Data processing methods, devices, and risk prediction model training methods
By performing feature processing and filtering optimization on customer attribute data and UnionPay acquiring data of individual business owners, risk prediction values are generated, which solves the problem that traditional credit assessment is unable to dynamically identify the credit status of individual business owners, and realizes dynamic identification of business behavior and accurate prediction of post-loan risks.
Patent Information
- Application Number
- CN202611021941.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-11-10
AI Technical Summary
Traditional credit assessment methods are unable to dynamically reflect the credit status of individual business owners, especially when business conditions change, making it difficult to identify risks in a timely manner, resulting in limited ability to identify post-loan risk trends.
By embedding and extracting features from customer attribute data and UnionPay acquiring data, a time-series feature sequence is generated. Then, the parameters are optimized using an expectation-maximization Kalman filter to generate a filtered feature sequence. Finally, a classification decision is made based on the filtered feature sequence to generate a risk prediction value.
It enables dynamic identification of the business activities of individual business owners and advance prediction of post-loan risks, thereby improving the accuracy and robustness of risk prediction.
Smart Images

Figure CN123071065A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of financial technology or other related fields, and in particular to a data processing method, apparatus, and risk prediction model training method. Background Technology
[0002] Business Quick Loan is an online credit loan product offered by banks and other financial institutions. It approves loans based on multi-dimensional data generated from customers' business activities (such as settlement, assets, transactions, taxes, etc.), thereby providing fast and convenient financing services for micro and small enterprises and individual businesses with high cash flow pressure.
[0003] With the increasing popularity of this product, the proportion of individual business owners has significantly increased. Their operating income sources are diversified and highly volatile, making it difficult for traditional credit assessment methods to comprehensively and dynamically reflect their creditworthiness. Financial institutions typically use credit scoring models based on historical transaction, tax, and asset information for risk assessment when lending to micro and small enterprises and individual business owners.
[0004] However, these credit scoring models tend to focus on static historical information, making it difficult to depict the temporal changes and abnormal patterns of transaction behavior. They also lack comprehensive utilization of multi-source behavioral data, resulting in limited ability to identify post-loan risk trends. When a customer's business situation or cash flow changes abruptly, it is often difficult to complete risk assessment and early warning in a timely and accurate manner. Summary of the Invention
[0005] This application provides a data processing method, apparatus, and risk prediction model training method to improve the ability to dynamically identify and predict trends in changes in customer behavior, thereby achieving more accurate risk warnings.
[0006] Firstly, this application provides a data processing method, including:
[0007] In response to the target customer's customer business data, the customer business data includes: customer attribute data and UnionPay acquiring data for the period to be analyzed, wherein the UnionPay acquiring data is used to indicate the target customer's transaction history.
[0008] The customer attribute data and the UnionPay acquiring data are subjected to feature embedding and feature extraction to generate a time-series feature sequence.
[0009] The time-series feature sequence is optimized by an expectation-maximization Kalman filter to generate a filtered feature sequence, which is obtained by filtering out the abnormal features corresponding to the customer business data.
[0010] Based on the filtered feature sequence, a classification decision is made to generate a risk prediction value for the target customer.
[0011] Secondly, this application provides a method for training a risk prediction model, including:
[0012] Obtain the sample training set;
[0013] For the customer business data of each sample customer in the sample training set, the customer business data is quantified into risk value labels using a preset quantification formula.
[0014] Based on the customer business data and the risk value labels, a risk prediction model is trained, which is used to implement the first aspect and / or various possible implementations of the first aspect.
[0015] Thirdly, this application provides a data processing apparatus, comprising:
[0016] The feature extraction module responds to the target customer's customer business data, which includes: customer attribute data and UnionPay acquiring data within the period to be analyzed. The UnionPay acquiring data is used to indicate the target customer's transaction history.
[0017] The feature extraction module is used to embed and extract features from the customer attribute data and the UnionPay acquiring data to generate a time-series feature sequence.
[0018] The parameter optimization module is used to optimize the parameters of the time-series feature sequence through the expectation-maximization Kalman filter to generate a filtered feature sequence, which is obtained by filtering out the abnormal features corresponding to the customer business data.
[0019] The classification decision module is used to make classification decisions based on the filtered feature sequence and generate the risk prediction value of the target customer.
[0020] Fourthly, embodiments of this application provide an electronic device, including: a memory and a processor;
[0021] The memory stores computer-executed instructions;
[0022] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as well as the second aspect and / or various possible implementations of the second aspect as described above.
[0023] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect as well as the second aspect and / or various possible implementations of the second aspect.
[0024] In a sixth aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect, as well as the second aspect and / or various possible implementations of the second aspect.
