Bidding and tendering data intelligent analysis method and system based on AI technology and storage medium

Abstract

The invention relates to the field of bidding risk assessment, and discloses an AI technology-based intelligent bidding data analysis method and system and a storage medium, and the method comprises the following steps: obtaining bidding data, constructing a multi-dimensional space-time tensor, and extracting low-dimensional features through dynamic tensor decomposition; establishing a dynamic enterprise relationship graph, modeling an enterprise cooperation relationship by using a graph neural network, and capturing a bidding behavior mode between enterprises; a multi-target joint optimization model is constructed, and market trend prediction, enterprise credit assessment and risk early warning are synchronously realized; performing adversarial training by adopting a generative adversarial network, and dynamically adjusting task weights in combination with Pareto optimization to balance multi-task conflicts; and outputting a market prediction report, an enterprise credit portrait and a risk early warning list, and providing policy and regulation change influence analysis and a risk evidence chain. According to the method, bidding and tendering data are deeply mined through the AI technology, the market insight and the abnormal bidding behavior recognition capability are improved, and the method has wide application value.

Description

Technical Field

[0001] The present invention relates to the technical field of bidding risk assessment, and specifically to an AI-based bidding data intelligent analysis method, system and storage medium. Background Art

[0002] In the bidding market, companies obtain projects through open competition, and the results of the bid are affected by many factors, including changes in market supply and demand, corporate credit status, and the standardization of bidding behavior. Market regulators and bidding companies need to accurately assess market trends, identify potential credit risks, and effectively detect abnormal bidding behaviors such as bid rigging and collusion to improve market transparency and optimize bidding strategies. However, as the scale of the bidding market expands, the amount of data grows rapidly, and traditional analysis methods face many challenges in processing massive heterogeneous data.

[0003] Existing technologies mainly rely on rule matching and statistical analysis methods, which have low adaptability to market trend forecasting and credit assessment, and it is difficult to effectively integrate corporate relationship networks, historical bidding behaviors, and market supply and demand information. In terms of market forecasting, traditional methods are often based on a single economic indicator or time series model, and fail to fully consider the dynamic relationship between enterprises and the multi-dimensional impact of industry characteristics, resulting in insufficient prediction accuracy. For corporate credit assessment, existing methods are usually based on financial data or a single scoring system, lacking comprehensive consideration of the company's historical behavior patterns, related corporate relationships, and industry background, making it difficult to fully characterize the company's credit status. In terms of abnormal bid detection, most existing solutions rely on manually set abnormal behavior rules, such as bid price similarity, time synchronization, etc., but their accuracy and generalization capabilities are limited when dealing with complex bid-rigging and collusion behaviors with constantly changing bidding strategies.

[0004] At the same time, existing technologies lack optimization mechanisms in multi-task learning. Market forecasting, credit assessment, and risk warning are usually processed as independent tasks, and cross-task information is not effectively utilized, resulting in limited overall performance of the model. Some studies have attempted to combine machine learning for multi-task modeling, but lack a dynamic optimization mechanism for weight allocation between tasks, making it difficult to maintain stable task coordination in different application scenarios. In addition, due to the complexity and concealment of abnormal bidding behavior, existing models still have deficiencies in generalization capabilities and are easily affected by changes in data distribution, resulting in unstable risk warning effects.

[0005] Therefore, the existing technology still has many shortcomings in terms of the accuracy of market trend prediction, the comprehensiveness of corporate credit assessment, the robustness of abnormal bid detection, and the coordination of multi-task optimization. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides an intelligent analysis method, system and storage medium for bidding data based on AI technology, which solves the problems of low accuracy in market trend prediction, incomplete corporate credit assessment, insufficient robustness in detecting abnormal bidding behavior and poor coordination in multi-task optimization in the existing bidding data analysis methods.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: an intelligent analysis method for bidding data based on AI technology, comprising the following steps: Obtaining bidding data, including structured data, unstructured text and external data; constructing the data into a multi-dimensional spatiotemporal tensor and extracting low-dimensional features based on dynamic tensor decomposition; Build a dynamic enterprise relationship map and model enterprise partnerships through graph neural networks; Based on the multi-objective joint optimization model, market trend forecasting, corporate credit assessment and risk warning are carried out simultaneously; Enhance the generalization ability of the model through adversarial training, and balance multi-task conflicts in combination with Pareto optimization; Output market forecast reports, corporate credit profiles and risk warning results.

[0008] Preferably, the structured data includes enterprise bidding records, winning bid amounts and timestamps, unstructured text extracts semantic features through natural language processing, and external data includes enterprise business information and judicial risk labels.

[0009] Preferably, the step of constructing the data into a multidimensional spatiotemporal tensor and extracting low-dimensional features based on dynamic tensor decomposition comprises: The bidding data is constructed into a four-dimensional spatiotemporal tensor, wherein the four dimensions include enterprise dimension, project dimension, time dimension and multimodal feature dimension, wherein the multimodal feature dimension integrates structured data, text semantic vectors and external risk labels; Performing non-negative tensor decomposition on the four-dimensional space-time tensor to obtain a low-rank core tensor and factor matrices of four dimensions: enterprise, project, time, and feature; Applying adaptive sparse constraints to the factor matrix and dynamically adjusting the sparsity strength according to the complexity of feature distribution, wherein the complexity is quantified by feature entropy value; Output the enterprise embedding matrix and time evolution matrix to characterize the dynamic characteristics of the enterprise and the evolution law of the time dimension.

