups that appear stably at multiple scales.

[0012] Step 3.1: Construct a cross-scale enterprise group graph based on community sets at different scales. The similarity between communities was detected at different scales.

[0013] Step 3.2: Based on the cross-scale robustness of connected components, extract robust enterprise groups from the connected components with high robustness, indicating that they have close relationships at different scales.

[0014] Step 4: Construct a hierarchical tree of bidding enterprise groups based on robust enterprise groups to reveal the inclusion relationships and hierarchical connections between different bidding enterprise groups.

[0015] Step 4.1: Construct an initial hierarchical structure diagram H based on the robust enterprise group.

[0016] Step 4.2: Optimize the hierarchical structure diagram H to obtain the final bidding enterprise group hierarchical tree.

[0017] Step 5: Traverse the hierarchical tree of bidding enterprise groups, extract the generalized modularity value of each enterprise group, and filter out candidate abnormal enterprise groups based on global sorting; for candidate abnormal enterprise groups, conduct collaborative analysis of abnormal bidding behavior by combining the bidding flow graph and the subgroup event sequence graph.

[0018] Step 5.1: Traverse all nodes in the bidding enterprise group hierarchy tree and extract the generalized modularity value of each node calculated by the community detection algorithm in Step 2; perform global sorting of subgroups at different levels in the hierarchy tree according to the generalized modularity value, and select the top K% of subgroups as candidate abnormal enterprise groups.

[0019] Step 5.2 For candidate abnormal enterprise groups, the joint bidding flow diagram and subgroup event sequence diagram are used to analyze the specific abnormal bidding behaviors within the enterprise groups from a micro perspective.

[0020] Advantages and effects of this application compared to existing technologies

[0021] This paper proposes a method for detecting abnormal behavior of bidding companies based on bidding data, effectively overcoming the limitations of traditional methods that rely solely on the similarity of bid content or single-scale community detection. The method constructs a relationship network of bidding companies based on historical bidding records and identifies recurring robust company groups at multiple scales. Building upon this, a hierarchical tree of bidding company groups is constructed to reveal the inclusion and relationship strength between groups of different sizes. Combined with multi-view collaborative visualization, this achieves more accurate identification of suspicious company groups. This method exhibits significant advantages in robustness, interpretability, and recognition accuracy, providing a novel technical approach for detecting abnormal behavior in bidding. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention;

[0023] Figure 2 This is a diagram of the overall network structure.

[0024] Figure 3 A hierarchical structure diagram of the enterprise group;

[0025] Figure 4 This is a diagram showing the bidding flow.

[0026] Figure 5 This is a sequence diagram of events for subgroups. Detailed Implementation

[0027] like Figure 1 As shown in the attached diagram, the following is a detailed explanation of a method for detecting abnormal behavior of bidding companies based on bidding data, including the following steps:

[0028] Step 1: Build a network of relationships among bidding companies based on bidding records.

[0029] Retrieve historical bidding records, including fields such as project number, project name, bidding company name, and tendering entity name. Build a network of relationships among bidding companies. ,in Let the set of nodes correspond to all participating companies, and the set of edges be... It encodes the relationship between any two companies participating in a bid. Adjacency matrix. elements in The connection weight between nodes i and j is defined as the cumulative number of projects in which the two companies jointly participate in bidding.

[0030] Step 2: Conduct community testing on the relationship network of bidding companies at different scales to obtain the community set at the corresponding scale.

[0031] Within a given resolution range Within, a series of resolution parameters are selected with a fixed step size. For each resolution parameter With a generalized modularity function with a resolution parameter As the optimization objective, the Louvain community detection algorithm is used to divide the bidding companies' relationship network into communities, obtaining the community structure under the corresponding resolution parameters. The generalized modularity function Defined as:

[0032]

[0033] In the formula, The adjacency matrix weights and Let m represent the weighted degree of nodes i and j, and m be half the sum of the edge weights of the bidding company's relationship network. This is used to determine whether nodes i and j are assigned to the same community. If they belong to the same community, then... ,otherwise .

