CN104750731B - A kind of method and device obtaining whole user portrait - Google Patents
A kind of method and device obtaining whole user portrait Download PDFInfo
- Publication number
- CN104750731B CN104750731B CN201310745711.7A CN201310745711A CN104750731B CN 104750731 B CN104750731 B CN 104750731B CN 201310745711 A CN201310745711 A CN 201310745711A CN 104750731 B CN104750731 B CN 104750731B
- Authority
- CN
- China
- Prior art keywords
- parameter matrix
- matrix
- user
- portrait
- customer parameter
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Transfer Between Computers (AREA)
Abstract
The invention discloses a kind of method of acquisition whole user portrait, the method for the present invention includes:Obtain incomplete user's portrait matrix, and random generation customer parameter matrix P and label matrix Q;Calculate the portrait error of first part user, update customer parameter matrix and tag parameter matrix, wherein, the first variation difference of the first part user of selection is more than the first variation difference of the first remaining users, first remaining users are the user in addition to first part user in multiple users, difference of the first variation difference between 1 newer first predicted value of user r and 2 newer first predicted values of user r;After updating customer parameter matrix P and tag parameter matrix Q at the R times, according to matrix decomposition as a result, obtaining complete user's portrait matrix.The present invention can accelerate the convergence rate of portrait error, operand be reduced, to reduce the operation time of system.
Description
Technical field
The present invention relates to field of information processing, and in particular to a kind of method and device obtaining whole user portrait.
Background technology
User tag plays an important role in practical applications.Such as, it is known that label includes the interest of user and often goes
Place can make and accurately marketing, while better service experience can be brought to user.In general, the label of user can be with
With a matrix rM×NIt indicates, wherein 1≤m≤M indicates user index, i.e., one shared M user, and 1≤n≤N indicates user's mark
Label index, i.e. each user one share N number of label.In general, the user tag observed is all incomplete, that is, there is unknown mark
The value of label, with "" symbolic indication.When there are a large number of users number and number of tags, matrix rM×NIt may abnormal sparse, square
It is filled with a large amount of 0 element in battle array, which represents Unknown Label, referred to as incomplete user portrait matrix.
In order to obtain the label information of each user, the method for using matrix decomposition at present, according to incomplete use
Family portrait matrix measures complete user's portrait matrix in advance.This method specifically includes:
Step 1:The incomplete portrait matrix r of inputM×N, and random initializtion customer parameter matrix P and tag parameter matrix Q.
Step 2:Calculating matrix rM×NThe portrait error e of middle nonzero elementM, n, calculation formula is:
eM, n=rM, n-pm Tqn。
Step 3:According to portrait error update parameter P and Q, calculation formula is:
pm=pm+γ(eM, nqn-λPpm), qn=qn+γ(eM, npm-λQqn)。
Wherein, γ is learning efficiency, is traditionally arranged to be γ=0.001.
Step 4:Step 2 and step 3 are repeated until matrix rM×NIn all nonzero elements scan one time, referred to as once
Cycle.
Step 5:1≤t of step 4≤T cycles are repeated until customer parameter matrix P and tag parameter matrix Q and portrait
Error is substantially constant in a state, i.e., repeatedly step 4 until portrait error convergence.
Step 6:Export whole user portrait matrix
But it is found in implementing this method:Cycle needs to scan input matrix r every timeM×NIn all nonzero elements,
Each nonzero element needs to calculate the portrait error on all class of subscribers, and reaches the cycle time that portrait error convergence needs
Number is more, therefore the operand of this method is big, takes very much.
Invention content
It is an object of the invention to a kind of method and devices of acquisition whole user portrait, and the method is according to first
Divide the portrait error of user, customer parameter matrix and tag parameter matrix is updated, so as to reduce operand.
A kind of method for acquisition whole user portrait that first aspect present invention provides, including:
Incomplete user's portrait matrix is obtained, and generates initial customer parameter matrix P and tag parameter matrix at random
Q;Wherein, user's portrait matrix, the customer parameter matrix P and the tag parameter matrix Q of the incompleteness are corresponded to identical
Multiple users, the customer parameter matrix P and the tag parameter matrix Q are the decomposition square of user's portrait matrix of the incompleteness
Battle array;
According to the user of incompleteness portrait matrix, the R update customer parameter matrix P and the tag parameter matrix
Q, R are the update times so that needed for the first portrait error convergence;The first portrait error is that the user of the incompleteness draws a portrait
Difference in matrix between nonzero element and corresponding first predicted value, first predicted value are the customer parameter matrix P
With the element for merging matrix of the tag parameter matrix Q;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than
Or it is equal to R, including:
As r=1, according to the user of incompleteness portrait matrix, the initial customer parameter matrix P and mark generated at random
Parameter matrix Q is signed, the first portrait error of nonzero element in user's portrait matrix of the incompleteness is calculated;It is drawn according to described first
As error, the customer parameter matrix P and the tag parameter matrix Q are updated;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, packet
It includes:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, calculate
The first portrait error of first part user in the multiple user;The first variation difference of the first part user is more than
First variation difference of the first remaining users, first remaining users are in the multiple user except the first part uses
User except family;
When r is equal to 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained with according to the initial customer parameter matrix P and institute generated at random
State the difference between the first predicted value that tag parameter matrix Q is obtained;
When r is more than 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and institute with according to the r-2 times
State the difference between the first predicted value that tag parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and label ginseng are updated
Matrix number Q;
After updating the customer parameter matrix P and tag parameter matrix Q at the R times, according to the use of the incompleteness
Draw a portrait matrix and the R times newer customer parameter matrix P and tag parameter matrix Q at family, obtains complete user
Portrait matrix.
In conjunction with first aspect present invention, in the first possible realization method of first aspect, described in the r times update
Customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than or equal to R, including:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, meter
Before the first portrait error for calculating the first part user in the multiple user, the variation difference of the multiple user is calculated,
And be ranked up the variation difference of the multiple user, select the user of predetermined ratio as the first part user, institute
The variation difference for stating the user of predetermined ratio is both greater than the variation difference of non-selected user in the multiple user.
