CN110310200A

CN110310200A - Take back the method and device of overdue loan completely

Info

Publication number: CN110310200A
Application number: CN201910604647.8A
Authority: CN
Inventors: 张鑫; 赵焕芳; 侯鑫磊; 张皓; 董朝霞; 刘晓龙; 朱伟伟; 李振
Original assignee: Agricultural Bank of China
Current assignee: Agricultural Bank of China
Priority date: 2019-07-05
Filing date: 2019-07-05
Publication date: 2019-10-08
Anticipated expiration: 2039-07-05
Also published as: CN110310200B

Abstract

The present invention provides a kind of method and device for taking back overdue loan completely, receive wait take back completely overdue loan take back request completely after, obtain respectively correspond multiple interlock accounts take back clue completely, it is analyzed using decision assistant model clue is taken back completely, it determines to recommend in clue to take back clue completely from multiple take back completely based on the analysis results, recommends to take back completely clue as the foundation for taking back strategy completely for formulating overdue loan to be taken back completely；Wherein, interlock account includes the savings account other than the agreement refund account of the client of overdue loan to be taken back completely, with the savings account of the guarantor of client, each taking back clue completely includes multiple class another characteristics, each class another characteristic determines all in accordance with account information and customer information, and one take back completely in clue each class another characteristic only one.Based on scheme provided by the invention, approving person can recommend take back completely clue formulation and take back strategy completely by analyzing, each without analyzing overdue loan to be taken back completely takes back clue completely, so that improves overdue loan takes back efficiency completely.

Description

Take back the method and device of overdue loan completely

Technical field

The present invention relates to data processing field, in particular to a kind of method and device for taking back overdue loan completely.

Background technique

Loan transaction refers to that business bank lends one fund of client, and arranges client and give back capital after a certain time And interest.It provides a loan for any one, client does not give back principal and interest on time, then the loan is known as overdue loan, takes back completely Overdue loan just refers to the process of the principal and interest for withdrawing overdue loan.

For business bank, a large amount of overdue loan can have an adverse effect to the capital turnover of bank, therefore, high Effect overdue loan is taken back completely, just at the important guarantee of business bank's stable operation.

Currently, be directed to overdue loan to be taken back completely, business bank mainly to take back method completely be that approving person is analyzed one by one to clear It receives the multiple of overdue loan and takes back clue completely, then formulate based on the analysis results and take back strategy completely, exceed according to strategy is taken back completely from wait take back completely It withholds in the agreement refund account of the client of phase loan, with the principal and interest of the payment funding overdue loan of deduction.However, one to clear Receipts overdue loan, which is usually corresponding with, largely takes back clue completely, and the mode that clue is each taken back in this manual analysis of the prior art completely is imitated Rate is lower, and limit overdue loan takes back efficiency completely.

Summary of the invention

Based on above-mentioned the deficiencies in the prior art, the present invention proposes a kind of method and device for taking back overdue loan completely, to taking back completely Clue is screened, and is taken back clue auxiliary completely using the recommendation after screening and is formulated and take back strategy completely, takes back effect completely with improve overdue loan Rate.

To solve the above problems, the scheme now proposed is as follows:

First aspect present invention discloses a kind of method for taking back overdue loan completely, comprising:

Receive wait take back completely overdue loan take back request completely after, obtain that overdue loan to be taken back completely is corresponding multiple to take back clue completely； Wherein, each interlock account for taking back the corresponding overdue loan to be taken back completely of clue completely, each is taken back clue completely and wraps Include: multiple class another characteristics, each class another characteristic is all in accordance with the account information of the corresponding interlock account and described The customer information of the client of overdue loan to be taken back completely determines, and one is taken back completely each classification in clue and only corresponds to a feature；Institute The interlock account for stating overdue loan to be taken back completely includes the savings account in addition to arranging refund account of the client, and/or, institute State the savings account of the guarantor of client；

Clue is taken back completely using decision assistant model analysis is each described, determines the multiple to take back line completely based on the analysis results Clue is taken back in recommendation in rope completely；Wherein, clue conduct is taken back in the recommendation completely, and that formulates the overdue loan to be taken back completely takes back strategy completely Foundation, the decision assistant model establishes according to multiple historical samples, and each historical sample includes one and takes back line completely Rope and the mark of withholing for taking back the corresponding interlock account of clue completely, the mark of withholing show depositing for corresponding interlock account Whether money is used to pay the principal and interest for taking back the corresponding any overdue loan of clue completely.

Optionally, described using clue is taken back completely described in decision assistant model analysis, it determines based on the analysis results described more Clue is taken back in a recommendation taken back completely in clue completely, comprising:

Clue is taken back completely for each, takes back clue completely using described in decision assistant model analysis, obtains corresponding analysis result；

Clue is taken back completely for each, is taken back the corresponding analysis result of clue described in judgement completely and whether is met and preset takes back item completely Part；

Clue is taken back completely for each, if described take back completely takes back condition completely described in the corresponding analysis result satisfaction of clue, by institute State take back completely clue be determined as recommend take back clue completely.

Optionally, described using clue is taken back completely described in decision assistant model analysis, it determines based on the analysis results described more A recommendation taken back completely in clue is taken back completely after clue, further includes:

Clue is taken back completely for each recommendation, determines that the recommendation takes back whether the corresponding interlock account of clue is used to pay completely It is described after the principal and interest for taking back overdue loan completely, using it is described recommendation take back completely clue and it is described recommendation take back the corresponding association of clue completely The mark of withholing of account optimizes the decision assistant model；Wherein, withholing for the corresponding interlock account of clue is taken back in the recommendation completely Whether mark is according to the deposit of corresponding interlock account for paying the principal and interest setting of the overdue loan to be taken back completely.

Optionally, the establishment process of the decision assistant model includes:

The crucial classification in the multiple classification is determined according to the different degree of each classification of the historical sample；Its In, the different degree of each classification is based on the multiple historical sample using random forests algorithm and is calculated；

For each historical sample, with the crucial class another characteristic of the historical sample and the historical sample Mark of withholing constructs the corresponding crucial sample of the historical sample；

An initial value is calculated as current submodel according to the mark of withholing of the multiple crucial sample, and by iteration Number is set as 1；

Using the current submodel, the negative gradient of each crucial sample is calculated；

According to the negative gradient of each crucial sample and each crucial sample, submodel is calculated more New value；

Based on the submodel updated value, the current submodel is updated, and passs the number of iterations Increase 1；

Using updated current submodel as current submodel, and it is default to judge whether current iteration number is greater than Threshold value, if the current iteration number be less than or equal to the threshold value, return execute it is described utilize the current submodel, The negative gradient of each crucial sample is calculated；It, will be current auxiliary if the current iteration number is greater than the threshold value Model is helped to be determined as decision assistant model.

Optionally, the different degree for calculating each classification based on the multiple historical sample using random forests algorithm Process, comprising:

M decision tree is established based on the multiple historical sample；Wherein, the M is preset positive integer, described in each Decision tree utilizes C4.5 algorithm to establish, and the M decision tree constitutes a random forest；

According to the multiple historical sample, first bag of outer data error of each decision tree in the random forest is calculated；

For each classification, random noise is applied to all features of the category in the multiple historical sample, then Corresponding with the classification the of each decision tree is calculated according to the multiple historical sample after random noise is applied Two bags of outer data errors；

For each classification, the different degree of the classification according to the second error calculation of first error and the classification； Wherein, the first error includes first bag of outer data error of each decision tree in the random forest, the classification Second error includes second bag of outer data error corresponding with the classification of each decision tree in the random forest.

Second aspect of the present invention discloses a kind of device for taking back overdue loan completely, comprising:

Acquiring unit, for receive wait take back completely overdue loan take back request completely after, it is corresponding to obtain overdue loan to be taken back completely It is multiple to take back clue completely；Wherein, each interlock account for taking back the corresponding overdue loan to be taken back completely of clue completely, each Taking back clue completely includes: multiple class another characteristics, account of each class another characteristic all in accordance with the corresponding interlock account The customer information of the client of family information and the overdue loan to be taken back completely determines, and one is taken back completely each classification in clue and only corresponded to One feature；The interlock account of the overdue loan to be taken back completely includes the deposit account in addition to arranging refund account of the client Family, and/or, the savings account of the guarantor of the client；

Construction unit, for establishing decision assistant model according to multiple historical samples；Wherein, each historical sample is equal Clue is taken back completely including one and the mark of withholing for taking back the corresponding interlock account of clue completely, the mark of withholing show to correspond to Interlock account deposit whether for paying the principal and interest for taking back the corresponding any overdue loan of clue completely；

Recommendation unit, for taking back clue completely using the decision assistant model analysis is each described, based on the analysis results really It makes the multiple recommendation taken back completely in clue and takes back clue completely；Wherein, it is described wait take back completely as formulating to take back clue completely for the recommendation The foundation for taking back strategy completely of overdue loan.

Optionally, the recommendation unit includes:

Analytical unit obtains pair for taking back clue completely using described in decision assistant model analysis for each clue is taken back completely The analysis result answered；

Judging unit, for taking back whether the corresponding analysis result of clue meets described in judgement completely for each clue is taken back completely It is preset to take back condition completely；

First determination unit each takes back clue completely for being directed to, if the corresponding analysis result of clue of taking back completely meets institute The condition of taking back completely is stated, then is determined as recommending to take back clue completely by the clue of taking back completely.

Optionally, further includes:

Optimize unit, for taking back clue completely for each recommendation, determines that the corresponding association account of clue is taken back in the recommendation completely It is described after the principal and interest for taking back overdue loan completely whether family is used to pay, and takes back clue completely using the recommendation and the recommendation is taken back completely The mark of withholing of the corresponding interlock account of clue optimizes the decision assistant model；Wherein, it is corresponding to take back clue completely for the recommendation Whether the mark of withholing of interlock account is according to the deposit of corresponding interlock account for paying the sheet of the overdue loan to be taken back completely Breath setting.

Optionally, the construction unit includes:

First computing unit, it is the multiple for being calculated using random forests algorithm based on the multiple historical sample The different degree of each classification of historical sample；

Second determination unit, the different degree for each classification according to the historical sample determine the multiple classification In crucial classification；

Sample construction unit, for being directed to each historical sample, with the crucial class another characteristic of the historical sample, with And the mark of withholing of the historical sample constructs the corresponding crucial sample of the historical sample；

Second computing unit, for calculating an initial value as current according to the mark of withholing of the multiple crucial sample Submodel, and 1 is set by the number of iterations；

The negative of each crucial sample is calculated for utilizing the current submodel in third computing unit Gradient；

4th computing unit, for the negative gradient according to each crucial sample and each crucial sample, meter Calculation obtains submodel updated value；

Updating unit is updated the current submodel, and make institute for being based on the submodel updated value It states the number of iterations and is incremented by 1；

Judging unit, for as the current submodel, and judging current change for updated current submodel Whether generation number is greater than preset threshold value, if the current iteration number is less than or equal to the threshold value, triggers described second Computing unit utilizes the current submodel, and the negative gradient of each crucial sample is calculated；If the iteration time Number is greater than the threshold value, then current submodel is determined as decision assistant model.

Optionally, first computing unit includes:

Decision tree establishes unit, for establishing M decision tree based on the multiple historical sample；Wherein, the M is default Positive integer, each described decision tree utilizes C4.5 algorithm to establish, the M decision tree one random forest of composition；

First error computing unit, for calculating each decision in the random forest according to the multiple historical sample First bag of outer data error of tree；

Second error calculation unit owns the category in the multiple historical sample for being directed to each classification Feature applies random noise, then calculates each described decision tree according to the multiple historical sample after application random noise Second bag of outer data error corresponding with the classification；

Different degree computing unit, for each classification, according to the second error calculation institute of first error and the classification State the different degree of classification；Wherein, the first error includes first bag of outer data of each decision tree in the random forest Error, the second error of the classification include second bag corresponding with the classification of each decision tree in the random forest Outer data error

The present invention provides a kind of method and device for taking back overdue loan completely, and receive overdue loan to be taken back completely takes back request completely Afterwards, obtain respectively correspond multiple interlock accounts take back clue completely, analyzed using decision assistant model clue is taken back completely, according to Analysis result determines to recommend to take back completely clue from multiple take back completely in clue, recommendation takes back clue conduct completely and formulates overdue loan to be taken back completely Take back completely strategy foundation；Wherein, interlock account include the client of overdue loan to be taken back completely agreement refund account other than deposit The savings account of the guarantor of money account and client, each taking back clue completely includes multiple class another characteristics, the spy of each classification Sign determines all in accordance with account information and customer information, and one take back completely in clue each class another characteristic only one.Based on this The scheme provided is invented, approving person can take back clue formulation completely by analysis recommendation and take back strategy completely, without analyzing wait take back completely Each of overdue loan takes back clue completely, so that improves overdue loan takes back efficiency completely.

Detailed description of the invention

In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis The attached drawing of offer obtains other attached drawings.

Fig. 1 is a kind of flow diagram of method for taking back overdue loan completely provided in an embodiment of the present invention；

Fig. 2 is a kind of flow diagram of method for constructing decision assistant model provided in an embodiment of the present invention；

Fig. 3 be another embodiment of the present invention provides a kind of method for taking back overdue loan completely flow diagram；

Fig. 4 is a kind of stream of the method for different degree for calculating each classification in historical sample provided in an embodiment of the present invention Journey schematic diagram；

Fig. 5 is a kind of structural schematic diagram of device for taking back overdue loan completely provided in an embodiment of the present invention；

Fig. 6 is a kind of structural schematic diagram of the construction unit of device for taking back overdue loan completely provided in an embodiment of the present invention；

Fig. 7 is a kind of structural representation of the first computing unit of device for taking back overdue loan completely provided in an embodiment of the present invention Figure.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

The embodiment of the present application provides a kind of method for taking back overdue loan completely, referring to FIG. 1, this method includes following step It is rapid:

S101, receive wait take back completely overdue loan take back request completely after, obtain that overdue loan to be taken back completely is corresponding multiple to be taken back completely Clue.

Wherein, each takes back the interlock account that clue corresponds to above-mentioned overdue loan to be taken back completely completely, each is taken back completely Clue includes: multiple class another characteristics, and each class another characteristic takes back the corresponding interlock account of clue completely all in accordance with this The customer information of the client of account information and overdue loan to be taken back completely determines, and one is taken back completely each classification in clue and only corresponded to One feature.

Taking back clue completely can be recorded by mode shown in table 1.

Table 1

It is each classification for taking back the feature in clue completely that 1st row as shown in table 1, which is shown, and such as the 2nd row to the 6th row is shown 5 including above-mentioned class another characteristic are taken back completely clue.

It should be noted that taking back the classification for including in clue completely is not limited to classification shown in table 1, in addition in table 1 Other than classification, taking back completely can also include: customer risk preference in clue, if case-involving client, if corporate boss, client segmentation and Client guarantor etc..

The interlock account of overdue loan to be taken back completely includes, the client of overdue loan to be taken back completely in addition to arranging refund account Savings account, and/or, the savings account of the guarantor of client.Savings account refers to, currently deposits the account being not zero.