[0025] The data processing method, apparatus, and risk prediction model training method provided in this application respond to the customer business data of the target customer, which includes customer attribute data and UnionPay acquiring data within the period to be analyzed. Feature embedding and feature extraction are performed on the customer attribute data and UnionPay acquiring data to generate a time-series feature sequence. The parameters of the time-series feature sequence are optimized using an Expectation-Maximization-Kalman filter to generate a filtered feature sequence. Classification decisions are made based on the filtered feature sequence to generate a risk prediction value for the target customer. By unifying the customer attribute data and UnionPay acquiring data within the period to be analyzed into a time-series feature sequence, and optimizing the sequence parameters using an Expectation-Maximization-Kalman filter, and then making classification decisions based on the stable filtered feature sequence, dynamic identification of risks in the target customer's business behavior and early prediction of post-loan risks are achieved, improving the accuracy and robustness of risk prediction. Attached Figure Description
[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0027] Figure 1 A flowchart illustrating a data processing method provided in this application embodiment. Figure 1 ;
[0028] Figure 2 A flowchart illustrating a data processing method provided in this application embodiment. Figure 2 ;
[0029] Figure 3 A flowchart illustrating a risk prediction model training method provided in an embodiment of this application;
[0030] Figure 4 A flowchart illustrating the processing of a risk prediction model provided in an embodiment of this application;
[0031] Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;
[0032] Figure 6 A schematic diagram of a risk prediction model training device provided in an embodiment of this application;
[0033] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0034] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0035] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0036] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.
[0037] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.
[0038] It should be noted that the data processing method, apparatus, and risk prediction model training method provided in this application can be used in the field of fintech, or in any field other than fintech. The application fields of the data processing method, apparatus, and risk prediction model training method in this application are not limited.
[0039] Existing loan risk control solutions typically build credit scoring models based on historical transaction, tax, and asset data to assess customer risk during pre-loan approval or post-loan review, thereby achieving risk monitoring. These solutions usually score, stratify, or classify based on existing characteristics, and their working method focuses on a static portrayal of past business performance.
[0040] Customer business behavior is highly dynamic, especially for individual business owners who may experience sudden increases or decreases in revenue, abnormal payment patterns, or changes in cash flow patterns during different operating cycles. Relying solely on static scoring is insufficient to accurately identify abnormal characteristics. When a customer's business situation suddenly deteriorates or abnormal transaction behavior accumulates rapidly, existing technologies often struggle to make timely and effective judgments, leading to delays in identifying potential risky customers.
[0041] In view of this, how to combine customer attribute data with UnionPay acquiring data within the period to be analyzed to improve the dynamic identification of changes in customer behavior and generate more accurate risk prediction results has become an urgent technical problem to be solved.
[0042] To address the aforementioned issues, this application provides a data processing method that, in response to the target customer's business data, performs feature embedding and feature extraction on customer attribute data and UnionPay acquiring data to generate a time-series feature sequence. Then, it optimizes the parameters of the time-series feature sequence using an expectation-maximization-Kalman filter to generate a filtered feature sequence for identifying anomalous features. Finally, it performs classification decisions based on the filtered feature sequence to generate a risk prediction value for the target customer.
[0043] Figure 1 A flowchart illustrating a data processing method provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:
[0044] S101: Customer business data in response to the target customer, which includes customer attribute data and UnionPay acquiring data for the period to be analyzed.
[0045] Customer attribute data describes the basic attribute information of target customers who have used the quick loan service, representing their basic business and credit attributes. Customer attribute data includes, but is not limited to, customer star rating, debt ratio, credit registration, customer type, registration date, industry, business location, license status, historical credit information, and repayment record identifiers.
[0046] UnionPay acquiring data is used to indicate the transaction history of target customers and is one of the core data sources reflecting their transaction behavior. UnionPay acquiring data includes transaction amount, number of transactions, refund records, cancellation records, transaction terminal identifier, transaction channel identifier, and settlement time information. The analysis period can be a preset fixed value or customized for each analysis. The analysis period can be 30, 60, or 90 consecutive days prior to the analysis date.
[0047] In this application, the implementing entity may be a data processing server, model service node, or post-loan monitoring platform with computing capabilities deployed in the bank's risk control system.
[0048] In this step, after receiving the target customer's business data, the risk prediction and analysis process begins.
[0049] Specifically, this could involve obtaining the target customer's business data after receiving a risk analysis request for a single target customer. Alternatively, it could be achieved by a scheduled task triggering a batch analysis process for customers already in the pool to obtain the target customer's business data.
[0050] In one possible embodiment, after receiving the customer identifier of the target customer, customer business data is retrieved from the customer master data system, the credit business system, and the acquiring transaction platform.
[0051] Optionally, before performing feature extraction and risk prediction analysis on customer business data, the collected customer business data can be preprocessed. Preprocessing includes integrity verification and time alignment.
[0052] Specifically, for missing fields in customer attribute data, they can be supplemented based on the corresponding values in the most recent valid file for the same customer; for abnormal data in UnionPay acquiring data, deduplication and removal can be performed. Abnormal data can include, for example, abnormal timestamps, duplicate transaction records, and obviously invalid test transaction records. UnionPay acquiring data within the analysis period is organized according to a preset time granularity, which can be daily, weekly, or hourly.
