Business risk prediction method, device, computer equipment and storage medium

Abstract

The present application relates to the technical field of big data analysis, and in particular to a business risk prediction method, apparatus, computer equipment, storage medium, and computer program product. The method comprises: obtaining customer information data and business status information data in multiple historical time series; mapping the customer information data and status information data to a low-dimensional dense vector space to obtain an embedded feature vector, the embedded feature vector comprising a customer feature vector and a state feature vector; performing weighted calculation on the state feature vector based on a multi-head attention mechanism to obtain an associated feature vector, the associated feature vector being used to characterize the relationship between the state of the business in the historical time series and the future state; fusing the associated feature vector with the customer feature vector to obtain a fused feature vector; inputting the fused feature vector into a preset risk classification model for identification to obtain a prediction result of the business risk. Accurate business risk prediction results can be obtained by using this method.

Description

Technical Field

[0001] The present application relates to the technical field of big data analysis, and in particular to a business risk prediction method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0002] As the demand for loans from enterprises and individuals gradually expands, the risks generated in the entire loan cycle are also increasing. Therefore, loan early warning work is of great significance before, during and after the loan.

[0003] When predicting loan risks, existing technologies usually only focus on the impact of loan funding risks at the current moment on future moments. This is too one-sided and cannot accurately and efficiently predict the subsequent evolution of the loan, and cannot obtain accurate risk prediction results.

[0004] It can be seen that the above-mentioned prior art has the problem of low accuracy of prediction results. Summary of the Invention

[0005] Based on this, it is necessary to provide a more accurate business risk prediction method, device, computer equipment, computer-readable storage medium and computer program product to address the above technical problems.

[0006] In a first aspect, the present application provides a business risk prediction method. The method comprises:

[0007] Obtain customer information data and business status information data in multiple historical time series;

[0008] Mapping customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, which includes a customer feature vector and a status feature vector;

[0009] Based on the multi-head attention mechanism, the state feature vector is weighted and calculated to obtain the correlation feature vector. The correlation feature vector is used to represent the correlation between the business state in the historical time series and the future state.

[0010] Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector;

[0011] The fused feature vector is input into the preset risk classification model for identification to obtain the prediction result of business risk.

[0012] In one embodiment, mapping the customer information data and the status information data into a low-dimensional dense vector space to obtain an embedded feature vector includes:

[0013] Preprocessing the customer information data and the status information data to obtain preprocessed customer information data and status information data;

[0014] The preprocessed customer information data and status information data are input into the feature embedding network for dimensionality reduction to obtain the embedded feature vector.

[0015] In one embodiment, the pre-processing includes:

[0016] At least one of data cleaning, time sorting, feature engineering, abnormal data identification, abnormal data removal, normalization, and one-hot encoding.

[0017] In one embodiment, weighted calculation is performed on the state feature vector based on the multi-head attention mechanism to obtain the associated feature vector including:

[0018] Input the state feature vector into the preset multi-head attention model for self-attention operation to obtain the attention distribution corresponding to the state feature vector;

[0019] Normalize the attention distribution to obtain a normalized attention distribution sequence;

[0020] The normalized attention distribution sequence is input into the feedforward neural network to obtain the associated feature vector.

[0021] In one embodiment, before inputting the fused feature vector into a preset risk classification model for identification and obtaining a prediction result of the business risk, the following steps are performed:

[0022] Obtain a preset sample data set, which carries the feature vector of the data to be classified and the actual classification result corresponding to the data to be classified;

[0023] Input the preset sample data set into the initial risk classification model for classification, and use the SoftMax function to obtain the classification results;

[0024] The loss value is calculated using the cross entropy loss function based on the classification results and the true classification results to obtain the loss value of the initial risk classification model;

[0025] If the loss value is not less than the preset classification loss value threshold, the loss value in the initial risk classification model is adjusted to obtain an adjusted initial risk classification model;

[0026] The adjusted initial risk classification model is used again as the initial risk classification model, and the preset sample data set is input into the initial risk classification model for classification, and the classification result is obtained by using the SoftMax function until the loss value is less than the preset classification loss value threshold, thereby obtaining the latest initial risk classification model;

[0027] The latest initial risk classification model is used as the default risk classification model.

[0028] In one embodiment, after obtaining the customer information data and the business status information data in multiple historical time series, the method further includes:

[0029] If the amount of status information data is not less than the preset data amount threshold, the process proceeds to the step of mapping the customer information data and status information data into a low-dimensional dense vector space;

[0030] If the amount of status information data is less than a preset data amount threshold, then obtaining target customer information data that matches the customer information data;

[0031] Performing data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data;

[0032] According to the similarity algorithm, the similarity between the feature vector and the target feature vector is obtained;

[0033] If the similarity is greater than the preset threshold, the target customer's business status information data is obtained;

[0034] Based on the business status information data, the business risk prediction results are obtained.