Wherein, the customer information of the client of the account information of interlock account and overdue loan to be taken back completely, can be by right It is handled to obtain from the business datum that multiple data sources obtain in advance, the data source that can be used for obtaining business datum includes but not It is limited to the customer data registered in banking system, the history that the history service for recording client and its guarantor and bank comes and goes Data, and by modes such as internets from the data outside the banking system of acquisition.Data outside banking system may include visitor The judicial data at family and its guarantor, customs's data and public sentiment data etc..

By the structuring processing and necessary data cleaning that carry out unstructured data to the business datum of acquisition, so that it may To obtain the account information of the interlock account of overdue loan to be taken back completely, and overdue loan to be taken back completely client customer information. Optionally, above-mentioned account information and customer information can be stored in Data Mart.

S102, clue is each taken back completely using decision assistant model analysis, determine above-mentioned multiple take back completely based on the analysis results Clue is taken back in recommendation in clue completely.

Clue is taken back in the recommendation determined, which completely, can be used as the foundation for taking back strategy completely for formulating overdue loan to be taken back completely.

Specifically, the strategy of taking back completely of overdue loan to be taken back completely is formulated by the approving person of bank.That is, utilizing this Shen Please embodiment provide method determine recommend take back clue completely after, approving person can analyze these recommendations and take back clue completely, thus Strategy is taken back completely according to what these recommendations took back that line formulates overdue loan to be taken back completely completely.

In general, clue is taken back in the recommendation that the embodiment of the present application finally determines completely, it is that overdue loan to be taken back completely is corresponding all The a part taken back completely in clue takes back clue completely, that is to say, that for an overdue loan to be taken back completely, recommends the quantity for taking back clue completely Less than the quantity for taking back clue completely, therefore, approving person's analysis recommendation is taken back the ratio of time used in clue completely and is analyzed wait take back overdue loan completely The all of money take back that the time used in clue is short completely, and according to recommending to take back completely, clue can be more efficient formulates overdue loan to be taken back completely Take back strategy completely.

Further, recommendation takes back the corresponding interlock account of clue completely and is properly termed as recommending to take back completely account, and approving person formulates Take back strategy completely, it is possible to specify what the multiple recommendations determined in step S102 were taken back completely in account one or more recommends to take back account completely Family, it is (overdue namely for taking back completely that principal and interest of the deposit of account for paying overdue loan to be taken back completely is taken back in appointed recommendation completely Loan), it is unappropriated to recommend the deposit for taking back account completely to be then not used in pay the principal and interest of overdue loan to be taken back completely and (do not have to namely In taking back overdue loan completely).

Certainly, what approving person formulated takes back strategy completely, any one recommendation can not also be specified to take back account completely, that is, It says, according to the decision of approving person, account is taken back in above-mentioned multiple recommendations completely, can be not used in and take back overdue loan completely.

Clue is taken back completely for each of step S101 acquisition, if one is taken back completely clue and is not determined to recommend to take back clue completely, Then this deposit for taking back the corresponding interlock account of clue completely is not used in the principal and interest for paying overdue loan to be taken back completely.

Above-mentioned decision assistant model is established according to multiple historical samples, each historical sample include one take back completely clue with And this takes back the mark of withholing of the corresponding interlock account of clue completely, mark of withholing shows whether the deposit of corresponding interlock account is used The principal and interest of the corresponding any overdue loan of clue is taken back completely in paying this.

What above-mentioned historical sample included takes back clue completely, is that bank has completed that the overdue loan taken back completely is corresponding to take back line completely Rope.

For any one historical sample, if the deposit of the corresponding interlock account of the historical sample be once used to pay it is corresponding The principal and interest of overdue loan, then the mark of withholing of this historical sample can be set to " receive ", if the corresponding pass of the historical sample The deposit of connection account not be used to pay the principal and interest of corresponding overdue loan, then the mark of withholing of this historical sample can be set For " refusal ",

The method provided by the embodiments of the present application for taking back overdue loan completely, receive treat take back overdue loan completely take back request completely Afterwards, obtain respectively correspond multiple interlock accounts take back clue completely, each is taken back clue completely and corresponds to an interlock account, then sharp It is analyzed with decision assistant model clue is taken back completely, so that it is clear as recommending to determine that a part takes back clue completely based on the analysis results Take-up rope, the recommendation determined take back completely clue can be used as approving person formulate overdue loan to be taken back completely take back completely strategy according to According to；Wherein, multiple interlock accounts include the savings account other than the agreement refund account of the client of overdue loan to be taken back completely, and visitor The savings account of the guarantor at family, it includes multiple class another characteristics that each, which takes back clue completely, and each class another characteristic is equal It is determined according to the account information of corresponding interlock account and customer information, takes back each class another characteristic in clue completely for one and only have One.Technical solution provided by the embodiments of the present application, can to overdue loan to be taken back completely it is multiple take back completely clue automatically into Row screening, then take back the obtained recommendation of screening completely clue and be supplied to approving person and analyze, compared with prior art, the application The method that embodiment provides makes approving person not need to treat each for taking back overdue loan completely and take back clue completely and analyze, and has The workload for reducing approving person of effect enables what approving person quickly formulated overdue loan to be taken back completely to take back plan completely Slightly, to improve the efficiency that bank takes back overdue loan completely.

Further, when strategy is taken back in formulation completely, approving person can recommend according to the analysis selection for taking back clue completely to recommendation It takes back account completely, the principal and interest of overdue loan to be taken back completely is paid with the deposit that account is taken back in the recommendation selected completely.By selection one or Account is taken back in multiple recommendations completely, can extend the channel that bank takes back overdue loan completely, and be more quickly completed overdue loan takes back work completely.

The method for taking back overdue loan completely that the application any embodiment provides, key are to utilize decision assistant model pair Multiple clues of taking back completely are analyzed, so that it is determined that clue is taken back in recommendation therein completely out.Therefore, it is given below a kind of according to history sample This establishes the method for decision assistant model as reference, referring to Fig. 2, the method for establishing decision assistant model includes following step It is rapid:

S201, crucial classification in multiple classifications is determined according to the different degree of each classification of historical sample.

The classification of historical sample, that is, the classification of each feature for taking back clue completely that includes in historical sample, i.e. 1 institute of table The association contract type shown, account purposes and account status etc..

The specific implementation procedure of step S201 can be, and each classification is successively sorted from big to small by different degree, important It spends maximum classification and comes first, it is subsequent and so on, then select top n classification therein as crucial classification.N is root According to the preset positive integer of the quantity of the classification of historical sample.Under normal circumstances, N can be set to 20, that is, will be each After classification presses importance sorting, preceding 20 classifications are chosen as crucial classification.

Wherein, the different degree of each classification is calculated using random forests algorithm based on multiple historical samples.

S202, it is directed to each historical sample, with the crucial class another characteristic of historical sample and withholing for historical sample The corresponding crucial sample of mark building historical sample.

Step S202 is the equal of to each historical sample, from all features for taking back clue completely of this historical sample Middle take out corresponds to aforementioned crucial class another characteristic, with this take back completely clue correspond to crucial class another characteristic and this It takes back the corresponding mark of withholing of clue completely, constructs this and take back the corresponding crucial sample of clue completely.

It is exemplified by Table 1, if taking back clue 1 in table 1 completely, to taking back completely, clue 5 is belonging respectively to historical sample 1, historical sample 2 ... is gone through History sample 5, and in 6 classifications enumerating of table 1, it is associated with contract type, three classifications of account purposes and account status are determined For crucial classification.So for each of above-mentioned 5 historical samples historical sample, clue is taken back completely with this historical sample Association contract type, the mark group of withholing of three class another characteristics of account purposes and account status and this historical sample It closes, just obtains the corresponding crucial sample of this historical sample.

It should be noted that needing to carry out the process of subsequent builds decision assistant model by each historical sample Mark of withholing shows in digital form.For example, can indicate " to refuse " with 0, that is, indicate depositing for corresponding interlock account Money not be used to pay the principal and interest of any overdue loan；It indicates " to receive " with 1, that is, indicates the deposit of corresponding interlock account It once be used to pay the principal and interest of corresponding overdue loan.

For taking back clue 1 and its corresponding historical sample 1 completely, it is assumed that withholing for historical sample 1 is identified as " refusal ", then The corresponding crucial sample of historical sample 1 is (loan, general, normally, 0), wherein the combination of all features of crucial sample, i.e., (loan, it is general, normal), it is believed that it is the feature vector of crucial sample.

S203, an initial value is calculated as current submodel according to the mark of withholing of multiple crucial samples, and will repeatedly Generation number is set as 1.

In the present embodiment, the number of iterations can be denoted as t.

It is generally basede on following formula (being denoted as formula 1) and calculates above-mentioned initial value:

In above-mentioned formula, m is the key that the quantity for constructing decision assistant model sample.Construct the decision assistant completed Model needs to be tested with Partial key sample, therefore under normal circumstances with the sum of crucial sample multiplied by 0.9, takes to result As the quantity for constructing crucial sample after whole, remaining 10% crucial sample is then used to test.For example, if currently having 100 historical samples construct 100 crucial samples according to these historical samples, are then used for 90 therein crucial samples Decision assistant model is constructed, remaining 10 are used to test the model built, that is, the m of above-mentioned formula is equal to 90.

Optionally, can also further adjust be previously used for multiple crucial samples of building decision assistant model composition it is logical It crosses to delete part therein and withhold and is identified as the crucial sample of " receiving ", to increase for constructing the multiple of decision assistant model Specific gravity shared by counter-example in crucial sample, to reach the counter-example identification pre-alerting ability for improving the decision assistant model that building obtains Effect.Wherein, counter-example refers to, withholds and is identified as the crucial sample of " refusal ".

Y therein_iIndicate the mark of withholing of crucial sample i, the conversion based on abovementioned steps to mark of withholing, y_iValue It is 1 or 0, c is exactly the initial value for needing to be calculated, L (y_i, c) and it then indicates according to withhold mark and the initial value of crucial sample i The initial abstraction function for the crucial sample i being calculated, the f on the left of formula₀(x) the current auxiliary of first time iteration is indicated Model.

The calculating process of formula 1 is the equal of, for the m crucial sample for constructing decision assistant model, calculating One meets the initial value of the following conditions:

Calculate separately the loss function of m crucial sample according to this initial value, m obtained calculated result and be most Small value.

The initial value being calculated can be denoted as c₀。

Due to withholing, mark is indicated with 1 and 0 respectively, the initial value that step S203 is calculated be one greater than 0 and Number less than 1.

Wherein, loss function can be exactly a kind of available as shown in following formula 2 there are many specific expression formula Loss function:

L(y_i, c) and=(y_i ²-c²)²

Loss function between two values can indicate the size of the difference between the two numerical value, such as above-mentioned formula 2 It is shown, y_iDifference between c is smaller, then for loss function between the two with regard to smaller, difference is bigger, then loss function is bigger.

Therefore, above-mentioned formula 1 is it is also understood that determine an initial value c₀, so that c₀With withholing for m crucial sample The whole difference of mark reaches minimum.

S204, using current submodel, the negative gradient of each crucial sample is calculated.

The negative gradient of crucial sample can be calculated based on following formula 3:

Wherein, r_iIndicate the negative gradient of crucial sample i, t indicates current the number of iterations, x_iIndicate the feature of crucial sample i Vector, f_t-1(x_i) indicate the feature vector of crucial sample i inputting the result obtained after current submodel.The t times iteration Current submodel, uses f_t-1(x) it indicates.For first time iteration, t=1, current submodel is exactly f₀(x), that is, before State the initial value being calculated in step S203, that is to say, that for first time iteration, what the feature vector no matter inputted is, Output is all the initial value being calculated in step S203.If t is greater than or equal to 2, the current submodel of the t times iteration f_t-1It (x) is exactly the updated current submodel obtained by subsequent step, for second of iteration in the t-1 times iteration And iteration each time later, the output f of current submodel_t-1(x_i) will be with feature vector x_iDifference and it is different, tool Body please refers to the explanation of subsequent step.

Above-mentioned formula 3 is equivalent to loss function and seeks local derviation to the output of current submodel, obtained result multiplied by -1 i.e. For the negative gradient of corresponding crucial sample.According to the difference of the expression formula of loss function, the result of above-mentioned local derviation is also different.If damage The expression formula that function uses aforementioned formula 2 is lost, then negative gradient calculation formula as shown in formula 3 can be converted to following formula 4:

r_i=2 × (f_t-1(x_i)-y_i)

S205, according to the negative gradient of each crucial sample and each crucial sample, submodel update is calculated Value.

Step S205, specifically according to the feature vector of each crucial sample and the negative gradient of this crucial sample, meter Calculation obtains submodel updated value.

The specific implementation process of step S205 may comprise steps of:

Current iteration is obtained according to the feature vector of each crucial sample and the negative gradient of each crucial sample, fitting Post-class processing (ClassificationAnd Regression Tree, CART), for the t times iteration, be fitted CART can be denoted as CART-t, i.e. post-class processing t.

It is fitted obtained CART-t, it can be according to the difference of the feature between each crucial sample, by aforementioned m crucial sample Originally it is divided into multiple set, each gathers a leaf node for being equivalent to CART-t, and each set includes at least One aforementioned crucial sample and the corresponding negative gradient of crucial sample.Each above-mentioned set, in other words leaf node, Ke Yiyong R_tjIt indicates, wherein subscript t indicates that this leaf node is the leaf node of post-class processing t, and j indicates this leaf node right The number in post-class processing answered, value range are greater than or equal to 1, less than or equal to the leaf node of post-class processing t The positive integer of sum.

For each leaf node of the post-class processing of current iteration, this leaf section is calculated according to following formula 5 The best-fit values of point:

Above-mentioned formula indicates, for each leaf node R of post-class processing t_tj, the best-fit values of this leaf node c_tjIt is the numerical value for meeting the following conditions:

For belonging to leaf node R_tjEach crucial sample i, spy of the current submodel based on this crucial sample i Levy vector x_iThe output f being calculated_t-1(x_i) and c_tjIt is added, the result being added according to the two is withholdd with key sample i's Identify y_iLoss function of the crucial sample i in current iteration is calculated.Best-fit values c_tjIt should make leaf node R_tj The sum of each loss function of the crucial sample in current iteration reach minimum value.

It should be noted that the expression formula of the loss function of above-mentioned formula 3 and formula 5 should calculate just with step S203 The expression formula of the loss function used when initial value is consistent.Method provided in this embodiment can be carried out based on a variety of loss functions, But during the entire realization of any one specific embodiment, the same loss function should be used.

Current iteration is fitted each leaf node of obtained post-class processing t and post-class processing t in the process Best-fit values constitute the submodel updated value being calculated during current iteration.

S206, it is based on submodel updated value, current submodel is updated, and the number of iterations is made to be incremented by 1.

Update to current submodel can be carried out based on following formula 6:

In above-mentioned formula, M indicates that current iteration is fitted the sum of the leaf node of obtained post-class processing t in the process, Due to being all based on identical m feature vector and its corresponding negative gradient when being fitted post-class processing in iterative process each time It carries out, therefore for each post-class processing t, the sum of leaf node is all M.

f_t(x) updated current submodel is indicated, it may also be said to be the t times updated current submodel, f_t-1 (x) the current submodel before the t time update, when t=1, f are indicated_t-1It (x) is exactly the initial value set in step S203, t is big When 2, f_t-1(x) the current submodel exported after an iteration before indicating, i.e., the t-1 times updated to work as Preceding submodel.

Formula 6 indicates, updated current auxiliary by the input of this feature vector the t times for any one feature vector x After helping model, the t times updated current submodel is first with post-class processing t, according to each of this feature vector Feature is classified, so that it is determined that going out this feature vector corresponding leaf node R in post-class processing t_tj, then use this Output of the feature vector in the t-1 times updated current submodel and leaf node R_tjBest-fit values be added, obtain To output of the feature vector x in the t times updated current submodel.Wherein, feature vector x takes back line completely by any one All crucial class another characteristic compositions in rope.

So that the number of iterations is incremented by 1, is the equal of increasing by 1 with current iteration number, then by obtained result assignment in working as Preceding the number of iterations.

S207, by updated current submodel, as current submodel.

Step S207 is the equal of by updated current submodel assignment in current submodel.

S208, judge whether current iteration number is greater than preset threshold value, if current iteration number is less than or equal to threshold value, Return to step S204；If current iteration number is greater than threshold value, S209 is thened follow the steps.

Above-mentioned preset threshold value is one according to accuracy requirement and time need it is also assumed that being maximum number of iterations It asks for help as determining positive integer, increases maximum number of iterations in a certain range, the decision assistant mould finally obtained can be made Type has higher precision, but the time needed for will increase building decision assistant model, and reduces maximum in a certain range The number of iterations, it is possible to reduce the time needed for building decision assistant model, but the decision assistant model finally obtained can be reduced Precision.

In general, maximum number of iterations can be set to be greater than or equal to 5, the positive integer less than or equal to 10.

The implementation procedure of step S204 to step S208 is equivalent to an iteration process, and step S208 execution terminates, quite In having executed an iteration process, S204 is returned to step from step S208, is equivalent to and starts next iteration process.

S209, current submodel is determined as decision assistant model.

The decision assistant model of step S209 output, can be expressed as following formula 7:

In above-mentioned formula, c₀Indicate that the initial value being calculated in step S203, f (x) indicate decision assistant model, T table Show the threshold value referred in step S208, i.e. maximum number of iterations, M indicates the sum of the leaf node of each post-class processing t.

The decision assistant model that can be seen that final output based on above-mentioned iterative process includes T post-class processing, successively It is post-class processing 1, post-class processing 2 ... post-class processing T, each of these post-class processing all corresponds to M leaf Node, each leaf node correspond to a best-fit values.

Optionally, method provided in this embodiment can also include:

S210, decision submodel is tested using the crucial sample for being not used for building decision assistant model.

As described in step S203, selected section key sample it can be used to construct decision assistant mould from all crucial samples Type, and the decision assistant model built is tested with another part key sample.

Test process includes, for a crucial sample for test, this crucial sample to decision assistant mode input Then this feature vector obtains output valve of the decision assistant model based on this feature vector, for determining as described in formula 7 Plan submodel, output valve will be one and be greater than 0 and the positive number less than 1, this output valve and the crucial sample for test Difference between mark of withholing is exactly analytical error of the decision assistant model to this crucial sample.Each use is calculated one by one In the analytical error of the crucial sample of test, the accuracy of decision assistant model can be evaluated according to all analytical errors.

In conjunction with the method for the building decision assistant model of previous embodiment introduction, another embodiment of the application provides a kind of clear The method for receiving overdue loan, referring to FIG. 3, method includes the following steps:

S301, receive wait take back completely overdue loan take back request completely after, obtain that overdue loan to be taken back completely is corresponding multiple to be taken back completely Clue.

S302, clue is taken back completely for each, using decision assistant model analysis, this takes back clue completely, obtains corresponding analysis As a result.

Clue is taken back completely using decision assistant model analysis one to obtain corresponding analysis as a result, being the equal of by this It takes back clue input decision assistant model completely, obtains the output that decision assistant model takes back clue completely based on this.

Below with reference to the realization process of the brief introduction step S302 of decision assistant model shown in formula 7.

Firstly the need of explanation, it is clear that each can be seen that according to the step S201 of previous embodiment and step S202 Take-up Suo Jun includes crucial class another characteristic each of being determined in previous embodiment, and take back completely in clue at one, arbitrarily One crucial class another characteristic one and only one.Therefore, clue is taken back completely to decision assistant mode input one, be equivalent to certainly Plan submodel inputs a feature vector of each crucial class another characteristic composition that this takes back clue completely.

Each post-class processing for forming decision assistant model, for each crucial class according to the feature vector of input Another characteristic classifies to the feature vector of input.Therefore, although input is taken back completely in clue in addition to crucial class another characteristic It in addition, further include multiple non-key class another characteristics, but these non-key class another characteristics can't be to decision submodel Analytic process and analysis result have an impact, thus input one take back completely clue be equivalent to input this to take back clue completely corresponding Feature vector.

For decision assistant model shown in formula 7, initial value is not just rechanged after the completion of building, so taking back clue quilt completely After input, decision assistant model calls the post-class processing of itself to take back the spy of each crucial classification of clue completely according to this one by one Sign takes back clue completely to this and classifies.

Above-mentioned calling process can be to be called in order, i.e., first calling classification regression tree 1 is classified to clue is taken back completely, really Make some leaf node R for taking back clue completely and corresponding to post-class processing 1 of input_1jAfterwards, so that it may determine the clear of input Take-up rope corresponding best-fit values c in post-class processing 1_1j, and so on, for decision assistant model shown in formula 7, Can determine input takes back clue completely in T best-fit values wherein, i.e. c_1j, c_2j……c_Tj, most by the T determined Good match value is added with the initial value of decision assistant model, and obtained numerical value is exactly decision assistant model shown in formula 7 to this A analysis result for taking back clue completely.

S303, take back clue completely for each, judge this take back completely the corresponding analysis result of clue whether meet it is preset clear Receipt part.

For the analysis that is obtained using decision model shown in formula 7 as a result, above-mentioned preset condition of taking back completely can be set For a threshold value.Step S303, is equivalent to and takes back clue completely to each, judges that this takes back the analysis of clue completely as a result, namely Whether the numerical value of decision assistant model output is greater than the threshold value of setting, if analysis result is greater than threshold value, then it is assumed that this takes back line completely Rope satisfaction takes back condition completely, on the contrary, then it is assumed that this is taken back clue completely and is unsatisfactory for taking back condition completely.

When constructing decision assistant model shown in formula 7, for the mark of withholing of crucial sample, use 1 indicates " to receive ", uses 0 indicates " refusal ", and therefore, above-mentioned threshold value can be set as one and be slightly less than 1 numerical value, for example, can be set as 0.9.

S304, clue is taken back completely for each, if this takes back clue completely, corresponding analysis result satisfaction takes back condition completely, this is clear Take-up rope is determined as recommending to take back clue completely.

Assuming that above-mentioned threshold value is set as 0.9, if an analysis result for taking back clue completely is 0.95, it is determined that this is clear Take-up rope is to recommend to take back clue completely, if an analysis result for taking back clue completely is 0.8, it is to recommend clearly that this, which takes back clue completely or not Take-up rope.

S305, clue is taken back for each recommendation completely, takes back whether the corresponding interlock account of clue is used for completely according to the recommendation It pays principal and interest and corresponding mark of withholing is set.

Wherein, clue is taken back for each recommendation completely, this recommendation takes back whether the corresponding interlock account of clue is used to prop up completely The principal and interest for paying corresponding overdue loan to be taken back completely, has the approving person of bank artificially to determine.

In step S304, determine after recommending to take back clue completely, clue is taken back in each recommendation completely, and clue is taken back in these recommendations completely Analysis pushes away approving person from multiple as a result, corresponding account information and customer information can be shown to the approving person of bank Recommend the interlock account taken back completely and determine the principal and interest for paying overdue loan to be taken back completely in clue.

Clue is taken back completely for each recommendation, is exceeded for paying wait take back completely if corresponding interlock account is determined through approving person The principal and interest of phase loan, then the corresponding mark of withholing of this interlock account is set as " receiving ", if corresponding interlock account is through examining Personnel's determination is not used in the principal and interest for paying overdue loan to be taken back completely, then the mark of withholing of this interlock account is determined as " refusing ".

S306, recommend to take back clue completely and its withhold using each to identify Optimal Decision-making submodel.

It should be noted that is utilized in step S306 is that clue is taken back in each recommendation that step S304 is determined completely, and No matter the principal and interest whether corresponding interlock account of clue is used to pay overdue loan to be taken back completely is taken back in this recommendation completely.That is, Clue is taken back in recommendation for Optimal Decision-making submodel completely, is identified as the recommendation of " refusal " including corresponding withhold and takes back clue completely, It is identified as the recommendation of " receiving " with corresponding withhold and takes back clue completely.

In the method for building decision assistant model as shown in Figure 2, it is related to the weight of each classification according to historical sample The process for determining crucial classification from multiple classifications of historical sample is spent, is described below and a kind of calculates the important of each classification The method of degree is as reference, and certainly, the method that others calculate different degree is readily applicable to the offer of the application any embodiment Method.Referring to FIG. 4, calculate different degree method the following steps are included:

S401, M decision tree is established based on historical sample set.

Historical sample set includes multiple historical samples, and step S401 is the equal of constructing M according to multiple historical samples Decision tree.

Wherein, M is preset positive integer, if it is desired to the different degree being calculated can accurately reflect actual conditions, A biggish M value can be set.

Each of step S401 decision tree is to be calculated using aforesaid plurality of historical sample as training sample using C4.5 The decision tree that method is established.

C4.5 algorithm is a kind of mature algorithm for being used to generate decision tree, therefore, for the tool of above-mentioned building decision tree Details are not described herein again for body process.

It should be noted that when constructing the decision tree, and not all historical sample is all for each above-mentioned decision tree Participate in building process.For each decision tree, need first to select at random from above-mentioned all historical samples when constructing the decision tree Partial history sample is selected as data in the bag of the decision tree, data in the bag of this decision tree is then based on and constructs the decision Tree.

It should also be noted that, the process of above-mentioned selected section historical sample is that have the selection process put back to, that is to say, that After constructing what a decision tree, data can be combined with other non-selected historical samples in the bag of this decision tree, be obtained just The historical sample set of beginning repeats the above process when constructing other decision trees, randomly chooses from this historical sample set Partial history sample.Wherein, for each decision tree, all historical samples that not be used to construct this decision tree, system Claim the bag of this decision tree outer data.

One decision tree is equivalent to a taxon.Based on the decision tree of above-mentioned historical sample set building, Ke Yi Obtain a historical sample take back completely clue as input after, export this mark of withholing for taking back the corresponding interlock account of clue completely Know, i.e. output " refusal " or " receiving ", is equivalent to and determines that the clue of taking back completely of input is specifically any in two types.

Certainly, the mark of withholing of decision tree output is only each feature taken back in clue of the special algorithm according to input One responded is as a result, mark of not necessarily really withholing (mark of withholing of i.e. corresponding historical sample). That is, the clue of taking back completely for the historical sample for being identified as " refusal " that one is withholdd inputs a decision tree, this decision tree The mark of withholing of output may be " receiving ", that is, not be inconsistent with mark of really withholing.

Above-mentioned M decision tree constitutes a random forest.

S402, according to historical sample set, calculate first bag of outer data error of each decision tree in random forest.

The outer data error of the bag of one decision tree, refers to according to the error that data are calculated outside the bag of this decision tree.

A kind of calculation method of the outer data error of bag is, for a decision tree, its bag each history in data outside one by one Sample takes back clue completely, and clue is taken back in every input one completely, and obtains this decision tree based on taking back clue output completely according to this After mark of withholing, output is withholdd into mark compared with this takes back the mark of withholing of the corresponding historical sample of clue completely, if the two It is inconsistent, show that this decision tree takes back clue completely to this, be in other words false judgment to the judgement of this historical sample, records The quantity of the false judgment of this decision tree, is denoted as a, it is assumed that the historical sample quantity that the outer data of the bag of this decision tree include is B, then the outer data error of the bag of this decision tree is equal to a divided by b.

S403, it is directed to each classification, random noise is applied to all features of the category in historical sample set, and count Calculate second bag of outer data error corresponding with the category of each decision tree.

It is exemplified by Table 1, it is assumed that historical sample set includes historical sample 1 to historical sample 5, these historical samples include It takes back clue completely successively and is in table 1 and take back clue 1 completely to taking back clue 5 completely.If necessary to other to account purposes this kind therein Feature applies random noise, it is necessary to take back clue (or perhaps historical sample) completely for each, this is taken back completely to the account of clue Family purposes class another characteristic random replacement at the category another feature.For example, the account purposes for taking back clue 1 completely is " logical With ", then need to take back completely this feature replacement of clue 1 at " financial direct subsidy " or " medical insurance family ", and it is similar, take back clue 2 completely " financial direct subsidy " needs be substituted for " general " or " medical insurance family ", and so on.

As described above, being exactly the method for applying random noise to any one class another characteristic in historical sample set.

It should be noted that aforementioned first bag of outer data error and second bag of outer data error, it is only for before showing Person is to be calculated according to data outside the bag before application noise, and the latter is calculated according to data outside the bag after application noise It arrives, both the outer data error of bag, the method for calculating are also identical.

It should also be noted that, only applying noise, the spy of other classifications to a class another characteristic every time in step S403 Sign then keeps consistent with the feature in original historical sample set.

S404, it is directed to each classification, according to the different degree of first error and the second error calculation category of the category.

Wherein, first error includes first bag of outer data error of each decision tree in random forest, any one class Other second error includes second bag of outer data error corresponding with the category of each decision tree in random forest.

A kind of method that the different degree of a classification is calculated in step S404 is, by each decision tree in random forest First bag of outer data error sums to obtain first error, and by corresponding with the category of each decision tree in random forest Two bags of outer data error summations, obtain the second error of the category, subtract first error with the second error of the category, obtain Difference is exactly the different degree of this classification divided by M.The process can be understood with reference to following formula 8:

Im therein indicates the different degree of a classification, errOOB1_iIndicate number outside first bag of any one decision tree i According to error, errOOB2_iIndicate second bag of outer data error corresponding with the category of any one decision tree i.

In conjunction with the method for taking back overdue loan completely that above-mentioned the application any embodiment provides, another embodiment of the application is provided A kind of device for taking back overdue loan completely, referring to FIG. 5, the device includes with lower unit:

Acquiring unit 501, for receive wait take back completely overdue loan take back request completely after, it is corresponding to obtain overdue loan to be taken back completely Multiple take back clue completely.

Wherein, each interlock account for taking back the corresponding overdue loan to be taken back completely of clue completely, each is taken back completely Clue includes: multiple class another characteristics, and each class another characteristic is believed all in accordance with the account of the corresponding interlock account The customer information of the client of breath and the overdue loan to be taken back completely determines, and one is taken back completely each classification in clue and only corresponds to one Feature；The interlock account of the overdue loan to be taken back completely includes the savings account in addition to arranging refund account of the client, And/or the savings account of the guarantor of the client.

Construction unit 502, for establishing decision assistant model according to multiple historical samples.

Wherein, each historical sample includes one and takes back clue completely and described take back the corresponding interlock account of clue completely Mark of withholing, the mark of withholing shows whether the deposit of corresponding interlock account is used to pay that described to take back clue completely corresponding The principal and interest of any overdue loan.

Recommendation unit 503, for taking back clue completely using the decision assistant model analysis is each described, based on the analysis results Determine that clue is taken back in the multiple recommendation taken back completely in clue completely.

Wherein, recommend to take back clue completely and be supplied to approving person to analyze, Examination and approval personnel formulate wait take back overdue loan completely Money takes back strategy completely.

Optionally, when strategy is taken back in approving person's formulation completely, one or more recommendations is can choose and take back account completely, selected Recommend the deposit for taking back account completely for paying the principal and interest of overdue loan to be taken back completely.It certainly, can also be with according to the analysis of approving person Without recommending the deposit for taking back account completely to pay the principal and interest of overdue loan to be taken back completely.Wherein, recommendation is taken back account completely and is referred to, recommendation unit The corresponding account of clue is taken back in 503 recommendations determined completely.

Optionally, recommendation unit 503 may include:

Analytical unit obtains pair for taking back clue completely using described in decision assistant model analysis for each clue is taken back completely The analysis result answered.

Judging unit, for taking back whether the corresponding analysis result of clue meets described in judgement completely for each clue is taken back completely It is preset to take back condition completely.

Optionally, above-mentioned apparatus further include:

Optimize unit 504, for taking back clue completely for each recommendation, determines that the corresponding association of clue is taken back in the recommendation completely It is described after the principal and interest for taking back overdue loan completely whether account is used to pay, and takes back clue completely using the recommendation and the recommendation is clear The mark of withholing of the corresponding interlock account of take-up rope optimizes the decision assistant model.

Wherein, the deposit identified according to corresponding interlock account of withholing for recommending to take back the corresponding interlock account of clue completely Whether for paying the principal and interest setting of the overdue loan to be taken back completely.

Referring to FIG. 6, construction unit 502 includes:

First computing unit 601, it is described for being calculated using random forests algorithm based on the multiple historical sample The different degree of each classification of multiple historical samples.

Second determination unit 602 is determined the multiple for the different degree according to each classification of the historical sample Crucial classification in classification.

Sample construction unit 603, for being directed to each historical sample, with the spy of the crucial classification of the historical sample The mark of withholing of sign and the historical sample constructs the corresponding crucial sample of the historical sample.

Second computing unit 604, for calculating an initial value conduct according to the mark of withholing of the multiple crucial sample Current submodel, and 1 is set by the number of iterations.

Each crucial sample is calculated for utilizing the current submodel in third computing unit 605 Negative gradient.

4th computing unit 606, for the negative ladder according to each crucial sample and each crucial sample Degree, is calculated submodel updated value.

Updating unit 607, for being updated to the current submodel based on the submodel updated value, and The number of iterations is set to be incremented by 1.

Judging unit 608, for as the current submodel, and judging to work as by updated current submodel Whether preceding the number of iterations is greater than preset threshold value, if the current iteration number be less than or equal to the threshold value, triggering described in Second computing unit 604 utilizes the current submodel, and the negative gradient of each crucial sample is calculated；If described The number of iterations is greater than the threshold value, then current submodel is determined as decision assistant model.

Referring to FIG. 7, above-mentioned first computing unit 601 may include with lower unit:

Decision tree establishes unit 701, for establishing M decision tree based on the multiple historical sample；Wherein, the M is Preset positive integer, each described decision tree utilize C4.5 algorithm to establish, and described M decision tree composition one random gloomy Woods.

First error computing unit 702, for calculating and each determining in the random forest according to the multiple historical sample First bag of outer data error of plan tree.

Second error calculation unit 703, for being directed to each classification, to the institute of the category in the multiple historical sample There is feature to apply random noise, each described decision is then calculated according to the multiple historical sample after application random noise Second bag of outer data error corresponding with the classification of tree.

Different degree computing unit 704, for each classification, according to the second error calculation of first error and the classification The different degree of the classification；Wherein, the first error includes number outside first bag of each decision tree in the random forest According to error, the second error of the classification includes corresponding with the classification second of each decision tree in the random forest The outer data error of bag.

The device provided by the embodiments of the present application for taking back overdue loan completely, concrete operating principle can be any with reference to the application The method for taking back overdue loan completely that embodiment provides, details are not described herein again.

Overdue loan provided by the embodiments of the present application takes back device completely, receive treat take back completely overdue loan take back request completely after, Have acquiring unit 501 obtain respectively correspond multiple interlock accounts take back clue completely, each takes back the corresponding association of clue completely Account, then recommendation unit 503 is analyzed using decision assistant model clue is taken back completely, and decision assistant model is by construction unit 502 establish according to multiple historical samples, are determined based on the analysis results from taking back completely for overdue loan to be taken back completely and determine recommendation in clue Take back clue completely, clue is taken back in recommendation, which completely, can be supplied to approving person, formulate taking back completely for overdue loan to be taken back completely as approving person The foundation of strategy；Wherein, multiple interlock accounts include the deposit other than the agreement refund account of the client of overdue loan to be taken back completely The savings account of the guarantor of account and client, each takes back clue completely and includes multiple class another characteristics, and each classification Feature is determined all in accordance with the account information of interlock account and the customer information of overdue loan client to be taken back completely.The embodiment of the present application mentions The method of confession, which can be treated automatically, to be taken back multiple clues of taking back completely of overdue loan completely and is screened, and line is taken back in the recommendation screened completely Rope can be used as the foundation for taking back strategy completely for formulating overdue loan to be taken back completely, and the approving person of bank is assisted to carry out overdue loan Take back work completely.Relative to the existing method for taking back overdue loan completely, method provided by the embodiments of the present application makes approving person only The recommendation for needing Analysis and Screening to come out, which is taken back clue completely and can be formulated, takes back strategy completely, without considering the every of overdue loan to be taken back completely One is taken back completely clue, therefore method provided by the embodiments of the present application can effectively reduce the workload of approving person, improve to That takes back overdue loan completely takes back efficiency completely.

Professional technician can be realized or using the present invention.Profession of the various modifications to these embodiments to this field It will be apparent for technical staff, the general principles defined herein can not depart from spirit or model of the invention In the case where enclosing, realize in other embodiments.Therefore, the present invention will not be limited to the embodiments shown herein, And it is to fit to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. a kind of method for taking back overdue loan completely characterized by comprising

Receive wait take back completely overdue loan take back request completely after, obtain that overdue loan to be taken back completely is corresponding multiple to take back clue completely；Wherein, Each takes back an interlock account of the corresponding overdue loan to be taken back completely of clue completely, and it includes: more that each, which takes back clue completely, A class another characteristic, each class another characteristic is all in accordance with the account information of the corresponding interlock account and described wait take back completely The customer information of the client of overdue loan determines, and one is taken back completely each classification in clue and only corresponds to a feature；It is described to clear The interlock account for receiving overdue loan includes the savings account in addition to arranging refund account of the client, and/or, the client Guarantor savings account；

Clue is taken back completely using decision assistant model analysis is each described, determines the multiple take back completely in clue based on the analysis results Recommendation take back clue completely；Wherein, clue conduct is taken back in the recommendation completely, formulate the overdue loan to be taken back completely take back completely strategy according to According to, the decision assistant model is established according to multiple historical samples, each historical sample include one take back completely clue with And the mark of withholing for taking back the corresponding interlock account of clue completely, the mark of withholing show that the deposit of corresponding interlock account is It is no to be used to pay the principal and interest for taking back the corresponding any overdue loan of clue completely.

2. the method according to claim 1, wherein described utilize takes back line completely described in decision assistant model analysis Rope determines that clue is taken back in the multiple recommendation taken back completely in clue completely based on the analysis results, comprising:

Clue is taken back completely for each, is taken back the corresponding analysis result of clue described in judgement completely and whether is met and preset takes back condition completely；

Clue is taken back completely for each, it, will be described clear if described take back completely takes back condition completely described in the corresponding analysis result satisfaction of clue Take-up rope is determined as recommending to take back clue completely.

3. the method according to claim 1, wherein described utilize takes back line completely described in decision assistant model analysis Rope, after determining that clue is taken back in the multiple recommendation taken back completely in clue completely based on the analysis results, further includes:

Take back clue completely for each recommendation, determine the recommendation take back completely the corresponding interlock account of clue whether be used to pay it is described After the principal and interest for taking back overdue loan completely, clue is taken back completely using the recommendation and the corresponding interlock account of clue is taken back in the recommendation completely Mark of withholing optimize the decision assistant model；Wherein, the mark of withholing for recommending to take back the corresponding interlock account of clue completely According to the deposit of corresponding interlock account whether for paying the principal and interest setting of the overdue loan to be taken back completely.

4. the method according to claim 1, wherein the establishment process of the decision assistant model includes:

The crucial classification in the multiple classification is determined according to the different degree of each classification of the historical sample；Wherein, often The different degree of a classification is based on the multiple historical sample using random forests algorithm and is calculated；

For each historical sample, with the crucial class another characteristic of the historical sample and withholing for the historical sample Mark constructs the corresponding crucial sample of the historical sample；

An initial value is calculated as current submodel according to the mark of withholing of the multiple crucial sample, and by the number of iterations It is set as 1；

According to the negative gradient of each crucial sample and each crucial sample, submodel updated value is calculated；

Based on the submodel updated value, the current submodel is updated, and the number of iterations is made to be incremented by 1；

Using updated current submodel as current submodel, and judge whether current iteration number is greater than preset threshold Value, if the current iteration number is less than or equal to the threshold value, return execution is described to utilize the current submodel, calculates Obtain the negative gradient of each crucial sample；If the current iteration number is greater than the threshold value, mould will be currently assisted Type is determined as decision assistant model.

5. according to the method described in claim 4, it is characterized in that, described be based on the multiple history using random forests algorithm Sample calculates the process of the different degree of each classification, comprising:

M decision tree is established based on the multiple historical sample；Wherein, the M is preset positive integer, each described decision Tree is established using C4.5 algorithm, and the M decision tree constitutes a random forest；

For each classification, random noise is applied to all features of the category in the multiple historical sample, then basis The multiple historical sample after applying random noise calculates second bag corresponding with the classification of each decision tree Outer data error；

For each classification, the different degree of the classification according to the second error calculation of first error and the classification；Wherein, The first error includes first bag of outer data error of each decision tree in the random forest, and the second of the classification misses Difference includes second bag of outer data error corresponding with the classification of each decision tree in the random forest.

6. a kind of device for taking back overdue loan completely characterized by comprising

Acquiring unit, for receive wait take back completely overdue loan take back request completely after, it is corresponding multiple to obtain overdue loan to be taken back completely Take back clue completely；Wherein, each interlock account for taking back the corresponding overdue loan to be taken back completely of clue completely, each is taken back completely Clue includes: multiple class another characteristics, and each class another characteristic is believed all in accordance with the account of the corresponding interlock account The customer information of the client of breath and the overdue loan to be taken back completely determines, and one is taken back completely each classification in clue and only corresponds to one Feature；The interlock account of the overdue loan to be taken back completely includes the savings account in addition to arranging refund account of the client, And/or the savings account of the guarantor of the client；

Construction unit, for establishing decision assistant model according to multiple historical samples；Wherein, each historical sample includes One is taken back completely clue and the mark of withholing for taking back the corresponding interlock account of clue completely, the mark of withholing show corresponding pass Whether the deposit of connection account is for paying the principal and interest for taking back the corresponding any overdue loan of clue completely；

Recommendation unit is determined based on the analysis results for taking back clue completely using the decision assistant model analysis is each described Clue is taken back in the multiple recommendation taken back completely in clue completely；Wherein, it is described overdue wait take back completely as formulating to take back clue completely for the recommendation The foundation for taking back strategy completely of loan.

7. device according to claim 6, which is characterized in that the recommendation unit includes:

Analytical unit, for taking back clue completely using described in decision assistant model analysis, obtaining corresponding for each clue is taken back completely Analyze result；

Judging unit, for for each clue is taken back completely, taken back completely described in judgement the corresponding analysis result of clue whether meet it is default Take back condition completely；

First determination unit, for for each taking back clue completely, if it is described take back completely the corresponding analysis result of clue meet it is described clearly The clue of taking back completely then is determined as recommending to take back clue completely by receipt part.

8. device according to claim 6, which is characterized in that further include:

Optimize unit, for taking back clue completely for each recommendation, determines that the recommendation takes back the corresponding interlock account of clue completely and is It is no described after the principal and interest for taking back overdue loan completely for paying, clue is taken back completely using the recommendation and clue is taken back in the recommendation completely The mark of withholing of corresponding interlock account optimizes the decision assistant model；Wherein, the corresponding association of clue is taken back in the recommendation completely The mark of withholing of account is set according to the principal and interest whether deposit of corresponding interlock account is used to pay the overdue loan to be taken back completely It sets.

9. device according to claim 6, which is characterized in that the construction unit includes:

The multiple history is calculated for being based on the multiple historical sample using random forests algorithm in first computing unit The different degree of each classification of sample；

Second determination unit, the different degree for each classification according to the historical sample are determined in the multiple classification Crucial classification；

Sample construction unit, for being directed to each historical sample, with the crucial class another characteristic of the historical sample, Yi Jisuo The mark of withholing for stating historical sample constructs the corresponding crucial sample of the historical sample；

Second computing unit, for calculating an initial value as current auxiliary according to the mark of withholing of the multiple crucial sample Model, and 1 is set by the number of iterations；

The negative gradient of each crucial sample is calculated for utilizing the current submodel in third computing unit；

4th computing unit is calculated for the negative gradient according to each crucial sample and each crucial sample To submodel updated value；

Updating unit is updated the current submodel, and make described change for being based on the submodel updated value Generation number is incremented by 1；

Judging unit, for as the current submodel, and judging current iteration time for updated current submodel Whether number is greater than preset threshold value, if the current iteration number is less than or equal to the threshold value, triggers described second and calculates Unit utilizes the current submodel, and the negative gradient of each crucial sample is calculated；If the number of iterations is big In the threshold value, then current submodel is determined as decision assistant model.

10. device according to claim 9, which is characterized in that first computing unit includes:

Decision tree establishes unit, for establishing M decision tree based on the multiple historical sample；Wherein, the M be it is preset just Integer, each described decision tree utilize C4.5 algorithm to establish, and the M decision tree constitutes a random forest；

First error computing unit, for calculating each decision tree in the random forest according to the multiple historical sample First bag of outer data error；

Second error calculation unit, for being directed to each classification, to all features of the category in the multiple historical sample Apply random noise, then according to apply random noise after the multiple historical sample calculate each decision tree with The corresponding second bag of outer data error of the classification；

Different degree computing unit, for each classification, the class according to the second error calculation of first error and the classification Other different degree；Wherein, the first error includes first bag of outer data error of each decision tree in the random forest, Second error of the classification includes number outside second bag corresponding with the classification of each decision tree in the random forest According to error.