CN116452320B - Credit risk prediction method based on continuous learning - Google Patents
Credit risk prediction method based on continuous learning Download PDFInfo
- Publication number
- CN116452320B CN116452320B CN202310385038.4A CN202310385038A CN116452320B CN 116452320 B CN116452320 B CN 116452320B CN 202310385038 A CN202310385038 A CN 202310385038A CN 116452320 B CN116452320 B CN 116452320B
- Authority
- CN
- China
- Prior art keywords
- credit risk
- training
- risk prediction
- model
- round
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000012549 training Methods 0.000 claims abstract description 152
- 238000012545 processing Methods 0.000 claims abstract description 13
- 238000013058 risk prediction model Methods 0.000 claims description 64
- 230000006870 function Effects 0.000 claims description 18
- 238000011156 evaluation Methods 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 10
- 238000007781 pre-processing Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 8
- 238000003062 neural network model Methods 0.000 claims description 3
- 238000009826 distribution Methods 0.000 abstract description 8
- 230000003111 delayed effect Effects 0.000 abstract description 3
- 238000012502 risk assessment Methods 0.000 abstract description 3
- 238000005516 engineering process Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000007596 consolidation process Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000012466 permeate Substances 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000012360 testing method Methods 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
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/03—Credit; Loans; Processing thereof
-
- 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
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Technology Law (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
The invention relates to a credit risk prediction technology, and discloses a credit risk prediction method based on continuous learning, which can continuously enhance the performance of a model and improve the accuracy of prediction under the condition of limited sample size caused by privacy protection. According to the invention, a continuous learning strategy is adopted to train the prediction model of each task, after the model of the previous task is trained, the parameter knowledge of the model is extracted and transmitted to the next task, so that the model training of the next task is assisted; in addition, three decisions are integrated in the model application, namely three decision threshold pairs of the model are adaptively set according to the confidence distribution condition of the task models on the sample prediction results in the corresponding task training data set, in the practical application, after the information to be predicted is received, the current latest model is adopted to output the risk assessment result of the information to be predicted and count the confidence, then the decision is executed according to the three decision threshold pairs, and the delayed decision processing on the uncertainty samples is realized, so that the credit risk is better controlled.
Description
Technical Field
The invention relates to a credit risk prediction technology, in particular to a credit risk prediction method based on continuous learning.
Background
In the scientific information age, big data is not only simple data, but also becomes an important production factor, and permeates every field nowadays. The generation, collection, storage, transmission and utilization of data are all over the industry, and while bringing convenience to us, highly informationized data also increase the risk of revealing personal private information, and personal privacy protection in the current big data age is becoming more important.
The trade object in the financial market is not common commodity, but money funds and derivatives thereof, and privacy protection has great significance in the financial market. In the field of credit loans, borrower credits are taken as a basis, and temporary separation or conditional yielding of the use right and ownership of funds is realized through a lending relationship and a proxy agent relationship. In this process, financial institutions such as banks use credit risk in exchange for future interest returns. The credit risk is also called default risk, which means the risk that a borrower, securities issuer, or transaction counterpart is unwilling or unable to fulfill contract conditions for various reasons to form default, and the bank, investor, or transaction counterpart is lost. The financial institution determines whether to lend to the borrower based on the predicted risk, and the general flow of loan approval is shown in fig. 1. Because of the direct relationship between the accuracy of credit risk prediction and the profit and loss of institutions, how to construct a credit risk assessment model has been an important research topic in the field of credit loans.
The conventional way to construct a credit risk assessment model is to construct static models from the existing data for each stage in the whole approval process, and then predict sample categories using these static models, see fig. 2. In the financial market where the sample is limited due to the stricter personal privacy protection, the traditional prediction mode is greatly limited. Due to the specificity of the transaction objects and the transaction modes in the financial market, the privacy protection of the participants is particularly emphasized by the financial market supervisor. In the field of credit, third party lending institutions such as banks are subject to privacy protection constraints and cannot permanently retain borrower information that must be destroyed after a certain time, which results in a limited amount of sample available for training the model. In the context of privacy protection, how to continuously enhance the model performance based on a small number of samples, the conventional credit risk prediction method has not solved the problem.
In addition, due to the current situation of privacy protection, the available data volume is small, and it is difficult to ensure the prediction accuracy of the constructed model. The traditional credit risk prediction method adopts two decisions, namely directly dividing samples into a certain category, which neglects the control of misclassification of the samples, and thus the target institution still faces higher credit risk.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: the credit risk prediction method based on continuous learning is provided, and under the condition that the sample size is limited due to privacy protection, the performance of a model is continuously enhanced, and the prediction accuracy is improved. And further control the uncertainty of the prediction result, facilitating the financial institution to control the credit risk faced.
