CN104361077B - The creation method and device of webpage scoring model - Google Patents
The creation method and device of webpage scoring model Download PDFInfo
- Publication number
- CN104361077B CN104361077B CN201410638360.4A CN201410638360A CN104361077B CN 104361077 B CN104361077 B CN 104361077B CN 201410638360 A CN201410638360 A CN 201410638360A CN 104361077 B CN104361077 B CN 104361077B
- Authority
- CN
- China
- Prior art keywords
- webpage
- adjusted
- feature
- scoring model
- loss function
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling techniques
Abstract
The embodiment of the invention discloses a kind of creation method and device of webpage scoring model.Methods described includes：Webpage training sample set is obtained, wherein webpage training sample set includes giving a mark with the characteristic vector and mark of multiple sample web pages under each query word at least one default query word；The mark marking of this webpage of various kinds and predetermined at least one webpage feature to be adjusted are concentrated according to webpage training sample, target loss function is generated；The characteristic vector of this webpage of various kinds is concentrated according to the target loss function and webpage training sample generated, webpage scoring model is created.Technical scheme provided in an embodiment of the present invention, it is possible to increase the accuracy rate of webpage sorting result, lifts the search experience of user.
Description
Technical field
The present embodiments relate to field of computer technology, more particularly to a kind of creation method and dress of webpage scoring model
Put.
Background technology
At present, search product can determine to return after the query word of user's input is received, first based on the query word
Then these related web pages are ranked up, finally by all webpages after sorting operation by the multiple related web pages returned
Link information constitutes a list, and user is presented to as search result.Webpage sorting it is whether accurate, to the standard of search result
True rate and user play vital effect to the satisfaction of search.The webpage more related to query word, its ranking should be got over
It is forward.
At present, search product is mostly to be pre-created a webpage scoring model, such as GBRank (Gradient
Boosting Rank, gradient lifting sequence) model, then according to the model to currently determined related to a certain query word
All webpages given a mark, and then according to the height of marking to these webpage sortings.Wherein, webpage scoring model is to use machine
The method of device study, concentrates the feature under the various dimensions of each webpage to be given a mark obtained from learning to substantial amounts of training sample
Rule.
But, because training sample set has certain limitation, it be may result in its learning process to webpage
Some feature learnings it is not abundant enough, so as to cause created webpage scoring model less reasonable so that the standard of ranking results
True rate is substantially reduced.For example, traditional GBRank models occur that the omission feature learning to webpage is not abundant enough, cause the spy
The effect levied in a model is not enough.The web pages under a certain query word are ranked up by the model, are easy to omit
The webpage that characteristic value is smaller, correlation is poor has been discharged to omission, and characteristic value is larger, before good relationship webpage, and this can be serious
Have influence on the search experience of user.
The content of the invention
The embodiment of the present invention provides a kind of creation method and device of webpage scoring model, can improve webpage sorting knot
The accuracy rate of fruit, lifts the search experience of user.
In a first aspect, the embodiments of the invention provide a kind of creation method of webpage scoring model, this method includes：
Obtain webpage training sample set, wherein the webpage training sample set include with least one default query word
The characteristic vector and mark of multiple sample web pages under each query word are given a mark；
The mark marking of this webpage of various kinds and at least one predetermined net are concentrated according to the webpage training sample
Page feature to be adjusted, generates target loss function；
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function generated and the webpage training sample,
Create webpage scoring model.
Second aspect, the embodiment of the present invention additionally provides a kind of creating device of webpage scoring model, and the device includes：
Webpage training sample acquiring unit, for obtaining webpage training sample set, wherein the webpage training sample set bag
The characteristic vector and mark included with multiple sample web pages under each query word at least one default query word is given a mark；
Target loss function generation unit, for being concentrated the mark of this webpage of various kinds to give a mark according to the webpage training sample
And predetermined at least one webpage feature to be adjusted, generate target loss function；
Webpage scoring model creating unit, for according to the target loss function and the webpage training sample generated
The characteristic vector of each sample web page is concentrated, webpage scoring model is created.
Technical scheme provided in an embodiment of the present invention, can predefine multiple webpages feature to be adjusted, then in combination with
Identified multiple webpages feature to be adjusted and webpage training sample concentrate the mark marking of this webpage of various kinds, are damaged to generate target
Function is lost, and then concentrates the characteristic vector of this webpage of various kinds to create webpage according to the target loss function and webpage training sample and is beaten
Sub-model, so as to adjust effect of the webpage feature to be adjusted in webpage scoring model, can overcome traditional webpage to beat
Sub-model occurs not reasonable to the learning process of webpage some features, and then causes these features in webpage scoring model
Effect it is not enough or the drawbacks of act on excessive.Therefore, using the webpage scoring model created in the embodiment of the present invention, to input
Either query word under multiple webpages given a mark, then according to the marking result carry out webpage sorting, it is possible to increase webpage
The accuracy rate of ranking results, lifts the search experience of user.
Brief description of the drawings
Figure 1A is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention one is provided；
Figure 1B is answering for webpage sorting used in the creation method for the webpage scoring model that the embodiment of the present invention one is provided
Use scene；
Fig. 2 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention two is provided；
Fig. 3 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention three is provided；
Fig. 4 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention four is provided
Fig. 5 show a kind of structural representation of the creating device of webpage scoring model of the offer of the embodiment of the present invention five.
Embodiment
The present invention is described in further detail with reference to the accompanying drawings and examples.It is understood that this place is retouched
The specific embodiment stated is used only for explaining the present invention, rather than limitation of the invention.It also should be noted that, in order to just
Part related to the present invention rather than entire infrastructure are illustrate only in description, accompanying drawing.
