Data processing method, device, electronic device and computer readable medium

Abstract

The present application discloses a data processing method, device, electronic device and computer-readable medium, and relates to the fields of artificial intelligence and big data, and specifically to the fields of machine learning and data preprocessing. The method comprises: receiving a business application request, determining the corresponding business identifier, extracting the data features of the provident fund data corresponding to the business application request, and then determining the data features corresponding to each business identifier from the data features; calling a prediction model, determining the classifier corresponding to each business identifier in the prediction model; inputting the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability; determining the corresponding work stability level according to each model prediction probability, and calling the corresponding strategy based on the work stability level to process the business application request. Thus, compared with other general work evaluation algorithms, it has higher efficiency and higher accuracy, and fully explores the use value of the provident fund value field.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and big data, specifically to the fields of machine learning and data preprocessing, and in particular to a data processing method, device, electronic device and computer-readable medium. Background Art

[0002] At present, the application of provident fund data in the market is still mostly focused on using provident fund loan data to analyze the risk level of customers. In some customer portraits, basic indicators that are easy to obtain are often used as the main presentation. Overall, there is still a lot of room for improvement in the data mining of provident fund value fields in the market.

[0003] In the process of implementing this application, the inventors found that there are at least the following problems in the prior art:

[0004] The provident fund value field is not fully utilized, and there is still much room for improvement in the data mining of provident fund value fields in the market. Summary of the invention

[0005] In view of this, the embodiments of the present application provide a data processing method, device, electronic device and computer-readable medium, which can solve the problem that the existing provident fund value field is insufficiently utilized and there is still a lot of room for improvement in the data mining of provident fund value fields in the market.

[0006] To achieve the above objective, according to one aspect of an embodiment of the present application, a data processing method is provided, including:

[0007] Receive a business application request, determine the corresponding business identifier, extract data features of provident fund data corresponding to the business application request, and then determine data features corresponding to each business identifier from the data features;

[0008] Calling the prediction model and determining the classifier corresponding to each business identifier in the prediction model;

[0009] Input the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability;

[0010] According to the prediction probability of each model, the corresponding work stability level is determined, and based on the work stability level, the corresponding strategy is called to process the business application request.

[0011] Optionally, before calling the prediction model, the data processing method further includes:

[0012] Get the initial gradient boosting model;

[0013] Obtain training sample data, which includes provident fund data with user identifiers and observation time points and annotated model prediction probabilities;

[0014] The provident fund data carrying user identification and observation time is used as the input of the initial gradient boosting model, the corresponding model prediction probability of the annotation is used as the expected output of the initial gradient boosting model, the gradient boosting tree algorithm is called, and the initial gradient boosting model is iteratively trained to obtain the prediction model.

[0015] Optionally, obtaining training sample data includes:

[0016] Obtain the provident fund data of each employee of the enterprise, format the initial provident fund data of each employee of the enterprise and unify the numerical field units to generate pre-processed provident fund data;

[0017] Determine the business identifier of the pre-processed provident fund data, and disassemble the pre-processed provident fund data based on the business identifier of the pre-processed provident fund;

[0018] The disassembled pre-processed provident fund data are cross-combined, and features are extracted from the cross-combined pre-processed provident fund data. The extracted features are derived and cleaned to obtain provident fund data carrying user identification and observation time points and annotated with corresponding model prediction probabilities, which are then determined as training sample data.

[0019] Optionally, iteratively train the initial gradient boosting model, including:

[0020] In response to the expected output not matching the labeled model prediction probability corresponding to the input, determining a training weight for the training sample data currently input into the initial gradient boosting model;

[0021] Based on a preset training weight level, a target training weight that is higher than the training weight is determined, and the training weight is updated to the target training weight.

[0022] Optionally, a prediction model is obtained, including:

[0023] The training sample data corresponding to the business identifiers of the pre-processed provident fund data are called to train the decision trees of the initial gradient boosting model respectively, thereby obtaining a trained classifier;

[0024] The classifiers are combined to obtain a prediction model.

[0025] Optionally, determining a corresponding working stability level includes:

[0026] Obtain the provident fund payment suspension data corresponding to the provident fund data;

[0027] Determine the probability of abnormal users based on the provident fund non-payment data, and then determine the probability of normal users based on the probability of abnormal users;

[0028] Determine the scoring value based on the abnormal user probability, normal user probability and preset coefficient;

[0029] Based on the correspondence between the preset scoring scores and the job stability levels and the scoring scores, the corresponding job stability levels are determined.

[0030] Optionally, before obtaining the prediction model, the method further includes:

[0031] The test sample data and the validation sample data are obtained, and the model parameters of the prediction model are updated according to the test sample data and the validation sample data.

[0032] In addition, the present application also provides a data processing device, including:

[0033] A receiving unit is configured to receive a service application request, determine a corresponding service identifier, extract data features of provident fund data corresponding to the service application request, and then determine data features corresponding to each service identifier from the data features;

[0034] A classifier determination unit is configured to call the prediction model and determine the classifier corresponding to each service identifier in the prediction model;

[0035] The probability prediction unit is configured to input the data features corresponding to each service identifier into the corresponding classifier to obtain the corresponding model prediction probability;

[0036] The data processing unit is configured to determine the corresponding working stability level according to the prediction probability of each model, and to call the corresponding strategy based on the working stability level to process the business application request.

