Information evaluation method and device, electronic equipment and storage medium

Abstract

The invention discloses an information evaluation method and device, electronic equipment and a storage medium, and relates to the field of financial credit evaluation, and the specific implementation scheme is as follows: obtaining financial business data of a user, and determining feature data of the financial business data; determining enhanced feature data based on the integrated learning model and the feature data; processing the enhanced feature data according to an attention model to obtain attention feature data; and determining risk assessment information of the user according to the financial risk assessment model and the attention feature data. In the embodiment of the invention, the generation of the feature data is enhanced, so that the sensitivity of the financial risk assessment model to data fluctuation is reduced, and the accuracy of risk identification is enhanced; by introducing the attention model to focus on the key information in the enhanced feature data, the attention feature data is extracted, and the extracted attention feature data is input into the financial risk assessment model, so that the risk assessment precision and the assessment efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of financial credit evaluation, and in particular to an information evaluation method, device, electronic equipment and storage medium. Background Art

[0002] In recent years, small and micro inclusive loans have become one of the core forms of Internet finance advocated by the country. Small and micro inclusive loans have promoted the prosperity and development of the market, but the default behavior of borrowers has caused significant financial losses to banks and other lending platforms, and also disrupted the orderly development of the market. In the credit assessment process, the traditional risk information assessment model mainly focuses on the borrower's target characteristics, which leads to the limited prediction ability of the traditional risk information assessment model and makes it difficult to identify high-risk borrowers who accumulate credit through small loans in the early stage to commit large-scale fraud. In addition, the data labeling and feature completion process before model prediction is cumbersome and lacks credibility, making it difficult to effectively reveal the borrower's true intention to default. Therefore, enhancing the accuracy of risk identification and improving the accuracy and efficiency of risk assessment have become urgent issues to be solved in the field of financial credit assessment. Summary of the invention

[0003] The present invention provides an information assessment method, device, electronic equipment and storage medium, which enhance the accuracy of risk identification and improve the efficiency of risk assessment.

[0004] In one aspect of an embodiment of the present invention, there is provided an information evaluation method, comprising:

[0005] Acquire the user's financial service data and determine characteristic data of the financial service data;

[0006] Determine enhanced feature data based on the integrated learning model and the feature data;

[0007] Processing the enhanced feature data according to the attention model to obtain attention feature data;

[0008] The risk assessment information of the user is determined based on the financial risk assessment model and the attention feature data.

[0009] Another aspect of an embodiment of the present invention provides an information evaluation device, including:

[0010] A feature extraction module, used to obtain the user's financial service data and determine feature data of the financial service data;

[0011] A feature enhancement module, used for determining enhanced feature data based on the integrated learning model and the feature data;

[0012] An attention module, used for processing the enhanced feature data according to an attention model to obtain attention feature data;

[0013] An assessment information module is used to determine the risk assessment information of the user based on the financial risk assessment model and the attention feature data.

[0014] Another aspect of an embodiment of the present invention provides an electronic device, including:

[0015] at least one processor;

[0016] and a memory communicatively coupled to the at least one processor;

[0017] The memory stores a computer program that can be executed by at least one processor, and the computer program is executed by at least one processor so that the at least one processor can execute any one of the information evaluation methods in the embodiments of the present invention.

[0018] According to another aspect of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the information evaluation method of any embodiment of the present invention when executed.

[0019] An embodiment of the present invention collects and records financial business data directly related to financial activities of users, and extracts feature data that plays a key role in assessing potential risks from the financial business data, obtains an integrated learning model that has been trained, inputs the feature data of the user into the integrated learning model that has been trained, the integrated learning model processes the feature data, and outputs enhanced feature data that is more expressive and discriminative, applies an attention model to process the enhanced feature data, and outputs attention feature data, obtains a financial risk assessment model that has been trained, inputs the attention feature data into the financial risk assessment model, the financial risk assessment model processes and analyzes the attention feature data, outputs risk assessment information containing various attributes of potential risks, and performs subsequent operations based on the risk assessment information. In an embodiment of the present invention, by processing financial business data, feature data that plays a key role in evaluating user information is extracted from the financial business data, so that the financial risk assessment model only processes the feature data in the financial business data during the assessment process, reducing the amount of data processed by the model, thereby improving the efficiency of risk assessment; the integrated learning model can enhance the expressiveness and discrimination of the input data, and by inputting the feature data into the integrated learning model, enhanced feature data with more expressiveness and discrimination can be obtained. The generation of enhanced feature data reduces the sensitivity of the financial risk assessment model to data fluctuations, so that the financial risk assessment model can more accurately identify potential risks, thereby enhancing the accuracy of risk identification; by introducing an attention model to focus on key information in the enhanced feature data, attention feature data is extracted from the enhanced feature data, and the extracted attention feature data is input into the financial risk assessment model, so that the financial risk assessment model can more accurately and efficiently assess the potential risks in the user information, thereby improving the assessment accuracy and assessment efficiency of the financial risk assessment model.