In conjunction with the possible realization method of the first of first aspect present invention or first aspect, second in first aspect can
In energy realization method, after updating the customer parameter matrix P and tag parameter matrix Q at the R times, the method is also wrapped
It includes:
The customer parameter matrix P is subjected to S deep decomposition, and obtain the customer parameter matrix of S level depth with
Tag parameter matrix;Described S times is the number set;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
It is random to generate customer parameter matrix PwWith tag parameter matrix Qw;Wherein, the customer parameter matrix Pw, the mark
Sign parameter matrix QwIdentical multiple users corresponding with the customer parameter matrix P;Customer parameter matrix PwWith tag parameter matrix
QwFor the split-matrix of the customer parameter matrix P;
According to the customer parameter matrix P, the Y update customer parameter matrix PwWith the tag parameter matrix Qw, Y
For the update times for making needed for the second portrait error convergence;The second portrait error is non-in the customer parameter matrix P
Difference between neutral element and corresponding second predicted value, second predicted value are the customer parameter matrix PwWith the mark
Sign parameter matrix QwMerging matrix element;
Wherein, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, y is greater than or equal to 1 and small
In or be equal to Y, including:
As y=1, according to the customer parameter matrix P, initial customer parameter matrix P is generated at randomwAnd tag parameter
Matrix Qw, calculate the second portrait error of nonzero element in user's portrait matrix of the incompleteness;
According to the second portrait error, the customer parameter matrix P is updatedwWith the tag parameter matrix Qw;
When y is greater than or equal to 2, according to the customer parameter matrix P, the customer parameter matrix PwJoin with the label
Matrix number Qw, calculate the second portrait error of the second part user in the multiple user;The of the second part user
Two variation differences are more than the second variation difference of the second remaining users, and second remaining users are removing in the multiple user
User except the second part user;
When y is equal to 2, the second variation difference is to update obtained customer parameter matrix P according to the y-1 timeswAnd mark
Sign parameter matrix QwObtain the second predicted value with according to the initial customer parameter matrix P and tag parameter matrix generated at random
QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained customer parameter matrix P and label according to the y-1 times
Parameter matrix QwThe second predicted value obtained updates obtained customer parameter matrix P with according to the y-2 timeswWith tag parameter matrix
QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the customer parameter matrix P is updatedwAnd institute
State tag parameter matrix Qw;
The obtained customer parameter matrix P is updated by Y timeswWith the tag parameter matrix QwAs w level depth
Customer parameter matrix PwWith tag parameter matrix Qw;
When the w is less than the S, the obtained customer parameter matrix P is updated by Y timeswWant deep as the w+1 times
Spend the customer parameter matrix P decomposed;
It is described according to the user of incompleteness portrait matrix and the R times newer customer parameter matrix P and described
Tag parameter matrix Q obtains complete user's portrait matrix, including:
The customer parameter matrix P is being subjected to S deep decomposition, and is obtaining the customer parameter matrix of S level depth
After tag parameter matrix, drawn a portrait matrix according to the user of the incompleteness, the R time newer customer parameter matrix P with
The tag parameter matrix Q, and obtain S level depth customer parameter matrix and tag parameter matrix, obtain completely
User draw a portrait matrix.
Second aspect of the present invention provides a kind of information processing unit, including:
Input module, user's portrait matrix for obtaining incompleteness, and user's portrait square of the incompleteness is generated at random
The customer parameter matrix P and tag parameter matrix Q of battle array;Wherein, user's portrait matrix, the customer parameter matrix of the incompleteness
P and the tag parameter matrix Q correspond to identical multiple users;
Matrix decomposition module is used for:Incomplete user's portrait matrix is obtained in the input module, and generates institute at random
After the customer parameter matrix P and tag parameter matrix Q that state incomplete user's portrait matrix, drawn a portrait according to the user of the incompleteness
Matrix, the R update customer parameter matrix P and the tag parameter matrix Q, R are so that needed for the first portrait error convergence
Update times;The first portrait error is nonzero element and corresponding first prediction in user's portrait matrix of the incompleteness
Difference between value, first predicted value are that the customer parameter matrix P and the tag parameter matrix Q merge matrix
Element;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than
Or it is equal to R, including:
As r=1, the r times update customer parameter matrix P and tag parameter matrix Q, including:According to institute
Incomplete user's portrait matrix is stated, the initial customer parameter matrix P and tag parameter matrix Q generated at random is calculated described residual
First portrait error of nonzero element in scarce user's portrait matrix;According to the first portrait error, user's ginseng is updated
The matrix number P and tag parameter matrix Q;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, packet
It includes:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, calculate
The first portrait error of first part user in the multiple user;The first variation difference of the first part user is more than
First variation difference of the first remaining users, first remaining users are in the multiple user except the first part uses
User except family;
When r is equal to 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained with according to the initial customer parameter matrix P and institute generated at random
State the difference between the first predicted value that tag parameter matrix Q is obtained;
When r is more than 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and institute with according to the r-2 times
State the difference between the first predicted value that tag parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and label ginseng are updated
Matrix number Q;
Output module, after updating the customer parameter matrix P and tag parameter matrix Q at the R times, according to
User's portrait matrix of the incompleteness and the R times newer customer parameter matrix P and tag parameter matrix Q, are obtained
Take complete user's portrait matrix.
In conjunction with second aspect of the present invention, in the first possible realization method of second aspect, the matrix decomposition module
Including dynamic dispatching module, the dynamic dispatching module is used to join according to the user of incompleteness portrait matrix, the user
The matrix number P and tag parameter matrix Q, calculate the first part user in the multiple user first portrait error it
Before, the variation difference of the multiple user is calculated, and the variation difference of the multiple user is ranked up, selects predetermined ratio
User as the first part user, the variation difference of the user of the predetermined ratio is both greater than in the multiple user not
The variation difference of selected user.
In conjunction with second aspect of the present invention or second aspect the first possible realization method, second in second aspect may
In realization method, the matrix decomposition module further includes deep decomposition module, and the deep decomposition module is used for:More at the R times
After the new customer parameter matrix P and tag parameter matrix Q, the customer parameter matrix P is subjected to S depth point
Solution, and obtain the customer parameter matrix and tag parameter matrix of S level depth;Described S times is the number set;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
It is random to generate customer parameter matrix PwWith tag parameter matrix Qw;Wherein, the customer parameter matrix Pw, the mark
Sign parameter matrix QwIdentical multiple users corresponding with the customer parameter matrix P;Customer parameter matrix PwWith tag parameter matrix
QwFor the split-matrix of the customer parameter matrix P;
According to the customer parameter matrix P, the Y update customer parameter matrix PwWith the tag parameter matrix Qw, Y
For the update times for making needed for the second portrait error convergence;The second portrait error is non-in the customer parameter matrix P
Difference between neutral element and corresponding second predicted value, second predicted value are the customer parameter matrix PwWith the mark
Sign parameter matrix QwMerging matrix element;
Wherein, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, y is greater than or equal to 1 and small
In or be equal to Y, including:
As y=1, according to the customer parameter matrix P, initial customer parameter matrix P is generated at randomwAnd tag parameter
Matrix Qw, calculate the second portrait error of nonzero element in user's portrait matrix of the incompleteness;
According to the second portrait error, the customer parameter matrix P is updatedwWith the tag parameter matrix Qw;
When y is greater than or equal to 2, according to the customer parameter matrix P, the customer parameter matrix PwJoin with the label
Matrix number Qw, calculate the second portrait error of the second part user in the multiple user;The of the second part user
Two variation differences are more than the second variation difference of the second remaining users, and second remaining users are removing in the multiple user
User except the second part user;
When y is equal to 2, the second variation difference is to update obtained customer parameter matrix P according to the y-1 timeswAnd mark
Sign parameter matrix QwObtain the second predicted value with according to the initial customer parameter matrix P and tag parameter matrix generated at random
QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained customer parameter matrix P and label according to the y-1 times
Parameter matrix QwThe second predicted value obtained updates obtained customer parameter matrix P with according to the y-2 timeswWith tag parameter matrix
QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the customer parameter matrix P is updatedwAnd institute
State tag parameter matrix Qw;
The obtained customer parameter matrix P is updated by Y timeswWith the tag parameter matrix QwAs w level depth
Customer parameter matrix PwWith tag parameter matrix Qw;
When the w is less than the S, the obtained customer parameter matrix P is updated by Y timeswWant deep as the w+1 times
Spend the customer parameter matrix P decomposed;
The output module is additionally operable to the customer parameter matrix P carrying out S deep decomposition, and obtains S level
After the customer parameter matrix and tag parameter matrix of depth, according to the user of incompleteness portrait matrix, the R times newer institute
State customer parameter matrix P and the tag parameter matrix Q, and the customer parameter matrix and label of S level depth of acquisition
Parameter matrix obtains complete user's portrait matrix.