[0037] Optionally, the data processing device further includes a training unit configured to:

[0038] Get the initial gradient boosting model;

[0039] Obtain training sample data, which includes provident fund data with user identifiers and observation time points and annotated model prediction probabilities;

[0040] The provident fund data carrying user identification and observation time is used as the input of the initial gradient boosting model, the corresponding model prediction probability of the annotation is used as the expected output of the initial gradient boosting model, the gradient boosting tree algorithm is called, and the initial gradient boosting model is iteratively trained to obtain the prediction model.

[0041] Optionally, the training unit is further configured to:

[0042] Obtain the provident fund data of each employee of the enterprise, format the initial provident fund data of each employee of the enterprise and unify the numerical field units to generate pre-processed provident fund data;

[0043] Determine the business identifier of the pre-processed provident fund data, and disassemble the pre-processed provident fund data based on the business identifier of the pre-processed provident fund;

[0044] The disassembled pre-processed provident fund data are cross-combined, and features are extracted from the cross-combined pre-processed provident fund data. The extracted features are derived and cleaned to obtain provident fund data carrying user identification and observation time points and annotated with corresponding model prediction probabilities, which are then determined as training sample data.

[0045] Optionally, the training unit is further configured to:

[0046] In response to the expected output not matching the labeled model prediction probability corresponding to the input, determining a training weight for the training sample data currently input into the initial gradient boosting model;

[0047] Based on a preset training weight level, a target training weight that is higher than the training weight is determined, and the training weight is updated to the target training weight.

[0048] Optionally, the training unit is further configured to:

[0049] The training sample data corresponding to the business identifiers of the pre-processed provident fund data are called to train the decision trees of the initial gradient boosting model respectively, thereby obtaining a trained classifier;

[0050] The classifiers are combined to obtain a prediction model.

[0051] Optionally, the data processing unit is further configured to:

[0052] Obtain the provident fund payment suspension data corresponding to the provident fund data;

[0053] Determine the probability of abnormal users based on the provident fund non-payment data, and then determine the probability of normal users based on the probability of abnormal users;

[0054] Determine the scoring value based on the abnormal user probability, normal user probability and preset coefficient;

[0055] Based on the correspondence between the preset scoring scores and the job stability levels and the scoring scores, the corresponding job stability levels are determined.

[0056] Optionally, the training unit is further configured to:

[0057] The test sample data and the validation sample data are obtained, and the model parameters of the prediction model are updated according to the test sample data and the validation sample data.

[0058] In addition, the present application also provides a data processing electronic device, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by one or more processors, the one or more processors implement the data processing method as described above.

[0059] In addition, the present application also provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the data processing method as described above is implemented.

[0060] To achieve the above objective, according to another aspect of the embodiments of the present application, a computer program product is provided.

[0061] A computer program product according to an embodiment of the present application includes a computer program, which, when executed by a processor, implements the data processing method provided by the embodiment of the present application.

[0062] One embodiment of the above invention has the following advantages or beneficial effects: the present application receives a business application request, determines the corresponding business identifier, extracts the data features of the provident fund data corresponding to the business application request, and then determines the data features corresponding to each business identifier from the data features; calls the prediction model to determine the classifier corresponding to each business identifier in the prediction model; inputs the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability; determines the corresponding work stability level according to each model prediction probability, and calls the corresponding strategy based on the work stability level to process the business application request. The implementation scheme is to input the corresponding data into the corresponding classifier of the prediction model according to the business identifier of the provident fund data, and obtain the evaluation result of the work stability of the corresponding user according to the model prediction probability output by each classifier. Therefore, compared with other general algorithms, it has higher efficiency and higher accuracy, and fully explores the use value of the provident fund value field.

[0063] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with specific implementation examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to better understand the present application and do not constitute an improper limitation on the present application.

[0065] Figure 1 is a schematic diagram of the main flow of the data processing method according to the first embodiment of the present application;

[0066] Figure 2 is a schematic diagram of the main flow of the data processing method according to the second embodiment of the present application;

[0067] Figure 3is a schematic diagram of an application scenario of a data processing method according to a third embodiment of the present application;

[0068] Figure 4 is a schematic diagram of main units of a data processing device according to an embodiment of the present application;

[0069] Figure 5 is an exemplary system architecture diagram to which the embodiments of the present application can be applied;

[0070] Figure 6 It is a structural diagram of a computer system of a terminal device or server suitable for implementing an embodiment of the present application. DETAILED DESCRIPTION

[0071] The following is a description of exemplary embodiments of the present application in conjunction with the accompanying drawings, including various details of the embodiments of the present application to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description. The acquisition, storage, use, and processing of data in the technical solution of the present application are in compliance with the relevant provisions of national laws and regulations.

[0072] Figure 1 is a schematic diagram of the main flow of the data processing method according to the first embodiment of the present application, such as Figure 1 As shown, the data processing method includes:

[0073] Step S101, receiving a service application request, determining a corresponding service identifier, extracting data features of provident fund data corresponding to the service application request, and then determining data features corresponding to each service identifier from the data features.

[0074] In this embodiment, the execution subject of the data processing method (for example, it can be a server) can receive a business application request through a wired connection or a wireless connection. Specifically, it can be a loan application request. The business identifier can be, for example, TQ, JC, ZFTQ, TXTQ, JE, BS, 3 / 6 / 9. The TQ and JC business identifiers can represent different behaviors, such as withdrawal and deposit; the ZFTQ, TXTQ, and business identifiers can represent different purposes, such as rental withdrawal and retirement withdrawal; the JE and BS business identifiers can represent different types of data, such as amount and number of transactions. The 3 / 6 / 9 business identifier can represent different time windows, such as March 6 / 9.

