Bidding risk value calculation method and system based on deep learning

Abstract

The invention discloses a bidding and tendering risk value calculation method and system based on deep learning, and relates to the technical field of data mining and risk assessment crossing. A weighted analysis and time sequence change identification method is adopted to carry out risk influence degree importance evaluation and dynamic change trend identification of the space-time dimension on the bidding and tendering risk time sequence data set to form a time sequence risk change result, and the time sequence risk change result is fused with the multi-dimensional risk feature vector to construct a fusion risk feature matrix; performing risk relevancy calculation on the fusion risk feature matrix by using an adaptive optimization algorithm to generate a comprehensive risk index set; and according to the comprehensive risk index set, carrying out risk level mapping through principal component risk compression and clustering mode identification, and generating a bidding risk assessment report. According to the invention, by generating the bidding and tendering risk assessment report, dimension reduction optimization of complex risk features and automatic identification of a high-risk mode are realized, and the accuracy and interpretability of risk assessment are enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to a deep learning-based bidding risk value calculation method and system, and belongs to the technical field of data mining and risk assessment. BACKGROUND

[0002] In the field of bidding risk management, traditional risk assessment methods mainly rely on statistical models and single-task machine learning models. These methods usually independently analyze structured information such as project business data and bidder behavior data, use classic algorithms such as logistic regression and decision trees to establish risk prediction models, and output risk levels through threshold setting rules or weighted scorecard mechanisms. With the development of big data technology, some studies have begun to introduce time series analysis methods to monitor the bidding process state in stages, combining static features and dynamic parameters to build a risk early warning system.

[0003] However, there is still room for optimization in existing methods. Traditional single-task machine learning models have difficulty in effectively coordinating the relevance and differences between multiple risk objectives, resulting in limited collaborative expression of multi-dimensional risk features; conventional time series analysis methods rely on fixed time window statistics for capturing risk dynamic evolution patterns, lacking adaptive adjustment of risk feature importance over time and spatial dimensions, affecting the sensitivity of risk mutation points and the accuracy of trend judgment. SUMMARY

[0004] According to the problems described in the background, the application aims to provide a deep learning-based bidding risk value calculation method and system to solve the problems described above.

[0005] To achieve the above purpose, the application provides the following technical scheme: a deep learning-based bidding risk value calculation method is provided, comprising the following steps:

[0006] (1) Collect and preprocess multi-dimensional bidding data, and train a multi-task learning framework using the preprocessed multi-dimensional bidding data to obtain a multi-objective deep learning model;

[0007] (2) Feature extraction is performed on the preprocessed multi-dimensional bidding data by the multi-objective deep learning model, and a multi-dimensional risk feature vector is outputted; the process state parameters of the bidding are continuously monitored and fused with the multi-dimensional risk feature vector to generate a bidding risk time series dataset;

[0008] (3) A weighted analysis and time series change identification method is used to evaluate the importance of risk influence degree in spatial and temporal dimensions and identify the dynamic change trend of the bidding risk time series dataset, form a time series risk change result, and fuse it with the multi-dimensional risk feature vector to construct a fused risk feature matrix;

[0009] (4) using an adaptive optimization algorithm to calculate the risk correlation degree of the fusion risk feature matrix, to generate a comprehensive risk index set;

[0010] (5) according to the comprehensive risk index set, through principal component risk compression and clustering pattern recognition, risk grade mapping is carried out to generate a bidding risk assessment report.

[0011] Preferably, the multi-dimensional data of bidding in the step (1) includes project business data, bidder behavior data and external environment data; the preprocessing includes data cleaning, outlier rejection, missing value filling and format standardization.

[0012] Preferably, the step (1) of training the multi-task learning framework by using the preprocessed multi-dimensional data of bidding to obtain a multi-target deep learning model includes the following steps:

[0013] (1.1) input the preprocessed multi-dimensional data of bidding into the shared layer and task-specific layer of the multi-task learning framework, perform forward calculation, and obtain a multi-risk type prediction result set;

[0014] (1.2) compare the multi-risk type prediction result set with the real risk label in the preset historical risk event database, and calculate the joint loss function value;

[0015] (1.3) according to the joint loss function value, calculate all trainable parameters in the multi-task learning framework by using the back propagation algorithm, and update all trainable parameters by using the Adam adaptive optimizer;

[0016] (1.4) repeatedly execute the forward propagation, joint loss function value calculation, back propagation and trainable parameter updating process until the joint loss function value converges, and obtain the trained multi-target deep learning model.

[0017] Preferably, the step (2) of outputting the multi-dimensional risk feature vector includes the following steps:

[0018] (2.1) input the preprocessed multi-dimensional data of bidding into the shared layer of the multi-target deep learning model, extract the general feature representation reflecting the common characteristics of different risk dimensions through full connection transformation, and output the shared feature matrix;

[0019] (2.2) input the shared feature matrix into each task-specific layer, calculate through multi-task parallel learning and feature mapping, extract the feature information of different risk targets, and obtain a multi-task feature output set;

[0020] (2.3) integrate and reduce the dimension of the multi-task feature output set to generate a multi-dimensional risk feature vector.

[0021] Preferably, the generating the bidding risk time series dataset in step (2) comprises the following steps:

[0022] (2.4) associating and storing the multi-dimensional risk feature vector with the project business data to generate a bidding risk feature snapshot;

[0023] (2.5) continuously monitoring the process state parameters of the bidding based on the pre-processed multi-dimensional bidding data, and aligning and organizing them in time sequence to generate a bidding process state time series dataset;

[0024] (2.6) splicing the bidding process state time series dataset with the corresponding bidding risk feature snapshot to generate a bidding risk time series dataset.

