CN116629612A - Risk prediction method and device, storage medium and electronic equipment - Google Patents
Risk prediction method and device, storage medium and electronic equipment Download PDFInfo
- Publication number
- CN116629612A CN116629612A CN202310619716.9A CN202310619716A CN116629612A CN 116629612 A CN116629612 A CN 116629612A CN 202310619716 A CN202310619716 A CN 202310619716A CN 116629612 A CN116629612 A CN 116629612A
- Authority
- CN
- China
- Prior art keywords
- data
- risk prediction
- domain
- target
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 93
- 238000013058 risk prediction model Methods 0.000 claims abstract description 62
- 238000013528 artificial neural network Methods 0.000 claims description 14
- 238000004590 computer program Methods 0.000 claims description 6
- 238000013508 migration Methods 0.000 description 18
- 230000005012 migration Effects 0.000 description 18
- 238000010586 diagram Methods 0.000 description 12
- 238000004458 analytical method Methods 0.000 description 11
- 230000000694 effects Effects 0.000 description 9
- 230000006870 function Effects 0.000 description 8
- 238000013526 transfer learning Methods 0.000 description 8
- 238000003062 neural network model Methods 0.000 description 7
- 230000006399 behavior Effects 0.000 description 6
- 238000012545 processing Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012797 qualification Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000012502 risk assessment Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Software Systems (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Mathematical Physics (AREA)
- Game Theory and Decision Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Tourism & Hospitality (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Technology Law (AREA)
- Educational Administration (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A risk prediction method, a risk prediction device, a storage medium and electronic equipment relate to the field of big data or the field of finance. The method comprises the following steps: acquiring data of a first target domain and data of a plurality of first source domains, wherein the data of the first target domain and each first source domain correspond to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product; according to the client characteristics and labels corresponding to the first source domain and the first target domain respectively, determining second source domains with similarity to the first target domain being greater than or equal to a preset threshold value, wherein the labels are used for indicating client risks; training to obtain a risk prediction model, wherein training data of the risk prediction model comprises data of each second source domain; and carrying out risk prediction on the target client of the target product according to the risk prediction model. The scheme can improve the accuracy of the client risk prediction result.
Description
Technical Field
The present application relates to the field of big data technologies, and in particular, to a risk prediction method, a risk prediction device, a storage medium, and an electronic device.
Background
Currently, financial institutions often need to conduct risk prediction for determining the risk of a customer's default when servicing the customer. Through risk prediction, the provision of services to high risk customers can be reduced, or the supervision of the behavior of high risk customers can be added. This is of great importance for maintaining the safety and stability of the financial market and for maintaining the stability of the socioeconomic development.
The traditional way of customer risk analysis is: and collecting the record of the violations of the historical occurrence of the target customer and the record of the violations of the customer corresponding to the product, and then carrying out modeling analysis. However, the number of high risk customers of a financial product is often scarce, resulting in less training data for modeling analysis and lower accuracy of the predicted results.
Disclosure of Invention
In order to solve the problems in the prior art, the application provides a risk prediction method, a risk prediction device, a risk prediction storage medium and electronic equipment, which can improve the accuracy of a client risk prediction result.
In a first aspect, the present application provides a method of risk prediction, the method comprising: acquiring data of a first target domain and data of a plurality of first source domains, wherein the data of the first target domain and each first source domain correspond to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product; according to the client characteristics and labels corresponding to the first source domain and the first target domain respectively, determining second source domains with similarity to the first target domain being greater than or equal to a preset threshold value, wherein the labels are used for indicating client risks; training to obtain a risk prediction model, wherein training data of the risk prediction model comprises data of each second source domain; and carrying out risk prediction on the target client of the target product according to the risk prediction model.
By using the method provided by the application, the first target domain and a plurality of first source domains are acquired first. The first target threshold includes data corresponding to a target product, and the first source domain includes data corresponding to a non-target product. According to the client characteristics and the labels, the method determines each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold value. And then adding the data of each second source domain into training data of model training to obtain a risk prediction model. According to the scheme, the existing data are utilized to a large extent, the data domain with higher migration learning value is found by measuring the similarity of the data, the occurrence probability of negative migration is greatly reduced, the problem of unbalanced data types is solved, the available data of model training is expanded, the analysis efficiency and the model effect are improved, and the accuracy of a client risk prediction result can be improved.
In one possible implementation manner, the determining, according to the client characteristics and the labels corresponding to each of the first source domain and the first target domain, each second source domain having a similarity with the first target domain greater than or equal to a preset threshold specifically includes: determining mutual information values of each first source domain and each first target domain by taking the client features and the labels as variables; and determining the mutual information value as the similarity, and determining a first source domain corresponding to the mutual information value larger than the preset threshold value as one second source domain.
In a possible implementation manner, the training obtains a risk prediction model, which specifically includes: training a fully connected neural network FCNN model by taking the data of the first target domain and the data of each second source domain as training data, wherein a hidden layer of the FCNN model comprises k layers, and k is an integer larger than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In a possible implementation manner, the training obtains a risk prediction model, which specifically includes: training the FCNN model by taking the data of each second source domain as training data, wherein the hidden layer of the FCNN model comprises k layers, and k is an integer greater than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In one possible implementation, the method further includes: updating the data in the first target domain to obtain a second target domain; and updating the last m layers of the hidden layers of the risk prediction model by utilizing the data of the second target domain so as to obtain an updated risk prediction model.
In a possible implementation manner, before the updating of the last m layers of the hidden layers of the FCNN model after training by using the data of the first target domain, the method further includes: determining a ratio value between the number of data of the first target domain and the sum of the number of data of each second source domain; and determining the value of m according to the ratio value and the corresponding relation between the predetermined ratio value and m.
In one possible implementation, the customer characteristics include one or more of the following: customer attributes and customer transaction behavior; the customer transaction actions include transaction times and transaction frequency.
In a second aspect, the present application also provides an apparatus for risk prediction, the apparatus comprising: the system comprises an acquisition unit, a determination unit, a model training unit and a risk prediction unit. The acquisition unit is used for acquiring data of a first target domain and data of a plurality of first source domains, wherein the data of the first target domain and each first source domain corresponds to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product. And the determining unit determines each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold according to the client characteristics and the labels corresponding to each first source domain and each first target domain respectively, wherein the labels are used for indicating client risks. The model training unit is used for training and obtaining a risk prediction model, and the training data of the risk prediction model comprise data of each second source domain. And the risk prediction unit is used for predicting the risk of the target customer of the target product according to the risk prediction model.
By using the device, the occurrence probability of negative migration is reduced, so that the problem of unbalanced data types is solved, available data of model training is expanded, analysis efficiency and model effect are improved, and the accuracy of a customer risk prediction result can be improved.
In a possible implementation manner, the determining unit is specifically configured to determine mutual information values of each of the first source domains and the first target domains by using the client feature and the tag as variables; and determining the mutual information value as the similarity, and determining a first source domain corresponding to the mutual information value larger than the preset threshold value as one second source domain.
In a possible implementation manner, the model training unit is specifically configured to train the fully-connected neural network FCNN model by using the data of the first target domain and the data of each second source domain as training data, where a hidden layer of the FCNN model includes k layers, and k is an integer greater than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In a possible implementation manner, the model training unit is specifically configured to train the fully-connected neural network FCNN model by using the data of each second source domain as training data, where a hidden layer of the FCNN model includes k layers, and k is an integer greater than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In a possible implementation manner, the device further includes an updating unit, where the updating unit is specifically configured to update data in the first target domain to obtain a second target domain; and updating the last m layers of the hidden layers of the risk prediction model by utilizing the data of the second target domain so as to obtain an updated risk prediction model.
In a possible implementation manner, the model training unit is further specifically configured to determine a ratio value between the number of data of the first target domain and the sum of the numbers of data of the second source domains; and determining the value of m according to the ratio value and the corresponding relation between the predetermined ratio value and m.
In a third aspect, the present application also provides a storage medium having stored thereon a program which when executed by a processor implements the method of risk prediction.
In a fourth aspect, the present application further provides an electronic device, where the electronic device may be a server, a PC, a PAD, a mobile phone, or the like. The electronic device is used for running the program. And executing the risk prediction method in any implementation mode when the program runs.
Drawings
FIG. 1 is a flow chart of a method for risk prediction according to an embodiment of the present application;
FIG. 2 is a flow chart of another method for risk prediction provided by an embodiment of the present application;
FIG. 3 is a schematic diagram of a fully connected neural network model according to an embodiment of the present application;
FIG. 4 is a flow chart of a method for risk prediction according to an embodiment of the present application;
FIG. 5 is a schematic diagram of an apparatus for risk prediction according to an embodiment of the present application;
FIG. 6 is a schematic diagram of another risk prediction apparatus according to an embodiment of the present application;
fig. 7 is a schematic diagram of an electronic device according to an embodiment of the present application.
Detailed Description
The risk prediction method, the risk prediction device, the storage medium and the electronic equipment provided by the application can be used in the big data field or the financial field. The foregoing is merely an example, and the application fields of the risk prediction method, the apparatus, the storage medium and the electronic device provided by the present application are not limited.
In order to make the person skilled in the art more clearly understand the solution of the present application, the following first describes an application scenario of the solution of the present application.
At present, when analyzing the customer risk of a target customer, conventionally, a high-risk customer label is obtained by collecting a customer record of a certain product violating or overdue, and modeling analysis is performed in combination with the historical violations of the target customer, but the customer characteristics and the customer behaviors of each product are often inconsistent, and the number of high-risk customers of each product type is often rare. And the historical violations of the target clients may be less or not, so that the training data of modeling analysis is less, and the accuracy of the prediction result is low.
In order to solve the above problems, the embodiments of the present application provide a risk prediction method, apparatus, storage medium, and electronic device. The method comprises the steps of firstly acquiring data of a first target domain and data of a plurality of first source domains. According to the client characteristics and the labels, the method determines each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold value. And then adding the data of each second source domain into training data of model training to obtain a risk prediction model. According to the scheme, the existing data are utilized to a large extent, the data domain with higher migration learning value is found by measuring the similarity of the data, the occurrence probability of negative migration is greatly reduced, the problem of unbalanced data types is solved, the available data of model training is expanded, the analysis efficiency and the model effect are improved, and the accuracy of a client risk prediction result can be improved.
According to the scheme, the existing data in each field are fully utilized, when the data and the target field are used, the analysis efficiency and the prediction effect of the prediction model are improved, the problems of insufficient data and unbalanced categories are also relieved, and the risk condition of a customer can be accurately predicted.
In order to make the solution of the present application more clearly understood by those skilled in the art, the following description will describe the solution of the present application in connection with the accompanying drawings in the embodiments of the present application.
The words "first," "second," and the like in the description of the application are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated.
The embodiment of the application provides a risk prediction method, and the risk prediction method is specifically described below with reference to the accompanying drawings.
Referring to fig. 1, a flowchart of a risk prediction method according to an embodiment of the present application is shown.
The method comprises the following steps:
s11: and acquiring the data of the first target domain and the data of a plurality of first source domains.
The traditional machine learning often needs to use a large amount of training data to train an established model, data labeling is needed, and the problem of unbalanced data is often solved by means of over-sampling, under-sampling and the like of the data, so that the original distribution characteristics of the data are destroyed.
In the prediction scene, the problem of unbalanced data is easy to occur, so that the scheme of the application integrates the client features and the labels in a plurality of other products through transfer learning, and further extracts and selects data capable of carrying out transfer learning from the client features and the labels.
Transfer learning is an important branch of machine learning, and focuses on applying knowledge transfer that has been learned to new problems, so as to enhance the ability to solve the new problems and increase the speed of solving the new problems. In the machine learning category, the transfer learning can apply models and knowledge learned in the old domain to the new domain using the similarity between data, tasks, or models. The core of the transfer learning is to find the similarity between the existing knowledge and the new knowledge.
Negative migration: generally, it is meant that one study interferes with or inhibits another study. Negative transfer is generally represented by one type of learning increasing the learning time or the number of exercises required for another type of learning or impeding the smooth progress of another type of learning and the correct grasp of knowledge.
Source domain/destination domain (destination domain), both the source domain and the destination domain are data domains, including data. In the transfer learning, the existing knowledge is called a source domain, and the new knowledge to be learned is called a target domain.
In the application, the data of the first target domain and each first source domain corresponds to historical transaction data, historical default data and customer information data of a class of products. That is, the data types corresponding to the first target domain and the first source domain are the same, and the difference is that the quantity of the data and the corresponding products are different.
The first target domain corresponds to a target product. Each first source domain corresponds to a non-target product, and the non-target products corresponding to each first source domain are different.
In practical applications, the total amount of data included in each first source domain is greater than the total amount of data included in the first destination domain.
S12: and determining each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold according to the client characteristics and the labels respectively corresponding to each first source domain and each first target domain.
The customer characteristics are used for representing customer attributes and transaction behaviors, the customer attributes can reflect qualification conditions of the customers, and the transaction behaviors of the customers mainly comprise historical transaction frequency, transaction times and the like of the customers for the product.
The tag is used to indicate the risk to the customer, i.e. the probability that the user may be surprised when using the product.
Customer characteristics and labels may be used as corresponding variables, e.g. X for customer characteristics i The representation is Y for label i I is a non-negative integer. The corresponding customer feature and label pair for the first source domain may be represented as (X 1 ，Y 1 ) The customer feature and label pairs corresponding to each second source domain may be represented in turn as (X 2 ，Y 2 )，(X 3 ，Y 3 )，…，(X n ，Y n ). n is a positive integer greater than 2.
The corresponding product of the first target domain is a target product, so as to avoid negative migration caused by directly utilizing the data of the first source domain. In the implementation of the application, a second source domain with the distribution of the client features and the labels close to the first source domain is selected from a plurality of first source domains, and then the data of the second source domain is utilized for migration learning.
In the embodiment of the application, when the similarity between the current first source domain and the first target domain is determined to be greater than or equal to the preset threshold according to the client characteristics and the labels of the current first source domain, the current first source domain is determined to be a second source domain.
And sequentially judging whether each first source domain can be determined to be the second source domain, and further obtaining one or more second source domains. The data of the second source domain may be applied for transfer learning, i.e. for model training.
The data of the first source domain with the similarity to the first target domain smaller than the preset threshold value can be omitted, because the client features and labels corresponding to the first source domain are greatly different from those of the first target domain, and negative migration may occur if the data of the first source domain is applied to model training.
S13: and training to obtain a risk prediction model, wherein the training data of the risk prediction model comprises data of each second source domain.
The specific type of the risk prediction model is not limited in the embodiments of the present application, and may be, for example, a fully connected neural network (Fully Connected Neural Network, FCNN) model, a convolutional neural network (Convolutional Neural Network, CNN) model, a recurrent neural network (Recurrent Neural Network, RNN) model, or other possible neural network models, which are not described herein.
When the model is trained, the training data comprise data of each second source domain, and as the distribution of client features and labels corresponding to the data of each second source domain is similar to that of the first target, the probability of negative migration is reduced, the expansion of the training data is realized, and the migration learning result is more accurate and reliable.
S14: and carrying out risk prediction on the target customer of the target product according to the risk prediction model.
After training, a risk prediction model is obtained, the client characteristics of the target client are used as the input of the model, and the risk of the client is predicted by using the risk prediction model.
In summary, by using the method provided by the embodiment of the present application, first, a first target domain and a plurality of first source domains are acquired. The first target threshold includes data corresponding to a target product, and the first source domain includes data corresponding to a non-target product. According to the client characteristics and the labels, the method determines each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold value. And then adding the data of each second source domain into training data of model training to obtain a risk prediction model. According to the scheme, the existing data are utilized to a large extent, the data domain with higher migration learning value is found by measuring the similarity of the data, the occurrence probability of negative migration is greatly reduced, the problem of unbalanced data types is solved, the available data of model training is expanded, the analysis efficiency and the model effect are improved, and the accuracy of a client risk prediction result can be improved.
The following description is made in connection with specific implementations. In the following description, the risk prediction model is taken as a fully connected neural network model as an example, and the principle is similar when the risk prediction model adopts other models, and will not be described herein.
Referring to fig. 2, a flowchart of another risk prediction method according to an embodiment of the present application is shown.
S21: and acquiring the data of the first target domain and the data of a plurality of first source domains.
The first target domain includes historical transaction data for the target product, historical breach data for the target product, and customer information data for the target product.
Each first source domain includes historical transaction data, historical breach data, and customer information data for a class of non-target products.
The customer information data may include customer attributes and customer transaction actions. The customer attributes are used to characterize the qualification of the customer of the product, and the transaction behavior of the customer is used to characterize the historical transaction frequency, the transaction times, etc. of the customer of the product.
In practical application, when acquiring data, the original data needs to be cleaned basically, the conditions of missing values, abnormal values and the like are processed, more client information data, historical transaction data, historical default data and the like are acquired as much as possible, and normalization and standardization processing are performed.
S22: and determining mutual information values of each first source domain and each first target domain respectively by taking the client characteristics and the labels as variables.
In the embodiment of the application, mutual information (mutual Information, MI) is used as a judgment basis for selecting a first source domain with the distribution of customer characteristics and labels similar to that of a first target domain.
MI measures the degree of interdependence between two variables. Specifically, for two random variables, MI is the "amount of information" that one random variable decreases due to the knowledge of the other random variable.
X for customer feature i The representation is Y for label i I is a positive integer.
An implementation of determining mutual information between the first target domain and the first source domain is described below.
The customer feature and label pair corresponding to the first target domain may be represented as (X 1 ，Y 1 ). The client features and tag pairs corresponding to the first source domain may be represented in turn as (X 2 ，Y 2 ) N is a positive integer. X1 is a set of customer features of the first target domain, Y1 is a set of labels of the first target domain; x2 is the set of client features of the first source domain and Y1 is the set of labels of the first source domain.
The mutual information value I (X; Y) can be determined by:
wherein x is 1 For a client feature of the first target domain, x 2 For a client feature of the first source domain, y 1 Is a tag feature, y, of the first target domain 2 Is a label of the first source domain, p (x 1 ,x 2 ) For joint probability mass of customer features in a first target domain and customer features in a first source domainFunction, p (y 1 ,y 2 ) For a joint probability mass function of a tag in a first target domain and a tag in a first source domain, p (x 1 ) An edge probability density function, p (x 2 ) An edge probability density function, p (y 1 ) An edge probability density function, p (y 2 ) Is a function of the edge probability density of the tag of the first source domain.
The manner of determining the mutual information value of the first target domain is similar to that of the other first source domains, and will not be described in detail herein.
The larger the mutual information value, the higher the similarity characterizing the first target domain and the first source domain at that time.
S23: and determining the mutual information value as similarity, and determining a first source domain corresponding to the mutual information value larger than a preset threshold value as a second source domain.
And determining a second source domain meeting the requirements by sequentially determining mutual information values of the first target domain and each first source domain.
It will be appreciated that the above use of mutual information values as similarity is only one possible implementation. In practical applications, the similarity may be measured in other manners, for example, determining a vector distance between a vector formed by the client feature and the tag in the first target domain and a vector formed by the client feature and the tag in the first source domain, when the obtained vector distances are added to obtain a vector distance sum, and when the vector distance sum is less than or equal to a preset threshold value, determining the first source domain at this time as a second source domain.
S24: and training the fully-connected neural network FCNN model by taking the data of the first target domain and the data of each second source domain as training data.
Referring to fig. 3, a schematic diagram of a fully connected neural network model according to an embodiment of the present application is shown.
The fully connected neural network (Fully Connected Neural Network, FCNN) model includes an input layer 31, a hidden layer 32, and an output layer 33, each of which is connected by a weight matrix.
Assuming that the input layer 31, the hidden layer 32 and the output layer 33 have j layers in total, j is an integer greater than 2, for the j-1 layer and the j layer, any node of the j-1 layer is connected to all nodes of the j layer. I.e. each node of the j-th layer, when performing the calculation, the input to the activation function is the weighting of all nodes of the j-1 layer.
The number of hidden layers 32 may be one or more, and only an implementation in which the number of hidden layers 32 is 5 is shown in fig. 3.
In the embodiment of the present application, in order to improve accuracy of risk prediction, the hidden layers of the FCNN model include k layers, where k is an integer greater than 1, that is, the number of hidden layers 32 is set to be multiple layers, and in the following description, the number of hidden layers is taken as 5 layers as an example.
When the FCNN model is trained, the training data is expanded because the training data is the data of the first target domain and the data of each second source domain.
And when the model is trained, randomly initializing parameters in the fully-connected neural network model, and updating the fully-connected neural network model by using a gradient descent algorithm. The specific process of training the fully connected neural network model is not described in detail herein.
S25: and updating the last m layers of the hidden layers of the FCNN model after training by using the data of the first target domain.
m is a positive integer less than k.
Let m be 2 as an example. And at this time, after the FCNN model training is completed, fixing parameters of the first three hidden layers, and updating parameters of the last two layers only by using data of the first target domain.
According to the scheme, the training data of the model is expanded by utilizing transfer learning, although the distribution of the client features and the labels of the second source domain is close to that of the first target domain, products corresponding to the data of the second source domain are non-target products, and certain deviation is possible when the model obtained through training is directly used for risk prediction of clients of target products, so that the parameter updating of hidden layers of the last two layers is realized by using the data of the first target domain, the calibration of applicable products of the model is realized, and the risk prediction effect of FCNN model on users of target products is more accurate.
The implementation of determining m in the embodiments of the present application is described below.
In one possible implementation, the value of m is a fixed value, for example, the value of m is 1 or 2.
In another possible implementation, the value of m is related to a proportional value between the number of data of the first destination and the sum of the number of data of each second source domain. The correspondence between the scale values and m may be determined in advance and stored in the form of a data table. And then determining the value of m according to the proportional value between the sum of the number of the data of the first target domain and the number of the data of each second source domain trained by the model and the corresponding relation. The following is an example.
Table 1: correspondence table between scale value and m
When the ratio between the number of the first target and the data and the sum of the number of the data of each second source domain is 0.3, the value of m is 2 in the interval of (0.2, 0.8), that is, the last 2 layers of the hidden layers of the FCNN model after training are updated by using the data of the first target domain, the correspondence shown in table 1 is only for convenience of description, and does not limit the technical scheme of the present application.
S26: and taking the updated FCNN model as a risk prediction model.
S27: and carrying out risk prediction on the target customer of the target product according to the risk prediction model.
Taking the client characteristics of the target client as the input of a model, and predicting the client risk by using the risk prediction model.
S28: updating the data in the first target domain to obtain the second target domain.
After the risk prediction model is obtained in S26, after a certain period, since the historical transaction data, the historical default data and the customer information data of the target product are expanded, in order to further improve the accuracy of risk prediction, the risk prediction model needs to be updated.
The method comprises the steps of firstly updating data in a first target domain, wherein the updating generally refers to adding newly generated data to original data, so as to realize the expansion of the data in the first target domain, and the expanded first target domain is a second target domain.
S29: and updating the last m layers of the hidden layers of the risk prediction model by using the data of the second target domain to obtain an updated risk prediction model.
In the embodiment of the application, the initial training data of model training is sufficient by adopting the migration learning method, so that only the data of the last m layers are required to be updated, thereby fine tuning the model, saving the computing power resource and time expenditure, and enabling the model to enable the risk prediction effect of the FCNN model on the users of the target product to be more accurate.
In summary, the method provided by the embodiment of the application utilizes the existing data to a greater extent, finds the data domain with higher migration learning value by measuring the similarity of the data, greatly reduces the occurrence probability of negative migration, thus alleviating the problem of unbalanced data category, expanding the available data of model training, improving the analysis efficiency and the model effect, and improving the accuracy of the client risk prediction result. And the updated data of the second target domain is used for updating the hidden layers with fewer layers, so that the risk prediction model can be updated conveniently, and the computational resource and the time cost are saved.
Referring to fig. 4, a flowchart of a method for risk prediction according to an embodiment of the present application is shown.
S31: and acquiring the data of the first target domain and the data of a plurality of first source domains.
S32: and determining mutual information values of each first source domain and each first target domain respectively by taking the client characteristics and the labels as variables.
S33: and determining the mutual information value as similarity, and determining a first source domain corresponding to the mutual information value larger than a preset threshold value as a second source domain.
S34: and training the fully-connected neural network FCNN model by taking the data of each second source domain as training data.
The hidden layer of the FCNN model includes k layers, k being an integer greater than 1.
The embodiment of the present application differs from the method shown in fig. 2 in that: in the training of the FCNN model, the training data in this embodiment is the data of the second source domain, and since the number of data of the second source domain is generally greater than the number of data of the first target domain, the training data is expanded. In the embodiment of the present application, in the subsequent S35, the data of the first target domain may be used to update the trained FCNN model, so that the data of the first target domain may not be used when the FCNN model is previously trained.
S35: and updating the last m layers of the hidden layers of the FCNN model after training by using the data of the first target domain.
m is a positive integer less than k.
S36: and taking the updated FCNN model as a risk prediction model.
S37: and carrying out risk prediction on the target customer of the target product according to the risk prediction model.
S38: updating the data in the first target domain to obtain the second target domain.
S39: and updating the last m layers of the hidden layers of the risk prediction model by using the data of the second target domain to obtain an updated risk prediction model.
For specific descriptions of S31-S33 and S35-S39, reference may be made to the descriptions in the corresponding embodiment of fig. 2, and the embodiments of the present application are not repeated here.
Based on the risk prediction method provided by the embodiment, the embodiment of the application also provides a risk prediction device, and the risk prediction device is specifically described below with reference to the accompanying drawings.
Referring to fig. 5, a schematic diagram of an apparatus for risk prediction according to an embodiment of the present application is shown.
The device comprises: an acquisition unit 51, a determination unit 52, a model training unit 53 and a risk prediction unit 54.
The acquiring unit 51 is configured to acquire data of a first target domain and data of a plurality of first source domains.
The first target domain and the data of each first source domain correspond to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product.
The determining unit 52 determines each second source domain having a similarity with the first target domain greater than or equal to a preset threshold according to the client feature and the label corresponding to each first source domain and the first target domain.
The tag is used to indicate customer risk.
The model training unit 53 is used for training the acquisition risk prediction model. The training data of the risk prediction model comprises data of each second source domain.
The risk prediction unit 54 is configured to perform risk prediction on a target customer of a target product according to a risk prediction model.
In a possible implementation manner, the determining unit 52 is specifically configured to determine, using the client feature and the tag as variables, a mutual information value of each of the first source domains and the first target domain, respectively; and determining the mutual information value as the similarity, and determining a first source domain corresponding to the mutual information value larger than the preset threshold value as one second source domain.
In a possible implementation manner, the model training unit 53 is specifically configured to train the fully-connected neural network FCNN model by using the data of the first target domain and the data of each second source domain as training data, where the hidden layer of the FCNN model includes k layers, and k is an integer greater than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In a possible implementation manner, the model training unit 53 is specifically configured to train the fully-connected neural network FCNN model by using the data of each second source domain as training data, where a hidden layer of the FCNN model includes k layers, and k is an integer greater than 1; updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k; and taking the updated FCNN model as the risk prediction model.
In a possible implementation manner, the model training unit 53 is further specifically configured to determine a ratio value between the number of data of the first target domain and the sum of the numbers of data of the second source domains; and determining the value of m according to the ratio value and the corresponding relation between the predetermined ratio value and m.
Referring to fig. 6, a schematic diagram of another risk prediction apparatus according to an embodiment of the present application is shown.
The device further includes an updating unit 55, which is specifically configured to update the data in the first target domain to obtain a second target domain; and updating the last m layers of the hidden layers of the risk prediction model by utilizing the data of the second target domain so as to obtain an updated risk prediction model.
By using the device provided by the embodiment of the application, the first target domain and a plurality of first source domains are acquired through the acquisition unit. The first target threshold includes data corresponding to a target product, and the first source domain includes data corresponding to a non-target product. And the determining unit determines each second source domain with the similarity with the first target domain being greater than or equal to a preset threshold according to the client characteristics and the labels. The model training unit then adds the data of each second source domain to the training data of the model training to obtain a risk prediction model. The device utilizes the existing data to a large extent, finds the data field with higher migration learning value by measuring the similarity of the data, greatly reduces the occurrence probability of negative migration, thus alleviating the problem of unbalanced data category, expanding the available data of model training, improving the analysis efficiency and model effect, and improving the accuracy of the client risk prediction result.
The risk prediction apparatus includes a processor and a memory, where the acquisition unit 51, the determination unit 52, the model training unit 53, the risk prediction unit 54, the update unit 55, and the like are stored as program units, and the processor executes the program modules stored in the memory to implement the corresponding functions.
The processor includes a kernel, and the kernel fetches the corresponding program unit from the memory. The kernel can be provided with one or more methods for realizing the risk prediction by adjusting kernel parameters.
The embodiment of the application provides a storage medium, on which a program is stored, which when executed by a processor implements the method of risk prediction.
The embodiment of the application provides a processor which is used for running a program, wherein the risk prediction method is executed when the program runs.
The embodiment of the application provides an electronic device, and the electronic device is specifically described below with reference to the accompanying drawings.
Referring to fig. 7, a schematic diagram of an electronic device according to an embodiment of the present application is shown.
The electronic device 70 comprises at least one processor 701, at least one memory 702 connected to the processor 701, and a bus 703.
The processor 701 and the memory 702 communicate with each other via a bus 703.
The processor 701 is configured to invoke the program instructions in the memory 702 to perform the above-described method of analyzing the flow of funds. The electronic device herein may be a server, a PC, a PAD, a mobile phone, etc.
The application also provides a computer program product adapted to perform a computer program for initializing the method of risk prediction as above when executed on a data processing device.
The risk prediction method, the risk prediction device, the risk prediction storage medium and the risk prediction electronic equipment provided by the application can be used in the technical field of big data, the technical field of computers or the financial field. The foregoing is merely an example, and the application fields of the risk prediction method, the apparatus, the storage medium and the electronic device provided by the present application are not limited.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
In one typical configuration, the device includes one or more processors (CPUs), memory, and a bus. The device may also include input/output interfaces, network interfaces, and the like.
The memory may include volatile memory, random Access Memory (RAM), and/or nonvolatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM), among other forms in computer readable media, the memory including at least one memory chip. Memory is an example of a computer-readable medium.
Computer readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media for a computer include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, does not include transitory computer-readable media (transmission media), such as modulated data signals and carrier waves.
It should be understood that in the present application, "at least one (item)" means one or more, and "a plurality" means two or more. "and/or" for describing the association relationship of the association object, the representation may have three relationships, for example, "a and/or B" may represent: only a, only B and both a and B are present, wherein a, B may be singular or plural.
In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. The apparatus embodiments described above are merely illustrative, wherein the units and modules illustrated as separate components may or may not be physically separate. In addition, some or all of the units and modules can be selected according to actual needs to achieve the purpose of the embodiment scheme. Those of ordinary skill in the art will understand and implement the present application without undue burden.
The foregoing is merely illustrative of the embodiments of this application and it will be appreciated by those skilled in the art that variations and modifications may be made without departing from the principles of the application, and it is intended to cover all modifications and variations as fall within the scope of the application.
Claims (10)
1. A method of risk prediction, the method comprising:
acquiring data of a first target domain and data of a plurality of first source domains, wherein the data of the first target domain and each first source domain correspond to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product;
according to the client characteristics and labels corresponding to the first source domain and the first target domain respectively, determining second source domains with similarity to the first target domain being greater than or equal to a preset threshold value, wherein the labels are used for indicating client risks;
training to obtain a risk prediction model, wherein training data of the risk prediction model comprises data of each second source domain;
and carrying out risk prediction on the target client of the target product according to the risk prediction model.
2. The risk prediction method according to claim 1, wherein the determining, according to the client characteristics and the labels corresponding to each of the first source domain and the first target domain, each second source domain having a similarity with the first target domain greater than or equal to a preset threshold value specifically includes:
determining mutual information values of each first source domain and each first target domain by taking the client features and the labels as variables;
And determining the mutual information value as the similarity, and determining a first source domain corresponding to the mutual information value larger than the preset threshold value as one second source domain.
3. The method of risk prediction according to claim 1, wherein the training obtains a risk prediction model, specifically comprising:
training a fully connected neural network FCNN model by taking the data of the first target domain and the data of each second source domain as training data, wherein a hidden layer of the FCNN model comprises k layers, and k is an integer larger than 1;
updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k;
and taking the updated FCNN model as the risk prediction model.
4. The method of risk prediction according to claim 1, wherein the training obtains a risk prediction model, specifically comprising:
training the FCNN model by taking the data of each second source domain as training data, wherein the hidden layer of the FCNN model comprises k layers, and k is an integer greater than 1;
updating the last m layers of the hidden layers of the FCNN model after training by utilizing the data of the first target domain, wherein m is a positive integer smaller than k;
And taking the updated FCNN model as the risk prediction model.
5. The method of risk prediction according to claim 3 or 4, further comprising:
updating the data in the first target domain to obtain a second target domain;
and updating the last m layers of the hidden layers of the risk prediction model by utilizing the data of the second target domain so as to obtain an updated risk prediction model.
6. The method of risk prediction according to claim 3 or 4, wherein before updating the last m layers of the trained hidden layers of the FCNN model with the data of the first target domain, the method further comprises:
determining a ratio value between the number of data of the first target domain and the sum of the number of data of each second source domain;
and determining the value of m according to the ratio value and the corresponding relation between the predetermined ratio value and m.
7. The method of risk prediction according to claim 1, wherein the customer characteristics include one or more of the following:
customer attributes and customer transaction behavior; the customer transaction actions include transaction times and transaction frequency.
8. An apparatus for risk prediction, the apparatus comprising: the system comprises an acquisition unit, a determination unit, a model training unit and a risk prediction unit;
the acquisition unit is used for acquiring data of a first target domain and data of a plurality of first source domains, wherein the data of the first target domain and each first source domain corresponds to historical transaction data, historical default data and customer information data of a class of products, and the first target domain corresponds to a target product;
the determining unit is configured to determine, according to the client characteristics and the labels corresponding to each of the first source domain and the first target domain, each second source domain with a similarity greater than or equal to a preset threshold value with respect to the first target domain, where the labels are used to indicate client risks;
the model training unit is used for training and obtaining a risk prediction model, and training data of the risk prediction model comprise data of each second source domain;
and the risk prediction unit is used for predicting the risk of the target customer of the target product according to the risk prediction model.
9. A storage medium having stored thereon a computer program which, when executed by a processor, implements the method of risk prediction of any of claims 1-7.
10. An electronic device for running a program, wherein the program is operative to perform the method of risk prediction between financial variables as claimed in any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310619716.9A CN116629612A (en) | 2023-05-29 | 2023-05-29 | Risk prediction method and device, storage medium and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310619716.9A CN116629612A (en) | 2023-05-29 | 2023-05-29 | Risk prediction method and device, storage medium and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116629612A true CN116629612A (en) | 2023-08-22 |
Family
ID=87596987
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310619716.9A Pending CN116629612A (en) | 2023-05-29 | 2023-05-29 | Risk prediction method and device, storage medium and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116629612A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117893215A (en) * | 2024-03-18 | 2024-04-16 | 花瓣支付(深圳)有限公司 | Risk control method, electronic device, server and storage medium |
-
2023
- 2023-05-29 CN CN202310619716.9A patent/CN116629612A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117893215A (en) * | 2024-03-18 | 2024-04-16 | 花瓣支付(深圳)有限公司 | Risk control method, electronic device, server and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| AU2021232839B2 (en) | Updating Attribute Data Structures to Indicate Trends in Attribute Data Provided to Automated Modelling Systems | |
| Ma et al. | Multi-valued collaborative QoS prediction for cloud service via time series analysis | |
| EP3923207A2 (en) | Clustering techniques for machine learning models | |
| US20170323216A1 (en) | Determining retraining of predictive models | |
| CN111181757B (en) | Information security risk prediction method and device, computing equipment and storage medium | |
| CN108830443A (en) | A kind of contract review method and device | |
| CN116629612A (en) | Risk prediction method and device, storage medium and electronic equipment | |
| CN110866698A (en) | Device for assessing service score of service provider | |
| CN113988458A (en) | Anti-money laundering risk monitoring method and model training method, device, equipment and medium | |
| US20220291840A1 (en) | Value-based replication of streaming data | |
| CN113850669A (en) | User grouping method and device, computer equipment and computer readable storage medium | |
| US20220129318A1 (en) | Quota Request Resolution On Computing Platform | |
| AU2022360356A1 (en) | Data compression techniques for machine learning models | |
| CN113743678B (en) | User credit score prediction method and related equipment | |
| CN112364181B (en) | Insurance product matching degree determining method and apparatus | |
| CN112214387B (en) | Knowledge graph-based user operation behavior prediction method and device | |
| Wang et al. | A DBN-Based Ensemble Method for Resource Usage Prediction in Clouds | |
| US20230136564A1 (en) | Agent assisted model development | |
| CN115409636A (en) | Product risk prediction method, device, equipment and medium | |
| CN113468142A (en) | Data transfer request processing method, device, equipment and storage medium | |
| CN113723710A (en) | Customer loss prediction method, system, storage medium and electronic equipment | |
| CN115439174A (en) | Method and device for determining recommendation information | |
| CN113743678A (en) | User credit score prediction method and related equipment | |
| CN114925770A (en) | Scene analysis method and device, storage medium and electronic equipment | |
| CN113094654A (en) | Product data processing method, device and equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |