CN110378434A - Training method, recommended method, device and the electronic equipment of clicking rate prediction model - Google Patents
Training method, recommended method, device and the electronic equipment of clicking rate prediction model Download PDFInfo
- Publication number
- CN110378434A CN110378434A CN201910673756.5A CN201910673756A CN110378434A CN 110378434 A CN110378434 A CN 110378434A CN 201910673756 A CN201910673756 A CN 201910673756A CN 110378434 A CN110378434 A CN 110378434A
- Authority
- CN
- China
- Prior art keywords
- clicking rate
- prediction model
- rate prediction
- weight parameter
- sample
- 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.)
- Pending
Links
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Economics (AREA)
- Artificial Intelligence (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Tourism & Hospitality (AREA)
- Entrepreneurship & Innovation (AREA)
- General Business, Economics & Management (AREA)
- Operations Research (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Quality & Reliability (AREA)
- Marketing (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention provides training method, recommended method, device, electronic equipment and the storage mediums of a kind of clicking rate prediction model;The training method of clicking rate prediction model includes: the click result for obtaining sample characteristics and sample characteristics corresponding to user;Clicking rate prediction model is initialized according to the weight parameter of setting;The sample characteristics are handled by the clicking rate prediction model to obtain prediction clicking rate;According to the error and zero norm regular terms building objective function between the click result and the prediction clicking rate;By the objective function in the clicking rate prediction model error described in backpropagation, and update in communication process the weight parameter of the clicking rate prediction model.By means of the invention it is possible to promote the sparsity of weight parameter, the file size of the clicking rate prediction model of generation is reduced, promotes the accuracy of clicking rate prediction.
Description
Technical field
The present invention relates to artificial intelligence technology more particularly to a kind of training method of clicking rate prediction model, recommended method,
Device, electronic equipment and storage medium.
Background technique
Artificial intelligence (AI, Artificial Intelligence) is to utilize digital computer or digital computer control
Machine simulation, extension and the intelligence for extending people of system, perception environment obtain knowledge and the reason using Knowledge Acquirement optimum
By, method, technology and application system.Machine learning (ML, Machine Learning) is a multi-field cross discipline, is people
The core of work intelligence, is the fundamental way for making computer have intelligence, and application spreads every field.
Clicking rate prediction is the important application branch of machine learning, for the provider of content-data, it will usually
Clicking rate prediction is carried out to various content-datas, so that the content-data of push is made to meet the hobby of user, it is such as military to preference
User recommend military news, makeups advertisement etc. is recommended to the user of preference makeups.
In the clicking rate prediction model scheme that the relevant technologies provide, the sparsity for the weight parameter that training obtains is poor, model
The volume of file is larger, model deployment the limited situation of device memory under, model may limited performance even can not
It uses, the applicability for carrying out clicking rate prediction is low.
Summary of the invention
The embodiment of the present invention provide the training method of clicking rate prediction model a kind of, recommended method, device, electronic equipment and
Storage medium, the sparsity for being able to ascend weight parameter, the applicability for reducing model volume and promoting clicking rate prediction.
The technical solution of the embodiment of the present invention is achieved in that
The embodiment of the present invention provides a kind of training method of clicking rate prediction model, comprising:
It obtains sample characteristics and sample characteristics corresponds to the click result of user;
Clicking rate prediction model is initialized according to the weight parameter of setting;
The sample characteristics are handled by the clicking rate prediction model to obtain prediction clicking rate;
According to the error and zero norm regular terms building target letter between the click result and the prediction clicking rate
Number;
By the objective function in the clicking rate prediction model error described in backpropagation, and in communication process
Update the weight parameter of the clicking rate prediction model.
In the above scheme, further includes:
Retain the numerical value of non-zero and corresponding dimension in the updated weight parameter.
In the above scheme, the click result for obtaining sample characteristics and sample characteristics corresponding to user, comprising:
Sample characteristics are obtained from database and sample characteristics correspond to the click result of user;Alternatively,
Data flow obtains sample characteristics from line and sample characteristics correspond to the click result of user.
The embodiment of the present invention provides a kind of recommended method based on clicking rate prediction model, comprising:
Obtain user data and at least two content-datas;
It combines each content-data with user data one-to-one correspondence, and feature is carried out to combined data and is mentioned
It takes, obtains user characteristics;
The user characteristics are handled by the clicking rate prediction model, obtain prediction clicking rate;
Recommend the corresponding content-data of prediction clicking rate for meeting clicking rate value condition.
The embodiment of the present invention provides a kind of training device of clicking rate prediction model, comprising:
Module is obtained, corresponds to the click result of user for obtaining sample characteristics and sample characteristics;
Initialization module, for initializing clicking rate prediction model according to the weight parameter of setting;
Processing module obtains prediction click for being handled by the clicking rate prediction model the sample characteristics
Rate;
Module is constructed, for clicking result and the error predicted between clicking rate and zero norm just according to described
Then item constructs objective function;
Update module, for by the objective function in the clicking rate prediction model error described in backpropagation,
And the weight parameter of the clicking rate prediction model is updated in communication process.
In the above scheme, the training device of the clicking rate prediction model further include:
Model modification module, for retaining the numerical value of non-zero and corresponding dimension in the updated weight parameter.
In the above scheme, the acquisition module is also used to:
Sample characteristics are obtained from database and sample characteristics correspond to the click result of user;Alternatively,
Data flow obtains sample characteristics from line and sample characteristics correspond to the click result of user.
The embodiment of the present invention provides a kind of recommendation apparatus based on clicking rate prediction model, comprising:
Data acquisition module, for obtaining user data and at least two content-datas;
Characteristic extracting module, for combining each content-data with user data one-to-one correspondence, and to combination
Data carry out feature extraction, obtain user characteristics;
Clicking rate prediction module is obtained for being handled by the clicking rate prediction model the user characteristics
Predict clicking rate;
Recommending module, the corresponding content-data of prediction clicking rate for recommending to meet clicking rate value condition.
The embodiment of the present invention provides a kind of electronic equipment, comprising:
Memory, for storing executable instruction;
Processor when for executing the executable instruction stored in the memory, is realized provided in an embodiment of the present invention
The training method of clicking rate prediction model, or the recommended method based on clicking rate prediction model.
The embodiment of the present invention provides a kind of storage medium, is stored with executable instruction, real when for causing processor to execute
The training method of existing clicking rate prediction model provided in an embodiment of the present invention, or the recommendation side based on clicking rate prediction model
Method.
The embodiment of the present invention has the advantages that
The embodiment of the present invention handles sample characteristics to obtain prediction click according to the clicking rate prediction model of initialization
Rate according to existing click result and predicts that error and zero norm regular terms between clicking rate construct objective function, according to
Objective function reverse propagated error updates the weight parameter of clicking rate prediction model in communication process, improves updated
The sparsity of weight parameter reduces the file size of clicking rate prediction model, improves clicking rate prediction for various electronics
The applicability of equipment.
Detailed description of the invention
Fig. 1 is that an optional framework of the training system 100 of clicking rate prediction model provided in an embodiment of the present invention shows
It is intended to;
Fig. 2A is an optional configuration diagram of server 200 provided in an embodiment of the present invention;
Fig. 2 B is another optional configuration diagram of server 200 provided in an embodiment of the present invention;
Fig. 3 is an optional process signal of the training method of clicking rate prediction model provided in an embodiment of the present invention
Figure;
Fig. 4 is another optional process signal of the training method of clicking rate prediction model provided in an embodiment of the present invention
Figure;
Fig. 5 is that an optional process of the recommended method provided in an embodiment of the present invention based on clicking rate prediction model is shown
It is intended to;
Fig. 6 is an optional configuration diagram of clicking rate forecasting system provided in an embodiment of the present invention;
Fig. 7 is an optional configuration diagram of model training module provided in an embodiment of the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention make into
It is described in detail to one step, described embodiment is not construed as limitation of the present invention, and those of ordinary skill in the art are not having
All other embodiment obtained under the premise of creative work is made, shall fall within the protection scope of the present invention.
In the following description, it is related to " some embodiments ", which depict the subsets of all possible embodiments, but can
To understand, " some embodiments " can be the same subsets or different subsets of all possible embodiments, and can not conflict
In the case where be combined with each other.
In the following description, related term " first second " be only be the similar object of difference, do not represent needle
To the particular sorted of object, it is possible to understand that specific sequence or successively can be interchanged in ground, " first second " in the case where permission
Order, so that the embodiment of the present invention described herein can be implemented with the sequence other than illustrating or describing herein.
Unless otherwise defined, all technical and scientific terms used herein and belong to technical field of the invention
The normally understood meaning of technical staff is identical.Term used herein is intended merely to the purpose of the description embodiment of the present invention,
It is not intended to limit the present invention.
Before the embodiment of the present invention is further elaborated, to noun involved in the embodiment of the present invention and term
It is illustrated, noun involved in the embodiment of the present invention and term are suitable for following explanation.
1) it clicks result: referring to the existing result clicked or do not clicked on.
2) predict clicking rate: referring to the probability that the content-data predicted is clicked by user, it is usual with a function come
To prediction clicking rate.
3) weight parameter: refer to one group of parameter for determining clicking rate anticipation function, usually vector form.
4) sample characteristics: the interactive information of user data and content-data is quantified to obtain multiple features, and is used
Multi-C vector indicates that multi-C vector, that is, sample characteristics, user characteristics are similarly to indicate multiple features.
5) gradient: refer to most fast decline (or rising) direction of the objective function defined.
6) propagated forward: referring to feature being input to machine learning model, obtains the process of prediction result.
7) back-propagating: refer to according to the error between the corresponding actual result of feature and prediction result, and utilize definition
Good objective function updates the parameter in machine learning model, so that the process that subsequent obtained error is gradually reduced.
Inventor has found in the practice of the invention, in order to give user's specific aim recommendation data, it will usually answer
With the means of machine learning, i.e., according to historical data come training pattern, and by model to user whether click on content data into
The prediction of row clicking rate.In the scheme that the relevant technologies provide, usually according to a norm regular terms and two norm regular terms come structure
Objective function is built, so that regularization is carried out to model, still, since in clicking rate prediction scene, the dimension of user characteristics is logical
Often higher, some even can reach hundred million grades, if being trained according to aforesaid way to model, finally obtained weight parameter
Sparsity is poor, and the volume of model file is larger, is deployed to mobile terminal in the case where limited memory, such as in model
When, model may limited performance even be not available.
The embodiment of the present invention provide the training method of clicking rate prediction model a kind of, recommended method, device, electronic equipment and
Storage medium is able to ascend the sparsity of weight parameter, reduces model file volume, promotes the applicability of clicking rate prediction, under
Face illustrates the exemplary application of electronic equipment provided in an embodiment of the present invention.
It is one of the training system 100 of clicking rate prediction model provided in an embodiment of the present invention optional referring to Fig. 1, Fig. 1
Configuration diagram, to realize the training application for supporting clicking rate prediction model, terminal 400 (illustrates terminal
400-1 and terminal 400-2) it is connect by network 300 with database 500, sample characteristics and clicking rate are deposited in into database
500, terminal 400 can also be connect by network 300 with server 200, and network 300 can be wide area network or local area network, and or
Person is combination, and server 200 obtains data from database 500 by connection database 500.
Server 200 is used to obtain sample characteristics and sample characteristics correspond to the click result of user;According to the power of setting
Weight parameter initialization clicking rate prediction model;The sample characteristics are handled to obtain by the clicking rate prediction model pre-
Survey clicking rate;According to the error and zero norm regular terms building target between the click result and the prediction clicking rate
Function;By the objective function in the clicking rate prediction model error described in backpropagation, and in communication process more
The weight parameter of the new clicking rate prediction model.
Terminal 400 (terminal 400-1 and terminal 400-2 is illustrated in Fig. 1) is used to send number of users to server
According to;Server 200 is also used to obtain user data and at least two content-datas, wherein content-data can be by specific content
Platform generates;It combines each content-data with user data one-to-one correspondence, and feature is carried out to combined data and is mentioned
It takes, obtains user characteristics;The user characteristics are handled by the clicking rate prediction model, obtain prediction clicking rate;
Recommend the corresponding content-data of prediction clicking rate for meeting clicking rate value condition;Terminal 400 is (exemplary in graphical interfaces 410
Show graphical interfaces 410-1 and graphical interfaces 410-2) in illustratively with tabular form illustrate server 200 recommendation
Content-data.
It is worth noting that obtaining sample characteristics and sample characteristics pair from database 500 with server 200 in Fig. 1
It should be certainly not limited to this for the click result of user, server 200 can also be directly from data flow on the line of terminal 400
The middle click result for obtaining sample characteristics and sample characteristics corresponding to user.
Illustrate the exemplary application of electronic equipment provided in an embodiment of the present invention below.Electronic equipment may be embodied as taking down notes
This computer, tablet computer, desktop computer, set-top box, (for example, mobile phone, portable music player is a for mobile device
Personal digital assistant, specific messages equipment, portable gaming device) etc. various types of terminal devices, also may be embodied as servicing
Device.In the following, being illustrated so that electronic equipment is server as an example.
A referring to fig. 2, Fig. 2A are servers provided in an embodiment of the present invention 200 (for example, it may be server shown in FIG. 1
200) configuration diagram, server 200 shown in Fig. 2A include: at least one processor 210, memory 250, at least one
Network interface 220 and user interface 230.Various components in server 200 are coupled by bus system 240.It can manage
Solution, bus system 240 is for realizing the connection communication between these components.Bus system 240 is in addition to including data/address bus, also
Including power bus, control bus and status signal bus in addition.But for the sake of clear explanation, in fig. 2 all by various buses
It is designated as bus system 240.
Processor 210 can be a kind of IC chip, the processing capacity with signal, such as general processor, number
Word signal processor (DSP, Digital Signal Processor) either other programmable logic device, discrete gate or
Transistor logic, discrete hardware components etc., wherein general processor can be microprocessor or any conventional processing
Device etc..
User interface 230 include make it possible to present one or more output devices 231 of media content, including one or
Multiple loudspeakers and/or one or more visual display screens.User interface 230 further includes one or more input units 232, packet
Include the user interface component for facilitating user's input, for example keyboard, mouse, microphone, touch screen display screen, camera, other are defeated
Enter button and control.
Memory 250 can be it is removable, it is non-removable or combinations thereof.Illustrative hardware device includes that solid-state is deposited
Reservoir, hard disk drive, CD drive etc..Memory 250 optionally includes one geographically far from processor 210
A or multiple storage equipment.
Memory 250 includes volatile memory or nonvolatile memory, may also comprise volatile and non-volatile and deposits
Both reservoirs.Nonvolatile memory can be read-only memory (ROM, Read Only Memory), and volatile memory can
To be random access memory (RAM, Random Access Memory).The memory 250 of description of the embodiment of the present invention is intended to
Memory including any suitable type.
In some embodiments, memory 250 can storing data to support various operations, the example of these data includes
Program, module and data structure or its subset or superset, below exemplary illustration.
Operating system 251, including for handle various basic system services and execute hardware dependent tasks system program,
Such as ccf layer, core library layer, driving layer etc., for realizing various basic businesses and the hardware based task of processing;
Network communication module 252, for reaching other calculating via one or more (wired or wireless) network interfaces 220
Equipment, illustrative network interface 220 include: bluetooth, Wireless Fidelity (WiFi) and universal serial bus (USB,
Universal Serial Bus) etc.;
Module 253 is presented, for via one or more associated with user interface 230 output device 231 (for example,
Display screen, loudspeaker etc.) make it possible to present information (for example, for operating peripheral equipment and showing the user of content and information
Interface);
Input processing module 254, for one to one or more from one of one or more input units 232 or
Multiple user's inputs or interaction detect and translate input or interaction detected.
In some embodiments, the training device of clicking rate prediction model provided in an embodiment of the present invention can use software
Mode realizes that Fig. 2A shows the training device 2550 for the clicking rate prediction model being stored in memory 250, can be journey
The software of the forms such as sequence and plug-in unit, including following software module: module 25501, initialization module 25502, processing module are obtained
25503, construct module 25504 and update module 25505, these modules be in logic, therefore can according to the function of being realized
Arbitrarily to be combined or further be split.
In some embodiments, the recommendation apparatus provided in an embodiment of the present invention based on clicking rate prediction model can also be adopted
Realized with software mode, referring to fig. 2 B, Fig. 2 B other than the recommendation apparatus 2551 based on clicking rate prediction model shown, remaining
Part is identical as Fig. 2A, and details are not described herein again.For the recommendation based on clicking rate prediction model being stored in memory 250
Device 2551 can be the software of the forms such as program and plug-in unit, including following software module: data acquisition module 25511, spy
Levy extraction module 25512, clicking rate prediction module 25513 and recommending module 25514, these modules are therefore roots in logic
It can be combined arbitrarily according to the function of being realized or further split.The function of modules will be described hereinafter.
In further embodiments, the training device of clicking rate prediction model provided in an embodiment of the present invention and based on click
The recommendation apparatus of rate prediction model can be realized using hardware mode, as an example, clicking rate provided in an embodiment of the present invention is pre-
The training device for surveying model can be the processor using hardware decoding processor form, be programmed to perform implementation of the present invention
The training method for the clicking rate prediction model that example provides;Recommendation dress provided in an embodiment of the present invention based on clicking rate prediction model
The processor that can be using hardware decoding processor form is set, is programmed to perform provided in an embodiment of the present invention based on point
Hit the recommended method of rate prediction model.For example, the processor of hardware decoding processor form can use one or more application
Specific integrated circuit (ASIC, Application Specific Integrated Circuit), DSP, programmable logic device
(PLD, Programmable Logic Device), Complex Programmable Logic Devices (CPLD, Complex Programmable
Logic Device), field programmable gate array (FPGA, Field-Programmable Gate Array) or other electronics
Element.
The training method of clicking rate prediction model provided in an embodiment of the present invention and recommendation based on clicking rate prediction model
Method can be executed by above-mentioned server, can also be by terminal device (for example, it may be terminal 400-1 shown in FIG. 1 and end
Hold 400-2) it executes, or executed jointly by server and terminal device.
Below in conjunction with the exemplary application and structure of the electronic equipment being described above, illustrate to pass through insertion in electronic equipment
Clicking rate prediction model training device and realize the process of the training method of clicking rate prediction model.
It is an optional stream of the training method of clicking rate prediction model provided in an embodiment of the present invention referring to Fig. 3, Fig. 3
Journey schematic diagram, the step of showing in conjunction with Fig. 3, are illustrated.
In step 301, sample characteristics are obtained and sample characteristics correspond to the click result of user.
Sample characteristics refer to feature relevant to scene is clicked, for example, sample characteristics may include age of user, user
Gender, content-data title and time of content-data publication etc., content-data such as news or advertisement etc..Together, sample is obtained
Eigen correspond to the click of user as a result, it is worth noting that, click result and do not imply that probability, but give directions and hit or non-point
The existing result hit.According to the difference of practical application scene, can pointedly obtain sample characteristics and click as a result, for example, if
It determines the click effect after some content to be thrown to some page, then can obtain different user generation when browsing the page
Sample characteristics and corresponding click result;To determine the interested content of some user, then it can obtain the user and use visitor
When family end (when as used some advertisement client side), the sample characteristics of generation and corresponding click result.Due to sample characteristics and
Result is clicked usually to obtain together, therefore for ease of description, sample characteristics and click result are collectively referred to as sample, in this base
On plinth, when clicking the quantity that result is the sample clicked and the quantity that click result is the sample not clicked on is unbalanced, utilize
Ratio between the two adjusts the weight (sample_weight) of the sample of negligible amounts, to guarantee the quantity one of positive negative sample
It causes, prevents from adversely affecting subsequent training.For example, clicking the quantity that result is the sample clicked is 200, is clicked
As a result the quantity of the sample not click on is 100, then is clicking the case where weight that result is the sample clicked is defaulted as 1
Under, it is 2 that the weight that result is the sample not clicked on is clicked in adjustment.
In some embodiments, above-mentioned acquisition sample characteristics and sample characteristics pair can be realized in this way
It should be in the click result of user: obtaining sample content data and sample of users data, and obtain relative users in the sample
Hold the click result of data;Feature extraction is carried out to the sample content data and the sample of users data, obtains sample spy
Sign.
Since sample characteristics generally involve two aspects of user and content, therefore in embodiments of the present invention, sample can be obtained
This content-data and sample of users data carry out feature extraction to sample content data and sample of users data, obtain sample spy
Sign, the sample characteristics of extraction can be preset, and sample content data are data relevant to content to be clicked, sample of users number
According to being data relevant to user attribute itself.Together, relative users are obtained to the clicks of sample content data as a result, i.e. point
It hits or does not click on.It is worth noting that feature can also be handled after carrying out feature extraction, it such as can be according to only heat
(One-Hot) coding carries out characteristic processing, and one-hot coding encodes N number of state using N bit status register, each state
There is independent register-bit, and when any, wherein only one is effective, such as, this feature of age is come
It says, if corresponding, there are three possible values, respectively low, medium and high, then after one-hot coding, low corresponding numerical value is
100, in corresponding numerical value be 010, high corresponding numerical value is 001, in this way, improving the sparsity of data.
In some embodiments, above-mentioned acquisition sample characteristics and sample characteristics pair can be realized in this way
It should be in the click result of user: obtaining sample characteristics from database and sample characteristics correspond to the click result of user;Alternatively, from
Data flow obtains sample characteristics on line and sample characteristics correspond to the click result of user.
The embodiment of the present invention does not limit sample characteristics and clicks the data source of result, for example uses news visitor in user
When family end, advertisement client side or browser, the click of availability data inventory's sample storage eigen and sample characteristics corresponding to user
As a result, obtaining corresponding data from database when needing to carry out the training of clicking rate prediction model, off-line learning is carried out;?
Data flow sample characteristics and click can be obtained as a result, realizing on-line study directly from line.Data are improved through the above way
The flexibility of acquisition.
In step 302, clicking rate prediction model is initialized according to the weight parameter of setting.
Weight parameter refers to for carrying out clicking rate prediction, and can be updated in the training process of clicking rate prediction model
Parameter, be specifically addressed below.Here, initializing clicking rate prediction model according to the weight parameter of setting.
In some embodiments, it can realize in this way above-mentioned according to the initialization of the weight parameter of setting
Clicking rate prediction model: determine that numerical value is the dimension of non-zero in the sample characteristics;It is corresponding according to the dimension determined
Weight parameter initializes the clicking rate prediction model.
Due to sample characteristics be a multidimensional characteristic vectors, therefore initialization when, need to set in weight parameter with sample
The corresponding numerical value of each dimension of feature.In embodiments of the present invention, the process of initialization clicking rate prediction model can be carried out
Optimization is specifically determined that numerical value is the dimension of non-zero in sample characteristics, and is set corresponding with the dimension determined in weight parameter
Numerical value, thus according to the weight parameter of setting initialize clicking rate prediction model.
For example, sample characteristics are four dimensional feature vectors, and the four-dimension is respectively age of user, user's gender, content
The time of data header and content-data publication, and wherein there was only the corresponding numerical value of content-data title is non-zero, then sets power
Numerical value corresponding with this dimension of content-data title in weight parameter, so that initializing clicking rate according to weight parameter predicts mould
Type.By the above-mentioned means, reducing the duration of setting weight parameter, avoid handling incoherent dimension, causes resource unrestrained
Take.
In step 303, the sample characteristics are handled by the clicking rate prediction model to obtain prediction click
Rate.
Sample characteristics are handled by clicking rate prediction model, specifically, are set according in clicking rate prediction model
Weight parameter construct clicking rate anticipation function, and sample characteristics are handled according to clicking rate anticipation function, are predicted
Clicking rate, the step be propagated forward process, the prediction clicking rate of the clicking rate prediction model of initialization with actually obtain
Error can be had by clicking between result.
In step 304, according to the error and zero norm canonical between the click result and the prediction clicking rate
Item building objective function.
In training clicking rate prediction model, not only need that clicking rate prediction model is made to be fitted sample characteristics and click as far as possible
As a result, guaranteeing that training error is sufficiently small, training error refers to actual click result and predicts the error between clicking rate, together
When, it is also desirable to the phenomenon that guaranteeing that the test error of clicking rate prediction model is sufficiently small, avoiding the occurrence of over-fitting.Implement in the present invention
In example, regularization is carried out to clicking rate prediction model, with specific reference to the error clicked between result and prediction clicking rate, Yi Jiling
Norm regular terms parameter constructs objective function, so that it is simple as far as possible to constrain clicking rate prediction model.Zero norm regular terms refers to base
In the regular terms of L0 norm building, L0 norm refers to the number of element non-zero in vector, itself is non-convexification, uses zero
Norm regular terms constructs objective function, substantially occurs in expectation weight parameter more for zero value.
In step 305, by the objective function in the clicking rate prediction model error described in backpropagation, and
The weight parameter of the clicking rate prediction model is updated in communication process.
By objective function in clicking rate prediction model reverse propagated error, the process of reverse propagated error, substantially
It is the solution according to objective function, determines the numerical value of corresponding weight parameter, and updates click using the new numerical value of weight parameter
The value of weight parameter in rate prediction model.
In some embodiments, it can realize in this way and above-mentioned update the click in communication process
The weight parameter of rate prediction model: in communication process, updating in the clicking rate prediction model, with the dimension determined
Corresponding weight parameter.
On the basis of numerical value is the dimension of non-zero in determining sample characteristics, in communication process, clicking rate is only updated
The weight parameter corresponding with the dimension determined of prediction model, to promote update efficiency.
Implemented by above-mentioned example of the inventive embodiments for Fig. 3 it is found that the embodiment of the present invention is according to zero norm canonical
Item building objective function, to realize weight parameter by objective function reverse propagated error in clicking rate prediction model
It updates, improves the sparsity of weight parameter, while also improving the accuracy for generating clicking rate prediction model.
In some embodiments, referring to fig. 4, Fig. 4 is the training side of clicking rate prediction model provided in an embodiment of the present invention
Another optional flow diagram of method before step 302, can also set dimension and institute in step 401 in Fig. 4
State that sample characteristics are consistent, and each dimension only includes the weight parameter of single number.
Specifically, it is assumed that sample characteristics and corresponding click result are expressed as (xt,yt), xtIt is special for the sample of t-th of sample
Sign, ytIt is that t-th sample corresponds to the click of user as a result, t is the integer greater than 1, then it can be according to Logic Regression Models
(Logistic Regression, LR) determines the format of weight parameter, and Logic Regression Models are shown in following formula:
In above-mentioned formula, ptFor the corresponding prediction clicking rate of t-th of sample, corresponded to if result will be clicked in sample to click
Numerical value be set as 1, will click on result is that the numerical value not clicked on is set as 0, then ptValue range between 0 to 1, above-mentioned public affairs
W in formula is then weight parameter.For the weight parameter determined using Logic Regression Models, dimension is consistent with sample characteristics,
And each dimension only includes single number, can set the numerical value of each dimension according to practical application scene in setting.
In step 402, setting dimension is consistent with the sample characteristics, and each dimension includes the weight ginseng of K numerical value
Number, wherein the K is the integer greater than 1.
In addition to aforesaid way, for each of sample characteristics dimension, K auxiliary vector can be also introduced, accordingly, power
Weight parameter also includes K numerical value (i.e. component) in each dimension.
Specifically, the lattice of weight parameter can be determined according to Factorization machine (Factorization Machine, FM) model
Formula, Factorization machine model are shown in following formula:
Similarly, ptFor the corresponding prediction clicking rate of t-th of sample, v(k)It is then weight parameter, k is integer.For utilizing
The weight parameter that Factorization machine model determines can set each dimension of weight parameter in setting according to practical application scene
Including K specific value.
In Fig. 4, step 304 includes: to be constructed in step 3041 according to the click result and the prediction clicking rate
Loss function.
When constructing objective function, according to result and prediction clicking rate building loss function is clicked, as illustratively, it is assumed that
Weight parameter is w, sees following formula:
lt(w)=- yt log(pt)+(1-yt)log(1-pt)
Wherein, ltIt (w) is the corresponding loss function of t-th of sample.
Finally, the loss function of comprehensive T sample, can be obtained:
Wherein, T is the total quantity of sample, is greater than or equal to the integer of t, and L (w) is the corresponding average loss of T sample
Function.
In step 3042, according to the error between the click result and the prediction clicking rate, the loss is determined
The gradient of function.
In the sample characteristics and corresponding click for obtaining t-th of sample as a result, i.e. (xt,yt) after, training goal is to search out
One weight parameter, so that the value of L (w) is minimum, therefore in embodiments of the present invention, according between click result and prediction clicking rate
Error, determine the gradient of loss function.It is worth noting that determining loss function for the weight parameter of different-format
The mode of gradient is also different.
When weight parameter is determined according to Logic Regression Models, it is determined that the formula of the gradient of loss function is as follows:
gt=(pt-yt)xt
gtFor the gradient of the corresponding loss function of t-th of sample.
When weight parameter is determined according to Factorization machine model, since each dimension of sample characteristics corresponds to K
Auxiliary vector, it is determined that the formula of the gradient of loss function is as follows:
gt 0=(pt-yt)xt
gt k=(pt-yt)(vk·xt-diag(vk)diag(xt))xt
Wherein, diag function is for constructing diagonal matrix, gt 0And gt kIt is the ladder of the corresponding loss function of t-th of sample
Degree.
In step 3043, objective function is constructed according to the gradient of the loss function, zero norm regular terms and bound term,
Wherein, the bound term is used to constrain the variation degree of the weight parameter.
Other than the gradient of loss function and zero norm regular terms, objective function is constructed also according to bound term, this is about
Beam item is used to constrain the variation degree of weight parameter, prevents the variation of weight parameter excessive, after constructing objective function, then may be used
By the process of training clicking rate prediction model, the process solved to objective function is converted to.
In some embodiments, can realize in this way the above-mentioned gradient according to the loss function,
Zero norm regular terms and bound term construct objective function: about according to the gradient of the loss function, zero norm regular terms and first
Beam item constructs objective function, wherein first bound term is used to constrain the difference between the weight parameter and zero.
The formula of objective function is as follows:
Wherein, <, > are inner product of vectors operation,For the first bound term, ηtFor step-length, for determining that weight is joined
Several convergences and convergence rate, λ | | w | |0It is zero norm regular terms, λ therein is hyper parameter, can rule of thumb or actually be answered
It is configured with scene.
In some embodiments, can realize in this way the above-mentioned gradient according to the loss function,
Zero norm regular terms and bound term construct objective function: about according to the gradient of the loss function, zero norm regular terms and second
Beam item constructs objective function, wherein second bound term is used to constrain the institute before the updated weight parameter and update
State the difference between weight parameter.
The formula of objective function is as follows:
Wherein, <, > are inner product of vectors operation,For the second bound term, σsFor step-length, it is equally used for determining
Determine the convergence and convergence rate of weight parameter, λ | | w | |0It is zero norm regular terms, λ therein is hyper parameter, can be rule of thumb
Or practical application scene is configured.Two kinds of above-mentioned objective functions can be selected according to practical application scene.
In Fig. 4, after step 305, non-zero in the updated weight parameter can also be retained in step 403
Numerical value and corresponding dimension.
It is worth noting that in training clicking rate prediction model, it will usually obtain the corresponding sample characteristics of multiple samples
And click result and be trained, after updating weight parameter according to the sample characteristics of current sample and click result, according to update
Weight parameter afterwards handles the corresponding sample characteristics of next sample to obtain prediction clicking rate, according to the target of building
Function backpropagation clicks result and predicts that the error between clicking rate repeats the above process to update weight parameter again,
Until completing the update to weight parameter according to all samples, or stop manually.
For the weight parameter for completing to update, since it is multi-C vector, wherein the numerical value for being zero can't be to future position
The rate of hitting impacts, therefore in embodiments of the present invention, only retain the numerical value of non-zero and corresponding dimension in updated weight parameter
Degree, to reduce the file size of clicking rate prediction model, convenient for being deployed in memory or the limited equipment of capacity.
Implemented by above-mentioned example of the inventive embodiments for Fig. 4 it is found that the embodiment of the invention provides two kinds of formats
Weight parameter and two kinds of formats objective function, improve the flexibility of trained clicking rate prediction model.
The content of embodiment to facilitate the understanding of the present invention, hereafter with the weight parameter and two kinds of formats of two kinds of formats
Objective function between correspond combination, the pseudocode of the mode of obtained four kinds trained clicking rate prediction models, progress
It illustrates.It is worth noting that parameter alpha hereinafter is the hyper parameter for adjusting step, can be carried out according to practical application scene
Setting;There is the case where molecule is 0, is usually arranged as 1 for preventing in material calculation in hyper parameter β;Z and n is intermediate ginseng
Number, specially dimension and the consistent multi-C vector of sample characteristics, this will not be repeated here for derivation process.
(1) by using the objective function of the first bound term building, it is applied in Logic Regression Models, weight parameter is with w table
Show, the pseudocode of the process of training clicking rate prediction model is as follows:
(2) by using the objective function of the second bound term building, it is applied in Logic Regression Models, weight parameter is with w table
Show, the pseudocode of the process of training clicking rate prediction model is as follows:
(3) by using the objective function of the first bound term building, it is applied in Factorization machine model, weight parameter is with v(k)It indicates, the pseudocode of the process of training clicking rate prediction model is as follows:
(4) by using the objective function of the second bound term building, it is applied in Factorization machine model, weight parameter is with v(k)It indicates, the pseudocode of the process of training clicking rate prediction model is as follows:
It is worth noting that in above-mentioned four kinds of pseudocodes, the formula of Truncate () function are as follows:
Parameter in the Truncate () functional expression is only used for indicating different variables, does not have practical significance.
In practical application scene, one of the mode of above-mentioned four kinds trained clicking rate prediction models can be chosen, is carried out
Training, according to finally obtained intermediate parameters z and n, updates weight parameter using Truncate () function after the completion of training.
Below in conjunction with the exemplary application and structure of the electronic equipment being described above, illustrate to pass through insertion in electronic equipment
The recommendation apparatus based on clicking rate prediction model and realize the process of the recommended method based on clicking rate prediction model.
It is one of the recommended method provided in an embodiment of the present invention based on clicking rate prediction model optional referring to Fig. 5, Fig. 5
Flow diagram, the step of showing in conjunction with Fig. 5 is illustrated.
In step 501, user data and at least two content-datas are obtained.
For example, obtaining user data and at least two content-datas to be recommended.
In step 502, it combines each content-data with user data one-to-one correspondence, and to combined data
Feature extraction is carried out, user characteristics are obtained.
Since user characteristics are related to user and content, therefore combine each content-data with user data one-to-one correspondence, and
Feature extraction is carried out to combined data, finally obtains at least two user characteristics.For example, user data includes A, content
Data include B and C, then carry out feature extraction to A+B, obtain corresponding user characteristics;Feature extraction is carried out to A+C, is obtained pair
The user characteristics answered.
In step 503, the user characteristics are handled by the clicking rate prediction model, obtains prediction and clicks
Rate.
Clicking rate anticipation function is constructed by the weight parameter in clicking rate prediction model, and according to clicking rate anticipation function
Each user characteristics are handled, corresponding prediction clicking rate, clicking rate anticipation function LR model or FM for example above are obtained
Model, the p of prediction clicking rate, that is, abovet。
In step 504, according to the Generalization bounds of setting, recommend the prediction clicking rate for meeting clicking rate value condition corresponding
Content-data.
For example, according to different application scenarios different clicking rate value conditions can be set, the embodiment of the present invention is to click
Rate value condition is without limitation.Assuming that the numerical value of the click result to click is set as in training clicking rate prediction model
1,0 will be set as the numerical value for the click result not clicked on, then the prediction clicking rate obtained by step 503 is bigger, Yong Hudian
The probability for hitting corresponding content-data is bigger.In these cases, clicking rate value condition can be set as the maximum prediction of numerical value
Clicking rate, alternatively, numerical value is greater than the prediction clicking rate of the clicking rate threshold value of setting, clicking rate threshold value such as 0.7.When meeting clicking rate
It, can be according to prediction clicking rate from big to small when the quantity of the corresponding content-data of prediction clicking rate of value condition is at least two
Sequence, with list or other forms in terminal device recommendation data.
Implemented by above-mentioned example of the inventive embodiments for Fig. 5 it is found that the embodiment of the present invention by each content-data and
User data corresponds combination, carries out feature extraction to combined data and obtains user characteristics, passes through trained clicking rate
Prediction model handles user characteristics to obtain prediction clicking rate, and recommends corresponding content-data according to prediction clicking rate, is promoted
The accuracys of commending contents.
In the following, will illustrate exemplary application of the embodiment of the present invention in an actual application scenarios.
It is a kind of optional configuration diagram of clicking rate forecasting system provided in an embodiment of the present invention referring to Fig. 6, Fig. 6.
Based on Fig. 6, in training clicking rate prediction model, by characteristic extracting module, the user data and content generate to user is flat
The content-data that platform generates carries out feature extraction, obtains sample characteristics, meanwhile, obtain the click knot that sample characteristics correspond to user
Fruit.In order to make it easy to understand, sample characteristics and click result are collectively referred to as training sample, for training sample, can be deposited into
Database, and training sample is obtained from database by model training module, carry out off-line training;It can also be by training sample with line
The form of upper data flow is sent to model training module, in this way, while user uses client, model training module
On-line training is carried out based on data flow on line.Model training module is being instructed according to training sample training clicking rate prediction model
After the completion of white silk, the rate prediction model of will click on is sent to recommendation platform, it is worth noting that, recommend platform can be and is deployed in terminal
The platform of equipment is also possible to be deployed in the platform of server, and it is not limited in the embodiment of the present invention.Carrying out commending contents
When, recommend platform to obtain the user data that user sends and the content-data that content platform provides, to user data and content
Data carry out same feature extraction (not shown), user characteristics identical with sample characteristics format are obtained, further according to point
It hits rate prediction model user characteristics are handled to obtain prediction clicking rate, recommendation results is obtained according to prediction clicking rate, will be pushed away
It recommends result and is back to user side, complete commending contents process.It can for the above-mentioned example implementation of Fig. 6 by inventive embodiments
Know, be trained by the historical data to user, to recommend the content-data for being bonded user interest, improves commending contents
Accuracy.
It is a kind of optional configuration diagram of model training module provided in an embodiment of the present invention referring to Fig. 7, Fig. 7.?
In Fig. 7, model training module is a distributed training platform, including parameter server and at least two calculate nodes, in order to
It is easy to understand, is hereafter illustrated with step format.
Step (1): on database or line in data flow, carrying out in random every a batch, separated in batches training sample,
Including at least one training sample.A kind of batch mode is, according to setting quantity to training sample carry out it is average in batches, set number
Amount is the maximum quantity of calculate node support while the training sample of processing.For example, quantity is set as 100, trains sample
This includes 1000, then training sample can be divided into 10 batches, every batch of includes 100 training samples.
Step (2): each of Fig. 7 calculate node is used to handle the training sample of a batch, after the completion of in batches,
Each batch of training sample is flowed into corresponding calculate node, is handled.It is worth noting that after the completion of in batches, the quantity of batch
It is likely larger than the quantity of calculate node, for example has separated 10 batches, but calculate node only includes 3, for this kind of situation, by more batches
Secondary training sample is divided into several rounds, and in each round, the open ended batch of calculate node is only sent to calculate node,
For example, 4 rounds can be divided into for above-mentioned example, in each round in preceding 3 rounds, 3 batches of training samples are sent to
Corresponding calculate node, in the 4th round, remaining 1 batch of training sample is sent in model training module wherein one
A calculate node.During processing, calculate node requests intermediate parameters z and n to parameter server, and is getting intermediate ginseng
After number, gradient is individually calculated to each of the batch training sample got training sample, then to batch training
The corresponding all gradients of sample are averaging.It wherein, can be in advance in calculate node and parameter server for remaining hyper parameter
Unified setting can also be handed down to calculate node by parameter server after parameter server setting.
Step (3): on the basis of obtaining average gradient, above four kinds of trained clicking rate predictions are can be used in calculate node
Any one of mode of model, update intermediate parameters z and n, particularly, for application the second bound term mode, due to
It is related to weight parameter during updating intermediate parameters, therefore calculate node can first find out each training in a batch training sample
The corresponding weight parameter of sample, then the corresponding weight parameter of a batch training sample is averaging, according to average gradient, it is averaging
The intermediate parameters that weight parameter and parameter server afterwards issues update intermediate parameters z and n.In addition, for batch training sample
Only include the case where a training sample in this, does not then need to calculate average gradient.
Step (4): intermediate parameters of the calculate node by updated intermediate parameters and before updating carry out difference operation, with
znewAnd nnewUpdated intermediate parameters are indicated, with zoldAnd noldIt indicates the intermediate parameters before updating, then parameter difference can be obtained
Δ z=znew-zold, Δ n=nnew-nold.Parameter difference is uploaded to parameter server by each calculate node, and parameter server is connecing
It, will original intermediate ginseng in each difference and parameter server after the parameter difference of calculate node upload for harvesting a complete round
Number carries out accumulated process, i.e. znew=zold+Δz1+…+ΔzN, nnew=nold+Δn1+…+ΔnN, wherein Δ z1Refer to first
The parameter difference for the intermediate parameters z that calculate node uploads, N are the sum of calculate node.
Step (5): it repeats step (2)~step (4), until the training sample of all rounds is all trained to finish.
Step (6): parameter server updates the weight ginseng in clicking rate prediction model according to the intermediate parameters completed are updated
Number.
Implemented by above-mentioned example of the inventive embodiments for Fig. 7 it is found that the embodiment of the present invention passes through building at least two
A calculate node realizes the distribution training to training sample, effectively improves the efficiency of trained clicking rate prediction model.
It continues with and illustrates that the training device 2550 of clicking rate prediction model provided in an embodiment of the present invention is embodied as software
The exemplary structure of module as shown in Figure 2 A, is stored in the clicking rate prediction model of memory 250 in some embodiments
Software module in training device 2550 may include: to obtain module 25501, corresponding for obtaining sample characteristics and sample characteristics
In the click result of user;Initialization module 25502, for initializing clicking rate prediction model according to the weight parameter of setting;
Processing module 25503 obtains prediction clicking rate for being handled by the clicking rate prediction model the sample characteristics;
Module 25504 is constructed, for according to the error and zero norm regular terms between the click result and the prediction clicking rate
Construct objective function;Update module 25505, for passing through objective function backpropagation in the clicking rate prediction model
The error, and update in communication process the weight parameter of the clicking rate prediction model.
In some embodiments, module 25504 is constructed, is also used to: according to the click result and the prediction clicking rate
Construct loss function;According to the error between the click result and the prediction clicking rate, the ladder of the loss function is determined
Degree;Objective function is constructed according to the gradient of the loss function, zero norm regular terms and bound term, wherein the bound term is used
In the variation degree for constraining the weight parameter.
In some embodiments, described that mesh is constructed according to the gradient of the loss function, zero norm regular terms and bound term
Scalar functions, comprising: objective function is constructed according to the gradient of the loss function, zero norm regular terms and the first bound term, wherein
First bound term is used to constrain the difference between the weight parameter and zero;Alternatively, according to the gradient of the loss function,
Zero norm regular terms and the second bound term construct objective function, wherein second bound term is updated described for constraining
The difference between the weight parameter before weight parameter and update.
In some embodiments, initialization module 25502 are also used to: setting dimension is consistent with the sample characteristics, and every
A dimension only includes the weight parameter of single number;Alternatively, setting dimension is consistent with the sample characteristics, and each dimension includes
The weight parameter of K numerical value, wherein the K is the integer greater than 1.
In some embodiments, initialization module 25502 are also used to: determining that numerical value is non-zero in the sample characteristics
Dimension;According to the corresponding weight parameter of the dimension determined, the clicking rate prediction model is initialized.
In some embodiments, update module 25505 are also used to: in communication process, updating the clicking rate prediction mould
In type, weight parameter corresponding with the dimension determined.
In some embodiments, module 25501 is obtained, is also used to: obtaining sample content data and sample of users data, and
Relative users are obtained to the click result of the sample content data;To the sample content data and the sample of users data
Feature extraction is carried out, sample characteristics are obtained.
In some embodiments, the training device 2550 of clicking rate prediction model further include: model modification module, for protecting
Stay the numerical value of non-zero and corresponding dimension in the updated weight parameter.
In some embodiments, module 25501 is obtained, is also used to: obtaining sample characteristics and sample characteristics pair from database
It should be in the click result of user;Alternatively, data flow obtains sample characteristics from line and sample characteristics correspond to the click knot of user
Fruit.
The explanation recommendation apparatus 2551 provided in an embodiment of the present invention based on clicking rate prediction model is continued with to be embodied as
The exemplary structure of software module as shown in Figure 2 B, is stored in the pre- based on clicking rate of memory 250 in some embodiments
Survey model recommendation apparatus 2551 in software module may include: data acquisition module 25511, for obtain user data and
At least two content-datas;Characteristic extracting module 25512, for corresponding each content-data and the user data
Combination, and feature extraction is carried out to combined data, obtain user characteristics;Clicking rate prediction module 25513, for by described
Clicking rate prediction model handles the user characteristics, obtains prediction clicking rate;Recommending module 25514, it is full for recommending
The corresponding content-data of prediction clicking rate of sufficient clicking rate value condition.
The embodiment of the present invention provides a kind of storage medium for being stored with executable instruction, wherein it is stored with executable instruction,
When executable instruction is executed by processor, processor will be caused to execute method provided in an embodiment of the present invention, for example, such as Fig. 3
Or the training method of the clicking rate prediction model shown in Fig. 4, alternatively, the recommendation as shown in Figure 5 based on clicking rate prediction model
Method.
In some embodiments, storage medium can be FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface and deposit
The memories such as reservoir, CD or CD-ROM;Be also possible to include one of above-mentioned memory or any combination various equipment.
In some embodiments, executable instruction can use program, software, software module, the form of script or code,
By any form of programming language (including compiling or interpretative code, or declaratively or process programming language) write, and its
It can be disposed by arbitrary form, including be deployed as independent program or be deployed as module, component, subroutine or be suitble to
Calculate other units used in environment.
As an example, executable instruction can with but not necessarily correspond to the file in file system, can be stored in
A part of the file of other programs or data is saved, for example, being stored in hypertext markup language (HTML, Hyper Text
Markup Language) in one or more scripts in document, it is stored in the single file for being exclusively used in discussed program
In, alternatively, being stored in multiple coordinated files (for example, the file for storing one or more modules, subprogram or code section).
As an example, executable instruction can be deployed as executing in a calculating equipment, or it is being located at one place
Multiple calculating equipment on execute, or, be distributed in multiple places and by multiple calculating equipment of interconnection of telecommunication network
Upper execution.
In conclusion through the embodiment of the present invention, on the one hand, carried out based on zero norm regular terms to clicking rate prediction model
Regularization greatly improves the sparsity of weight parameter, reduces the file size of the clicking rate prediction model ultimately generated, mentions
The applicability of clicking rate prediction model is risen, and inventor is found through experiments that, using training points provided in an embodiment of the present invention
The mode for hitting rate prediction model, in the case where clicking rate predicts scene, the problem of there is no NP-hard, and in accuracy rate and model
Compression aspect, all has better performance than traditional approach;On the other hand, content-data is realized by clicking rate prediction model
Specific aim is recommended, and the accuracy of commending contents is improved.
The above, only the embodiment of the present invention, are not intended to limit the scope of the present invention.It is all in this hair
Made any modifications, equivalent replacements, and improvements etc. within bright spirit and scope, be all contained in protection scope of the present invention it
It is interior.
Claims (10)
1. a kind of training method of clicking rate prediction model characterized by comprising
It obtains sample characteristics and sample characteristics corresponds to the click result of user;
Clicking rate prediction model is initialized according to the weight parameter of setting;
The sample characteristics are handled by the clicking rate prediction model to obtain prediction clicking rate;
According to the error and zero norm regular terms building objective function between the click result and the prediction clicking rate;
By the objective function in the clicking rate prediction model error described in backpropagation, and updated in communication process
The weight parameter of the clicking rate prediction model.
2. the training method of clicking rate prediction model according to claim 1, which is characterized in that described according to the click
As a result the error between the prediction clicking rate and zero norm regular terms construct objective function, comprising:
Loss function is constructed according to the click result and the prediction clicking rate;
According to the error between the click result and the prediction clicking rate, the gradient of the loss function is determined;
Objective function is constructed according to the gradient of the loss function, zero norm regular terms and bound term, wherein the bound term is used
In the variation degree for constraining the weight parameter.
3. the training method of clicking rate prediction model according to claim 2, which is characterized in that described according to the loss
The gradient of function, zero norm regular terms and bound term construct objective function, comprising:
Objective function is constructed according to the gradient of the loss function, zero norm regular terms and the first bound term, wherein described first
Bound term is used to constrain the difference between the weight parameter and zero;Alternatively,
Objective function is constructed according to the gradient of the loss function, zero norm regular terms and the second bound term, wherein described second
Bound term is used to constrain the difference between the weight parameter before the updated weight parameter and update.
4. the training method of clicking rate prediction model according to claim 1, which is characterized in that the power according to setting
Before weight parameter initialization clicking rate prediction model, further includes:
It is consistent with the sample characteristics to set dimension, and each dimension only includes the weight parameter of single number;Alternatively,
It is consistent with the sample characteristics to set dimension, and each dimension includes the weight parameter of K numerical value, wherein the K is big
In 1 integer.
5. the training method of clicking rate prediction model according to claim 1, which is characterized in that the power according to setting
Weight parameter initialization clicking rate prediction model, comprising:
Determine that numerical value is the dimension of non-zero in the sample characteristics;
According to the corresponding weight parameter of the dimension determined, the clicking rate prediction model is initialized.
6. the training method of clicking rate prediction model according to claim 5, which is characterized in that described in communication process
Update the weight parameter of the clicking rate prediction model, comprising:
It in communication process, updates in the clicking rate prediction model, weight parameter corresponding with the dimension determined.
7. the training method of clicking rate prediction model according to claim 1, which is characterized in that the acquisition sample characteristics
And sample characteristics correspond to the click result of user, comprising:
Sample content data and sample of users data are obtained, and obtain relative users to the click knot of the sample content data
Fruit;
Feature extraction is carried out to the sample content data and the sample of users data, obtains sample characteristics.
8. a kind of recommended method based on the described in any item clicking rate prediction models of claim 1 to 7, which is characterized in that packet
It includes:
Obtain user data and at least two content-datas;
It combines each content-data with user data one-to-one correspondence, and feature extraction is carried out to combined data, obtain
To user characteristics;
The user characteristics are handled by the clicking rate prediction model, obtain prediction clicking rate;
Recommend the corresponding content-data of prediction clicking rate for meeting clicking rate value condition.
9. a kind of training device of clicking rate prediction model characterized by comprising
Module is obtained, corresponds to the click result of user for obtaining sample characteristics and sample characteristics;
Initialization module, for initializing clicking rate prediction model according to the weight parameter of setting;
Processing module obtains prediction clicking rate for being handled by the clicking rate prediction model the sample characteristics;
Module is constructed, for according to the error and zero norm regular terms between the click result and the prediction clicking rate
Construct objective function;
Update module, for by the objective function in the clicking rate prediction model error described in backpropagation, and
The weight parameter of the clicking rate prediction model is updated in communication process.
10. a kind of electronic equipment characterized by comprising
Memory, for storing executable instruction;
Processor when for executing the executable instruction stored in the memory, is realized described in any one of claim 1 to 7
Clicking rate prediction model training method or recommended method according to any one of claims 8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910673756.5A CN110378434A (en) | 2019-07-24 | 2019-07-24 | Training method, recommended method, device and the electronic equipment of clicking rate prediction model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910673756.5A CN110378434A (en) | 2019-07-24 | 2019-07-24 | Training method, recommended method, device and the electronic equipment of clicking rate prediction model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110378434A true CN110378434A (en) | 2019-10-25 |
Family
ID=68255780
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910673756.5A Pending CN110378434A (en) | 2019-07-24 | 2019-07-24 | Training method, recommended method, device and the electronic equipment of clicking rate prediction model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110378434A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110825966A (en) * | 2019-10-31 | 2020-02-21 | 广州市百果园信息技术有限公司 | Information recommendation method and device, recommendation server and storage medium |
CN111079006A (en) * | 2019-12-09 | 2020-04-28 | 腾讯科技(深圳)有限公司 | Message pushing method and device, electronic equipment and medium |
CN111177644A (en) * | 2019-12-20 | 2020-05-19 | 中国移动通信集团江苏有限公司 | Model parameter optimization method, device, equipment and storage medium |
CN111260142A (en) * | 2020-01-19 | 2020-06-09 | 杭州网易再顾科技有限公司 | Commodity index data prediction method and device, storage medium and electronic equipment |
CN111274473A (en) * | 2020-01-13 | 2020-06-12 | 腾讯科技(深圳)有限公司 | Training method and device for recommendation model based on artificial intelligence and storage medium |
CN112115300A (en) * | 2020-09-28 | 2020-12-22 | 北京奇艺世纪科技有限公司 | Text processing method and device, electronic equipment and readable storage medium |
CN112182382A (en) * | 2020-09-28 | 2021-01-05 | 上海宏路数据技术股份有限公司 | Data processing method, electronic device, and medium |
CN112182381A (en) * | 2020-09-28 | 2021-01-05 | 上海宏路数据技术股份有限公司 | Data processing method, electronic device, and medium |
CN112258285A (en) * | 2020-10-26 | 2021-01-22 | 北京沃东天骏信息技术有限公司 | Content recommendation method and device, equipment and storage medium |
CN112437349A (en) * | 2020-11-10 | 2021-03-02 | 杭州时趣信息技术有限公司 | Video stream recommendation method and related device |
CN112801708A (en) * | 2021-02-05 | 2021-05-14 | 通联数据股份公司 | Business income prediction model determination method and device, and prediction method and device |
CN112949864A (en) * | 2021-02-01 | 2021-06-11 | 北京三快在线科技有限公司 | Training method and device for pre-estimation model |
CN113065066A (en) * | 2021-03-31 | 2021-07-02 | 北京达佳互联信息技术有限公司 | Prediction method, prediction device, server and storage medium |
CN113657926A (en) * | 2021-07-28 | 2021-11-16 | 上海明略人工智能(集团)有限公司 | Advertisement effect prediction method, system, electronic device and storage medium |
WO2021233015A1 (en) * | 2020-05-19 | 2021-11-25 | 腾讯科技(深圳)有限公司 | Information processing method and apparatus, and computer readable storage medium |
CN114219346A (en) * | 2021-12-24 | 2022-03-22 | 江苏童能文化科技有限公司 | Method and system for improving service quality of network learning environment |
-
2019
- 2019-07-24 CN CN201910673756.5A patent/CN110378434A/en active Pending
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110825966B (en) * | 2019-10-31 | 2022-03-04 | 广州市百果园信息技术有限公司 | Information recommendation method and device, recommendation server and storage medium |
CN110825966A (en) * | 2019-10-31 | 2020-02-21 | 广州市百果园信息技术有限公司 | Information recommendation method and device, recommendation server and storage medium |
CN111079006B (en) * | 2019-12-09 | 2024-05-10 | 腾讯科技(深圳)有限公司 | Message pushing method and device, electronic equipment and medium |
CN111079006A (en) * | 2019-12-09 | 2020-04-28 | 腾讯科技(深圳)有限公司 | Message pushing method and device, electronic equipment and medium |
CN111177644A (en) * | 2019-12-20 | 2020-05-19 | 中国移动通信集团江苏有限公司 | Model parameter optimization method, device, equipment and storage medium |
CN111177644B (en) * | 2019-12-20 | 2023-10-27 | 中国移动通信集团江苏有限公司 | Model parameter optimization method, device, equipment and storage medium |
CN111274473A (en) * | 2020-01-13 | 2020-06-12 | 腾讯科技(深圳)有限公司 | Training method and device for recommendation model based on artificial intelligence and storage medium |
CN111274473B (en) * | 2020-01-13 | 2021-03-12 | 腾讯科技(深圳)有限公司 | Training method and device for recommendation model based on artificial intelligence and storage medium |
CN111260142B (en) * | 2020-01-19 | 2023-06-09 | 杭州网易再顾科技有限公司 | Commodity index data prediction method and device, storage medium and electronic equipment |
CN111260142A (en) * | 2020-01-19 | 2020-06-09 | 杭州网易再顾科技有限公司 | Commodity index data prediction method and device, storage medium and electronic equipment |
WO2021233015A1 (en) * | 2020-05-19 | 2021-11-25 | 腾讯科技(深圳)有限公司 | Information processing method and apparatus, and computer readable storage medium |
CN112182381A (en) * | 2020-09-28 | 2021-01-05 | 上海宏路数据技术股份有限公司 | Data processing method, electronic device, and medium |
CN112182382A (en) * | 2020-09-28 | 2021-01-05 | 上海宏路数据技术股份有限公司 | Data processing method, electronic device, and medium |
CN112115300A (en) * | 2020-09-28 | 2020-12-22 | 北京奇艺世纪科技有限公司 | Text processing method and device, electronic equipment and readable storage medium |
CN112182381B (en) * | 2020-09-28 | 2022-09-13 | 上海嗨普智能信息科技股份有限公司 | Data processing method, electronic device, and medium |
CN112258285A (en) * | 2020-10-26 | 2021-01-22 | 北京沃东天骏信息技术有限公司 | Content recommendation method and device, equipment and storage medium |
CN112437349A (en) * | 2020-11-10 | 2021-03-02 | 杭州时趣信息技术有限公司 | Video stream recommendation method and related device |
CN112949864A (en) * | 2021-02-01 | 2021-06-11 | 北京三快在线科技有限公司 | Training method and device for pre-estimation model |
CN112949864B (en) * | 2021-02-01 | 2022-04-22 | 海南两心科技有限公司 | Training method and device for pre-estimation model |
CN112801708B (en) * | 2021-02-05 | 2024-01-16 | 通联数据股份公司 | Business income prediction model determining method and device, and business income prediction model determining method and device |
CN112801708A (en) * | 2021-02-05 | 2021-05-14 | 通联数据股份公司 | Business income prediction model determination method and device, and prediction method and device |
CN113065066A (en) * | 2021-03-31 | 2021-07-02 | 北京达佳互联信息技术有限公司 | Prediction method, prediction device, server and storage medium |
CN113065066B (en) * | 2021-03-31 | 2024-05-07 | 北京达佳互联信息技术有限公司 | Prediction method, prediction device, server and storage medium |
CN113657926A (en) * | 2021-07-28 | 2021-11-16 | 上海明略人工智能(集团)有限公司 | Advertisement effect prediction method, system, electronic device and storage medium |
CN114219346A (en) * | 2021-12-24 | 2022-03-22 | 江苏童能文化科技有限公司 | Method and system for improving service quality of network learning environment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110378434A (en) | Training method, recommended method, device and the electronic equipment of clicking rate prediction model | |
US11531867B2 (en) | User behavior prediction method and apparatus, and behavior prediction model training method and apparatus | |
US20210256403A1 (en) | Recommendation method and apparatus | |
CN109902708A (en) | A kind of recommended models training method and relevant apparatus | |
CN106023015B (en) | Course learning path recommendation method and device | |
CN110781391A (en) | Information recommendation method, device, equipment and storage medium | |
CN110909182B (en) | Multimedia resource searching method, device, computer equipment and storage medium | |
CN111241311A (en) | Media information recommendation method and device, electronic equipment and storage medium | |
CN108780518A (en) | Web interface generation and test are carried out using artificial neural network | |
CN108664651A (en) | A kind of pattern recommends method, apparatus and storage medium | |
CN112765373B (en) | Resource recommendation method and device, electronic equipment and storage medium | |
CN111611488B (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
US11748452B2 (en) | Method for data processing by performing different non-linear combination processing | |
CN110008397A (en) | A kind of recommended models training method and device | |
CN113536097B (en) | Recommendation method and device based on automatic feature grouping | |
CN110516815A (en) | The characteristic processing method, apparatus and electronic equipment of artificial intelligence recommended models | |
CN112085615A (en) | Method and device for training graph neural network | |
CN110515986A (en) | A kind of processing method of social network diagram, device and storage medium | |
CN107807935B (en) | Using recommended method and device | |
CN108052670A (en) | A kind of recommendation method and device of camera special effect | |
CN108475256A (en) | Feature insertion is generated from homologous factors | |
CN110490274A (en) | Assess the method and device of alternative events | |
KR20230095796A (en) | Joint personalized search and recommendation with hypergraph convolutional networks | |
CN109446431A (en) | For the method, apparatus of information recommendation, medium and calculate equipment | |
CN104239421A (en) | Method and system for pushing application to terminal |
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 |