[0025] Preferably, the forming the time series risk change result in step (3) comprises the following steps:

[0026] (3.1) calculating the information entropy and difference coefficient of the multi-dimensional risk feature vector in the bidding risk time series dataset by using the entropy weight method, and arranging them in descending order according to the difference coefficient to generate a risk feature importance list;

[0027] (3.2) calculating the mean, standard deviation and first-order difference slope of the risk features with high ranking in the risk feature importance list within different time windows to generate a risk feature dynamic index set;

[0028] (3.3) calculating the Spearman rank correlation coefficient of each index in the risk feature dynamic index set and the preset risk benchmark sequence to generate a dynamic-benchmark correlation degree set;

[0029] (3.4) assigning a decreasing weight to each risk feature in the risk feature importance list, and correcting the weight using the dynamic-benchmark correlation degree set to generate a feature quantization weight set;

[0030] (3.5) performing linear projection calculation on the original risk features at each time point according to the feature quantization weight set to generate a risk index time series curve;

[0031] (3.6) applying the cumulative sum control chart algorithm to statistically monitor the risk index time series curve to identify the mutation point time sequence;

[0032] (3.7) aligning and integrating the risk feature importance list, the risk feature dynamic index set and the mutation point time sequence to generate a time series risk change result.

[0033] Preferably, the constructing the fusion risk feature matrix in step (3) comprises the following steps:

[0034] (3.8) Align the time series risk change results with the multi-dimensional risk feature vector by timestamp, and splice along the feature dimension to generate a fusion feature vector sequence;

[0035] (3.9) Arrange the fusion feature vector sequence in chronological order to construct a fusion risk feature matrix.

[0036] Preferably, the step (4) of generating a comprehensive risk indicator set comprises the following steps:

[0037] (4.1) Based on the historical risk event database, the frequency and impact degree of each type of risk target are counted, and an initial risk target weight vector is calculated and output;

[0038] (4.2) Taking the initial risk target weight vector as the starting point, the fusion risk feature matrix is subjected to sparse regression by an adaptive optimization algorithm to generate a feature-risk correlation vector;

[0039] (4.3) According to the feature-risk correlation vector, the initial risk target weight is updated, and the initial risk target weight lower than the preset weight convergence threshold is screened out to generate a stable risk target weight vector;

[0040] (4.4) Linearly weighting and combining the stable risk target weight vector and the fusion risk feature matrix to generate a comprehensive risk indicator set.

[0041] Preferably, the step (5) of generating a bidding risk assessment report comprises the following steps:

[0042] (5.1) Map the comprehensive risk indicator set to an orthogonal basis vector by orthogonal transformation, and extract the orthogonal basis vector with the maximum variance to generate a core risk feature set;

[0043] (5.2) Perform unsupervised clustering analysis on the core risk feature set by K-means clustering algorithm to identify different risk pattern categories and generate a risk pattern clustering identifier;

[0044] (5.3) Map the risk pattern clustering identifier to a risk level in combination with the historical risk event database to output a risk level determination result;

[0045] (5.4) Integrate the core risk feature set and the risk level determination result to generate a bidding risk assessment report.

[0046] The application also provides a bidding risk value calculation system based on deep learning, comprising a data acquisition module, a feature extraction module, a risk time series analysis module, a risk indicator calculation module, and a risk level mapping module.

[0047] The data acquisition module is used to collect multi-dimensional bidding data, preprocess it, and use the preprocessed multi-dimensional bidding data to train the multi-task learning framework to obtain a multi-objective deep learning model.

[0048] The feature extraction module is used to extract features from the preprocessed multi-dimensional bidding data using a multi-objective deep learning model, output a multi-dimensional risk feature vector, continuously monitor the process status parameters of bidding and integrate them with the multi-dimensional risk feature vector to generate a bidding risk time series dataset.

[0049] The risk time series analysis module is used to assess the importance of risk impact in the spatiotemporal dimensions and identify dynamic change trends in the bidding risk time series dataset using weighted analysis and time series change identification methods, forming time series risk change results, and fusing them with multi-dimensional risk feature vectors to construct a fused risk feature matrix;

[0050] The risk indicator calculation module is used to calculate the risk correlation of the fused risk feature matrix using an adaptive optimization algorithm to generate a comprehensive risk indicator set.

[0051] The risk level mapping module is used to map risk levels based on a comprehensive risk index set through principal component risk compression and clustering pattern recognition, and generate a bidding risk assessment report.

[0052] The beneficial effects of this invention are:

[0053] By generating a time-series dataset of bidding and tendering risks, dynamic tracking and spatiotemporal integration of risks throughout the entire bidding and tendering process were achieved, capturing the evolution patterns of risks and thus improving the real-time performance and data coverage of risk monitoring. By generating bidding and tendering risk assessment reports, dimensionality reduction optimization of complex risk characteristics and automatic identification of high-risk patterns were achieved, enhancing the accuracy and interpretability of risk assessment. Attached Figure Description

[0054] Fig. 1 This is a flowchart of a deep learning-based method for calculating bidding risk values.

[0055] Fig. 2 This is a schematic diagram of a deep learning-based bidding risk value calculation system.

[0056] Fig. 3 A flowchart for generating the results of time-series risk changes.

[0057] Fig. 4 A flowchart for generating a bidding risk assessment report. Detailed Implementation

[0058] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0059] Example 1

[0060] like Figs. 1 to 4 As shown, this invention provides a method for calculating bidding risk value based on deep learning, including the following steps:

[0061] (1) Collect and preprocess multi-dimensional bidding data, and use the preprocessed multi-dimensional bidding data to train the multi-task learning framework to obtain a multi-objective deep learning model.

[0062] The multi-dimensional bidding data in step (1) includes project business data, bidder behavior data, and external environment data;

[0063] It should be noted that project business data includes basic project information, budget amount, schedule requirements, contract terms, and execution progress data. Among them, the execution progress data is further subdivided into fields such as bidding status, contract execution progress, fund payment records, and risk event trigger records during the bidding monitoring process, which are extracted from the bidding documents, procurement plans, and project descriptions. Bidder behavior data includes bid price records, bidding frequency, performance status, default records, and credit score data, which are collected in real time through the operation logs and historical record databases of the bidding platform. External environment data includes industry market data, policy and regulatory change data, and macroeconomic fluctuation data, which are obtained by accessing government open databases, industry data platforms, and third-party economic index API interfaces.

[0064] The preprocessing includes data cleaning, outlier removal, missing value filling, and format standardization.

[0065] It should be noted that data cleaning refers to filtering and unifying redundant fields and inconsistent records in multi-dimensional bidding data to eliminate data noise and improve data accuracy; outlier removal refers to using Z-Score statistical methods to detect and filter values ​​in project business data and bidder behavior data that significantly deviate from the normal range to eliminate noise interference and improve data quality; missing value imputation refers to using interpolation to complete missing data items in external environment data to ensure data integrity; and format standardization refers to the unified conversion of data fields, units, and codes from different sources to eliminate differences in units.

[0066] The step (1) of training the multi-task learning framework using preprocessed multi-dimensional bidding data to obtain a multi-objective deep learning model includes the following steps:

[0067] (1.1) Input the preprocessed multi-dimensional bidding data into the shared layer and task-specific layer of the multi-task learning framework, perform forward computation, and obtain a set of prediction results for multiple risk types;

[0068] Furthermore, the preprocessed multi-dimensional bidding data is input into the shared layer and task-specific layer of the multi-task learning framework. In the shared layer, linear transformation, nonlinear activation, and feature normalization are performed layer by layer on the preprocessed multi-dimensional bidding data to extract general feature representations that can simultaneously characterize cost risk, time risk, and quality risk. The general feature representations are then passed to the task-specific layer, where independent task branches are set up for cost risk, time risk, and quality risk, respectively. Each task branch receives the general feature representations and further performs weighted combination, nonlinear mapping, and risk correlation amplification processing to highlight the highly correlated feature patterns of the corresponding risk targets, generating prediction outputs for cost risk, time risk, and quality risk. The prediction outputs for cost risk, time risk, and quality risk are then arranged and merged according to the risk target category to generate a multi-risk type prediction result set.

[0069] (1.2) Compare the multi-risk type prediction result set with the real risk labels in the preset historical risk event database, and calculate the joint loss function value;

[0070] Furthermore, for the prediction output of each risk target in the multi-risk type prediction result set, the true risk label of the same risk target in the preset historical risk event database is searched, and the deviation between the prediction output and the true risk label is calculated; the cost risk deviation, time risk deviation and quality risk deviation are weighted and summed according to a unified loss measurement rule to obtain the joint loss function value that can simultaneously measure the overall error level of multi-task prediction.

[0071] It should be noted that the historical risk event database is a structured data set built based on records of risk events that have occurred in historical bidding projects. By aggregating multi-source data from project archives, contract execution records, credit rating reports, and regulatory notices, it organizes and standardizes the risk target identification, occurrence time, event type, loss amount, duration of impact, and scope of impact for each risk event. The unified loss measurement rule is based on the weighted mean square error calculation rule of the prediction errors of each risk target in the multi-task learning framework. It is obtained by squaring the difference between the predicted output of the three types of targets, namely cost risk, time risk, and quality risk, and the actual risk label, and then summing them according to the task importance coefficient.

[0072] (1.3) Based on the joint loss function value, the backpropagation algorithm is used to calculate all trainable parameters in the multi-task learning framework, and the Adam adaptive optimizer is used to iteratively update all trainable parameters;

[0073] Furthermore, using the joint loss function value as the target signal, the gradient value of each trainable parameter in the task-specific layer with respect to the joint loss function value is calculated layer by layer in reverse along the task-specific layer. The gradient value is then passed up the connection path to the shared layer, where the gradient value of each trainable parameter with respect to the joint loss function value is calculated. After obtaining the gradient values ​​of all trainable parameters in the shared layer and the task-specific layer of the multi-task learning framework, the Adam adaptive optimizer is used to iteratively adjust the current values ​​of the trainable parameters based on the gradient values, historical gradient accumulation, and adaptive learning rate. This allows the shared layer and the task-specific layer of the multi-task learning framework to gradually converge in the direction of reducing the joint loss function value, completing one round of parameter update.

[0074] (1.4) Repeatedly perform forward propagation, joint loss function value calculation, back propagation and trainable parameter update process until the joint loss function value converges and the trained multi-objective deep learning model is obtained.

[0075] Furthermore, after completing one round of parameter updates, the forward propagation calculation, joint loss function value calculation, backpropagation, and trainable parameter update process are repeated based on the updated multi-task learning framework. That is, the preprocessed multi-dimensional bidding data is input into the shared layer and task-specific layer of the multi-task learning framework again to generate a new multi-risk type prediction result set. The new multi-risk type prediction result set is compared with the real risk labels in the historical risk event database to obtain a new joint loss function value. Based on the new joint loss function value, new gradient calculations and new parameter updates are performed. This process is iterated until the joint loss function value gradually decreases and stabilizes in multiple iterations. When the joint loss function value remains within the preset convergence threshold in several iterations, it is determined that the trainable parameters of the shared layer and task-specific layer of the multi-task learning framework have effectively converged, and the trained multi-objective deep learning model is output.

[0076] It should be noted that the convergence threshold is a numerical limit set based on the rate of change of the joint loss function value. It is determined by statistical analysis of the decreasing trend of the joint loss function value during the early stage of training the multi-objective deep learning model. An exemplary value range is 1×10. -4 Up to 1×10 -6 .

[0077] (2) The preprocessed multi-dimensional bidding data is extracted using a multi-objective deep learning model, and a multi-dimensional risk feature vector is output. The process state parameters of bidding are continuously monitored and fused with the multi-dimensional risk feature vector to generate a bidding risk time series dataset.

[0078] The output multidimensional risk feature vector in step (2) includes the following steps:

[0079] (2.1) Input the preprocessed multi-dimensional bidding data into the shared layer of the multi-objective deep learning model, extract the general feature representation reflecting the common features of different risk dimensions through fully connected transformation, and output the shared feature matrix;

[0080] Furthermore, the preprocessed multi-dimensional bidding data is input into the shared layer of the multi-objective deep learning model. Fully connected matrix multiplication is performed in the shared layer, multiplying each input feature vector element-wise with the shared layer weight matrix and summing the results to obtain a linear transformation. The ReLU activation function is applied to the linear mapping result for non-linear transformation to generate feature activation values. The feature activation values ​​are then batch normalized to adjust the numerical distribution and maintain consistent feature scale. The calculation output of continuous multi-layer mapping and activation forms a general feature representation that can reflect the project's business characteristics, the bidder's behavioral characteristics, and the external environment characteristics. These features are then arranged in order of feature dimensions to form a shared feature matrix.

[0081] (2.2) Input the shared feature matrix into the specific layer of each task, and extract the feature information of different risk targets through multi-task parallel learning and feature mapping calculation to obtain the multi-task feature output set;

[0082] Furthermore, the shared feature matrix is ​​input into each task-specific layer of the multi-objective deep learning model. Each task-specific layer independently receives the shared feature matrix and uses the corresponding risk objective as its learning direction. A linear weight mapping is performed in each task-specific layer, multiplying the shared feature matrix with the task-specific layer weight matrix and adding a bias term to obtain the initial task feature representation. The sigmoid activation function is applied to compress the feature value range to highlight the risk-related region, and then Dropout random deactivation is performed to prevent overfitting. Feature weighting is performed on the initial task feature representation, using task-specific weight coefficients to enhance the feature components related to the current risk objective and suppress irrelevant components, outputting a task-specific risk feature vector. The three types of tasks—cost risk, time risk, and quality risk—are processed separately, and the task-specific risk feature vectors of the three types are arranged and merged in order of risk category to form a multi-task feature output set.

[0083] (2.3) Integrate and reduce the dimensionality of the multi-task feature output set to generate a multi-dimensional risk feature vector.

[0084] Furthermore, the feature vectors of each task are directly concatenated according to their feature dimensions to form a high-dimensional joint feature matrix containing all risk target features. Variance calculations are performed on each column of the high-dimensional joint feature matrix, followed by normalization to obtain the variance contribution rate. The high-dimensional joint feature matrix is ​​then sorted from largest to smallest according to its variance contribution, and the top few features with the highest variance contribution are selected as principal components to form the principal feature matrix. Covariance matrix operations are performed on the principal feature matrix, and eigenvalue decomposition is performed to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors are sorted according to their eigenvalues, and the top few eigenvectors with the largest eigenvalues ​​are used to form the principal component matrix. Matrix multiplication is then performed between the high-dimensional joint feature matrix and the principal component matrix to obtain a low-dimensional projection matrix. Each element in the low-dimensional projection matrix is ​​a linear combination coefficient of the corresponding feature along the principal component direction. The linear combination coefficients are rearranged according to the index of the principal component direction to form a single ordered feature sequence. All ordered feature sequences are then concatenated into a vector structure according to the sample order to output a multi-dimensional risk feature vector.

[0085] It should be noted that the linear combination coefficients are weight coefficients calculated from the mapping relationship between the eigenvectors of each principal component and the original feature basis when performing eigenvalue decomposition on the high-dimensional joint feature matrix. They are used to represent the degree of contribution of each original feature in the direction of constructing the principal components.

[0086] The step (2) of generating the bidding risk time series dataset includes the following steps:

[0087] (2.4) Associate and store the multidimensional risk feature vector with the project business data to generate a snapshot of the bidding risk features;

[0088] Furthermore, a matching index is established between the multidimensional risk feature vector and the project business data based on the project number, bidding time, and bidder identifier. During the matching process, the corresponding project business record is searched for each multidimensional risk feature vector to generate a one-to-one matching result. The matching results are arranged in chronological order and stored in a structured format. Each record contains the project number, bidder identifier, bidding timestamp, and corresponding multidimensional risk feature vector, forming a snapshot of bidding risk features.

[0089] (2.5) Based on the preprocessed multi-dimensional bidding data, continuously monitor the process status parameters of bidding, and align and organize them in chronological order to generate a bidding process status time series dataset;

[0090] Furthermore, based on the preprocessed multi-dimensional bidding data, the bidding status, contract execution progress, fund payment records, and risk event trigger records are continuously and periodically read from the project business data, and the status change values ​​are recorded at fixed time intervals. The collected process status parameters are sorted according to the timestamp order, and missing records are resampled at fixed intervals. The supplemented records are indexed by project number and timestamp to form a time-series dataset of bidding process status covering the entire process.

[0091] (2.6) The bidding process status time series dataset is spliced ​​with the corresponding bidding risk feature snapshot to generate the bidding risk time series dataset.

[0092] Furthermore, the bidding process status time series dataset and the corresponding bidding risk feature snapshot are spliced ​​together using project number and timestamp as matching conditions. In the splicing process, the status records of the same project at the same time point in the bidding process status time series dataset and the multi-dimensional risk feature vectors of the same time point in the bidding risk feature snapshot are merged into extended feature row vectors according to feature dimensions. All extended feature row vectors are arranged in chronological order to generate the bidding risk time series dataset.

[0093] (3) Using weighted analysis and time series change identification methods, the importance of risk impact in the spatiotemporal dimension and the dynamic change trend identification are carried out on the bidding risk time series dataset to form time series risk change results, and then fused with multi-dimensional risk feature vectors to construct a fused risk feature matrix;

[0094] The formation of the time-series risk change results in step (3) includes the following steps:

[0095] (3.1) The information entropy and difference coefficient of the multidimensional risk feature vector in the time series dataset of bidding risk are calculated by the entropy weight method, and the vectors are arranged in descending order of the difference coefficient to generate a list of risk feature importance;

[0096] Furthermore, each risk feature in the bidding risk time series dataset is normalized along the time dimension to obtain a normalized proportion value sequence; the product relationship between the proportion value and the logarithm is calculated at each time point of the normalized proportion value sequence to obtain the information entropy of the risk feature; the information entropy value is converted into a proportion to obtain the difference coefficient that reflects the fluctuation range of the risk feature; the difference coefficients of all risk features are arranged in order of numerical value to form a list of risk feature importance.

[0097] The expression for calculating the coefficient of difference is:

[0098] ;

[0099] in,

[0100] ;

[0101] in, It is the first The coefficient of difference for each risk characteristic; It is the first Information entropy of each risk characteristic; It is the first Risk characteristics within the time window The normalized ratio value; It represents the total number of time windows; It is the total number of risk characteristics; It is an index based on time location; It is the first Information entropy of each risk characteristic; It is a traversal index of risk characteristics.

[0102] (3.2) Calculate the mean, standard deviation and first difference slope of the risk features ranked first in the risk feature importance list in different time windows, and generate a dynamic index set of risk features;

[0103] Furthermore, a fixed-length sliding window is defined on the bidding risk time series dataset. The arithmetic mean and standard deviation of the numerical sequences of the risk features ranked first in the risk feature importance list within the window are calculated to measure the fluctuation range. The first-order difference slope is calculated by the numerical difference between adjacent time nodes to reflect the speed of risk change. The mean, standard deviation and first-order difference slope of each time window are combined to form a dynamic index set of risk features.

[0104] (3.3) Calculate the Spearman rank correlation coefficients of each indicator in the dynamic risk characteristic index set with the preset risk benchmark sequence, and generate a dynamic-benchmark correlation set;

[0105] Furthermore, the indicator sequences to be evaluated are selected one by one from the dynamic indicator set of risk characteristics. All values ​​of the indicator sequences to be evaluated are sorted according to their numerical values ​​and assigned a corresponding rank for each time window position, resulting in the rank sequence of the indicator sequences. Reference values ​​within the time window range corresponding to the indicator sequences to be evaluated are taken from the preset risk benchmark sequence, sorted and assigned ranks in the same way, resulting in the rank sequence of the risk benchmark sequence. The rank sequences of the indicator sequences and the rank sequences of the risk benchmark sequence are matched one-to-one according to the time window, and the sum of squared differences at the same time window position is calculated. The Spearman rank correlation coefficient is calculated according to the Spearman formula. The Spearman rank correlation coefficients of all risk characteristics are summarized to generate a dynamic-benchmark correlation set.

[0106] The expression for calculating the Spearman rank correlation coefficient is:

[0107] ;

[0108] in, It is the first Spearman's rank correlation coefficient for each risk characteristic; It is in the At each time window position, the rank difference between the rank of the indicator sequence to be evaluated and the rank of the risk benchmark sequence; It refers to the number of time windows; It is the time window position index; The Spearman constant coefficient parameter is set based on the mathematical derivation that the sum of squared rank differences maintains a consistent proportion within the range of standardized correlation coefficient values ​​(−1 to 1). It is used to ensure that the calculation results of the Spearman rank correlation coefficient have comparability and normalization characteristics.

[0109] It should be noted that the risk benchmark sequence is a standard reference sequence constructed based on the long-term average and trend characteristics of various typical risk features in the historical risk event database. It includes three parts: the average risk index sequence, the volatility trend sequence, and the abnormal deviation sequence. It is used to measure the degree of consistency between the risk features to be assessed and historical risk patterns in terms of time trends.

[0110] (3.4) Assign decreasing weights to each risk feature in the risk feature importance list, and use the dynamic-benchmark correlation set to correct the weights to generate a feature quantification weight set;

[0111] Furthermore, based on the ranking position in the list of risk feature importance, initial weight values ​​are assigned to each risk feature according to the principle of decreasing weight from higher to lower ranking. The Spearman rank correlation coefficient of each risk feature is then weighted and averaged with the initial weight value to obtain the corrected weight value for each risk feature. These corrected weight values ​​are then arranged in the order of the risk feature index to obtain the feature quantification weight set.

[0112] (3.5) Perform linear projection calculation on the original risk characteristics at each time point based on the feature quantization weight set to generate the risk index time series curve;

[0113] Furthermore, all risk feature values ​​at each time point are extracted from the bidding risk time series dataset, and weighted summation is performed with the corresponding modified weight values ​​in the feature quantification weight set to obtain a weighted comprehensive score as the risk index at each time point; the risk indices of all time points are arranged in ascending order of time and connected on the time axis to form a curve, and the risk index time series curve is output.

[0114] (3.6) Apply the cumulative sum control chart algorithm to perform statistical process monitoring on the risk index time series curve and identify the time series of abrupt change points;

[0115] Furthermore, the population mean and standard deviation of the risk index time series curve are calculated, with the population mean serving as the control center line and the allowable deviation constant set based on the standard deviation. At each time point, the deviation between the risk index and the control center line is calculated, and the deviation is corrected according to a preset reference deviation threshold to obtain the corrected deviation amount. Simultaneously, the corrected deviation amount is progressively accumulated in chronological order, and positive and negative cumulative sums are calculated separately to form a cumulative sum sequence that reflects the rising and falling trends of the risk level. When the positive cumulative sum exceeds the upper limit control threshold or the negative cumulative sum exceeds the lower limit control threshold, the time point is recorded as the sudden anomaly location. All sudden anomaly locations are arranged in chronological order to obtain the time series of abrupt change points.

[0116] It should be noted that the reference deviation threshold is a judgment limit set based on the historical stable range fluctuation amplitude of the risk index time series curve, used to identify abnormal fluctuations deviating from the mean. An exemplary range is ±1 to ±3 times the mean standard deviation. The upper and lower control thresholds are set based on the long-term fluctuation statistics of the risk index time series curve, used to identify abnormal fluctuations with significant increases and decreases in risk levels. An exemplary range for the upper control threshold is 0.5 to 2 times the standard deviation above the control center line, and an exemplary range for the lower control threshold is 0.5 to 2 times the standard deviation below the control center line. The cumulative sum control chart algorithm is a dynamic process monitoring method based on the statistical deviation of time series data. Its principle is to continuously track the fluctuation trend of process variables by gradually accumulating the deviation of each observation point relative to the center line, under the condition of setting a control center line and an allowable offset constant. The cumulative sum control chart algorithm is used to sensitively identify long-term, small risk level shifts in the monitoring of the risk index time series curve.

[0117] (3.7) Align and integrate the list of risk feature importance, the set of dynamic indicators of risk features and the time series of mutation points to generate time series risk change results.

[0118] Furthermore, using timestamps and risk feature identifiers as a joint index, the mean, standard deviation, and first-order difference slope of each time window position are selected from the dynamic indicator set of risk features. At the same time, it is compared whether there are mutation points in the same time window position in the mutation point time series, and the time window position with mutation points is marked as the risk mutation period. The difference coefficient of the corresponding risk feature is read from the risk feature importance list, and merged with the dynamic indicators and mutation marker information in the same time window into a record. The records are arranged in chronological order to form a multidimensional structured data table, generating the time series risk change results.

[0119] The construction of the fusion risk feature matrix in step (3) includes the following steps:

[0120] (3.8) Align the time series risk change results with the multidimensional risk feature vector by timestamp and concatenate them along the feature dimension to generate a fused feature vector sequence;

[0121] Furthermore, for each timestamp, the corresponding risk feature time series description record is read from the time series risk change results, and the corresponding multidimensional risk feature vector is read from the multidimensional risk feature vector; the risk feature time series description record is appended to the tail of the corresponding multidimensional risk feature vector to form a fusion feature vector corresponding to a single timestamp, and the fusion feature vectors of all timestamps are arranged in chronological order to generate a fusion feature vector sequence.

[0122] (3.9) Arrange the fusion feature vector sequence in chronological order to construct the fusion risk feature matrix.

[0123] Furthermore, using time sequence as the row direction and the feature dimension in the fused feature vector as the column direction, the fused feature vector sequence is written into a two-dimensional matrix structure row by row. This allows fused feature vectors with the same time sequence to be aggregated in the same row, and records with different timestamps to be arranged in chronological order, thus obtaining the fused risk feature matrix.

[0124] (4) Use an adaptive optimization algorithm to calculate the risk correlation of the fused risk feature matrix and generate a comprehensive risk index set;

[0125] The generation of the comprehensive risk index set in step (4) includes the following steps:

[0126] (4.1) Based on the historical risk event database, statistically analyze the frequency and impact of various risk targets, and calculate and output the initial risk target weight vector;

[0127] Furthermore, the system reads the risk target identifier, project to which it belongs, occurrence time, actual impact scope, and impact intensity fields recorded in the historical risk event database. It then statistically analyzes the frequency of occurrence of each risk target across different projects to calculate the occurrence frequency index for each risk target. The system also statistically analyzes the impact intensity of each risk target and calculates the average impact intensity level of the same risk target as an impact degree index. For each risk target, the occurrence frequency index and impact degree index are standardized and then weighted to obtain the initial weight value of the risk target. Finally, the initial weight values ​​of all risk targets are arranged in order of their risk target identifiers to form an initial risk target weight vector.

[0128] (4.2) Starting from the initial risk target weight vector, the fused risk feature matrix is ​​subjected to sparse regression through an adaptive optimization algorithm to generate a feature-risk correlation vector;

[0129] Furthermore, the fused risk feature matrix is ​​matched with the initial risk target weight vector, establishing a mapping relationship between each column of fused risk features in the fused risk feature matrix and the corresponding risk target in the initial risk target weight vector. The sparse regression optimization process is as follows: In the sparse regression optimization process, the fused risk feature matrix is ​​used as the input structure, and the initial risk target weight vector is used as the initial parameter. The correlation between each fused risk feature and its corresponding risk target is calculated sequentially, and sparse constraints are applied to gradually reduce the influence of fused risk features with low correlation during iteration, while maintaining a higher influence of fused risk features with high correlation during iteration. The adaptive optimization algorithm automatically adjusts the iteration step size in each iteration based on the change in correlation obtained in the previous iteration, accelerating the convergence process of fused risk features with high correlation and their corresponding risk targets, while suppressing the interference of irrelevant fused risk features. The sparse regression optimization is continuously executed until the change in correlation tends to stabilize, and the correlation value of each fused risk feature relative to each risk target is output and arranged in order of fused risk features to form a feature-risk correlation vector.

[0130] It should be noted that the adaptive optimization algorithm is an optimization method that improves parameter convergence efficiency by dynamically adjusting the learning rate or weight update step size. It is used to calculate the risk correlation of the fused risk feature matrix. By iteratively updating and converging the weights of each risk target, it dynamically optimizes the weight distribution relationship between risk features and risk targets. Sparse constraint is a calculation strategy that highlights the contribution of key features by limiting the intensity of feature participation. The principle is to set sparsity conditions during the correlation calculation process, so that the correlation of some features gradually approaches zero, and only feature values ​​that are obviously related to the risk target are retained for subsequent calculations. Through sparsity constraint, the interference of redundant features can be effectively reduced, making the correlation between the fused risk features and risk targets more representative and distinguishable, and improving the stability and accuracy of risk correlation calculation.

[0131] (4.3) Update the initial risk target weights based on the feature-risk correlation vector, and select the initial risk target weights that are lower than the preset weight convergence threshold to generate a stable risk target weight vector;

[0132] Furthermore, taking the same initial risk target as the corresponding object, the correlation value in the feature-risk correlation vector is compared with the initial weight value in the initial risk target weight vector. For initial risk targets with higher correlation values, weight enhancement processing is performed to increase the corresponding initial weight value, while for initial risk targets with lower correlation values, weight suppression processing is performed to decrease the corresponding initial weight value. The weights of all risk targets are iteratively optimized. When the change in the initial weight value is lower than the preset weight convergence threshold in multiple consecutive iterations, it is determined that the initial weight value of the initial risk target has converged. The initial weight values ​​that have not yet reached the weight convergence threshold are removed, and the converged initial weight values ​​are retained to form a stable risk target weight vector.

[0133] It should be noted that the weight convergence threshold is a numerical limit set based on the stability requirement of the change amplitude between two consecutive rounds of weight update results. It is used to determine whether the risk target weights have reached a convergence state, and the exemplary value range is 1×10. -4 Up to 1×10 -6 .

[0134] (4.4) The stable risk target weight vector and the fusion risk feature matrix are linearly weighted and combined to generate a comprehensive risk index set.

[0135] Furthermore, in the fusion risk feature matrix, column vectors corresponding to the stable risk target weights are selected according to the order of the fusion risk features. The value of each fusion risk feature is proportionally weighted with the corresponding stable risk target weight to obtain the fusion risk feature weighted vector. All fusion risk feature weighted vectors are aggregated in the row direction to generate a comprehensive risk index set.

[0136] (5) Based on the comprehensive risk index set, risk level mapping is performed through principal component risk compression and clustering pattern recognition to generate a bidding risk assessment report.

[0137] The generation of the bidding risk assessment report in step (5) includes the following steps:

[0138] (5.1) The comprehensive risk index set is mapped to orthogonal basis vectors through orthogonal transformation, and the orthogonal basis vector with the largest variance is extracted to generate the core risk feature set;

[0139] Furthermore, the covariance matrix of the comprehensive risk index set is calculated using the Pearson product-moment correlation method, and eigenvalue decomposition is performed on the covariance matrix to extract eigenvalues ​​and corresponding eigenvectors. The eigenvectors are normalized to construct a set of mutually orthogonal basis vectors, and the set of mutually orthogonal basis vectors is sorted according to the size of the eigenvalues. Several orthogonal basis vectors with the largest variance are selected as the main risk directions to form the core risk feature set.