Embodiment one
Figure 1A is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention one is provided.Utilize
Multiple webpages under the either query word that is inputted can be given a mark by the webpage scoring model that the present embodiment is created, search
Product can be ranked up according to marking result to the multiple webpage, and then by the chain corresponding to multiple webpages after sequence
Connect information and be presented to user.Figure 1B is webpage used in the creation method for the webpage scoring model that the embodiment of the present invention one is provided
The application scenarios of sequence.Referring to Figure 1B, the basic procedure of webpage sorting can be divided into training and two processes of prediction, line on line under line
The input of lower training is webpage training sample set, and by one webpage scoring model of learning system output, wherein learning system is adopted
The creation method of the webpage scoring model provided with the present embodiment.The input predicted on line is not arranging under any one query word
Multiple webpages of sequence, it is sequential according to the size of marking output one after webpage scoring model is given a mark to these webpages
Web page listings.
Referring to Figure 1A, the method that the present embodiment is provided can be performed by the creating device of webpage scoring model, specific bag
Include following operation：
Operation 110, acquisition webpage training sample set, wherein webpage training sample set include and at least one default inquiry
The characteristic vector and mark of multiple sample web pages under the query word of each in word are given a mark.
In the present embodiment, to create webpage scoring model, webpage training sample set need to be obtained.In the sample set, bag
Characteristic vector containing multiple sample web pages under different query words and mark marking.The characteristic vector of any sample web page by
Corresponding value is constituted the sample web page in multiple web page characteristics of setting respectively.Wherein, the multiple web page characteristics can be
From multiple angle initializations such as web page contents, web page interrogation word and web page interlinkages, for example, it may include the number of keyword in webpage
Measure feature, the average length feature of word, word frequency against document frequency feature, omit feature, webpage level characteristics and webpage degree of belief
Any number of combination in index characteristic etc..In embodiments of the present invention, it is characterized in be used according to corresponding with the webpage that webpage, which is omitted,
The query word of family input, and the target search word used when searching the webpage, the proportionate relationship between both are obtained
Arrive.Wherein, target search word is the word obtained after being operated to the omission that the query word that user inputs carries out part lexical item.Webpage
Omit feature smaller, illustrate that the lexical item dispensed is more, the possibility that escape occurs for query word is also higher.Each sample web page
Mark marking, can be beforehand through manual or automatic mode, according to sample web page content and/or correlation query word,
Marking result to made by the sample web page.
For example, acquired webpage training sample set includes and each query word difference corresponding m in n query word
Sample web page, is specifically included：
Q1,For by the characteristic vector of m sample web page under the 1st query word
The set of composition；Q1,To be beaten by the mark of m sample web page under the 1st query word
The set being grouped；
Q2,For by the characteristic vector of m sample web page under the 2nd query word
The set of composition；Q2,To be beaten by the mark of m sample web page under the 2nd query word
The set being grouped；
… …
Qn,For by the characteristic vector of m sample web page under n-th of query word
The set of composition；Qn,To be beaten by the mark of m sample web page under the 1st query word
Divide set.
Operation 120, the mark marking and predetermined at least one according to this webpage of webpage training sample concentration various kinds
Individual webpage feature to be adjusted, generates target loss function.
Operation 130, the feature for concentrating according to the target loss function and webpage training sample that are generated this webpage of various kinds
Vector, creates webpage scoring model.
In the present embodiment, it can concentrate the mark marking generation of this webpage of various kinds is original to damage previously according to webpage training sample
Function is lost, then according to machine learning algorithm, the feature to each sample web page in webpage sample set learns, and makes it towards most
Optimize the angle development of loss function, and then obtain an original web page scoring model.
Because webpage training sample set has certain limitation, it is likely to result in its learning process to sample
Some feature learnings of webpage are not abundant enough, so as to cause created original web page scoring model less reasonable, more such as
Web page characteristics act on too strong, other web page characteristics institute in original web page scoring model played in original web page scoring model
Role Shortcomings.
Therefore, too strong web page characteristics (webpage feature to be strengthened) and effect will can be acted in original web page scoring model
Not enough web page characteristics (webpage feature to be slackened), as at least one webpage feature to be adjusted, then regenerate one accordingly
Individual target loss function, new webpage scoring model is created using the target loss function using above-mentioned machine learning algorithm.This
Sample can put on the effect of webpage feature to be adjusted in target loss function, by influenceing target loss function to adjust it
Effect in webpage scoring model, so as to overcome above-mentioned drawback.Wherein, webpage feature to be adjusted can be above-mentioned setting
Multiple web page characteristics in arbitrary characteristics or other web page characteristics for being not comprised in sample web page characteristic vector.
After it is determined that finishing multiple webpages feature to be adjusted, the mark of this webpage of various kinds is concentrated according to webpage training sample
Marking and webpage feature to be adjusted generate target loss function, may particularly include：It will be used to characterize webpage feature to be adjusted
Acting factor be added in the decision factor of primary loss function, to update primary loss function, obtain added with effect because
The target loss function of son.Wherein, decision factor concentrates identical look into be used to weigh webpage training sample in primary loss function
Ask the factor of difference degree between different web pages under word.If webpage is to be adjusted to be characterized as webpage feature to be strengthened, can be direct
Its acting factor is made an addition in decision factor in the way of addition, to strengthen webpage feature to be strengthened in webpage marking mould
Effect played in type；If webpage is to be adjusted to be characterized as webpage feature to be slackened, directly it can be made in the way of subtracting each other
Made an addition to the factor in decision factor, to slacken effect of the webpage feature to be slackened played in webpage scoring model.
It is, of course, also possible to come by other means according to webpage training sample concentrate various kinds this webpage mark marking and
Webpage feature to be adjusted generation target loss function, as long as can to strengthen webpage accordingly to be added for the target loss function generated
Strong feature and slacken effect of the webpage feature to be slackened played in webpage scoring model.
It should be noted that the present embodiment can need not also previously generate original web page scoring model, and directly obtain in advance
Multiple webpages feature to be adjusted of setting, the mark for then concentrating this webpage of various kinds according to webpage training sample is given a mark and obtained
The multiple webpages feature to be adjusted taken, generates target loss function；And then according to target loss function and webpage training sample
The characteristic vector of each sample web page is concentrated, webpage scoring model is created.Wherein target loss function should include：For weighing webpage
Training sample concentrates the factor of difference degree between different web pages under same queries word, and for characterizing webpage feature to be adjusted
Acting factor.
The present embodiment provide technical scheme, multiple webpages feature to be adjusted can be predefined, then in combination with really
Fixed multiple webpages feature to be adjusted and webpage training sample concentrate the mark marking of this webpage of various kinds, to generate target loss letter
Number, and then concentrate the characteristic vector of this webpage of various kinds to create webpage marking mould according to the target loss function and webpage training sample
Type, so as to improve effect degree of the webpage feature to be adjusted in webpage scoring model, can overcome traditional webpage to beat
Sub-model occurs not reasonable to the learning process of webpage some features, and then causes these features in webpage scoring model
Effect it is not enough or the drawbacks of act on excessive.Therefore, using the webpage scoring model created in the present embodiment, input is appointed
Multiple webpages under one query word are given a mark, and then carry out webpage sorting according to the marking result, it is possible to increase webpage sorting
As a result accuracy rate, lifts the search experience of user.
Embodiment two
Fig. 2 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention two is provided.This reality
Example is applied on the basis of above-described embodiment one, operation 120 is optimized.Referring to Fig. 2, the method that the present embodiment is provided specifically is wrapped
Include following operation：
Operation 210, acquisition webpage training sample set, wherein webpage training sample set include and at least one default inquiry
The characteristic vector and mark of multiple sample web pages under the query word of each in word are given a mark.
Operation 220, acquisition concentrate the primary loss letter that the mark marking of this webpage of various kinds is obtained according to webpage training sample
Number.
In the present embodiment, the mark marking generation primary loss of this webpage of various kinds is concentrated previously according to webpage training sample
Function, then according to machine learning algorithm, the feature to each sample web page in webpage sample set learns, and makes it towards optimal
Change the angle development of loss function, and then obtain an original web page scoring model.Wherein, primary loss function can be any
The loss function used during for creating webpage scoring model, as long as being included in the loss function is used to weigh webpage training sample
Under this concentration same queries word between different web pages difference degree decision factor.For example, primary loss function is intersection
Entropy loss function, or Hinge loss loss functions.
Operate 230, determine that being used in primary loss function weighs webpage training sample and concentrate different under same queries word
The decision factor of difference degree between webpage.
Operate 240, be directed to each webpage feature to be adjusted in predetermined at least one webpage feature to be adjusted respectively, will
Acting factor for characterizing webpage feature to be adjusted is added in decision factor, to generate target loss function.
In the present embodiment, original web page scoring model can be analyzed by manual or automatic mode, it is determined that needing
Strengthen or slacken its web page characteristics acted on played in original web page scoring model, be used as webpage feature to be adjusted.Wherein, net
Page feature to be adjusted can be arbitrary characteristics in multiple web page characteristics of above-mentioned setting or be not comprised in sample web page
Other web page characteristics in characteristic vector.Certainly, original web page scoring model can also be used under some query words of input
Each webpage is given a mark, then by the risk given a mark present in interpretation of result original web page scoring model, it is determined that need to strengthen or
Its web page characteristics acted on played in original web page scoring model is slackened, webpage feature to be adjusted is used as.
If webpage is to be adjusted to be characterized as webpage feature to be strengthened, it will directly can be used to characterize this feature in the way of addition
Acting factor make an addition in the decision factor of primary loss function, to strengthen this feature played in webpage scoring model
Effect；If webpage is to be adjusted to be characterized as webpage feature to be slackened, it will directly can be used to characterize this feature in the way of subtracting each other
Acting factor make an addition in the decision factor of primary loss function, to slacken this feature played in webpage scoring model
Effect.
The present embodiment it is a kind of preferred embodiment in, by the acting factor for characterizing webpage feature to be adjusted
It is added in decision factor, including：Corresponding function coefficient will be multiplied by for the acting factor for characterizing webpage feature to be adjusted
Afterwards, it is added in the decision factor of primary loss function.
Wherein, if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened, corresponding effect system
Number is positive coefficient；If the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be weakened, corresponding function coefficient
For negative coefficient.In the present embodiment, function coefficient can be redefined for fixed value based on experience value；Also first it can set at random
A fixed initial value, then come continuous optimization function coefficient by way of iteration, it implements process reference can be made to this hair
Technical scheme described in bright embodiment three.
Above-mentioned embodiment can be by setting function coefficient, flexibly to control its corresponding webpage feature to be adjusted
Effect played in webpage scoring model, can be significantly more efficient gram in the way of being directly added or subtract each other
Take the drawbacks of too strong web page characteristics effect or not enough effect.
For example, being used in primary loss function weighs webpage training sample and concentrates different web pages under same queries word
Between the decision factor of difference degree can be：H(xqi,xqj)=- h (xqi)+h(xqj)；
By addition operation after decision factor can be：
Wherein, q is the integer more than or equal to 1 and less than or equal to Q, and Q is all query word numbers at least one query word；
h(xqi) and h (xqj) be webpage scoring model predicted value, xqiAnd xqjRespectively q at least one query word
Different web pages under individual query word；
T is more than or equal to 1 and less than or equal to the integer of at least one webpage feature total number to be adjusted；
reduce_difft* label_diff waits to adjust for characterizing t-th of webpage at least one webpage feature to be adjusted
The acting factor of whole feature, εtFor with for characterizing t-th of webpage feature to be adjusted at least one webpage feature to be adjusted
The corresponding function coefficient of acting factor；
reduce_difft=reducet,qi-reducet,qj, it is xqiT-th of webpage characteristic value reduce to be adjustedt,qi
With xqjT-th of webpage characteristic value reduce to be adjustedt,qjBetween difference；
Label_diff=labelqi-labelqjIt is xqiMark marking labelqiWith xqjMark marking labelqjIt
Difference.
For example, at least one webpage feature to be adjusted, which is a webpage, omits feature (webpage feature to be strengthened), it is used for
It is 1 to characterize the webpage and omit the corresponding function coefficient of acting factor of feature, then the decision factor after addition operation is：
Hnew(xqi,xqj)=- h (xqi)+h(xqj)+reduce_diff*label_diff
Wherein, reduce_diff is xqiWebpage omit characteristic value and xqjWebpage omit characteristic value between difference.
Specifically, if primary loss function is cross entropy loss function L (h), its corresponding mathematic(al) representation is：
The target loss function L then generatednew(h) it is：
Wherein, p (xqi＞ xqj) it is x in primary loss functionqiCompare xqjPoint high probable value of mark；
τ is default parameter value；pnew(xqi＞ xqj) it is x in target loss functionqiCompare xqjPoint high probable value of mark；
If primary loss function is Hinge loss loss functions R (h, τ), its corresponding mathematic(al) representation is：
Then target loss function Rnew(h, τ) is：
Wherein, τ is the predicted value of webpage scoring model, and λ is default parameter.
Operation 250, the feature for concentrating according to the target loss function and webpage training sample that are generated this webpage of various kinds
Vector, creates webpage scoring model.
In the present embodiment, after generation target loss function, it should use and generation original web page scoring model identical
Machine learning algorithm, come concentrated according to the target loss function and webpage training sample that are generated the feature of this webpage of various kinds to
Amount, creates a webpage scoring model for being different from original web page scoring model.
The technical scheme that the present embodiment is provided, can previously generate original object loss function and the marking of corresponding original web page
Model, and determine to act on the not enough or excessive web page characteristics of effect in original web page scoring model by the analysis means of setting,
As webpage feature to be adjusted, then primary loss function will be added to for the acting factor for characterizing webpage feature to be adjusted
In decision factor, target loss function is obtained to be subject to adjustment to primary loss function, so according to the target loss function and
Webpage training sample concentrates the characteristic vector of various kinds this webpage, based on original web page scoring model identical training algorithm, weight
A webpage scoring model is newly created, so as to improve effect degree of the webpage feature to be adjusted in webpage scoring model,
Traditional webpage scoring model can be overcome to occur not reasonable to the learning process of webpage some features, and then cause these
The drawbacks of effect of the feature in webpage scoring model is not enough or acts on excessive.Therefore, based on the net created in the present embodiment
Multiple webpages under the either query word of input are given a mark and sorted, it is possible to increase webpage sorting result by page scoring model
Accuracy rate, lifted user search experience.
Embodiment three
Fig. 3 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention three is provided.This reality
Example is applied on the basis of above-described embodiment one and embodiment two, further increase renewal target loss function in function coefficient with
And create the operation of new webpage scoring model.Referring to Fig. 3, the method that the present embodiment is provided specifically includes following operation：
Operation 310, acquisition webpage training sample set, wherein webpage training sample set include and at least one default inquiry
The characteristic vector and mark of multiple sample web pages under the query word of each in word are given a mark.
Operation 320, acquisition concentrate the primary loss letter that the mark marking of this webpage of various kinds is obtained according to webpage training sample
Number.
Operate 330, determine that being used in primary loss function weighs webpage training sample and concentrate different under same queries word
The decision factor of difference degree between webpage.
Operate 340, be directed to each webpage feature to be adjusted in predetermined at least one webpage feature to be adjusted respectively, will
Acting factor for characterizing webpage feature to be adjusted is multiplied by after corresponding function coefficient, is added in decision factor, with life
Into target loss function.
In the present embodiment, if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened, accordingly
Function coefficient be positive coefficient, 0 value can be set greater than at random；It is characterized as if the webpage of acting factor sign is to be adjusted
Webpage feature to be weakened, then corresponding function coefficient is negative coefficient, and 0 value can be set smaller than at random.
Operation 350, the feature for concentrating according to the target loss function and webpage training sample that are generated this webpage of various kinds
Vector, creates webpage scoring model.
Operation 360, the function coefficient in setting Policy Updates target loss function.
In the present embodiment, performing after operation 350 obtains webpage scoring model, can be by manual or automatic
Mode, is analyzed the webpage scoring model, using the webpage scoring model to each webpage under some query words of input
Given a mark, then determine that each webpage is waited to adjust at least one webpage feature to be adjusted according to analysis result and/or marking result
Active state of the whole feature in the webpage scoring model.
And then, obtain and determine result, according to the function coefficient in setting Policy Updates target loss function.Wherein, it is described
Setting rule includes：If getting any webpage feature to be adjusted at least one webpage feature to be adjusted is in effect deficiency
State, then increase the corresponding function coefficient of target elements for characterizing any webpage feature to be adjusted；If got at least
Any webpage feature to be adjusted, which is in, in one webpage feature to be adjusted acts on larger state, then reduces for characterizing any webpage
The corresponding function coefficient of target elements of feature to be adjusted.
Operate 370, various kinds Home Network is concentrated according to the target loss function and webpage training sample after updating operation
The characteristic vector of page, creates new webpage scoring model.
In the present embodiment, after the operation 370 i.e. establishment that is finished has new webpage scoring model, it can obtain again
Active state of each webpage feature to be adjusted in the new webpage scoring model at least one webpage feature to be adjusted, so
Afterwards based on the result obtained again, continued to update the function coefficient in target loss function according to setting rule, and then continue root
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function and webpage training sample after updating operation, is created again
Build a new webpage scoring model.Such loop iteration goes down.
If still located specifically, getting some webpage feature to be strengthened at least one webpage feature to be adjusted again
In the not enough state of effect, then it can continue the corresponding function coefficient of target elements that increase is used to characterize webpage feature to be strengthened；
, whereas if it is larger in acting on to get some webpage feature to be strengthened at least one webpage feature to be adjusted again
State, then can reduce the corresponding function coefficient of target elements for characterizing webpage feature to be strengthened.
Accordingly, located if getting some webpage feature to be slackened at least one webpage feature to be adjusted again
In the not enough state of effect, then it can increase the corresponding function coefficient of target elements for characterizing webpage feature to be strengthened；If
Some webpage feature to be slackened at least one webpage feature to be adjusted is got again to be still within acting on larger state, then may be used
Continue to reduce the corresponding function coefficient of target elements for being used for characterizing webpage feature to be strengthened.
Wherein, the step-length that each iteration increases or reduced can be redefined for fixed value, also with iterations
Increase adjusted in real time, for example iterations is bigger, and the step-length for increasing or reducing is smaller, and the present embodiment is not made to this
Limit.
The technical scheme that the present embodiment is provided, presets corresponding for the acting factor that characterizes webpage feature to be adjusted
Function coefficient, then constantly optimizes the function coefficient by way of iteration, so that the net finally created
The active state of webpage feature to be adjusted tends to the perfect condition of setting in page scoring model.Carried out using the webpage scoring model
The marking of multiple webpages under arbitary inquiry word, can greatly improve the accuracy rate of marking, and then cause search engine to multiple
The sequence of webpage is more reasonable, lifts user's search experience.
On the basis of above-mentioned technical proposal, in a kind of embodiment of the embodiment of the present invention, original web page is beaten
Sub-model is GBRank models, and the spy of this webpage of various kinds is concentrated according to the target loss function and webpage training sample generated
Vector is levied, webpage scoring model is created, may particularly include：
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function of generation and webpage training sample, based on ladder
Degree lifting sort algorithm creates at least one decision tree；At least one decision tree created is utilized to create webpage scoring model.
Example IV
Fig. 4 is a kind of schematic flow sheet of the creation method for webpage scoring model that the embodiment of the present invention four is provided.This reality
Apply that example can there is provided a kind of preferred embodiment based on above-described embodiment.Referring to Fig. 4, the method that the present embodiment is provided specifically is wrapped
Include following operation：
Operation 400, acquisition webpage training sample set, wherein webpage training sample set include and at least one default inquiry
The characteristic vector and mark of multiple sample web pages under the query word of each in word are given a mark.
Operate 410, concentrated the mark marking of this webpage of various kinds to obtain primary loss function according to webpage training sample, be based on
The primary loss function concentrates the characteristic vector of this webpage of various kinds to be trained webpage training sample using GBRank algorithms, with
Create original GBRank models.
Wherein, primary loss function is cross entropy loss function.
Cross entropy loss function is to h (xqi) gradientIt is
And haveSet up.
Work as Sqi,qjWhen=1, above-mentioned gradient can be expressed as：
In the present embodiment, the characteristic vector for this webpage of various kinds being concentrated to webpage training sample using GBRank algorithms is carried out
Training, to create GBRank models, including：
1st, each predicted value { h in initialization webpage scoring model0(xqw), q is whole more than or equal to 1 and less than or equal to Q
Number, Q is all query word numbers at least one query word；xqwFor w-th under q-th of query word at least one query word
Webpage；
2nd, k is set successively from 1 value to K, and following steps are performed respectively：
Calculate negative gradient
According to { rk,qwCreate a decision tree gk；
Update hk(xqw)=hk-1(xqw)+ηgk(xqw)。
Operate 420, according to the original GBRank models created, determine at least one webpage feature to be adjusted.
Operation 430, at least one webpage feature to be adjusted determined by, judge whether to change primary loss function.
If it is determined that any webpage feature to be adjusted is not present under the original GBRank models created, then judging need not
Primary loss function is changed, otherwise, judgement need to change primary loss function.
If it is, performing operation 450, operation 440 is otherwise performed.
Operate 440, directly exported original GBRank models as final webpage scoring model, terminated.
Operate 450, determine whether first time iteration；
Iterations initial value is pre-set to 0 in the present embodiment.If it is 0 to obtain iterations initial value,
It is judged as first time iteration, is otherwise non-first time iteration.
If it is, performing operation 460, operation 470 is otherwise performed.
Operation 460, generation include acting factor for characterizing webpage feature to be adjusted and its respective action coefficient
Target loss function；Wherein, it is corresponding to make if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened
Positive coefficient with coefficient to set at random；If the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be weakened,
Corresponding function coefficient is the negative coefficient set at random.Perform operation 490.
Operation 460 is specifically included in the present embodiment：
Primary loss function is obtained, determines that being used in primary loss function weighs webpage training sample and concentrate same queries
Under word between different web pages difference degree decision factor；
Each webpage feature to be adjusted at least one webpage feature to be adjusted is directed to respectively, will be used to characterizing the webpage and be waited to adjust
The acting factor of whole feature is multiplied by after corresponding function coefficient, is added in decision factor, to generate target loss function.
Description in embodiment two to operating 220-240 can be found in the specific detailed description for operating 460, will not be repeated here.
Operate 470, obtain effect of at least one webpage feature to be adjusted in the webpage scoring model currently created
State；
If operation 480, any webpage characteristic action state to be adjusted are the not enough state of effect or the larger state of effect,
According to the function coefficient in setting Policy Updates target loss function.
Wherein, setting rule includes：Increase is used to characterize in the not enough shape of effect at least one webpage feature to be adjusted
The corresponding function coefficient of target elements of the webpage feature to be adjusted of state；Being reduced at least a webpage feature to be adjusted is used to characterize
The corresponding function coefficient of target elements in the webpage feature to be adjusted for acting on larger state.
The spy of this webpage of various kinds is concentrated in operation 490, the target loss function and webpage training sample that are generated according to this
Vector is levied, webpage scoring model is created, and iterations is added 1.Return and perform operation 450.
The webpage scoring model finally created using the present embodiment, to carry out beating for multiple webpages under arbitary inquiry word
Point, the phenomenon that the webpage characteristic action to be adjusted for such as omitting feature etc can be caused not enough is improved, so as to carry significantly
The accuracy rate of height marking so that sequence of the search engine to multiple webpages is more reasonable, lifts user's search experience.
Embodiment five
Fig. 5 show a kind of structural representation of the creating device of webpage scoring model of the offer of the embodiment of the present invention five.
Referring to Fig. 5, the device can realize that concrete structure is as follows by software and/hardware：
Webpage training sample acquiring unit 510, for obtaining webpage training sample set, wherein the webpage training sample set
Including the characteristic vector with multiple sample web pages under each query word at least one default query word and mark marking；
Target loss function generation unit 520, the mark for concentrating this webpage of various kinds according to the webpage training sample
Marking and predetermined at least one webpage feature to be adjusted, generate target loss function；
Webpage scoring model creating unit 530, for according to the target loss function generated and webpage training
The characteristic vector of each sample web page in sample set, creates webpage scoring model.
Further, the target loss function generation unit 520, including：
Primary loss function obtains subelement 5201, and this webpage of various kinds is concentrated according to the webpage training sample for obtaining
The obtained primary loss function of mark marking；
Decision factor determination subelement 5202, for determining that being used in the primary loss function weighs the webpage instruction
Practice in sample set the decision factor of difference degree between different web pages under same queries word；
Acting factor adds subelement 5203, for being directed to respectively in predetermined at least one webpage feature to be adjusted
Each webpage feature to be adjusted, will be added in the decision factor for the acting factor for characterizing webpage feature to be adjusted, with
Generate target loss function.
Further, the acting factor addition subelement 5203, specifically for：
Each webpage feature to be adjusted in predetermined at least one webpage feature to be adjusted is directed to respectively, will be used to characterize
The acting factor of webpage feature to be adjusted is multiplied by after corresponding function coefficient, is added in the decision factor；
Wherein, if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened, corresponding effect system
Number is positive coefficient；If the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be weakened, corresponding function coefficient
For negative coefficient.
Further, the device also includes：
Function coefficient updating block 540, in the webpage scoring model creating unit 530 according to the target generated
Loss function and the webpage training sample are concentrated after the characteristic vector of this webpage of various kinds, establishment webpage scoring model, root
According to the function coefficient in target loss function described in setting Policy Updates；
Wherein, the setting rule includes：If getting any webpage at least one described webpage feature to be adjusted
Feature to be adjusted is in the not enough state of effect, then increases the target elements correspondence for characterizing any webpage feature to be adjusted
Function coefficient；If get any webpage feature to be adjusted at least one webpage feature to be adjusted be in effect compared with
Big state, then reduce the corresponding function coefficient of target elements for characterizing any webpage feature to be adjusted；
New web page scoring model creating unit 550, for according to the target loss function after updating operation and institute
The characteristic vector that webpage training sample concentrates this webpage of various kinds is stated, new webpage scoring model is created.
Further, the decision factor H (xqi,xqj)=- h (xqi)+h(xqj)；
The decision factor after addition operation
Wherein, q is the integer more than or equal to 1 and less than or equal to Q, and Q is all query words at least one described query word
Number；
h(xqi) and h (xqj) be the webpage scoring model predicted value, xqiAnd xqjRespectively it is described at least one look into
Ask the different web pages under q-th of query word in word；
T is more than or equal to 1 and less than or equal to the integer of at least one webpage feature total number to be adjusted；
reduce_difft* label_diff is used to characterize t-th of webpage at least one described webpage feature to be adjusted
The acting factor of feature to be adjusted, εtFor with waiting to adjust for characterizing t-th of webpage at least one described webpage feature to be adjusted
The corresponding function coefficient of acting factor of whole feature；
reduce_difft=reducet,qi-reducet,qj, it is xqiT-th of webpage characteristic value reduce to be adjustedt,qi
With xqjT-th of webpage characteristic value reduce to be adjustedt,qjBetween difference；
Label_diff=labelqi-labelqjIt is xqiMark marking labelqiWith xqjMark marking labelqjIt
Difference.
On the basis of above-mentioned technical proposal, the primary loss function is cross entropy loss function, or Hinge
Loss loss functions.
On the basis of above-mentioned technical proposal, the webpage scoring model creating unit 530, specifically for：According to generation
Target loss function and webpage training sample concentrate the characteristic vector of various kinds this webpage, based on gradient lifting sort algorithm wound
Build at least one decision tree；At least one decision tree created is utilized to create webpage scoring model.
The said goods can perform the method that any embodiment of the present invention is provided, and possess the corresponding functional module of execution method
And beneficial effect.
Note, above are only presently preferred embodiments of the present invention and institute's application technology principle.It will be appreciated by those skilled in the art that
The invention is not restricted to specific embodiment described here, can carry out for a person skilled in the art it is various it is obvious change,
Readjust and substitute without departing from protection scope of the present invention.Therefore, although the present invention is carried out by above example
It is described in further detail, but the present invention is not limited only to above example, without departing from the inventive concept, also
Other more equivalent embodiments can be included, and the scope of the present invention is determined by scope of the appended claims.
Claims (14)
1. a kind of creation method of webpage scoring model, it is characterised in that including：
Webpage training sample set is obtained, wherein the webpage training sample set includes and each at least one default query word
The characteristic vector and mark of multiple sample web pages under query word are given a mark；
The mark marking of this webpage of various kinds and at least one predetermined webpage is concentrated to treat according to the webpage training sample
Feature is adjusted, target loss function is generated；
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function generated and the webpage training sample, is created
Webpage scoring model.
2. the creation method of webpage scoring model according to claim 1, it is characterised in that sample is trained according to the webpage
The mark marking of each sample web page of this concentration and predetermined at least one webpage feature to be adjusted, generate target loss letter
Number, including：
Obtain the primary loss function for concentrating the mark marking of this webpage of various kinds to obtain according to the webpage training sample；
Determine that being used in the primary loss function weighs the webpage training sample and concentrate different web pages under same queries word
Between difference degree decision factor；
Each webpage feature to be adjusted in predetermined at least one webpage feature to be adjusted is directed to respectively, will be used to characterize the net
The acting factor of page feature to be adjusted is added in the decision factor, to generate target loss function.
3. the creation method of webpage scoring model according to claim 2, it is characterised in that described to be used to characterize the net
The acting factor of page feature to be adjusted is added in the decision factor, including：
To be multiplied by for the acting factor for characterizing webpage feature to be adjusted after corresponding function coefficient, be added to the decision-making because
In son；
Wherein, if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened, corresponding function coefficient is
Positive coefficient；If the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be weakened, corresponding function coefficient is negative
Coefficient.
4. the creation method of webpage scoring model according to claim 3, it is characterised in that according to the target generated
Loss function and the webpage training sample concentrate the characteristic vector of this webpage of various kinds, create after webpage scoring model, go back
Including：
Function coefficient in the target loss function according to setting Policy Updates；
The feature of this webpage of various kinds is concentrated according to the target loss function and the webpage training sample after updating operation
Vector, creates new webpage scoring model.
5. the creation method of webpage scoring model according to claim 3, it is characterised in that the decision factor H (xqi,
xqj)=- h (xqi)+h(xqj)；
The decision factor after addition operation
<mrow>
<msup>
<mi>H</mi>
<mi>new</mi>
</msup>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qi</mi>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>qj</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mi>h</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qi</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<mi>h</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qj</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<munder>
<mi>&Sigma;</mi>
<mi>t</mi>
</munder>
<msub>
<mi>&epsiv;</mi>
<mi>t</mi>
</msub>
<mo>*</mo>
<mrow>
<mo>(</mo>
<mi>reduce</mi>
<mo>_</mo>
<msub>
<mi>diff</mi>
<mi>t</mi>
</msub>
<mo>*</mo>
<mi>label</mi>
<mo>_</mo>
<mi>diff</mi>
<mo>)</mo>
</mrow>
</mrow>
Wherein, q is the integer more than or equal to 1 and less than or equal to Q, and Q is all query word numbers at least one described query word；
h(xqi) and h (xqj) be the webpage scoring model predicted value, xqiAnd xqjAt least one respectively described query word
In different web pages under q-th of query word；
T is more than or equal to 1 and less than or equal to the integer of at least one webpage feature total number to be adjusted；
reduce_difft* label_diff is to be adjusted for characterizing t-th of webpage at least one described webpage feature to be adjusted
The acting factor of feature, εtFor with for characterizing t-th of webpage feature to be adjusted at least one webpage feature to be adjusted
The corresponding function coefficient of acting factor；
reduce_difft=reducet,qi-reducet,qj, it is xqiT-th of webpage characteristic value reduce to be adjustedt,qiWith xqj
T-th of webpage characteristic value reduce to be adjustedt,qjBetween difference；
Label_diff=labelqi-labelqjIt is xqiMark marking labelqiWith xqjMark marking labelqjDifference.
6. according to the creation method of any described webpage scoring model in claim 2-5, it is characterised in that the original damage
Mistake function is cross entropy loss function, or Hinge loss loss functions.
7. according to the creation method of any described webpage scoring model in claim 1-5, it is characterised in that
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function generated and the webpage training sample, is created
Webpage scoring model, including：
The characteristic vector of this webpage of various kinds is concentrated according to the target loss function of generation and webpage training sample, is carried based on gradient
Rise sort algorithm and create at least one decision tree；At least one decision tree created is utilized to create webpage scoring model.
8. a kind of creating device of webpage scoring model, it is characterised in that including：
Webpage training sample acquiring unit, for obtaining webpage training sample set, wherein the webpage training sample set include with
The characteristic vector and mark of multiple sample web pages at least one default query word under each query word are given a mark；
Target loss function generation unit, for concentrated according to the webpage training sample various kinds this webpage mark marking and
Predetermined at least one webpage feature to be adjusted, generates target loss function；
Webpage scoring model creating unit, for according to the target loss function generated and webpage training sample concentration
The characteristic vector of each sample web page, creates webpage scoring model.
9. the creating device of webpage scoring model according to claim 8, it is characterised in that the target loss function life
Into unit, including：
Primary loss function obtains subelement, concentrates the mark of this webpage of various kinds to beat according to the webpage training sample for obtaining
The primary loss function got；
Decision factor determination subelement, for determining that being used in the primary loss function weighs the webpage training sample set
Under middle same queries word between different web pages difference degree decision factor；
Acting factor adds subelement, is treated for being directed to each webpage in predetermined at least one webpage feature to be adjusted respectively
Feature is adjusted, will be added to for the acting factor for characterizing webpage feature to be adjusted in the decision factor, to generate target
Loss function.
10. the creating device of webpage scoring model according to claim 9, it is characterised in that the acting factor addition
Subelement, specifically for：
Each webpage feature to be adjusted in predetermined at least one webpage feature to be adjusted is directed to respectively, will be used to characterize the net
The acting factor of page feature to be adjusted is multiplied by after corresponding function coefficient, is added in the decision factor；
Wherein, if the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be strengthened, corresponding function coefficient is
Positive coefficient；If the webpage of acting factor sign is to be adjusted to be characterized as webpage feature to be weakened, corresponding function coefficient is negative
Coefficient.
11. the creating device of webpage scoring model according to claim 10, it is characterised in that also include：
Function coefficient updating block, for the webpage scoring model creating unit according to the target loss function generated with
And the webpage training sample concentrates the characteristic vector of this webpage of various kinds, after creating webpage scoring model, according to setting rule
Update the function coefficient in the target loss function；
New web page scoring model creating unit, for being instructed according to the target loss function and the webpage after updating operation
Practice the characteristic vector of each sample web page in sample set, create new webpage scoring model.
12. the creating device of webpage scoring model according to claim 10, it is characterised in that the decision factor H
(xqi,xqj)=- h (xqi)+h(xqj)；
The decision factor after addition operation
<mrow>
<msup>
<mi>H</mi>
<mi>new</mi>
</msup>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qi</mi>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>qj</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mi>h</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qi</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<mi>h</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>qj</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<munder>
<mi>&Sigma;</mi>
<mi>t</mi>
</munder>
<msub>
<mi>&epsiv;</mi>
<mi>t</mi>
</msub>
<mo>*</mo>
<mrow>
<mo>(</mo>
<mi>reduce</mi>
<mo>_</mo>
<msub>
<mi>diff</mi>
<mi>t</mi>
</msub>
<mo>*</mo>
<mi>label</mi>
<mo>_</mo>
<mi>diff</mi>
<mo>)</mo>
</mrow>
</mrow>
Wherein, q is the integer more than or equal to 1 and less than or equal to Q, and Q is all query word numbers at least one described query word；
h(xqi) and h (xqj) be the webpage scoring model predicted value, xqiAnd xqjAt least one respectively described query word
In different web pages under q-th of query word；
T is more than or equal to 1 and less than or equal to the integer of at least one webpage feature total number to be adjusted；
reduce_difft* label_diff is to be adjusted for characterizing t-th of webpage at least one described webpage feature to be adjusted
The acting factor of feature, εtFor with for characterizing t-th of webpage feature to be adjusted at least one webpage feature to be adjusted
The corresponding function coefficient of acting factor；
reduce_difft=reducet,qi-reducet,qj, it is xqiT-th of webpage characteristic value reduce to be adjustedt,qiWith xqj
T-th of webpage characteristic value reduce to be adjustedt,qjBetween difference；
Label_diff=labelqi-labelqjIt is xqiMark marking labelqiWith xqjMark marking labelqjDifference.
13. according to the creating device of any described webpage scoring model in claim 9-12, it is characterised in that described original
Loss function is cross entropy loss function, or Hinge loss loss functions.
14. according to the creating device of any described webpage scoring model in claim 8-12, it is characterised in that the webpage
Scoring model creating unit, specifically for：Various kinds Home Network is concentrated according to the target loss function of generation and webpage training sample
The characteristic vector of page, at least one decision tree is created based on gradient lifting sort algorithm；Utilize at least one decision-making created
Tree creates webpage scoring model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410638360.4A CN104361077B (en) | 2014-11-06 | 2014-11-06 | The creation method and device of webpage scoring model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410638360.4A CN104361077B (en) | 2014-11-06 | 2014-11-06 | The creation method and device of webpage scoring model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104361077A CN104361077A (en) | 2015-02-18 |
| CN104361077B true CN104361077B (en) | 2017-11-03 |
Family
ID=52528338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410638360.4A Active CN104361077B (en) | 2014-11-06 | 2014-11-06 | The creation method and device of webpage scoring model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104361077B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104715063B (en) * | 2015-03-31 | 2018-11-02 | 百度在线网络技术（北京）有限公司 | search ordering method and device |
| CN107622056B (en) * | 2016-07-13 | 2021-03-02 | 百度在线网络技术（北京）有限公司 | Training sample generation method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043776A (en) * | 2009-10-14 | 2011-05-04 | 南开大学 | Inquiry-related multi-ranking-model integration algorithm |
| CN103984733A (en) * | 2014-05-20 | 2014-08-13 | 国家电网公司 | Direct optimizing performance index sequencing method capable of embodying query difference |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7647314B2 (en) * | 2006-04-28 | 2010-01-12 | Yahoo! Inc. | System and method for indexing web content using click-through features |
-
2014
- 2014-11-06 CN CN201410638360.4A patent/CN104361077B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043776A (en) * | 2009-10-14 | 2011-05-04 | 南开大学 | Inquiry-related multi-ranking-model integration algorithm |
| CN103984733A (en) * | 2014-05-20 | 2014-08-13 | 国家电网公司 | Direct optimizing performance index sequencing method capable of embodying query difference |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104361077A (en) | 2015-02-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Esfandyari et al. | A tuned version of genetic algorithm for efficient test suite generation in interactive t-way testing strategy | |
| CN104598611B (en) | The method and system being ranked up to search entry | |
| CN108509408A (en) | A kind of sentence similarity judgment method | |
| CN107967255A (en) | A kind of method and system for judging text similarity | |
| CN108170736A (en) | A kind of document based on cycle attention mechanism quickly scans qualitative method | |
| CN106228183A (en) | A kind of semi-supervised learning sorting technique and device | |
| CN106355192A (en) | Support vector machine method based on chaos and grey wolf optimization | |
| Fazzolari et al. | A study on the application of instance selection techniques in genetic fuzzy rule-based classification systems: Accuracy-complexity trade-off | |
| CN104391860A (en) | Content type detection method and device | |
| CN108459999B (en) | Font design method, system, equipment and computer readable storage medium | |
| CN106056209B (en) | A kind of inquiry lexical item weight learning method based on Recognition with Recurrent Neural Network | |
| CN105550746A (en) | Training method and training device of machine learning model | |
| CN107870964A (en) | A kind of sentence sort method and system applied to answer emerging system | |
| CN109344994A (en) | A kind of prediction model method based on improvement moth optimization algorithm | |
| CN110222848A (en) | The determination method and device for the integrated model that computer executes | |
| Ali et al. | A modified cultural algorithm with a balanced performance for the differential evolution frameworks | |
| CN111768004A (en) | Model self-adaption method and system based on intelligent computing framework | |
| CN112000770A (en) | Intelligent question and answer oriented sentence-to-sentence matching method based on semantic feature map | |
| CN106909972A (en) | A kind of learning method of sensing data calibrating patterns | |
| CN110008332A (en) | The method and device of trunk word is extracted by intensified learning | |
| CN104361077B (en) | The creation method and device of webpage scoring model | |
| CN112084307A (en) | Data processing method and device, server and computer readable storage medium | |
| Hekanaho | DOGMA: A GA-based relational learner | |
| CN106446011A (en) | Data processing method and device | |
| Kadavy et al. | The ensemble of strategies and perturbation parameter in self-organizing migrating algorithm solving CEC 2019 100-digit challenge |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |