CN105761093A

CN105761093A - Knowledge-space-based behavior result evaluation method and device

Info

Publication number: CN105761093A
Application number: CN201410789962.XA
Authority: CN
Inventors: 杨强; 袁明轩; 曾嘉; 戴文渊
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2014-12-17
Filing date: 2014-12-17
Publication date: 2016-07-13

Abstract

The invention provides a knowledge-space-based behavior result evaluation method. The method is applied based on the analysis result of a user offline prediction model provided by an operator, an airline or any other service providers, and the behavior result of a user that might be off the line is compared and evaluated finally. Through training an offline prediction model, the data of the user is subjected to offline prediction and probability prediction. Meanwhile, through evaluating the influence of marketing or promotion means on the offline prediction probability change, a most efficient marketing means is selected. Therefore, the retention is conducted for users that are mostly likely to get off the line. Meanwhile, the maximum revenue is realized on the premise that the cost is limited to a certain degree.

Description

The behavior outcome appraisal procedure in a kind of knowledge based space and device

Technical field

The present invention relates to data analysis field, particularly relate to behavior outcome appraisal procedure and the device in a kind of knowledge based space.

Background technology

For any service provider, such as telecom operators (China Mobile, UNICOM etc.), airline and other service provider (travel service, E-business service etc.), the business model of core all includes two parts: storage operation and increment operation.Storage is managed and is referred to maintaining and servicing of active user, and one of them most crucial task is how farthest to make user especially high-value user not run off.Increment operation refers to creates new business model to obtain more profit.Such as telecom operators, outside completing offer call and data basic service, provide Users'Data Analysis to support (such as providing the information such as certain zone user classification, flow for the third company of shop addressing class) for third-party service company and obtain profit.

Storage manage in user to maintain be the primary core business to solve of service provider, traditional mode is experienced customer manager has high loss (off-network) risk by which user of empirical estimating, and provides the user the service of maintaining.Customer manager rule of thumb selects to maintain (marketing) action for certain user.Such as customer manager is chosen as part 20 to 25 years old, and the networking duration user more than 3 years provides the favorable sale adding 1 yuan of increase 100M surfing flow following three months every months to maintain.The main weakness of this mode is that efficiency is low and poor accuracy.Along with the development of mobile internet, user data presents explosive growth in recent years, analyzes user by rule of thumb inoperable.Increasing service provider begins with the technology of big data analysis and is modeled user analyzing and maintaining.Basic method is to set up the big Data Analysis Model of user's off-network in a future/loss, and the user of a period of time (such as 1 month) prediction in advance will be likely to off-grid probability future.Model output is according to off-network prediction probability user list from high to low.Customer manager empirically carries out telemarketing etc. according to the information of list output and relative users and maintains action.Such as, the existing user's off-network of telecommunication carrier analyzes system, sets up the big data model of user's off-network early signal, analyzes the off-network prediction probability after user January.Then high off-network user is maintained according to the result that model exports and keeps by customer manager targetedly.

Current system is applied only for the analysis result of model, customer manager selects some marketing behavior means to carry out user further according to experience to maintain, this action selection maintaining method is only that the experience with customer manager carries out, there is no certain data support, it is difficult to the result of Accurate Prediction behavior act, and carrying out effective action assessment, it is achieved minimum cost realizes maximum client's dimension and draws effect.

Summary of the invention

In view of this, the embodiment of the present invention provides the behavior outcome appraisal procedure in a kind of knowledge based space, for the result to Accurate Prediction behavior act, and carries out effective action assessment.

Embodiment of the present invention first aspect provides the behavior outcome appraisal procedure in a kind of knowledge based space, including

nullObtain the second user by user's off-network forecast model concentrate on user that in the 3rd time period, off-network prediction probability is the highest collection and obtain the 3rd user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest，Wherein said 3rd user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period，Described second user's collection includes not implemented at least one user of marketing or sales promotion action within the second time period，Described user's off-network forecast model is for obtaining user's off-network prediction probability based on user's historical data，Described user's historical data includes user basic information、Customer consumption information、User's off-network information and user's historical action information，Described historical action information indicates marketing or sales promotion action that described user was implemented，Described 3rd time period is the subsequent time period of described second time；

Described second user's collection off-network prediction probability is the highest within the 3rd time period user collection and the described 3rd user's collection user that off-network prediction probability is the highest within the 3rd time period collection are carried out matching ratio compared with the off-network prediction probability difference to obtain between two users, the off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user's collection and off-network prediction probability difference between two users that the 3rd user collects；

According to comparative result, the off-network prediction probability difference setting up described two users and the described two users carrying out mathematic interpolation belong to described 3rd user collection user user's historical action information between mapping relations；

According to described mapping relations, it is thus achieved that at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum.