[0053] Customer attribute data can be attached as static features throughout the entire analysis period, or it can be replicated to each time unit to form a static extended vector with the same length as the time series. Through the above processing, a customer business data set with a unified structure, continuous time, and clearly defined customer dimensions is obtained, triggering subsequent feature embedding and feature extraction steps.
[0054] By simultaneously collecting customer attribute information and UnionPay acquiring data for the period to be analyzed during the input phase, continuous change data representing the business activities of target customers is provided for subsequent risk prediction analysis of target customers.
[0055] S102: Perform feature embedding and feature extraction on customer attribute data and UnionPay acquiring data to generate a time-series feature sequence.
[0056] Among them, the temporal feature sequence is used to carry the serialized features after feature embedding, global feature extraction and temporal modeling, and is the basis for subsequent parameter optimization and classification decision.
[0057] In this step, customer attribute data and UnionPay acquiring data are mapped into numerical representations under a unified computing space, and time-varying feature information is extracted to form a feature sequence arranged in chronological order.
[0058] Specifically, customer attribute data and UnionPay acquiring data are input into a pre-trained risk prediction model, and the corresponding time-series feature sequences are extracted through the feature extraction layer in the risk prediction model. The feature extraction layer may include a Long Short-Term Memory (LSTM) network or a Bidirectional Long Short-Term Memory (Bi-LSTM) network.
[0059] The risk prediction model is obtained by quantifying customer business data into risk value labels using a preset quantification formula, and then training the risk prediction model based on each customer's business data and risk value labels.
[0060] In one possible implementation, the process of embedding and extracting features from the aforementioned customer attribute data and UnionPay acquiring data to generate a time-series feature sequence is described in detail. This includes: embedding features from the customer attribute data and UnionPay acquiring data to generate an embedded feature sequence; extracting global features from the embedded feature sequence using an attention mechanism to generate an attention feature sequence; and modeling the global temporal dependency of the attention feature sequence using a pre-trained long short-term memory network to generate a time-series feature sequence.
[0061] In this application, the embedded feature sequence refers to a dense vector arranged in chronological order, formed by mapping customer attribute data and UnionPay acquiring data through an embedding layer. The attention feature sequence refers to a feature sequence obtained by dynamically weighting the embedded feature sequence using an attention mechanism. The attention mechanism can be a self-attention mechanism or a multi-head attention mechanism.
[0062] Specifically, customer attribute data and UnionPay acquiring data are input into the embedding layer for feature embedding, generating an embedded feature sequence. The embedding vectors of the static customer attribute data and the embedding vectors of each UnionPay acquiring data are concatenated along the feature dimension to form the embedded feature sequence. The transform network in the feature extraction layer extracts attention feature sequences from the embedded feature sequence, and the LSTM network in the feature extraction layer models temporal relationships to ensure that time periods of sudden increases, continuous declines, and abnormal concentrations of refunds receive effective responses during feature aggregation, resulting in a temporal feature sequence.
[0063] For example, customer attribute data may include discrete fields, such as customer type, industry code, etc. Each discrete field can be mapped to a fixed-dimensional vector using an embedding matrix.
[0064] By introducing location coding or time index coding, the chronological relationship of each time unit in the UnionPay acquiring data within the period to be analyzed can be characterized.
[0065] Feature extraction was achieved through embedding layers, transform networks, and LTSM networks, enabling the final time-series feature sequences to accurately reflect the dynamic evolution of customer business behavior and provide more discriminative input features for the generation of subsequent risk prediction values.
[0066] S103: Optimize the parameters of the time-series feature sequence using an expectation-maximization Kalman filter to generate a filtered feature sequence.
[0067] The filtered feature sequence is obtained by filtering out anomalous features corresponding to customer business data. The Expectation-Maximization-Kalman filter is used to iteratively optimize the time-series feature sequence, generating a filtered feature sequence corrected by state estimation. This sequence is then used to identify and filter out anomalous behaviors or patterns in the customer business data, retaining more stable and effective information that reflects anomalous changes. Anomalous features correspond to characteristic manifestations in customer business data that differ from normal change patterns.
[0068] In this step, the time-series feature sequence generated in step S102 is used as input, and the parameters are optimized through an expectation-maximization Kalman filter to distinguish between trend changes and anomalous disturbances in the time-series feature sequence. Since the state evolution characteristics differ for different customers, different business types, and different time windows, the relevant parameters can be adaptively adjusted based on the current time-series feature sequence.
[0069] In one possible implementation, a detailed explanation is provided on optimizing the parameters of a time-series feature sequence using an expectation-maximization Kalman filter to generate a filtered feature sequence. This includes: using the time-series feature sequence as observations, establishing a Kalman filter that includes state transition equations and observation equations, and initializing the state parameters of the Kalman filter; performing maximum likelihood estimation on the state parameters using the expectation-maximization algorithm to filter out outliers in the time-series feature sequence until the change in state parameters between adjacent iterations is less than a convergence threshold or a preset number of iterations is reached; and performing Kalman filtering recursion on the time-series feature sequence based on the converged state parameters to generate the filtered feature sequence.