[0010] Preferably, the steps of constructing a dynamic enterprise relationship graph and modeling enterprise cooperation relationships through a graph neural network include: An initial static enterprise relationship map is constructed based on historical bidding data, wherein the nodes of the initial static enterprise relationship map are enterprise entities, and the edge weights are calculated and generated by the number of historical cooperation times and cooperation amounts between enterprises; The graph edge weights are dynamically updated through a temporal attention mechanism, which includes: Calculate the attention score based on the embedded features of the enterprise’s previous time slice; Combined with the actual number of collaborations between enterprises in the current time slice, the attention weight is adjusted by the domain knowledge coefficient, which is pre-trained based on the impact intensity of policies and regulations; Output a sequence of dynamic enterprise relationship graphs that evolve in time slices for task modeling of multi-objective graph neural networks.

[0011] Preferably, the steps of simultaneously performing market trend forecasting, enterprise credit assessment and risk warning based on a multi-objective joint optimization model include: Construct a shared graph neural network layer to extract the temporal features of enterprise nodes through a dynamic enterprise relationship graph sequence. The shared graph neural network layer adopts a gated graph convolution structure to fuse the features of the current node with the weighted features of its neighboring nodes. Based on the enterprise node features output by the shared graph neural network layer, the following tasks are performed through independent branch networks: Market trend prediction task: average pooling of enterprise node features across the industry is performed and input into a multi-layer perceptron to generate future market trend prediction results; Enterprise credit assessment task: concatenate the enterprise node features with the maximum pooled features of its neighboring nodes and input them into a multi-layer perceptron to generate a credit score and risk label; Risk warning task: Concatenate the enterprise node features with the standard deviation pooling features of its neighboring nodes and input them into a multi-layer perceptron to detect abnormal bidding behavior.

[0012] Preferably, the step of enhancing the generalization ability of the model through adversarial training includes: Construct a generative adversarial network, in which the generator generates simulated abnormal bidding samples based on the normal bidding patterns in historical bidding data. The abnormal bidding samples include the bidding collusion behavior patterns of multiple companies with abnormally close bidding prices and highly synchronized bidding times. The discriminator shares some network layers with the multi-objective graph neural network to distinguish between real bidding data and abnormal samples generated by the generator; During adversarial training, the generator and discriminator are jointly optimized through the following loss function: Generator loss function: maximize the probability of misjudgment of the generated sample by the discriminator; Discriminator loss function: minimize the classification error rate of real samples and generated samples; The adversarial training loss function and the multi-task loss function are weightedly fused to improve the model's ability to identify abnormal bidding behaviors.

[0013] Preferably, the step of balancing multi-task conflicts in combination with Pareto optimization includes: Define domain-driven multi-objective evaluation indicators, including the combined calculation of market prediction accuracy, enterprise credit assessment AUC value and risk warning false alarm rate, where the risk warning false alarm rate is constrained by a policy compliance threshold; Dynamically adjust the weight of each task loss function based on the Pareto frontier theory. The weight adjustment is achieved through the gradient descent algorithm to ensure the synchronous optimization of the performance of each task. During the optimization process, if the conflict between tasks exceeds the preset threshold, the task weight redistribution mechanism will be triggered to prioritize the false alarm rate of risk warning tasks to meet policy compliance requirements; Output optimized multi-task model parameters and analysis results.

[0014] Preferably, the output market forecast report, corporate credit profile and risk warning results include: Market trend forecast report: Based on the industry characteristic mean output by the multi-objective graph neural network, the market supply and demand forecast curve for the next 1-4 quarters is generated, and the impact range of policy and regulatory changes on the forecast results is marked; Enterprise credit profile: Integrate enterprise node characteristics, neighbor node characteristics and historical bidding records to generate credit scores and multi-dimensional risk labels, including the frequency of judicial litigation, volatility of bid winning rate, and intensity of risk transmission of related enterprises; Risk warning list: detect abnormal bidding behavior, output the list of companies suspected of bid rigging and collusion and the associated chain of evidence, which includes bid price difference threshold, bid time synchronization, and historical cooperation correlation indicators.

[0015] The present invention also provides an AI-based bidding data intelligent analysis system, including: Data acquisition module, used to integrate structured data, unstructured text and external data; Data processing module, which performs dynamic tensor decomposition and low-dimensional feature extraction; Dynamic graph modeling module, building and updating enterprise relationship graph; Multi-objective analysis module, which jointly optimizes market forecasting, credit assessment and risk warning tasks through graph neural networks; The optimization and output module performs adversarial training and Pareto optimization and generates analysis reports.

[0016] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above method is implemented.

[0017] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described above is implemented.

[0018] The present invention provides a bidding data intelligent analysis method, system and storage medium based on AI technology. It has the following beneficial effects: 1. This invention uses a multi-objective graph neural network (GNN) to model enterprise and industry characteristics, and combined with historical market data, can effectively predict the market supply and demand trends in the next 1-4 quarters. Based on the mean of industry characteristics and the influencing factors of policy changes, it provides an accurate market forecast report to reduce decision-making errors caused by market fluctuations.

[0019] 2. The present invention combines the characteristics of the enterprise itself, the characteristics of neighboring enterprises and historical bidding behavior to build a multi-dimensional credit scoring system. Through multi-dimensional risk labels such as the frequency of judicial litigation, the volatility of the winning bid rate and the intensity of risk transmission of related enterprises, a comprehensive assessment of the credit status of the enterprise is achieved, and the scientificity and reliability of the credit rating are improved.

[0020] 3. This invention uses a generative adversarial network (GAN) to simulate abnormal bidding patterns, and through key features such as bid price differences, time synchronization, and historical cooperation correlation, it can achieve efficient detection of bid-rigging and bid-collusion behaviors. It can effectively reduce the misjudgment rate of traditional rule matching methods and improve the accuracy of abnormal bidding identification.

[0021] 4. The present invention dynamically adjusts the weights of market forecasting, credit assessment and risk warning tasks through Pareto optimization, so that each task achieves an optimal balance in performance. It solves the weight conflict problem in multi-task learning, ensures that the model can be adaptively adjusted in different application scenarios, and improves the stability and applicability of the overall system.

[0022] 5. The present invention makes the risk warning results more transparent and reliable by outputting a chain of evidence including bid price anomalies, bid time synchronization and corporate historical relevance, helping regulatory agencies and corporate users to quickly understand the source of risks and improve the pertinence and effectiveness of risk control measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the method flow of the present invention; Figure 2 It is a schematic diagram of the structure of the device of the present invention; Figure 3 It is a schematic diagram of the computer device structure of the present invention.

[0024] Among them, 100, data acquisition module; 200, data processing module; 300, dynamic graph modeling module; 400, multi-objective analysis module; 500, optimization and output module; 40, computer equipment; 41, processor; 42, memory; 43, storage medium. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Please refer to the attached Figure 1 The present invention provides an intelligent analysis method for bidding data based on AI technology, which can achieve efficient and accurate market trend prediction, enterprise credit assessment and bidding risk warning by deeply mining and analyzing bidding data. This method aims to improve the transparency and fairness of the bidding market, reduce malicious bidding behaviors such as bid rigging and collusion, and provide scientific decision-making basis for enterprises, governments and regulatory authorities.

[0027] like Figure 1 As shown, the bidding data intelligent analysis method based on AI technology may include the following steps: S1. Obtain bidding data; S2, construct a multidimensional space-time tensor and use dynamic tensor decomposition to extract low-dimensional features; S3. Build a dynamic enterprise relationship map and use graph neural network to model enterprise cooperation relationships; S4. Based on the multi-objective joint optimization model, market trend forecasting, corporate credit assessment and risk warning are carried out simultaneously; S5. Use adversarial training to enhance the generalization ability of the model, and combine Pareto optimization to balance multi-task conflicts; S6. Finally, the market forecast report, corporate credit profile and risk warning results are output.

[0028] The following is a detailed description of each step in the method of the present invention, which comprehensively describes the specific implementation principle, technical details and process of each step.

[0029] For step S1, this step is to ensure that the data quality used in subsequent steps can meet the requirements of model training by reasonably integrating, cleaning, standardizing and extracting features of the bidding data.

[0030] First, the obtained bidding data includes structured data, unstructured text data, and external data. These data come from different sources in practical applications, such as enterprise bidding records, public bidding announcements, contract texts, legal litigation records, etc.

[0031] Structured data mainly includes the following categories: Enterprise basic information: such as enterprise name, registered capital, legal representative, registered address, etc.; Project bidding information: including bidding project name, budget amount, number of bidding enterprises, historical winning enterprises, etc.; Bidding history records: including bidding amount, bidding time, historical bidding projects, status of bidding enterprises, etc.

[0032] These structured data are usually stored in tabular form with a clear data format. After obtaining these data, data cleaning is first performed to remove duplicate values, missing values, and outliers. To ensure data consistency, all numerical data is standardized. For example, the bidding amount is normalized to ensure its equal contribution in the model.

[0033] Unstructured text data usually comes from bidding announcements, enterprise news reports, contract texts, enterprise litigation records, etc. To convert these text data into computable features, natural language processing (NLP) techniques are adopted.

[0034] The text preprocessing steps include: Word segmentation: Perform word segmentation on the original text to split sentences into words or phrases; Stop word removal: Remove common words without practical meaning, such as "de" (的), "le" (了), etc.; Stemming: Restore words to their base form. For example, "bidding" and "bidding of" are combined into "bidding".

[0035] Then, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm or BERT model is used to convert the text into word vectors. The TF-IDF value measures the importance of a word in a document, while the BERT model can capture semantic relationships in the text. Through these techniques, the text data is converted into vector form to provide input for subsequent models.

[0036] External data includes enterprise industrial and commercial registration information, judicial litigation records, industry statistical data, etc. These data can provide rich background knowledge for the model and improve the accuracy of prediction. External data usually comes from public government databases, industry association websites, etc.

[0037] The processing methods of external data include: Data alignment: Match external data with structured data and establish a corresponding relationship through information such as company name or registration number; Data normalization: Normalize the numerical fields in external data, such as the number of negative litigation records and market share of a company, to ensure that their numerical ranges are compatible with the model.

[0038] Outliers in the data may have a negative impact on the training of the model. Therefore, in the process of data preprocessing, outlier detection is also required. The isolation forest algorithm or box plot method is used to identify and remove abnormal data points that do not conform to the norm.

[0039] The Isolation Forest algorithm recursively splits data points and builds multiple decision trees to identify isolated points. The Box Plot Rule calculates the upper and lower quartile intervals based on the quartiles of the data and identifies abnormal data that exceeds the interval.

[0040] After completing the above steps, structured data, unstructured text data, and external data will be integrated into a unified data framework. This framework includes the following data columns: Enterprise information: such as enterprise ID, enterprise name, legal representative, registered capital, etc.; Bidding records: such as bidding project ID, bidding amount, bidding date, etc.; Text features: such as the TF-IDF vector or BERT embedding vector of the announcement text; External data: such as industrial and commercial data, judicial risk labels, etc.

[0041] During the data integration process, ensure the alignment and merging of different data sources, such as connecting structured data with external data through enterprise ID, and using text vectors and historical bidding records to jointly describe the characteristics of the enterprise.

[0042] After data preprocessing is completed, the final multimodal dataset is generated. This dataset will be used for subsequent feature extraction and model training. The dataset contains information including but not limited to enterprise characteristics, bidding history, legal litigation information, etc., and has been standardized, cleaned and denoised, which can provide high-quality input for subsequent machine learning models.

[0043] For step S2, the bidding data is constructed into a multi-dimensional spatiotemporal tensor, and low-dimensional features are extracted based on dynamic tensor decomposition. The core of this step is to effectively model multimodal and time-series data through tensor representation methods, and extract low-rank representations using tensor decomposition technology, thereby providing accurate features for subsequent analysis tasks.

[0044] First, a four-dimensional space-time tensor is constructed, whose dimensions include enterprise dimension, project dimension, time dimension and multimodal feature dimension. This four-dimensional space-time tensor represents the bidding data and related features under different time, different enterprises and projects. The specific process is as follows: 1. Construction of multidimensional space-time tensor The four-dimensional spacetime tensor is denoted by , whose dimensions are set as: ; in: Represents the enterprise dimension, including all enterprises participating in the bidding; Represents the project dimension, covering all bidding projects; It represents the time dimension, indicating bidding activities at different time points or time periods; Represents a multimodal feature dimension, which includes features integrated from structured data, unstructured text data (such as TF-IDF vectors or BERT embedding vectors), and external data (such as enterprise risk labels).

[0045] Through data preprocessing, we first unify the structured data (such as bid amount, bid number, etc.), text data (such as the semantic vector of the bid announcement text), and external data (such as judicial risk labels, industry statistics) and integrate them into a unified feature vector. These vectors will serve as the feature dimensions of the tensor. , thus obtaining a complete four-dimensional tensor , used for subsequent dynamic tensor decomposition.

[0046] 2. Dynamic Tensor Decomposition Next, for the four-dimensional space-time tensor The purpose of performing non-negative tensor decomposition (NMF) is to obtain a low-rank core tensor and its factor matrices of each dimension by decomposing the tensor. The specific process is as follows: First, the non-negative tensor factorization (NMF) model is used to Decompose. Given a tensor The decomposition formula is: ; in: Represents the core tensor, which contains low-rank information; , , , The factor matrices representing the enterprise, project, time and feature dimensions respectively, is the rank of tensor decomposition, E represents entity, P represents attribute, T represents time, and F represents feature. The rank R of tensor decomposition is usually much smaller than E, P, T, F, that is, much smaller than the total size of data dimensions, thereby reducing computational complexity and improving the generalization ability of the model while maintaining information integrity.

[0047] In this model, the core tensor It provides interactive information between multidimensional data, while the factor matrix captures the potential features of each dimension. Through tensor decomposition, the dimension of the data can be effectively reduced and important multidimensional interactive information can be retained.

[0048] 3. Imposition of Adaptive Sparse Constraints In order to further improve the effectiveness of tensor decomposition, adaptive sparse constraints are used to dynamically adjust the sparse strength according to the complexity of the feature. The complexity of the feature is quantified by the feature entropy value, and the formula is as follows: ; in, Indicates According to the entropy value of the feature, more complex features (i.e., more information) will be assigned lower sparsity strength, while simpler features (i.e., less information) will be imposed with higher sparsity constraints.

[0049] In this way, the model can dynamically adjust the importance of features and reduce the impact of redundant features, ultimately outputting a more concise factor matrix with stronger generalization ability.

[0050] 4. Output of Enterprise Embedding Matrix and Time Evolution Matrix After completing tensor decomposition and sparse constraints, the enterprise embedding matrix and time evolution matrix are output to characterize the dynamic characteristics of the enterprise and the evolution law of the time dimension. Specifically, the enterprise embedding matrix Contains the low-dimensional embedding features of each enterprise, which can reflect the behavior patterns and credit characteristics of enterprises in the bidding market; the time evolution matrix It describes the changing trends of enterprises in different time periods and can capture the time evolution laws of the market.

[0051] For example, for enterprises , which in time The embedding features can be expressed as: ; in, Indicates the enterprise In time The embedding vector of can reflect the dynamic behavior of the enterprise.

[0052] Through the above process, the enterprise embedding matrix and time evolution matrix obtained will provide high-quality input for subsequent market trend forecasting, enterprise credit assessment and risk early warning.

[0053] For step S3, this step is to build a dynamic enterprise relationship graph and model enterprise cooperation relationships based on graph neural network (GNN). The core of this step is to use the historical cooperation data between enterprises to build an enterprise relationship graph that evolves over time, and adjust the edge weights through the temporal attention mechanism to capture the dynamic changes in enterprise relationships.