In the present invention, the first variation difference of the first part user of selection is more than the first variation of the first remaining users
Difference updates customer parameter matrix and tag parameter matrix according to the portrait error of first part user, can accelerate portrait and miss
The convergence rate of difference reduces operand, to reduce the operation time of system.
Description of the drawings
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, other are can also be obtained according to these attached drawings
Attached drawing.
Fig. 1 is a kind of method flow schematic diagram obtaining whole user portrait provided in an embodiment of the present invention;
Fig. 2 is deep decomposition matrix procedures schematic diagram;
Fig. 3 is a kind of structural schematic diagram of information processing unit provided in an embodiment of the present invention;
Fig. 4 is a kind of another structural schematic diagram of information processing unit provided in an embodiment of the present invention;
Fig. 5 is the structural schematic diagram of another information processing unit provided in an embodiment of the present invention.
Specific implementation mode
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation describes, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other without creative efforts
Embodiment shall fall within the protection scope of the present invention.
Embodiment is exemplified below the present invention is introduced.
As shown in Figure 1, the embodiment of the present invention provides a kind of method of acquisition whole user portrait, including:
101, incomplete user's portrait matrix is obtained, and generates initial customer parameter matrix P and tag parameter at random
Matrix Q;Wherein, user's portrait matrix, the customer parameter matrix P and the tag parameter matrix Q of the incompleteness correspond to phase
Same multiple users, the customer parameter matrix P and the tag parameter matrix Q are point of user's portrait matrix of the incompleteness
Dematrix.
102, according to the user of incompleteness portrait matrix, the R update customer parameter matrix P and the tag parameter
Matrix Q, R are the update times so that needed for the first portrait error convergence;The first portrait error is the user of the incompleteness
Difference in portrait matrix between nonzero element and corresponding first predicted value, first predicted value are the customer parameter square
The element for merging matrix of battle array P and the tag parameter matrix Q.
Wherein, 103, the r times update customer parameter matrix P and the tag parameter matrix Q, r be greater than or equal to 1 and
Less than or equal to R.Step 103 includes:
As r=1, the r times update customer parameter matrix P and tag parameter matrix Q, including:
According to the user of incompleteness portrait matrix, initial customer parameter matrix P and tag parameter matrix are generated at random
Q calculates the first portrait error of nonzero element in user's portrait matrix of the incompleteness;
According to the first portrait error, the customer parameter matrix P and the tag parameter matrix Q are updated;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, packet
It includes:
103a, draw a portrait matrix, the customer parameter matrix P and the tag parameter matrix Q according to the user of the incompleteness,
Calculate the first portrait error of the first part user in the multiple user;The first variation difference of the first part user
Change difference more than the first of the first remaining users, first remaining users are to remove described first in the multiple user
Divide the user except user.
When r is equal to 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained with according to the initial customer parameter matrix P and institute generated at random
State the difference between the first predicted value that tag parameter matrix Q is obtained;
When r is more than 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and institute with according to the r-2 times
State the difference between the first predicted value that tag parameter matrix Q is obtained.
103b, error of drawing a portrait according to the first of the first part user, update the customer parameter matrix P and the mark
Sign parameter matrix Q;
104, after updating the customer parameter matrix P and tag parameter matrix Q at the R times, according to the incompleteness
User draw a portrait matrix and the R time newer customer parameter matrix P and tag parameter matrix Q, obtain completely
User's portrait matrix.
In embodiments of the present invention, the first variation difference of first part user is more than the first variation of the first remaining users
Difference updates customer parameter matrix and tag parameter matrix according to the portrait error of first part user, can accelerate portrait and miss
The convergence rate of difference reduces operand, to reduce the operation time of system.
Preferably, step 103 can also include:According to the user of incompleteness portrait matrix, the customer parameter square
Battle array the P and tag parameter matrix Q, calculate the first part user in the multiple user first portrait error before, meter
The variation difference of the multiple user is calculated, and the variation difference of the multiple user is ranked up, selects the use of predetermined ratio
As the first part user, the variation difference of the user of the predetermined ratio, which is both greater than in the multiple user, not to be chosen at family
The variation difference of the user selected.Therefore, the embodiment of the present invention can accurately determine first part user.
As shown in Fig. 2, after updating the customer parameter matrix P and tag parameter matrix Q at the R times, the side
Method can also include:
201, the customer parameter matrix P is subjected to S deep decomposition, and obtains the customer parameter square of S level depth
Battle array and tag parameter matrix;Described S times is the number set;
Wherein, 202, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S.Step
Rapid 202, including:
202a, random generation customer parameter matrix PwWith tag parameter matrix Qw;Wherein, the customer parameter matrix Pw, institute
State tag parameter matrix QwIdentical multiple users corresponding with the customer parameter matrix P;Customer parameter matrix PwAnd tag parameter
Matrix QwFor the split-matrix of the customer parameter matrix P.
202b, according to the customer parameter matrix P, the Y update customer parameter matrix PwWith the tag parameter square
Battle array Qw, Y is the update times so that needed for the second portrait error convergence;The second portrait error is the customer parameter matrix
Difference in P between nonzero element and corresponding second predicted value, second predicted value are the customer parameter matrix PwWith
The tag parameter matrix QwMerging matrix element.
Wherein, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, y is greater than or equal to 1 and small
In or be equal to Y, including:
As y=1, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, including:According to described
Customer parameter matrix P generates initial customer parameter matrix P at randomwWith tag parameter matrix Qw, calculate the user of the incompleteness
Second portrait error of nonzero element in portrait matrix;
According to the second portrait error, the customer parameter matrix P is updatedwWith the tag parameter matrix Qw;
When y is greater than or equal to 2, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, including:
According to the customer parameter matrix P, the customer parameter matrix PwWith the tag parameter matrix Qw, calculate the multiple user
In second part user second portrait error;The second variation difference of the second part user is more than the second remaining users
Second variation difference, second remaining users be the multiple user in the use in addition to the second part user
Family;
When y is equal to 2, the second variation difference is to update obtained customer parameter matrix P according to the y-1 timeswAnd mark
Sign parameter matrix QwObtain the second predicted value with according to the initial customer parameter matrix P and tag parameter matrix generated at random
QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained customer parameter matrix P and label according to the y-1 times
Parameter matrix QwThe second predicted value obtained updates obtained customer parameter matrix P with according to the y-2 timeswWith tag parameter matrix
QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the customer parameter matrix P is updatedwAnd institute
State tag parameter matrix Qw;
The obtained customer parameter matrix P is updated by Y timeswWith the tag parameter matrix QwAs w level depth
Customer parameter matrix PwWith tag parameter matrix Qw;
When the w is less than the S, the obtained customer parameter matrix P is updated by Y timeswWant deep as the w+1 times
Spend the customer parameter matrix P decomposed;
203, the customer parameter matrix P is subjected to S deep decomposition, and obtains the customer parameter of S level depth
After matrix and tag parameter matrix, according to the user of incompleteness portrait matrix, the R times newer customer parameter matrix
The P and tag parameter matrix Q, and obtain S level depth customer parameter matrix and tag parameter matrix, obtained
Whole user's portrait matrix.
Implementation steps 201 to 203, after updating the customer parameter matrix P and tag parameter matrix Q at the R times,
Customer parameter matrix P is subjected to further multi-level deep decomposition, merge multi-level matrix decomposition as a result, improving use
The precision of family portrait.
It is exemplified below specific embodiment, the present invention is described further.
The embodiment of the present invention provides a kind of method of acquisition whole user portrait, including:
Step 1:The incomplete portrait matrix r of inputM, N, and random initializtion customer parameter matrix P and tag parameter matrix Q.
Step 2:It calculates predicted value and is recycling the variation difference c between t and t-1 twicem, and it is ranked up, it selects
The variation maximum 20% ratio user of difference carries out portrait error calculation, and the formula specifically calculated is as follows:
cM, n=|(pm Tqn)t-(pm Tqn)t-1|;
cm=ΣncM, n。
Step 3:Calculate the portrait error e of certain customers' nonzero element of selectionM, n, the formula specifically calculated is as follows:
eM, n=rM, n-pm Tqn。
Step 4:According to portrait error update parameter P and Q, the formula specifically calculated is as follows:
pm=pm+γ(eM, nqn-λPpm)。
Step 5:It repeats Step 2: three and four, Step 2: three and four be one cycle, cycles through sequence c every timemIt is dynamic
State dispatches scanned user's nonzero element, but user's nonzero element of all selected sections calculates.
Step 6:1≤t of step 4≤T cycles are repeated until customer parameter matrix P and tag parameter matrix Q and portrait
Error is substantially constant in a state, i.e., until portrait error convergence.
After executing step 6, deep decomposition will be carried out to customer parameter matrix P, the user to improve output matrix draws
As precision.
Step 7:To customer parameter matrix P as input matrix, step 1 is repeated to step 6, obtains the portrait of matrix P
Matrix P1、Q1, P2And Q2, until PSAnd QS, the depth s of matrix decomposition specifies by user, usual s=2.
Below in conjunction with concrete application example, invention is further explained.
For example, by user's portrait matrix of the film matrix of user incompleteness as input, the row of wherein matrix, which represents, to be used
Family, row indicate that user likes film and other relevant parameters that can be provided such as:{ K=100, λP=λQ=0.05, γ=0.001,
S=0, ρm=0.2 }, K indicates class of subscriber number, λPIndicate vector matrix P={ pmAdjusting parameter, λQExpression vector matrix Q=
{qnAdjusting parameter, γ indicate learning efficiency, s indicate the deep layer matrix decomposition number of plies, ρmIndicate selection ratio.The present embodiment is only
ρ is worked as in testmWhen=0.2, the acceleration effect of system operations.In the present embodiment, system is by the non-zero entry in the matrix according to input
Element, calculates the gradient of portrait error, and is decomposed into matrix P and Q, and 20% user is carried out gradient updating.Pass through cycle 100
Secondary, speed is than about 2 times soon of existing matrix decomposition scheme.
The matrix further for example, user of the music matrix of user incompleteness as input draws a portrait, and offer following parameters K=
100, λP=λQ=0.05, γ=0.001, s=2, ρm=1}.The number of plies of deep layer matrix decomposition module is set to 2 by the present embodiment, cycle
It provides user for 100 times to draw a portrait precision 0.81, the error 0.85 than script matrix decomposition scheme is low, reflects the precision of user's portrait
It greatly improves.
As shown in figure 3, the present invention provides a kind of information processing unit 301, including:
Input module 302, user's portrait matrix for obtaining incompleteness, and user's portrait of the incompleteness is generated at random
The customer parameter matrix P and tag parameter matrix Q of matrix;Wherein, user's portrait matrix, the customer parameter square of the incompleteness
The battle array P and tag parameter matrix Q corresponds to identical multiple users;
Matrix decomposition module 303, user's portrait matrix for obtaining incompleteness in the input module 302, and it is random
After the customer parameter matrix P and tag parameter matrix Q of user's portrait matrix for generating the incompleteness, according to the use of the incompleteness
Matrix, the R update customer parameter matrix P and the tag parameter matrix Q, R draw a portrait as so that the first portrait error is received in family
Hold back required update times;The first portrait error be incompleteness user's portrait matrix in nonzero element and corresponding the
Difference between one predicted value, first predicted value are merging for the customer parameter matrix P and the tag parameter matrix Q
The element of matrix;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than
Or it is equal to R, including:
As r=1, the r times update customer parameter matrix P and tag parameter matrix Q, including:According to institute
Incomplete user's portrait matrix is stated, the initial customer parameter matrix P and tag parameter matrix Q generated at random is calculated described residual
First portrait error of nonzero element in scarce user's portrait matrix;According to the first portrait error, user's ginseng is updated
The matrix number P and tag parameter matrix Q;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, packet
It includes:The r times update customer parameter matrix P and tag parameter matrix Q, including:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, calculate
The first portrait error of first part user in the multiple user;The first variation difference of the first part user is more than
First variation difference of the first remaining users, first remaining users are in the multiple user except the first part uses
User except family;
When r is equal to 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained with according to the initial customer parameter matrix P and institute generated at random
State the difference between the first predicted value that tag parameter matrix Q is obtained;
When r is more than 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and institute with according to the r-2 times
State the difference between the first predicted value that tag parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and label ginseng are updated
Matrix number Q;
Output module 304, for updating the customer parameter matrix P and described in the matrix decomposition module 303 the R time
After tag parameter matrix Q, according to the user of incompleteness portrait matrix and the R times newer customer parameter matrix P
With the tag parameter matrix Q, complete user's portrait matrix is obtained.