[0020] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 is a flow chart of an information evaluation method provided according to Embodiment 1 of the present invention;

[0023] Figure 2is a flow chart of another information evaluation method provided according to Embodiment 2 of the present invention;

[0024] Figure 3 This is a schematic diagram of a feature combination generation method provided in Embodiment 3 of the present invention;

[0025] Figure 4 This is a schematic diagram of a characteristic combination provided according to Embodiment 3 of the present invention;

[0026] Figure 5 is a schematic diagram of a training data combination provided according to Embodiment 3 of the present invention;

[0027] Figure 6 is a schematic diagram of a Transformer model architecture provided according to Embodiment 3 of the present invention;

[0028] Figure 7 It is a schematic diagram of the structure of a device for implementing an information evaluation method provided by an embodiment of the present invention;

[0029] Figure 8 It is a schematic diagram of the structure of an electronic device for implementing the information evaluation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] Embodiment 1

[0033] Figure 1The present invention provides a flowchart of an information evaluation method. This embodiment is applicable to the situation of evaluating financial credit loan risk information. The method can be executed by an information evaluation device. The information evaluation device can be implemented in the form of hardware and / or software. The information evaluation device can be configured in an electronic device. Figure 1 As shown, the method includes:

[0034] S101. Obtain the user's financial service data and determine characteristic data of the financial service data.

[0035] In the embodiment of the present invention, financial business data may be understood as information directly related to financial activities. For example, financial business data may include: financial status information or transaction records, etc.

[0036] Feature data can be understood as key information extracted from financial business data. Extracting feature data from financial business data helps financial risk assessment models predict potential risks more efficiently. For example, feature data may include: loan amount or real estate information.

[0037] Specifically, collect and record the user's financial business data directly related to financial activities, and extract characteristic data that plays a key role in assessing potential risks from the financial business data. The methods of obtaining user financial business data may include: obtaining through application programming interface collection or crawling user financial business data from relevant web pages using general web crawler technology; the methods of extracting characteristic data from financial business data may include: data cleaning methods or machine learning methods, etc.

[0038] S102: Determine enhanced feature data based on the integrated learning model and the feature data.

[0039] In the embodiment of the present invention, the ensemble learning model can be understood as a strong learner, which is formed by combining multiple weak learners. The ensemble learning model uses an ensemble learning algorithm to process input data. For example, the ensemble learning algorithm includes but is not limited to: Bagging algorithm or Boosting algorithm.

[0040] Enhanced feature data can be understood as a set of new features obtained after processing feature data. Enhanced feature data is more expressive and discriminative than feature data.

[0041] Specifically, an integrated learning model that has been trained is obtained, and the user's feature data is obtained. The user's feature data is input into the integrated learning model that has been trained. The integrated learning model processes the feature data and outputs enhanced feature data that is more expressive and discriminative.

[0042] S103. Process the enhanced feature data according to the attention model to obtain attention feature data.

[0043] In an embodiment of the present invention, the attention model can be understood as a machine learning model with an attention mechanism added. The attention model can dynamically focus on important parts of the input data to improve the prediction performance of the model. For example, the form of the attention mechanism may include: self-attention mechanism or multi-head attention mechanism, etc.

[0044] The attention feature data can be understood as the result data obtained after applying the attention model to process the enhanced feature data. The attention feature data highlights the important parts of the enhanced feature data and helps to improve the accuracy of risk prediction.

[0045] Specifically, the enhanced feature data output by the integrated learning model is input into the attention model with the attention mechanism added. The attention model processes the input enhanced feature data and outputs the attention feature data.

[0046] For example, the step of processing the enhanced feature data by the attention model may include: the attention model assigns an attention weight to each feature in the enhanced feature data according to the importance of each feature in the enhanced feature data, and the attention model performs weighted summation or weighted averaging on the enhanced feature data according to the attention weight to obtain the final attention feature data.

[0047] S104: Determine the user's risk assessment information based on the financial risk assessment model and attention feature data.

[0048] In an embodiment of the present invention, the financial risk assessment model can be understood as a tool for quantifying and managing financial risks. For example, the financial risk assessment model can be based on theories such as statistics, mathematics or economics, and use various risk measurement methods to conduct risk assessment on the user's financial business data.

[0049] The risk assessment information may be understood as a comprehensive data set, and the risk assessment information includes various attributes of potential risks. For example, the various attributes of potential risks included in the risk assessment information may include: risk level attributes or risk management suggestion attributes.

[0050] Specifically, a trained financial risk assessment model is obtained, and the attention feature data is input into the financial risk assessment model. The financial risk assessment model processes and analyzes the attention feature data, and outputs risk assessment information containing various attributes of potential risks. Subsequent operations can be performed based on the risk assessment information. For example, the subsequent operations may include: sending the determined risk assessment information to relevant personnel or relevant equipment, and the relevant personnel or relevant equipment take corresponding measures to reduce or eliminate the risk based on the risk assessment information. It is understandable that the architecture and parameters of the financial risk assessment model can be designed according to business needs and the characteristics of the attention feature data.

[0051] An embodiment of the present invention collects and records financial business data directly related to financial activities of users, and extracts feature data that plays a key role in assessing potential risks from the financial business data, obtains an integrated learning model that has been trained, inputs the feature data of the user into the integrated learning model that has been trained, the integrated learning model processes the feature data, and outputs enhanced feature data that is more expressive and discriminative, applies an attention model to process the enhanced feature data, and outputs attention feature data, obtains a financial risk assessment model that has been trained, inputs the attention feature data into the financial risk assessment model, the financial risk assessment model processes and analyzes the attention feature data, outputs risk assessment information containing various attributes of potential risks, and performs subsequent operations based on the risk assessment information. In an embodiment of the present invention, by processing financial business data, feature data that plays a key role in evaluating user information is extracted from the financial business data, so that the financial risk assessment model only processes the feature data in the financial business data during the assessment process, reducing the amount of data processed by the model, thereby improving the efficiency of risk assessment; the integrated learning model can enhance the expressiveness and discrimination of the input data, and by inputting the feature data into the integrated learning model, enhanced feature data with more expressiveness and discrimination can be obtained. The generation of enhanced feature data reduces the sensitivity of the financial risk assessment model to data fluctuations, so that the financial risk assessment model can more accurately identify potential risks, thereby enhancing the accuracy of risk identification; by introducing an attention model to focus on key information in the enhanced feature data, attention feature data is extracted from the enhanced feature data, and the extracted attention feature data is input into the financial risk assessment model, so that the financial risk assessment model can more accurately and efficiently assess the potential risks in the user information, thereby improving the assessment accuracy and assessment efficiency of the financial risk assessment model.

[0052] On the basis of the above-mentioned embodiment, the enhanced feature data is processed according to the attention model, and before the attention feature data is obtained, the method further includes: splitting and combining the enhanced feature data according to the dimensions.

[0053] Specifically, when splitting and combining enhanced feature data, it is necessary to pay attention to the dimension of the enhanced feature data, decompose the enhanced feature data into features of multiple dimensions, and recombine these features according to certain rules after the enhanced feature data is decomposed to form new combined features. Through reasonable splitting and combination, the model can better capture the complex relationships in the data and ultimately improve the overall prediction effect.

[0054] Embodiment 2

[0055] Figure 2 This is a flow chart of another information evaluation method provided by an embodiment of the present invention. Based on the above embodiment, the embodiment of the present invention provides another information evaluation method. Figure 2 As shown, the method includes:

[0056] S201. Obtain the user's financial service data and determine characteristic data of the financial service data.

[0057] S202. Input the feature data into a gradient boosted decision tree model to obtain enhanced feature data, wherein the gradient boosted decision tree model is generated based on training of a training set, and the training set at least includes user financial business data with a positive risk assessment label.

[0058] In an embodiment of the present invention, the gradient boosting decision tree model refers to an integrated learning algorithm model, and the gradient boosting decision tree model applies a gradient boosting decision tree algorithm to process feature data.

[0059] The training set can be understood as a data set used to train a gradient enhanced decision tree model. The training set at least includes the user's financial business data, wherein each data sample in the financial business data is labeled with a corresponding risk assessment label. The risk assessment label can be understood as an indicator for measuring the level of risk. For example, the risk assessment label at least includes: a positive risk assessment label and a negative risk assessment label. The positive risk assessment label can be understood as an indicator indicating that the user's financial business data has a low risk.

[0060] Specifically, feature data extracted from user financial business data is obtained, and the obtained feature data is input into a gradient boosting decision tree model, and the gradient boosting decision tree model processes and analyzes the feature data and outputs enhanced feature data. For example, the training process of the gradient boosting decision tree model at least includes: dividing the feature data into a training set and a test set, and using the training set to train the gradient boosting decision tree model, wherein the training set at least includes financial business data labeled as a positive risk assessment label.

[0061] For example, the step of processing and analyzing the feature data by the gradient enhanced decision tree model may include: the gradient enhanced decision tree model performs nonlinear combination and transformation on the feature data based on the learned decision tree structure, and uses the feature data after the nonlinear combination and transformation as enhanced feature data.

[0062] S203: Input the enhanced feature data into the sequence embedding layer of the attention model for dimension upgrading to obtain an embedding vector.

[0063] In the embodiment of the present invention, the sequence embedding layer is a component of the attention model, which can be understood as a mechanism of feature transformation and dimensionality enhancement, which is used to map the input data to a higher dimensional vector space.

[0064] The embedding vector is the output of the sequence embedding layer and can be understood as a high-dimensional vector that contains the semantic relationship of all elements in the enhanced feature data.

[0065] Specifically, the enhanced feature data output by the gradient enhanced decision tree model is obtained, and the enhanced feature data is input into the attention model. The sequence embedding layer in the attention model processes and analyzes the enhanced feature data and outputs an embedding vector.

[0066] For example, the step of processing and analyzing the enhanced feature data by the sequence embedding layer may include: the sequence embedding layer may perform a dimensionality-increasing operation on the enhanced feature data through a corresponding mapping function to obtain a high-dimensional vector of the enhanced feature data, and use the high-dimensional vector as an embedding vector.

[0067] For example, the step of increasing the dimension of the enhanced feature data by using a mapping function at least includes: converting the enhanced feature data into a vector of a fixed length and increasing the dimension of the fixed length vector.

[0068] S204: Input the embedded vector into the position encoding layer of the attention model for encoding to obtain a position encoding vector.

[0069] In an embodiment of the present invention, the position encoding layer is another component of the attention model, which can be understood as a mechanism for adding position information to the embedding vector. The position information is added to the embedding vector so that the attention model can capture the sequential relationship of elements in the data, thereby improving the accuracy of the attention model.

[0070] The position encoding vector is the vector output by the position encoding layer. Adding position information to the embedding vector allows the attention model to better understand the contextual relationship of elements in the enhanced feature data.

[0071] Specifically, when the embedding vector output by the sequence embedding layer is input to the position encoding layer, the position encoding layer can encode the embedding vector through a mathematical function, incorporate the position information into each embedding vector, and generate a position encoding vector.

[0072] S205. Input the position encoding vector into the multi-head attention layer of the attention model, and obtain the attention vector by setting attention weights for different positions in the position encoding vector in the multi-head attention layer.

[0073] In an embodiment of the present invention, the multi-head attention layer is another component of the attention model, which can be understood as an attention mechanism. In the multi-head attention layer, the output of each attention head is concatenated or weighted summed to form the final output.

[0074] The attention weight can be understood as a numerical value that measures the importance of different position information in the position encoding vector.

[0075] The attention vector is the output vector of the multi-head attention layer, which can reflect the importance of different positions in the position encoding vector.

[0076] Specifically, when the position encoding vector is input to the multi-head attention layer, the multi-head attention layer uses the multi-head attention mechanism to set different attention weights for different positions in the position encoding vector according to the importance of different positions in the position encoding vector. The outputs of multiple attention heads can be concatenated or weighted summed to form an attention vector.

[0077] S206. Input the attention vector into the linear and regression layers of the attention model to obtain attention feature data.

[0078] In the embodiment of the present invention, the linear and regression layers are components of the attention model, wherein in the linear and regression layers, operations such as linear transformation are mainly performed on the input data.

[0079] Attention feature data is the output data of the linear and regression layers. In risk assessment, attention feature data is used as the input of the financial risk assessment model for subsequent risk prediction.

[0080] Specifically, the attention vector is input into the linear and regression layers, where the attention vector is processed and analyzed to obtain the attention feature data.

[0081] For example, the steps of processing and analyzing the attention vector in the linear and regression layers include at least: linear transformation, wherein the specific steps of the linear transformation may include: applying a weight matrix and a bias vector, performing operations such as feature extraction or feature combination on the attention vector, thereby obtaining attention feature data.

[0082] S207. Determine the user's risk assessment information based on the financial risk assessment model and the attention feature data, wherein the financial risk assessment model includes at least one of the following: a convolutional neural network model and a decision tree model.

[0083] In an embodiment of the present invention, a convolutional neural network model refers to a deep learning model, and the convolutional neural network model extracts features of input data through a convolution operation.

[0084] A decision tree model is a tree-structured model used for data classification and regression. The decision tree model divides data through a series of decision rules.

[0085] Specifically, a trained financial risk assessment model is obtained, and the attention feature data output by the linear and regression layers is used as the input of the financial risk assessment model, and is input into the financial risk assessment model. The financial risk assessment model processes and analyzes the attention feature data and outputs risk assessment information. It is understandable that the architecture and parameters of the financial risk assessment model can be designed according to business needs and the characteristics of the attention feature data, wherein the architecture of the financial risk assessment model includes at least two types: a convolutional neural network model architecture and a decision tree model architecture.

[0086] In an embodiment of the present invention, financial business data of a user is obtained, and feature data of the financial business data is determined, the obtained feature data is input into a gradient boosting decision tree model, the gradient boosting decision tree model processes and analyzes the feature data and outputs enhanced feature data, the enhanced feature data is input into an attention model, the sequence embedding layer in the attention model processes and analyzes the enhanced feature data and outputs an embedding vector, when the embedding vector output by the sequence embedding layer is input into a position encoding layer, the position encoding layer can encode the embedding vector through a mathematical function, incorporate position information into each embedding vector, generate a position encoding vector, the attention vector is input into a linear and regression layer, in the linear and regression layer, the attention vector is processed and analyzed to obtain attention feature data, the attention feature data output by the linear and regression layer is input into a financial risk assessment model, the financial risk assessment model processes and analyzes the attention feature data and outputs risk assessment information. In an embodiment of the present invention, deep mining and risk assessment of user financial business data are achieved by using three models, namely, a gradient enhanced decision tree model, an attention model and a financial risk assessment model. The feature data of the financial business data is processed and analyzed by the gradient enhanced decision tree model, key features are effectively extracted and enhanced feature data is generated, thereby enhancing the ability to understand the data and the prediction accuracy. The enhanced feature data is processed by the sequence embedding layer in the attention model, so that the output embedding vector contains the semantic relationship of all elements in the enhanced feature data. According to the semantic relationship of all elements, the connection between features in the enhanced feature data can be deeply understood, providing more comprehensive data support for subsequent risk assessment. The embedding vector is encoded by the position encoding layer, and the position information is integrated into each embedding vector to generate a position encoding vector, so that the model can capture the position dependence in the sequence data, improve the sensitivity of risk assessment to the order of elements, and improve the assessment accuracy of the financial risk assessment model. By introducing the attention mechanism, the attention model can automatically focus on the most critical features for risk assessment. The attention feature data is input into the financial risk assessment model, so that the model can more accurately and efficiently assess financial risks, providing strong decision support for financial institutions.

[0087] On the basis of the above embodiments, before determining the enhanced feature data based on the integrated learning model and the feature data, the embodiment of the present invention also includes: obtaining a training set of the gradient enhanced decision tree model, and determining the result prediction value according to the training set and the result prediction formula; constructing a decision tree of the training set, and determining the negative gradient of the decision tree for each sample data in the training set according to the result prediction formula; substituting each negative gradient into the result prediction formula to obtain a decision prediction result; determining a loss value between the result prediction value and the decision prediction result, and when the loss value meets the preset conditions, determining that the gradient enhanced decision tree training is completed, otherwise the decision tree is reconstructed.

[0088] In an embodiment of the present invention, the result prediction formula can be understood as a mathematical expression for converting input feature data into a result prediction value, wherein the result prediction value can be understood as a specific numerical value calculated by a gradient enhanced decision tree model according to the result prediction formula. For example, in the context of financial risk assessment, the result prediction value may include: a user's risk score value or a user's default probability value.

[0089] Sample data refers to every data record in the training set. Sample data is the basis for model training and evaluation. By analyzing sample data, the model can learn effective prediction rules.

[0090] The negative gradient can be understood as a direction vector. In the process of building a decision tree, the negative gradient is used to guide the learning direction of the new decision tree to minimize the loss function.

[0091] The decision prediction result refers to the final output result calculated based on the learned prediction rule and negative gradient. For example, the decision prediction result may be expressed in the form of a category label or a numerical value.

[0092] The loss value can be understood as the difference between the result prediction value and the decision prediction result, which is used to compare with the preset conditions to determine whether the gradient enhancement decision tree is trained.

[0093] The preset condition can be understood as a predefined criterion for judging whether the training of the gradient boosting decision tree is complete.

[0094] Specifically, feature data of financial business data is obtained, and the feature data of the financial business data is divided into two types of data, namely, a training set and a test set. The gradient enhanced decision tree model is trained based on the training set, and a result prediction formula for converting input features into result prediction values ​​is obtained. During the training process, the result prediction formula is used to calculate the result prediction value. A decision tree is constructed based on the training set in the feature data, and the result prediction formula is used to determine the negative gradient of each sample data. The calculated negative gradient is substituted into the result prediction formula to obtain the decision prediction result of each sample. The difference between the result prediction value and the decision prediction result is calculated to obtain a loss value, and the loss value is compared with a preset condition. If the loss value meets the preset condition, it means that the constructed decision tree reaches a preset performance standard, and it is considered that the training of the gradient enhanced decision tree model has been completed. Otherwise, the decision tree is reconstructed, and the iterative training process continues until the model performance reaches the preset performance standard.

[0095] Based on the above embodiments, the attention model in the embodiment of the present invention is a multi-head attention model, and the loss function used in the training process of the multi-head attention model is a cross entropy loss function.

[0096] In an embodiment of the present invention, the multi-head attention model can be understood as an extended attention mechanism. The multi-head attention model includes multiple attention heads, each of which can focus on different parts of the input data, thereby improving the performance of the model on specific tasks.

[0097] The loss function can be understood as a function used to measure the gap between the model's predicted results and the actual results. During the model training process, the model optimizes its performance by minimizing the loss function.

[0098] Specifically, in an embodiment of the present invention, a multi-head attention model can be used to process and analyze enhanced feature data, and the multi-head attention model can be trained based on a cross-entropy loss function.

[0099] Embodiment 3

[0100] The embodiment of the present invention provides an information evaluation model based on the above embodiment mode, and the information evaluation model consists of two parts: Gradient Boosting Decision Tree (GBDT) and Transformer model.

[0101] The GBDT algorithm is an iterative decision tree algorithm. The GBDT algorithm constructs a set of weak learning trees and accumulates the results of multiple decision trees as the final prediction output. The GBDT algorithm continuously creates a new decision tree in the gradient direction of the residual reduction in an iterative manner to improve the accuracy of the prediction. In the embodiment of the present invention, H = {(x 1 ,y 1 ), (x 2 ,y 2 ), …(x |H| ,y |H| )} represents the target data set of small and micro inclusive loans, where x i Represents the i-th sample data with n characteristic indicators, y i Represents its corresponding label data, y i indicates whether the lender defaults, and |H| represents the number of data sets.

[0102] Set H = {([3800, 0] T ,0),([5000,1] T ,0),([20000,1] T ,0),([32000,0] T , 1)}, then x 1 =[3800,0] T The first feature item in represents the loan amount, and the second feature item represents whether there is a house, where 1 means there is a house and 0 means there is no house; 1=0 means there is no default on the loan, y 1 =1 indicates a default. Now we construct a GBDT algorithm tree based on this data. Here we define the number of trees m = 2, the depth d = 2, and the learning rate lr = 1. The tree construction process is as follows: First, initialize the first tree. The specific initialization formula is Among them, F 0 (x) represents the prediction result of the first tree, and P(Y=1|x) represents the probability that the labeled data in the training set is normal performance, that is, the proportion of all data. Substituting the above data into the formula, we can get F 0 The initial value of (x) is -0.477. The above data are summarized in the following table, and the sample numbers are set to 1, 2, 3 and 4 respectively.