The technical scheme adopted for solving the technical problems is as follows:
a credit risk prediction method based on continuous learning comprises the following steps:
A1, constructing a credit risk prediction model and initializing;
A2, constructing a training data set of the current training round, and training a credit risk prediction model based on the training data set of the current training round;
A3, executing a credit risk prediction task by using a credit risk prediction model of the current training round obtained by training;
a4, judging whether the set model updating period is reached, if so, returning to the step A2, entering the next training round, otherwise, returning to the step A3;
wherein, step A2 comprises the following steps:
A21, judging whether the training is a first training task, if so, executing a step A22, otherwise, executing a step A23;
A22, preprocessing the current existing sample data to construct a training data set of a first training round; training the initialized credit risk prediction model by using a training data set of the first training round to obtain a credit risk prediction model corresponding to the first training round, and executing the step A24;
A23, preprocessing newly added sample data in an updating period time period to obtain a training data set of a current training round; performing incremental training on the credit risk prediction model obtained from the previous training round by using the training data set of the current training round, performing parameter optimization by using parameter knowledge extracted from the credit risk prediction model obtained from the previous training round in the incremental training process to obtain a credit risk prediction model corresponding to the current training round, and executing step A24;
A24, extracting parameter knowledge of a credit risk prediction model of the current training turn obtained by training.
Further, in step A1, the credit risk prediction model uses a fully connected neural network model including an input layer, a plurality of hidden layers, and an output layer.
Further, in steps a22 and a23, the preprocessing includes: processing missing data, deleting redundant features, constructing new features and balancing data categories.
Further, in step a24, the parameter knowledge of the credit risk prediction model of the current training round obtained by training is extracted by using an EWC (ELASTIC WEIGHT Consolidation ) method.
Further, in step A4, the model update period is a fixed time period or is an update period when the current sample data available for incremental training reaches a certain data amount.
Further, in step a23, the performing parameter optimization by using the parameter knowledge extracted from the credit risk prediction model obtained from the previous training round specifically includes:
Comparing the parameters of the model in the current round of training task with the parameter knowledge extracted from the credit risk prediction model obtained from the previous round of training task, incorporating the difference between the parameters into the loss function of model parameter optimization of the current round of training task, and enabling the parameters with larger weights in the credit risk prediction model obtained from the previous round of training task to keep larger weights in the loss function in the current round of training task.
Further, the step A3 includes the following steps:
A31, taking credit risk information to be predicted as input, and outputting an evaluation result of the credit risk category by using a credit risk prediction model of the current training turn obtained by training;
A32, dividing credit risk information to be predicted into a positive domain, a negative domain or a boundary domain according to the confidence level of the credit risk category assessment result and by combining three decision threshold pairs of a preset credit risk prediction model for the current training round obtained by training, and executing the following decisions:
If the decision is divided into the positive domain, a preset decision corresponding to the low credit risk is executed, if the decision is divided into the negative domain, a preset decision corresponding to the high credit risk is executed, if the decision is divided into the boundary domain, the credit risk prediction is executed again on the credit risk information to be predicted after the credit risk prediction model is trained for the next round.
Further, in step a32, the method for presetting three decision threshold pairs of the credit risk prediction model for the current training round includes:
Sample data in a training data set of a current training round are input into a credit risk prediction model of the current training round obtained through training, credit risk category assessment results for all sample data are obtained, confidence degrees of all assessment results are counted, and then the confidence degrees of all assessment results are ranked;
Obtaining two confidence boundary points from the ordered confidence sequence according to the proportion of the preset sample data quantity entering the boundary domain to the total input sample data quantity;
And taking the confidence boundary point with relatively lower confidence in the two confidence boundary points as a lower threshold value, and taking the confidence boundary point with relatively higher confidence in the two confidence boundary points as an upper threshold value, wherein the lower threshold value and the upper threshold value form three decision threshold value pairs.
Further, in step a32, the dividing the credit risk information to be predicted into a positive domain, a negative domain or a boundary domain according to the confidence level of the credit risk category evaluation result and in combination with three decision threshold pairs of the preset credit risk prediction model for the current training round obtained by training specifically includes:
The confidence level of the credit risk category assessment result of the credit risk information to be predicted is compared with three decision threshold pairs of the current latest credit risk prediction model, and the credit risk information to be predicted is divided according to the comparison condition:
If the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is higher than the upper threshold value in the three decision threshold value pairs, dividing the credit risk information to be predicted into a positive domain;
If the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is lower than the lower threshold value of the three decision threshold value pairs, dividing the credit risk information to be predicted into a negative domain;
and if the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is between the lower limit threshold and the upper limit threshold of the three decision threshold pairs, dividing the credit risk information to be predicted into a boundary domain.
The beneficial effects of the invention are as follows:
(1) The continuous learning is adopted to realize the transmission of knowledge between tasks, so that the defect that the traditional credit risk prediction can only rely on data to construct a model is overcome, and meanwhile, the problem that an ideal model cannot be constructed due to limited available samples under privacy protection is solved, so that the credit risk prediction capability of a financial institution is improved;
(2) By adding three decisions in the continuous learning framework and adaptively and dynamically setting decision thresholds according to the distribution characteristics of the data in each task, subjective influence caused by manually and directly setting the decision thresholds is avoided, and the three decisions are introduced, so that a financial institution can be helped to realize delayed decision processing on uncertain samples, and the credit risk faced by the financial institution is reduced.
Drawings
FIG. 1 is a general loan approval flow chart for a financial institution;
FIG. 2 is a diagram illustrating a conventional credit risk prediction method;
FIG. 3 is a schematic diagram of a continuous learning framework in an embodiment of the present invention;
FIG. 4 is a schematic diagram of three decisions in an embodiment of the present invention;
fig. 5 is a schematic diagram of a three-support learning framework in an embodiment of the invention.
Detailed Description
The invention aims to provide a credit risk prediction method based on continuous learning, which can continuously enhance the performance of a model and improve the accuracy of prediction under the condition of limited sample size caused by privacy protection. And further control the uncertainty of the prediction result, facilitating the financial institution to control the credit risk faced.
In the process of training a credit risk prediction model, firstly, a credit risk prediction model is constructed and initialized, then, each round of training model is used as a training task, the model is trained by constructing a training data set of the current training round, wherein for the first training task, the current existing sample data of a target financial institution are directly preprocessed to obtain a corresponding training data set, and for the subsequent other training tasks, the new sample data in a time period of an update period (the time interval between the last training round and the current training round) are preprocessed to obtain the corresponding training data set. And after each round of training is completed, the parameter knowledge of the extracted model is transferred to the next task, and the next task performs parameter optimization by combining the parameter knowledge extracted from the credit risk prediction model obtained from the previous training task in the process of training the model, so as to guide the updating of the model, thereby realizing the continuous learning of the model. When the model is applied, three decisions are integrated, namely, three decision threshold pairs corresponding to the current task model are adaptively set according to the confidence coefficient distribution condition of the sample prediction result in the corresponding training data set by the current task model, when the credit risk information to be predicted is received, the risk category assessment result of the credit risk information to be predicted is obtained by using the current task model (the latest model), the confidence coefficient of the assessment result is counted, and then the confidence coefficient is compared with the three decision threshold pairs corresponding to the model, so that three decision tasks of the credit risk information to be predicted are executed: if the decision is divided into a positive domain, a preset decision corresponding to a low credit risk is executed (e.g. through application), if the decision is divided into a negative domain, a preset decision corresponding to a high credit risk is executed (e.g. application is refused), if the decision is divided into a boundary domain, the credit risk prediction model is waited for to execute credit risk prediction on the credit risk information to be predicted again after the next round of training, and thus, delay decision processing of an uncertainty sample is realized so as to better control the credit risk faced by a financial institution.
And (3) continuously learning (Continual Learning), accumulating knowledge of each stage based on a deep neural network (Deep Neural Network, DNN), and constructing a model with more stable performance, so that the model can help future tasks to obtain better effects and can also have better performance on the previous tasks. Continuous learning breaks the traditional machine learning paradigm of "modeling a single task in isolation," but instead emphasizes overcoming knowledge learned by forgotten past tasks in a series of tasks and passes that knowledge to upcoming tasks. In recent years, continuous learning has been widely applied to the fields of emotion classification, image recognition, man-machine interaction, and the like, and continuous learning has not been studied to be applied to credit risk prediction. Current credit risk predictions are still modeled based only on existing data, without consideration of the knowledge contained in the use of existing models through continuous learning. Under privacy protection scene, borrower's information can not be preserved forever, leads to the fact in the past most data can not be used, and traditional mode based on data modeling alone is no longer effective. For this purpose, knowledge contained in the existing model is transmitted by means of a continuous learning framework to assist the modeling of the next stage. In a privacy protection scene, in order to balance two factors of model effect and resource consumption, the invention adopts an elastic weight consolidation (ELASTIC WEIGHT Consolidation, EWC) based regularization method to realize continuous learning. According to the method, the mode of restraining the change amplitude of the important parameters is adopted, the learned knowledge in the previous task is transferred to the next task, and a solution is provided for improving the performance of the credit risk prediction model under privacy protection.
Meanwhile, the uncertainty of the prediction result is controlled through three decisions, so that the credit risk faced by a financial institution is reduced. The softMax function in the neural network model is used as an evaluation function of three decisions, so that the objectivity of the evaluation function is ensured, and the combination of continuous learning and three decisions is innovatively realized. Moreover, unlike the traditional three-branch decision which always adopts a constant decision threshold, the invention combines the data distribution of each task to dynamically make a proper decision threshold for each task, thereby realizing the effective control of an uncertainty sample.
Examples:
The credit risk prediction method based on continuous learning in this embodiment includes two parts, namely training a credit risk prediction model and executing a credit risk prediction task by using the model, specifically described as follows:
1. training a credit risk prediction model:
The training risk prediction model adopts a continuous learning strategy, each round of model training is used as a training task, and the whole continuous learning training process comprises the following steps:
1. First training task:
(1) Constructing a credit risk prediction model:
A fully connected neural network is constructed that includes an input layer, a plurality of hidden layers, and an output layer. After a series of linear processing and activation processing are performed on the input object x through the weight matrix theta and the bias vector theta, softMax processing is performed, and finally a final result is output through an output layer. The activation operation of SoftMax for the kth neuron can be represented by the following equation:
wherein e (·) represents an exponential function; n L is the number of neurons on the output layer L, i.e. the number of categories of category labels; is the value of the kth output on output layer L; /(I) Representing the probability that the sample belongs to the kth class after SoftMax activation processing.
(2) Training a credit risk prediction model of the first task:
preprocessing currently existing sample data which can be used for training when a financial institution trains a credit risk prediction model of a first task to obtain a training sample set; the data preprocessing mode includes, but is not limited to, processing missing data, deleting redundant features, constructing new features, balancing data categories, etc., and aims to provide for training a high-quality model.
And training the credit risk prediction model by adopting a back propagation mode based on the training sample set to obtain a credit risk prediction model corresponding to the first task.
(3) Extracting parameter knowledge of a model:
In the embodiment, the EWC method is used to extract the parameter knowledge of the model, and in the privacy protection context, the EWC method uses the importance degree of the parameters in the model as knowledge, and the information of the importance of the parameters of the model is transmitted between the models. This knowledge is passed on to the next task, limiting the magnitude of the change in the important parameters to guide the next task in model training, with reference to this knowledge.
2. Model continuous training:
(1) When the model updating period set by the target financial institution arrives, preprocessing newly added sample data which can be used for training in the updating period time period to obtain a training data set of a current model training task;
The update period may be a fixed time period (for example, 1 week or 1 month), and since the number of sample data obtained in each fixed time period may be different, for example, the number of sample data newly added in a certain period is smaller, and the number of sample data newly added in a certain period is larger, in order to update the model in time to obtain a more accurate prediction, the update period may be used when the number of sample data currently available for incremental training reaches a certain data amount. The data preprocessing mode includes, but is not limited to, processing missing data, deleting redundant features, constructing new features, balancing data categories, and the like.
(2) Performing incremental training on the credit risk prediction model obtained by the previous training task by utilizing a training data set of the current model training task, performing parameter optimization by utilizing parameter knowledge extracted from the credit risk prediction model obtained by the previous training task in the incremental training process, obtaining a credit risk prediction model corresponding to the current training task, and extracting the parameter knowledge of the model;
The parameter optimization is performed by utilizing the parameter knowledge extracted from the credit risk prediction model obtained from the previous training task, and specifically comprises the following steps: and comparing the parameters of the model in the current training task with the parameter knowledge extracted from the credit risk prediction model obtained from the previous task, incorporating the difference between the parameters into a loss function of the model parameter optimization of the current task, and enabling the parameters with larger weight in the credit risk prediction model obtained from the previous task to keep larger weight in the loss function of the current task. When minimizing the loss function of the current task, the magnitude of the change in parameters is limited because the difference between the new and old task parameters is taken into account in the loss function. The more important parameters in the previous task have larger weights in the current loss function, and the influence on the loss function of the current task is larger than that of the general parameters. By minimizing the loss function, a severe limitation of the amplitude of variation of different parameters is achieved, thus enabling preservation of knowledge learned from previous tasks.
Taking two tasks A, B as an example, the total training data set D of the two tasks includes training data sets D A and D B corresponding to the two tasks, respectively. After the task A builds a model, the knowledge of the model is transmitted to the task B through the EWC. Weight information of model parametersAs knowledge, the knowledge is extracted by the task A through the EWC and is transmitted to the task B, and the task B parameter weight information/> islimitedThe specific principle of constructing the loss function is presented as follows:
① According to the bayesian formula, using the prior probability p (θ) of the parameter θ and the conditional probability p (d|θ), the conditional probability p (θ|d) can be calculated by:
log p(θ|D)＝log p(D|θ)+log p(θ)-log p(D)。
② Since dataset D is fully partitioned into task A and task B, the above equation is equivalent to:
log p(θ|D)＝log p(DB|θ)+log p(DA|θ)。
③ Since each subset affects the posterior probability, the loss function of EWC can be expressed as:
Wherein i represents each parameter; Representing total loss; /(I) Representing only the loss of task B; f i is laplace approximation information of the posterior probability calculated based on the data set D A; lambda is the relative importance of the task.
④ When task B builds a model, the loss of the parameter in task B is a part of the total loss, and the variation amplitude of the parameter relative to the parameter theta A of task A is another part of the total loss. The final selected parameter is to letAcquiring a minimum parameter, wherein in an ideal case, the task B has better performance; while the more important parameters in task a are the smaller the amplitude of the variation compared to the current parameters.
Therefore, the continuous learning is introduced to enable the performance of the prediction model to be more stable, the prediction effect to be more accurate, and a continuous learning framework constructed based on EWC is shown in FIG. 3.
2. Performing a credit risk prediction task using the model:
in the model specific application, for the target financial institution, application information is taken as input, and the evaluation result of the credit risk category can be output by using the current latest credit risk prediction model, and in order to control the credit risk of the target financial institution within a certain range, three decision processing uncertainty samples are adopted in the embodiment. The Three-way decision (Three-way decisions,3 WD) divides all samples into Three independent regions, i.e. positive, negative or boundary, according to the probability that the object belongs to a certain class, see fig. 4 for schematic diagram. For the samples in the boundary domain with lower confidence of the evaluation result, three decisions do not immediately assign the category to the samples, but the model updated later makes a judgment for the samples, and when the confidence of the samples meets the requirement, the samples are assigned the corresponding category. This mode provides the ability for the 3WD to handle sample uncertainty, which can handle more complex decision problems than two-way decisions.
The three-branch decision divides the whole set of samples Ω= { x 1,...,xn } into a Positive (POS), negative (NEG) or Boundary (BND) domain without crossing under the boundary threshold pair (β, α) and the evaluation function f (x), the sample division can be described as:
POS＝{x∈Ω|f(x)＞α},
NEG＝{x∈Ω|f(x)＜β},
BND＝{x∈Ω|β≤f(x)≤α}
the relationship between the positive, negative and boundary domains can be described as:
POS∩BND＝％,
BND∩NEG＝％,
NEG∩POS＝％,
Ω＝POS∪BND∪NEG
to avoid subjective impact of human set threshold pairs (β, α) on predictions, the present embodiment generates threshold pairs based on data characteristics. Because the data distribution of each task may change in a real scene, in this embodiment, in combination with the data distribution situation, a threshold pair conforming to the data distribution situation is adaptively generated for each task, so that the robustness of the three decision frames is improved, and the adaptive mode of the threshold pair is as follows:
(1) Setting the proportion p% of the boundary field sample to the total sample input by the corresponding task, wherein the larger the proportion is, the more attention is paid to the uncertainty of the prediction result, the more the target mechanism can strengthen the control of risks, and the target mechanism can adjust the proportion according to actual requirements, for example, when the application amount of applying loan is more, the proportion of the boundary field sample can be properly improved.
(2) According to the sample proportion of the corresponding task occupied by the preset boundary field sample of the task and the confidence level of the risk prediction model of the corresponding task on the training sample prediction result, three decision threshold pairs of the corresponding task are determined:
specifically, training sample data Te (k) of a task T (k) is put into a risk prediction model h k of the task where the training sample data Te (k) is located, and confidence degrees of sample risk categories are divided by a statistical model;
then, the confidence obtained is ordered from low to high to form a sequence Where N k represents the sample size of training sample Te (k) for task T (K);
Then, according to the predetermined sample boundary domain proportion p%, calculating the sample size N k =p%;
finally, the threshold value of the task T (k) is determined to be respectively the left boundary beta and the right boundary alpha And/>
Based on the above, the model for each task adaptively generates corresponding three decision threshold pairs.
In the model specific application, the current latest model is utilized to evaluate the risk category of the credit application of the client, the confidence level of the evaluation result is counted, and three decision-making threshold pairs corresponding to the current latest model are compared, so that three decision-making can be realized:
If the confidence coefficient of the credit risk category assessment result of the application information is higher than the upper limit threshold (right boundary) in the three decision threshold pairs, dividing the application information into positive domains, wherein the application information indicates that the risk is relatively smaller and the application can be passed;
If the confidence coefficient of the credit risk category assessment result of the application information is lower than the lower threshold (left boundary) in the three decision threshold pairs, dividing the application information into a negative domain, wherein the risk is relatively large, and the application can be refused;
if the confidence level of the credit risk category assessment result of the application information is between the lower limit threshold and the upper limit threshold in the three decision threshold pairs, the application information is divided into a boundary domain, which indicates that the risk is tentatively determined, the application can be delayed to be processed, and prediction processing is performed after the next model update.
In summary, as shown in fig. 5, the framework integrated with continuous learning and three decisions provided in this embodiment is configured to train a prediction model by using a corresponding training data set for each training task, extract parameter knowledge of the model, transmit the parameter knowledge to the next training task, and adaptively set three decision threshold pairs according to the confidence level distribution condition of risk category assessment results predicted by the model according to data in the training data set, when a credit application to be evaluated (test data in a corresponding graph) is received in actual application of the model, execute three decisions according to the confidence level of the model on risk category assessment results of the application information and the comparison condition of the three decision threshold pairs, if the application enters a positive domain, the application is refused if the application enters a negative domain, and delay the decision if the application information enters a boundary domain, thereby controlling the risk.
Although the application has been described herein with reference to the above examples, which are only preferred embodiments of the present application, the embodiments of the present application are not limited by the above examples, and it should be understood that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the scope and spirit of the principles of this disclosure.
Claims (6)
1. A credit risk prediction method based on continuous learning, comprising the steps of:
A1, constructing a credit risk prediction model and initializing; the credit risk prediction model adopts a fully-connected neural network model comprising an input layer, a plurality of hidden layers and an output layer;
A2, constructing a training data set of the current training round, and training a credit risk prediction model based on the training data set of the current training round;
A3, executing a credit risk prediction task by using a credit risk prediction model of the current training round obtained by training, wherein the method specifically comprises the following steps of A31-A32:
A31, taking credit risk information to be predicted as input, and outputting an evaluation result of the credit risk category by using a credit risk prediction model of the current training turn obtained by training;
A32, dividing credit risk information to be predicted into a positive domain, a negative domain or a boundary domain according to the confidence level of the credit risk category assessment result and by combining three decision threshold pairs of a preset credit risk prediction model for the current training round obtained by training, and executing the following decisions:
If the decision is divided into a positive domain, executing a preset decision corresponding to low credit risk, if the decision is divided into a negative domain, executing a preset decision corresponding to high credit risk, if the decision is divided into a boundary domain, waiting for the credit risk prediction model to execute credit risk prediction on the credit risk information to be predicted again after the next round of training;
a4, judging whether the set model updating period is reached, if so, returning to the step A2, entering the next training round, otherwise, returning to the step A3;
wherein, step A2 comprises the following steps:
A21, judging whether the training is a first training task, if so, executing a step A22, otherwise, executing a step A23;
A22, preprocessing the current existing sample data to construct a training data set of a first training round; training the initialized credit risk prediction model by using a training data set of the first training round to obtain a credit risk prediction model corresponding to the first training round, and executing the step A24;
A23, preprocessing newly added sample data in an updating period time period to obtain a training data set of a current training round; performing incremental training on the credit risk prediction model obtained from the previous training round by using the training data set of the current training round, performing parameter optimization by using parameter knowledge extracted from the credit risk prediction model obtained from the previous training round in the incremental training process to obtain a credit risk prediction model corresponding to the current training round, and executing step A24;
A24, extracting parameter knowledge of a credit risk prediction model of the current training turn obtained by training;
In step a23, the performing parameter optimization by using the parameter knowledge extracted from the credit risk prediction model obtained from the previous training round specifically includes:
Comparing the parameters of the model in the current round of training task with the parameter knowledge extracted from the credit risk prediction model obtained from the previous round of training task, incorporating the difference between the parameters into the loss function of model parameter optimization of the current round of training task, and enabling the parameters with larger weights in the credit risk prediction model obtained from the previous round of training task to keep larger weights in the loss function in the current round of training task.
2. A method for credit risk prediction based on continuous learning as claimed in claim 1,
In steps a22 and a23, the preprocessing includes: processing missing data, deleting redundant features, constructing new features and balancing data categories.
3. A method for credit risk prediction based on continuous learning as claimed in claim 1,
In step a24, the parameter knowledge of the credit risk prediction model of the current training round obtained by training is extracted by the EWC method.
4. A method for credit risk prediction based on continuous learning as claimed in claim 1,
In step A4, the model update period is a fixed time period or is used as an update period when the current sample data available for incremental training reaches a certain data amount.
5. A method for credit risk prediction based on continuous learning as claimed in claim 4,
In step a32, the method for presetting three decision threshold pairs of the credit risk prediction model for the current training round includes:
Sample data in a training data set of a current training round are input into a credit risk prediction model of the current training round obtained through training, credit risk category assessment results for all sample data are obtained, confidence degrees of all assessment results are counted, and then the confidence degrees of all assessment results are ranked;
Obtaining two confidence boundary points from the ordered confidence sequence according to the proportion of the preset sample data quantity entering the boundary domain to the total input sample data quantity;
And taking the confidence boundary point with relatively lower confidence in the two confidence boundary points as a lower threshold value, and taking the confidence boundary point with relatively higher confidence in the two confidence boundary points as an upper threshold value, wherein the lower threshold value and the upper threshold value form three decision threshold value pairs.
6. A method for credit risk prediction based on continuous learning as claimed in claim 5,
In step a32, the dividing the credit risk information to be predicted into a positive domain, a negative domain or a boundary domain according to the confidence level of the credit risk category evaluation result and in combination with three decision threshold pairs of the preset credit risk prediction model for the current training round obtained by training specifically includes:
The confidence level of the credit risk category assessment result of the credit risk information to be predicted is compared with three decision threshold pairs of the current latest credit risk prediction model, and the credit risk information to be predicted is divided according to the comparison condition:
If the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is higher than the upper threshold value in the three decision threshold value pairs, dividing the credit risk information to be predicted into a positive domain;
If the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is lower than the lower threshold value of the three decision threshold value pairs, dividing the credit risk information to be predicted into a negative domain;
and if the confidence coefficient of the credit risk category assessment result of the credit risk information to be predicted is between the lower limit threshold and the upper limit threshold of the three decision threshold pairs, dividing the credit risk information to be predicted into a boundary domain.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310385038.4A CN116452320B (en) | 2023-04-12 | 2023-04-12 | Credit risk prediction method based on continuous learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310385038.4A CN116452320B (en) | 2023-04-12 | 2023-04-12 | Credit risk prediction method based on continuous learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116452320A CN116452320A (en) | 2023-07-18 |
| CN116452320B true CN116452320B (en) | 2024-04-30 |
Family
ID=87123216
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310385038.4A Active CN116452320B (en) | 2023-04-12 | 2023-04-12 | Credit risk prediction method based on continuous learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116452320B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273912A (en) * | 2017-05-10 | 2017-10-20 | 重庆邮电大学 | A kind of Active Learning Method based on three decision theories |
| CN109670940A (en) * | 2018-11-12 | 2019-04-23 | 深圳壹账通智能科技有限公司 | Credit Risk Assessment Model generation method and relevant device based on machine learning |
| CN111652710A (en) * | 2020-06-03 | 2020-09-11 | 北京化工大学 | Personal credit risk assessment method based on ensemble tree feature extraction and Logistic regression |
| CN112580785A (en) * | 2020-12-18 | 2021-03-30 | 河北工业大学 | Neural network topological structure optimization method based on three-branch decision |
| CN112734038A (en) * | 2021-01-20 | 2021-04-30 | 清华大学 | Training method, medium, device and computing equipment for small sample continuous learning model |
| CN113177839A (en) * | 2021-05-20 | 2021-07-27 | 中国建设银行股份有限公司 | Credit risk assessment method, device, storage medium and equipment |
| CN113487403A (en) * | 2021-06-29 | 2021-10-08 | 百维金科(上海)信息科技有限公司 | Credit risk assessment system, method, device and medium |
| CN113792874A (en) * | 2021-09-08 | 2021-12-14 | 清华大学 | Continuous learning method and device based on innate knowledge |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107798390B (en) * | 2017-11-22 | 2023-03-21 | 创新先进技术有限公司 | Training method and device of machine learning model and electronic equipment |
-
2023
- 2023-04-12 CN CN202310385038.4A patent/CN116452320B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273912A (en) * | 2017-05-10 | 2017-10-20 | 重庆邮电大学 | A kind of Active Learning Method based on three decision theories |
| CN109670940A (en) * | 2018-11-12 | 2019-04-23 | 深圳壹账通智能科技有限公司 | Credit Risk Assessment Model generation method and relevant device based on machine learning |
| CN111652710A (en) * | 2020-06-03 | 2020-09-11 | 北京化工大学 | Personal credit risk assessment method based on ensemble tree feature extraction and Logistic regression |
| CN112580785A (en) * | 2020-12-18 | 2021-03-30 | 河北工业大学 | Neural network topological structure optimization method based on three-branch decision |
| CN112734038A (en) * | 2021-01-20 | 2021-04-30 | 清华大学 | Training method, medium, device and computing equipment for small sample continuous learning model |
| CN113177839A (en) * | 2021-05-20 | 2021-07-27 | 中国建设银行股份有限公司 | Credit risk assessment method, device, storage medium and equipment |
| CN113487403A (en) * | 2021-06-29 | 2021-10-08 | 百维金科(上海)信息科技有限公司 | Credit risk assessment system, method, device and medium |
| CN113792874A (en) * | 2021-09-08 | 2021-12-14 | 清华大学 | Continuous learning method and device based on innate knowledge |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116452320A (en) | 2023-07-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yan et al. | Research on financial assets transaction prediction model based on LSTM neural network | |
| WO2021155706A1 (en) | Method and device for training business prediction model by using unbalanced positive and negative samples | |
| Blanco et al. | Credit scoring models for the microfinance industry using neural networks: Evidence from Peru | |
| Shen et al. | A cost-sensitive logistic regression credit scoring model based on multi-objective optimization approach | |
| Ala’raj et al. | Modelling customers credit card behaviour using bidirectional LSTM neural networks | |
| CN110599336B (en) | Financial product purchase prediction method and system | |
| Zhao et al. | The study on the text classification for financial news based on partial information | |
| Chung et al. | Bankruptcy prediction using cerebellar model neural networks | |
| Ebiaredoh-Mienye et al. | Artificial neural network technique for improving prediction of credit card default: A stacked sparse autoencoder approach | |
| Su et al. | Multi-factor RFG-LSTM algorithm for stock sequence predicting | |
| JP2023162346A (en) | Framework for learning to transfer learn | |
| CN113240506A (en) | Financial wind-controlled cold start modeling method based on unsupervised field self-adaptation | |
| CN113435204A (en) | Stock price fluctuation prediction method based on news information | |
| Yuan et al. | A dynamic clustering ensemble learning approach for crude oil price forecasting | |
| Wang et al. | Leveraging Multisource Heterogeneous Data for Financial Risk Prediction: A Novel Hybrid-Strategy-Based Self-Adaptive Method. | |
| Kang et al. | A CWGAN-GP-based multi-task learning model for consumer credit scoring | |
| Zhang et al. | Choice of the number of hidden layers for back propagation neural network driven by stock price data and application to price prediction | |
| Yang | Research on Financial Credit Evaluation and Early Warning System of Internet of Things Driven by Computer-Aided Technology | |
| CN116452320B (en) | Credit risk prediction method based on continuous learning | |
| Verboven et al. | Hydalearn: Highly dynamic task weighting for multitask learning with auxiliary tasks | |
| Liu et al. | A Novel Methodology for Credit Spread Prediction: Depth‐Gated Recurrent Neural Network with Self‐Attention Mechanism | |
| Chen et al. | The credit scoring model based on logistic-bp-adaboost algorithm and its application in p2p credit platform | |
| Lin et al. | Soft computing algorithms in price of Taiwan real estates | |
| Ciola et al. | Search for profits and business fluctuations: How does banks’ behaviour explain cycles? | |
| Petelin et al. | Financial modeling using Gaussian process models |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |