Bill risk data analysis method and device based on big model and context information
The method and device leverage a large model and context information to address data and model limitations in bill risk analysis, achieving accurate and explainable risk assessments.
Patent Information
- Application Number
- CN202411222987.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-03
- Publication Date
- 2024-12-24
- Estimated Expiration
- 2044-09-03
AI Technical Summary
Traditional bill risk analysis methods face limitations in data collection scope, data quality, feature construction, model generalization, and result comprehensibility, leading to inaccurate and unreliable risk assessments.
A method and device utilizing a large model and context information, incorporating distributed crawler technology, advanced data cleaning, feature engineering with PCA and GCN, mixed precision training, and explainable AI to enhance data quality, model accuracy, and result clarity.
Enables comprehensive and accurate bill risk assessment with enhanced model representation and generalization, providing clear explanations for improved decision-making.
Smart Images
Figure CN118735687B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bill risk data processing, and more specifically, to a bill risk data analysis method and device based on a large model and context information. Background Art
[0002] In today's economic activities, bills are an important financial instrument, and their transaction scale and complexity are increasing. However, with the rapid development of bill business, the risks associated with it are becoming increasingly prominent. Accurate and timely analysis and evaluation of bill risks are of vital importance to ensure the stability of the financial market, maintain the capital security of enterprises, and promote the healthy development of the economy.
[0003] Traditional bill risk analysis methods often rely on manual experience and simple statistical models, which have many limitations. These methods can usually only process limited bill data and features, and it is difficult to fully explore the hidden information and complex relationships in the data. For example, traditional methods are often unable to effectively extract and utilize the contextual information of bills, such as the time series of transactions, location distribution, and the relationship network of related parties, resulting in insufficient accuracy and comprehensiveness of risk assessment.
[0004] In terms of data collection, traditional methods may have problems such as narrow collection scope, incomplete data or untimely updates, which makes it difficult to adapt to the rapid changes in bill business and the needs of large-scale data processing. In the data cleaning and feature engineering stages, the lack of efficient and intelligent technical means can easily lead to low data quality and unreasonable feature construction, affecting the subsequent risk analysis results.
[0005] In terms of model training and risk assessment, traditional machine learning models show performance bottlenecks and insufficient generalization capabilities when processing massive, high-dimensional bill data. Moreover, these models often find it difficult to capture the dynamic changes and uncertainties of bill risks, which calls into question the reliability and timeliness of risk assessment results.
[0006] In addition, the existing bill risk analysis methods also have shortcomings in result interpretation and comprehensibility, making it difficult to provide users with clear and easy-to-understand risk explanations, which is not conducive to decision-making and risk management.
[0007] With the development of big data technology, deep learning and artificial intelligence, new opportunities have been brought to the analysis of bill risk data. The combination of big models and contextual information can more comprehensively and deeply explore the potential risk patterns in bill data and improve the accuracy and reliability of risk assessment. However, in this field, the relevant technologies and methods are still in the stage of continuous exploration and development, and a mature and complete solution has not yet been formed.
[0008] In summary, in order to cope with the increasingly complex bill risks, there is an urgent need for an innovative bill risk data analysis method and device based on large models and contextual information to achieve more accurate, efficient and explainable bill risk assessment and provide strong support for risk management in the financial field. Summary of the invention
[0009] The purpose of the present invention is to provide a bill risk data analysis method and device based on a large model and context information to solve the problems raised in the above background technology.
[0010] A bill risk data analysis method based on a large model and context information includes the following steps:
[0011] S1. Data collection: Collect basic information of bills, transaction records, related party information, and bill endorsement details. Suppose the total amount of data collected is ;
[0012] S2. Data cleaning: Clean the collected data to remove duplicate, erroneous and incomplete data. Suppose the data after cleaning is effective. ;
[0013] S3. Feature Engineering: Construct features from the cleaned data. The features include statistical features of bill amount, transaction frequency features, and credit features of related parties. Suppose the number of constructed features is ;
[0014] S4. Context information extraction: Extract context information from bill data, including time series characteristics of transactions, geographical distribution characteristics of transaction locations, and topological characteristics of the relationship network of related parties. Suppose the number of context features is ;
[0015] S5. Large model training: Use the extracted context information and constructed features to train a large model based on deep learning. Suppose the loss function of the model training is:
[0016] ;
[0017] in is the sample size, It is a real risk label. is the model's predicted value, is the mean square error, which is used to measure the difference between the predicted value and the true value. is the L1 regularization term, which is used to control the complexity of the model and avoid overfitting. is the regularization parameter, is the L2 regularization term, is the regularization parameter, is the KL divergence term used to deal with differences in the distribution of contextual information, is the weight parameter, and They are the true distribution of context features and the model predicted distribution;
[0018] S6. Model optimization: Use early stopping method to prevent overfitting, and set the patience value of early stopping to , dynamically adjust the training process based on the performance of the validation set;
[0019] S7. Risk assessment: Input the bill data to be analyzed into the trained large model and calculate the risk assessment index. The calculation formula of the risk assessment index is:
[0020] ;
[0021] in and Respectively context features and The weight of the constructed features, and Based on the large model context features and A prediction function that constructs features, The input bill data;
[0022] S8. Result interpretation step: Use explainable artificial intelligence (XAI) technology to explain and illustrate the risk assessment results. Set the clarity index of the explanation as .
[0023] A bill risk data analysis device based on a large model and context information, comprising:
[0024] Data collection module: used to collect bill-related data;
[0025] Data cleaning module: cleans and filters the collected data;
[0026] Feature Engineering Module: Construct effective features from cleaned data;
[0027] Context information extraction module: extracts context information from bill data;
[0028] Large model training module: Use the extracted context information and constructed features to train the large model;
[0029] Model optimization module: uses early stopping technology to optimize the model training process;
[0030] Risk assessment module: used to input the bill data to be analyzed into the trained large model and calculate the risk assessment index;
[0031] Result interpretation module: Use explainable artificial intelligence technology to explain the evaluation results.
[0032] Preferably, in the data collection step, a distributed crawler technology is used, and in the data cleaning step, a cleaning method combining rule-based and machine learning is applied.
[0033] Preferably, in the feature engineering step, principal component analysis PCA is introduced to perform feature dimensionality reduction and reduce feature redundancy, and in the context information extraction step, graph convolutional neural network GCN is used.
[0034] Preferably, in the large model training step, a mixed precision training strategy is adopted and a self-supervised learning task is introduced to enhance the representation ability of the model.
[0035] Preferably, the data cleaning module uses Hadoop or Spark distributed computing framework, distributes cleaning tasks to multiple computing nodes for parallel processing, and adopts multiple data verification and error correction mechanisms.
[0036] Preferably, the model optimization module adopts a variant optimization algorithm of Adagrad or Adadelta or RMSProp stochastic gradient descent SGD to adaptively adjust the learning rate according to the training history of the model, improve the convergence speed and stability of the model, and avoid falling into a local optimal solution.
[0037] Preferably, in the distributed crawler technology, an intelligent scheduling algorithm is used to dynamically allocate crawler resources according to the priority of the data source, the frequency of data update, and network status factors; in the feature engineering step, the generative adversarial network GAN or the autoencoder AE automatic generation technology is used to automatically mine valuable features.
[0038] Preferably, in principal component analysis, a principal component analysis method based on a Gaussian kernel function or a polynomial kernel function is used to process nonlinear data structures and extract principal components; when using a graph convolutional neural network, a multi-scale graph convolution operation is defined to capture information at different levels and granularities by performing convolution operations on the graph structure at different scales.
[0039] Preferably, in the mixed precision training strategy, an adaptive threshold adjustment mechanism for precision switching is set to automatically adjust the threshold for precision switching according to the training progress of the model, data distribution, and changes in computing resources.
[0040] Compared with the prior art, the advantages of the present invention are:
[0041] The present invention adopts distributed crawler technology to widely collect various detailed information of bills in the data collection step, including basic information, transaction records, related party information and bill endorsement details, thereby achieving comprehensiveness and efficiency of data collection, providing a rich and accurate data basis for subsequent risk analysis, and greatly improving the ability to grasp the overall picture of bill risks.
[0042] The present invention applies a cleaning method based on a combination of rules and machine learning and uses Hadoop or Spark distributed computing frameworks to perform data processing in the data cleaning module. It can not only quickly remove duplicate, erroneous and incomplete data, but also adopts multiple data verification and error correction mechanisms to ensure the high quality of the cleaned data, providing a reliable data source for subsequent analysis work, and effectively ensuring the accuracy of risk assessment.
[0043] The present invention effectively reduces feature redundancy, mines potential value features, and enriches the feature space by introducing advanced technologies in the feature engineering step, such as principal component analysis (PCA) for feature dimensionality reduction and generative adversarial network (GAN) or autoencoder (AE) to automatically generate features. This allows the constructed features to more accurately reflect the risk characteristics of the bill, providing strong support for the training and risk assessment of large models.
[0044] The present invention greatly enhances the representation and generalization capabilities of the model by adopting a mixed precision training strategy in large model training, introducing self-supervised learning tasks, and using a large model of deep learning to combine the extracted rich contextual information and constructed high-quality features. It can accurately capture the complex patterns and dynamic changes of bill risks, thereby achieving accurate assessment of bill risks and providing a scientific basis for risk management decisions.
[0045] The present invention uses explainable artificial intelligence (XAI) technology to clearly explain and illustrate the risk assessment results in the result interpretation step, thereby improving the clarity index of the explanation, enabling users to intuitively understand the process and results of risk assessment, and facilitating users to take corresponding risk management measures based on the interpretation results. At the same time, it also enhances users' trust and acceptance of the risk analysis system, and promotes the effective implementation of risk analysis work in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a schematic block diagram of the bill risk data analysis method based on a large model and context information of the present invention. DETAILED DESCRIPTION
[0047] Example: See Figure 1 , a bill risk data analysis method based on a large model and context information, comprising the following steps:
[0048] S1. Data collection: Collect basic information of bills, transaction records, related party information, and bill endorsement details. Suppose the total amount of data collected is ;
[0049] S2. Data cleaning: Clean the collected data to remove duplicate, erroneous and incomplete data. Suppose the data after cleaning is effective. ;
[0050] S3. Feature Engineering: Construct features from the cleaned data. The features include statistical features of bill amount, transaction frequency features, and credit features of related parties. Suppose the number of constructed features is ;
[0051] S4. Context information extraction: Extract context information from bill data, including time series characteristics of transactions, geographical distribution characteristics of transaction locations, and topological characteristics of the relationship network of related parties. Suppose the number of context features is ;
[0052] S5. Large model training: Use the extracted context information and constructed features to train a large model based on deep learning. Suppose the loss function of the model training is:
[0053] ;
[0054] in is the sample size, It is a real risk label. is the model's predicted value, is the mean square error, which is used to measure the difference between the predicted value and the true value. is the L1 regularization term, which is used to control the complexity of the model and avoid overfitting. is the regularization parameter, is the L2 regularization term, is the regularization parameter, is the KL divergence term used to deal with differences in the distribution of contextual information, is the weight parameter, and They are the true distribution of context features and the model predicted distribution;
[0055] S6. Model optimization: Use early stopping method to prevent overfitting, and set the patience value of early stopping to , dynamically adjust the training process based on the performance of the validation set;
[0056] S7. Risk assessment: Input the bill data to be analyzed into the trained large model and calculate the risk assessment index. The calculation formula of the risk assessment index is:
[0057] ;
[0058] in and Respectively context features and The weight of the constructed features, and Based on the large model context features and A prediction function that constructs features, The input bill data;
[0059] S8. Result interpretation step: Use explainable artificial intelligence (XAI) technology to explain and illustrate the risk assessment results. Set the clarity index of the explanation as .
[0060] A bill risk data analysis device based on a large model and context information, comprising:
[0061] Data collection module: used to collect bill-related data;
[0062] Data cleaning module: cleans and filters the collected data;
[0063] Feature Engineering Module: Construct effective features from cleaned data;
[0064] Context information extraction module: extracts context information from bill data;
[0065] Large model training module: Use the extracted context information and constructed features to train the large model;
[0066] Model optimization module: uses early stopping and other techniques to optimize the model training process;
[0067] Risk assessment module: used to input the bill data to be analyzed into the trained large model and calculate the risk assessment index;
[0068] Result interpretation module: Use explainable artificial intelligence technology to explain the evaluation results.
[0069] In the present invention, in the data collection step, a distributed crawler technology is used, and in the data cleaning step, a cleaning method based on a combination of rules and machine learning is applied.
[0070] In the present invention, in the feature engineering step, principal component analysis PCA is introduced to perform feature dimensionality reduction and reduce feature redundancy, and in the context information extraction step, graph convolutional neural network GCN is used.
[0071] In the present invention, in the large model training step, a mixed precision training strategy is adopted and a self-supervised learning task is introduced to enhance the representation ability of the model.
[0072] In the present invention, the data cleaning module uses the Hadoop or Spark distributed computing framework, distributes the cleaning tasks to multiple computing nodes for parallel processing, and adopts multiple data verification and error correction mechanisms.
[0073] In the present invention, the model optimization module adopts a variant optimization algorithm of Adagrad or Adadelta or RMSProp stochastic gradient descent SGD to adaptively adjust the learning rate according to the training history of the model, improve the convergence speed and stability of the model, and avoid falling into a local optimal solution.
[0074] In the present invention, in the distributed crawler technology, an intelligent scheduling algorithm is adopted to dynamically allocate crawler resources according to the priority of the data source, the data update frequency, and the network status factors; in the feature engineering step, the generative adversarial network GAN or the autoencoder AE automatic generation technology is adopted to automatically mine valuable features.
[0075] In the present invention, in principal component analysis, a principal component analysis method based on a Gaussian kernel function or a polynomial kernel function is used to process nonlinear data structures and extract principal components; when using a graph convolutional neural network, a multi-scale graph convolution operation is defined to capture information of different levels and granularities by performing convolution operations on the graph structure at different scales.
[0076] In the present invention, in the mixed precision training strategy, an adaptive threshold adjustment mechanism for precision switching is set to automatically adjust the threshold of precision switching according to the training progress of the model, data distribution, and changes in computing resources.
[0077] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A bill risk data analysis method based on a large model and context information, characterized in that: The following steps are involved: S1. Data collection: Collect basic information of bills, transaction records, related party information, and bill endorsement details. Suppose the total amount of data collected is ; S2. Data cleaning: Clean the collected data to remove duplicate, erroneous and incomplete data. Suppose the data after cleaning is effective. ; S3. Feature engineering: construct features from the cleaned data. The features include statistical features of bill amount, transaction frequency features, and credit features of related parties. Suppose the number of constructed features is ; S4. Context information extraction: Extract context information from bill data, including time series characteristics of transactions, geographical distribution characteristics of transaction locations, and topological characteristics of the relationship network of related parties. Suppose the number of context features is ; S5. Large model training: Use the extracted context information and constructed features to train a large model based on deep learning. Suppose the loss function of the model training is: in is the sample size, It is a real risk label. is the model's predicted value, is the mean square error, which is used to measure the difference between the predicted value and the true value. is the L1 regularization term, which is used to control the complexity of the model and avoid overfitting. is the regularization parameter, is the L2 regularization term, is the regularization parameter, is the KL divergence term used to deal with differences in the distribution of contextual information, is the weight parameter, and They are the true distribution of context features and the model predicted distribution; S6. Model optimization: Use early stopping method to prevent overfitting, and set the patience value of early stopping to , dynamically adjust the training process based on the performance of the validation set; S7. Risk assessment: Input the bill data to be analyzed into the trained large model and calculate the risk assessment index. The calculation formula of the risk assessment index is: in and Respectively context features and The weight of the constructed features, and Based on the large model context features and A prediction function that constructs features, The input bill data; S8. Result interpretation step: Use explainable artificial intelligence (XAI) technology to explain and illustrate the risk assessment results. Set the clarity index of the explanation as .
2. The method for analyzing bill risk data based on a large model and contextual information according to claim 1 is characterized by: In the data collection step, distributed crawler technology is used, and in the data cleaning step, a cleaning method combining rule-based and machine learning is applied.
3. The method for analyzing bill risk data based on a large model and contextual information according to claim 2 is characterized by: In distributed crawler technology, an intelligent scheduling algorithm is used to dynamically allocate crawler resources based on the priority of the data source, data update frequency, and network status factors; in the feature engineering step, the generative adversarial network GAN or autoencoder AE automatic generation technology is used to automatically mine valuable features.
4. The bill risk data analysis method based on a large model and contextual information according to claim 1 is characterized by: In the feature engineering step, principal component analysis (PCA) is introduced to perform feature dimensionality reduction and reduce feature redundancy. In the context information extraction step, graph convolutional neural network (GCN) is used.
5. The method for analyzing bill risk data based on a large model and contextual information according to claim 4 is characterized by: In principal component analysis, the principal component analysis method based on Gaussian kernel function or polynomial kernel function is used to process nonlinear data structures and extract principal components. When using graph convolutional neural networks, multi-scale graph convolution operations are defined to capture information at different levels and granularities by performing convolution operations on graph structures at different scales.
6. The method for analyzing bill risk data based on a large model and contextual information according to claim 1 is characterized by: In the large model training step, a mixed precision training strategy is adopted and a self-supervised learning task is introduced to enhance the representation ability of the model.
7. The method for analyzing bill risk data based on a large model and contextual information according to claim 6 is characterized by: In the mixed precision training strategy, an adaptive threshold adjustment mechanism for precision switching is set to automatically adjust the threshold for precision switching according to the model's training progress, data distribution, and changes in computing resources.
8. A bill risk data analysis device based on a large model and context information, characterized in that: include: Data collection module: used to collect bill-related data; Data cleaning module: cleans and filters the collected data; Feature Engineering Module: Construct effective features from cleaned data; Context information extraction module: extracts context information from bill data; Large model training module: Use the extracted context information and constructed features to train the large model; Model optimization module: uses early stopping technology to optimize the model training process; Risk assessment module: used to input the bill data to be analyzed into the trained large model and calculate the risk assessment index; Result interpretation module: Use explainable artificial intelligence technology to explain the evaluation results; The device is used to apply the bill risk data analysis method based on a large model and context information as described in claim 1.
9. The bill risk data analysis device based on a large model and context information according to claim 8 is characterized by: The data cleaning module uses Hadoop or Spark distributed computing framework, distributes cleaning tasks to multiple computing nodes for parallel processing, and adopts multiple data verification and error correction mechanisms.
10. The bill risk data analysis device based on a large model and context information according to claim 8, characterized in that: The model optimization module uses a variant optimization algorithm of Adagrad or Adadelta or RMSProp stochastic gradient descent SGD to adaptively adjust the learning rate according to the training history of the model, improve the convergence speed and stability of the model, and avoid falling into the local optimal solution.