CN110322085A

CN110322085A - A kind of customer churn prediction method and apparatus

Info

Publication number: CN110322085A
Application number: CN201810272573.8A
Authority: CN
Inventors: 刘军; 张帆
Original assignee: Nine Chapter Yunji Technology Co Ltd Beijing
Current assignee: Nine Chapter Yunji Technology Co Ltd Beijing
Priority date: 2018-03-29
Filing date: 2018-03-29
Publication date: 2019-10-11

Abstract

The present invention provides a kind of customer churn prediction method and apparatus, this method comprises: obtaining customer data to be predicted；The customer data to be predicted is input to the data that customer revenue is calculated in Model of customer churn prediction, wherein the Model of customer churn prediction is obtained using the historical customer data training of multiple clients.By the above-mentioned means, the present invention can accurately predict customer churn, the client for being possible to be lost is found, to be kept in time to client, the cost of customer retention is saved, and improve the effect kept, effectively solves the losing issue of client.

Description

A kind of customer churn prediction method and apparatus

Technical field

The present invention relates to technical field of data processing more particularly to a kind of customer churn prediction method and apparatus.

Background technique

Since customer churn is huge on enterprise profit influence, existing customer losing issue is widely paid attention to.It is competing More and more fierce banking, Problem of Customer-Churn are equally urgently to be resolved.Existing customer churn prediction method is monitoring client Total assets, when client's total assets volume decline in a short time it is larger for example decline 250,000 when, determine client be possible to be lost, to client into Row is kept.Using this customer churn prediction method, prediction is not accurate enough, and customer retention is at high cost and income is smaller.

Summary of the invention

In view of this, the present invention provides a kind of customer churn prediction method and apparatus, for solving existing customer churn The true problem of prediction technique forecasting inaccuracy.

In order to solve the above technical problems, the present invention provides a kind of customer churn prediction method, comprising:

Obtain customer data to be predicted；

The customer data to be predicted is input to the data that customer revenue is calculated in Model of customer churn prediction, In, the Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

Preferably, the described customer data to be predicted is input in Model of customer churn prediction calculates customer revenue Data the step of before, further includes:

Training sample set is obtained, the training sample set is the set of the historical customer data of multiple clients；

Obtain at least two algorithm models to be selected；

For each algorithm model to be selected, according to the parameter of the algorithm model to be selected of setting and input Training sample set is trained the algorithm model to be selected, obtains the data of customer revenue；

The true loss that the data of the customer revenue compared and the threshold value defined based on scheduled customer churn are determined The data of client, obtain comparison result；

When the comparison result is unsatisfactory for preset condition, the parameter of the algorithm model to be selected, new root of laying equal stress on are adjusted The algorithm model to be selected is instructed according to the parameter of the algorithm model to be selected of adjustment and the training sample set of input Practice, until the comparison result meets the preset condition, obtains the algorithm model of training completion；

The algorithm model completed to training is assessed, and assessment result is obtained；

Compare at least two training complete algorithm models assessment results, select one training complete algorithm model as The Model of customer churn prediction.

Preferably, the algorithm model to be selected include logistic regression algorithm model, it is Bagging algorithm model, random gloomy At least two in woods algorithm model, AdaBoost algorithm model, Voting Model, Stack Model and neural network algorithm model.

Preferably, algorithm model the step of assessing, obtaining assessment result that described pair of training is completed includes:

According to default evaluation index, the algorithm model completed to training is assessed, the default evaluation index includes Training sample set predictablity rate, test sample collection predictablity rate, area under the curve AUC score, F score and Kappa coefficient At least one of, the test sample collection is the set of the historical customer data of multiple clients, what the test sample was concentrated Customer data is different from the customer data that the training sample is concentrated.

Preferably, the default evaluation index includes: that training sample set predictablity rate and test sample collection prediction are accurate Rate；

Described pair includes: the step of training the algorithm model completed to assess, obtain assessment result

Obtain test sample collection；

The test sample collection is input in the algorithm model that the training is completed, obtains the data of customer revenue；

The data of obtained customer revenue and the data of the true customer revenue, meter are trained according to the training sample set Calculate the training sample set predictablity rate；

The data for the customer revenue predicted according to the test sample collection and the data of the true customer revenue, meter Calculate the test sample collection predictablity rate；

Compare the training sample set predictablity rate and the test sample collection predictablity rate, obtains assessment result.

Preferably, the algorithm model to be selected includes an algorithm model, alternatively, including at least two algorithm models；

When the algorithm model to be selected includes at least two algorithm model, wherein at least two algorithm model In, algorithm model of the algorithm model as the second layer, algorithm model of remaining algorithm model as first layer；It is described to be directed to Each algorithm model to be selected, according to the training sample set of the parameter of the algorithm model to be selected of setting and input, The step of being trained to the algorithm model to be selected, obtain the data of customer revenue include:

The algorithm model that the training sample set is input to the first layer is trained, pilot process data are obtained；

The algorithm model that the pilot process data are input to the second layer is trained, the number of customer revenue is obtained According to.

Preferably, it is described select one training complete algorithm model as the Model of customer churn prediction the step of it Afterwards, further includes:

Obtain the client characteristics of the Model of customer churn prediction output；

The client characteristics are inputted in decision-tree model, show the decision tree diagram based on the client characteristics.

Preferably, described including predetermined observation phase and preset table current interior customer data in the historical customer data The predetermined observation phase is current earlier than the preset table.

It preferably, further include customer data in the pre-determined stability phase, the pre-determined stability phase in the historical customer data Positioned at the predetermined observation phase and preset table it is current between.

Preferably, the step of acquisition training sample set includes:

Data prediction is carried out to training sample set to be processed, the training sample set after obtaining data prediction.

Preferably, the data prediction includes missing values calculating, outlier exclusion, data transformation, nondimensionalization and returns At least one of one change.

Preferably, the step of acquisition training sample set includes:

The client characteristics that training sample to be processed is concentrated are screened using feature selection module, determine that the client of selection is special Sign screens training sample set using the client characteristics of selection.

Preferably, the feature selection module is using Chi-square Test, Pearson correlation coefficients method, extremely tree Method for Feature Selection The client characteristics concentrated at least one of recursive feature null method screening training sample to be processed.

The material information of the client characteristics of selection is calculated by the Model of customer churn prediction；

According to the material information of client characteristics, the client characteristics that the feature selection module uses are adjusted.

Preferably, the step of acquisition training sample set includes:

Unbalanced data processing is carried out to training sample set to be processed, the training sample set that obtains that treated, after processing Training sample concentrate the difference of the quantity of customer revenue and non-attrition customer to be less than preset threshold.

Preferably, the described customer data to be predicted is input in Model of customer churn prediction calculates customer revenue Data the step of after further include:

Obtain the data for retrieving client；

According to the data for retrieving client, the parameter of the Model of customer churn prediction is adjusted.

Obtain the data for retrieving client；

According to the data for retrieving client, the threshold value that the customer churn defines is adjusted.

The present invention also provides a kind of customer churn prediction devices, comprising:

Module is obtained, for obtaining customer data to be predicted；

Computing module is lost visitor for the customer data to be predicted to be input to calculate in Model of customer churn prediction The data at family, wherein the Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

Preferably, the customer churn prediction device further include:

Training module, for obtaining training sample set, the training sample set is the historical customer data of multiple clients Set；Obtain at least two algorithm models to be selected；For each algorithm model to be selected, according to the described to be selected of setting Algorithm model parameter and input training sample set, the algorithm model to be selected is trained, customer revenue is obtained Data；The true loss visitor that the data of the customer revenue compared and the threshold value defined based on scheduled customer churn are determined The data at family, obtain comparison result；When the comparison result is unsatisfactory for preset condition, the algorithm model to be selected is adjusted Parameter, new root of laying equal stress on is according to the parameter of the algorithm model to be selected of adjustment and the training sample set of input to the calculation to be selected Method model is trained, until the comparison result meets the preset condition, obtains the algorithm model of training completion；

Evaluation module, the algorithm model for completing to training are assessed, and assessment result is obtained；

Selecting module selects a training to complete for comparing the assessment result for the algorithm model that at least two training are completed Algorithm model as the Model of customer churn prediction.

Preferably, the evaluation module, for according to evaluation index is preset, the algorithm model completed to training to be commented Estimate, the default evaluation index includes training sample set predictablity rate, test sample collection predictablity rate, area under the curve At least one of AUC score, F score and Kappa coefficient, the test sample collection are the historical customer data of multiple clients Set, the customer data that the test sample is concentrated are different from the customer data that the training sample is concentrated.

The evaluation module includes: data submodule, computational submodule and Comparative sub-module,

The data submodule, for obtaining test sample collection；The test sample collection is input to the training to complete Algorithm model in, obtain the data of customer revenue；

The computational submodule, the data of customer revenue for being obtained according to training sample set training and described true The data of real customer revenue calculate the training sample set predictablity rate；The stream predicted according to the test sample collection The data of client and the data of the true customer revenue are lost, the test sample collection predictablity rate is calculated；

The Comparative sub-module, it is accurate for the training sample set predictablity rate and test sample collection prediction Rate obtains assessment result.

The training module, for when the algorithm model to be selected includes at least two algorithm model, wherein described In at least two algorithm models, algorithm model of the algorithm model as the second layer, remaining algorithm model is as first layer Algorithm model；The algorithm model that the training sample set is input to the first layer is trained, pilot process data are obtained； The algorithm model that the pilot process data are input to the second layer is trained, the data of customer revenue are obtained.

Preferably, the customer churn prediction device further include:

Image display module, for obtaining the client characteristics of the Model of customer churn prediction output；The client is special In sign input decision-tree model, the decision tree diagram based on the client characteristics is shown.

Preferably, the training module includes the first processing submodule, for counting to training sample set to be processed Data preprocess, the training sample set after obtaining data prediction.

Preferably, the training module includes screening submodule, for screening instruction to be processed using feature selection module Practice the client characteristics in sample set, determine the client characteristics of selection, screens training sample set using the client characteristics of selection.

Preferably, the customer churn prediction device further include:

The first adjustment module, the importance letter of the client characteristics for calculating selection by the Model of customer churn prediction Breath；According to the material information of client characteristics, the client characteristics that the feature selection module uses are adjusted.

Preferably, the training module includes second processing submodule, for carrying out not to training sample set to be processed Equalization data processing, the training sample set that obtains that treated, treated, and training sample concentrates customer revenue and non-attrition customer Quantity difference be less than preset threshold.

Preferably, the customer churn prediction device further include:

Second adjustment module, for obtaining the data for retrieving client；According to the data for retrieving client, client's stream is adjusted Lose the parameter of prediction model.

Preferably, the customer churn prediction device further include:

Third adjusts module, for obtaining the data for retrieving client；According to the data for retrieving client, client's stream is adjusted Lose the threshold value of definition.

The present invention also provides a kind of customer churn prediction device, including memory, processor and it is stored in the storage On device and the computer program that can run on the processor；It is characterized in that, reality when the processor executes described program Existing above-mentioned customer churn prediction method.

The present invention also provides a kind of computer readable storage mediums, are stored thereon with computer program, which is characterized in that The program realizes the step in above-mentioned customer churn prediction method when being executed by processor.

The advantageous effects of the above technical solutions of the present invention are as follows:

By obtaining customer data to be predicted；The customer data to be predicted is input to Model of customer churn prediction The middle data for calculating customer revenue, the present invention can accurately predict customer churn, find the client for being possible to be lost, from And client is kept in time, the cost of customer retention is saved, and improve the effect kept, can effectively solve the problem that the stream of client Mistake problem.

Detailed description of the invention

Fig. 1 is the flow diagram of the customer churn prediction method of the embodiment of the present invention one；

Fig. 2 is the comparison result signal of an application scenarios difference prediction algorithm model in some preferred embodiments of the invention Figure；

Fig. 3 is the confusion matrix schematic diagram of the display model error of some preferred embodiments of the invention；

Fig. 4 is the single model training flow diagram of some preferred embodiments of the invention；

Fig. 5 is the Stack Model training flow diagram of some preferred embodiments of the invention；

Fig. 6 is the sequence schematic diagram of the client characteristics importance of an application scenarios of some preferred embodiments of the invention；

Fig. 7 is the structural schematic diagram of the customer churn prediction device of the embodiment of the present invention two；

Fig. 8 is the structural schematic diagram of the customer churn prediction device of some preferred embodiments of the invention.

Specific embodiment

In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention Attached drawing, the technical solution of the embodiment of the present invention is clearly and completely described.Obviously, described embodiment is this hair Bright a part of the embodiment, instead of all the embodiments.Based on described the embodiment of the present invention, ordinary skill Personnel's every other embodiment obtained, shall fall within the protection scope of the present invention.

Referring to Fig. 1, Fig. 1 is the flow diagram of the customer churn prediction method of the embodiment of the present invention one, this method packet It includes:

Step 10: obtaining customer data to be predicted；

Step 20: customer data to be predicted is input to the data that customer revenue is calculated in Model of customer churn prediction, Wherein, Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

Customer churn prediction method provided in an embodiment of the present invention, can accurately predict customer churn, find The client that may be lost saves the cost of customer retention, and improve the effect kept, energy to keep in time to client Enough losing issues for effectively solving client.

Below to how Model of customer churn prediction being trained to be illustrated.

In some embodiment of the invention, before step 20 can include:

Step 11: obtaining training sample set, the training sample set is the set of the historical customer data of multiple clients；

It may include that predetermined observation phase and preset table are current interior in some concrete scenes, in the historical customer data Customer data, the predetermined observation phase are current earlier than the preset table.

It further, further include customer data in the pre-determined stability phase, the pre-determined stability in the historical customer data Phase be located at the predetermined observation phase and preset table it is current between.Such as: obtain the continuous predetermined observation phase (such as: in January, 2017 To March), the history of pre-determined stability phase (such as: in April, 2017 to May), preset table current (such as: in June, 2017 to August) Customer data.

Certainly, the predetermined observation phase, pre-determined stability phase and preset table is current can carry out freely setting according to the demand of user It is fixed, it is more flexible and convenient.

Step 12: obtaining at least two algorithm models to be selected；

Step 13: for each algorithm model to be selected, according to the parameter of the algorithm model to be selected of setting With the training sample set of input, the algorithm model to be selected is trained, the data of customer revenue are obtained；

The parameter of the algorithm model to be selected can be the parameter that user sets according to demand, is also possible to software and is locating The parameter set automatically during reason, the present invention are not construed as limiting.

Step 14: the data of the customer revenue compared and the threshold value defined based on scheduled customer churn determine true The data of real customer revenue, obtain comparison result；

Step 15: when the comparison result is unsatisfactory for preset condition, the parameter of the algorithm model to be selected is adjusted, and Again according to the parameter of the algorithm model to be selected of adjustment and the training sample set of input to the algorithm model to be selected It is trained, until the comparison result meets the preset condition, obtains the algorithm model of training completion；

The purpose of step 14 and step 15 be the obtained data of customer revenue of verifying prediction whether with based on scheduled visitor The data for the true customer revenue that the threshold value of family definitions of attrition determines are almost the same.Wherein, the data of true customer revenue can root The threshold calculations for defining according to customer churn and being defined based on scheduled customer churn are obtained.In training process, it can be used and be based on The cross validation method of grid search, set a preset condition, according to the result of the comparison with the preset condition of setting, to algorithm mould The parameter of type carries out tuning, until training is completed.For example, preset condition can be set are as follows: the accuracy rate of prediction is greater than or equal to When 75%, training is completed, and the data of the data for the customer revenue that comparison prediction obtains and true customer revenue calculate the standard of prediction True rate, when the accuracy rate of prediction is greater than or equal to 75%, training is completed；When the accuracy rate of prediction is less than 75%, adjustment is calculated The parameter of method model, re-starts training, until the accuracy rate of prediction is greater than or equal to 75%, obtains the algorithm mould of training completion Type.For another example, preset condition can be set are as follows: the AUC score of algorithm model, i.e. area under the curve (area under the Curve, abbreviation AUC) when being greater than or equal to 0.8, training is completed, and the data of customer revenue that are obtained according to prediction and true is lost The data of client obtain the AUC score of algorithm model, and when the AUC score of algorithm model is greater than or equal to 0.8, training is completed； When the AUC score of algorithm model is less than 0.8, the parameter of adjustment algorithm model re-starts training, until algorithm model AUC score is greater than or equal to 0.8, obtains the algorithm model of training completion.

Step 16: the algorithm model completed to training is assessed, and assessment result is obtained；

Specifically, user can set default evaluation index according to demand, the algorithm model completed to training is assessed, and is obtained To assessment result.

Step 17: comparing the assessment result for the algorithm model that at least two training are completed, the algorithm for selecting a training to complete Model is as the Model of customer churn prediction.

In above-described embodiment, customer churn define can demand based on user, professional standard and experience setting, and can and When be adjusted optimization.Such as: can set customer churn be client preset table it is current in total assets mean value compared with predetermined observation Phase is reduced, and is set as 20% based on the threshold value that scheduled customer churn defines, i.e. total assets of the client in preset table is current is equal Value reduces by more than 20% compared with the predetermined observation phase as customer churn.Customer capital includes the deposit of client, financing etc..This can be set Value is annual average, month or the season mean value of client.If setting the mean value as the annual average of client, i.e., more than the average daily deposit of client It is current, pre- in preset table to calculate separately out client using formula " annual average=daily cash in banks remaining sum adds up to/number of days " for volume If the annual average of observation period.If a client preset table is current, the annual average in the predetermined observation phase is respectively 7800 yuan, 10000 yuan, then annual average of the client in preset table is current reduces by more than 20% compared with the predetermined observation phase, then the client is to be lost Client.

It certainly, also may include multiple numerical value, such as given threshold V1=based on the threshold value that scheduled customer churn defines 30%, V2=35% set total assets mean value of the client within the pre-determined stability phase and reduce compared with the predetermined observation phase less than V2, and pre- If it is customer churn that the total assets mean value in the performance phase, which reduces by more than V1 compared with the predetermined observation phase,.That is, in the pre-determined stability phase Interior total assets mean value reduces by more than 35% client compared with the predetermined observation phase, it is considered to be is difficult to the client retrieved, is not paying close attention to Range.I.e. client rapidly losing within the pre-determined stability phase is directly filtered, and reduces data processing amount, promotes data processing Speed and efficiency.Client is calculated again preset table is current, the total assets mean value in the predetermined observation phase, obtains customer revenue.If The value of V1 and V2 is excessively high, and only seldom people can be defined as being lost, even if prediction model is absolutely accurate and is keeping Stage can retrieve all clients, and the income for keeping the stage is still less；, whereas if the value of V1 and V2 is too low, although client has The fluctuation normally consumed, a large amount of clients can be defined as being lost, and such income also can be seldom.Therefore, can according to the actual situation and Demand adjusts threshold value V1, V2 in real time, more effectively to retrieve client.

Preferably, can according to actual needs, scheduled customer churn definition may include multiple conditions, according to different conditions The threshold value that customer churn defines is respectively set.Such as: the client that customer churn is more than 50,000 yuan as total assets mean value can be set, and Total assets mean value of the client in preset table is current reduces 25% compared with the predetermined observation phase.In this way, being clicked through first to rising for total assets Row limitation, pays close attention to the loss of high assets client, can further increase the effect kept.

For example, in an application scenarios, to 2016 and 2 years 2017 9210000 historic customer numbers of certain bank According to being sampled, 2,760,000 datas are obtained, can be used as the data of training sample set and test sample collection.It extracts wherein to preset and see The client that phase annual average is greater than 50,000 is examined, 280,000 datas are obtained.Pre-determined stability phase annual average is filtered out to reduce compared with the predetermined observation phase Client more than 35%, remaining 210,000 datas.Define further according to scheduled customer churn: client is total in preset table is current Assets mean value reduces by more than 30% compared with the predetermined observation phase, classifies to 210,000 datas, is lost and is labeled as 1, non-streaming lose-submission note It is 0, obtaining non-attrition customer is 180,000 people, and it is the data of true customer revenue that customer revenue, which is 30,000 people,.Based on identical pre- Fixed customer churn definition, the data that above-mentioned training sample set and test sample collection can be used are trained, the stream predicted The data of client are lost, and are compared with the data of true customer revenue, until comparison result meets preset condition, are trained The Model of customer churn prediction of completion.

In above-described embodiment, algorithm model to be selected may include logistic regression algorithm model, Bagging algorithm model, In random forests algorithm model, AdaBoost algorithm model, Voting Model, Stack Model and neural network algorithm model at least Two.

Wherein, Bagging algorithm model, random forests algorithm model, AdaBoost algorithm model, Voting Model and storehouse Model belongs to Integrated Algorithm model, and Integrated Algorithm model is by multiple single algorithm model (only comprising single algorithm model) collection At what is obtained.

Logistic regression algorithm model is the generalized linear model for being traditionally used for two classification problems, it is by sigmoid function Applied to the output of linear model, output area is compressed to [0,1] range.

Bagging (Bootstrap aggregating, abbreviation Bagging) algorithm model, i.e. bagging method, is a kind of collection At algorithm, it uses multiple strong learners, such as decision tree or nearest neighbor algorithm (K-Nearest Neighbor, abbreviation KNN), uses In solution overfitting problem.It first concentrates from initial data with there is the methods of sampling put back to extract n training sample, repeatedly k Wheel, obtains k number according to collection.It is used for training pattern using a training set every time later, obtains k model.All models are all adopted With consistent algorithm.Finally, for classification problem, final knot is obtained by the way of ballot to the result of k model prediction Fruit.Wherein, n and k is positive integer.

Random forests algorithm model is the evolution version of Bagging, and by combining multiple Weak Classifiers, final result passes through Mean value is voted or takes, so that the result of overall model accuracy with higher and Generalization Capability.It can obtain good achievement, It is mainly attributed to " random " and " forest ", one makes it have anti-over-fitting ability, and one keeps it more accurate.Random forest makes Use CART decision tree as weak learner.When generating each tree, the feature that each tree is chosen is only random choosing A few features out, general default takes the evolution of feature sum m, to ensure that feature randomness, reduces the phase between tree Guan Xing.

AdaBoost algorithm model is iteratively established on weak learner, in each iteration, can all add one New learner, and all existing learners all remain unchanged.All learners according to their performance (such as: it is accurate Property) be weighted, and after new learner is added, data are weighted again: the sample of mistake classification obtains more More weights, and the sample correctly classified reduces weight.Therefore, the learner mistake before new learner can increasingly focus on The sample of classification.

Voting Model chooses gradient and promotes sub-model, multinomial model-naive Bayesian and random forests algorithm model, adopts With the mode of soft ballot, each model prediction probability is averaged.

Stack Model is trained using two layers of algorithm model, using the output of the algorithm model of first layer as second The input of the single algorithm model of layer carrys out training algorithm model, can merge multiple models and be predicted that prediction result is more acurrate.

Neural network algorithm model, using three-layer neural network, middle layer activation primitive uses line rectification function (Rectified Linear Unit, abbreviation ReLU), also known as amendment linear unit, output layer activation primitive use Softmax letter Number, learning rate are preferably 0.001, and the number of iterations is preferably 1500 times, and batch data size is preferably 32.Optimized parameter Selection is then based on trellis search method.Grid refers to the multi-dimensional grid sky formed after different parameters difference value intersection Between, grid search is exactly all situations traversed in mesh space, is trained and verifies to model, and final choice goes out effect most The parameter combination of excellent (such as: test set accuracy rate highest).0.003 wherein, by attempting different learning rate: 0.001, 0.01,0.03,0.1, different the number of iterations: 500 times, 1000 times and 1500 times, different batch data sizes: 32, 64 and 128, determine learning rate, the number of iterations, the preferred value of batch data size be respectively 0.001,1500 time, 32.

In some currently preferred embodiments of the present invention, default evaluation index includes training sample set predictablity rate, test At least one of sample set predictablity rate, area under the curve AUC score, F score and Kappa coefficient, the test sample Collection is the set of the historical customer data of multiple clients, and the customer data and the training sample that the test sample is concentrated are concentrated Customer data it is different.

Wherein it is possible to the data that prediction obtains customer revenue be carried out, in conjunction with the training according to the algorithm model that training is completed The data of true customer revenue, calculate training sample set predictablity rate in sample set.Similarly, test sample collection can be calculated Predictablity rate.

The numerical value of AUC score, AUC score is bigger, shows that the precision of classification is higher.

F score is a kind of index for being used to measure two disaggregated model accuracy in statistics.It has combined classification mould The accuracy rate and recall rate of type.F score can be regarded as a kind of weighted average of model accuracy rate and recall rate, its maximum Value is 1, and minimum value is 0.F score is higher, illustrates that disaggregated model is more steady.

Kappa coefficient is a kind of index for measuring nicety of grading, and the calculating of Kappa coefficient is based on confusion matrix.

In order to avoid there is the case where over-fitting in model, it is preferable that default evaluation index includes: that training sample set prediction is quasi- True rate and test sample collection predictablity rate；At this point, step 16 the following steps are included:

Step 161: obtaining test sample collection；

Specifically, the historical customer data that can be will acquire is divided into two parts, the historical customer data of a portion is made For training sample set, the historical customer data of another part as test sample collection, such as: wherein 70% historical customer data As training sample set, 30% historical customer data so may insure the consistency of data as test sample collection.

Step 162: the algorithm model completed for each training, input test sample set obtain the data of customer revenue；

Step 163: the data of obtained customer revenue and the data of true customer revenue, meter are trained according to training sample set Calculate training sample set predictablity rate；

Step 164: the data for the customer revenue predicted according to test sample collection and the data of true customer revenue, meter Calculate test sample collection predictablity rate；

Step 165: comparing training sample set predictablity rate and test sample collection predictablity rate, obtain assessment result.

The algorithm model completed using training sample set training may only prediction effect be good on the training sample set, and Prediction effect is poor on other samples.Therefore, it is necessary to calculate training sample set predictablity rate and test sample collection predictablity rate, There is the case where over-fitting in the algorithm model for avoiding training from completing.

If the difference of calculated training sample set predictablity rate and test sample collection predictablity rate is less than or equal to Almost, then there is not the case where over-fitting, prediction effect in the algorithm model that training is completed for the comparison threshold value of setting, i.e. the two It is good.If test sample collection predictablity rate is more much smaller than training sample set predictablity rate, the difference of the two is greater than the ratio of setting Compared with threshold value, then illustrates algorithm model the case where there are over-fittings of training completion, further progress is needed to adjust.

Further, default evaluation index includes: test sample collection predictablity rate and AUC score.At this point, step 16 is wrapped Include following steps:

Step 1601: obtaining the AUC score for the algorithm model that each training is completed；

Step 1602: the AUC score for the algorithm model that each training is completed being arranged from high to low, chooses predetermined number The algorithm model that training is completed；

Step 1603: obtaining test sample collection；

Step 1604: the algorithm model that the training for the predetermined number of selection is completed, input test sample set are flowed Lose the data of client；

In the step, test sample collection can be separately input into the algorithm model that the training selected in step 1602 is completed In, for each selected model, obtain the data of customer revenue.

Step 1605: the data for the customer revenue predicted according to test sample collection and the data of true customer revenue, Calculate test sample collection predictablity rate；

Step 1606: comparing test sample collection predictablity rate, obtain assessment result.

Referring to Fig. 2, in an application scenarios, algorithm model to be selected is respectively as follows: logistic regression algorithm model, random gloomy Woods algorithm model, Bagging algorithm model, AdaBoost algorithm model, Voting Model and neural network algorithm model.For every A algorithm model to be selected, input training sample set are trained, and obtain the algorithm model of training completion.It is complete to obtain each training At algorithm model AUC score, select AUC score it is highest 3 training complete algorithm models, be respectively as follows: random forest Algorithm model, Bagging algorithm model and neural network algorithm model.Test sample collection is separately input into selected this 3 again In the algorithm model that a training is completed, the data of customer revenue are obtained, in conjunction with the data of true customer revenue, calculate separately out this The test sample collection predictablity rate of 3 models, and be compared, obtain the test sample collection prediction of random forests algorithm model Accuracy rate highest is 81.66%, so selecting random forests algorithm model as the model of prediction customer churn probability.

In some currently preferred embodiments of the present invention, during model training, confusion matrix display model error can be used, Carry out the error evaluation of prediction model.Model error in the present embodiment be predicted as being lost but the practical client not being lost and It is predicted as the customer quantity that is not lost but actually lost.Referring to Fig. 3, the longitudinal axis is actual value, i.e., flowed based on scheduled client The data for the true customer revenue that the threshold value defined determines are lost, horizontal axis is predicted value, i.e. the stream based on prediction algorithm model prediction Lose the data of client.It is labeled as 1 by being lost, non-streaming lose-submission is denoted as 0, then the region in the upper left corner indicates predicted value and actual value is Non- loss, data markers are 00 (first position is predicted value, and second position is actual value)；The region in the lower right corner indicates pre- Measured value and actual value are to be lost, data markers 11；The region in the lower left corner indicates that predicted value is non-loss, but actual value is stream It loses, data markers 01；The region in the upper right corner indicates that predicted value is to be lost, but actual value is non-loss, data markers 10.It calculates Method model error is the smaller the better, therefore in model error data, and 11 and 00 accounting is the bigger the better.In the present embodiment, 4 in figure It is marked with customer quantity respectively in a region, convenient for allowing user more intuitively to experience the precision of classification results.

In some currently preferred embodiments of the present invention, algorithm model to be selected may include an algorithm model, alternatively, packet Include at least two algorithm models.

When algorithm model to be selected only includes an algorithm model, referring to Fig. 4, training sample set is input to this It is trained in algorithm model, exports the data of customer revenue.

When algorithm model to be selected includes at least two algorithm model, can be instructed using the method for Stack Model Practice, the algorithm model used in Stack Model includes at least two algorithm models, the algorithm at least two algorithm model Model can be Integrated Algorithm model, or other algorithm models, such as: traditional machine learning algorithm model and nerve net Network algorithm model.Referring specifically to Fig. 5, Stack Model is trained using two layers of algorithm model, by the algorithm mould of first layer Single algorithm model (i.e. model 4 in Fig. 5) of the output of type (i.e. model 1, model 2 and model 3 in Fig. 5) as the second layer Input carry out training algorithm model.

Specifically, using an algorithm model at least two algorithm models as the algorithm model of the second layer, remaining Algorithm model of the algorithm model as first layer；The algorithm model that training sample set is input to first layer is trained, is obtained Pilot process data；The algorithm model that pilot process data are input to the second layer is trained, the data of customer revenue are obtained.

The algorithm model of first layer can be gradient boosted tree algorithm model, neural network algorithm model, random forest and calculate At least one of method model, logistic regression algorithm model and KNN algorithm model, the algorithm model of the second layer can be gradient and mention Rise tree algorithm model, neural network algorithm model, random forests algorithm model, logistic regression algorithm model and KNN algorithm model Any of.Such as: first layer is 2 gradient boosted tree algorithm models, 2 neural network algorithm models, 2 random forests Algorithm model, 2 logistic regression algorithm models and 1 KNN algorithm model, the second layer are that 1 gradient promotes tree-model.

Customer revenue is predicted using Stack Model, has merged multiple models, than the method only with single model Prediction effect is more preferably.

In some currently preferred embodiments of the present invention, step 11 can include: data are carried out to training sample set to be processed Pretreatment, the training sample set after obtaining data prediction.

Wherein, the process of data prediction includes data cleansing and data normalization, the purpose is to comparing convenient for data and Assessment.Specifically, data prediction may include in missing values calculating, outlier exclusion, data transformation, nondimensionalization and normalization At least one of.Nondimensionalization includes Standardization Act, section pantography etc., and missing values calculate the data such as including Missing Data Filling Transformation includes polynomial data conversion etc..For example, data are deleted or are filled, wherein filling can be to fill default value, Such as: zero or average value.

In other embodiments of the invention, data can not also be pre-processed, be conducive to training sample set directly into Row training, the present invention are not construed as limiting.

In other preferred embodiment of the invention, step 11 can include: screened using feature selection module wait locate The client characteristics that the training sample of reason is concentrated, determine the client characteristics of selection, screen training sample using the client characteristics of selection Collection.

Specifically, the client characteristics filtered out are that have correlation more important for a possibility that customer revenue with client Feature.Such as: for bank, the client characteristics filtered out can be at least one of following client characteristics: become silver The time span of row client, personal fixed deposit, following 3 months finance product amount dues or the transaction of nearest 3 months monthly average The amount of money.

Common, for becoming the time span of bank client, it is lower to be lost possibility by time longer client；For Personal fixed deposit has the client of fixed deposit compared to the client of not fixed deposit, it is lower to be lost possibility；For future 3 There is within a month the overdue client of finance product, when deposit expires, client faces the selection that whether continue to participate in banking, therefore There is the client of loss intention that will withdraw consciously；For nearest 3 months monthly average transaction amount, with the increasing of transaction amount Add, customer churn can be significantly raised, therefore bank's reply frequently enhance your vigilance by transaction, because client before loss, has meter It withdraws fund with drawing, causes more frequently to trade.

Unscreened client characteristics are numerous, may include: client's essential information, consumer product holds information, client purchases Buy behavioural information, client trading behavior etc..Wherein, client's essential information include: open an account duration, as bank client time it is long Degree, residence, gender, age, occupation, marital status etc..Consumer product hold information include: total assets, deposit season it is average daily, The balance of deposits, whether generation hair client, whether current row financing client, whether hold fixed deposit, whether hold determine work it is convenient to both, whether I manage it withhold client, hold product quantity (including deposit product quantity, loan product quantity), this month financing expire stroke count, be It is no to open Internetbank, whether open Mobile banking etc..Customers buying behavior information includes: the regular number of accumulative purchase, accumulative purchase The regular amount of money, the accumulative amount of the loan, accumulative purchase financing number, adds up the purchase financing amount of money, is 3 nearest accumulative number of providing a loan The moon purchase finance product number, the nearest 3 months purchase finance product amount of money, nearest 3 months purchase Determined products numbers, nearest 3 A month purchase Determined products amount of money.Client trading behavior includes: this month transaction stroke count, this month transaction amount, last time transaction The amount of money, the nearest stroke count of transaction in 4th month to 6th month, is traded at nearest 3 months transaction stroke counts for nearest 7th month to 9th month It is stroke count, the nearest stroke count of transaction in 10th month to 12nd month, nearest 3 months monthly average transaction amount, 4th month to the 6th nearest It is a month monthly average transaction amount, nearest 7th month to 9th month monthly average transaction amount, 10th month to 12nd month nearest Monthly average transaction amount, last time transaction amount, nearest 3 months internet bank trade numbers, nearest 3 Ge Yue Web bank hand over The easy amount of money, nearest 3 months mobile banking transaction numbers, the nearest 3 months mobile banking transaction amount of money, nearest 3 months third parties hand over Easy number, the nearest 3 months third party transaction amount of money, nearest 6 months internet bank trade numbers, nearest 6 Ge Yue Web bank hand over The easy amount of money, nearest 6 months mobile banking transaction numbers, the nearest 6 months mobile banking transaction amount of money, nearest 6 months third parties hand over Easy number, the nearest 6 months third party transaction amount of money, nearest 1 year internet bank trade number, nearest 1 year internet bank trade The amount of money, nearest 1 year mobile banking transaction number, the nearest 1 year mobile banking transaction amount of money, nearest 1 year third party transaction time Several, nearest 1 year third party transaction amount of money etc..

Unscreened client characteristics are numerous, therefore are screened using feature selection module, and selecting can with customer churn The relevant Very Important Person feature of energy property, and redundancy feature is removed, to be screened to training sample set.

Further, feature selection module using Chi-square Test, Pearson correlation coefficients method, extreme tree Method for Feature Selection and At least one of recursive feature null method screens the client characteristics that training sample to be processed is concentrated.

Wherein, Chi-square Test is feature selection approach, calculates the chi-square statistics amount between independent variable and target variable, retains card Side is worth relatively large variable.In addition the value of characteristic variable must be non-negative.For example, if a client characteristics and loss correlation Close to 0, then it is assumed that this client characteristics does not have any predictive ability, is not put into prediction model.

Pearson correlation coefficients (Pearson Correlation Coefficient) are also referred to as Pearson product-moment correlation coefficient (Pearson product-moment correlation coefficient), is a kind of linearly dependent coefficient.Pearson's phase Relationship number is the statistic for reflecting two linear variable displacement degrees of correlation.For example, if two client characteristics correlations are close In 1, it is believed that the two client characteristics be it is same, one of client characteristics are only put into prediction model.

Extreme tree Method for Feature Selection is a kind of method (embedded embedded category method) in feature selecting.This method is base In trained machine learning model, is deleted according to the importance of client characteristics and select variable.

Recursive feature null method (Recursive Feature Elimination) is a kind of feature selection approach, is based on The variation coefficient or feature importance of algorithm output are deleted the small variable of importance, are fitted again later, delete, so It repeats.

Preferably, scalping can be carried out to client characteristics using Chi-square Test and Pearson correlation coefficients method, after scalping Client characteristics further screened using extreme tree Method for Feature Selection and recursive feature null method.

Specifically, the correlation between each client characteristics and customer churn possibility first can be calculated with Chi-square Test, sieve Select the high client characteristics of correlation；The correlation between client characteristics is calculated further according to Pearson correlation coefficients, rejects those There are the client characteristics of strong correlation with other client characteristics；It is eliminated again according to extreme tree Method for Feature Selection and recursive feature later Method further screens the importance of obtained each client characteristics, determines the client characteristics of final choice.Final choice The numbers of client characteristics be denoted as M, M is positive integer.Such as: as the time span of bank client, gender, age, assets Total value, deposit product quantity, this month financing expire stroke count, this month transaction amount, nearest 3 months purchase Determined products numbers, most Nearly 3 months monthly average transaction amount and nearest 3 months transaction stroke counts.

Since the Model of customer churn prediction that the present invention uses belongs to classified calculating model, and under normal conditions, it is lost Client is fewer than non-attrition customer very much, i.e., sample data is unbalanced, and will lead to customer revenue cannot effectively be classified calculating mould Type identification, prediction result be not accurate, it is therefore desirable to carry out unbalanced data processing.

In some currently preferred embodiments of the present invention, step 11 can include: training sample set to be processed is carried out uneven It weighs data processing, the training sample set that obtains that treated, treated training sample concentrates customer revenue and non-attrition customer The difference of quantity is less than preset threshold.

Differentiate that characteristic standard (uses Kappa coefficient instead as differentiation feature specifically, over-sampling can be taken and changed Method) carry out unbalanced data processing.The method for using over-sampling first, for example, synthesizing a small number of oversampling techniques (Synthetic Minority Oversampling Technique, abbreviation SMOTE) handles initial data, makes to flow The difference for losing the quantity of client and non-attrition customer is less than preset threshold, that is, keeps customer revenue and non-attrition customer ratio close 1:1.Further, the consistent degree between predicted value and actual value can also be differentiated using Kappa coefficient come assessment prediction model, The nicety of grading for measuring model, verifies the effect of unbalanced data processing, that is, carries out unbalanced data processing, it is pre- to improve model Consistent degree between measured value and actual value.

In some currently preferred embodiments of the present invention, client characteristics included in prediction model in order to further increase Interpretation is convenient for subsequent processing, after step 17, further includes:

Step 18: obtaining the client characteristics of Model of customer churn prediction output；

Step 19: client characteristics being inputted in decision-tree model, show the decision tree diagram based on client characteristics.

In this step, decision-tree model is a kind of more original Evolving State of random forests algorithm model, by client spy In sign input decision-tree model, the client characteristics for best embodying customer churn are selected according to statistic (such as: gini index), Such as: client is greater than 50 years old, and it is bigger to be lost possibility；The nearly three months turnovers of client are greater than 10,000 yuan, and loss possibility is bigger, Decision tree diagram is drawn out again, can more intuitively explain how the client characteristics for best embodying customer churn selected lead to client It is lost.

In some currently preferred embodiments of the present invention, after step 17, further includes:

Step 31: the material information of the client characteristics of selection is calculated by Model of customer churn prediction；

Step 32: according to the material information of client characteristics, adjusting the client characteristics that feature selection module uses.

It in an application scenarios, is predicted using random forests algorithm model as Model of customer churn prediction, training During, the material information of M selected client characteristics is calculated, as shown in Figure 6.Geordie impurity level (Gini can be passed through Impurity) calculate the importance of each client characteristics of selection, then to the Geordie impurity level of each client characteristics from big to small into Row sequence filters out the X impure maximum client characteristics of angle value；Alternatively, being counted by information gain (Information Gain) The importance of each client characteristics of selection is calculated, then the information gain of each client characteristics is ranked up from big to small, is screened The maximum field of X information gain value out.X is the positive integer no more than M.It further, can be according to calculated each client characteristics Importance, selected client characteristics are carried out to delete choosing, so that prediction result is more accurate.

In some currently preferred embodiments of the present invention, in step 20, the loss calculated using Model of customer churn prediction is objective The data at family may include: the quantity of the probability of customer churn, the list of customer revenue and customer revenue.It can be sought according to site The strength of pin further determines that the data of customer revenue, so that client is drawn according to the actual conditions that itself markets in site It stays, improves the effect kept.

Such as: following 3 months loss probability that the model that customer churn is defined as using AUC score higher than A1 calculates Client higher than P1.Threshold value A 1 and P1 can be set based on experience and professional standard.If setting A1=0.8, P1=0.5, If excessive using the customer revenue quantity that the two threshold values are predicted, the strength beyond site marketing, i.e., to the net For point, subsequent workload of keeping is too big, can only select to be lost probability to be preceding N of client, N is positive integer, and N is based on site The strength of marketing determines；The threshold value of setting can also be adjusted, be predicted again, such as P1 is adjusted to 0.7, so that The customer revenue quantity of prediction is reduced.

In some currently preferred embodiments of the present invention, after step 20 further include:

Step 30: obtaining the data for retrieving client；

Step 40: according to the data for retrieving client, adjusting the parameter of Model of customer churn prediction.

That is, the customer revenue data using prediction safeguard client, it is actual keep during, get The higher data for retrieving client of accuracy rate, such as: having some clients although being predicted to be loss, reality is not lost but, by these Client, which is modified to, not to be lost.Revised customer data can be input in Model of customer churn prediction, adjustment customer churn is pre- Survey the parameter of model.Such as: Model of customer churn prediction is random forests algorithm model, and revised customer data is input to In random forests algorithm model, at least one of following parameter: classifier number (n_estimators), maximum variable is adjusted Number (max_features), tree node smallest sample number (min_samples_leaf).It is of course also possible to repeatedly correct client's number According to carry out successive ignition, to adjust the parameter of Model of customer churn prediction, the acquisition better customer churn of prediction effect is pre- Survey model.It after the parameter for adjusting Model of customer churn prediction, then is predicted, the accuracy rate of prediction result is higher.

After the data for obtaining the customer revenue of prediction, two similar sites are chosen, such as: two teller's numbers and self-service The similar bank outlets of number of devices.One of them is as intervention group, according to the data of the customer revenue of prediction, to client into Row is kept, and specifically, according to customer churn probability, the different clients for being lost probability is divided into five grades, for example, one grade is lost generally Rate is 80%-100%, and two grades are 60%-80%, and third gear 40%-60%, fourth gear 20%-40%, five grades are 0%- 20%, limited site marketing strength is focused on into one grade of client of maintenance, due emotional care, Presents Giving and preferential activity can be passed through Equal various ways.Another is control group, is kept using conventional method.After 3 months, the customer churn of two sites is compared Rate, it is found that using after method of the invention, compared with the same period of last year the churn rate of intervention group declines.And control group Churn rate remain basically stable with the same period last year or higher.This is the result shows that reduce client using method of the invention Loss, improve the effect kept, be better than conventional method.

Step 300: obtaining the data for retrieving client；

Step 400: according to the data for retrieving client, adjusting the threshold value that the customer churn defines.

That is, the customer revenue data using prediction safeguard client, it is actual keep during, get The higher data for retrieving client of accuracy rate.Revised customer data can be input in Model of customer churn prediction, and adjusted The threshold value that whole customer churn defines.Such as: customer churn is defined as being lost within following 3 months the client that probability is higher than P2, threshold value P2 It can be set based on experience and professional standard.Certain site is estimated before keeping client itself at most to keep 10,000 clients, if P2=0.4 is set, the customer revenue quantity for using Model of customer churn prediction to predict is 9500 people, but the site is being kept It finds in the process, some clients are practical not to be lost but although being predicted to be loss, these clients are modified to and are not lost, correct Customer revenue quantity afterwards is 9000 people, does not give full play to the marketing strength of site, can be by revised customer data again It is input in Model of customer churn prediction, and the threshold value P2 that customer churn defines is adjusted to 0.35, predicted again, predicted Obtained customer revenue quantity be 10500 people so that prediction customer revenue list it is more accurate, more give full play to the battalion of site Strength is sold, the loss of client is avoided.

In another situation, which finds that the strength of site marketing is thinner than what is expected before during actually keeping It is weak, 6000 clients can only may be at most kept, therefore, revised customer data is re-entered to customer churn prediction mould In type, and the threshold value P2 that customer churn defines is adjusted to 0.6, is predicted again, so that the customer revenue list of prediction is more To be accurate, and the customer revenue quantity predicted is reduced.

Of course, it is possible to customer data is repeatedly corrected, and the threshold value that repeatedly adjustment customer churn defines according to actual needs, from And successive ignition is carried out, the customer revenue data of prediction that is more accurate and being more suitable for site are obtained, to give full play to the battalion of site Strength is sold, the loss of client is further reduced.

Referring to Fig. 7, Fig. 7 is the structural schematic diagram of the customer churn prediction device of the embodiment of the present invention two, the device 70 Include:

Module 71 is obtained, for obtaining customer data to be predicted；

Computing module 72 is lost for the customer data to be predicted to be input to calculate in Model of customer churn prediction The data of client, wherein the Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

Preferably, referring to Fig. 8, the customer churn prediction device 70 further include:

Training module 73, for obtaining training sample set, the training sample set is the historical customer data of multiple clients Set；Obtain at least two algorithm models to be selected；For each algorithm model to be selected, according to setting to The parameter of the algorithm model of choosing and the training sample set of input are trained the algorithm model to be selected, obtain being lost visitor The data at family；The true loss that the data of the customer revenue compared and the threshold value defined based on scheduled customer churn are determined The data of client, obtain comparison result；When the comparison result is unsatisfactory for preset condition, the algorithm model to be selected is adjusted Parameter, new root of laying equal stress on is according to the parameter of the algorithm model to be selected of adjustment and the training sample set of input to described to be selected Algorithm model is trained, until the comparison result meets the preset condition, obtains the algorithm model of training completion；

Evaluation module 74, the algorithm model for completing to training are assessed, and assessment result is obtained；

Selecting module 75 selects an instruction for the assessment result for the algorithm model that at least two training is completed Practice the algorithm model completed as the Model of customer churn prediction.

Preferably, the evaluation module 74, for according to evaluation index is preset, the algorithm model completed to training to be commented Estimate, the default evaluation index includes training sample set predictablity rate, test sample collection predictablity rate, area under the curve At least one of AUC score, F score and Kappa coefficient, the test sample collection are the historical customer data of multiple clients Set, the customer data that the test sample is concentrated are different from the customer data that the training sample is concentrated.

The evaluation module 74 includes: data submodule, computational submodule and Comparative sub-module,

The training module 73, for when the algorithm model to be selected includes at least two algorithm model, wherein institute It states at least two algorithm models, algorithm model of the algorithm model as the second layer, remaining algorithm model is as first layer Algorithm model；The algorithm model that the training sample set is input to the first layer is trained, pilot process number is obtained According to；The algorithm model that the pilot process data are input to the second layer is trained, the data of customer revenue are obtained.

Preferably, the customer churn prediction device 70 further include:

Preferably, the training module 73 includes the first processing submodule, for training sample set progress to be processed Data prediction, the training sample set after obtaining data prediction.

Preferably, the training module 73 includes screening submodule, to be processed for being screened using feature selection module The client characteristics that training sample is concentrated, determine the client characteristics of selection, screen training sample set using the client characteristics of selection.

Preferably, the customer churn prediction device 70 further include:

Preferably, the training module 73 includes second processing submodule, for training sample set progress to be processed Unbalanced data processing, the training sample set that obtains that treated, treated, and training sample concentrates customer revenue and non-losss objective The difference of the quantity at family is less than preset threshold.

Preferably, the customer churn prediction device 70 further include:

It should be noted that the customer churn prediction device of the present embodiment belong to above-mentioned customer churn prediction method it is identical Inventive concept, each module in the device can execute step corresponding in above-mentioned customer churn prediction embodiment of the method respectively, Therefore details are not described herein, please refers to the explanation of the above corresponding customer churn prediction method and step in detail.

The embodiment of the invention also provides a kind of computer readable storage mediums, are stored thereon with computer program, special Sign is that the program realizes the step in above-mentioned customer churn prediction method when being executed by processor.

In conclusion the present invention can accurately predict customer churn, the client for being possible to be lost is found, thus and When client is kept, save the cost of customer retention, and improve the effect kept, can effectively solve the problem that the loss of client is asked Topic.

Unless otherwise defined, the technical term or scientific term used herein should be in fields of the present invention and has The ordinary meaning that the personage of general technical ability is understood.Used in present patent application specification and claims " the One ", " second " and similar word are not offered as any sequence, quantity or importance, and are used only to distinguish different Component part." connection " either the similar word such as " connected " is not limited to physics or mechanical connection, but can be with It is either direct or indirect including electrical connection."upper", "lower", "left", "right" etc. are only used for indicating relative position Relationship, after the absolute position for being described object changes, then the relative positional relationship also correspondingly changes.

The above is a preferred embodiment of the present invention, it is noted that for those skilled in the art For, without departing from the principles of the present invention, it can also make several improvements and retouch, these improvements and modifications It should be regarded as protection scope of the present invention.

Claims

1. a kind of customer churn prediction method characterized by comprising

Obtain customer data to be predicted；

The customer data to be predicted is input to the data that customer revenue is calculated in Model of customer churn prediction, wherein institute Stating Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

2. customer churn prediction method according to claim 1, which is characterized in that described by client's number to be predicted Before the step of being input to the data for calculating customer revenue in Model of customer churn prediction, further includes:

Obtain at least two algorithm models to be selected；

For each algorithm model to be selected, according to the training of the parameter of the algorithm model to be selected of setting and input Sample set is trained the algorithm model to be selected, obtains the data of customer revenue；

The true customer revenue that the data of the customer revenue compared and the threshold value defined based on scheduled customer churn are determined Data, obtain comparison result；

When the comparison result is unsatisfactory for preset condition, the parameter of the algorithm model to be selected is adjusted, new root of laying equal stress on is according to tune The parameter of the whole algorithm model to be selected and the training sample set of input are trained the algorithm model to be selected, directly Meet the preset condition to the comparison result, obtains the algorithm model of training completion；

The assessment result for comparing the algorithm model that at least two training are completed, the algorithm model for selecting a training to complete is as described in Model of customer churn prediction.

3. customer churn prediction method according to claim 2, which is characterized in that the algorithm model to be selected includes patrolling Collect regression algorithm model, Bagging algorithm model, random forests algorithm model, AdaBoost algorithm model, Voting Model, heap At least two in stack model and neural network algorithm model.

4. customer churn prediction method according to claim 2, which is characterized in that the algorithm model that described pair of training is completed The step of being assessed, obtaining assessment result include:

According to default evaluation index, the algorithm model completed to training is assessed, and the default evaluation index includes training sample In this collection predictablity rate, test sample collection predictablity rate, area under the curve AUC score, F score and Kappa coefficient extremely One of few, the test sample collection is the set of the historical customer data of multiple clients, client's number that the test sample is concentrated It is different according to the customer data concentrated from the training sample.

5. customer churn prediction method according to claim 2, which is characterized in that the step of the acquisition training sample set Include:

The client characteristics that training sample to be processed is concentrated are screened using feature selection module, the client characteristics of selection is determined, adopts Training sample set is screened with the client characteristics of selection.

6. a kind of customer churn prediction device characterized by comprising

Module is obtained, for obtaining customer data to be predicted；

Computing module calculates customer revenue for the customer data to be predicted to be input in Model of customer churn prediction Data, wherein the Model of customer churn prediction is obtained using the historical customer data training of multiple clients.

7. customer churn prediction device according to claim 6, which is characterized in that further include:

Training module, for obtaining training sample set, the training sample set is the set of the historical customer data of multiple clients； Obtain at least two algorithm models to be selected；For each algorithm model to be selected, according to the calculation to be selected of setting The parameter of method model and the training sample set of input are trained the algorithm model to be selected, obtain the number of customer revenue According to；The true customer revenue that the data of the customer revenue compared and the threshold value defined based on scheduled customer churn are determined Data obtain comparison result；When the comparison result is unsatisfactory for preset condition, the ginseng of the algorithm model to be selected is adjusted Number, new root of laying equal stress on is according to the parameter of the algorithm model to be selected of adjustment and the training sample set of input to the algorithm to be selected Model is trained, until the comparison result meets the preset condition, obtains the algorithm model of training completion；

Selecting module, for comparing the assessment result for the algorithm model that at least two training are completed, the calculation for selecting a training to complete Method model is as the Model of customer churn prediction.

8. customer churn prediction device according to claim 7, which is characterized in that the algorithm model to be selected includes patrolling Collect regression algorithm model, Bagging algorithm model, random forests algorithm model, AdaBoost algorithm model, Voting Model, heap At least two in stack model and neural network algorithm model.

9. customer churn prediction device according to claim 7, which is characterized in that

The evaluation module is used for according to evaluation index is preset, and the algorithm model completed to training is assessed, and described preset is commented Estimate index include training sample set predictablity rate, test sample collection predictablity rate, area under the curve AUC score, F score and At least one of Kappa coefficient, the test sample collection are the set of the historical customer data of multiple clients, the test specimens The customer data of this concentration is different from the customer data that the training sample is concentrated.

10. customer churn prediction device according to claim 7, which is characterized in that the training module includes:

Submodule is screened, for screening the client characteristics that training sample to be processed is concentrated using feature selection module, determines choosing The client characteristics selected screen training sample set using the client characteristics of selection.