CN108805613A - Electronic device promotes list recommendation method and computer readable storage medium - Google Patents
Electronic device promotes list recommendation method and computer readable storage medium Download PDFInfo
- Publication number
- CN108805613A CN108805613A CN201810502331.3A CN201810502331A CN108805613A CN 108805613 A CN108805613 A CN 108805613A CN 201810502331 A CN201810502331 A CN 201810502331A CN 108805613 A CN108805613 A CN 108805613A
- Authority
- CN
- China
- Prior art keywords
- client
- feature vector
- distribution
- machine learning
- promoted
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Strategic Management (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Entrepreneurship & Innovation (AREA)
- General Engineering & Computer Science (AREA)
- Game Theory and Decision Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention, which discloses a kind of electronic device, promotes list recommends the method and computer readable storage medium, this method to include:It is receiving after promoting client's list, is obtaining the customer data of each client to be promoted, be respectively converted into corresponding feature vector and store;Each feature vector is substituted into trained online machine learning model in advance and carries out analysis prediction, obtains the distribution prediction result of each client to be promoted;The list for meeting the client to be promoted for presetting screening conditions is filtered out, and is issued to the center of attending a banquet and carries out distribution processing;After the distribution feedback data for receiving the center of attending a banquet, obtains each and promote the feature vector for recording corresponding client, and the feature vector of acquisition is divided into positive and negative samples;Each feature vector of acquisition is substituted into the online machine learning model, iteratively faster training is carried out to each feature vector using FTRL algorithms, updates model parameter.The present invention program lift scheme prediction effect, reduces human cost.
Description
Technical field
The present invention relates to technical field of data processing, more particularly to a kind of electronic device promotes list recommendation method and meter
Calculation machine readable storage medium storing program for executing.
Background technology
Traditional commending system is typically to be come from using a batch in the modeling process for business (such as telemarketing)
The data with existing of business side, study to a cured model;The generalization ability of the model depends not only upon well-designed mould
Type ensures with greater need for disposable perfusion mass data.But it in the actual recommendation process of business, is sat especially by phone
Seat is collected under the scene of sample, it is difficult to disposably obtain a large amount of training datas;On the other hand, after model is reached the standard grade, it is for a period of time
Static, even if prediction effect is deteriorated, it is also desirable to update, influence pre- again after being collected into enough data next time
Effect is surveyed, and generally by artificial re -training, the period is week or the moon, and height, and model timeliness are required to human cost
Property is poor.
Invention content
A kind of electronic device of present invention offer promotes list recommendation method and computer readable storage medium, it is intended to be promoted
Forecast result of model and timeliness reduce human cost.
To achieve the above object, electronic device proposed by the present invention, including memory and processor are deposited on the memory
The distribution list commending system that can be run on the processor is contained, the distribution list commending system is held by the processor
Following steps are realized when row:
It is receiving after promoting client's list, is obtaining the customer data of each client to be promoted, waiting promoting visitor by each
The customer data at family is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is carried out
Analysis prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to
The center of attending a banquet carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote note
The feature vector of corresponding client is recorded, and the feature vector of acquisition is divided by positive and negative sample according to each distribution result for promoting record
This;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL (Follow The
Regularized Leader) to the progress iteratively faster training of each feature vector, iterative solution every time allows all before algorithm
The model parameter of the sum of loss function minimum and update.
Preferably, the initial training process of the online machine learning model is:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each visitor
User data is converted to corresponding feature vector, according to it is each promote sample distribution result by the corresponding feature of each customer data to
Amount is divided into positive and negative samples;
The positive and negative samples are inputted the online machine by the model parameter for initializing preset online machine learning model
In device learning model, training is iterated using FTRL algorithms, the sum of all loss functions minimum before iterative solution allows every time
Model parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Preferably, the object function of the model parameter vector of the online machine learning model is:
Wherein, W is exactly model parameter vector, σsIndicate learning rate;
First item g1:t* w is an estimation of the contribution to loss function;
Section 2This is cumulative and is to limit w in each iteration not change too greatly;
Section 3 λ1||w||1For L1 canonicals.
Preferably, the calculation of the model parameter vector of the online machine learning model is:
To the feature vector, X t that the t times iteration is newly added, hyper parameter α, β, λ are introduced1、λ2, it is calculated by the following formula:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then take back object function, then solve wt+1。
The present invention also proposes that a kind of distribution list recommends method, and the method comprising the steps of:
It is receiving after promoting client's list, is obtaining the customer data of each client to be promoted, waiting promoting visitor by each
The customer data at family is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is carried out
Analysis prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to
The center of attending a banquet carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote note
The feature vector of corresponding client is recorded, and the feature vector of acquisition is divided by positive and negative sample according to each distribution result for promoting record
This;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL algorithms to each feature
Vector carries out iteratively faster training, the minimum model parameter of the sum of all loss functions and update before iterative solution allows every time.
Preferably, the initial training process of the online machine learning model is:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each visitor
User data is converted to corresponding feature vector, according to it is each promote sample distribution result by the corresponding feature of each customer data to
Amount is divided into positive and negative samples;
The positive and negative samples are inputted the online machine by the model parameter for initializing preset online machine learning model
In device learning model, training is iterated using FTRL algorithms, the sum of all loss functions minimum before iterative solution allows every time
Model parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Preferably, the object function of the model parameter vector of the online machine learning model is:
Wherein, W is exactly model parameter vector, σsIndicate learning rate;
First item g1:t* w is an estimation of the contribution to loss function;
Section 2This is cumulative and is to limit w in each iteration not change too greatly;
Section 3 λ1||w||1For L1 canonicals.
Preferably, the calculation of the model parameter vector of the online machine learning model is:
The feature vector x that the t times iteration is newly addedt, introduce hyper parameter α, β, λ1、λ2, pass through following calculation processing:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then g, σ and w for will solvingtObject function is taken back, then solves wt+1。
The present invention also proposes that a kind of computer readable storage medium, the computer-readable recording medium storage have distribution name
Single commending system, the distribution list commending system can be executed by least one processor, so that at least one processor
Execute following steps:
It is receiving after promoting client's list, is obtaining the customer data of each client to be promoted, waiting promoting visitor by each
The customer data at family is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is carried out
Analysis prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to
The center of attending a banquet carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote note
The feature vector of corresponding client is recorded, and the feature vector of acquisition is divided by positive and negative sample according to each distribution result for promoting record
This;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL algorithms to each feature
Vector carries out iteratively faster training, the minimum model parameter of the sum of all loss functions and update before iterative solution allows every time.
Preferably, the initial training process of the online machine learning model is:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each visitor
User data is converted to corresponding feature vector, according to it is each promote sample distribution result by the corresponding feature of each customer data to
Amount is divided into positive and negative samples;
The positive and negative samples are inputted the online machine by the model parameter for initializing preset online machine learning model
In device learning model, training is iterated using FTRL algorithms, the sum of all loss functions minimum before iterative solution allows every time
Model parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Technical solution of the present invention, the distribution that client to be promoted is predicted by using the analysis of online machine learning model are predicted
As a result, the list that selected client is filtered out according to distribution prediction result, which is issued to the center of attending a banquet, carries out distribution processing, and in real time
It receives center of attending a banquet and promotes the distribution feedback data being disposed, it is right from the positive and negative samples sorted out promoted in feedback data
Online machine learning model is iterated training with more new model.For existing conventional recommendation model, side of the present invention
Case carries out recommendation prediction using online machine learning model, and requirement of the model to amount of training data is small, enough without waiting for collecting
The data of retraining after more data, the center feedback that can be attended a banquet with real-time reception avoid model for a long time to be iterated update
The prediction effect for not updating and bringing is deteriorated, and ensures the good prediction effect and timeliness of model, and the repetitive exercise of model
It is automatic on-line iteration, is not necessarily to offline expert along training, human cost is low.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
The structure shown according to these attached drawings obtains other attached drawings.
Fig. 1 is the flow diagram that the present invention promotes that list recommends one embodiment of method;
Fig. 2 is the flow diagram that the present invention promotes the initial training of online machine learning model in list recommendation method;
Fig. 3 is the running environment schematic diagram that the present invention promotes list commending system preferred embodiment;
Fig. 4 is the Program modual graph that the present invention promotes one embodiment of list commending system.
The embodiments will be further described with reference to the accompanying drawings for the realization, the function and the advantages of the object of the present invention.
Specific implementation mode
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the present invention.
The present invention proposes that a kind of distribution list recommends method, and the processor in electronic device is executing its distribution list recommendation
Realize that the distribution list recommends method when system.
As shown in FIG. 1, FIG. 1 is the present invention to promote the flow diagram that list recommends one embodiment of method.
In the present embodiment, distribution list recommendation method includes:
Step S10 is being received after promoting client's list, obtains the customer data of each client to be promoted, will be each
The customer data of client to be promoted is respectively converted into corresponding feature vector and stores;
Customer database in electronic device is stored with the customer data of each client, and customer data includes many of client
The data and label of dimension, for example, gender, the age, occupation, income level, consumption purchaser record (comprising buying pattern, channel,
Quantity, the amount of money, payment method etc.) etc. correlated characteristics data.Electronic device receive it is to be uploaded after promoting client's list,
The customer data for getting each client from customer database according to the Customer ID in client's list to be promoted, then will be each
The customer data of client is converted to corresponding feature vector, and the corresponding feature vector of each client is stored to customer database
In.
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning mould in advance by step S20
Analysis prediction is carried out in type, obtains the distribution prediction result of each client to be promoted;
Electronic device obtain it is each wait promote client feature vector after, by obtained each feature vector substitute into electronics
In device in advance trained online machine learning model, which analyzes each feature vector
Prediction, obtains the distribution prediction result of each client to be promoted.In the present embodiment, which can be:It promotes into
Power (for example, 30%, 80% etc.) is promoted difficulty (such as easy, general, difficulty etc.) or is promoted and recommends score value etc..
Step S30 is filtered out and is promoted the list that prediction result meets the client to be promoted for presetting screening conditions, by the name
It is singly issued to the center of attending a banquet and carries out distribution processing;
Electronic device utilizes pre-set screening conditions pair after the distribution prediction result for obtaining each client to be recommended
Each distribution prediction result is screened, with the distribution prediction result that screening obtains meeting screening conditions, the distribution screened
The corresponding client of prediction result is then selected client, the list of this selected client is issued to center of attending a banquet and carries out sales promotion department
Reason.
Step S40 is obtained each in the distribution feedback data after receiving the distribution feedback data at the center of attending a banquet
Item promotes the feature vector for recording corresponding client, and is divided into the feature vector of acquisition according to each distribution result for promoting record
Positive and negative samples;
After the center of attending a banquet gets the list of the selected client of electronic device recommendation, distributes to seat personnel and go to be pushed away
Pin, due to by phone or network promote usual closing time differ (some may be struck a bargain on the same day, some may be week or
Month);After distribution result (crack a prospect or promote failure) of the center of attending a banquet in the client that processing has been promoted in confirmation, center of attending a banquet
Timing or the promotional data of newest determination can be fed back into electronic device in real time, that is, the center of attending a banquet may disposably feed back one or
A collection of promotional data is to electronic device.After electronic device receives the distribution feedback data at the center of attending a banquet, then from customer database
It is middle obtain to promote each in feedback data and promote record the feature vector of corresponding client (customer data of the client be to this
Client had carried out feature vector before being predicted and has converted and store, therefore here can be directly according to Customer ID in client
Searched in database to obtain), each feature vector of acquisition (is cracked a prospect or is promoted according to the distribution result for promoting record
Failure), the distribution cracked a prospect is recorded into corresponding feature vector as positive sample, the distribution for promoting failure is recorded corresponding
Feature vector is as negative sample, for being trained update to online machine learning model.
Step S50 substitutes into each feature vector of acquisition in the online machine learning model, using FTRL
(Follow The Regularized Leader) algorithm carries out iteratively faster training to each feature vector, and each iteration is asked
The minimum model parameter of the sum of all loss functions and update before solution allows.
The positive and negative samples (feature vector obtained) of acquisition are substituted into the online machine learning model by electronic device
In, by FTRL algorithms to each sample iteratively faster to be updated to the online machine learning model, each iteration is all asked
The model parameter that the sum of all loss functions are minimum before allowing is released, model parameter is updated after solving and is changed next time again
In generation, so repeats, until all samples all iteration are completed, finally obtains updated online machine learning model.Since this is obtained
The feature vector taken is that the online machine learning model carries out the feature vector predicted, the feature vector of the acquisition is confirmed
After sample type (confirm be positive sample or negative sample), then for being trained to online machine learning model, correct and
Learning effect is more preferably.In addition, FTRL algorithms can generate the effect of rarefaction, the model trained can be smaller so that online
Machine learning model is more conducive to storage and prediction in real time on line.
The present embodiment technical solution, the distribution that client to be promoted is predicted by using the analysis of online machine learning model are pre-
Survey carries out distribution processing as a result, the list for filtering out selected client according to distribution prediction result is issued to the center of attending a banquet, and real
When receive center of attending a banquet and promote the distribution feedback data that is disposed, from the positive and negative samples sorted out promoted in feedback data,
Training is iterated with more new model to online machine learning model.For existing conventional recommendation model, the present invention
Scheme carries out recommendation prediction using online machine learning model, and requirement of the model to amount of training data is small, sufficient without waiting for collecting
Retraining after enough data, the data for the center feedback that can be attended a banquet with real-time reception are to be iterated update, when avoiding model long
Between do not update and the prediction effect that brings is deteriorated, ensure the good prediction effect and timeliness of model, and the iteration instruction of model
White silk is automatic on-line iteration, is not necessarily to offline expert along training, and human cost is low.
As shown in Fig. 2, Fig. 2 is the stream of the initial training of online machine learning model in present invention distribution list recommendation method
Journey schematic diagram;In this embodiment, the initial training process of the online machine learning model is:
Step S60 obtains the distribution sample of preset quantity, obtains each customer data for promoting the corresponding client of sample,
Each customer data is converted into corresponding feature vector, is corresponded to each customer data according to each distribution result for promoting sample
Feature vector be divided into positive and negative samples;
The distribution sample is history promotional data, is all the distribution result for having confirmation, i.e., each distribution sample is all
Know and cracked a prospect or fail, according to the affiliated client of each distribution sample, each client is obtained from customer database
Customer data, each customer data is converted into corresponding feature vector.It is the distribution sample cracked a prospect by result is promoted,
Its corresponding feature vector is as positive sample;It is to promote the distribution sample of failure, corresponding feature vector conduct to promote result
Negative sample.
Step S70 initializes the model parameter of preset online machine learning model, and the positive and negative samples are inputted institute
It states in online machine learning model, training is iterated using FTRL algorithms, all loss functions before iterative solution allows every time
The sum of minimum model parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Online machine learning model is pre-established in electronic device, to the online machine learning model repetitive exercise
Before, first initialize the model parameter (can with indirect assignment, can also random initializtion) of the online machine learning model, electronics dress
It sets and inputs above-mentioned positive and negative samples in the online machine learning model, iteratively faster is calculated by FTRL, an iteration obtains once
New model parameter once updates online machine learning model, next iteration is then in newer online engineering
It is carried out on the basis of habit model, after the completion of to sample data repetitive exercise, eventually confirms and obtain online machine learning model
Updated model parameter is to get to newest online machine learning model, for predicting the analysis of client to be recommended.
In the present embodiment, the object function of the model parameter vector of the online machine learning model is:
Wherein, W is exactly model parameter vector (w generally use random initializtions, can also indirect assignment), σsIndicate study
Rate;
First item g1:t* w is one of the contribution to loss function estimation namely gradient or accumulative gradient;
Section 2This is cumulative and is to limit w in each iteration not change too greatly, also defines
New iteration result should not be too far from the solution that iteration is crossed;
Section 3 λ1||w||1For L1 canonicals.
Specifically, the solution mode of the model parameter vector of the online machine learning model is:
In the feature vector x that the t times iteration is newly addedt, introduce hyper parameter α, β, λ1、λ2, pass through following calculation processing:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then g, σ and w for will solvingtObject function is taken back, then solves wt+1(i.e. t+1 iteration
Model parameter vector afterwards).
In addition, the present invention also proposes a kind of distribution list commending system.
Referring to Fig. 3, being the running environment schematic diagram that the present invention promotes 10 preferred embodiment of list commending system.
In the present embodiment, list commending system 10 is promoted to install and run in electronic device 1.Electronic device 1 can be with
It is the computing devices such as desktop PC, notebook, palm PC and server.The electronic device 1 may include, but not only limit
In memory 11, processor 12 and display 13.Fig. 3 illustrates only the electronic device 1 with component 11-13, it should be understood that
Be, it is not required that implement all components shown, the implementation that can be substituted is more or less component.
Memory 11 can be the internal storage unit of electronic device 1 in some embodiments, such as the electronic device 1
Hard disk or memory.Memory 11 can also be the External memory equipment of electronic device 1, such as electronics dress in further embodiments
Set the plug-in type hard disk being equipped on 1, intelligent memory card (Smart Media Card, SMC), secure digital (Secure
Digital, SD) card, flash card (Flash Card) etc..Further, memory 11 can also both include the interior of electronic device 1
Portion's storage unit also includes External memory equipment.Memory 11 is for storing the application software for being installed on electronic device 1 and all kinds of
Data, such as promote the program code etc. of list commending system 10.Memory 11 can be also used for temporarily storing and export
Or the data that will be exported.
Processor 12 can be in some embodiments a central processing unit (Central Processing Unit,
CPU), microprocessor or other data processing chips, the program code for being stored in run memory 11 or processing data, example
It such as executes and promotes list commending system 10.
Display 13 can be in some embodiments light-emitting diode display, liquid crystal display, touch-control liquid crystal display and
OLED (Organic Light-Emitting Diode, Organic Light Emitting Diode) touches device etc..Display 13 is for being shown in
The information that is handled in electronic device 1 and for showing visual user interface.The component 11-13 of electronic device 1, which passes through, is
System bus is in communication with each other.
Referring to Fig. 4, being the Program modual graph that the present invention promotes 10 preferred embodiment of list commending system.In the present embodiment
In, one or more modules can be divided by promoting list commending system 10, one or more module is stored in storage
In device 11, and it is performed by one or more processors (the present embodiment is processor 12), to complete the present invention.For example, in Fig. 4
In, acquisition module 101, analysis prediction module 102, screening module 103, feedback can be divided by promoting list commending system 10
Data preparation module 104 and on-line training module 105.The so-called module of the present invention is to refer to complete a series of of specific function
Computer program instructions section promotes the implementation procedure of list commending system 10 in the electronic apparatus 1 than program more suitable for description,
Wherein:
Acquisition module 101 obtains client's number of each client to be promoted for receiving after promoting client's list
According to the customer data of each client to be promoted is respectively converted into corresponding feature vector and is stored;
Customer database in electronic device is stored with the customer data of each client, and customer data includes many of client
The data and label of dimension, for example, gender, the age, occupation, income level, consumption purchaser record (comprising buying pattern, channel,
Quantity, the amount of money, payment method etc.) etc. correlated characteristics data.Electronic device receive it is to be uploaded after promoting client's list,
The customer data for getting each client from customer database according to the Customer ID in client's list to be promoted, then will be each
The customer data of client is converted to corresponding feature vector, and the corresponding feature vector of each client is stored to customer database
In.
Analyze prediction module 102, for by the corresponding feature vector of each client to be promoted substitute into it is advance it is trained
Analysis prediction is carried out in line machine learning model, obtains the distribution prediction result of each client to be promoted;
Obtain it is each after the feature vector for promoting client, will obtained each feature vector substitute into it is pre- in electronic device
First in trained online machine learning model, which carries out analysis prediction to each feature vector, obtains
Go out the distribution prediction result of each client to be promoted.In the present embodiment, which can be:Rate of cracking a prospect (example
Such as, it 30%, 80% etc.), promotes difficulty (such as easy, general, difficulty etc.) or promotes and recommend score value etc..
Screening module 103, for filtering out the list promoted prediction result and meet the client to be promoted for presetting screening conditions,
The list is issued to the center of attending a banquet and carries out distribution processing;
Electronic device utilizes pre-set screening conditions pair after the distribution prediction result for obtaining each client to be recommended
Each distribution prediction result is screened, with the distribution prediction result that screening obtains meeting screening conditions, the distribution screened
The corresponding client of prediction result is then selected client, the list of this selected client is issued to center of attending a banquet and carries out sales promotion department
Reason.
Feedback data sorting module 104, for after receiving the distribution feedback data at the center of attending a banquet, obtaining the distribution
Each in feedback data promotes the feature vector for recording corresponding client, and will be obtained according to each distribution result for promoting record
The feature vector taken is divided into positive and negative samples;
After the center of attending a banquet gets the list of the selected client of electronic device recommendation, distributes to seat personnel and go to be pushed away
Pin, due to by phone or network promote usual closing time differ (some may be struck a bargain on the same day, some may be week or
Month);After distribution result (crack a prospect or promote failure) of the center of attending a banquet in the client that processing has been promoted in confirmation, center of attending a banquet
Timing or the promotional data of newest determination can be fed back into electronic device in real time, that is, the center of attending a banquet may disposably feed back one or
A collection of promotional data is to electronic device.After electronic device receives the distribution feedback data at the center of attending a banquet, then from customer database
It is middle obtain to promote each in feedback data and promote record the feature vector of corresponding client (customer data of the client be to this
Client had carried out feature vector before being predicted and has converted and store, therefore here can be directly according to Customer ID in client
Searched in database to obtain), each feature vector of acquisition (is cracked a prospect or is promoted according to the distribution result for promoting record
Failure), the distribution cracked a prospect is recorded into corresponding feature vector as positive sample, the distribution for promoting failure is recorded corresponding
Feature vector is as negative sample, for being trained update to online machine learning model.
On-line training module 105, each feature vector for that will obtain are substituted into the online machine learning model, are adopted
Iteratively faster training is carried out to each feature vector with FTRL (Follow The Regularized Leader) algorithm, every time
The minimum model parameter of the sum of all loss functions and update before iterative solution allows.
The positive and negative samples (feature vector obtained) of acquisition are substituted into the online machine learning model by electronic device
In, by FTRL algorithms to each sample iteratively faster to be updated to the online machine learning model, each iteration is all asked
The model parameter that the sum of all loss functions are minimum before allowing is released, model parameter is updated after solving and is changed next time again
In generation, so repeats, until all samples all iteration are completed, finally obtains updated online machine learning model.Since this is obtained
The feature vector taken is that the online machine learning model carries out the feature vector predicted, the feature vector of the acquisition is confirmed
After sample type (confirm be positive sample or negative sample), then for being trained to online machine learning model, correct and
Learning effect is more preferably.In addition, FTRL algorithms can generate the effect of rarefaction, the model trained can be smaller so that online
Machine learning model is more conducive to storage and prediction in real time on line.
The present embodiment technical solution, the distribution that client to be promoted is predicted by using the analysis of online machine learning model are pre-
Survey carries out distribution processing as a result, the list for filtering out selected client according to distribution prediction result is issued to the center of attending a banquet, and real
When receive center of attending a banquet and promote the distribution feedback data that is disposed, from the positive and negative samples sorted out promoted in feedback data,
Training is iterated with more new model to online machine learning model.For existing conventional recommendation model, the present invention
Scheme carries out recommendation prediction using online machine learning model, and requirement of the model to amount of training data is small, sufficient without waiting for collecting
Retraining after enough data, the data for the center feedback that can be attended a banquet with real-time reception are to be iterated update, when avoiding model long
Between do not update and the prediction effect that brings is deteriorated, ensure the good prediction effect and timeliness of model, and the iteration instruction of model
White silk is automatic on-line iteration, is not necessarily to offline expert along training, and human cost is low.
In the present embodiment, the initial training process of the online machine learning model includes:
1, the distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, it will be each
Customer data is converted to corresponding feature vector, according to each distribution result for promoting sample by the corresponding feature of each customer data
Vector is divided into positive and negative samples;
The distribution sample is history promotional data, is all the distribution result for having confirmation, i.e., each distribution sample is all
Know and cracked a prospect or fail, according to the affiliated client of each distribution sample, each client is obtained from customer database
Customer data, each customer data is converted into corresponding feature vector.It is the distribution sample cracked a prospect by result is promoted,
Its corresponding feature vector is as positive sample;It is to promote the distribution sample of failure, corresponding feature vector conduct to promote result
Negative sample.
2, the model parameter for initializing preset online machine learning model, positive and negative samples input is described online
In machine learning model, training is iterated using FTRL algorithms, the sum of all loss functions are most before iterative solution allows every time
Small model parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Online machine learning model is pre-established in electronic device, to the online machine learning model repetitive exercise
Before, first initialize the model parameter (can with indirect assignment, can also random initializtion) of the online machine learning model, electronics dress
It sets and inputs above-mentioned positive and negative samples in the online machine learning model, iteratively faster is calculated by FTRL, an iteration obtains once
New model parameter once updates online machine learning model, next iteration is then in newer online engineering
It is carried out on the basis of habit model, after the completion of to sample data repetitive exercise, eventually confirms and obtain online machine learning model
Updated model parameter is to get to newest online machine learning model, for predicting the analysis of client to be recommended.
In the present embodiment, the object function of the model parameter vector of the online machine learning model is:
Wherein, W is exactly model parameter vector (w generally use random initializtions, can also indirect assignment), σsIndicate study
Rate;
First item g1:t* w is one of the contribution to loss function estimation namely gradient or accumulative gradient;
Section 2This is cumulative and is to limit w in each iteration not change too greatly, also defines
New iteration result should not be too far from the solution that iteration is crossed;
Section 3 λ1||w||1For L1 canonicals.
Specifically, the solution mode of the model parameter vector of the online machine learning model is:
In the feature vector x that the t times iteration is newly addedt, introduce hyper parameter α, β, λ1、λ2, pass through following calculation processing:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then g, σ and w for will solvingtObject function is taken back, then solves wt+1(i.e. t+1 iteration
Model parameter vector afterwards).
The distribution list commending system of the electronic device of the present embodiment preferably using based on spark big datas technology and
Redis memory storage technologies, have so been greatly improved operational performance, reduce response business time cost and manpower at
This.
Further, the present invention also proposes that a kind of computer readable storage medium, the computer readable storage medium are deposited
Contain and promote list commending system, the distribution list commending system can execute by least one processor so that it is described at least
One processor executes the distribution list in any of the above-described embodiment and recommends method.
The foregoing is merely the preferred embodiment of the present invention, are not intended to limit the scope of the invention, every at this
Under the inventive concept of invention, using equivalent structure transformation made by description of the invention and accompanying drawing content, or directly/use indirectly
In the scope of patent protection that other related technical areas are included in the present invention.
Claims (10)
1. a kind of electronic device, which is characterized in that the electronic device includes memory and processor, is stored on the memory
There are the distribution list commending system that can be run on the processor, the distribution list commending system to be executed by the processor
Shi Shixian following steps:
It is receiving after promoting client's list, the customer data of each client to be promoted is being obtained, by each client's to be promoted
Customer data is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is analyzed
Prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to and is attended a banquet
Center carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote record pair
The feature vector of the client answered, and the feature vector of acquisition is divided by positive and negative samples according to each distribution result for promoting record;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL (Follow The
Regularized Leader) to the progress iteratively faster training of each feature vector, iterative solution every time allows all before algorithm
The model parameter of the sum of loss function minimum and update.
2. electronic device as described in claim 1, which is characterized in that the initial training process of the online machine learning model
For:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each client's number
According to corresponding feature vector is converted to, the corresponding feature vector of each customer data is divided according to each distribution result for promoting sample
At positive and negative samples;
The positive and negative samples are inputted the online engineering by the model parameter for initializing preset online machine learning model
It practises in model, training is iterated using FTRL algorithms, the minimum mould of the sum of all loss functions before iterative solution allows every time
Shape parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
3. electronic device as claimed in claim 1 or 2, which is characterized in that the model parameter of the online machine learning model
Vector object function be:
Wherein, W is exactly model parameter vector, σsIndicate learning rate;
First item g1:t* w is an estimation of the contribution to loss function;
Section 2This is cumulative and is to limit w in each iteration not change too greatly;
Section 3 λ1||w||1For L1 canonicals.
4. electronic device as claimed in claim 3, which is characterized in that the model parameter vector of the online machine learning model
Calculation be:
To the feature vector, X t that the t times iteration is newly added, hyper parameter α, β, λ are introduced1、λ2, it is calculated by the following formula:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then take back object function, then solve wt+1。
5. a kind of distribution list recommends method, which is characterized in that the method comprising the steps of:
It is receiving after promoting client's list, the customer data of each client to be promoted is being obtained, by each client's to be promoted
Customer data is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is analyzed
Prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to and is attended a banquet
Center carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote record pair
The feature vector of the client answered, and the feature vector of acquisition is divided by positive and negative samples according to each distribution result for promoting record;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL algorithms to each feature vector
Iteratively faster training is carried out, the minimum model parameter of the sum of all loss functions and update before iterative solution allows every time.
6. as claimed in claim 5 promote list recommend method, which is characterized in that the online machine learning model it is initial
Training process is:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each client's number
According to corresponding feature vector is converted to, the corresponding feature vector of each customer data is divided according to each distribution result for promoting sample
At positive and negative samples;
The positive and negative samples are inputted the online engineering by the model parameter for initializing preset online machine learning model
It practises in model, training is iterated using FTRL algorithms, the minimum mould of the sum of all loss functions before iterative solution allows every time
Shape parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
7. as distribution list described in claim 5 or 6 recommends method, which is characterized in that the online machine learning model
The object function of model parameter vector is:
Wherein, W is exactly model parameter vector, σsIndicate learning rate;
First item g1:t* w is an estimation of the contribution to loss function;
Section 2This is cumulative and is to limit w in each iteration not change too greatly;
Section 3 λ1||w||1For L1 canonicals.
8. promoting list as claimed in claim 7 recommends method, which is characterized in that the model of the online machine learning model
The calculation of parameter vector is:
The feature vector x that the t times iteration is newly addedt, introduce hyper parameter α, β, λ1、λ2, pass through following calculation processing:
Predict pt=σ (xt·w)using the wt,i computed above
Observe label yt∈{0,1}
for all i∈I do
gi=(pt-yt)xi#gradient of loss w.r.t.wi
zi←zi+gi-σiwt,i
ni←ni+gi 2
Solve g, σ and wt, then g, σ and w for will solvingtObject function is taken back, then solves wt+1。
9. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has distribution list
Commending system, the distribution list commending system can be executed by least one processor, so that at least one processor is held
Row following steps:
It is receiving after promoting client's list, the customer data of each client to be promoted is being obtained, by each client's to be promoted
Customer data is respectively converted into corresponding feature vector and stores;
The corresponding feature vector of each client to be promoted is substituted into trained online machine learning model in advance and is analyzed
Prediction, obtains the distribution prediction result of each client to be promoted;
It filters out and promotes the list that prediction result meets the client to be promoted for presetting screening conditions, the list is issued to and is attended a banquet
Center carries out distribution processing;
After receiving the distribution feedback data at the center of attending a banquet, obtains each promoted in feedback data and promote record pair
The feature vector of the client answered, and the feature vector of acquisition is divided by positive and negative samples according to each distribution result for promoting record;
Each feature vector of acquisition is substituted into the online machine learning model, using FTRL algorithms to each feature vector
Iteratively faster training is carried out, the minimum model parameter of the sum of all loss functions and update before iterative solution allows every time.
10. computer readable storage medium as claimed in claim 9, which is characterized in that the online machine learning model
Initial training process is:
The distribution sample of preset quantity is obtained, each customer data for promoting the corresponding client of sample is obtained, by each client's number
According to corresponding feature vector is converted to, the corresponding feature vector of each customer data is divided according to each distribution result for promoting sample
At positive and negative samples;
The positive and negative samples are inputted the online engineering by the model parameter for initializing preset online machine learning model
It practises in model, training is iterated using FTRL algorithms, the minimum mould of the sum of all loss functions before iterative solution allows every time
Shape parameter, to confirm the updated model parameter of the online machine learning model after the completion of repetitive exercise.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810502331.3A CN108805613A (en) | 2018-05-23 | 2018-05-23 | Electronic device promotes list recommendation method and computer readable storage medium |
PCT/CN2018/102111 WO2019223145A1 (en) | 2018-05-23 | 2018-08-24 | Electronic device, promotion list recommendation method and system, and computer-readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810502331.3A CN108805613A (en) | 2018-05-23 | 2018-05-23 | Electronic device promotes list recommendation method and computer readable storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108805613A true CN108805613A (en) | 2018-11-13 |
Family
ID=64092807
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810502331.3A Withdrawn CN108805613A (en) | 2018-05-23 | 2018-05-23 | Electronic device promotes list recommendation method and computer readable storage medium |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN108805613A (en) |
WO (1) | WO2019223145A1 (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110400209A (en) * | 2019-07-03 | 2019-11-01 | 上海上湖信息技术有限公司 | Determination method and device, storage medium, the terminal of user credit |
CN111797852A (en) * | 2019-04-09 | 2020-10-20 | Oppo广东移动通信有限公司 | Method and device for adjusting algorithm model, storage medium and electronic equipment |
CN112001434A (en) * | 2020-08-14 | 2020-11-27 | 深圳世间乐见科技有限公司 | Color value prediction method, device, terminal and medium |
CN112084408A (en) * | 2020-09-08 | 2020-12-15 | 中国平安财产保险股份有限公司 | List data screening method and device, computer equipment and storage medium |
CN115329084A (en) * | 2022-08-26 | 2022-11-11 | 苏州大学 | Junk mail classification method and system based on sparse linear online learning |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105989374B (en) * | 2015-03-03 | 2019-12-24 | 阿里巴巴集团控股有限公司 | Method and equipment for training model on line |
CN105184321B (en) * | 2015-09-10 | 2018-12-14 | 北京金山安全软件有限公司 | Data processing method and device for ftrl model |
CN107194532A (en) * | 2017-04-07 | 2017-09-22 | 广东精点数据科技股份有限公司 | Insurance business analysis method based on big data |
CN107609060A (en) * | 2017-08-28 | 2018-01-19 | 百度在线网络技术(北京)有限公司 | Resource recommendation method and device |
-
2018
- 2018-05-23 CN CN201810502331.3A patent/CN108805613A/en not_active Withdrawn
- 2018-08-24 WO PCT/CN2018/102111 patent/WO2019223145A1/en active Application Filing
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111797852A (en) * | 2019-04-09 | 2020-10-20 | Oppo广东移动通信有限公司 | Method and device for adjusting algorithm model, storage medium and electronic equipment |
CN110400209A (en) * | 2019-07-03 | 2019-11-01 | 上海上湖信息技术有限公司 | Determination method and device, storage medium, the terminal of user credit |
CN112001434A (en) * | 2020-08-14 | 2020-11-27 | 深圳世间乐见科技有限公司 | Color value prediction method, device, terminal and medium |
CN112084408A (en) * | 2020-09-08 | 2020-12-15 | 中国平安财产保险股份有限公司 | List data screening method and device, computer equipment and storage medium |
CN112084408B (en) * | 2020-09-08 | 2023-11-21 | 中国平安财产保险股份有限公司 | List data screening method, device, computer equipment and storage medium |
CN115329084A (en) * | 2022-08-26 | 2022-11-11 | 苏州大学 | Junk mail classification method and system based on sparse linear online learning |
Also Published As
Publication number | Publication date |
---|---|
WO2019223145A1 (en) | 2019-11-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108805613A (en) | Electronic device promotes list recommendation method and computer readable storage medium | |
US11836615B2 (en) | Bayesian nonparametric learning of neural networks | |
Haas et al. | Clamshell: Speeding up crowds for low-latency data labeling | |
CN103502899B (en) | Dynamic prediction Modeling Platform | |
Garcia-Molina et al. | Challenges in data crowdsourcing | |
US9639807B2 (en) | Method and system for forecasting future events | |
WO2019109790A1 (en) | Sales volume prediction method and device, and computer-readable storage medium | |
CN110956272A (en) | Method and system for realizing data processing | |
CN110163723A (en) | Recommended method, device, computer equipment and storage medium based on product feature | |
CN109544197A (en) | A kind of customer churn prediction technique and device | |
CN114663198A (en) | Product recommendation method, device and equipment based on user portrait and storage medium | |
CN107862551B (en) | Method and device for predicting network application promotion effect and terminal equipment | |
CN107729915A (en) | For the method and system for the key character for determining machine learning sample | |
CN108108820A (en) | For selecting the method and system of the feature of machine learning sample | |
CN108846695A (en) | The prediction technique and device of terminal replacement cycle | |
CN110069629A (en) | House transaction task processing method, equipment, storage medium and device | |
Gao et al. | Discrete harmony search algorithm for scheduling and rescheduling the reprocessing problems in remanufacturing: a case study | |
CN111814056A (en) | Supplier recommendation method based on information processing and related equipment | |
CN110276642A (en) | Billboard recommended method, device and computer readable storage medium | |
CN105869017A (en) | Method and system for predicting ticket prices | |
CN110852785A (en) | User grading method, device and computer readable storage medium | |
CN109740036A (en) | OTA platform hotel's sort method and device | |
Perera et al. | A rule-based system for automated generation of serverless-microservices architecture | |
CN109255629A (en) | A kind of customer grouping method and device, electronic equipment, readable storage medium storing program for executing | |
CN108846687A (en) | Client segmentation method, apparatus and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20181113 |