Big model technology-based anti-bidding risk control system platform

Abstract

The invention relates to the technical field of electric power risk control, and discloses an anti-bidding risk control system platform based on a large model technology, and the system platform comprises the steps: collecting bidding data, splitting the bidding data into bidding text data, behavior data and enterprise background data, carrying out the semantic extraction of the bidding text data, obtaining a core data set, and carrying out the calculation of the core data set; performing similarity score calculation on the core data set to obtain a similarity score; constructing a bidding relationship network by utilizing the behavior data and the enterprise background data; performing data synthesis by using the bidding relationship network, the similarity score and the core data set to obtain contrast data; and performing comparative analysis on the comparison data and the bidding data to obtain an abnormal behavior, and performing risk control scoring on the abnormal behavior to obtain an anti-bidding risk control result. According to the invention, the problems of low efficiency and poor adaptability of an existing anti-flipping risk control system platform based on a large model technology can be solved.

Description

Technical Field

[0001] The present invention relates to the field of electric power risk control technology, and in particular to an anti-fraud risk control system platform based on large model technology. Background Art

[0002] In the power industry, bidding, as an important part of corporate procurement and engineering contracting, has always been the core means of market competition. However, due to the huge amount of bidding projects and the complexity of the process, unfair behaviors such as bid rigging and bid collusion often occur. These behaviors not only undermine the fair competition market environment, but also may lead to high-cost, low-quality projects or equipment entering the power system, affecting the stability and sustainable development of the power industry.

[0003] Traditional means of preventing bid rigging and bid collusion mainly rely on rule models and manual verification. However, when faced with large-scale, multi-dimensional data, traditional rules are difficult to accurately identify bid rigging and bid collusion due to the constant changes in new bid rigging and bid collusion methods, which leads to low processing accuracy and poor adaptability of rule models. Manual verification requires proofreading a large amount of data one by one, resulting in low processing efficiency. In summary, the current risk control system platform for preventing bid rigging and bid collusion still has problems of low efficiency and poor adaptability. Summary of the invention

[0004] The present invention provides a risk control system platform for preventing bid fraud based on large model technology, and its main purpose is to solve the problems of low efficiency and poor adaptability of the existing risk control system platform for preventing bid fraud based on large model technology.

[0005] To achieve the above objectives, the present invention provides a risk control system platform for preventing fraudulent bidding based on large model technology, including: a semantic analysis module, a relationship construction module, a data generation module, and a result generation module. Specifically:

[0006] A semantic analysis module is used to collect bidding data, split the bidding data into bidding text data, behavior data and enterprise background data using a preset language model, perform semantic extraction on the bidding text data to obtain a core data set, and perform similarity score calculation on the core data set to obtain a similarity score;

[0007] A relationship building module, used to build a bidding relationship network using the behavior data and enterprise background data;

[0008] A data generation module, used to synthesize data using the bidding relationship network, the similarity score and the core data set to obtain control data;

[0009] A result generation module is used to compare and analyze the control data with the bidding data to obtain a risk control result for preventing bid rigging and collusion.

[0010] Optionally, when the semantic analysis module performs the function of performing semantic extraction on the bidding text data to obtain a core data set, it is specifically used to:

[0011] Performing stop word processing on the bidding text data to obtain text segmentation;

[0012] Count the frequency of each text segmentation word in the corresponding bidding text data to obtain the word frequency;

[0013] Calculate the inverse document frequency of the text segmentation words in the bidding text data;

[0014] Calculate the importance weight of the text segmentation according to the word frequency and the inverse document frequency;

[0015] Integrate the text segmentation and the importance weights into a feature matrix of the bidding text data;

[0016] Perform data dimension reduction on the feature matrix to obtain a core data set.

[0017] Optionally, when the semantic analysis module performs the function of performing data dimension reduction on the feature matrix to obtain a core data set, it is specifically used to:

[0018] Performing singular value decomposition on the characteristic matrix using a singular value matrix algorithm to obtain a singular value matrix, a transpose of a singular vector matrix, and a singular vector matrix;

[0019] Sorting the singular value matrix according to the size of the singular value to obtain a sorting result;

[0020] Selecting a preset number of singular values ​​from the sorting results in sequence as the selected singular values;

[0021] Filter out singular vectors corresponding to the filtered singular values ​​from the transpose of the singular vector matrix and the singular vector matrix as dimensionality reduction vectors;

[0022] The dimension-reduced vector and the feature matrix are reconstructed to obtain a core data set.