[0054] First, an initial static enterprise relationship graph is constructed based on historical bidding data. In this graph, enterprises are nodes, and the cooperation relationships between enterprises are connected as edges. The weight of each edge is calculated based on the number of historical cooperation times and the amount of cooperation. The specific construction method is as follows: Assume that the enterprise relationship map is ,in: Represents the set of all enterprise nodes; For time The collection of corporate cooperation edges at the location; is the edge weight matrix, which reflects the cooperation intensity among enterprises.

[0055] The edge weight is calculated as follows: ; in: Indicates time Enterprise and Enterprise The intensity of cooperation; Indicates the enterprise and In time The number of historical collaborations; Indicates the enterprise and In time The total amount of cooperation; and They are normalization functions respectively; and It is a coefficient used to balance the impact of the number of collaborations and the amount of collaboration.

[0056] After the static graph is constructed, the edge weights of the enterprise relationship graph are dynamically updated through the temporal attention mechanism. The core of this mechanism is to adjust the edge weights according to the historical behavior of the enterprise and dynamically optimize it in combination with the cooperation data of the current time slice.

[0057] The implementation of the temporal attention mechanism includes the following steps: 1. Calculate the embedding features of the enterprise in the previous time slice: Assume that the enterprise In time The embedding representation of , then the attention score The calculation is as follows: ; in: It is an activation function, used to introduce nonlinear characteristics; is a trainable linear transformation matrix; Represents the time step Time Node The eigenvector of Represents the time step Time Node The eigenvector of Represents a vector concatenation operation; is the attention parameter vector.

[0058] 2. Adjust the attention weight based on the number of collaborations in the current time slice: ; in: Neighborhood gathering for businesses; is the weight coefficient of the number of cooperation times in the time slice.

[0059] 3. Adjust the attention score by domain knowledge coefficient: ; in: is the domain knowledge coefficient, which indicates the impact of policies and regulations on corporate partnerships; This coefficient is pre-trained based on the intensity of policy impact and can dynamically adjust corporate relationships when policies in a specific industry change.

[0060] After the above calculations, the final output is a dynamic enterprise relationship graph sequence that evolves by time slices. This graph sequence can be used for task modeling of graph neural networks (GNNs).

[0061] In the process of graph neural network training, a multi-objective learning strategy is adopted to jointly optimize tasks such as enterprise partnership prediction, enterprise credit scoring, and market competition relationship mining. The specific GNN model optimization goals are: ; in: forecasting losses for business partnerships; loss of credit scores for businesses; Modeling losses for competitive market relationships; , , is the task weight coefficient.

[0062] Through the above methods, dynamic enterprise cooperation relationship prediction based on graph neural network can be realized, and enterprise market behavior patterns can be effectively explored.

[0063] As for step S4, this step is based on a multi-objective joint optimization model to simultaneously perform market trend forecasting, enterprise credit assessment and risk warning tasks. The core of this step is to build a shared graph neural network layer, extract the time series features of enterprise nodes using the enterprise relationship graph sequence, and complete different prediction tasks through independent task branches.

[0064] First, a shared graph neural network layer is constructed based on the dynamic enterprise relationship graph sequence in step S3. This layer adopts a gated graph convolution structure to fuse the current features of the enterprise node with the weighted features of its neighbor nodes. It is expressed as , then the calculation method of the graph neural network layer is as follows: Establish an enterprise In time The characteristics of , then the weighted sum of its neighbor node features is expressed as: ; in: For time Enterprise and Enterprise The edge weights between ; For Node At time step The eigenvector of For Enterprises The set of neighbor nodes.

[0065] Enterprise node characteristics that use a gating mechanism to calculate updates: ; ; ; ; in: and They are update gate and reset gate respectively; is the sigmoid activation function; represents element-wise multiplication; , , , , , is the trainable parameter matrix.

[0066] Through the above calculation, we get the time Timing characteristics of enterprise nodes , used for different prediction tasks.

[0067] On the basis of the shared graph neural network layer, independent task branch networks are adopted to perform market trend forecasting, corporate credit assessment and risk warning tasks respectively.

[0068] The market trend prediction task uses mean pooling to calculate the mean of all enterprise node features: ; The mean feature As input, a multi-layer perceptron (MLP) is used to generate the future market trend forecast results: ; The enterprise credit assessment task uses maximum pooling to concatenate the enterprise node features with the maximum pooling features of its neighboring nodes: ; ; Input the concatenated features into the multi-layer perceptron to generate the enterprise credit score and risk label: ; The risk warning task uses standard deviation pooling to calculate the standard deviation of enterprise node characteristics and neighbor node characteristics: ; ; Input the concatenated features into the multi-layer perceptron to detect abnormal bidding behavior: ; Finally, a multi-objective joint optimization strategy is adopted to optimize the above tasks simultaneously. The objective function is set as follows: ; in: predicting losses for market trends; Assess losses for corporate credit; To warn of risks and losses; , , is the weight hyperparameter of each task.

[0069] Through the above calculation process, market trend forecasting, corporate credit assessment and risk warning tasks can be optimized simultaneously to improve the overall modeling effect.

[0070] For step S5, this step enhances the generalization ability of the model through adversarial training and balances multi-task conflicts by combining Pareto optimization. The key to this step is to generate abnormal bidding behavior samples and model training through generative adversarial networks (GANs), combine multi-task loss function and adversarial loss function optimization, and use Pareto optimization to dynamically adjust the weights of the multi-task model to ensure the balance of task performance.

[0071] First, a generative adversarial network (GAN) is constructed, which consists of a generator and a discriminator. The generator generates simulated abnormal bidding samples based on the normal bidding patterns in historical bidding data. The abnormal samples include bid-rigging behavior patterns such as abnormally close bidding prices of multiple companies and highly synchronized bidding times.

[0072] 1. Generator design: The input of the generator is random noise and normal bidding data, and it generates abnormal bidding samples through its internal network. The generator learns to simulate bidding samples that conform to bid rigging behavior by comparing the normal bidding patterns in historical data. The specific output of the generator is abnormal bidding samples, which contain abnormal information of bidding price, time, and other behavior patterns.

[0073] Assume that the abnormal bidding sample generated by the generator is , the goal of the generator is to make the discriminator mistakenly judge these abnormal samples as real bidding data.

[0074] 2. Discriminator design: The discriminator shares some network layers with the multi-objective graph neural network (GNN) and is responsible for distinguishing between real bidding data and abnormal samples generated by the generator. The output of the discriminator is ,in Indicates the input bidding data. is the probability of the discriminator judging the authenticity of the data. A value close to 1 indicates a real sample, and a value close to 0 indicates a generated abnormal sample.

[0075] 3. Generator loss function: The goal of the generator is to maximize the probability of the discriminator misjudging the generated samples. The loss function of the generator is: ; in, represents the distribution of abnormal bidding samples generated by the generator, is the judgment of the discriminator on the generated samples.

[0076] 4. Discriminator loss function: The goal of the discriminator is to minimize the classification error rate of real samples and generated samples. The loss function of the discriminator is: ; in, represents the real bidding sample distribution, and They are the discriminator's judgment on real samples and generated samples respectively.

[0077] 5. Adversarial training loss function: During adversarial training, the generator and discriminator are jointly optimized through the above loss function. The final loss function is: ; in, is the overall loss function of the Generative Adversarial Network (GAN), is the generator loss, is the discriminator loss.

[0078] This loss function will be weighted and fused with the multi-task loss function to improve the model's ability to identify abnormal bidding behavior.

[0079] Subsequently, Pareto optimization is used to balance the conflicts between tasks. By defining domain-driven multi-objective evaluation indicators, combined with market trend prediction, enterprise credit assessment AUC value and risk warning false alarm rate, the performance of the multi-task model is optimized synchronously.

[0080] 1. Define multi-objective evaluation indicators: including market forecast accuracy, corporate credit assessment AUC value and risk warning false alarm rate, where the risk warning false alarm rate is constrained by the policy compliance threshold. The specific calculation method is as follows: False alarm rate = number of false alarm samples / number of all warning samples; Among them, false positive samples refer to samples that are judged as risky when they fail to meet policy compliance requirements.

[0081] 2. Dynamically adjust the weight of the task loss function: Based on the Pareto frontier theory, dynamically adjust the weight of each task loss function to balance the conflicts between tasks. The weight adjustment is achieved through the gradient descent algorithm to ensure the synchronous optimization of the performance of each task. Assume that the loss function of each task is , , , then the total loss function is: ; in, , , It is the weight coefficient of each task, which is adjusted dynamically.

[0082] 3. Task weight redistribution mechanism: During the optimization process, if the conflict between tasks exceeds the preset threshold, the task weight redistribution mechanism will be triggered to prioritize ensuring that the false alarm rate of risk warning tasks meets policy compliance requirements.

[0083] 4. Output optimized multi-task model parameters and analysis results: After Pareto optimization adjustment, output optimized model parameters. This parameter includes the results of market trend forecast, corporate credit assessment and risk warning after weight adjustment, as well as the final multi-task optimization analysis results.

[0084] By combining adversarial training with Pareto optimization, this step can significantly improve the model's ability to identify abnormal bidding behavior. The abnormal samples generated by adversarial training can enhance the generalization ability of the model, and Pareto optimization can effectively balance the conflicts between multiple tasks, ensuring that the performance of each task is optimized simultaneously, and ultimately achieving efficient and accurate market trend forecasting, corporate credit assessment, and risk warning.

[0085] For step S6, this step generates a market forecast report, corporate credit profile and risk warning results based on the output generated by the multi-objective graph neural network (GNN). This step combines the output of multiple tasks and processes them through different algorithms to transform complex multi-dimensional information into easy-to-understand reports and lists, helping users make more accurate market and credit assessment decisions.

[0086] First, the industry characteristic mean output by the multi-objective graph neural network model in the aforementioned step S4 is used to generate the market supply and demand forecast curve for the next 1 to 4 quarters.

[0087] 1. Calculate the mean of industry characteristics: The enterprise node characteristics output by the multi-objective graph neural network model are , then the industry characteristic mean The calculation formula is as follows: ; in, is the set of all enterprise nodes, is the total number of nodes.

[0088] 2. Generate market supply and demand forecast curve: Based on the industry characteristic mean, the market supply and demand change trend in the next 1-4 quarters is predicted through the regression model. The prediction function is , then the market supply and demand forecast curve is: ; in, Indicates the market supply and demand forecast results for the next 1-4 quarters.

[0089] 3. Mark the impact range of policy and regulatory changes: Through the analysis of historical policy changes and market data, the impact range of policy and regulatory changes on the forecast results is obtained. The calculation of the impact range is based on the volatility of historical data, assuming that the impact coefficient of policy changes is , then the policy impact range is: ; Finally, the generated market trend forecast report includes the changing trends of market supply and demand in the next 1-4 quarters, and marks the impact range of policy and regulatory changes.

[0090] Next, by integrating the enterprise node characteristics, neighbor node characteristics and historical bidding records, the enterprise's credit score and multi-dimensional risk label are generated.

[0091] 1. Credit score calculation: The enterprise credit score is calculated based on the enterprise node characteristics and its neighbor node characteristics. The enterprise node characteristics are , the neighbor node characteristics are , then the credit score It can be obtained by the weighted sum formula: ; in, and is the weight coefficient, which reflects the importance of the enterprise’s own characteristics and the characteristics of neighboring nodes.

[0092] 2. Multi-dimensional risk label generation: Generate multi-dimensional risk labels based on the company's historical bidding records and partnerships. These labels include: Judicial litigation frequency: calculate the frequency of litigation in the company's historical bidding; Bid success rate volatility: Calculate the standard deviation of the company's historical bid success rate to reflect the volatility of its bid success rate; Risk transmission intensity of affiliated enterprises: Generate a risk transmission intensity index by calculating the cooperation intensity between an enterprise and its affiliated enterprises.

[0093] Ultimately, the generated corporate credit profile includes the company's credit score and the above-mentioned multi-dimensional risk labels, providing a basis for further credit assessment and decision-making.

[0094] Finally, by detecting abnormal bidding behavior, a list of companies suspected of bid rigging and collusion and the associated chain of evidence are output.

[0095] 1. Abnormal bidding behavior detection: Detect abnormal bidding behavior based on characteristics such as bid price differences and bidding time synchronization. and Enterprise The bid price is and The bidding time is and , then the calculation formula for bid price difference and bid time synchronization is: ; ; Set the bid price difference threshold to , the bidding time synchronization threshold is ,like and , it is judged as abnormal bidding behavior.

[0096] 2. Generate the chain of evidence: Generate the chain of evidence of abnormal bidding behavior by calculating the strength of historical cooperation relationships. and Enterprise The historical cooperation correlation is , then the associated evidence chain is: ; in, Calculated from historical bidding records, it reflects the strength of the cooperative relationship between the two companies.

[0097] Ultimately, the generated risk warning list includes a list of companies suspected of bid rigging or collusion, along with relevant chains of evidence to help identify potential risky behaviors.

[0098] By combining market forecasts, corporate credit assessments and risk warnings, this step can generate detailed market trend forecast reports, corporate credit profiles and risk warning lists, providing important basis for corporate decision-making.

[0099] In general, this invention realizes market supply and demand forecasting, enterprise credit profile generation and abnormal bidding behavior detection by constructing a multi-task learning framework, combining enterprise node characteristics, historical bidding data and market dynamic information. This method uses Pareto optimization to dynamically balance multi-task conflicts, and combines generative adversarial networks (GAN) to enhance the model's ability to identify bid rigging and collusion. Finally, it outputs market forecast reports, enterprise credit scores and risk warning lists, providing accurate decision-making support for market supervision, enterprise risk control and policy formulation.

[0100] The intelligent analysis system for bidding data based on AI technology described below and the intelligent analysis method for bidding data based on AI technology described above can be referenced to each other.

[0101] Please refer to the attached Figure 2 The present invention also provides an AI-based bidding data intelligent analysis system, including: Data acquisition module 100, used to integrate structured data, unstructured text and external data; Data processing module 200, performing dynamic tensor decomposition and low-dimensional feature extraction; Dynamic graph modeling module 300, constructing and updating enterprise relationship graph; Multi-objective analysis module 400, which jointly optimizes market forecasting, credit assessment and risk warning tasks through graph neural networks; The optimization and output module 500 performs adversarial training and Pareto optimization, and generates an analysis report.

[0102] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, which will not be repeated here.

[0103] Please refer to the attached Figure 3 The present invention further provides a computer device 40, comprising: a processor 41 and a memory 42, wherein the memory 42 stores a computer program executable by the processor, and when the computer program is executed by the processor, the above method is executed.

[0104] The present invention further provides a storage medium 43 on which a computer program is stored. When the computer program is run by the processor 41, the above method is executed.

[0105] Among them, the storage medium 43 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable red-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, disk or optical disk.

[0106] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent analysis method of bidding data based on AI technology is characterized by: The following steps are involved: Acquiring bidding data, wherein the bidding data includes structured data, unstructured text and external data; Constructing the bidding data into a multi-dimensional spatiotemporal tensor, and extracting low-dimensional features based on dynamic tensor decomposition; Build a dynamic enterprise relationship map and model enterprise partnerships through graph neural networks; Based on the multi-objective joint optimization model, market trend forecasting, corporate credit assessment and risk warning are carried out simultaneously; Enhance the generalization ability of the model through adversarial training, and balance multi-task conflicts in combination with Pareto optimization; Output market forecast reports, corporate credit profiles and risk warning results.

2. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The structured data includes corporate bidding records, winning bid amounts and timestamps; the unstructured text extracts semantic features through natural language processing; and the external data includes corporate business information and judicial risk labels.

3. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The step of constructing the bidding data into a multi-dimensional spatiotemporal tensor and extracting low-dimensional features based on dynamic tensor decomposition includes: The bidding data is constructed into a four-dimensional spatiotemporal tensor, wherein the four dimensions include enterprise dimension, project dimension, time dimension and multimodal feature dimension, wherein the multimodal feature dimension integrates structured data, text semantic vectors and external risk labels; Performing non-negative tensor decomposition on the four-dimensional space-time tensor to obtain a low-rank core tensor and factor matrices of four dimensions: enterprise, project, time, and feature; Applying adaptive sparse constraints to the factor matrix and dynamically adjusting the sparsity strength according to the complexity of feature distribution, wherein the complexity is quantified by feature entropy value; Output the enterprise embedding matrix and time evolution matrix to characterize the dynamic characteristics of the enterprise and the evolution law of the time dimension.

4. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The steps of constructing a dynamic enterprise relationship graph and modeling enterprise cooperation relationships through a graph neural network include: An initial static enterprise relationship map is constructed based on historical bidding data, wherein the nodes of the initial static enterprise relationship map are enterprise entities, and the edge weights are calculated and generated by the number of historical cooperation times and cooperation amounts between enterprises; The graph edge weights are dynamically updated through a temporal attention mechanism, which includes: Calculate the attention score based on the embedded features of the enterprise’s previous time slice; Combined with the actual number of collaborations between enterprises in the current time slice, the attention weight is adjusted by the domain knowledge coefficient, which is pre-trained based on the impact intensity of policies and regulations; Output a sequence of dynamic enterprise relationship graphs that evolve in time slices for task modeling of multi-objective graph neural networks.

5. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The steps of simultaneously performing market trend forecasting, enterprise credit assessment and risk early warning based on the multi-objective joint optimization model include: Construct a shared graph neural network layer to extract the temporal features of enterprise nodes through a dynamic enterprise relationship graph sequence. The shared graph neural network layer adopts a gated graph convolution structure to fuse the features of the current node with the weighted features of its neighboring nodes. Based on the enterprise node features output by the shared graph neural network layer, the following tasks are performed through independent branch networks: Market trend prediction task: average pooling of enterprise node features across the industry is performed and input into a multi-layer perceptron to generate future market trend prediction results; Enterprise credit assessment task: concatenate the enterprise node features with the maximum pooled features of its neighboring nodes and input them into a multi-layer perceptron to generate a credit score and risk label; Risk warning task: Concatenate the enterprise node features with the standard deviation pooling features of its neighboring nodes and input them into a multi-layer perceptron to detect abnormal bidding behavior.

6. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The step of enhancing the generalization ability of the model through adversarial training includes: Construct a generative adversarial network, in which the generator generates simulated abnormal bidding samples based on the normal bidding patterns in historical bidding data. The abnormal bidding samples include the bidding collusion behavior patterns of multiple companies with abnormally close bidding prices and highly synchronized bidding times. The discriminator shares some network layers with the multi-objective graph neural network to distinguish between real bidding data and abnormal samples generated by the generator; During adversarial training, the generator and discriminator are jointly optimized through the following loss function: Generator loss function: maximize the probability of misjudgment of the generated sample by the discriminator; Discriminator loss function: minimize the classification error rate of real samples and generated samples; The adversarial training loss function and the multi-task loss function are weightedly fused to improve the model's ability to identify abnormal bidding behaviors.

7. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The step of combining Pareto optimization to balance multi-task conflicts includes: Define domain-driven multi-objective evaluation indicators, which include a combination of market prediction accuracy, enterprise credit assessment AUC value, and risk warning false alarm rate, where the risk warning false alarm rate is constrained by a policy compliance threshold; Dynamically adjust the weight of each task loss function based on the Pareto frontier theory. The weight adjustment is achieved through the gradient descent algorithm to ensure the synchronous optimization of the performance of each task. During the optimization process, if the conflict between tasks exceeds the preset threshold, the task weight redistribution mechanism will be triggered to prioritize the false alarm rate of risk warning tasks to meet policy compliance requirements; Output optimized multi-task model parameters and analysis results.

8. The AI-based bidding data intelligent analysis method according to claim 1 is characterized in that: The output market forecast report, corporate credit profile and risk warning results include: Market trend forecast report: Based on the industry characteristic mean output by the multi-objective graph neural network, the market supply and demand forecast curve for the next 1-4 quarters is generated, and the impact range of policy and regulatory changes on the forecast results is marked; Enterprise credit profile: Integrate enterprise node characteristics, neighbor node characteristics and historical bidding records to generate credit scores and multi-dimensional risk labels, including the frequency of judicial litigation, volatility of bid winning rate, and intensity of risk transmission of related enterprises; Risk warning list: detect abnormal bidding behavior, output the list of companies suspected of bid rigging and collusion and the associated chain of evidence, which includes bid price difference threshold, bid time synchronization, and historical cooperation correlation indicators.

9. An AI-based bidding data intelligent analysis system, used to execute the AI-based bidding data intelligent analysis method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, used to integrate structured data, unstructured text and external data; Data processing module, which performs dynamic tensor decomposition and low-dimensional feature extraction; Dynamic graph modeling module, building and updating enterprise relationship graph; Multi-objective analysis module, which jointly optimizes market forecasting, credit assessment and risk warning tasks through graph neural networks; The optimization and output module performs adversarial training and Pareto optimization and generates analysis reports.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent analysis method for bidding data based on AI technology as described in any one of claims 1 to 8 is implemented.

Images