[0075] After receiving the business application request, the executing entity can extract the corresponding provident fund data, and then extract the corresponding features of the provident fund data based on the business identifier to obtain the corresponding data features.

[0076] Step S102: calling the prediction model to determine the classifier corresponding to each service identifier in the prediction model.

[0077] Specifically, before calling the prediction model, the method further includes:

[0078] Get the initial gradient boosting model, such as LightGBM. LightGBM adopts the Leaf-wise growth strategy, which is to find the leaf with the largest split gain from all current leaves each time, and then split it, and so on.

[0079] The training sample data is obtained, and the training sample data includes provident fund data carrying user identifiers and observation time points and annotated model prediction probabilities.

[0080] The provident fund data with user identification and observation time is used as the input of the initial gradient boosting model, and the corresponding model prediction probability of the annotation is used as the expected output of the initial gradient boosting model. The gradient boosting tree algorithm is called to iteratively train the initial gradient boosting model, specifically by repeatedly iteratively training the decision tree to minimize the loss function to obtain the prediction model. Specifically, the gradient boosting tree will iteratively train a series of decision trees in sequence. In each iteration, the algorithm uses the existing integration to predict the category of each training instance, and then compares the predicted result with the true label value. For instances with inaccurate prediction results, the algorithm will give them higher weights by re-labeling. As a result, these "errors" will be seriously concerned in the next iteration, and the previous errors will be corrected. After N iterations, N simple classifiers will be obtained, and they can be combined to obtain a final model.

[0081] Specifically, obtaining training sample data includes:

[0082] Obtain the provident fund data of each employee of the enterprise, format the initial provident fund data of each employee of the enterprise and unify the units of the numerical fields to generate pre-processed provident fund data.

[0083] For example, the certificate numbers that do not meet the standards are eliminated; the time field format is standardized and time outliers are eliminated; the units of numeric fields are unified, the amount type unit is unified as yuan, and the unit of ratio type variables is unified as percentage. For example, if the loan interest rate is 3.25%, the recorded value is 3.25; the absolute value of the numeric variable is taken to solve the abnormal situation of negative values ​​in the numeric variable; the categorical variables are cleaned and the outliers are converted into missing values.

[0084] Determine the business identifier of the pre-processed provident fund data, and disassemble the pre-processed provident fund data based on the business identifier of the pre-processed provident fund.

[0085] For example, according to the different business characteristics of different data sources, the data sources are reclassified into 11 categories, including personal information, personal deposits, personal loans, repayment behavior, overdue behavior, joint borrowing, unit information, unit deposits, tax information, guarantee information and policy information. For each type of information, the data can be disassembled and mixed from different business perspectives. Specifically, the data can be disassembled and mixed based on the business identifier, and the window data and data fields are combined to complete the feature design. For example, provident fund deposits can be divided into different behaviors (such as withdrawals, deposits, etc.), different purposes (such as rental withdrawals, retirement, etc.), different time windows (such as March / June / September, etc.), different data types (such as amount, number of items, etc.), etc., which can be cross-combined to form a huge feature network.

[0086] The disassembled pre-processed provident fund data are cross-combined, and the features of the cross-combined pre-processed provident fund data are extracted. The extracted features are derived and cleaned to obtain provident fund data with user identification and observation time points and marked with corresponding model prediction probabilities, which are then determined as training sample data. Among them, feature derivation: according to the designed derivation ideas, the derivation code of each category of features is developed, and finally more than 900 features can be derived. Feature cleaning: solve the problem of abnormal values ​​of features by determining the normal value range of features. Generally, the abnormal values ​​are selected and replaced with null or the maximum or minimum value for cleaning. The specific measures include: for statistical variables of the number of people, the normal value range of the features is determined after observing the data distribution, and the abnormal values ​​are replaced; for the time interval features, if negative values ​​appear or exceed the 99th percentile, they are replaced with null; for the proportion variables, they are uniformly processed into percentages; for numerical variables, if they exceed the 99th percentile, they are replaced with the 99th percentile, and if negative values ​​appear, they are replaced with 0.

[0087] Specifically, the initial gradient boosting model is iteratively trained, including:

[0088] In response to the mismatch between the model prediction probabilities of the annotations corresponding to the expected output and the input, the training weights of the training sample data currently input into the initial gradient boosting model are determined; based on the preset training weight level, a target training weight higher than the training weight is determined, and the training weight is updated to the target training weight. The target training weight is the weight that needs to be achieved in the next model iteration training.

[0089] For example, the gradient boosting tree will iteratively train a series of decision trees. In each iteration, the algorithm uses the existing ensemble to predict the category of each training instance, and then compares the prediction result with the true label value. For instances with inaccurate prediction results, the algorithm will give them higher weights by relabeling them. As a result, in the next iteration, these "errors" will be paid more attention to, and the previous errors will be corrected. After N iterations, N simple classifiers will be obtained, which can be combined to get a final model.

[0090] Specifically, the prediction model is obtained, including:

[0091] The training sample data corresponding to the business identifiers of the pre-processed provident fund data are called to train the decision trees of the initial gradient boosting model respectively, and then the trained classifiers are obtained; the classifiers are combined to obtain the prediction model.

[0092] Specifically, before obtaining the prediction model, the method further includes:

[0093] The test sample data and the validation sample data are obtained, and the model parameters of the prediction model are updated according to the test sample data and the validation sample data.

[0094] Specifically, the number of iterations of the algorithm, the maximum depth of the decision tree, the maximum number of continuous feature discretizations, the minimum information gain required when splitting a node, the minimum number of instances contained in the node after splitting, etc. can all be adjusted and set by yourself. The acquired training sample data is put into the selected initial gradient boosting model for training, and compared with the output results of the prediction model corresponding to the test sample data and the verification sample data, the model results can be obtained and verified, and the model parameters can be corrected and updated based on the model results.

[0095] Step S103: input the data features corresponding to each service identifier into the corresponding classifier to obtain the corresponding model prediction probability.

[0096] The classifier is a classifier in the prediction model. Based on the combined classifiers in the prediction model, the stability probability prediction of the input data features is made.

[0097] Step S104, determining the corresponding working stability level according to the prediction probability of each model, and calling the corresponding strategy based on the working stability level to process the service application request.

[0098] The execution subject can determine the stability level corresponding to the model prediction probability output by each classifier of the prediction model according to the correspondence between the preset probability and the stability level. Specifically, the execution subject can take the mode of the stability levels output by each classifier as the final working stability level. Of course, in the embodiment of the present application, the median of the stability levels output by each classifier can also be taken as the final working stability level. The embodiment of the present application does not specifically limit the method for determining the stability level.

[0099] Specifically, the execution entity can call the policy corresponding to the work stability level to process the business application request. For example, if the work stability level is A among A, B, C, and D, which is also the lowest level, the most stringent risk control policy can be called to process the business application request to strictly control the user's business application.

[0100] This embodiment receives a business application request, determines the corresponding business identifier, extracts the data features of the provident fund data corresponding to the business application request, and then determines the data features corresponding to each business identifier from the data features; calls the prediction model to determine the classifier corresponding to each business identifier in the prediction model; inputs the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability; determines the corresponding work stability level based on each model prediction probability, and calls the corresponding strategy based on the work stability level to process the business application request. The implementation scheme is to input the corresponding data into the corresponding classifier of the prediction model according to the business identifier of the provident fund data, and obtain the evaluation result of the work stability of the corresponding user according to the model prediction probability output by each classifier. Therefore, compared with other general algorithms, it has higher efficiency and higher accuracy, and fully explores the use value of the provident fund value field.

[0101] Figure 2 is a schematic diagram of the main flow of the data processing method according to the second embodiment of the present application. Figure 2 As shown, the data processing method includes:

[0102] Step S201, receiving a business application request, determining a corresponding business identifier, extracting data features of provident fund data corresponding to the business application request, and then determining data features corresponding to each business identifier from the data features.

[0103] Step S202: calling the prediction model to determine the classifier corresponding to each service identifier in the prediction model.

[0104] Specifically, the construction of the prediction model first uses data analysis and mining technology to complete the definition of good and bad customer labels and sample division of the job stability model. Then, the relevant features are extracted and derived based on the housing provident fund data. Finally, the model is built using machine learning technology, and the prediction probability value of the job stability of each employee is output. The construction of the prediction model involves all customers of the national housing provident fund data platform. The construction of the model mainly involves six aspects: determining the observation period and performance period, defining good and bad labels, formulating sample exclusion rules, obtaining and dividing samples, screening features, and selecting training models.

[0105] For example, the observation period is a rolling window to ensure that the "good" and "bad" behaviors of loans can be fully observed. The performance period is the time period for monitoring the performance of existing loans at the observation point in the future. According to the performance of the loan during the performance period, it is marked as "good", "uncertain" and "bad" respectively. The performance period must be long enough to ensure that the good and bad conditions of the sample are fully reflected. Due to the time limit of the sample data, the prediction model can select 4 observation points, namely 2019.10.31, 2019.11.30, 2019.12.31, and 2020.01.31, and define the 6 months after the observation point as the performance period. And determine that the observation window for provident fund change data is 3 years, and the observation window for other data is 6 months.

[0106] Bad customer definition:

[0107] Since the construction of the job stability model originates from the provident fund data platform and its goal is to predict the client's job stability risk, the prediction model uses the behavior of providingnt fund suspension as the standard for judging the client's job stability.

[0108] Housing fund non-payment behavior involves non-payment of the unit account and non-payment of the individual account. In order to more accurately assess the risk of job stability caused by personal factors of the client, it is stipulated that only when the individual stops paying and the unit does not stop paying, the client is judged to have truly stopped paying in that month. If "1" represents the real non-payment mark of the client in that month, and "0" represents the normal payment mark of the client's housing fund in that month. Then there is the following corresponding relationship:

[0109] The personal account status of the month (normal payment) - the unit account status of the month (normal payment) - the actual payment mark of the month (0);

[0110] The personal account status of the month (normal payment) - the unit account status of the month (suspended payment) - the actual payment mark of the month (0);

[0111] The personal account status of the month (suspended payment) - the unit account status of the month (normal payment) - the actual payment mark of the month (1);

[0112] Personal account status for the current month (suspended payment) - unit account status for the current month (suspended payment) - actual payment mark for the current month (0).

[0113] For example, as the number of consecutive non-payment months of the customer increases, the probability that the customer will continue to default on payment in the next month gradually increases. When the customer's consecutive non-payment mark reaches 3 months (m3), there is a more than 60% probability that it will get worse in the next month, that is, it will enter the stage of continuous non-payment for 4 months (m4). Therefore, it can be considered that after the customer's non-payment stage reaches m3, the rolling rate is relatively stable, and the customer has a greater probability of continuing to deteriorate. Therefore, the final definition of the working stability model is that the bad customer label is "1", and the standard is m3+ (that is, the customer has defaulted on payment for 3 months or more); the good customer label is "0", and the standard is that the customer has not defaulted on payment or has defaulted on payment for 2 months or less.

[0114] Sample Exclusions:

[0115] The model development sample should be representative of the group and must have accurate prediction information and performance information, so that the model is more accurate during development and can be applied in the future. As a loan model, the prediction model needs to exclude some special customer sample data, such as customers with abnormal data, customers with unpredictable behavior, and customers who can be defined as bad customers at the time of observation, to prevent special data from interfering with the real effect of the model.

[0116] The prediction model has established the following sample exclusion rules: exclude customers with abnormal ID numbers; exclude customers with empty personal accounts or abnormal status; exclude customers with personal accounts less than or equal to 180 days; exclude customers who have closed their accounts; exclude customers with empty unit accounts; exclude customers whose individual-associated unit accounts have not paid provident fund for three consecutive months before the observation point; exclude customers whose individual accounts have not paid provident fund for three consecutive months before the observation point.

[0117] Sample division:

[0118] Starting from the personal information table and the personal account information table, the model selected the overall sample data of the four observation time points of 201910, 201911, 201912, and 202001 to build the model, and used the customer ID number plus the observation time point as the primary key to define the unique identifier of the sample. At the same time, in order to better verify the model effect, the overall sample is divided into a development training set, a development test set, and a validation set. The training set is used for training and fitting of the work stability model, the test set is used to evaluate the generalization ability of the model, and the validation set is used to verify the performance and stability of the model in the extrapolated sample.

[0119] The prediction model can use samples from the two observation points of 201910 and 201911 for sampling the training set and the test set. Since there are too many good sample data, while retaining all bad samples, the good samples are sampled at a good-to-bad ratio of 9:1, and finally the training and test sets are divided at a ratio of 7:3. In addition, the prediction model can randomly select 1% of the samples from the two observation points of 201912 and 202001 as the validation set.

[0120] Feature screening:

[0121] The feature screening part can be divided into two modules. The first module is the preliminary feature screening, including missing rate screening (eliminating features with missing rate greater than or equal to 90%), iv screening (eliminating features with iv less than or equal to 0.01, the full name of IV is Information Value, which means information value or information amount in Chinese), related item screening (related item screening does not include categorical features, only sorts numerical features from high to low by iv for correlation calculation, and eliminates features with correlation greater than or equal to 0.7) and psi screening (eliminating features with psi value greater than or equal to 0.1); the second module is feature importance screening, which is to put the features after the first step of preliminary screening and the selected sample data into the model for training, obtain the model feature importance result ranking, and select the top 99% of the cumulative feature importance as the final selected model features.

[0122] Model selection and training:

[0123] The prediction model can be developed using the GBTModel algorithm. GBTModel is a gradient boosting tree algorithm, which is an integrated algorithm for decision trees. It minimizes the loss function by repeatedly training decision trees, and can be used for classification and regression. The gradient boosting tree algorithm is suitable for continuous features and categorical features. It has the properties of being able to handle categorical features, being easy to expand to multi-classification problems, and not requiring feature scaling.

[0124] Gradient boosting trees train a series of decision trees in sequence. In each iteration, the algorithm uses the existing ensemble to predict the category of each training instance, and then compares the prediction with the true label value. For instances with inaccurate predictions, the algorithm will give them higher weights by relabeling them. As a result, in the next iteration, these "errors" will be paid more attention to, and the previous errors will be corrected. After N iterations, N simple classifiers will be obtained, which can be combined to get a final model.

[0125] Among them, the number of iterations of the algorithm, the maximum depth of the tree, the maximum number of continuous feature discretizations, the minimum information gain required when splitting a node, the minimum number of instances contained in the node after splitting, etc. can all be adjusted and set by yourself. The acquired training set samples and selected features are put into the selected model for training, and compared with the results of the test set and validation set, the model results can be obtained and verified.

[0126] Step S203: input the data features corresponding to each service identifier into the corresponding classifier to obtain the corresponding model prediction probability.

[0127] The principles of steps S201 to S203 are similar to those of steps S101 to S103 and will not be described in detail herein.

[0128] Step S204, obtaining the provident fund suspension payment data corresponding to the provident fund data.

[0129] Step S205, determining the probability of abnormal users based on the provident fund non-payment data, and then determining the probability of normal users based on the probability of abnormal users.

[0130] Step S206, determining a scoring value based on the abnormal user probability, the normal user probability and a preset coefficient.

[0131] Step S207, determining the corresponding work stability level based on the preset correspondence between the scoring score and the work stability level and the scoring score.

[0132] In order to improve the readability and interpretability of the model output results, the model prediction probability is converted into a score, which can facilitate business personnel to make quick judgments. For example, the data of provident fund suspension corresponds to "bad users". Users who pay provident funds normally correspond to "good users". The probability that the user corresponding to the business application request is a bad user (that is, the probability of an abnormal user) can be P(Y=1|X)=p, and the probability that the user corresponding to the business application request is a good user (that is, the probability of a normal user) can be P(Y=0|X)=1-p. The ratio of good and bad users is calculated as odds=(1-p) / p. The score can be expressed as a linear expression of the logarithm of the ratio:

[0133] score=A+B*ln(odds)

[0134] Set the score p0 when odd = Θ0, and set the increased score to PDO when odds doubles, so the following combination formula can be listed:

[0135]

[0136] Finally, the values ​​of A and B can be calculated:

[0137]

[0138] In the embodiment of the present application, p0 (i.e., the preset coefficient) can be set to 600, odd=1 / 17 (i.e., the good / bad ratio of the training set sample data), PDO=100, and the score range is limited to between 300 and 700 (all scores below 300 are set to 300, and all scores above 700 are set to 700). According to the above-mentioned scoring conversion method, the final customer job stability score can be obtained.

[0139] The executing entity can determine the work stability level corresponding to the business application request based on the correspondence between the pre-set scoring points and the work stability level.

[0140] Step S208, determining the corresponding work stability level according to the prediction probability of each model, and calling the corresponding strategy based on the work stability level to process the service application request.

[0141] The embodiment of the present application can more accurately determine the work stability level corresponding to the business application request through the final scoring score.

[0142] Figure 3 : is a schematic diagram of an application scenario of the data processing method according to the third embodiment of the present application. The data processing method of the embodiment of the present application can be applied to the development and utilization scenario of provident fund data, that is, it can be a scenario of using provident fund data to determine work stability. Figure 3 As shown, the provident fund data of the national housing provident fund data platform (provident fund data) is used as the underlying data. After completing the analysis of the original data, the original data is sorted and cleaned to enter the feature construction link. The feature construction link includes three parts: feature design, feature derivation and feature cleaning. Some basic features are output for the next step of model building and platform display of features with significant importance. Then, relying on the features obtained by the construction, the construction of the job stability prediction model is completed after the feature importance screening, and two job stability level scores are output to the decision analysis platform, which are finally applied to the target functions such as employee risk view and special indicator audit. Among them, the original data may include: personal data, unit data, and other data. Among them, personal data includes personal information, account information, personal deposits and change information, etc. Unit data may include: unit deposits and changes, etc. Other data may include: co-borrowers, policy information, etc. The embodiment of this application does not limit the specific content contained in the original data.

[0143] The embodiments of the present application have achieved data innovation: making full use of the rich information value contained in the provident fund data, classifying and integrating the provident fund data source, and deriving a large number of new features with business significance through feature derivation technology, and finally applying them to the prediction of job stability, so as to achieve innovative development, deep mining and application of the value of provident fund data. Definition innovation: the user's job stability is accurately defined based on the provident fund data source, combining the dual data behaviors of unit payment and personal payment, accurately defining the concept of payment, so that job stability has a quantitative measurement indicator, and improving the reliability and interpretability of the results. Unlike traditional statistical analysis methods and logistic regression algorithms, the embodiments of the present application use a machine learning GBTModel algorithm, which is suitable for continuous features and category features, can better show the nonlinear relationship between data, and perform data modeling more quickly and efficiently, so as to ultimately achieve an overall improvement in the risk control capabilities of provident fund users under the output application of the job stability scoring index. The data usage level is richer: Different from the business scenario in the market where provident fund data is used to analyze loan delinquency and thus judge the customer risk level, the embodiment of the present application applies provident fund data to describe the user's job stability, and thereby evaluates the user's liquidity level, thereby identifying user risks. The new boundaries expanded at the data application level provide a new direction for future provident fund data mining. The model is more reliable: Compared with traditional statistical analysis and logistic regression models, GBTModel, as an advanced machine learning integration model, not only completes the error correction step by step by repeatedly iteratively training decision trees and continuously adjusting the weights of predicted error samples, but is also very friendly to large amounts of data, which makes the model have extremely high model development efficiency while also ensuring high accuracy and reliability of the results, and the business explanation is more persuasive: The embodiment of the present application quantifies the vague conceptual term of job stability and corresponds to indicators, so that the entire invention has a basis and reason when introducing and explaining the business. In practical applications, the guarantee of underlying data brings stronger persuasiveness and credibility to business explanations.

[0144] Figure 4 Schematic diagram of the main units of the data processing device according to an embodiment of the present application. Figure 4 As shown, the data processing device includes a receiving unit 401, a classifier determination unit 402, a probability prediction unit 403 and a data processing unit 404.

[0145] The receiving unit 401 is configured to receive a service application request, determine a corresponding service identifier, extract data features of the provident fund data corresponding to the service application request, and then determine data features corresponding to each service identifier from the data features.

[0146] The classifier determination unit 402 is configured to call the prediction model and determine the classifier corresponding to each service identifier in the prediction model.

[0147] The probability prediction unit 403 is configured to input the data features corresponding to each service identifier into the corresponding classifier to obtain the corresponding model prediction probability.

[0148] The data processing unit 404 is configured to determine the corresponding working stability level according to the prediction probability of each model, and to call the corresponding strategy based on the working stability level to process the service application request.

[0149] In some embodiments, the data processing device further includes Figure 4 The training unit not shown in the figure is configured to: obtain an initial gradient boosting model; obtain training sample data, the training sample data including provident fund data carrying user identification and observation time and annotated model prediction probability; use the provident fund data carrying user identification and observation time as the input of the initial gradient boosting model, use the annotated corresponding model prediction probability as the expected output of the initial gradient boosting model, call the gradient boosting tree algorithm, and iteratively train the initial gradient boosting model to obtain a prediction model.

[0150] In some embodiments, the training unit is further configured to: obtain the provident fund data of each employee of the enterprise, format the initial provident fund data of each employee of the enterprise and unify the units of the numerical fields to generate pre-processed provident fund data; determine the business identifier of the pre-processed provident fund data, and disassemble the pre-processed provident fund data based on the business identifier of the pre-processed provident fund; cross-combine the disassembled pre-processed provident fund data, extract features from the cross-combined pre-processed provident fund data, derive features and clean features on the extracted features to obtain provident fund data carrying user identifiers and observation time points and annotated with corresponding model prediction probabilities, and then determine them as training sample data.

[0151] In some embodiments, the training unit is further configured to: in response to the expected output not matching the labeled model prediction probability corresponding to the input, determine the training weight of the training sample data of the current input initial gradient boosting model; based on a preset training weight level, determine a target training weight that is higher than the training weight, and update the training weight to the target training weight.

[0152] In some embodiments, the training unit is further configured to: call the training sample data corresponding to the business identifier of each pre-processed provident fund data to train each decision tree of the initial gradient boosting model respectively, and then obtain a trained classifier; combine each classifier to obtain a prediction model.

[0153] In some embodiments, the data processing unit 404 is further configured to: obtain provident fund default payment data corresponding to the provident fund data; determine the probability of abnormal users based on the provident fund default payment data, and then determine the probability of normal users based on the abnormal user probability; determine the scoring value based on the abnormal user probability, the normal user probability and a preset coefficient; determine the corresponding job stability level based on the correspondence between the preset scoring value and the job stability level and the scoring value.

[0154] In some embodiments, the training unit is further configured to: obtain test sample data and validation sample data, and update model parameters of the prediction model according to the test sample data and validation sample data.

[0155] It should be noted that the data processing method and the data processing device in the present application have a corresponding relationship in terms of specific implementation content, so the repeated content will not be described again.

[0156] Figure 5 An exemplary system architecture 500 to which the data processing method or data processing device according to the embodiment of the present application can be applied is shown.

[0157] like Figure 5 As shown, system architecture 500 may include terminal devices 501, 502, 503, a network 504 and a server 505. Network 504 is used to provide a medium for communication links between terminal devices 501, 502, 503 and server 505. Network 504 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0158] Users can use terminal devices 501, 502, 503 to interact with server 505 through network 504 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 501, 502, 503, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0159] The terminal devices 501 , 502 , and 503 may be various electronic devices having a data processing screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0160] Server 505 can be a server that provides various services, such as a background management server that provides support for business application requests submitted by users using terminal devices 501, 502, and 503 (only as an example). The background management server can receive a business application request, determine the corresponding business identifier, extract the data features of the provident fund data corresponding to the business application request, and then determine the data features corresponding to each business identifier from the data features; call the prediction model to determine the classifier corresponding to each business identifier in the prediction model; input the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability; determine the corresponding work stability level according to each model prediction probability, and call the corresponding strategy based on the work stability level to process the business application request. By inputting the corresponding data into the corresponding classifier of the prediction model according to the business identifier of the provident fund data, the implementation scheme of obtaining the evaluation result of the work stability of the corresponding user according to the model prediction probability output by each classifier is implemented. Therefore, compared with other general algorithms, it has higher efficiency and higher accuracy, and fully explores the use value of the provident fund value field.

[0161] It should be noted that the data processing method provided in the embodiment of the present application is generally executed by the server 505 , and accordingly, the data processing device is generally set in the server 505 .

[0162] It should be understood that Figure 5 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0163] Reference below Figure 6 , which shows a schematic diagram of the structure of a computer system 600 of a terminal device suitable for implementing an embodiment of the present application. Figure 6 The terminal device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0164] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The CPU 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0165] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0166] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 609, and / or installed from the removable medium 611. When the computer program is executed by the central processing unit (CPU) 601, the above-mentioned functions defined in the system of the present application are executed.

[0167] It should be noted that the computer-readable medium shown in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. Computer-readable storage media may include, for example, but are not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0168] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0169] The units involved in the embodiments described in the present application may be implemented by software or hardware. The units described may also be set in a processor, for example, it may be described as: a processor includes a receiving unit, a classifier determination unit, a probability prediction unit, and a data processing unit. The names of these units do not, in some cases, constitute limitations on the units themselves.

[0170] As another aspect, the present application also provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently and not be assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device receives a business application request, determines the corresponding business identifier, extracts the data features of the provident fund data corresponding to the business application request, and then determines the data features corresponding to each business identifier from the data features; calls the prediction model to determine the classifier corresponding to each business identifier in the prediction model; inputs the data features corresponding to each business identifier into the corresponding classifier to obtain the corresponding model prediction probability; determines the corresponding working stability level according to each model prediction probability, and calls the corresponding strategy based on the working stability level to process the business application request.

[0171] The computer program product of the present application includes a computer program, and the computer program implements the data processing method in the embodiment of the present application when executed by a processor.

[0172] According to the technical solution of the embodiment of the present application, by inputting the corresponding data into the corresponding classifier of the prediction model according to the business identifier of the provident fund data, the implementation scheme of obtaining the evaluation result of the work stability of the corresponding user according to the model prediction probability output by each classifier is implemented. Therefore, compared with other general algorithms, it has higher efficiency and higher accuracy, and fully explores the use value of the provident fund value field.

[0173] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A data processing method, characterized in that: include: Receiving a service application request, determining a corresponding service identifier, extracting data features of provident fund data corresponding to the service application request, and then determining data features corresponding to each service identifier from the data features; Calling a prediction model to determine a classifier corresponding to each of the service identifiers in the prediction model; Inputting the data features corresponding to each of the service identifiers into the corresponding classifier to obtain the corresponding model prediction probability; Determine a corresponding work stability level according to the prediction probability of each model, and call a corresponding strategy based on the work stability level to process the service application request; Before calling the prediction model, the method further includes: obtaining an initial gradient boosting model; obtaining training sample data, the training sample data including provident fund data carrying a user identifier and an observation time point and annotated model prediction probability; using the provident fund data carrying the user identifier and the observation time as input of the initial gradient boosting model, using the annotated corresponding model prediction probability as the expected output of the initial gradient boosting model, calling the gradient boosting tree algorithm, and iteratively training the initial gradient boosting model to obtain the prediction model; Among them, the obtaining of training sample data includes: obtaining provident fund data of each employee of the enterprise, formatting the initial provident fund data of each employee of the enterprise and unifying the units of numerical fields to generate pre-processed provident fund data; determining the business identifier of the pre-processed provident fund data, and disassembling the pre-processed provident fund data based on the business identifier of the pre-processed provident fund; cross-combining the disassembled pre-processed provident fund data, extracting features from the cross-combined pre-processed provident fund data, and deriving and cleaning the extracted features to obtain provident fund data carrying user identifiers and observation time points and annotated with corresponding model prediction probabilities, and then determining them as training sample data.

2. The method according to claim 1, characterized in that The iterative training of the initial gradient boosting model comprises: In response to the expected output not matching the labeled model prediction probability corresponding to the input, determining a training weight for the training sample data currently input into the initial gradient boosting model; Based on a preset training weight level, a target training weight higher than the training weight is determined, and the training weight is updated to the target training weight.

3. The method according to claim 1, characterized in that The obtaining of the prediction model comprises: Calling the training sample data corresponding to the business identifier of each pre-processed provident fund data to train each decision tree of the initial gradient boosting model, thereby obtaining a trained classifier; The classifiers are combined to obtain the prediction model.

4. The method according to claim 1, characterized in that: The determining of the corresponding working stability level includes: Obtain the provident fund suspension payment data corresponding to the provident fund data; Determine the probability of abnormal users based on the provident fund non-payment data, and then determine the probability of normal users based on the abnormal user probability; Determine a scoring value based on the abnormal user probability, the normal user probability and a preset coefficient; Based on the preset correspondence between the scoring score and the work stability level and the scoring score, the corresponding work stability level is determined.

5. The method according to claim 1, characterized in that Before obtaining the prediction model, the method further includes: Acquire test sample data and validation sample data, and update model parameters of the prediction model according to the test sample data and validation sample data.

6. A data processing device, characterized in that: include: A receiving unit is configured to receive a service application request, determine a corresponding service identifier, extract data features of provident fund data corresponding to the service application request, and then determine data features corresponding to each service identifier from the data features; A classifier determination unit is configured to call the prediction model and determine the classifier corresponding to each of the service identifiers in the prediction model; A probability prediction unit is configured to input the data features corresponding to each of the service identifiers into a corresponding classifier to obtain a corresponding model prediction probability; A data processing unit is configured to determine a corresponding work stability level according to each model prediction probability, and to call a corresponding strategy based on the work stability level to process the service application request; The training unit is configured to obtain an initial gradient boosting model; obtain training sample data, the training sample data including provident fund data carrying user identifiers and observation time points and annotated model prediction probabilities; use the provident fund data carrying user identifiers and observation time points as inputs to the initial gradient boosting model, use the annotated corresponding model prediction probabilities as expected outputs of the initial gradient boosting model, call a gradient boosting tree algorithm, and iteratively train the initial gradient boosting model to obtain the prediction model; wherein, obtaining training sample data includes: obtaining provident fund data of each employee of the enterprise; The initial provident fund data of each employee of the enterprise is formatted and the units of the numerical fields are unified to generate pre-processed provident fund data; the business identifier of the pre-processed provident fund data is determined, and the pre-processed provident fund data is disassembled based on the business identifier of the pre-processed provident fund; the disassembled pre-processed provident fund data are cross-combined, and features are extracted from the cross-combined pre-processed provident fund data, and feature derivation and feature cleaning are performed on the extracted features to obtain the provident fund data carrying the user identifier and the observation time point and annotated with the corresponding model prediction probability, and then determined as the training sample data.

7. The device according to claim 6, characterized in that The training unit is further configured to: In response to the expected output not matching the labeled model prediction probability corresponding to the input, determining a training weight for the training sample data currently input into the initial gradient boosting model; Based on a preset training weight level, a target training weight higher than the training weight is determined, and the training weight is updated to the target training weight.

8. The device according to claim 6, characterized in that The training unit is further configured to: Calling the training sample data corresponding to the business identifier of each pre-processed provident fund data to train each decision tree of the initial gradient boosting model, thereby obtaining a trained classifier; The classifiers are combined to obtain the prediction model.

9. The device according to claim 6, characterized in that The data processing unit is further configured to: Obtain the provident fund suspension payment data corresponding to the provident fund data; Determine the probability of abnormal users based on the provident fund non-payment data, and then determine the probability of normal users based on the abnormal user probability; Determine a scoring value based on the abnormal user probability, the normal user probability and a preset coefficient; Based on the preset correspondence between the scoring score and the work stability level and the scoring score, the corresponding work stability level is determined.

10. A data processing electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 5.

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

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

Images