Preferably, as shown in figure 4, the matrix decomposition module 303 in information processing unit of the present invention 301 further includes moving
State scheduler module 305, the dynamic dispatching module 305 are used to join according to the user of incompleteness portrait matrix, the user
The matrix number P and tag parameter matrix Q, calculate the first part user in the multiple user first portrait error it
Before, the variation difference of the multiple user is calculated, and the variation difference of the multiple user is ranked up, selects predetermined ratio
User as the first part user, the variation difference of the user of the predetermined ratio is both greater than in the multiple user not
The variation difference of selected user.
Preferably, the matrix decomposition module 303 further includes deep decomposition module 306, and the deep decomposition module 306 is used
After updating the customer parameter matrix P and tag parameter matrix Q at the R times, the customer parameter matrix P is carried out
S deep decomposition, and obtain the customer parameter matrix and tag parameter matrix of S level depth;Described S times is the secondary of setting
Number;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
It is random to generate customer parameter matrix PwWith tag parameter matrix Qw;Wherein, the customer parameter matrix Pw, the mark
Sign parameter matrix QwIdentical multiple users corresponding with the customer parameter matrix P;Customer parameter matrix PwWith tag parameter matrix
QwFor the split-matrix of the customer parameter matrix P;
According to the customer parameter matrix P, the Y update customer parameter matrix PwWith the tag parameter matrix Qw, Y
For the update times for making needed for the second portrait error convergence;The second portrait error is non-in the customer parameter matrix P
Difference between neutral element and corresponding second predicted value, second predicted value are the customer parameter matrix PwWith the mark
Sign parameter matrix QwMerging matrix element;
Wherein, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, y is greater than or equal to 1 and small
In or be equal to Y, including:
As y=1, according to the customer parameter matrix P, initial customer parameter matrix P is generated at randomwAnd tag parameter
Matrix Qw, calculate the second portrait error of nonzero element in user's portrait matrix of the incompleteness;
According to the second portrait error, the customer parameter matrix P is updatedwWith the tag parameter matrix Qw;
When y is greater than or equal to 2, according to the customer parameter matrix P, the customer parameter matrix PwJoin with the label
Matrix number Qw, calculate the second portrait error of the second part user in the multiple user;The of the second part user
Two variation differences are more than the second variation difference of the second remaining users, and second remaining users are removing in the multiple user
User except the second part user;
When y is equal to 2, the second variation difference is to update obtained customer parameter matrix P according to the y-1 timeswAnd mark
Sign parameter matrix QwObtain the second predicted value with according to the initial customer parameter matrix P and tag parameter matrix generated at random
QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained customer parameter matrix P and label according to the y-1 times
Parameter matrix QwThe second predicted value obtained updates obtained customer parameter matrix P with according to the y-2 timeswWith tag parameter matrix
QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the customer parameter matrix P is updatedwAnd institute
State tag parameter matrix Qw;
The obtained customer parameter matrix P is updated by Y timeswWith the tag parameter matrix QwAs w level depth
Customer parameter matrix PwWith tag parameter matrix Qw;
When the w is less than the S, the obtained customer parameter matrix P is updated by Y timeswWant deep as the w+1 times
Spend the customer parameter matrix P decomposed;
The output module 304 is additionally operable to carry out the customer parameter matrix P S times in the matrix decomposition module 303
Deep decomposition, and after obtaining customer parameter matrix and the tag parameter matrix of S level depth, according to the user of the incompleteness
Portrait matrix, the R times newer customer parameter matrix P and tag parameter matrix Q, and the S level obtained are deep
The customer parameter matrix and tag parameter matrix of degree obtain complete user's portrait matrix.
As shown in figure 5, the present invention provides a kind of information processing unit 401, including:Input interface 402,403 and of processor
Output interface 404, the processor are separately connected input interface and output interface.
The processor 403 is used to obtain incomplete user's portrait matrix by input interface 402, and generates institute at random
State the customer parameter matrix P and tag parameter matrix Q of incomplete user's portrait matrix;Wherein, user's portrait square of the incompleteness
Battle array, the customer parameter matrix P and the tag parameter matrix Q correspond to identical multiple users;
And it is used for:Incomplete user's portrait matrix is obtained in the input module, and generates the incompleteness at random
User draw a portrait matrix customer parameter matrix P and tag parameter matrix Q after, according to the user of incompleteness portrait matrix, R times
The customer parameter matrix P and the tag parameter matrix Q are updated, R is the update time so that needed for the first portrait error convergence
Number;It is described first portrait error be incompleteness user portrait matrix between nonzero element and corresponding first predicted value
Difference, first predicted value are the element for merging matrix of the customer parameter matrix P and the tag parameter matrix Q;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than
Or it is equal to R, including:
As r=1, the r times update customer parameter matrix P and tag parameter matrix Q, including:According to institute
Incomplete user's portrait matrix is stated, the initial customer parameter matrix P and tag parameter matrix Q generated at random is calculated described residual
First portrait error of nonzero element in scarce user's portrait matrix;According to the first portrait error, user's ginseng is updated
The matrix number P and tag parameter matrix Q;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, packet
It includes:The r times update customer parameter matrix P and tag parameter matrix Q, including:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, calculate
The first portrait error of first part user in the multiple user;The first variation difference of the first part user is more than
First variation difference of the first remaining users, first remaining users are in the multiple user except the first part uses
User except family;
When r is equal to 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained with according to the initial customer parameter matrix P and institute generated at random
State the difference between the first predicted value that tag parameter matrix Q is obtained;
When r is more than 2, the first variation difference be according to update for the r-1 time the obtained customer parameter matrix P with
The first predicted value that the tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and institute with according to the r-2 times
State the difference between the first predicted value that tag parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and label ginseng are updated
Matrix number Q;
After the processor 403 is used to update the customer parameter matrix P and tag parameter matrix Q at the R times,
According to the user of incompleteness portrait matrix and the R times newer customer parameter matrix P and the tag parameter matrix
Q obtains complete user's portrait matrix.
The processor 403 can export complete user's portrait matrix by output interface 404.
Preferably, the processor 403 is additionally operable to according to the user of incompleteness portrait matrix, the customer parameter square
Battle array the P and tag parameter matrix Q, calculate the first part user in the multiple user first portrait error before, meter
The variation difference of the multiple user is calculated, and the variation difference of the multiple user is ranked up, selects the use of predetermined ratio
As the first part user, the variation difference of the user of the predetermined ratio, which is both greater than in the multiple user, not to be chosen at family
The variation difference of the user selected.
Preferably, the processor 403 is additionally operable to update the customer parameter matrix P and the tag parameter at the R times
After matrix Q, the customer parameter matrix P is subjected to S deep decomposition, and obtain the customer parameter matrix of S level depth
With tag parameter matrix;Described S times is the number set;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
It is random to generate customer parameter matrix PwWith tag parameter matrix Qw;Wherein, the customer parameter matrix Pw, the mark
Sign parameter matrix QwIdentical multiple users corresponding with the customer parameter matrix P;Customer parameter matrix PwWith tag parameter matrix
QwFor the split-matrix of the customer parameter matrix P;
According to the customer parameter matrix P, the Y update customer parameter matrix PwWith the tag parameter matrix Qw, Y
For the update times for making needed for the second portrait error convergence;The second portrait error is non-in the customer parameter matrix P
Difference between neutral element and corresponding second predicted value, second predicted value are the customer parameter matrix PwWith the mark
Sign parameter matrix QwMerging matrix element;
Wherein, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, y is greater than or equal to 1 and small
In or be equal to Y, including:
As y=1, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, including:According to described
Customer parameter matrix P generates initial customer parameter matrix P at randomwWith tag parameter matrix Qw, calculate the user of the incompleteness
Second portrait error of nonzero element in portrait matrix;
According to the second portrait error, the customer parameter matrix P is updatedwWith the tag parameter matrix Qw;
When y is greater than or equal to 2, the y times update customer parameter matrix PwWith the tag parameter matrix Qw, including:
According to the customer parameter matrix P, the customer parameter matrix PwWith the tag parameter matrix Qw, described in calculating
The second portrait error of second part user in multiple users;The second variation difference of the second part user is more than second
Remaining users second variation difference, second remaining users be the multiple user in except the second part user it
Outer user;
When y is equal to 2, the second variation difference is to update obtained customer parameter matrix P according to the y-1 timeswAnd mark
Sign parameter matrix QwObtain the second predicted value with according to the initial customer parameter matrix P and tag parameter matrix generated at random
QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained customer parameter matrix P and label according to the y-1 times
Parameter matrix QwThe second predicted value obtained updates obtained customer parameter matrix P with according to the y-2 timeswWith tag parameter matrix
QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the customer parameter matrix P is updatedwAnd institute
State tag parameter matrix Qw;
The obtained customer parameter matrix P is updated by Y timeswWith the tag parameter matrix QwAs w level depth
Customer parameter matrix PwWith tag parameter matrix Qw;
When the w is less than the S, the obtained customer parameter matrix P is updated by Y timeswWant deep as the w+1 times
Spend the customer parameter matrix P decomposed;
The processor 403 is additionally operable to the customer parameter matrix P carrying out S deep decomposition, and obtains S level
After the customer parameter matrix and tag parameter matrix of depth, according to the user of incompleteness portrait matrix, the R times newer institute
State customer parameter matrix P and the tag parameter matrix Q, and the customer parameter matrix and label of S level depth of acquisition
Parameter matrix obtains complete user's portrait matrix.
One of ordinary skill in the art will appreciate that all or part of step in the various methods of above-described embodiment is can
It is completed with instructing relevant hardware by program, which can be stored in a computer readable storage medium, storage
Medium may include:Read-only memory (ROM, Read Only Memory), random access memory (RAM, Random
Access Memory), disk or CD etc..
Be provided for the embodiments of the invention above it is a kind of acquisition whole user portrait method and device carried out in detail
It introduces, principle and implementation of the present invention are described for specific case used herein, the explanation of above example
It is merely used to help understand the method and its core concept of the present invention;Meanwhile for those skilled in the art, according to the present invention
Thought, there will be changes in the specific implementation manner and application range, in conclusion the content of the present specification should not be understood
For limitation of the present invention.
Claims (6)
1. a kind of method obtaining whole user portrait, which is characterized in that including:
Incomplete user's portrait matrix is obtained, and generates initial customer parameter matrix P and tag parameter matrix Q at random;Its
In, the user of the incompleteness draw a portrait matrix, the customer parameter matrix P and the tag parameter matrix Q correspond to it is identical multiple
User, the customer parameter matrix P and the tag parameter matrix Q are the split-matrix of user's portrait matrix of the incompleteness;
According to the user of incompleteness portrait matrix, the R update customer parameter matrix P and tag parameter the matrix Q, R
For the update times for making needed for the first portrait error convergence;The first portrait error is user's portrait matrix of the incompleteness
Difference between middle nonzero element and corresponding first predicted value, first predicted value are the customer parameter matrix P and institute
State the element of the merging matrix of tag parameter matrix Q;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than or wait
In R, including:
As r=1, according to the user of incompleteness portrait matrix, the initial customer parameter matrix P and label generated at random joins
Matrix number Q calculates the first portrait error of nonzero element in user's portrait matrix of the incompleteness;It is missed according to first portrait
Difference updates the customer parameter matrix P and the tag parameter matrix Q;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, including:
It is drawn a portrait matrix, the customer parameter matrix P and the tag parameter matrix Q according to the user of the incompleteness, described in calculating
The first portrait error of first part user in multiple users;The first variation difference of the first part user is more than first
Remaining users first variation difference, first remaining users be the multiple user in except the first part user it
Outer user;
When r is equal to 2, the first variation difference is according to the r-1 time obtained customer parameter matrix P of update and described
The first predicted value that tag parameter matrix Q is obtained with according to the initial customer parameter matrix P generated at random and the mark
Sign the difference between the first predicted value that parameter matrix Q is obtained;
When r is more than 2, the first variation difference is according to the r-1 time obtained customer parameter matrix P of update and described
The first predicted value that tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and the mark with according to the r-2 times
Sign the difference between the first predicted value that parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and the tag parameter square are updated
Battle array Q;
After updating the customer parameter matrix P and tag parameter matrix Q at the R times, drawn according to the user of the incompleteness
As matrix and the R times newer customer parameter matrix P and tag parameter matrix Q, complete user's portrait is obtained
Matrix.
2. the method according to claim 1 for obtaining whole user portrait, which is characterized in that described in the r times update
Customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than or equal to R, including:
According to the user of incompleteness portrait matrix, the customer parameter matrix P and the tag parameter matrix Q, institute is calculated
Before the first portrait error for stating the first part user in multiple users, the variation difference of the multiple user is calculated, and will
The variation difference of the multiple user is ranked up, and selects the user of predetermined ratio as the first part user, described pre-
The variation difference of the user of certainty ratio is both greater than the variation difference of non-selected user in the multiple user.
3. the method according to claim 1 or 2 for obtaining whole user portrait, which is characterized in that at the R times described in update
After the customer parameter matrix P and tag parameter matrix Q, the method further includes:
The customer parameter matrix P is subjected to S deep decomposition, and obtains the customer parameter matrix and label of S level depth
Parameter matrix;Described S times is the number set;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
The split-matrix P of the customer parameter matrix P is generated at randomwAnd Qw;Wherein, the Pw, the QwWith the customer parameter
Matrix P corresponds to identical multiple users;
According to the customer parameter matrix P, the Y update PwWith the Qw, Y is so that needed for the second portrait error convergence
Update times;It is described second portrait error be the customer parameter matrix P between nonzero element and corresponding second predicted value
Difference, second predicted value be the PwWith the QwMerging matrix element;
Wherein, the y times update PwWith the Qw, y is more than or equal to 1 and is less than or equal to Y, including:
As y=1, according to the customer parameter matrix P, initial P is generated at randomwAnd Qw, calculate user's portrait of the incompleteness
Second portrait error of nonzero element in matrix;
According to the second portrait error, the P is updatedwWith the Qw;
When y is greater than or equal to 2, according to the customer parameter matrix P, the PwWith the Qw, calculate in the multiple user
The second portrait error of second part user;The second variation difference of the second part user is more than the of the second remaining users
Two variation differences, second remaining users are the user in addition to the second part user in the multiple user;
When y is equal to 2, the second variation difference is to update obtained P according to the y-1 timeswAnd QwObtain the second predicted value with
According to the initial P generated at randomwAnd QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained P according to the y-1 timeswAnd QwObtain the second predicted value with
Obtained P is updated according to the y-2 timeswAnd QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the P is updatedwWith the Qw;
The obtained P is updated by Y timeswWith the QwP as w level depthwAnd Qw;
When the w is less than the S, the obtained P is updated by Y timeswCustomer parameter as the w+1 times wanted deep decomposition
Matrix P;
It is described to be drawn a portrait matrix and the R times newer customer parameter matrix P and the label according to the user of the incompleteness
Parameter matrix Q obtains complete user's portrait matrix, including:
The customer parameter matrix P is being subjected to S deep decomposition, and is obtaining the customer parameter matrix and mark of S level depth
After signing parameter matrix, according to the user of incompleteness portrait matrix, the R times newer customer parameter matrix P and described
Tag parameter matrix Q, and the customer parameter matrix and tag parameter matrix of the S level depth that obtain, obtain complete use
Family portrait matrix.
4. a kind of information processing unit, which is characterized in that including:
Input module, user's portrait matrix for obtaining incompleteness, and generate user's portrait matrix of the incompleteness at random
Customer parameter matrix P and tag parameter matrix Q;Wherein, the incompleteness user draw a portrait matrix, the customer parameter matrix P and
The tag parameter matrix Q corresponds to identical multiple users;
Matrix decomposition module is used for:Incomplete user's portrait matrix is obtained in the input module, and is generated at random described residual
After the customer parameter matrix P and tag parameter matrix Q of scarce user's portrait matrix, according to the user of incompleteness portrait square
Battle array, the R update customer parameter matrix P and the tag parameter matrix Q, R are so that needed for the first portrait error convergence
Update times;The first portrait error is nonzero element and corresponding first predicted value in user's portrait matrix of the incompleteness
Between difference, first predicted value be the customer parameter matrix P and the tag parameter matrix Q the member for merging matrix
Element;
Wherein, the r times update customer parameter matrix P and the tag parameter matrix Q, r are greater than or equal to 1 and are less than or wait
In R, including:
As r=1, according to the user of incompleteness portrait matrix, the initial customer parameter matrix P and label generated at random joins
Matrix number Q calculates the first portrait error of nonzero element in user's portrait matrix of the incompleteness;It is missed according to first portrait
Difference updates the customer parameter matrix P and the tag parameter matrix Q;
When r is greater than or equal to 2, the r times update customer parameter matrix P and tag parameter matrix Q, including:
It is drawn a portrait matrix, the customer parameter matrix P and the tag parameter matrix Q according to the user of the incompleteness, described in calculating
The first portrait error of first part user in multiple users;The first variation difference of the first part user is more than first
Remaining users first variation difference, first remaining users be the multiple user in except the first part user it
Outer user;
When r is equal to 2, the first variation difference is according to the r-1 time obtained customer parameter matrix P of update and described
The first predicted value that tag parameter matrix Q is obtained with according to the initial customer parameter matrix P generated at random and the mark
Sign the difference between the first predicted value that parameter matrix Q is obtained;
When r is more than 2, the first variation difference is according to the r-1 time obtained customer parameter matrix P of update and described
The first predicted value that tag parameter matrix Q is obtained updates the obtained customer parameter matrix P and the mark with according to the r-2 times
Sign the difference between the first predicted value that parameter matrix Q is obtained;
According to the first of the first part user the portrait error, the customer parameter matrix P and the tag parameter square are updated
Battle array Q;
Output module, after updating the customer parameter matrix P and tag parameter matrix Q at the R times, according to described
Incomplete user's portrait matrix and the R times newer customer parameter matrix P and tag parameter matrix Q, have obtained
Whole user's portrait matrix.
5. device according to claim 4, which is characterized in that the matrix decomposition module includes dynamic dispatching module, institute
Dynamic dispatching module is stated for joining according to the user of incompleteness portrait matrix, the customer parameter matrix P and the label
Matrix number Q, calculate the first part user in the multiple user first portrait error before, calculate the multiple user's
Change difference, and the variation difference of the multiple user is ranked up, selects the user of predetermined ratio as described first
User, the variation difference of the user of the predetermined ratio is divided to be both greater than the difference in change of non-selected user in the multiple user
Value.
6. device according to claim 4 or 5, which is characterized in that the matrix decomposition module further includes deep decomposition mould
Block, the deep decomposition module are used for:After the customer parameter matrix P and tag parameter matrix Q being updated at the R times,
The customer parameter matrix P is subjected to S deep decomposition, and obtains the customer parameter matrix and tag parameter of S level depth
Matrix;Described S times is the number set;
Wherein, customer parameter matrix P described in the w times deep decomposition, w are greater than or equal to 1 and are less than or equal to S, including:
The split-matrix P of the customer parameter matrix P is generated at randomwAnd Qw;Wherein, the Pw, the QwWith the customer parameter
Matrix P corresponds to identical multiple users;
According to the customer parameter matrix P, the Y update PwWith the Qw, Y is so that needed for the second portrait error convergence
Update times;It is described second portrait error be the customer parameter matrix P between nonzero element and corresponding second predicted value
Difference, second predicted value be the PwWith the QwMerging matrix element;
Wherein, the y times update PwWith the Qw, y is more than or equal to 1 and is less than or equal to Y, including:
As y=1, according to the customer parameter matrix P, initial P is generated at randomwAnd Qw, calculate user's portrait of the incompleteness
Second portrait error of nonzero element in matrix;
According to the second portrait error, the P is updatedwWith the Qw;
When y is greater than or equal to 2, according to the customer parameter matrix P, the PwWith the Qw, calculate in the multiple user
The second portrait error of second part user;The second variation difference of the second part user is more than the of the second remaining users
Two variation differences, second remaining users are the user in addition to the second part user in the multiple user;
When y is equal to 2, the second variation difference is to update obtained P according to the y-1 timeswAnd QwObtain the second predicted value with
According to the initial P generated at randomwAnd QwDifference between the second predicted value obtained;
When y is more than 2, the second variation difference is to update obtained P according to the y-1 timeswAnd QwObtain the second predicted value with
Obtained P is updated according to the y-2 timeswAnd QwDifference between the second predicted value obtained;
According to the second of the second part user nonzero element the portrait error, the P is updatedwWith the Qw;
The obtained P is updated by Y timeswWith the QwP as w level depthwAnd Qw;
When the w is less than the S, the obtained P is updated by Y timeswCustomer parameter as the w+1 times wanted deep decomposition
Matrix P;
The output module is additionally operable to the customer parameter matrix P carrying out S deep decomposition, and obtains S level depth
Customer parameter matrix and tag parameter matrix after, according to the user of the incompleteness draw a portrait matrix, the R times newer use
The family parameter matrix P and tag parameter matrix Q, and obtain S level depth customer parameter matrix and tag parameter
Matrix obtains complete user's portrait matrix.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310745711.7A CN104750731B (en) | 2013-12-30 | 2013-12-30 | A kind of method and device obtaining whole user portrait |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310745711.7A CN104750731B (en) | 2013-12-30 | 2013-12-30 | A kind of method and device obtaining whole user portrait |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104750731A CN104750731A (en) | 2015-07-01 |
CN104750731B true CN104750731B (en) | 2018-09-21 |
Family
ID=53590430
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310745711.7A Active CN104750731B (en) | 2013-12-30 | 2013-12-30 | A kind of method and device obtaining whole user portrait |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104750731B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106503015A (en) * | 2015-09-07 | 2017-03-15 | 国家计算机网络与信息安全管理中心 | A kind of method for building user's portrait |
CN105893406A (en) * | 2015-11-12 | 2016-08-24 | 乐视云计算有限公司 | Group user profiling method and system |
CN107341679A (en) * | 2016-04-29 | 2017-11-10 | 腾讯科技(深圳)有限公司 | Obtain the method and device of user's portrait |
CN106489159A (en) * | 2016-06-29 | 2017-03-08 | 深圳狗尾草智能科技有限公司 | A kind of user's portrait based on deep neural network represents learning system and method |
CN107644047B (en) * | 2016-07-22 | 2021-01-15 | 华为技术有限公司 | Label prediction generation method and device |
CN107480271B (en) * | 2017-08-18 | 2020-09-18 | 晶赞广告(上海)有限公司 | Crowd image drawing method and system based on sampling search and index search |
CN108846097B (en) * | 2018-06-15 | 2021-01-29 | 北京搜狐新媒体信息技术有限公司 | User interest tag representation method, article recommendation device and equipment |
CN111322716B (en) * | 2020-02-24 | 2021-08-03 | 青岛海尔工业智能研究院有限公司 | Air conditioner temperature automatic setting method, air conditioner, equipment and storage medium |
CN111523026B (en) * | 2020-04-15 | 2023-10-17 | 咪咕文化科技有限公司 | User portrait updating method, system, network equipment and storage medium |
CN111522828B (en) * | 2020-04-23 | 2023-08-01 | 中国农业银行股份有限公司 | User portrait tag value analysis method and device |
CN111931107B (en) * | 2020-07-31 | 2024-03-22 | 博泰车联网科技(上海)股份有限公司 | Digital citizen system construction method, system and storage medium |
CN117440192B (en) * | 2023-12-21 | 2024-02-23 | 辽宁云科智造产业技术研究院有限公司 | User demand analysis method and system based on intelligent cloud service platform |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103345474A (en) * | 2013-07-25 | 2013-10-09 | 苏州大学 | Method for online tracking of document theme |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090112859A1 (en) * | 2007-10-25 | 2009-04-30 | Dehlinger Peter J | Citation-based information retrieval system and method |
-
2013
- 2013-12-30 CN CN201310745711.7A patent/CN104750731B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103345474A (en) * | 2013-07-25 | 2013-10-09 | 苏州大学 | Method for online tracking of document theme |
Non-Patent Citations (1)
Title |
---|
主题模型的在线消息传递算法研究;叶芸;《中国优秀硕士学位论文全文数据库信息科技辑》;20131115(第11期);第I138-975页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104750731A (en) | 2015-07-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104750731B (en) | A kind of method and device obtaining whole user portrait | |
US20220414544A1 (en) | Parallel Development and Deployment for Machine Learning Models | |
CN105205180B (en) | A kind of evaluation method and device of knowledge mapping | |
CN107688493B (en) | Method, device and system for training deep neural network | |
US11443217B2 (en) | Information processing apparatus and information processing method | |
Ludl et al. | A complete survey of texture zeros in the lepton mass matrices | |
CN110366734A (en) | Optimization neural network framework | |
KR20180091043A (en) | Method and apparatus for obtaining user portraits | |
Jesus et al. | Newtonian perturbations on models with matter creation | |
WO2019084560A1 (en) | Neural architecture search | |
CN103327092A (en) | Cell discovery method and system on information networks | |
CN107229966A (en) | A kind of model data update method, apparatus and system | |
CN106373112A (en) | Image processing method, image processing device and electronic equipment | |
Andreassen et al. | Parameter estimation using neural networks in the presence of detector effects | |
US20230051237A1 (en) | Determining material properties based on machine learning models | |
CN106250110A (en) | Set up the method and device of model | |
JP6291041B2 (en) | Search recommendation method and apparatus | |
Li et al. | Trajectory scheduling methods for minimizing total tardiness in a flowshop | |
US20210149985A1 (en) | Method and apparatus for processing large-scale distributed matrix product | |
CN106355301A (en) | Method and device for determining workflow template | |
CN116962438A (en) | Gradient data synchronization method, system, electronic equipment and readable storage medium | |
Ilten | Pythia 8: Simulating Tau-Lepton Decays | |
JP4362594B2 (en) | State transition analysis method and apparatus in pseudo-random number generation mechanism and period length measurement apparatus | |
CN105227635B (en) | cloud data matching method and device | |
JP5854924B2 (en) | Power consumption estimation apparatus, electronic device, power consumption estimation method, and program |
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 |