[0035] In a second aspect, the present application further provides a business risk prediction device. The device comprises:

[0036] Information acquisition module, used to obtain customer information data and business status information data in multiple historical time series;

[0037] A preprocessing module is used to map customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, which includes a customer feature vector and a status feature vector;

[0038] The multi-head attention module is used to perform weighted calculations on the state feature vectors based on the multi-head attention mechanism to obtain the correlation feature vectors. The correlation feature vectors are used to represent the correlation between the business state in the historical time series and the future state.

[0039] The classification prediction module is used to fuse the associated feature vector with the customer feature vector to obtain a fused feature vector; the fused feature vector is input into the preset risk classification model for identification to obtain the prediction result of business risk.

[0040] In one embodiment, the preprocessing module is further used to preprocess the customer information data and status information data to obtain preprocessed customer information data and status information data; the preprocessed customer information data and status information data are input into the feature embedding network for dimensionality reduction processing to obtain an embedded feature vector.

[0041] In one embodiment, the multi-head attention module is also used to input the state feature vector into a preset multi-head attention model to perform self-attention operation to obtain the attention distribution corresponding to the state feature vector; normalize the attention distribution to obtain a normalized attention distribution sequence; input the normalized attention distribution sequence into the feedforward neural network to obtain the associated feature vector.

[0042] In one embodiment, the classification prediction module is also used to obtain a preset sample data set, which carries the feature vector of the data to be classified and the actual classification result corresponding to the data to be classified; the preset sample data set is input into the initial risk classification model for classification, and the classification result is obtained using the SoftMax function; the loss value is calculated based on the classification result and the actual classification result using the cross entropy loss function to obtain the loss value of the initial risk classification model; if the loss value is not less than the preset classification loss value threshold, the loss value in the initial risk classification model is adjusted to obtain the adjusted initial risk classification model; the adjusted initial risk classification model is used again as the initial risk classification model, and the step of inputting the preset sample data set into the initial risk classification model for classification and obtaining the classification result using the SoftMax function is returned until the loss value is less than the preset classification loss value threshold to obtain the latest initial risk classification model; the latest initial risk classification model is used as the preset risk classification model.

[0043] In one embodiment, the classification prediction module is also used to, if the amount of status information data is not less than a preset data volume threshold, enter the step of mapping the customer information data and the status information data to a low-dimensional dense vector space; if the amount of status information data is less than the preset data volume threshold, obtain target customer information data that matches the customer information data; perform data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data; obtain the similarity between the feature vector and the target feature vector based on a similarity algorithm; if the similarity is greater than a preset threshold, obtain the business status information data of the target customer; and obtain a prediction result of the business risk based on the business status information data.

[0044] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0045] Obtain customer information data and business status information data in multiple historical time series;

[0046] Mapping customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, which includes a customer feature vector and a status feature vector;

[0047] Based on the multi-head attention mechanism, the state feature vector is weighted and calculated to obtain the correlation feature vector. The correlation feature vector is used to represent the correlation between the business state in the historical time series and the future state.

[0048] Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector;

[0049] The fused feature vector is input into the preset risk classification model for identification to obtain the prediction result of business risk.

[0050] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0051] Obtain customer information data and business status information data in multiple historical time series;

[0052] Mapping customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, which includes a customer feature vector and a status feature vector;

[0053] Based on the multi-head attention mechanism, the state feature vector is weighted and calculated to obtain the correlation feature vector. The correlation feature vector is used to represent the correlation between the business state in the historical time series and the future state.

[0054] Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector;

[0055] The fused feature vector is input into the preset risk classification model for identification to obtain the prediction result of business risk.

[0056] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0057] Obtain customer information data and business status information data in multiple historical time series;

[0058] Mapping customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, which includes a customer feature vector and a status feature vector;

[0059] Based on the multi-head attention mechanism, the state feature vector is weighted and calculated to obtain the correlation feature vector. The correlation feature vector is used to represent the correlation between the business state in the historical time series and the future state.

[0060] Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector;

[0061] The fused feature vector is input into the preset risk classification model for identification to obtain the prediction result of business risk.

[0062] The above-mentioned business risk prediction method, device, computer equipment, storage medium and computer program product first obtain customer information data and status information data of the loan business in multiple historical time series, then map the obtained customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, and then perform weighted calculation on the state feature vector based on the multi-head attention mechanism to obtain the correlation between the state of the business in multiple historical time series and the future state, and then fuse the correlation feature vector with the customer feature vector to obtain a fused feature vector, and finally input the fused feature vector into a preset risk classification model for identification. During the entire process, the impact of the status information data of multiple historical time series on the business risk is comprehensively considered to ensure the accuracy of the risk prediction, and the obtained information data is converted into an embedded feature vector, which can ensure the quality of subsequent data processing and is conducive to obtaining accurate risk prediction results to a certain extent. In addition, the multi-head attention mechanism can obtain the correlation between the state of the business in multiple historical time series and the future state, and fuse the correlation feature vector with the customer feature vector, further ensuring the accuracy of the final risk prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is an application environment diagram of a business risk prediction method in one embodiment;

[0064] Figure 2 1 is a flow chart of a business risk prediction method according to an embodiment;

[0065] Figure 3 A schematic diagram of a process for obtaining a preset risk classification model in one embodiment;

[0066] Figure 4 Schematic diagram of the process after S100 in another embodiment;

[0067] Figure 5Schematic diagram of the entire data processing process for business risk prediction in one embodiment;

[0068] Figure 6 is a structural block diagram of a business risk prediction device in one embodiment;

[0069] Figure 7 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application, and that the acquisition, storage, use, and processing of data in the technical solutions of this application comply with the relevant provisions of national laws and regulations.

[0071] The business risk prediction method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store data that server 104 needs to process. The data storage system can be integrated with server 104, or placed on a cloud or other network server. Terminal 102 sends customer information data and business status information data from multiple historical time series to server 104. Server 104 maps the obtained customer information data and status information data into a low-dimensional dense vector space to obtain an embedded feature vector, where the embedded feature vector includes a customer feature vector and a state feature vector. The state feature vector is then weighted based on a multi-head attention mechanism to obtain a correlation feature vector. The obtained correlation feature vector is fused with the customer feature vector to obtain a fused feature vector. Finally, the fused feature vector is input into a preset risk classification model for identification, thereby obtaining a business risk prediction result. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart car devices, etc. Portable wearable devices can include smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0072] In one embodiment, Figure 2 As shown, a business risk prediction method is provided, which is applied to Figure 1 The following steps are used as an example to illustrate the server in the example:

[0073] S100, obtaining customer information data and business status information data in multiple historical time series.

[0074] Among them, customer information data refers to the basic information of customers who have handled a certain business; status information data refers to the status changes of the business handled by the customer in multiple historical time series; historical time series refers to a sequence composed of multiple historical sampling time points for obtaining status information data; specifically, when predicting the risk of a certain business, the information of the customer handling the business and the changes in the status information of the business at historical moments are closely related to the business risk prediction. Therefore, when predicting business risks, it is necessary to comprehensively consider the impact of customer information data and business status information data on the final business risk. At the same time, since the same business may have different information and status at different times, for example, a loan business was in a normal state two months ago, and may become suspicious two months later. This state change does not occur suddenly at this moment, but is a subtle change. In order to obtain accurate prediction results, it is necessary to predict the possible risks of the subsequent business based on the previous state (historical state) of the current business. Therefore, customer information data and business status information data in multiple historical time series are obtained. The customer information and business status information mainly include content. Customer information and business status information will be different in different fields. For example: when used in the risk prediction of the bank's loan business, customer information mainly includes age, household registration, education level, occupation, income, mortgage assets, credit points, purpose of loan funds and transaction loss, while business status information mainly includes the handling bank, the maximum number of overdue periods, available credit limit, expenditure amount, fund pool number, interest calculation value, deduction method, number of overdue periods, loan interest rate, cumulative recovery of liquidated damages, cumulative recovery of interest and cumulative write-off amount.

[0075] S200 , mapping the customer information data and the status information data into a low-dimensional dense vector space to obtain an embedded feature vector.

[0076] The embedded feature vector is a low-dimensional feature vector that can represent the semantic information of the information data; specifically, since the data dimensions of the acquired customer information data and status information data are different from the data dimensions that can be identified subsequently, in order to facilitate the subsequent model to operate on the information data and ensure the effect of data processing to obtain accurate prediction results, it is necessary to first perform a series of preprocessing operations on the multiple information data, extract the features of the status data, convert it from general data into a feature vector, and then map it to a low-dimensional dense vector space through the embedding layer to obtain an embedded feature vector. The obtained embedded feature vector includes the customer feature vector corresponding to the customer information data and multiple state feature vectors corresponding to the status information data.

[0077] S300, performing weighted calculation on the state feature vector based on the multi-head attention mechanism to obtain the associated feature vector.

[0078] Among them, the multi-head attention mechanism is a mechanism that can capture the impact of the business state information in multiple historical time series on the change of the business state at the next moment; the associated feature vector is a feature vector that can represent the impact of the business state at the past moment on the future state. Specifically, the state feature vector reflects the changes in the business state over time. The state feature vectors of different historical time series have different effects on the business state at future moments. Generally speaking, the state information data of the more recent historical time series has a greater impact on the change of the business state at future moments. According to the multi-head attention mechanism, the state feature vector is weighted and calculated, and the attention distribution of the state information data of each historical time series is allocated to obtain the associated feature vector, which can characterize the correlation between the state of the business in each historical time series and the future state.

[0079] S400: Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector.

[0080] Specifically, the obtained associated feature vector and customer feature vector are input into the trained deep neural network, and the features are spliced ​​and fused to obtain a total feature vector. The calculation method is as follows:

[0081] Tanh(Concat(X,U)W+b)

[0082] Among them, X is the output generated by the multi-head attention mechanism, that is, the associated feature vector, U represents the customer feature vector, Tanh is the activation function, and W and b represent trainable parameters.

[0083] S500: Input the fused feature vector into a preset risk classification model for identification to obtain a prediction result of the business risk.

[0084] Specifically, business risks are divided into different levels according to the degree of risk. For example, in the loan business, loan business risks can be divided into: normal, special mention, substandard, doubtful, and loss. The latter three are classified as non-performing loans. The fused feature vector is input into the preset risk classification model for identification. The SoftMax function is used to operate on the input feature vector. The calculation method is as follows:

[0085]

[0086] Where S is a five-bit vector, and the value of each element represents the probability of the corresponding business risk category. V is the output generated by the deep neural network module. The category corresponding to the maximum value calculated by the preset risk classification model is the predicted risk status of the business at the next moment. For example, in the loan business, the probabilities corresponding to the five risks of normal, special attention, substandard, doubtful, and loss are: 80%, 60%, 30%, 15%, and 5%, indicating that the business will be in a normal risk state at the next moment.

[0087] In the above-mentioned business risk prediction method, on the one hand, customer information data and status information data of loan business in multiple historical time series are obtained, and the impact of status information data of multiple historical time series on business risk is taken into account to ensure the accuracy of risk prediction. On the other hand, in the process of information data processing, the obtained customer information data and status information data are first mapped to a low-dimensional dense vector space to obtain an embedded feature vector to ensure the quality of subsequent data processing. Then, based on the multi-head attention mechanism, the state feature vector is weightedly calculated to obtain the correlation relationship between the state of the business in multiple historical time series and the future state. Then, the correlation feature vector is fused with the customer feature vector to obtain the fused feature vector. Finally, the fused feature vector is input into the preset risk classification model for identification to further ensure the accuracy of the prediction results.

[0088] In one embodiment, the customer information data and the status information data are mapped to a low-dimensional dense vector space to obtain an embedded feature vector including:

[0089] Step 1: pre-process the customer information data and status information data to obtain pre-processed customer information data and status information data;

[0090] Step 2: Input the pre-processed customer information data and status information data into the feature embedding network for dimensionality reduction to obtain the embedded feature vector.

[0091] Among them, preprocessing refers to some processing performed on information data before the main processing to ensure the effectiveness of data processing; feature embedding network refers to a neural network that can identify the dimension of input information data and reduce its dimensionality. Specifically, customer information data and status information data are first subjected to a series of preprocessing actions such as data cleaning, time sorting, feature engineering, abnormal data identification, abnormal data removal, normalization, and one-hot encoding to complete the feature extraction of customer information data and status information data. The feature vector obtained at this time is a high-dimensional sparse feature vector. The high-dimensional sparse feature vector is input into the feature embedding network. The data dimension of the input high-dimensional sparse feature vector is identified by the feature embedding network. The high-dimensional sparse feature vector obtained is then embedded through the embedding layer to map the data into a low-dimensional dense vector space to obtain an embedded feature vector. The embedded feature vector includes the customer feature vector and the status feature vector. The dimensions of the customer feature vector and the status feature vector at this time are the same as the dimensional information of the identifiable data in the subsequent processing process.

[0092] In this embodiment, by preprocessing the customer information data and status information data first, information data that can be truly used to predict risks can be extracted from a large amount of relatively messy information data, which can greatly improve the efficiency of data processing and ensure the effect of subsequent data processing to ensure the accuracy of the final risk prediction results. At the same time, the preprocessed information data is subjected to dimensionality reduction processing, and the data is processed in a low-dimensional dense space. Key features are extracted from high-dimensional data, and redundant features and noise are eliminated, which is beneficial to subsequent processing and can effectively improve the accuracy of the final predicted risk, greatly reduce the computational cost, and improve efficiency.

[0093] In one embodiment, pre-processing includes:

[0094] At least one of data cleaning, time sorting, feature engineering, abnormal data identification, abnormal data removal, normalization, and one-hot encoding.

[0095] Data cleaning refers to the final step of identifying and correcting identifiable errors in data files, including checking data consistency and handling invalid and missing values. Time sorting refers to sorting data based on the order in which it was acquired. Feature engineering is essentially an engineering activity for extracting features from data, aiming to maximize the extraction of features from the raw data for use in algorithms and models. Anomaly identification refers to identifying significantly deviating or erroneous data within the data. Anomaly removal refers to removing significantly deviating or erroneous data from the data. Normalization converts data into decimals between (0, 1) or (1, 1), primarily for ease of data processing, mapping the data to a range of 0 to 1. One-hot encoding uses an N-bit state register to encode N states, with each state having a separate register bit, and only one bit active at any given time. Specifically, the acquired customer information and status data undergo data cleaning, time sorting, feature engineering, anomaly identification, anomaly removal, normalization, and one-hot encoding to obtain a data format that can be recognized and processed by the feature embedding network.

[0096] In this embodiment, by performing a series of preprocessing on the information data, including data cleaning, time sorting, feature engineering, abnormal data identification, abnormal data removal, normalization, and one-hot encoding, the efficiency of data processing can be improved and the effect of subsequent data processing can be guaranteed, so as to ensure the accuracy of the final risk prediction results.

[0097] In one embodiment, the state feature vector is weightedly calculated based on the multi-head attention mechanism to obtain the associated feature vector including:

[0098] Step 1: Input the state feature vector into the preset multi-head attention model to perform self-attention operation and obtain the attention distribution corresponding to the state feature vector;

[0099] Step 2: normalize the attention distribution to obtain a normalized attention distribution sequence;

[0100] Step 3: Input the normalized attention distribution sequence into the feedforward neural network to obtain the associated feature vector.

[0101] Among them, attention distribution refers to selecting information related to a specific task from multiple input vectors, and obtaining the feature weight corresponding to each input feature vector, that is, the importance of each input feature vector to a specific task; the preset multi-head attention model for each input vector is a model composed of multiple self-attention operation modules that can perform weighted calculations on the state feature vector; the feedforward neural network is a neural network composed of two layers of fully connected neural networks that can summarize the normalized attention distribution sequence; specifically, the state feature vector is input into the preset multi-head attention model for self-attention operation to obtain the attention distribution corresponding to the state feature vector. The operation method is as follows:

[0102]

[0103]

[0104] MultiHead(Q,K,V)=Concat(head1,…,head i ,…,head n )W O

[0105] Where W represents a parameterizable matrix, Att() represents the attention operation, d represents the vector dimension, MultiHead() represents the final output of multi-head attention, and Q, K, and V are the query, key, and value vectors generated by the input vector for calculating attention. Since the obtained attention distribution sequence corresponds to each historical time series, each historical time series includes multiple moments, and each moment has a corresponding attention distribution. In order to speed up the efficiency of subsequent data processing, the attention distribution corresponding to the state feature vector under each historical time series is normalized, and the attention distribution of all moments in each historical time series is normalized into a total attention distribution, obtaining a normalized attention distribution sequence including multiple historical time series. The normalization operation is performed as follows:

[0106] LayerNorm(X+MultiHead(Q,K,V))

[0107] Here, X represents the state feature vector. The LayerNorm function approximates the mean and variance of the inputs to each layer of neurons, accelerating model convergence. The resulting normalized attention distribution sequence is then fed into a feedforward neural network for processing to generate a correlation feature vector. This correlation feature vector essentially summarizes the state information of the business across multiple historical time series and represents the overall impact of the business state across all historical time series on the business state at the next moment.

[0108] In this embodiment, the state feature vector is subjected to self-attention operation through a multi-head attention model to obtain the attention distribution corresponding to the state feature vector. The attention distribution is normalized to obtain a normalized attention distribution sequence, which is input into the feedforward neural network to obtain an associated feature vector. This can capture the influence and connection between each time series and the impact on the business status changes at future moments, thereby realizing the prediction of business status changes at future moments based on the business status of historical time series to obtain accurate risk prediction results.

[0109] In one embodiment, Figure 3 As shown, S500 previously included:

[0110] S420, obtaining a preset sample data set, the preset sample data set carrying the feature vector of the data to be classified and the actual classification result corresponding to the data to be classified;

[0111] S440, inputting the preset sample data set into the initial risk classification model for classification, and obtaining the classification result using the SoftMax function;

[0112] S460, calculating a loss value using a cross entropy loss function based on the classification result and the true classification result to obtain a loss value of the initial risk classification model;

[0113] S470, if the loss value is not less than the preset classification loss value threshold, adjusting the loss value in the initial risk classification model to obtain an adjusted initial risk classification model;

[0114] S472, if the loss value is less than the preset classification loss value threshold, the initial risk classification model is used as the preset risk classification model;

[0115] S480, the adjusted initial risk classification model is used again as the initial risk classification model, and the process returns to the step of inputting the preset sample data set into the initial risk classification model for classification, and using the SoftMax function to obtain the classification result, until the loss value is less than the preset classification loss value threshold, thereby obtaining the latest initial risk classification model;

[0116] S490, using the latest initial risk classification model as the preset risk classification model.

[0117] Among them, the preset sample data set refers to the data set for training the risk classification model; the SoftMax function is the gradient logarithm normalization function of the finite-term discrete probability distribution; the cross entropy loss function is a classification objective function existing in the convolutional neural network; the loss value refers to the error between the actual result of the inference result calculated according to the model; the preset classification loss value threshold is the critical condition for the risk classification model to complete convergence; specifically, a preset sample data set is obtained, which carries the feature vector of the data to be classified and the true classification result corresponding to the data to be classified, and the preset sample data set is input into the initial risk classification model for classification, and the SoftMax function is used to obtain the classification result, and the obtained classification result is compared with the true classification result in the data set. According to the error between the two, the cross entropy loss function is used to calculate the loss value to obtain the loss value of the initial risk classification model, and the calculated loss value is compared with the preset classification loss value threshold, and the comparison result is used as the judgment of the initial risk classification model. The basis for whether the risk classification model has completed training is that if the loss value is not less than the preset classification loss value threshold, the loss value in the initial risk classification model is adjusted through the error back propagation algorithm to reduce the loss value and obtain the adjusted initial risk classification model. The adjusted initial risk classification model is used as the initial risk classification model again, and the preset sample data set is continued to be input into the initial risk classification model for classification. The classification result is obtained by using the SoftMax function until the loss value is less than the preset classification loss value threshold. At this time, the risk classification model completes convergence and the latest initial risk classification model is obtained. The latest initial risk classification model is used as the preset risk classification model. Conversely, if the loss value of the initial classification model is less than the preset classification loss value threshold, the initial classification model can be directly used as the preset risk classification model, and the fused feature vector is subsequently processed to obtain the final prediction result of the business risk.

[0118] In this embodiment, before inputting the fused feature vector into the preset risk classification model for identification and classification, it is necessary to first obtain a preset sample data set to train the risk classification model to ensure the accuracy of the preset risk classification model, so as to ensure that after the fused feature vector is input into the preset risk classification model for identification, accurate risk prediction results can be obtained.

[0119] In one embodiment, Figure 4 As shown, S100 and later also include:

[0120] S210, if the amount of status information data is not less than the preset data amount threshold, then enter the step of mapping the customer information data and status information data into a low-dimensional dense vector space;

[0121] S220, if the amount of status information data is less than the preset data amount threshold, obtaining target customer information data that matches the customer information data;

[0122] S240, performing data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data;

[0123] S260, obtaining the similarity between the feature vector and the target feature vector according to a similarity algorithm;

[0124] S280, if the similarity is greater than a preset threshold, obtaining the target customer's business status information data;

[0125] S282, if the similarity is not greater than the preset threshold, returning to the step of obtaining target customer information data that matches the customer information data;

[0126] S290: Obtain a prediction result of the business risk based on the business status information data.

[0127] Among them, target customer information data refers to the information data of target customers that matches the customer information data; similarity algorithm refers to an algorithm for calculating the similarity between the customer feature vector of the current customer and the target customer feature vector of the target customer. Specifically, if the amount of status information data is not less than the preset data volume threshold, it means that the customer is an old customer and there is enough status information data to predict the risk of the business it handles, which can ensure accurate prediction results. Then enter the above-mentioned step of mapping customer information data and status information data to low-dimensional dense vector space. If the amount of status information data is less than the preset data volume threshold, it means that the customer is a new customer and there is not enough historical time series status information data to predict the risk of its business. It faces the cold start problem. Based on insufficient data, the accuracy of the final prediction result cannot be guaranteed. In this case, the target customer information data that matches the customer information data can be obtained, that is, the information data of the target customer who has handled the same business and whose basic situation is basically similar to that of the customer is obtained, and it is used as a reference for predicting the business risk of the customer. The obtained customer information data and target customer information data are converted into the form of feature vectors to facilitate subsequent data processing, and the feature vector corresponding to the customer information data and the target feature vector corresponding to the target customer information data are obtained. According to the similarity algorithm, the similarity between the feature vector and the target feature vector is obtained, and it is judged whether the similarity meets the preset threshold. If the similarity is greater than the preset threshold, the business status information data of the target user is obtained. According to the business status information data, combined with the customer information data of the new customer, the previous steps are performed to obtain the final business risk prediction result. On the contrary, if the similarity is not greater than the preset threshold, the step of obtaining the target customer information data that matches the customer information data is returned until the target customer whose similarity meets the conditions is found, and then the final business risk prediction result is obtained according to the business status information data of the target user.

[0128] In this embodiment, when the acquired status information data is insufficient, target customer information data that matches the customer information data is acquired and used as a reference to predict the business risk of the customer. This can effectively solve the cold start problem of insufficient analysis data, and ensure that relatively accurate prediction results can be obtained even when the status information data is insufficient.

[0129] Figure 5This is a schematic diagram of the data processing process of the business risk prediction method. It should be understood that although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0130] Based on the same inventive concept, the embodiments of the present application also provide a business risk prediction device for implementing the business risk prediction method involved above. The implementation solution provided by this device is similar to the implementation solution described in the above method. Therefore, the specific limitations of one or more business risk prediction device embodiments provided below can be referred to the limitations of the business risk prediction method above and will not be repeated here.

[0131] In one embodiment, Figure 6 As shown, a business risk prediction device is provided, comprising: an information acquisition module 100, a preprocessing module 200, a multi-head attention module 300 and a classification prediction module 400, wherein:

[0132] The information acquisition module 100 is used to obtain customer information data and business status information data in multiple historical time series;

[0133] A preprocessing module 200 is configured to map the customer information data and the status information data into a low-dimensional dense vector space to obtain an embedded feature vector, where the embedded feature vector includes a customer feature vector and a status feature vector;

[0134] The multi-head attention module 300 is used to perform weighted calculation on the state feature vector based on the multi-head attention mechanism to obtain the correlation feature vector. The correlation feature vector is used to represent the correlation relationship between the state of the business in the historical time series and the future state;

[0135] The classification prediction module 400 is used to fuse the associated feature vector with the customer feature vector to obtain a fused feature vector; input the fused feature vector into a preset risk classification model for identification to obtain a prediction result of the business risk.

[0136] The above-mentioned business risk prediction device, on the one hand, obtains customer information data and status information data of loan business in multiple historical time series, and takes into account the impact of status information data of multiple historical time series on business risk to ensure the accuracy of risk prediction. On the other hand, in the process of processing information data, the obtained customer information data and status information data are first mapped to a low-dimensional dense vector space to obtain an embedded feature vector to ensure the quality of subsequent data processing. Then, based on the multi-head attention mechanism, the state feature vector is weightedly calculated to obtain the correlation relationship between the state of the business in multiple historical time series and the future state. Then, the correlation feature vector is fused with the customer feature vector to obtain the fused feature vector. Finally, the fused feature vector is input into the preset risk classification model for identification to further ensure the accuracy of the prediction result.

[0137] In one embodiment, the preprocessing module 200 is also used to preprocess the customer information data and status information data to obtain preprocessed customer information data and status information data; the preprocessed customer information data and status information data are input into the feature embedding network for dimensionality reduction processing to obtain an embedded feature vector.

[0138] In one embodiment, the multi-head attention module 300 is also used to input the state feature vector into a preset multi-head attention model to perform self-attention operation to obtain the attention distribution corresponding to the state feature vector; normalize the attention distribution to obtain a normalized attention distribution sequence; input the normalized attention distribution sequence into the feedforward neural network to obtain the associated feature vector.

[0139] In one embodiment, the classification prediction module 400 is also used to obtain a preset sample data set, which carries the feature vector of the data to be classified and the actual classification result corresponding to the data to be classified; the preset sample data set is input into the initial risk classification model for classification, and the classification result is obtained using the SoftMax function; the loss value is calculated based on the classification result and the actual classification result using the cross entropy loss function to obtain the loss value of the initial risk classification model; if the loss value is not less than the preset classification loss value threshold, the loss value in the initial risk classification model is adjusted to obtain the adjusted initial risk classification model; the adjusted initial risk classification model is used again as the initial risk classification model, and the step of inputting the preset sample data set into the initial risk classification model for classification and obtaining the classification result using the SoftMax function is returned until the loss value is less than the preset classification loss value threshold to obtain the latest initial risk classification model; the latest initial risk classification model is used as the preset risk classification model.

[0140] In one embodiment, the classification prediction module 400 is also used to, if the amount of status information data is not less than a preset data volume threshold, enter the step of mapping the customer information data and the status information data to a low-dimensional dense vector space; if the amount of status information data is less than the preset data volume threshold, obtain target customer information data that matches the customer information data; perform data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data; obtain the similarity between the feature vector and the target feature vector based on a similarity algorithm; if the similarity is greater than a preset threshold, obtain the business status information data of the target customer; and obtain a prediction result of the business risk based on the business status information data.

[0141] Each module in the business risk prediction device described above may be implemented in whole or in part through software, hardware, or a combination thereof. Each module may be embedded in or independent of a processor in a computer device in hardware form, or may be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0142] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store customer information data and business status information data in multiple historical time series. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it implements a business risk prediction method.

[0143] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0144] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0145] In one embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above-mentioned method embodiments when executed by a processor.

[0146] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of each of the above-described method embodiments.

[0147] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a programmable logic unit (PLC), a data processing logic unit based on quantum computing, and the like.

[0148] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A business risk prediction method, characterized in that: The method comprises: Obtain customer information data and business status information data in multiple historical time series; Preprocessing the customer information data and the status information data to obtain preprocessed customer information data and status information data, inputting the preprocessed customer information data and status information data into a feature embedding network as high-dimensional sparse feature vectors, performing data dimension identification and embedding operations on the input high-dimensional sparse feature vectors through the feature embedding network, mapping the high-dimensional sparse feature vectors into a low-dimensional dense vector space, and obtaining an embedded feature vector, wherein the embedded feature vector includes a customer feature vector and a status feature vector; Performing weighted calculation on the state feature vector based on a multi-head attention mechanism to obtain a correlation feature vector, where the correlation feature vector is used to represent the correlation relationship between the state of the business in the historical time series and the future state; Fusing the associated feature vector with the customer feature vector to obtain a fused feature vector; The fused feature vector is input into a preset risk classification model for identification to obtain a prediction result of the business risk.

2. The method according to claim 1, characterized in that The pretreatment includes: At least one of data cleaning, time sorting, feature engineering, abnormal data identification, abnormal data removal, normalization, and one-hot encoding.

3. The method according to claim 1, characterized in that The weighted calculation of the state feature vector based on the multi-head attention mechanism to obtain the associated feature vector includes: Input the state feature vector into a preset multi-head attention model to perform self-attention operation to obtain the attention distribution corresponding to the state feature vector; Normalizing the attention distribution to obtain a normalized attention distribution sequence; The normalized attention distribution sequence is input into a feedforward neural network to obtain a correlation feature vector.

4. The method according to claim 1, wherein Before inputting the fused feature vector into a preset risk classification model for identification and obtaining a prediction result of the business risk, the process includes: Obtaining a preset sample data set, wherein the preset sample data set carries a feature vector of the data to be classified and a true classification result corresponding to the data to be classified; Input the preset sample data set into the initial risk classification model for classification, and use the SoftMax function to obtain the classification result; Calculating a loss value using a cross entropy loss function based on the classification result and the true classification result to obtain a loss value of the initial risk classification model; If the loss value is not less than the preset classification loss value threshold, adjusting the loss value in the initial risk classification model to obtain an adjusted initial risk classification model; The adjusted initial risk classification model is used again as the initial risk classification model, and the step of inputting the preset sample data set into the initial risk classification model for classification is returned to, and a classification result is obtained by using a SoftMax function, until the loss value is less than a preset classification loss value threshold, thereby obtaining the latest initial risk classification model; The latest initial risk classification model is used as the preset risk classification model.

5. The method according to claim 1, wherein After obtaining the customer information data and the business status information data in multiple historical time series, the method further includes: If the amount of the status information data is not less than the preset data amount threshold, then entering the step of mapping the customer information data and the status information data into a low-dimensional dense vector space; If the amount of the status information data is less than a preset data amount threshold, obtaining target customer information data that matches the customer information data; Performing data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data; Obtaining the similarity between the feature vector and the target feature vector according to a similarity algorithm; If the similarity is greater than a preset threshold, obtaining the business status information data of the target customer; A prediction result of the business risk is obtained based on the business status information data.

6. A business risk prediction device, characterized in that: The device comprises: Information acquisition module, used to obtain customer information data and business status information data in multiple historical time series; a preprocessing module, configured to preprocess the customer information data and the status information data to obtain preprocessed customer information data and status information data, input the preprocessed customer information data and status information data as high-dimensional sparse feature vectors into a feature embedding network, perform data dimension identification and embedding operations on the input high-dimensional sparse feature vectors through the feature embedding network, map the high-dimensional sparse feature vectors into a low-dimensional dense vector space, and obtain an embedded feature vector, wherein the embedded feature vector includes a customer feature vector and a status feature vector; a multi-head attention module, configured to perform weighted calculation on the state feature vector based on a multi-head attention mechanism to obtain a correlation feature vector, wherein the correlation feature vector is used to represent the correlation relationship between the state of the business in the historical time series and the future state; The classification prediction module is used to fuse the associated feature vector with the customer feature vector to obtain a fused feature vector; input the fused feature vector into a preset risk classification model for identification to obtain a prediction result of business risk.

7. The device according to claim 6, characterized in that The multi-head attention module is also used to input the state feature vector into a preset multi-head attention model to perform self-attention operation to obtain the attention distribution corresponding to the state feature vector; normalize the attention distribution to obtain a normalized attention distribution sequence; input the normalized attention distribution sequence into a feedforward neural network to obtain an associated feature vector.

8. The device according to claim 6, characterized in that The classification prediction module is further used to obtain a preset sample data set, which carries the feature vector of the data to be classified and the actual classification result corresponding to the data to be classified; Input the preset sample data set into the initial risk classification model for classification, and use the SoftMax function to obtain the classification result; Calculating a loss value using a cross entropy loss function based on the classification result and the true classification result to obtain a loss value of the initial risk classification model; if the loss value is not less than a preset classification loss value threshold, adjusting the loss value in the initial risk classification model to obtain an adjusted initial risk classification model; The adjusted initial risk classification model is re-used as the initial risk classification model, and the step of inputting the preset sample data set into the initial risk classification model for classification is returned to, and the classification result is obtained by using the SoftMax function until the loss value is less than the preset classification loss value threshold, thereby obtaining the latest initial risk classification model; and the latest initial risk classification model is used as the preset risk classification model.

9. The device according to claim 6, characterized in that The classification prediction module is also used to enter the step of mapping the customer information data and the status information data to a low-dimensional dense vector space if the amount of the status information data is not less than the preset data amount threshold; if the amount of the status information data is less than the preset data amount threshold, obtain target customer information data that matches the customer information data; perform data conversion on the customer information data and the target customer information data to obtain a feature vector corresponding to the customer information data and a target feature vector corresponding to the target customer information data; obtain the similarity between the feature vector and the target feature vector based on a similarity algorithm; if the similarity is greater than the preset threshold, obtain the business status information data of the target customer; and obtain a prediction result of the business risk based on the business status information data.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

12. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images