[0103] serial number Loan Amount Whether there is real estate Whether it is a breach of contract <![CDATA[F 0 (x)]]> 1 3800 0 0 -0.477 2 5000 1 0 -0.477 3 20000 1 0 -0.477 4 32000 0 1 -0.477

[0104] Calculate the negative gradient that needs to be fitted for the mth tree Among them, y i Indicates the real label data, that is, whether the contract is fulfilled normally, and normal fulfillment is y i =1, default is y i = 0. Substituting the above data into the formula, we can get the negative gradient value, as shown in the following table:

[0105] serial number Loan Amount Whether there is real estate Whether it is a breach of contract Negative Gradient 1 3800 0 0 -0.383 2 5000 1 0 -0.383 3 20000 1 0 -0.383 4 32000 0 1 0.617

[0106] Find the best partition point of the regression tree and traverse the value of each feature to get the square loss of the left node and the squared loss of the right node Add the two together to get the total square loss value S under this division method. Sum =S L +S R , where Size L Indicates the number of data divided into the left node, r m，avgL Represents the average negative gradient of the data divided into the left node, Size R Indicates the number of data divided into the right node, r m，avgR It represents the average value of the negative gradient of the data divided into the right node. Exhaustively divide the nodes and substitute the data into the above formula to get the following table:

[0107]

[0108] Calculate the best fit value, the specific calculation formula Among them, c m，j represents the fitted value of the jth leaf node of the mth subtree. Substituting the data into the calculation, we can get the fitted value c of the first leaf node of the first tree. 1，1 =-1.621, the fitted value of the second leaf node of the first tree is c1，2 =2.611.

[0109] Establish the mth regression tree for classification, the calculation formula is where F m (x) represents the prediction result of the mth tree, F m-1 (x) represents the prediction result of the m-1th tree, lr represents the learning rate, J m represents the number of leaf nodes, I(x∈R m，j ) indicates whether the prediction result falls on the current leaf node, if so, it is 1, otherwise it is 0. Substituting the data into the calculation, the F of each data can be obtained. 1 (x) As shown in the following table:

[0110] serial number Loan Amount Whether there is real estate Whether it is a breach of contract <![CDATA[F 1 (x)]]> 1 3800 0 0 -2.098 2 5000 1 0 -2.098 3 20000 1 0 -2.098 4 32000 0 1 2.134

[0111] Repeat the above steps until m trees are built. The predicted results falling on the leaf nodes of each tree are set to 1, and the other parts are set to 0. The values ​​of all leaf nodes are combined to obtain the combined features enhanced by the GBDT algorithm. The specific method of generating the combined features is as follows: Figure 3 As shown, in Figure 3 In the figure, the leaf node where the prediction result falls is marked in black, and its combined feature is [1, 0, 0, 0, 0, 0, 1, …, 0, 0, 1, 0].

[0112] Data H = {([3800, 0] T ,0),([5000,1] T ,0),([20000,1] T ,0),([32000,0] T , 1)} The feature combination process is as follows Figure 4 As shown in the figure, the sample data of the decision tree on the right is divided by whether there is a property or not. When the loan amount is 32,000, F 1 (x) is the largest, so the sample data of the decision tree on the left uses the loan amount = 32000 as the dividing point of the decision tree. Figure 4 It can be seen that the valid feature corresponding to the first feature data in data H is [1,0,0,1], the valid feature corresponding to the second feature data in data H is [1,0,1,0], the valid feature corresponding to the third feature data in data H is [1,0,1,0], and the valid feature corresponding to the fourth feature data in data H is [0,1,0,1]. Based on the above valid features, the combined feature of data H after GBDT algorithm is [1,0,0,1,1,0,1,0,0,1,0,1,1,0,1,0].

[0113] The Transformer model is used to predict the possibility of customer default, and the combined features output by the GBDT model are used as the input of the Transformer model. In order to expand the amount of data and enhance the availability of data training, the combined features can be preprocessed, the input data of the model can be split, and different training data inputs can be combined.

[0114] like Figure 5 As shown, the input data is Z = {[1, 0, 0, 1] T , [1, 0, 1, 0] T , [1, 0, 1, 0] T , [0, 1, 0, 1] T}, the corresponding label Y = {0, 0, 0, 1}, and the input data can also be expressed as Z = {z 1 , z 2 , z 3 , z 4}, if the number of data is 1, the Z 1 ={z 1}, Z 2 ={z 2}, Z 3 ={z 3}, Z 4 ={z 4}, corresponding to label Y 1 ={y 1},Y 2 ={y 2},Y 3 ={y 3},Y 4 = {Y 4}, Z can be generated by setting the number of data to 2 5 ={z 1 , z 2},Y 5 ={y 1 ,y 2}, and so on, and the number of other combinations is similar, where z i Represents the feature vector after GBDT tree enhancement.

[0115] The architecture of the Transformer model is as follows Figure 6 As shown, it includes sequence embedding (Embedding) layer, position encoding (Position Embedding) layer, encoding (Encoder) part, decoding (Decoder) part, linear and regression (Linear&SoftMax) layer.

[0116] In the sequence embedding layer, the input data Z is upgraded in dimension, and E=Embedding(Z) is the output vector after the upgrade.

[0117] The position embedding vector is generated in the position encoding layer, using a fixed position encoding method. The calculation formula is as follows:

[0118]

[0119] Among them, pos represents the position of the first dimension of the data Z. In the position embedding matrix, the position of even numbers is calculated by sin, and the position of odd numbers is calculated by cos. Finally, the generated position embedding vector is added to the vector output by the sequence embedding layer to obtain the vector P embedded with the position information.

[0120] In the encoding part, the position embedding vector of the multi-head attention layer input is linearly transformed to produce three tensors Q i , K i and V i , where Q i =P (e) W i Q , K i =P (e) W i K , V i =P (e) W i V , calculate Q i With K i T In order to prevent the result from being too large, a scale is introduced where d k is the matrix K i The size of the second dimension is then normalized by softmax and multiplied by V i Get the sum of the weights: All heads i The final output G is generated by a linear projection transformation that can be learned. The calculation formula of G is G = MHA (Q, K, V) = [head 1 ,…head h ]·W o , where W i Q , W i K , W i V is the weight matrix of the i-th self-attention layer, W o is a linear projection transformation.

[0121] In the encoding part, the feed forward layer provides nonlinear transformation. The feed forward layer mainly consists of three layers. The first and third layers are linear transformation layers, the second layer is a nonlinear activation layer, and the activation function is Relu. Finally, after the residual connection, the feed forward layer outputs the result O (e) =Relu(GW 1 +b 1 )W 3 +b 3 +G, where W 1 , W 3 、b 1 and b 3 are the weight parameters and bias parameters of the first and third layers.

[0122] In the decoding part, the Masked Multi-Head Attention layer introduces a mask based on the multi-head attention layer in the encoding part to eliminate the interference of the input data after a certain moment when predicting the result at that moment. It is necessary to calculate the sum of the weights to represent the head i When adding the mask matrix M, the calculation formula is The difference between the Encoder-Decoder Multi-Head Attention layer and the Multi-Head Attention layer of the encoding part is that the input data comes not only from the information of the decoding part, but also from the output information of the encoding part.

[0123] In the decoding part, the feed forward layer is consistent with the corresponding layer in the encoding part, and the output is O (d) .

[0124] In the linear and regression layers, O (d) Input the fully connected layer and perform softmax normalization to get the output sequence Among them, W L is the weight matrix of the fully connected layer, For the prediction results.

[0125] The Transformer model is trained using the cross entropy loss function. The specific calculation formula of the cross entropy loss function is: Among them, the parameter settings of the Transformer model are shown in the following table:

[0126] name parameter MHA number 8 Number of encoders 6 Number of decoders 6 Number of neurons 512 Training Method SGD Learning Rate 1×

[0127] Embodiment 4

[0128] Figure 7Schematic diagram of the structure of an information evaluation device provided by an embodiment of the present invention. Figure 7 As shown, the device comprises:

[0129] A feature extraction module is used to obtain the user's financial service data and determine the feature data of the financial service data;

[0130] A feature enhancement module, used for determining enhanced feature data based on the integrated learning model and the feature data;

[0131] An attention module, used for processing the enhanced feature data according to the attention model to obtain attention feature data;

[0132] The assessment information module is used to determine the user's risk assessment information based on the financial risk assessment model and attention feature data.

[0133] An embodiment of the present invention collects and records financial business data directly related to financial activities of users, and extracts feature data that plays a key role in assessing potential risks from the financial business data, obtains an integrated learning model that has been trained, inputs the feature data of the user into the integrated learning model that has been trained, the integrated learning model processes the feature data, and outputs enhanced feature data that is more expressive and discriminative, applies an attention model to process the enhanced feature data, and outputs attention feature data, obtains a financial risk assessment model that has been trained, inputs the attention feature data into the financial risk assessment model, the financial risk assessment model processes and analyzes the attention feature data, outputs risk assessment information containing various attributes of potential risks, and performs subsequent operations based on the risk assessment information. In an embodiment of the present invention, by processing financial business data, feature data that plays a key role in evaluating user information is extracted from the financial business data, so that the financial risk assessment model only processes the feature data in the financial business data during the assessment process, reducing the amount of data processed by the model, thereby improving the efficiency of risk assessment; the integrated learning model can enhance the expressiveness and discrimination of the input data, and by inputting the feature data into the integrated learning model, enhanced feature data with more expressiveness and discrimination can be obtained. The generation of enhanced feature data reduces the sensitivity of the financial risk assessment model to data fluctuations, so that the financial risk assessment model can more accurately identify potential risks, thereby enhancing the accuracy of risk identification; by introducing an attention model to focus on key information in the enhanced feature data, attention feature data is extracted from the enhanced feature data, and the extracted attention feature data is input into the financial risk assessment model, so that the financial risk assessment model can more accurately and efficiently assess the potential risks in the user information, thereby improving the assessment accuracy and assessment efficiency of the financial risk assessment model.

[0134] On the basis of the above embodiments, before determining the enhanced feature data based on the integrated learning model and the feature data, the embodiment of the present invention also includes: obtaining a training set of the gradient enhanced decision tree model, and determining the result prediction value according to the training set and the result prediction formula; constructing a decision tree of the training set, and determining the negative gradient of the decision tree for each sample data in the training set according to the result prediction formula; substituting each negative gradient into the result prediction formula to obtain a decision prediction result; determining a loss value between the result prediction value and the decision prediction result, and when the loss value meets the preset conditions, determining that the gradient enhanced decision tree training is completed, otherwise the decision tree is reconstructed.

[0135] Based on the above embodiments, the attention model in the embodiment of the present invention is a multi-head attention model, and the loss function used in the training process of the multi-head attention model is a cross entropy loss function.

[0136] On the basis of the above-mentioned embodiment, in the embodiment of the present invention, the enhanced feature data is processed according to the attention model, and before the attention feature data is obtained, the enhanced feature data is also split and combined according to the dimensions.