[0023] Optionally, when executing the function of constructing a bidding relationship network using the behavior data and the enterprise background data, the relationship construction module is specifically used to:

[0024] Performing data association on the behavior data and the enterprise background data to obtain associated data;

[0025] Constructing a data association matrix between the associated data;

[0026] Based on the data association matrix, association edges are added to each enterprise in the enterprise background data to obtain a bidding relationship network.

[0027] Optionally, when executing the function of constructing the data association matrix between the associated data, the relationship construction module is specifically used to:

[0028] Calculating the correlation between the enterprise relationship correlation indicators in the correlation data to obtain the relationship correlation;

[0029] Calculating the comprehensive correlation between the related data according to the relationship correlation;

[0030] A data association matrix of the associated data is constructed according to the comprehensive association degree.

[0031] Optionally, when the data generation module performs the function of synthesizing data using the bidding relationship network, the similarity score and the core data set to obtain the comparison data, it is specifically used to:

[0032] Performing text synthesis on the core data set according to the similarity score to obtain similar text content;

[0033] Performing vector conversion on the bidding relationship network to obtain a network vector;

[0034] Performing behavior simulation on the network vector using the behavior data to obtain simulation data;

[0035] Data fusion is performed based on the simulation data and the similar text content to obtain comparison data.

[0036] Optionally, when the data generation module performs the function of synthesizing texts of the core data set according to the similarity score to obtain similar text content, it is specifically used to:

[0037] Converting the core data set into a vector form to obtain a text vector;

[0038] Encoding the text vector and the similarity score to obtain context information;

[0039] Similar texts are predicted according to the context information and the text vector.

[0040] Optionally, when the data generation module performs the function of using the behavior data and the network vector to perform behavior simulation to obtain simulation data, it is specifically used to:

[0041] adding noise to the behavior data to obtain diversified data;

[0042] The network vectors are updated with the diversified data to obtain simulation data.

[0043] Optionally, when the result generation module performs the function of using the control data to compare and analyze the bidding data to obtain the anti-bidding risk control result, it is specifically used to:

[0044] Performing a similarity comparison between the control data and the bidding data to obtain a similarity result;

[0045] Performing time series analysis according to the control data and the bidding data to obtain a time series result;

[0046] Performing data integration on the similarity results and the time series results to obtain abnormal behaviors;

[0047] A comprehensive score is given based on the abnormal behaviors to obtain the risk control results for preventing bid rigging.

[0048] Optionally, when the result generation module performs the function of performing time series analysis according to the comparison data and the bidding data to obtain a time series result, it is specifically used to:

[0049] Calculating the time density of the bidding data and the comparison data;

[0050] Performing time series clustering on the bidding data and the control data to obtain clustering results;

[0051] The time density and the clustering result are used to perform collaborative relationship detection to obtain a time series result.

[0052] The present invention constructs a bidding relationship network through behavioral data and enterprise background data. This network-based analysis method can adapt to the complex and diverse bidding environment of the power industry, and the bidding relationship network can dynamically display the relationship between enterprises, can intuitively reflect problems, and effectively solve the problem that traditional systems cannot adapt to complex and changeable bidding environments; using the bidding relationship network, similarity scores, and core data to synthesize control data can more comprehensively and systematically reveal the characteristics of abnormal behavior and improve the system's ability to identify collusion in bidding. Therefore, the anti-collusion and bid-rigging risk control system platform based on large model technology proposed by the present invention can solve the problems of low efficiency and poor adaptability of the existing anti-collusion and bid-rigging risk control system platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A functional module diagram of a risk control system platform for preventing fraudulent bidding based on a large model technology provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of a flow chart of semantic extraction provided by an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of a bidding relationship network flow diagram provided by an embodiment of the present invention;

[0056] Figure 4 A flowchart of a risk control method for preventing bid rigging based on large model technology is provided for one embodiment of the present invention.