[0070] The state parameters of the Kalman filter include: initial state value, state covariance, process covariance, and observation covariance.
[0071] In practical implementation, a time-series feature sequence is used as the observation value. A corresponding state-space model is established based on the sequence length and feature dimension of the time-series feature sequence. Initial values of the state, state covariance, process covariance, and observation covariance are initialized, and these initialization parameters can be set based on historical stable sample statistical results. The expectation-maximization algorithm is used to iteratively update the expectation and maximize the update within the Kalman filter framework, and maximum likelihood estimation is performed on the state parameters to gradually approximate values that can explain the time-series changes. Simultaneously, abnormal features deviating from the normal change pattern are suppressed or eliminated. When the parameter difference between two adjacent iterations is less than the convergence threshold, or when the preset number of iterations has been reached, parameter updates are stopped. The converged state parameters are then fed into the Kalman filter recursive process, updating the hidden state and observation estimates point by point in time sequence, and outputting a smoothed filtered feature sequence.
[0072] By optimizing and iteratively estimating state parameters, and then performing recursive filtering accordingly, the filtering model can fit the dynamic changes of the time-series feature sequence, weaken the interference of abnormal fluctuations on feature expression, and improve the stability of the input data for subsequent classification decisions.
[0073] S104: Based on the filtered feature sequence, make classification decisions and generate risk prediction values for target customers.
[0074] The classification decision is used to identify target customers based on the filtered feature sequence and output a risk prediction value for each target customer. The risk prediction value is used to characterize the likelihood of loan delinquency for the target customer.
[0075] In this application, the filtered feature sequence output in step S103 is input into a pre-trained classification model for risk assessment. The classification model can be an ensemble tree model, such as a random forest model or an XGBoost model.
[0076] Specifically, the filtered feature sequence is first aggregated, and then each decision tree makes a decision on the aggregated input feature. Each decision tree outputs the probability value or category label of the target customer belonging to the risk category. Finally, the outputs of all decision trees are arithmetically averaged to obtain the risk prediction value.
[0077] In one possible implementation, the classification decision based on the filtered feature sequence to generate the risk prediction value of the target customer includes: making decisions on the filtered feature sequence through multiple decision trees of a pre-trained random forest network; and statistically averaging the decision results of each decision tree to generate the risk prediction value of the target customer.
[0078] After receiving the filtered feature sequence, it is input into each decision tree in the random forest network. Each decision tree outputs a corresponding risk probability value according to its own node partitioning rules. The arithmetic mean of the risk probability values output by each decision tree is then calculated, and this arithmetic mean is used as the risk prediction value for the target customer.
[0079] By using a random forest network to perform parallel discrimination on the aggregated input features, risk prediction values for target customers are generated, thus enabling the prediction of loan delinquency risk for target customers.
[0080] This application provides a data processing method that responds to customer business data of a target customer, including customer attribute data and UnionPay acquiring data within the analysis period. The method involves feature embedding and feature extraction of the customer attribute data and UnionPay acquiring data to generate a time-series feature sequence. The time-series feature sequence is then optimized using an Expectation-Maximization-Kalman filter to generate a filtered feature sequence. Based on the filtered feature sequence, a classification decision is made to generate a risk prediction value for the target customer. By unifying the customer attribute data and UnionPay acquiring data within the analysis period into a time-series feature sequence, optimizing the sequence parameters using an Expectation-Maximization-Kalman filter, and then performing classification decisions based on the stable filtered feature sequence, this method achieves dynamic identification of short-term mutations, continuous anomalies, and trend deterioration in the target customer's business behavior, as well as early prediction of post-loan risks, thus improving the accuracy and robustness of risk prediction.
[0081] Figure 2 A flowchart illustrating a data processing method provided in this application embodiment. Figure 2 In this embodiment Figure 1 Based on the embodiments, a data processing method is further described, such as... Figure 2 As shown, the method includes:
[0082] S201. Compare the risk prediction value with the preset probability threshold, and discretize the first comparison result into a risk prediction level.
[0083] The basic risk level is established during the loan application approval process for target customers. This basic risk level can be derived from pre-loan screening, credit verification, operational stability assessment, or manual review results, and is pre-stored in the customer's risk profile using the customer's identifier as an index. Preset probability thresholds can be pre-set by financial institutions based on business rules, historical default sample distribution, and post-loan early warning requirements, and are stored in the risk control parameter table. The risk prediction level is used to represent the current risk level of the target customer.
[0084] Specifically, after obtaining the risk prediction value, the risk prediction value is compared with the probability threshold. When the risk prediction value is lower than the threshold, it is determined to be a low-risk prediction level. When the risk prediction value reaches or exceeds the threshold, it is determined to be a high-risk prediction level. In this way, the continuous prediction result is mapped to a discrete level, which is convenient for subsequent rule-based processing.
[0085] For example, with a probability threshold of 0.7, when the risk prediction value outputs a probability of... In this case, the customer is a low-risk prediction level customer; when the risk prediction value is... In this case, the customer is classified as a high-risk prediction level customer.
[0086] S202. By comparing the risk prediction level with the pre-existing basic risk level of the target customer, determine the risk category of the target customer.
[0087] Among them, risk categories are used to characterize states such as normal, attention, warning, or high risk.
[0088] Specifically, after outputting the risk prediction level, it is compared with the basic risk level. The risk category of the target customer is determined based on the combination relationship between the two levels, and post-loan monitoring is carried out according to the risk category.
[0089] Optionally, the risk category of the target customer can be determined based on the changing trend of the risk category. For example, if the risk prediction level is higher than the basic risk level, it can be determined that the risk status of the target customer has increased; if the two are consistent, it can be determined that the risk status remains stable; if the risk prediction level is lower than the basic risk level, it can be determined that the current risk is not higher than the pre-loan baseline.
[0090] Optionally, a detailed explanation can be provided regarding the determination of the risk category of a target customer by comparing the risk prediction level with the pre-existing basic risk level of the target customer, including:
[0091] The risk level is determined by comparing the predicted risk level with the target customer's basic risk level. The risk level deviation is then compared with the preset benchmark deviation, and the second comparison result is discretized into a comparison deviation level. Based on the target customer's risk prediction level and comparison deviation level, the risk category of the target customer is determined by matching it with a pre-constructed multi-dimensional decision matrix.
[0092] The risk level deviation characterizes the degree of deviation of the current prediction result from the baseline risk level, and its value can be obtained through level difference, mapping difference, or correspondence difference. The benchmark deviation is preset by business rules and serves as the boundary for judging whether the deviation has entered the abnormal range.
[0093] The comparison deviation level is used to convert continuous or semi-continuous deviation relationships into discrete levels, facilitating subsequent matrix matching. The multi-dimensional decision matrix can be pre-built as a two-dimensional table or a multi-dimensional lookup table, where one dimension corresponds to the risk prediction level and the other dimension corresponds to the comparison deviation level. The matrix cells record the risk category of the target customer and its explanatory identifier. In practical applications, this matrix can also be a database table, a rule table, or a cache mapping table, which is not limited in this application.
[0094] In practical implementation, after obtaining the risk prediction level of the target customer, the corresponding basic risk level is read from the customer's risk profile, and a level comparison is performed between the two to obtain the risk level deviation. The risk level deviation is compared with a preset benchmark deviation and mapped to different comparison deviation levels, which are then used as discrete input quantities.
[0095] The risk prediction level and the comparison deviation level are then input into a multi-dimensional decision matrix for matching, and the corresponding risk category is found and the output result is obtained.
[0096] For example, Table 1 is a customer classification table provided in an embodiment of this application. When the risk level deviation is greater than or equal to the benchmark deviation, the comparison deviation level of the target customer is determined to be large; when the risk level deviation is less than the benchmark deviation, the comparison deviation level of the target customer is determined to be small.
[0097]
[0098] By discretizing the deviation relationship between the risk prediction level and the basic risk level, and combining it with a pre-constructed matrix for joint judgment, the correlation between risk category and the customer's current business fluctuations and existing credit background is realized, thereby improving the structured expression ability and business interpretability of the classification results.
[0099] This application provides a data processing method that compares a risk prediction value with a preset probability threshold and discretizes the first comparison result into a risk prediction level. By comparing the risk prediction level with the pre-stored basic risk level of the target customer, the risk category of the target customer is determined. By comparing the risk prediction level with the pre-stored basic risk level of the target customer, the basic judgment formed before lending is correlated with the prediction result of the post-loan model, thereby improving the interpretability of the risk category of the target customer.
[0100] Figure 3 This is a flowchart illustrating a risk prediction model training method provided in an embodiment of this application, as shown below. Figure 3 As shown, the method includes:
[0101] S301: Obtain the sample training set.
[0102] The training set serves as the foundation for training the risk prediction model, comprising customer business data corresponding to multiple sample customers. Sample customers are the analysis objects within the training set and can be self-employed individuals, micro-enterprises, or other business entities with established loan relationships and verified historical risk results. Customer business data includes customer attribute data and UnionPay acquiring data for the period to be analyzed.
[0103] When acquiring a sample training set, the executing entity can be a training server, data processing platform, or model development device deployed in the bank's risk control system, which extracts historical sample data from relevant data sources and forms a sample record set based on customers.
[0104] S302: For the customer business data of each sample customer in the sample training set, the customer business data is quantified into risk value labels through a preset quantification formula.
[0105] The pre-defined quantification formula is used to convert customer business data into supervised training risk value labels. These risk value labels characterize the risk level of sample customers within the corresponding observation period. In this application, the risk value labels are paired with customer business data as the supervision target and input into the risk prediction model. Based on this, the risk prediction model learns the correspondence between business behavior and risk outcomes.
[0106] In this step, when quantifying the customer business data of sample customers, the data can be calculated based on a preset quantification formula to obtain corresponding risk value labels. These risk value labels can characterize the risk level of different sample customers in a uniform numerical form, providing clear supervision signals for subsequent model training.
[0107] By introducing a computable quantitative formula, business behavior data is transformed into structured risk labels, enabling subsequent models to learn the relationship between business signals and risk outputs.
[0108] In one possible implementation, the above-mentioned customer business data is quantified into risk value labels through a preset quantitative formula, including: extracting multi-dimensional risk indicators of sample customers from UnionPay acquiring data in customer business data. The multi-dimensional risk indicators include: fund flow frequency, transaction amount distribution, and account correlation; weighting and summing the fund flow frequency, transaction amount distribution, and account correlation to obtain the amplification factor; and calculating the ratio of the customer contribution of the sample customer to the exponential amplification factor to obtain the risk value label of the sample customer.
[0109] The risk value label is determined using the following formula:
[0110]
[0111] in, Risk value label; It is customer contribution, which is determined based on basic customer attributes such as customer star rating and debt ratio; Weighting coefficients for pre-set risk indicators. To determine whether the client has met this risk indicator, The label value for the risk indicator. , The threshold value for risk indicators. This refers to the indicator value of the risk indicator.
[0112] The following formula is used to determine the value of the capital flow frequency index:
[0113]
[0114] in, It is an indicator of the frequency of capital flows. This refers to the number of large-value transactions transferred in during the period to be analyzed. This refers to the number of large-value outflow transactions during the period to be analyzed. This is the end time of the period to be analyzed; This is the start time of the period to be analyzed.
[0115] The indicator value for the distribution of transaction amounts is determined using the following formula:
[0116]
[0117] in, It is an indicator value for the distribution of transaction amounts; This represents the number of small-amount transactions during the period to be analyzed. This represents the number of large-scale transactions during the period to be analyzed.
[0118] The following formula is used to determine the value of the account correlation index:
[0119]
[0120] in, This is an indicator value for account correlation; For the period to be analyzed, this customer and other accounts The number of transactions.
[0121] By determining the frequency of fund flows, the distribution of transaction amounts, account correlations, and risk label values, a data foundation is provided for model training, thereby improving the ability to identify abnormal business fluctuations.
[0122] S303: Train a risk prediction model based on each customer's business data and risk value labels.
[0123] The risk prediction model is used to implement the data processing method provided in the aforementioned embodiments. The risk prediction model is a target model for predicting loan risk. Its training process uses customer business data of each sample customer as input and corresponding risk value labels as supervision information. Through iterative optimization, model parameters capable of outputting risk prediction results are obtained. After the risk prediction model is trained, risk prediction results can be output based on the target customer business data.
[0124] In one implementation, to ensure the model training results are consistent with actual risk control scenarios, the sample training set can be divided into a training subset and a validation subset. During training, the error between the predicted value and the risk value label can be calculated using the validation subset after each iteration, and the training parameters or stopping conditions can be adjusted accordingly. After training is complete, the model parameters and related configurations are saved to form a deployable risk prediction model.
[0125] This application provides a risk prediction model training method. The method involves acquiring a sample training set; quantifying the customer business data of each sample customer in the training set into risk value labels using a preset quantization formula; and training a risk prediction model based on the customer business data and risk value labels. This risk prediction model implements the method described above. By acquiring a sample training set, generating supervised risk value labels using the quantization formula, and then training the risk prediction model, the model learns the correspondence between business data and risk outcomes, thereby outputting risk prediction values in subsequent actual business processing.
[0126] This application can be applied to risk assessment and post-loan management scenarios for online quick credit loans to individual business owners by banks. Based on customer business data from multiple customers who applied for and received approval for quick business loans on online platforms, a sample training set is obtained to build the model. The sample training set contains customer business data from multiple sample customers. By quantifying the customer business data, risk label values are determined, and a risk prediction model is trained.
[0127] For example, Figure 4 A flowchart of a risk prediction model provided in this application embodiment is shown below. Figure 4 As shown, the risk prediction model includes an embedding layer, a feature extraction layer, a feature optimization layer, and a decision layer. The feature extraction layer includes a transform network and an LSTM network.
[0128] The trained risk prediction model receives customer business data from the target customer, which includes customer attribute data and UnionPay acquiring data for the period to be analyzed.
[0129] The embedding layer performs feature embedding on customer attribute data and UnionPay acquiring data to generate embedded feature sequences.
[0130] After generating the embedded feature sequence, a transform network is used to perform weighted calculations on the embedded feature sequence to obtain the attention feature sequence. Then, an LSTM network is used to model the dependencies of the attention feature sequence and output the corresponding temporal feature sequence.