[0137] Based on the above embodiments, the financial risk assessment model in the embodiments of the present invention includes at least one of the following: a convolutional neural network model and a decision tree model.

[0138] Based on the above embodiment, the feature enhancement module 402 is specifically used to: input the feature data into the gradient enhancement decision tree model respectively to obtain enhanced feature data, wherein the gradient enhancement decision tree model is generated based on training of a training set, and the training set at least includes user financial data with a positive risk assessment label.

[0139] Based on the above embodiments, the attention module 403 is specifically used to: input the enhanced feature data into the sequence embedding layer of the attention model for dimensionality upgrading to obtain an embedding vector; input the embedding vector into the position encoding layer of the attention model for encoding to obtain a position encoding vector; input the position encoding vector into the multi-head attention layer of the attention model, and set the attention weights for different positions in the position encoding vector in the multi-head attention layer to obtain the attention vector; input the attention vector into the linear and regression layers of the attention model to obtain attention feature data.

[0140] Embodiment 5

[0141] Figure 8A schematic diagram of an electronic device that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown in the embodiments of the present invention, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present invention described and / or required herein.

[0142] like Figure 8 As shown, the electronic device includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the ROM 12 or the computer program loaded from the storage unit 18 to the RAM 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. The input / output (I / O) interface 15 is also connected to the bus 14.

[0143] A number of components in the electronic device are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0144] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit, a graphics processing unit, various special artificial intelligence computing chips, various processors running machine learning model algorithms, digital signal processors, and any appropriate processors, controllers, microcontrollers, etc. The processor 11 performs the various methods and processes described above, for example, the information evaluation method.

[0145] In some embodiments, the information evaluation method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the information evaluation method may be performed. Alternatively, in other embodiments, the processor 11 may be configured as the information evaluation method in any other suitable manner (e.g., by means of firmware).

[0146] Various implementations of the systems and techniques described above in the embodiments of the present invention may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays, application specific integrated circuits, application specific standard products, systems on a chip, load programmable logic devices, computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: being implemented in one or more computer programs, which may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general programmable processor, which may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0147] The computer programs for implementing the methods of the embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer programs are executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer programs may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0148] In the context of an embodiment of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0149] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a cathode ray tube or a liquid crystal display monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including: acoustic input, voice input, or tactile input).

[0150] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local area networks, wide area networks, blockchain networks, and the Internet.

[0151] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and virtual private server services.

[0152] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0153] The above specific implementations do not constitute a limitation on the protection scope of the embodiments of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An information evaluation method, characterized in that: The method comprises: Acquire the user's financial service data and determine characteristic data of the financial service data; Determine enhanced feature data based on the integrated learning model and the feature data; Processing the enhanced feature data according to the attention model to obtain attention feature data; The risk assessment information of the user is determined based on the financial risk assessment model and the attention feature data.

2. The method according to claim 1, characterized in that: The integrated learning model includes a gradient enhanced decision tree model, and the step of determining enhanced feature data based on the integrated learning model and the feature data includes: Inputting the feature data into the gradient enhanced decision tree model respectively to obtain the enhanced feature data; The gradient enhanced decision tree model is generated based on training of a training set, and the training set at least includes user financial business data with a positive risk assessment label.

3. The method according to claim 2, characterized in that: Before determining the enhanced feature data based on the integrated learning model and the feature data, the method further includes: Obtaining a training set of the gradient enhanced decision tree model, and determining a result prediction value according to the training set and a result prediction formula; Constructing a decision tree of the training set, and determining the negative gradient of the decision tree for each sample data in the training set according to the result prediction formula; Substituting each of the negative gradients into the result prediction formula to obtain a decision prediction result; Determine the loss value of the result prediction value and the decision prediction result, and when the loss value meets a preset condition, determine that the gradient enhancement decision tree training is completed, otherwise reconstruct the decision tree.

4. The method according to claim 1, characterized in that: The step of processing the enhanced feature data according to the attention model to obtain the attention feature data includes: Inputting the enhanced feature data into the sequence embedding layer of the attention model for dimension increase to obtain an embedding vector; Inputting the embedding vector into the position encoding layer of the attention model for encoding to obtain a position encoding vector; Inputting the position encoding vector into the multi-head attention layer of the attention model to obtain an attention vector by setting attention weights for different positions in the position encoding vector in the multi-head attention layer; The attention vector is input into the linear and regression layers of the attention model to obtain the attention feature data.

5. The method according to claim 1, characterized in that: The attention model is a multi-head attention model, and the loss function used in the training process of the multi-head attention model is a cross-entropy loss function.

6. The method according to claim 1, characterized in that: Before the enhanced feature data is processed according to the attention model to obtain the attention feature data, the method further includes: The enhanced feature data is split and combined according to dimensions.

7. The method according to claim 1, characterized in that: The financial risk assessment model includes at least one of the following: a convolutional neural network model and a decision tree model.

8. An information evaluation device, characterized in that: include: A feature extraction module, used to obtain the user's financial service data and determine feature data of the financial service data; A feature enhancement module, used for determining enhanced feature data based on the integrated learning model and the feature data; An attention module, used for processing the enhanced feature data according to an attention model to obtain attention feature data; An assessment information module is used to determine the risk assessment information of the user based on the financial risk assessment model and the attention feature data.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the information evaluation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the information evaluation method according to any one of claims 1 to 7 when executed.

Images