In conjunction with first aspect, in the first possible implementation of first aspect,

Aforementioned in steps before include:

The first user collection in described user's historical data is utilized to train the described user's off-network forecast model of generation in the historical data of first time period, described first user collection includes not implemented at least one user of marketing or sales promotion action in first time period, and described first time period is any time period in the past；

The historical data of second user's collection described in the second time period is input to described user's off-network forecast model, it is thus achieved that described second user concentrates on user's collection that in the 3rd time period, off-network prediction probability is the highest；

The historical data of the 3rd user's collection described in the second time period is input to described user's off-network forecast model, it is thus achieved that described 3rd user concentrates on user's collection that in the 3rd time period, off-network prediction probability is the highest.

The first in conjunction with first aspect or first aspect is likely to, in the second possible implementation of first aspect, described described second user's collection off-network prediction probability is the highest within the 3rd time period user collection carried out matching ratio with the described second user's collection user that off-network prediction probability is the highest within the 3rd time period collection relatively include: by identical for out of Memory except user basic information and historical action information in described user's historical data and be belonging respectively to described second user's collection and the 3rd user collection two users' coupling together, and the off-network prediction probability of the described two users of comparison is to obtain the off-network prediction probability difference between described two users.

The first or the second in conjunction with first aspect or first aspect are likely to, in the third possible implementation of first aspect, described according to comparative result, set up the historical action information that described 3rd user concentrates to include with the mapping relations of user's off-network prediction probability difference of corresponding user: set up described second user's collection and the 3rd user concentrates the mapping relations between the historical action information belonging to the user that the 3rd user collects in the off-network prediction probability difference belonged between described two users that user that in the 3rd time period, off-network prediction probability is the highest collects and out of Memory is identical except user basic information and historical action information in described user's historical data and described two users carrying out mathematic interpolation.

The first or the second in conjunction with first aspect or first aspect are likely to, in the third possible implementation of first aspect, described utilize first user collection in described user's historical data to train in the historical data of first time period to generate user's off-network forecast model and include utilizing first user collection in described user's historical data to train neural network model or Logic Regression Models to generate described user's off-network forecast model in the historical data of first time period.

Any one in being likely to the third in conjunction with the first of first aspect or first aspect is likely to, in the 4th kind of possible implementation of first aspect, described according to described mapping relations, at least one the historical action information obtaining the off-network prediction probability difference of corresponding described user maximum includes according to described mapping relations, obtain at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum and conforms to a predetermined condition, described predetermined condition is the marketing that described user was implemented of described historical action information instruction or the cost-range of sales promotion action and income range.

Embodiment of the present invention second aspect provides the behavior outcome appraisal procedure in a kind of knowledge based space, including

Obtaining user's historical data, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, and described historical action information indicates marketing or sales promotion action that described user was implemented；

The first user collection in described user's historical data is utilized to train generation first user off-network forecast model in the historical data of first time period, described first user collection includes not implemented at least one user of marketing or sales promotion action in first time period, described first time period is any time period in the past, and described first user off-network forecast model is for obtaining user's off-network probability according to the user's historical data outside removing historical action information；

Utilize the second user's collection historical data in the second time period in described user's historical data to train and generate second user's off-network forecast model, described second user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period, described second time period is any time period in the past, and described second user's off-network forecast model is for obtaining user's off-network probability according to except user's historical data；

The historical data of the 3rd user's collection of the 3rd time period is input to described first user off-network forecast model, it is thus achieved that in the 3rd time period, the highest and without marketing or sales promotion action the user of off-network prediction probability collects；

By the 3rd time period the 3rd user collection historical data be input to described second user's off-network forecast model, it is thus achieved that in the 3rd time period off-network prediction probability the highest and have marketing or sales promotion action user collection；

By described 3rd user collection within the 3rd time period off-network prediction probability the highest and without marketing or sales promotion action user collection and described 3rd user collection within the 3rd time period off-network prediction probability the highest and have marketing or sales promotion action user collection carry out matching ratio compared with the off-network prediction probability difference to obtain between two users, the off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user collection and the 3rd user collect two users between off-network prediction probability difference；

The embodiment of the present invention third aspect provides the behavior outcome assessment system in a kind of knowledge based space, including:

nullOff-network prediction module，Concentrate on user that in the 3rd time period, off-network prediction probability is the highest collection for obtaining the second user by user's off-network forecast model and obtain the 3rd user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest，Wherein said 3rd user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period，Described second user's collection includes not implemented at least one user of marketing or sales promotion action within the second time period，Described user's off-network forecast model is for obtaining user's off-network prediction probability based on user's historical data，Described user's historical data includes user basic information、Customer consumption information、User's off-network information and user's historical action information，Described historical action information indicates marketing or sales promotion action that described user was implemented，Described 3rd time period is the subsequent time period of described second time；

Coupling mapping block, matching ratio is carried out compared with the off-network prediction probability difference to obtain between two users for the described second user's collection user that off-network prediction probability is the highest within the 3rd time period collection and the described 3rd user's collection user that off-network prediction probability is the highest within the 3rd time period collection, and according to comparative result, the off-network prediction probability difference setting up described two users and the described two users carrying out mathematic interpolation belong to the mapping relations between user's historical action information of the user that described 3rd user collects, off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user collection and the 3rd user collection two users between off-network prediction probability difference；

Select module, for according to described mapping relations, it is thus achieved that at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum.

Embodiment of the present invention fourth aspect provides the behavior outcome assessment system in a kind of knowledge based space, including:

Data acquisition module, for obtaining user's historical data, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, and described historical action information indicates marketing or sales promotion action that described user was implemented；

nullTraining module，For utilizing the first user collection in described user's historical data to train generation first user off-network forecast model in the historical data of first time period，And utilize the second user's collection historical data in the second time period in described user's historical data to train and generate second user's off-network forecast model，Wherein said first user collection includes not implemented at least one user of marketing or sales promotion action in first time period，Described second user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period，Described first time period and described second time period are any time period in the past，Described first user off-network forecast model is for obtaining user's off-network probability according to the user's historical data outside removing historical action information，，Described second user's off-network forecast model is for obtaining user's off-network probability according to except user's historical data；

Off-network prediction module, obtain that off-network prediction probability in the 3rd time period is the highest and user's collection without marketing or sales promotion action for the historical data of the 3rd user's collection of the 3rd time period being input to described first user off-network forecast model, and the historical data of the 3rd user's collection of the 3rd time period is input to described second user's off-network forecast model, and to obtain off-network prediction probability in the 3rd time period the highest and have user's collection of marketing or sales promotion action；

Coupling mapping block, for by described 3rd user collection within the 3rd time period off-network prediction probability the highest and without marketing or sales promotion action user collection and described 3rd user collection within the 3rd time period off-network prediction probability the highest and have marketing or sales promotion action user collection carry out matching ratio compared with the off-network prediction probability difference to obtain between two users, and belong in setting up the off-network prediction probability difference of described two users and carrying out described two users of mathematic interpolation described 3rd user's collection user user's historical action information between mapping relations, off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user collection and the 3rd user collection two users between off-network prediction probability difference；

The embodiment of the present invention the 5th aspect provides the behavior outcome assessment computing equipment in a kind of knowledge based space, including network interface, memorizer and processor, wherein said network interface is used for obtaining user's historical data, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, and described historical action information indicates marketing or sales promotion action that described user was implemented；

Described memorizer is used for storing described user's historical data；

Described processor is used for performing following steps: obtains the second user by user's off-network forecast model and concentrates on user that in the 3rd time period, off-network prediction probability is the highest collection and obtain the 3rd user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest, wherein said 3rd user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period, described second user's collection includes not implemented at least one user of marketing or sales promotion action within the second time period, described user's off-network forecast model obtains user's off-network prediction probability for the user's historical data based on memorizer, described 3rd time period is the subsequent time period of described second time；

The advantage of the embodiment of the present invention will partly illustrate in the following description, and a part is apparent from according to description, or can be known by the enforcement of the embodiment of the present invention.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the behavior outcome appraisal procedure in a kind of knowledge based space in the embodiment of the present invention one.

Fig. 2 is the operation principle schematic diagram of off-network forecast model in the embodiment of the present invention.

Fig. 3 is the schematic flow sheet of the behavior outcome appraisal procedure in a kind of knowledge based space in the embodiment of the present invention two.

Fig. 4 is the structural representation of the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention three.

Fig. 5 is the schematic diagram of the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention four.

Fig. 6 is the schematic diagram of the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention five.

Fig. 7 is the schematic diagram of the behavior outcome assessment computing equipment in a kind of knowledge based space in the embodiment of the present invention six.

Detailed description of the invention

The following stated is the preferred implementation of the embodiment of the present invention; it should be pointed out that, for those skilled in the art, under the premise without departing from embodiment of the present invention principle; can also making some improvements and modifications, these improvements and modifications are also considered as the protection domain of the embodiment of the present invention.

The behavior outcome appraisal procedure in a kind of knowledge based space of the present invention is stated the churn prediction forecast model of the service provider such as operator or airline before application and is analyzed on the basis of result, the last off-grid user of possibility is carried out behavior outcome assessment predict, draw and produce the maximum return under certain cost restriction to select the most efficient marketing methods that these most possible off-grid users are tieed up.nullThe behavior outcome appraisal procedure in a kind of knowledge based space of the present invention uses and maintains (marketing) action executing recruitment evaluation knowledge，This maintains (marketing) action executing recruitment evaluation knowledge and refers to if user carrying out certain marketing maintain action，Such as a given user historian database D，A is gathered in operable action (marketing actions)，Off-network forecast model M，Calculate marketing actions A and be applied to the Profit Assessment of each user，And recommend the user of optimum to maintain action for customer manager on this basis，Wherein historian database includes user's history consumption information，Information on services，Off-network information，User basic information，History marketing message etc.，Wherein without user off-network forecast model M，Generation M can be trained in database D，By the big data analysis retrievable user intended change of future features of historical data.After the knowledge of model own refers to user characteristics change, from the prospective earnings change that "current" model knowledge analysis obtains.The embodiment of the present invention can obtain in conjunction with action executing recruitment evaluation knowledge and model knowledge and maintain (marketing) earnings forecast more accurately.In practical application scene, under certain limited fund, the embodiment of the present invention can be made the user of optimization according to the earnings forecast of marketing actions and maintain (marketing) action and recommend with maximum gain.

Embodiment one

As it is shown in figure 1, the behavior outcome appraisal procedure in a kind of knowledge based space in the embodiment of the present invention one, the method includes step:

nullStep 101、Obtain the second user by user's off-network forecast model concentrate on user that in the 3rd time period, off-network prediction probability is the highest collection and obtain the 3rd user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest，Wherein said 3rd user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period，Described second user's collection includes not implemented at least one user of marketing or sales promotion action within the second time period，Described user's off-network forecast model is for obtaining user's off-network prediction probability based on user's historical data，Described user's historical data includes user basic information、Customer consumption information、User's off-network information and user's historical action information，Described historical action information indicates marketing or sales promotion action that described user was implemented，Described 3rd time period is the subsequent time period of described second time；

Step 102, described second user's collection off-network prediction probability is the highest within the 3rd time period user collection and the described 3rd user's collection user that off-network prediction probability is the highest within the 3rd time period collection are carried out matching ratio compared with the off-network prediction probability difference to obtain between two users, the off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user's collection and off-network prediction probability difference between two users that the 3rd user collects；

Step 103, according to comparative result, the off-network prediction probability difference setting up described two users and the described two users carrying out mathematic interpolation belong to described 3rd user collection user user's historical action information between mapping relations；

Step 104, according to described mapping relations, it is thus achieved that at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum.

Described in steps 101 also before include:

In step above-mentioned steps, from the data base of service provider obtain user's historical data, wherein service provider include as telecom operators (China Mobile, UNICOM etc.), airline, other take the service provider such as travel service and ecommerce.

In step above-mentioned steps, off-network forecast model is trained by input user historical data base D to user's off-network predictor (ChurnPredictor).User's off-network predictor mainly uses user's history consumption information, information on services, user basic information, and history marketing message etc. builds user characteristics, uses the user off-network information labeling off-network user in historical data base D simultaneously.Then training generates user off-network forecast model M, it was predicted that the following off-grid probability of each user.The general principle of training is, the user historical data constantly corresponding with the same period by the data that predict the outcome of different time sections compares, the algorithm of described user's off-network forecast model is constantly adjusted according to comparative result, until prediction data is minimum with historical data deviation, determine optimal models thus obtaining last user's off-network forecast model.

Referring specifically to Fig. 2, generating the decision tree of off-network forecast model M, the off-network forecast model M of output can be characterized as being the set of one group of rule:

Feature (feature) 1 numerical range (valuerange) x1 ∧ feature (feature) 2 numerical range (valuerange) x2 ∧ ...-> off-network probability (Churnprobability) p1

Model in such as Fig. 2, it is possible to represent with following regular collection:

The duration of call: low ∧ call meters: low-> 0.1

The duration of call: low ∧ call meters: in-> 0.6

The duration of call: low ∧ call meters: high-> 0.9

The duration of call: middle ∧ call meters: in-> 0.3

The duration of call: high ∧ call meters: low-> 0.1

The duration of call: high ∧ call meters: in-> 0.2

The duration of call: high ∧ call meters: high-> 0.6

The input of off-network forecast model is user characteristics, it is output as the following off-network prediction probability according to user characteristics prediction, off-network forecast model knowledge can regard the combination of some rules as, and these acting rules, on user characteristics, represent which feature value can cause certain off-network prediction probability.In such as Fig. 2, the digitized representation off-network prediction probability on leaf node side, non-leaf nodes is the feature of user, one rule is exactly a path from root node to leaf node, and such as a rule therein is exactly: user's off-network prediction probability of the duration of call ' height ' and call meters ' low ' is 0.1 (10%).

In 101-102 step, user's historical data of this service provider can be divided into the historical data of different time sections, wherein first and second and three the time period be the continuously or discontinuously time period that time span is identical, the 3rd time period can be a nearest time period.In order to provide the up-to-date property of data, the next time period of the 3rd period can be the nearest time period that will occur future.In step 101, the time period of input is different thus obtaining user's off-network prediction data of different time sections, first obtains user's off-network prediction data of the second time period, user's off-network prediction data of the 3rd time period of reentrying in such as step 101.

In step 106 and 107, the change of the user's off-network prediction probability in observation different time sections, mapping association relation again through user's off-network prediction probability situation of change with the marketing in the corresponding time period, this user implemented and sales promotion action, search out and user's off-network prediction rate is reduced maximally effective marketing and sales promotion action, thus helping service provider to select suitable or optimum marketing and sales promotion action or scheme.

In addition, in step 103-104, can also be brought Selection In the restraining factors of marketing actions, such as can consider cost and factor of profit, the marketing such as selected and sales promotion action or scheme are not only only capable of being that user's off-network prediction probability reduction amplitude is maximum, also need to meet predetermined cost or income restriction, as, need to meet cost less than 10 yuan or earning rate more than 20%, wherein the calculating by promoting the increment that customer consumption increase brings to deduct marketing due to marketing or marketing tool or promotional cost comes calculated of earning rate.

Embodiment two

As it is shown on figure 3, the behavior outcome appraisal procedure in a kind of knowledge based space in the embodiment of the present invention two, the method includes step:

Step 101, acquisition user's historical data, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, and described historical action information indicates marketing or sales promotion action that described user was implemented；

Step 102, utilize first user collection in described user's historical data to train in the historical data of first time period to generate first user off-network forecast model, described first user collection includes not implemented at least one user of marketing or sales promotion action in first time period, described first time period is any time period in the past, and described first user off-network forecast model is for obtaining user's off-network probability according to the user's historical data outside removing historical action information；

Step 103, utilize in described user's historical data the second user's collection historical data in the second time period to train to generate second user's off-network forecast model, described second user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period, described second time period is any time period in the past, and described second user's off-network forecast model is for obtaining user's off-network probability according to except user's historical data；

Step 104, the historical data of the 3rd user's collection of the 3rd time period is input to described first user off-network forecast model obtains that off-network prediction probability in the 3rd time period is the highest and user's collection without marketing or sales promotion action；

Step 105, the historical data of the 3rd user's collection of the 3rd time period is input to described second user's off-network forecast model, and to obtain off-network prediction probability in the 3rd time period the highest and have user's collection of marketing or sales promotion action；

Step 106, by described 3rd user collection within the 3rd time period off-network prediction probability the highest and without marketing or sales promotion action user collection and described 3rd user collection within the 3rd time period off-network prediction probability the highest and have marketing or sales promotion action user collection carry out matching ratio compared with the off-network prediction probability difference to obtain between two users, the off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user collection and the 3rd user collect two users between off-network prediction probability difference；

Step 107, according to comparative result, the off-network prediction probability difference setting up described two users and the described two users carrying out mathematic interpolation belong to described 3rd user collection user user's historical action information between mapping relations；

Step 108, according to described mapping relations, it is thus achieved that at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum.

Method And Principle in above-described embodiment two is similar to embodiment one, and the scheme of training pattern is also identical, and the mode that data are processed is also identical, and the model of use is also identical, no longer repeats in detail at this.

Embodiment three

As shown in Figure 4, the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention three, including:

Off-network predictor, for utilizing the training of described user's historical data generate user's off-network forecast model and the historical data of first time period be input to and obtain, with described family off-network forecast model, the user that in the second time period, off-network prediction probability is the highest, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, described historical action information indicates marketing or sales promotion action that described user was implemented, and described second time period is the subsequent time period of the described very first time；

Action effect evaluator, for setting up the mapping relations of the historical action information of the user that off-network prediction probability is the highest and user's off-network prediction probability changing value in described second time period, described user's off-network prediction probability changing value is user that in described second time period, off-network prediction probability the is the highest off-network prediction probability changing value within the second time period to the 3rd time period, wherein said 3rd time period is the subsequent time period of the second time period, described user's off-network prediction probability changing value is that in described 3rd time period, off-network prediction probability obtains by the historical data of the second time period is input to described user's off-network forecast model；

Optimum action selector, for according to described mapping relations, it is thus achieved that make the off-network prediction probability of the user that off-network prediction probability is the highest in described second time period reduce at least one maximum historical action information.

Above-mentioned user's off-network predictor mainly uses user's history consumption information, information on services, user basic information, and history marketing message etc. builds user characteristics, uses the user off-network information labeling off-network user in historical data base D simultaneously.Then training generates user off-network forecast model M, the following off-grid probability of each user of model prediction, the decision tree in the model of generation such as Fig. 2, the set that can be characterized as being one group of rule of user off-network forecast model M output:

feature1valuerangex1∧feature2valuerangex2∧…->Churnprobabilityp1

…

Model output in such as Fig. 2, it is possible to represent with following regular collection,

The duration of call: low ∧ call meters: low-> 0.1

The duration of call: low ∧ call meters: in-> 0.6

The duration of call: low ∧ call meters: high-> 0.9

The duration of call: middle ∧ call meters: *-> 0.3

The duration of call: high ∧ call meters: low-> 0.1

The duration of call: high ∧ call meters: in-> 0.2

The duration of call: high ∧ call meters: high-> 0.6

Above-mentioned action Profit Assessment device comprises two submodules: action effect evaluator and action income calculation device.

Described action effect evaluator (ActionResultEstimator): training generation is maintained forecast model that user characteristics value changes by (marketing) action and exports the table T that predicts the outcome that user characteristics is changed by adoptable action.Having two actions in Fig. 2, the table of Fig. 2 lower left is exactly the example exporting result, and output format is,

Action id, user id, feature value list, change to the probability of preceding features value

…

Described action income calculation device (ActionGainComputer): the table T that predicts the outcome, user's off-network forecast model and user user characteristics changed according to action be worth (if user's non-off-network, every month is the value that service provider brings), calculate action income statement, the form of action income statement T ' is

Action id, user id, income

…

Action income statement T ' and budget limit budget recommends action selector (OptimalActionSetCalculator) to optimum.Export the optimum marketing actions recommended for user.

Such as shown in Fig. 2, under the restriction of budget $ 10, user 1 is arrived in embodiment of the present invention recommendation action 1, it is desirable to income is $ 39 (analysis is referred to 1.1 analyses in page 2), and output format is,

Recommend Action Selection 1:

User id action id

…

Recommend Action Selection 2:

…

Noting, the embodiment of the present invention allows to export the Action Selection of multiple recommendation and be more than single here.Such customer manager can have multiple choices, such as output expected revenus institute's likely Action Selection more than current 20%.

Specifically, user's off-network predictor main target is to set up a model, and this model can pass through the early sign of user and predict following off-network prediction probability.Such as, the embodiment of the present invention uses user 4, and the historical record in May sets up feature, the off-network prediction probability (be one month earlier because reserving user and maintain the time) predicting user's July by these features.The feature extracted includes a series of descriptions of consumer consumption behavior, service quality etc..Such as, caller charging duration, networking duration, caller number yardage, free duration, called duration, voice call number of times, caller number of times, GPRS total flow and descending point-to-point note number etc..

The embodiment of the present invention extracts user's historical sample training pattern.Being such as August now, the embodiment of the present invention extracts 3, and the data in April do feature, and the data in May are done mark and trained a model (the test tuning of model adopts 4, does feature May, and July marks).Mark is meant to the embodiment of the present invention and indicates which user in off-network in July, which user not off-network.Because July has been historical record, the embodiment of the present invention can mark the off-network user of off-network/not clearly.Use these historical datas, next step one forecast model of training of the embodiment of the present invention.Each user is a sample, and each user is described as a series of feature.For a simplest example, in Fig. 2, user 1 is represented as after feature extraction: [user 1, (duration of call ' in ', call meters ' height ') feature, (July off-network: 1) predictive value].The purpose of model is exactly that the feature using user is to predict off-grid mark.The process of model training is an iterative process, and the feature and the model parameter that repeatedly adjust extraction ultimately produce optimum user's off-network forecast model.After generating optimal models, the embodiment of the present invention extracts the record that user is nearest, and such as current August, the embodiment of the present invention extracts user 6, and the data in July do feature, uses the probability that run off the future that user's off-network model calculates user

The work process that effect knowledge assessment device is maintained in above-mentioned action is similar with user's off-network predictor, the difference is that the selection of sample and feature.

Fig. 4 describes action and maintains the logic flow of effect knowledge assessment device, and each sample is the combination (user, action) of a user and action here.The mark of data is that this action is applied to the value of certain feature after user, such as ' local caller duration '.Because each sample here comprises user and action, sample statement just comprises user characteristics and motion characteristic (and combination of both features) simultaneously.Namely the embodiment of the present invention uses the feature of user and action to characterize each (user, action).For a simplest example, in Fig. 2, assume in history to after user 1 usage operation 1, his ' local caller duration ' becomes ' in ', action A has a feature: cost ' in ', this sample just can be described as that [(user 1, action 1), ((the duration of call ' low ', call meters ' height ') user characteristics, (cost ' in ') motion characteristic, (duration of call and cost: ' in ') user action assemblage characteristic), (' local caller duration ': in) predictive value].The embodiment of the present invention is each attribute that can passively adjust one forecast model of training.The action of history and effect thereof are used for training and generate model, and model is used to the prediction current adoptable action change to the attribute of each user.Because action is represented as motion characteristic, even if action 1 did not occur in history, the feature of action 1 can also by other manual expression used in history, such that it is able to the action variation effect to user property value that the plan that dopes accurately adopts.

The probability Estimation that user characteristics is changed by above-mentioned action income calculation device according to action, calculates user and how to move at the knowledge space of off-network forecast model.Namely user is from current off-network prediction probability, changes to the probability of additionally different off-network prediction probabilities.Finally calculating by reducing user's retrievable income of off-network prediction probability, the cost deducting employing action can obtain action income estimation table.Example in such as Fig. 2, after action is applied to user, the embodiment of the present invention can calculate shown in action-expected revenus table following table:

Table 1: action-expected revenus result of calculation example

Optimum recommendation action selector (OptimalActionSetCalculator)

The embodiment of the present invention calculates the set of actions of recommendation according to budgetary restraints.Such as budgetary restraints is $ 10 (action 1 cost $ 5, action 2 cost $ 10), and embodiment of the present invention optimum recommends [(user 1, action 1)].If customer manager wants all recommendation set of actions obtaining income $ more than 20, the embodiment of the present invention recommends [(user 1, action 1)], and [(user 2, action 2)] select for it.The optimum problem recommending action selector to solve can be represented as equation:

Maximize:

\underset{u}{Σ} g_{u} - c_{u}

or

\underset{u}{Σ} g_{u} - c_{u} &GreaterEqual; lower_bound

Subjectto:

&ForAll; u \underset{i}{Σ} x_{u, i} \leq 1

\underset{u}{Σ} \underset{i}{Σ} c_{u} \times x_{u, i} \leq budget

&ForAll; u g_{u} = x_{u, i} \times Ga {in}_{u}

&ForAll; u c_{u} = x_{u, i} \times C_{u, i}

x_{U, i}∈ { 0,1}

In the equation, optimum recommendation action selector xu, i is one (0,1) integer, 1 is represented as user u recommendation action i, the equation target is to find the set of actions of optimum (maximum) income (equation the 1st row), it can be again the institute's likely set of actions (equation the 2nd row) finding income more than lower_bound, in the equation, in the 4th row, constraint only recommends an action to a user, retraining always expending less than budget Budget of employing action in the equation the 5th row, equation 6-8 row is exactly the calculating of action-expected revenus table.

Visible, the behavior outcome appraisal procedure in a kind of knowledge based space that the embodiment of the present invention provides is analyzed on the basis of result at churn prediction forecast model, the last off-grid user of possibility is carried out behavior outcome assessment predict, and compare the relation between the situation of user's off-network prediction probability change and marketing and sales promotion action that user is implemented, draw and produce the maximum return under certain cost restriction to select the most efficient marketing methods that these most possible off-grid users are tieed up.

Embodiment four

As it is shown in figure 5, the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention four, including:

The behavior outcome assessment system in described knowledge based space also includes:

Training module, for utilizing the first user collection in described user's historical data to train the described user's off-network forecast model of generation in the historical data of first time period, described first user collection includes not implemented at least one user of marketing or sales promotion action in first time period, and described first time period is any time period in the past.

Described prediction module, specifically for the historical data of second user's collection described in the second time period is input to described user's off-network forecast model, obtain described second user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest, and the historical data of the 3rd user's collection described in the second time period is input to described user's off-network forecast model, it is thus achieved that described 3rd user concentrates on user's collection that in the 3rd time period, off-network prediction probability is the highest.Described selection module, being additionally operable to obtain at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum and conforms to a predetermined condition, described predetermined condition is the marketing that described user was implemented of described historical action information instruction or the cost-range of sales promotion action and income range.

Embodiment five

As shown in Figure 6, the behavior outcome assessment system in a kind of knowledge based space in the embodiment of the present invention four, including:

nullTraining module，For utilizing first user collection in described user's historical data to train generation first user off-network forecast model in the historical data of first time period，And utilize the second user's collection historical data in the second time period in described user's historical data to train and generate second user's off-network forecast model，Wherein said first user collection includes not implemented at least one user of marketing or sales promotion action in first time period，Described second user's collection includes being implemented at least one user of marketing or sales promotion action within the second time period，Described first time period and described second time period are any time period in the past，Described first user off-network forecast model is for obtaining user's off-network probability according to the user's historical data outside removing historical action information，，Described second user's off-network forecast model is for obtaining user's off-network probability according to except user's historical data；

Coupling mapping block, for by described 3rd user collection within the 3rd time period off-network prediction probability the highest and without marketing or sales promotion action user collection and described 3rd user collection within the 3rd time period off-network prediction probability the highest and have marketing or sales promotion action user collection carry out matching ratio compared with the off-network prediction probability difference to obtain between two users, and belong in setting up the off-network prediction probability difference of described two users and carrying out described two users of mathematic interpolation described 3rd user's collection user user's historical action information between mapping relations, off-network prediction probability difference between described two users refer to except the user basic information in described user's historical data and except historical action information out of Memory identical and be belonging respectively to second user collection and the 3rd user collection two users between off-network prediction probability difference value；

The behavior outcome assessment system in described knowledge based space utilizes first user collection in described user's historical data to train generation first user off-network forecast model in the historical data of first time period described, and utilize in described user's historical data the second user's collection historical data in the second time period train generate second user's off-network forecast model in, described training module trains neural network model or Logic Regression Models specifically for utilizing first user collection in described user's historical data in the historical data of first time period, to generate first user off-network forecast model, and utilize the second user's collection historical data in the second time period in described user's historical data to train neural network model or Logic Regression Models, to generate second user's off-network forecast model.

Embodiment six

As it is shown in fig. 7, the behavior outcome assessment computing equipment in a kind of knowledge based space in the embodiment of the present invention four, including network interface, memorizer and processor, wherein:

Described network interface is used for obtaining user's historical data, described user's historical data includes user basic information, customer consumption information, user's off-network information and user's historical action information, and described historical action information indicates marketing or sales promotion action that described user was implemented；

Described memorizer is used for storing described user's historical data；

Although preferred embodiments of the present invention have been described, but those skilled in the art are once know basic creative concept, then these embodiments can be made other change and amendment.So, claims are intended to be construed to include preferred embodiment and fall into all changes and the amendment of the scope of the invention.

Obviously, the present invention can be carried out various change and modification without deviating from the spirit and scope of the present invention by those skilled in the art.So, if these amendments of the present invention and modification belong within the scope of the claims in the present invention and equivalent technologies thereof, then the present invention is also intended to comprise these change and modification.

Claims

1. the behavior outcome appraisal procedure in a knowledge based space, it is characterised in that including:

2. the behavior knot outcome evaluation in knowledge based space as claimed in claim 1, it is characterised in that described in steps before include:

3. the behavior outcome appraisal procedure in knowledge based space as claimed in claim 1 or 2, it is characterized in that, described described second user's collection off-network prediction probability is the highest within the 3rd time period user collection carried out matching ratio with the described second user's collection user that off-network prediction probability is the highest within the 3rd time period collection relatively include: by identical for out of Memory except user basic information and historical action information in described user's historical data and be belonging respectively to described second user's collection and the 3rd user collection two users' coupling together, and the off-network prediction probability of the described two users of comparison is to obtain the off-network prediction probability difference between described two users.

4. the behavior outcome Forecasting Methodology in knowledge based space as claimed in claim 1 or 2, it is characterized in that, described according to comparative result, set up the historical action information that described 3rd user concentrates to include with the mapping relations of user's off-network prediction probability difference of corresponding user: set up described second user's collection and the 3rd user concentrates the mapping relations between the historical action information belonging to the user that the 3rd user collects in the off-network prediction probability difference belonged between described two users that user that in the 3rd time period, off-network prediction probability is the highest collects and out of Memory is identical except user basic information and historical action information in described user's historical data and described two users carrying out mathematic interpolation.

5. the behavior outcome appraisal procedure in knowledge based space as claimed in claim 1 or 2, it is characterized in that, described utilize first user collection in described user's historical data to train in the historical data of first time period to generate user's off-network forecast model and include: utilize first user collection in described user's historical data to train neural network model or Logic Regression Models in the historical data of first time period, to generate described user's off-network forecast model.

6. the behavior outcome appraisal procedure in knowledge based space as claimed in claim 1 or 2, it is characterized in that, described according to described mapping relations, at least one the historical action information obtaining the off-network prediction probability difference of corresponding described user maximum includes: according to described mapping relations, obtaining at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum and conforms to a predetermined condition, described predetermined condition is the marketing that described user was implemented of described historical action information instruction or the cost-range of sales promotion action and income range.

7. the behavior outcome appraisal procedure in a knowledge based space, it is characterised in that include

8. the behavior outcome appraisal procedure in knowledge based space as claimed in claim 7, it is characterized in that, described utilize first user collection in described user's historical data to train in the historical data of first time period to generate user's off-network forecast model and include: utilize first user collection in described user's historical data to train neural network model or Logic Regression Models in the historical data of first time period, to generate described user's off-network forecast model.

9. the behavior outcome appraisal procedure in knowledge based space as claimed in claim 7 or 8, it is characterized in that, described according to described mapping relations, at least one the historical action information obtaining the off-network prediction probability difference of corresponding described user maximum includes: according to described mapping relations, obtaining at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum and conforms to a predetermined condition, described predetermined condition is the marketing that described user was implemented of described historical action information instruction or the cost-range of sales promotion action and income range.

10. the behavior outcome assessment system in a knowledge based space, it is characterised in that including:

11. the behavior outcome assessment system in knowledge based space as claimed in claim 10, it is characterised in that also include:

12. the behavior outcome assessment system in the knowledge based space as described in claim 10 or 11, it is characterized in that, described prediction module, specifically for the historical data of second user's collection described in the second time period is input to described user's off-network forecast model, obtain described second user and concentrate on user's collection that in the 3rd time period, off-network prediction probability is the highest, and the historical data of the 3rd user's collection described in the second time period is input to described user's off-network forecast model, it is thus achieved that described 3rd user concentrates on user's collection that in the 3rd time period, off-network prediction probability is the highest.

13. the behavior outcome assessment system in knowledge based space as claimed in claim 10, it is characterized in that, described selection module, being additionally operable to obtain at least one historical action information that the off-network prediction probability difference of corresponding described user is maximum and conforms to a predetermined condition, described predetermined condition is the marketing that described user was implemented of described historical action information instruction or the cost-range of sales promotion action and income range.

14. the behavior outcome assessment system in a knowledge based space, it is characterised in that including:

15. the behavior outcome assessment system in knowledge based space as claimed in claim 14, it is characterized in that, first user collection in described user's historical data is utilized to train generation first user off-network forecast model in the historical data of first time period described, and utilize in described user's historical data the second user's collection historical data in the second time period train generate second user's off-network forecast model in, described training module trains neural network model or Logic Regression Models specifically for utilizing first user collection in described user's historical data in the historical data of first time period, to generate first user off-network forecast model, and utilize the second user's collection historical data in the second time period in described user's historical data to train neural network model or Logic Regression Models, to generate second user's off-network forecast model.

16. the behavior outcome assessment computing equipment in knowledge based space, including network interface, memorizer and processor, it is characterised in that

Described memorizer is used for storing described user's historical data；