[0131] After obtaining the temporal feature sequence, the feature optimization layer uses the temporal feature sequence as observations to establish a Kalman filter and initializes the state parameters of the Kalman filter. Then, the expectation-maximization algorithm is used to perform maximum likelihood estimation on the state parameters. During the iteration process, outliers in the temporal feature sequence are filtered out, and the state parameters are continuously updated until the change in state parameters between adjacent iterations is less than the convergence threshold, or the number of iterations reaches the preset number of iterations. Based on the converged state parameters, Kalman filtering is recursively applied to the temporal feature sequence.
[0132] After filtering, the filtered feature sequence is input into the decision layer, where multiple decision trees in the random forest network make classification decisions. The decision results output by each decision tree are statistically analyzed and arithmetically averaged to obtain the risk prediction value for the target customer.
[0133] Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application, as shown below. Figure 5 As shown, the data processing device 50 provided in this embodiment includes: a feature extraction module 501, a parameter optimization module 502, and a classification decision module 503;
[0134] The feature extraction module 501 responds to the target customer's customer business data, which includes: customer attribute data and UnionPay acquiring data for the period to be analyzed. The UnionPay acquiring data is used to indicate the target customer's transaction history.
[0135] The feature extraction module 501 is used to embed and extract features from customer attribute data and UnionPay acquiring data to generate a time-series feature sequence.
[0136] The parameter optimization module 502 is used to optimize the parameters of the time series feature sequence through the expectation-maximization Kalman filter to generate a filtered feature sequence. The filtered feature sequence is obtained by filtering out the abnormal features corresponding to the customer business data.
[0137] The classification decision module 503 is used to make classification decisions based on the filtered feature sequence and generate risk prediction values for target customers.
[0138] In one possible implementation, the parameter optimization module 502 is further configured to optimize the parameters of the time-series feature sequence using an expectation-maximization Kalman filter to generate a filtered feature sequence. This includes: establishing a Kalman filter comprising a state transition equation and an observation equation, using the time-series feature sequence as the observation, and initializing the state parameters of the Kalman filter; performing maximum likelihood estimation on the state parameters using the expectation-maximization algorithm to filter out abnormal features in the time-series feature sequence until the change in state parameters in adjacent iterations is less than the convergence threshold or the preset number of iterations is reached; and performing Kalman filtering recursion on the time-series feature sequence based on the converged state parameters to generate the filtered feature sequence.
[0139] In one possible implementation, the classification decision module 503 is also used to embed features into customer attribute data and UnionPay acquiring data to generate an embedded feature sequence; to extract global features from the embedded feature sequence through a self-attention mechanism to generate an attention feature sequence; and to model the global temporal dependency of the attention feature sequence through a pre-trained long short-term memory network to generate a temporal feature sequence.
[0140] In one possible implementation, the classification decision module 503 is further used to make decisions on the filtered feature sequence using multiple decision trees of a pre-trained random forest network; statistically, it performs an arithmetic average of the decision results of each decision tree to generate a risk prediction value for the target customer. The risk prediction value is compared with a preset probability threshold, and the first comparison result is discretized into a risk prediction level; by comparing the risk prediction level with the pre-stored basic risk level of the target customer, the risk category of the target customer is determined. The basic risk level is constructed when approving the target customer's loan application.
[0141] In one possible implementation, the classification decision module 503 is further used to compare the risk prediction level with the target customer's basic risk level to determine the risk level deviation; compare the risk level deviation with the preset benchmark deviation, and discretize the second comparison result into a comparison deviation level; and match the target customer's risk prediction level and comparison deviation level with a pre-constructed multi-dimensional decision matrix to determine the target customer's risk category.
[0142] The data processing device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0143] Figure 6 This is a schematic diagram of the structure of a risk prediction model training device provided in an embodiment of this application, as shown below. Figure 6 As shown, the risk prediction model training device 60 provided in this embodiment includes: an acquisition module 601, a quantization module 602, and a training module 603;
[0144] Module 601 is used to acquire the sample training set;
[0145] The quantization module 602 is used to quantify the customer business data of each sample customer in the sample training set into risk value labels using a preset quantization formula.
[0146] Training module 603 is used to train a risk prediction model based on each customer's business data and risk value labels. The risk prediction model is used to implement the method as described in any one of claims 1-6.
[0147] In one possible implementation, the quantification module 602 is specifically used to extract multi-dimensional risk indicators for sample customers from UnionPay acquiring data in customer business data. The multi-dimensional risk indicators include: frequency of fund flows, distribution of transaction amount, and account correlation. The frequency of fund flows, distribution of transaction amount, and account correlation are weighted and summed to obtain the amplification factor. The ratio of the customer contribution of the sample customer to the exponential amplification factor is calculated to obtain the risk value label of the sample customer.
[0148] This embodiment provides a risk prediction model training device that can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effects are similar, and will not be described in detail here.
[0149] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 70 provided in this embodiment includes at least one processor 701 and a memory 702. Optionally, the electronic device 70 further includes a communication component 703. The processor 701, memory 702, and communication component 703 are connected via a bus 704.
[0150] In a specific implementation, at least one processor 701 executes computer execution instructions stored in memory 702, causing at least one processor 701 to perform the above-described method.
[0151] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0152] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0153] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0154] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0155] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0156] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.
[0157] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.
[0158] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.
[0159] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.
[0160] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.
[0161] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.
[0162] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.
[0163] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.
[0164] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0165] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A data processing method, characterized in that, include: In response to the target customer's customer business data, the customer business data includes: customer attribute data and UnionPay acquiring data for the period to be analyzed, wherein the UnionPay acquiring data is used to indicate the target customer's transaction history. The customer attribute data and the UnionPay acquiring data are subjected to feature embedding and feature extraction to generate a time-series feature sequence. The time-series feature sequence is optimized by an expectation-maximization Kalman filter to generate a filtered feature sequence, which is obtained by filtering out the abnormal features corresponding to the customer business data. Based on the filtered feature sequence, a classification decision is made to generate a risk prediction value for the target customer.
2. The method according to claim 1, characterized in that, The step of optimizing the parameters of the time-series feature sequence using an expectation-maximization Kalman filter to generate a filtered feature sequence includes: Using the time-series feature sequence as the observation value, a Kalman filter including a state transition equation and an observation equation is established, and the state parameters of the Kalman filter are initialized. The state parameters are estimated by the expectation-maximization algorithm to filter out abnormal features in the time series feature sequence until the change in state parameters in adjacent iterations is less than the convergence threshold or the preset number of iterations is reached. Based on the state parameters after iterative convergence, Kalman filtering is performed on the time-series feature sequence to generate the filtered feature sequence.
3. The method according to claim 1, characterized in that, The step of embedding and extracting features from the customer attribute data and the UnionPay acquiring data to generate a time-series feature sequence includes: The customer attribute data and the UnionPay acquiring data are used to embed features to generate an embedded feature sequence; Global feature extraction is performed on the embedded feature sequence using a self-attention mechanism to generate an attention feature sequence; The attention feature sequence is modeled globally using a pre-trained long short-term memory network to generate a temporal feature sequence.
4. The method according to claim 1, characterized in that, The step of making a classification decision based on the filtered feature sequence to generate a risk prediction value for the target customer includes: The filtered feature sequence is evaluated using multiple decision trees from a pre-trained random forest network. The statistical analysis is performed by averaging the decision results of each decision tree to generate the risk prediction value for the target customer.
5. The method according to claim 1, characterized in that, The method further includes: The predicted risk value is compared with a preset probability threshold, and the first comparison result is discretized into a risk prediction level. The risk category of the target customer is determined by comparing the risk prediction level with the pre-stored basic risk level of the target customer. The basic risk level is constructed when approving the loan application of the target customer.
6. The method according to claim 5, characterized in that, The step of determining the risk category of the target customer by comparing the risk prediction level with the pre-stored basic risk level of the target customer includes: Compare the predicted risk level with the basic risk level of the target customer to determine the risk level deviation; The risk level deviation is compared with a preset benchmark deviation, and the second comparison result is discretized into a comparison deviation level; The risk category of the target customer is determined by matching the risk prediction level and the comparison deviation level of the target customer with a pre-constructed multi-dimensional decision matrix.
7. A method for training a risk prediction model, characterized in that, include: Obtain the sample training set; For the customer business data of each sample customer in the sample training set, the customer business data is quantified into risk value labels using a preset quantification formula. Based on the customer business data and the risk value labels, a risk prediction model is trained, which is used to implement the method as described in any one of claims 1-6.
8. The method according to claim 7, characterized in that, The step of quantifying the customer business data into risk value labels using a preset quantification formula includes: From the UnionPay acquiring data in the customer business data, multi-dimensional risk indicators are extracted for the sample customers. These multi-dimensional risk indicators include: frequency of fund flows, distribution of transaction amounts, and account correlation. The amplification factor is obtained by weighting and summing the frequency of fund flows, the distribution of transaction amounts, and the account correlations. The risk value label of the sample customer is obtained by calculating the ratio of the customer contribution to the exponential amplification factor.
9. A data processing apparatus, characterized in that, include: The feature extraction module responds to the target customer's customer business data, which includes: customer attribute data and UnionPay acquiring data within the period to be analyzed. The UnionPay acquiring data is used to indicate the target customer's transaction history. The feature extraction module is used to embed and extract features from the customer attribute data and the UnionPay acquiring data to generate a time-series feature sequence. The parameter optimization module is used to optimize the parameters of the time-series feature sequence through the expectation-maximization Kalman filter to generate a filtered feature sequence, which is obtained by filtering out the abnormal features corresponding to the customer business data. The classification decision module is used to make classification decisions based on the filtered feature sequence and generate the risk prediction value of the target customer.
10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 8.
12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.