[0034] In a single resolution parameter The resulting community structure This study describes the closely related bidding groups of firms at the analytical scale induced by this resolution parameter, characterizing the joint bidding relationships among firms at this scale. As the resolution parameter changes, the original bidding firm relationship network is mapped into a series of community sets induced by different resolution parameters. This allows us to characterize the organizational structure of corporate joint bidding relationships under multi-scale conditions.

[0035] Step 3: Based on community sets at different scales, identify robust corporate groups that appear stably at multiple scales.

[0036] Step 3.1: Construct a cross-scale enterprise group graph based on community sets at different scales. The similarity between communities was detected at different scales.

[0037] Building cross-scale enterprise group diagrams Each node corresponds to a community detected at a different scale, and the node set is as follows: The edges encode the similarity between different nodes. community and Based on the node sets of both and The similarity is calculated using the Jaccard coefficient:

[0038]

[0039] in and Representing communities and The set of nodes contained therein This indicates the number of nodes in the set. It measures the degree of overlap in the membership composition of two communities. Based on the Jaccard similarity threshold. ,like This indicates that the two communities have similar community structures at different scales. An edge is created between two elements, and no edge is created between the two elements.

[0040] Then, by traversing the cross-scale corporate group diagram All connected components in , where each connected component represents a set of similar communities that repeat at different resolutions.

[0041] Step 3.2: Based on the cross-scale robustness of connected components, extract robust enterprise groups from the connected components with high robustness, indicating that they have close relationships at different scales.

[0042] The number of different resolutions covered by a connected component is defined as its cross-scale robustness index. If the number of resolutions covered by a connected component exceeds a given threshold... This indicates that communities within the connected component can be identified at multiple scales. Based on this, in The system filters out all connected components that meet the robustness requirements.

[0043] For each connected component that satisfies the condition, let its set of communities be denoted as . If a node (bidding company) appears more than a threshold number of times in the community of set C. If a core member is identified, it is considered a core member of that connected component. Constructing a robust enterprise group using these core members demonstrates the strong stability of its membership structure and its ability to be repeatedly detected across multiple scales.

[0044] Step 4: Construct a hierarchical tree of bidding enterprise groups based on robust enterprise groups to reveal the inclusion relationships and hierarchical connections between different bidding enterprise groups.

[0045] Step 4.1: Construct an initial hierarchical structure diagram H based on the robust enterprise group.

[0046] Construct an initial hierarchical graph H based on robust firm groups, where each node represents a robust firm group. Add a root node to represent the set of all robust firm groups. For any two robust firm groups... and Calculate its inclusion index:

[0047]

[0048] in, and Representing stable corporate groups and The set of nodes. express What percentage of nodes appear in In the middle. If Greater than a given threshold Then add a line from H point to A directed edge indicates that yes Subgroups, otherwise no edges are added.

[0049] Step 4.2: Optimize the hierarchical structure diagram H to obtain the final bidding enterprise group hierarchical tree.

[0050] Simplify the hierarchical diagram H. If... and ,in If the index threshold is included, it indicates that the two robust groups of firms are highly similar. and Merged into a new corporate group And add it to H, while removing the original community node from H. and Subsequently, a transitive reduction operation is performed on H to remove redundant containing edges, retaining only the most direct hierarchical connections, making the overall structure clearer and simpler.

[0051] After simplifying the structure, the node set of each parent node is expanded downwards by closure, that is, all the nodes contained in its child nodes are included in the parent node, thus constructing the final bidding enterprise group hierarchy tree. The leaf nodes represent the most closely related enterprise subgroups in the bidding enterprise relationship network, while the upper-level nodes correspond to enterprise groups with a wider coverage and a more macro structure.

[0052] Step 5: Traverse the hierarchical tree of bidding enterprise groups, extract the generalized modularity value of each enterprise group, and filter out candidate abnormal enterprise groups based on global sorting; for candidate abnormal enterprise groups, conduct micro-level collaborative analysis of abnormal bidding behavior by combining the bidding flow graph and the subgroup event sequence graph.

[0053] Step 5.1: Traverse all nodes in the bidding enterprise group hierarchy tree and extract the generalized modularity value of each node calculated by the community detection algorithm in Step 2; perform global sorting of subgroups at different levels in the hierarchy tree according to the generalized modularity value, and select the top K% of subgroups as candidate abnormal enterprise groups.

[0054] Step 5.2 For candidate abnormal enterprise groups, the joint bidding flow diagram and subgroup event sequence diagram are used to analyze the specific abnormal bidding behaviors within the enterprise groups from a micro perspective.

[0055] For candidate abnormal enterprise groups, a bidding flow diagram is constructed. The left vertical bars represent bidding enterprises, and the right vertical bars represent bidding units. The total width of the flow line between them represents the total number of bids made by the bidding enterprises for the corresponding bidding units. The width of the flow line, indicated by different colors, corresponds to the number of successful bids. Based on this, the bidding concentration of the current enterprise group for a specific bidding unit is calculated, defined as the ratio of the sum of the number of bids made by the group members for that bidding unit to the total number of bids made by the group. If the bidding concentration is greater than a preset threshold α, it is determined that the bidding flow of the enterprise group is highly concentrated in a few bidding units. If the bidding flow of a certain enterprise group member is highly concentrated in a few bidding units, and the number of successful bids of a few internal members shows an abnormal monopolistic distribution, then the enterprise group is highly suspected of bid rigging.

[0056] For candidate groups of companies exhibiting abnormal behavior, a subgroup event timeline is constructed. The horizontal axis represents the bid opening time, and the vertical axis represents all bidding companies within a subgroup. Each cell is represented by a pie chart with an outer ring. The pie chart is divided according to the projects that bidding companies participated in on the bid opening day: different sectors in the pie chart correspond to those who participated in bidding but did not win, and those who won projects. The outer ring color is used to distinguish the bidding entities corresponding to different projects. Based on the changing patterns of bidding results over time, the winning roles of companies in different projects or at different times are analyzed to determine whether there is any suspicion of rotating winning bids. If some companies show in the timeline that they continuously participate in bidding but never win a bid, they are suspected of bid-rigging.

[0057] The views described above support flexible interactive linkage. For example, clicking on a subgroup in the enterprise group hierarchy diagram will simultaneously update the corresponding subgroup event sequence diagram and bidding flow diagram. Clicking on a node in the overall network structure diagram will display the specific information of the corresponding enterprise. Through multi-perspective collaborative analysis, users can comprehensively explore the bidding group hierarchy, thereby effectively identifying potentially abnormal bidding enterprise groups.

[0058] Example:

[0059] An analysis of the historical records of a certain bidding system was conducted. The dataset contains more than 1,300 bidding records, involving 512 bidding companies and 89 tendering units.

[0060] Figure 2 This is the overall network structure diagram. All bidding companies are divided into several subgroups based on their industry sector. The color of each subgroup is coded according to the modularity of its corresponding subnetwork. Purple subgroups indicate closer bidding relationships between the companies. (See diagram.) Indicates the first One first-level enterprise group, among which For hierarchical indexes, This is used as a group number within the same level.

[0061] Figure 3 This diagram presents a hierarchical structure of bidding companies, visually illustrating the hierarchical tree of their groups. It allows for understanding the hierarchical relationships between subgroups, with the color of each node indicating the modularity of its corresponding subnetwork, thus revealing the degree of connection between subgroups. (See diagram.) The number before the hyphen represents the level of the identified enterprise group in the hierarchy tree, and the number after the hyphen represents the group number at that level.

[0062] Figure 4To identify an abnormal group of companies, a bidding flow diagram is used. The eight vertical bars on the left correspond to eight bidding companies, and the vertical bars on the right correspond to the tendering entities. The yellow flow line corresponds to the total number of bids, and the green flow line in the middle corresponds to the number of successful bids. The diagram shows that all bidding activities of the eight companies on the left are directed to the same tendering entity, U61. Furthermore, these companies did not participate in bidding for any other tendering entities besides U61. Among all the bidding companies, company BK never won a bid, while company BI had the highest overall success rate.

[0063] Figure 5 for Figure 4 A sequence diagram of events in the subgroups corresponding to the anomalous group. (and...) Figure 4 The color meaning of the streamline should remain consistent. Figure 5 In the pie chart, the yellow sectors represent those participating in the bidding, and the green sectors represent those winning bids. The chart shows that the bidding companies within this group participated in the bidding for tenderer U61 at different times, and these companies participated in multiple projects for tenderer U61 on the same day. The chart clearly shows the participation in each bidding process and the winning companies.

Claims

1. A method for detecting abnormal behavior of bidding enterprises based on bidding data, characterized in that, Includes the following steps: Step 1: Construct a network of relationships among bidding companies based on bidding records; Step 2: Perform community detection on the relationship network of bidding companies at different scales to obtain the community set at the corresponding scale; Step 3: Based on community sets at different scales, identify robust corporate groups that consistently appear at multiple scales; Step 4: Construct a hierarchical tree of bidding enterprise groups based on robust enterprise groups to reveal the inclusion relationships and hierarchical connections between different bidding enterprise groups; Step 5: Traverse the hierarchical tree of bidding enterprise groups, filter out candidate abnormal enterprise groups, and combine the bidding flow diagram and the subgroup event sequence diagram to conduct collaborative analysis of abnormal bidding behavior of candidate abnormal enterprise groups.

2. The method for detecting abnormal behavior of bidding enterprises based on bidding data according to claim 1, characterized in that, The specific implementation of constructing the bidding company relationship network is as follows: obtaining historical bidding records and constructing the bidding company relationship network. ,in Let the set of nodes correspond to all participating companies, and the set of edges be... Encodes the relationship between any two companies participating in a bid; adjacency matrix elements in The connection weight between nodes i and j is defined as the cumulative number of projects in which the two companies jointly participate in bidding.

3. The method for detecting abnormal behavior of bidding enterprises based on bidding data according to claim 2, characterized in that, Step 2 is implemented as follows: Within a given resolution range Within, a series of resolution parameters are selected with a fixed step size. ; For each resolution parameter With a generalized modularity function with a resolution parameter As the optimization objective, the Louvain community detection algorithm is used to divide the bidding companies' relationship network into communities, obtaining the community structure under the corresponding resolution parameters. Generalized modularity function Defined as: ， In the formula, and Let m represent the weighted degree of nodes i and j, and m be half the sum of the edge weights of the bidding company's relationship network. This is used to determine whether nodes i and j are assigned to the same community. If they belong to the same community, then... ,otherwise ; In a single resolution parameter The resulting community structure This describes the closely related bidding groups of enterprises at the analytical scale induced by the resolution parameter, and characterizes the joint bidding relationships of enterprises at this scale. As the resolution parameter changes, the original bidding enterprise relationship network is mapped into a series of community sets induced by different resolution parameters. This characterizes the organizational structure of joint bidding relationships among enterprises under multi-scale conditions.

4. The method for detecting abnormal behavior of bidding enterprises based on bidding data according to claim 3, characterized in that, Step 3 is implemented as follows: Step 3.1: Construct a cross-scale enterprise group graph based on community sets at different scales. The similarity between communities was detected at different scales using encoding. Building cross-scale enterprise group diagrams Each node corresponds to a community detected at a different scale, and the node set is as follows: Edge encoding determines the similarity between different nodes; for any two nodes from different resolutions... community and Based on the node sets of both and Calculating similarity using the Jaccard coefficient ; Based on Jaccard similarity threshold ,like This indicates that the two communities have similar community structures at different scales. An edge is established between two of them, and no edge is established if they are not in the same position. Then, by traversing the cross-scale corporate group diagram Identify All connected components in the array, each connected component represents a set of similar communities that repeat at different resolutions; Step 3.2: Based on the cross-scale robustness of connected components, extract robust enterprise groups, showing that they have close relationships at different scales; The number of different resolutions covered by a connected component is defined as its cross-scale robustness index; if the number of resolutions covered by a connected component is greater than a given threshold... This indicates that communities within the connected component can be identified at multiple scales; based on this, in The system filters out all connected components that meet the robustness requirements. For each connected component that satisfies the condition, let its set of communities be denoted as . ; If a node appears more than a threshold number of times in the community of set C. If so, then it is considered a core member of the connected component; A robust corporate group built around core members indicates that its membership structure is highly stable and can be repeatedly detected across multiple scales.

5. The method for detecting abnormal behavior of bidding enterprises based on bidding data according to claim 4, characterized in that, Step 4 is implemented as follows: Step 4.1: Construct the initial hierarchical structure diagram H based on the robust enterprise group: Construct an initial hierarchical graph H based on robust firm groups, where each node represents a robust firm group; add a root node to represent the set of all robust firm groups; for any two robust firm groups... and Calculate its inclusion index: in, and Representing stable corporate groups and The set of nodes; express What percentage of nodes appear in In the middle; if Greater than a given threshold Then add a line from H point to A directed edge indicates that yes Subgroups, otherwise no edges are added; Step 4.2: Optimize the hierarchical diagram H to obtain the bidding company group hierarchy tree: Simplify the hierarchical diagram H. If the following occurs... and ,in If the index threshold is included, it indicates that the two robust groups of firms are highly similar; and Merged into a new corporate group And add it to H, while removing the original community node from H. and Subsequently, a transitive reduction operation is performed on H to remove redundant containing edges, retaining only the most direct hierarchical connections. After simplifying the structure, the node set of each parent node is expanded downwards by closure, that is, all the nodes contained in its child nodes are included in the parent node, thus constructing the final bidding enterprise group hierarchy tree. The leaf nodes represent the most closely related enterprise subgroups in the bidding enterprise relationship network, while the upper-level nodes correspond to enterprise groups with a wider coverage and a more macro structure.

6. The method for detecting abnormal behavior of bidding enterprises based on bidding data according to claim 5, characterized in that, Step 5 is implemented as follows: Step 5.1: Traverse all nodes in the bidding enterprise group hierarchy tree and extract the generalized modularity value of each node calculated by the community detection algorithm in Step 2; perform global sorting of subgroups at different levels in the hierarchy tree according to the generalized modularity value, and select the top K% of subgroups as candidate abnormal enterprise groups. Step 5.2: For candidate anomalous enterprise groups, analyze the specific anomalous bidding behaviors within the enterprise groups from a micro-level perspective using the joint bidding flow diagram and subgroup event sequence diagram: For candidate abnormal enterprise groups, a bidding flow diagram is constructed. The left vertical bar represents the bidding enterprise, the right vertical bar represents the bidding unit, and the total width of the flow line between the two represents the total number of bids of the bidding enterprise on the corresponding bidding unit. The width of the flow line in the middle, which is represented by different colors, corresponds to the number of times the bid is won. Based on this, the current concentration of bids by the enterprise group to the bidding entity is calculated, which is defined as the proportion of the sum of the number of bids made by the group members to the bidding entity to the total number of bids made by the group. If the concentration of bids is greater than the preset threshold α, it is determined that the bid flow of the enterprise group is highly concentrated in a few bidding units; If the bidding flow of a group of companies is highly concentrated in a few bidding units, and the number of winning bids by internal members shows an abnormal monopolistic distribution, then the group of companies is suspected of bid rigging. For candidate abnormal enterprise groups, a subgroup event timeline is constructed. The horizontal axis represents the bid opening time, and the vertical axis represents all bidding enterprises within a certain subgroup. Each cell is represented by a pie chart with an outer ring. The pie chart is divided according to the projects that bidding enterprises participated in bidding on the bid opening day: the colors of different sectors in the pie chart correspond to those who participated in bidding but did not win, and those who won the project. The outer ring color is used to distinguish the bidding units corresponding to different projects. Based on the changing patterns of bidding results over time, the winning roles of enterprises in different projects or at different times are analyzed to determine whether there is any suspicion of rotating winning bids. If some companies appear in the timeline to be continuously participating in bidding but never winning a bid, they are suspected of bid rigging.

Images