[0140] It should be noted that the Pearson product-moment correlation method is a classic statistical method for measuring the degree of linear correlation between two variables. It can not only be used to assess the correlation strength between variables, but also to form a complete covariance matrix by calculating the correlation between multiple variables pairwise, providing a basic data structure for subsequent eigenvalue decomposition, principal component extraction and multidimensional risk feature compression.

[0141] (5.2) The core risk feature set is subjected to unsupervised clustering analysis by K-means clustering algorithm to identify different risk pattern categories and generate risk pattern clustering labels;

[0142] Furthermore, all feature records are read from the core risk feature set, and the number of clusters is set for the K-means clustering algorithm. Several feature records are randomly selected as initial cluster centers. For each feature record in the core risk feature set, the Euclidean distance between it and each initial cluster center is calculated. Each feature record is temporarily assigned to the cluster category corresponding to the initial cluster center with the shortest Euclidean distance. After one assignment, the average feature value of all feature records under each cluster category is calculated as the new cluster center position. The Euclidean distance calculation, record assignment, and cluster center position update operations are repeated using the new cluster center until the change in cluster center position in two consecutive iterations is lower than the preset stopping condition. The final cluster category number is assigned to each feature record. The cluster category number is output in the order of the feature records to form the risk pattern cluster identifier.

[0143] It should be noted that the stopping condition is a convergence threshold set based on the magnitude of the change in the cluster center positions between two adjacent iterations. The clustering result is considered stable when the average rate of change is calculated by statistically analyzing the positional shift of the cluster centers after each update and when the average rate of change is lower than the convergence threshold. An exemplary value range is 1 × 10⁻⁶. -3 Up to 1×10 -5 .

[0144] (5.3) Combining the historical risk event database, the risk pattern clustering identifiers are mapped to risk levels, and the risk level determination results are output;

[0145] Furthermore, the historical risk event database is queried for historical samples with the same clustering identifier as the risk pattern clustering identifier, and the actual distribution frequency of different risk levels in the historical samples is counted. Based on the actual distribution frequency, the mapping rule between the risk pattern clustering identifier and the risk level is determined (e.g., if there are 100 samples marked as "clustering identifier A" in the historical risk event database, of which 85 have an actual risk level of "high risk", 10 have "medium risk" and 5 have "low risk", then based on the principle of highest frequency, the mapping rule is determined as: risk pattern clustering identifier A corresponds to the "high risk" level). That is, each risk pattern clustering identifier corresponds to a risk level that appears most frequently in the historical samples. The mapping rule is then applied to convert the current risk pattern clustering identifier into the corresponding risk level, generating a risk level determination result.

[0146] (5.4) Integrate the core risk feature set with the risk level determination results to generate a bidding risk assessment report.

[0147] Furthermore, each feature record in the core risk feature set is paired one-to-one with the corresponding risk level label in the risk level determination result, and the corresponding risk level label is attached to form a feature record entry with risk level information. Using the project number, timestamp, and risk level label as index fields, all feature record entries are rearranged in chronological order to form a time-progressing risk level description sequence. The risk level description sequence is then organized into a reportable output format, which includes the risk feature structure corresponding to the core risk feature set, the risk pattern category corresponding to the risk pattern clustering identifier, and the risk level label corresponding to the risk level determination result, generating a bidding risk assessment report.

[0148] The present invention also provides a bidding risk value calculation system based on deep learning, including a data acquisition module, a feature extraction module, a risk time series analysis module, a risk indicator calculation module, and a risk level mapping module;

[0149] The data acquisition module is used to collect multi-dimensional bidding data, preprocess it, and use the preprocessed multi-dimensional bidding data to train the multi-task learning framework to obtain a multi-objective deep learning model.

[0150] The feature extraction module is used to extract features from the preprocessed multi-dimensional bidding data using a multi-objective deep learning model, output a multi-dimensional risk feature vector, continuously monitor the process status parameters of bidding and integrate them with the multi-dimensional risk feature vector to generate a bidding risk time series dataset.

[0151] The risk time series analysis module is used to assess the importance of risk impact in the spatiotemporal dimensions and identify dynamic change trends in the bidding risk time series dataset using weighted analysis and time series change identification methods, forming time series risk change results, and fusing them with multi-dimensional risk feature vectors to construct a fused risk feature matrix;

[0152] The risk indicator calculation module is used to calculate the risk correlation of the fused risk feature matrix using an adaptive optimization algorithm to generate a comprehensive risk indicator set.

[0153] The risk level mapping module is used to map risk levels based on a comprehensive risk index set through principal component risk compression and clustering pattern recognition, and generate a bidding risk assessment report.

[0154] This invention achieves dynamic tracking and spatiotemporal integration of risks throughout the entire bidding process by generating a time-series dataset of bidding risks, capturing the evolution patterns of risks, and thus improving the real-time performance and data coverage of risk monitoring; and by generating a bidding risk assessment report, it achieves dimensionality reduction optimization of complex risk characteristics and automatic identification of high-risk patterns, enhancing the accuracy and interpretability of risk assessment.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for calculating bidding risk value based on deep learning, characterized in that, Includes the following steps: (1) Collect and preprocess multi-dimensional bidding data, and use the preprocessed multi-dimensional bidding data to train the multi-task learning framework to obtain a multi-objective deep learning model. (2) The preprocessed multi-dimensional bidding data is extracted using a multi-objective deep learning model, and a multi-dimensional risk feature vector is output. The process state parameters of bidding are continuously monitored and fused with the multi-dimensional risk feature vector to generate a bidding risk time series dataset. (3) Using weighted analysis and time series change identification methods, the importance of risk impact in the spatiotemporal dimension and the dynamic change trend identification are carried out on the bidding risk time series dataset to form time series risk change results, and then fused with multi-dimensional risk feature vectors to construct a fused risk feature matrix; (4) Use an adaptive optimization algorithm to calculate the risk correlation of the fused risk feature matrix and generate a comprehensive risk index set; (5) Based on the comprehensive risk index set, risk level mapping is performed through principal component risk compression and clustering pattern recognition to generate a bidding risk assessment report.

2. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The multi-dimensional bidding data in step (1) includes project business data, bidder behavior data and external environment data; the preprocessing includes data cleaning, outlier removal, missing value filling and format standardization.

3. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The step (1) of training the multi-task learning framework using preprocessed multi-dimensional bidding data to obtain a multi-objective deep learning model includes the following steps: (1.1) Input the preprocessed multi-dimensional bidding data into the shared layer and task-specific layer of the multi-task learning framework, perform forward computation, and obtain a set of prediction results for multiple risk types; (1.2) Compare the multi-risk type prediction result set with the real risk labels in the preset historical risk event database, and calculate the joint loss function value; (1.3) Based on the joint loss function value, the backpropagation algorithm is used to calculate all trainable parameters in the multi-task learning framework, and the Adam adaptive optimizer is used to iteratively update all trainable parameters; (1.4) Repeatedly perform forward propagation, joint loss function value calculation, back propagation and trainable parameter update process until the joint loss function value converges and the trained multi-objective deep learning model is obtained.

4. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The output multidimensional risk feature vector in step (2) includes the following steps: (2.1) Input the preprocessed multi-dimensional bidding data into the shared layer of the multi-objective deep learning model, extract the general feature representation reflecting the common features of different risk dimensions through fully connected transformation, and output the shared feature matrix; (2.2) Input the shared feature matrix into the specific layer of each task, and extract the feature information of different risk targets through multi-task parallel learning and feature mapping calculation to obtain the multi-task feature output set; (2.3) Integrate and reduce the dimensionality of the multi-task feature output set to generate a multi-dimensional risk feature vector.

5. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The step (2) of generating the bidding risk time series dataset includes the following steps: (2.4) Associate and store the multidimensional risk feature vector with the project business data to generate a snapshot of the bidding risk features; (2.5) Based on the preprocessed multi-dimensional bidding data, continuously monitor the process status parameters of bidding, and align and organize them in chronological order to generate a bidding process status time series dataset; (2.6) The bidding process status time series dataset is spliced ​​with the corresponding bidding risk feature snapshot to generate the bidding risk time series dataset.

6. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The formation of the time-series risk change results in step (3) includes the following steps: (3.1) The information entropy and difference coefficient of the multidimensional risk feature vector in the time series dataset of bidding risk are calculated by the entropy weight method, and the vectors are arranged in descending order of the difference coefficient to generate a list of risk feature importance; (3.2) Calculate the mean, standard deviation and first difference slope of the risk features ranked first in the risk feature importance list in different time windows, and generate a dynamic index set of risk features; (3.3) Calculate the Spearman rank correlation coefficients of each indicator in the dynamic risk characteristic index set with the preset risk benchmark sequence, and generate a dynamic-benchmark correlation set; (3.4) Assign decreasing weights to each risk feature in the risk feature importance list, and use the dynamic-benchmark correlation set to correct the weights to generate a feature quantification weight set; (3.5) Perform linear projection calculation on the original risk characteristics at each time point based on the feature quantization weight set to generate the risk index time series curve; (3.6) Apply the cumulative sum control chart algorithm to perform statistical process monitoring on the risk index time series curve and identify the time series of abrupt change points; (3.7) Align and integrate the list of risk feature importance, the set of dynamic indicators of risk features and the time series of mutation points to generate time series risk change results.

7. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The construction of the fusion risk feature matrix in step (3) includes the following steps: (3.8) Align the time series risk change results with the multidimensional risk feature vector by timestamp and concatenate them along the feature dimension to generate a fused feature vector sequence; (3.9) Arrange the fusion feature vector sequence in chronological order to construct the fusion risk feature matrix.

8. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The generation of the comprehensive risk index set in step (4) includes the following steps: (4.1) Based on the historical risk event database, statistically analyze the frequency and impact of various risk targets, and calculate and output the initial risk target weight vector; (4.2) Starting from the initial risk target weight vector, the fused risk feature matrix is ​​subjected to sparse regression through an adaptive optimization algorithm to generate a feature-risk correlation vector; (4.3) Update the initial risk target weights based on the feature-risk correlation vector, and select the initial risk target weights that are lower than the preset weight convergence threshold to generate a stable risk target weight vector; (4.4) The stable risk target weight vector and the fusion risk feature matrix are linearly weighted and combined to generate a comprehensive risk index set.

9. The method for calculating bidding risk value based on deep learning according to claim 1, characterized in that, The generation of the bidding risk assessment report in step (5) includes the following steps: (5.1) The comprehensive risk index set is mapped to orthogonal basis vectors through orthogonal transformation, and the orthogonal basis vector with the largest variance is extracted to generate the core risk feature set; (5.2) The core risk feature set is subjected to unsupervised clustering analysis by K-means clustering algorithm to identify different risk pattern categories and generate risk pattern clustering labels; (5.3) Combining the historical risk event database, the risk pattern clustering identifiers are mapped to risk levels, and the risk level determination results are output; (5.4) Integrate the core risk feature set with the risk level determination results to generate a bidding risk assessment report.

10. A deep learning-based system for calculating bidding risk values, characterized in that, It includes a data acquisition module, a feature extraction module, a risk time series analysis module, a risk indicator calculation module, and a risk level mapping module; The data acquisition module is used to collect multi-dimensional bidding data, preprocess it, and use the preprocessed multi-dimensional bidding data to train the multi-task learning framework to obtain a multi-objective deep learning model. The feature extraction module is used to extract features from the preprocessed multi-dimensional bidding data using a multi-objective deep learning model, output a multi-dimensional risk feature vector, continuously monitor the process status parameters of bidding and integrate them with the multi-dimensional risk feature vector to generate a bidding risk time series dataset. The risk time series analysis module is used to assess the importance of risk impact in the spatiotemporal dimensions and identify dynamic change trends in the bidding risk time series dataset using weighted analysis and time series change identification methods, forming time series risk change results, and fusing them with multi-dimensional risk feature vectors to construct a fused risk feature matrix; The risk indicator calculation module is used to calculate the risk correlation of the fused risk feature matrix using an adaptive optimization algorithm to generate a comprehensive risk indicator set. The risk level mapping module is used to map risk levels based on a comprehensive risk index set through principal component risk compression and clustering pattern recognition, and generate a bidding risk assessment report.

Images