[0057] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0058] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0059] Reference Figure 1 As shown, it is a functional module diagram of a risk control system platform for anti-fraudulent bidding based on large model technology provided by one embodiment of the present invention. In this embodiment, the risk control system platform 100 for anti-fraudulent bidding based on large model technology can be installed in an electronic device. According to the functions implemented, the risk control system platform 100 for anti-fraudulent bidding based on large model technology can include a data semantic analysis module 101, a relationship construction module 102, a data generation module 103 and a result generation module 104. The module described in the present invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can complete fixed functions, which are stored in the memory of the electronic device.

[0060] In the embodiment of the present invention, the semantic analysis module 101 includes collecting bidding data, splitting the bidding data into bidding text data, behavior data and enterprise background data using a preset language model, performing semantic extraction on the bidding text data to obtain a core data set, and performing similarity score calculation on the core data set to obtain a similarity score;

[0061] In the embodiment of the present invention, the relationship construction module 102 includes constructing a bidding relationship network using the behavior data and the enterprise background data;

[0062] In the embodiment of the present invention, the data generation module 103 comprises using the bidding relationship network, the similarity score and the core data set to perform data synthesis to obtain control data;

[0063] In the embodiment of the present invention, the result generation module 104 includes using the control data to compare and analyze with the bidding data to obtain the risk control result of anti-bidding collusion.

[0064] In detail, the modules described in the anti-fraud and bid-rigging risk control system platform 100 based on big model technology in the embodiment of the present invention adopt the same technical means as the anti-fraud and bid-rigging risk control system platform based on big model technology described in the accompanying drawings when used, and can produce the same technical effects, which will not be repeated here.

[0065] In combination with specific embodiments, the following describes the various components and specific workflows of the anti-fraud risk control system platform based on large model technology:

[0066] The semantic analysis module 101 is used to collect bidding data, split the bidding data into bidding text data, behavior data and enterprise background data using a preset language model, perform semantic extraction on the bidding text data to obtain a core data set, and perform similarity score calculation on the core data set to obtain a similarity score.

[0067] In the embodiment of the present invention, by dividing the bidding data into text data, behavior data and enterprise background data, it is helpful for classification and structured analysis, and reduces data confusion and processing difficulty.

[0068] In an embodiment of the present invention, bidding data generally refers to various information generated in the bidding and tendering process related to power projects, including: information on products or services provided by the bidder, the bidder's participation history, bidding time, the bidding company's power engineering design capabilities, construction capabilities, etc.

[0069] In an embodiment of the present invention, the preset language model refers to a language model that has been trained with a large-scale corpus, such as BERT, RoBERTa, T5, GPT series, etc.

[0070] In the embodiment of the present invention, the bidding text data refers to the content related to the project description, technical solution, quotation, etc. in the bidding document. This part of data usually contains a large amount of natural language text, and may also include tables, charts, etc. By using the language model to identify the key entities in the bidding data, such as project name, technical terms, amount, etc., and then classifying the text through the language model, identifying different paragraph types, such as "technical requirements", "quotation", "project background", etc., combined with the context understanding ability of the language model, locate and extract the content of specific fields (such as "technical requirements", "quotation"), and store it in a structured data format to obtain the bidding text data.

[0071] In an embodiment of the present invention, behavior data refers to timestamp information and related behavior records involved in the bidding process, such as registration time, submission time, bid opening time, etc. This part of data usually exists in a structured manner in a bidding document or database. By using a language model to identify the time information in the text (such as date, time), and associate it with the corresponding behavior (such as registration, submission, bid opening), and then classify the behavior in the text through the language model, identify different bidding behaviors (such as registration, submission, bid opening, winning bid, etc.), and extract these behaviors and their corresponding timestamps to construct behavior data.

[0072] In the embodiment of the present invention, the enterprise background data refers to the fixed information about the enterprise in the bidding document, such as company size, registered capital, business scope, credit rating, etc. This part of data usually exists in the form of a table or a fixed format. By using a language model combined with a table parsing tool (such as PyPDF2, pdfminer.six), the enterprise background information in the table is extracted from the bidding document, and then the fixed fields in the text (such as company name, registered capital, business scope, etc.) are identified by the language model, and these fields are extracted and stored in a structured data format to obtain.

[0073] Ginseng Figure 2 In the embodiment of the present invention, when the semantic analysis module performs the function of performing semantic extraction on the bidding text data to obtain the core data set, it is specifically used to:

[0074] S21, performing stop word processing on the bidding text data to obtain text segmentation;

[0075] S22, counting the frequency of each text segmentation word in the corresponding bidding text data to obtain the word frequency;

[0076] S23, calculating the inverse document frequency of the text segmentation words appearing in the bidding text data;

[0077] S24, calculating the importance weight of the text segmentation according to the word frequency and the inverse document frequency;

[0078] S25, integrating the text segmentation and the importance weight into a feature matrix of the bidding text data;

[0079] S26. Perform data dimension reduction on the feature matrix to obtain a core data set.

[0080] In the embodiments of the present invention, common words that are not helpful for semantic analysis, such as "de", "shi", "zai", etc., are removed from the text to reduce noise and improve the efficiency of subsequent analysis. By using a Chinese word segmentation tool (such as Jieba segmentation), the bid text data is segmented into words or phrases, and then a preset stop word list, which contains common meaningless words, is used to traverse each word or phrase to remove the stop words, obtaining the text word segmentation.

[0081] In the embodiments of the present invention, the inverse document frequency can be calculated using the following formula: ;

[0082] Wherein, is the inverse document frequency corresponding to the text word segmentation , represents the total number of the bid text data, represents the number of texts in the bid text data that contain the text word segmentation , is the logarithm symbol.

[0083] In the embodiments of the present invention, the product of the word frequency and the inverse document frequency can be used as the value of the importance weight.

[0084] In the embodiments of the present invention, each document is represented as a vector, where each dimension corresponds to a word, and the value of the vector is the importance weight of the word, obtaining a feature matrix. Specifically, it is a sparse matrix, where each row represents a document, each column represents a word, and the elements in the matrix are the values of the importance weights of the word in the corresponding document.

[0085] In the embodiments of the present invention, when the semantic analysis module executes the function of performing data dimensionality reduction on the feature matrix to obtain the core data set, it specifically is used for:

[0086] Performing singular value decomposition on the feature matrix using the singular value matrix algorithm to obtain a singular value matrix, the transpose of the singular vector matrix, and the singular vector matrix;

[0087] Sorting the singular value matrix according to the magnitudes of the singular values to obtain a sorting result;

[0088] Sequentially screening out a preset number of singular values from the sorting result as the screened singular values;

[0089] Screening out the singular vectors corresponding to the screened singular values from the transpose of the singular vector matrix and the singular vector matrix as the dimensionality reduction vectors;

[0090] Performing matrix reconstruction on the dimensionality reduction vectors and the feature matrix to obtain the core data set.

[0091] In detail, the singular value decomposition can be performed using the following formula: ;

[0092] in, is the feature matrix, is the singular vector matrix, is the singular value matrix, Represents the transpose of the singular vector matrix.

[0093] In detail, the singular values ​​reflect the importance of each principal component, and larger singular values ​​correspond to more important principal components. By sorting in descending order, you can determine which singular values ​​are most important.

[0094] In detail, a preset number of largest singular values ​​are selected to retain the most important information while reducing the dimensionality.

[0095] In detail, according to the selected singular values, the corresponding singular vectors and the transpose of the singular vectors are extracted as the representation after dimensionality reduction.

[0096] In detail, a new low-dimensional matrix is ​​reconstructed using the selected singular values ​​and singular vectors as a core data set, and the new low-dimensional matrix is ​​reconstructed through matrix multiplication.

[0097] In the embodiment of the present invention, by calculating the similarity score, it is possible to quickly determine whether there are abnormal similarities between different bids (such as consistent quotations, repeated text templates, etc.), which can improve the efficiency of processing bid information, especially when processing large-scale data, and quickly realize automated analysis of bid data.

[0098] The relationship building module 102 is used to build a bidding relationship network using the behavior data and enterprise background data.

[0099] In the embodiment of the present invention, the constructed relationship network can graphically display the relationship between the enterprises, clearly revealing potential bid-rigging groups or enterprise alliances.

[0100] Ginseng Figure 3 In the embodiment of the present invention, when the relationship building module executes the function of building a bidding relationship network using the behavior data and the enterprise background data, it is specifically used to:

[0101] S31, performing data association on the behavior data and the enterprise background data to obtain associated data;

[0102] S32, constructing a data association matrix between the associated data;

[0103] S33. Add associated edges to each enterprise in the enterprise background data based on the data association matrix to obtain a bidding relationship network.

[0104] In the embodiment of the present invention, the behavior data and the enterprise background data are combined to identify the potential associations between different bidders, extract key features from the behavior data, such as the bidding time, number of revisions, interaction records, etc.; extract key information from the enterprise background data, such as company size, registered capital, shareholder information, past projects, etc., and match the behavior data and enterprise background data based on common identifiers (such as company name, unified social credit code). For incomplete matches, fuzzy matching techniques (such as Levenshtein distance, Jaccard similarity) can be used to improve matching accuracy.

[0105] In the embodiment of the present invention, when executing the function of constructing the data association matrix between the associated data, the relationship construction module is specifically used to:

[0106] Calculating the correlation between the enterprise relationship correlation indicators in the correlation data to obtain the relationship correlation;

[0107] Calculating the comprehensive correlation between the related data according to the relationship correlation;

[0108] A data association matrix of the associated data is constructed according to the comprehensive association degree.

[0109] In detail, a variety of association indicators are calculated for each pair of bidders (enterprises) to quantify the strength of the relationship between them, including the number of times the two bidders bid simultaneously on the same project, whether the two bidders have the same shareholders or affiliated companies, and the number of interactions between the two bidders (e.g., asking questions and answering questions), to obtain the degree of association.

[0110] In detail, multiple correlation indicators are weighted and integrated into a comprehensive correlation to ensure a comprehensive measurement of the relationship between bidders.

[0111] In an embodiment of the present invention, a bidding relationship network with a graph structure is constructed based on a data association matrix, in which nodes represent bidders and edges represent associations between bidders. Based on the values ​​in the data association matrix, an edge is added for each pair of bidders, and the weight of the edge is the corresponding value in the data association matrix. Only edges with association strength exceeding a certain threshold are selected to reduce noise.

[0112] In the embodiment of the present invention, by constructing a bidding relationship network, the enterprise group that may be involved in bid rigging can be quickly identified, and a behavioral evidence chain of bid rigging can be provided, which is helpful for the reliability and traceability of risk control results.

[0113] The data generation module 103 is used to perform data synthesis using the bidding relationship network, the similarity score and the core data set to obtain control data.

[0114] In the embodiment of the present invention, multiple data sources (core data set, relationship network, similarity score) are combined to generate control data, providing a more comprehensive analysis perspective.

[0115] In the embodiment of the present invention, when the data generation module performs the function of synthesizing data using the bidding relationship network, the similarity score and the core data set to obtain the comparison data, it is specifically used to:

[0116] Performing text synthesis on the core data set according to the similarity score to obtain similar text content;

[0117] Performing vector conversion on the bidding relationship network to obtain a network vector;

[0118] Performing behavior simulation using the behavior data and the network vector to obtain simulation data;

[0119] Data fusion is performed based on the simulation data and the similar text content to obtain comparison data.

[0120] In the embodiment of the present invention, when the data generation module performs the function of synthesizing texts of the core data set according to the similarity score to obtain similar text content, it is specifically used to:

[0121] Converting the core data set into a vector form to obtain a text vector;

[0122] Encoding the text vector and the similarity score to obtain context information;

[0123] Similar texts are predicted according to the context information and the text vector.

[0124] In detail, the core dataset can be converted into a text vector through a word embedding method, or through a pre-trained word vector model (such as Word2Vec, GloVe) or a context-related language model (such as BERT, GPT) to convert the words in the text into a vector of fixed dimension, which is the text vector of the core dataset.

[0125] In detail, the similarity score is a scalar that indicates the semantic similarity between the core dataset and the target text (for example, 0.8 indicates high similarity). By directly using the similarity score as a conditional input, it is concatenated with the text vector to form a new input vector, and then the input vector is encoded using a preset neural network, where the neural network can be a simple fully connected network or a Transformer encoder (such as BERT). The context information vector is a vector representation that combines text semantics and similarity scores.

[0126] In detail, similar texts are generated word by word using a generative network based on context information and text vectors. The generative network generates similar texts by initializing a decoder, taking context information as initial conditions, and gradually generating texts according to conditional probabilities. Each time, an optimal word is selected from the probability distribution (or by random sampling) and used as input for the next step until an end symbol is generated, thereby finally obtaining similar texts.

[0127] In an embodiment of the present invention, a graph neural network (such as GCNs, GATs, etc.) can be used to encode the bidding relationship network, and each node (i.e., the bidding enterprise) and its edge (i.e., the association relationship between enterprises) can be converted into a low-dimensional vector representation.

[0128] In the embodiment of the present invention, when the data generation module performs the function of using the behavior data and the network vector to perform behavior simulation to obtain simulation data, it is specifically used to:

[0129] adding noise to the behavior data to obtain diversified data;

[0130] The network vectors are updated with the diversified data to obtain simulation data.

[0131] In detail, a random noise vector with the same dimension as the original behavior data is generated by using Gaussian distribution, and the generated random noise vector is added element by element to the original behavior data to obtain new diversified data.

[0132] In detail, the diversified data are merged with the original data, the data association matrix is ​​reconstructed, and the bidding relationship network is updated using the new data association matrix to obtain the simulated data.

[0133] In the embodiment of the present invention, the simulation data and the similar text content are directly spliced ​​into a joint feature, and the joint feature is the comparison data.

[0134] In the embodiment of the present invention, through data synthesis, noise data can be eliminated and a higher quality control data set can be generated, which helps to improve the accuracy of the analysis results.

[0135] The result generation module 104 is used to compare and analyze the control data with the bidding data to obtain the risk control result of anti-bidding collusion.

[0136] In the embodiment of the present invention, the bidding data and the control data are systematically compared, so that the results can be generated quickly, and the time and error rate required for manual analysis can be reduced.

[0137] In the embodiment of the present invention, when the result generation module performs the function of using the control data and the bidding data to compare and analyze and obtain abnormal behavior, it is specifically used to:

[0138] Performing a similarity comparison between the control data and the bidding data to obtain a similarity result;

[0139] Performing time series analysis according to the control data and the bidding data to obtain a time series result;

[0140] Performing data integration on the similarity results and the time series results to obtain abnormal behaviors;

[0141] A comprehensive score is given based on the abnormal behaviors to obtain the risk control results for preventing bid rigging.

[0142] In the embodiment of the present invention, by comparing the similarity between the reference data and the actual bidding data, potential abnormal behavior is identified. For example, if the text or behavior pattern of some power bidders is highly similar to the reference data, there may be a risk of bid rigging or collusion. Natural language processing (NLP) techniques, such as TF-IDF, cosine similarity, Jaccard similarity, etc., can be used to calculate the similarity between the reference text and the actual bidding text data.

[0143] In the embodiment of the present invention, when the result generation module performs the function of performing time series analysis according to the comparison data and the bidding data to obtain the time series result, it is specifically used to:

[0144] Calculating the time density of the bidding data and the comparison data;

[0145] Performing time series clustering on the bidding data and the control data to obtain clustering results;

[0146] The time density and the clustering result are used to perform collaborative relationship detection to obtain a time series result.

[0147] In detail, by calculating the activity frequencies (such as the number of bids, the number of modifications, the number of interactions, etc.) of the bidders in different time periods, the time density of their behaviors is quantified. By dividing the bidding data into multiple time windows of fixed length, counting the number of activities of each bidder in each time window, such as the number of bids, the number of modifications, the number of interactions, etc., the activity density in each time window is calculated, that is, the ratio of the number of activities to the length of the time window.

[0148] In detail, features such as activity frequency, time interval, activity type, etc. are extracted from the bidding data and the control data, the extracted features are standardized to ensure comparability between different features, a time series is obtained, and a clustering algorithm (such as K-means, DBSCAN, Gaussian Mixture Model) is used to cluster the time series data to obtain a clustering result.

[0149] In detail, the coordination relationship between bidders is detected by analyzing the time density and clustering results. If multiple bidders show high-density activities in the same time window and are clustered into the same cluster, there may be suspicion of bid rigging or collusion. By traversing each time window, it is checked whether multiple bidders meet the preset coordination relationship criteria at the same time, and a risk score is calculated for each bidder, taking into account the time density, clustering results and other relevant factors. The coordination relationship criteria are as follows: within a certain time window, the time density of multiple bidders exceeds a certain threshold, and they belong to the same cluster.

[0150] In detail, the time series results refer to the behavioral characteristic information extracted after analyzing the time characteristics of the bidding data and the control data, which is used to reveal the time patterns, coordination and anomalies of the bidding behavior. Specifically, the time series results are the output results of the time series analysis, which are mainly used to assist in determining whether the enterprise has engaged in collusion in bidding.

[0151] In the embodiment of the present invention, the similarity results and the time series results are combined to comprehensively evaluate the behavior of the bidders and identify potential abnormal behaviors. For example, some bidders not only have similar texts, but also show abnormally consistent behaviors in the time series, which may be a sign of bid rigging or collusion. A comprehensive score is calculated for each bidder, taking into account text similarity, behavior similarity, and time series abnormality. By using the weighted summation method, different weights are assigned to different indicators. According to a preset threshold, when the comprehensive score exceeds the threshold, it is marked as abnormal behavior.

[0152] In the embodiment of the present invention, a preset risk scoring table is used to perform risk scoring, to obtain a risk score and its corresponding risk level, and to obtain a risk control result for preventing bid rigging.

[0153] In the embodiments of the present invention, since abnormal behaviors (such as bid collusion) are usually carefully designed, their patterns may not be easy to identify directly from real data. Therefore, by comparing the control data with the bid data, misjudgments and missed judgments can be significantly reduced.

[0154] like Figure 4 FIG. 1 is a flow chart of a risk control method for preventing bid-rigging based on a large model technology provided by an embodiment of the present invention. In an embodiment of the present invention, the risk control method for preventing bid-rigging based on a large model technology includes:

[0155] S401, collecting bidding data, splitting the bidding data into bidding text data, behavior data and enterprise background data using a preset language model, performing semantic extraction on the bidding text data to obtain a core data set, and performing similarity score calculation on the core data set to obtain a similarity score;

[0156] S402, constructing a bidding relationship network using the behavior data and enterprise background data;

[0157] S403, synthesizing data using the bidding relationship network, the similarity score and the core data set to obtain control data;

[0158] S404: Compare and analyze the control data with the bidding data to obtain risk control results for preventing bid rigging and collusion.

[0159] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology.

[0160] Among them, Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0161] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or systems stated in the system can also be implemented by one unit or system through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A risk control system platform for preventing fraudulent bidding based on large model technology, characterized in that: The system platform includes: a semantic analysis module, a relationship construction module, a data generation module, and a result generation module. Specifically: A semantic analysis module is used to collect bidding data, split the bidding data into bidding text data, behavior data and enterprise background data using a preset language model, perform semantic extraction on the bidding text data to obtain a core data set, and perform similarity score calculation on the core data set to obtain a similarity score; A relationship building module, used to build a bidding relationship network using the behavior data and enterprise background data; A data generation module, used to synthesize data using the bidding relationship network, the similarity score and the core data set to obtain control data; A result generation module is used to compare and analyze the control data with the bidding data to obtain a risk control result for preventing bid rigging and collusion.

2. The anti-fraud risk control system platform based on large model technology as claimed in claim 1 is characterized in that: When the semantic analysis module performs the function of performing semantic extraction on the bidding text data to obtain the core data set, it is specifically used to: Performing stop word processing on the bidding text data to obtain text segmentation; Count the frequency of each text segmentation word in the corresponding bidding text data to obtain the word frequency; Calculate the inverse document frequency of the text segmentation words in the bidding text data; Calculate the importance weight of the text segmentation according to the word frequency and the inverse document frequency; Integrate the text segmentation and the importance weights into a feature matrix of the bidding text data; Perform data dimension reduction on the feature matrix to obtain a core data set.

3. The anti-fraud risk control system platform based on large model technology as claimed in claim 2 is characterized in that: When the semantic analysis module performs the function of reducing the dimension of the feature matrix to obtain the core data set, it is specifically used to: Performing singular value decomposition on the characteristic matrix using a singular value matrix algorithm to obtain a singular value matrix, a transpose of a singular vector matrix, and a singular vector matrix; Sorting the singular value matrix according to the size of the singular value to obtain a sorting result; Selecting a preset number of singular values ​​from the sorting results in sequence as the selected singular values; Filter out singular vectors corresponding to the filtered singular values ​​from the transpose of the singular vector matrix and the singular vector matrix as dimensionality reduction vectors; The dimension reduction vector and the feature matrix are reconstructed to obtain a core data set.

4. The anti-fraud risk control system platform based on large model technology as claimed in claim 1 is characterized in that: When executing the function of constructing a bidding relationship network using the behavior data and the enterprise background data, the relationship construction module is specifically used to: Performing data association on the behavior data and the enterprise background data to obtain associated data; Constructing a data association matrix between the associated data; Based on the data association matrix, association edges are added to each enterprise in the enterprise background data to obtain a bidding relationship network.

5. The anti-fraud risk control system platform based on large model technology as claimed in claim 4 is characterized in that: When executing the function of constructing the data association matrix between the associated data, the relationship construction module is specifically used to: Calculating the correlation between the enterprise relationship correlation indicators in the correlation data to obtain the relationship correlation; Calculating the comprehensive correlation between the related data according to the relationship correlation; A data association matrix of the associated data is constructed according to the comprehensive association degree.

6. The anti-fraud risk control system platform based on large model technology as claimed in claim 1 is characterized in that: When the data generation module performs the function of synthesizing data using the bidding relationship network, the similarity score and the core data set to obtain the comparison data, it is specifically used to: Performing text synthesis on the core data set according to the similarity score to obtain similar text content; Performing vector conversion on the bidding relationship network to obtain a network vector; Performing behavior simulation on the network vector using the behavior data to obtain simulation data; Data fusion is performed based on the simulation data and the similar text content to obtain comparison data.

7. The anti-fraud risk control system platform based on large model technology as claimed in claim 6 is characterized in that: When the data generation module performs the function of synthesizing texts of the core data set according to the similarity score to obtain similar text content, it is specifically used to: Converting the core data set into a vector form to obtain a text vector; Encoding the text vector and the similarity score to obtain context information; Similar texts are predicted according to the context information and the text vector.

8. The anti-fraud risk control system platform based on large model technology as claimed in claim 6 is characterized in that: When the data generation module performs the function of using the behavior data to simulate the behavior of the network vector to obtain simulation data, it is specifically used to: adding noise to the behavior data to obtain diversified data; The network vectors are updated with the diversified data to obtain simulation data.

9. The anti-fraud risk control system platform based on large model technology as claimed in claim 1 is characterized in that: When the result generation module performs the function of using the control data to compare and analyze with the bidding data to obtain the risk control result of anti-bidding collusion, it is specifically used to: Performing a similarity comparison between the control data and the bidding data to obtain a similarity result; Performing time series analysis according to the control data and the bidding data to obtain a time series result; Performing data integration on the similarity results and the time series results to obtain abnormal behaviors; The abnormal behaviors are comprehensively scored to obtain risk control results for preventing bid rigging.

10. The anti-fraud risk control system platform based on large model technology as claimed in claim 9, characterized in that: When the result generation module performs the function of performing time series analysis according to the comparison data and the bidding data to obtain the time series result, it is specifically used to: Calculating the time density of the bidding data and the comparison data; Performing time series clustering on the bidding data and the control data to obtain clustering results; The time density and the clustering result are used to perform collaborative relationship detection to obtain a time series result.

Images