CN109410080A - A kind of social image recommended method based on level attention mechanism - Google Patents
A kind of social image recommended method based on level attention mechanism Download PDFInfo
- Publication number
- CN109410080A CN109410080A CN201811200827.1A CN201811200827A CN109410080A CN 109410080 A CN109410080 A CN 109410080A CN 201811200827 A CN201811200827 A CN 201811200827A CN 109410080 A CN109410080 A CN 109410080A
- Authority
- CN
- China
- Prior art keywords
- user
- image
- matrix
- social
- formula
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 239000011159 matrix material Substances 0.000 claims abstract description 100
- 239000013598 vector Substances 0.000 claims description 53
- 238000013527 convolutional neural network Methods 0.000 claims description 11
- 238000003780 insertion Methods 0.000 claims description 11
- 230000037431 insertion Effects 0.000 claims description 11
- 230000000007 visual effect Effects 0.000 claims description 8
- 230000000153 supplemental effect Effects 0.000 claims description 7
- 230000005284 excitation Effects 0.000 claims description 6
- 241000208340 Araliaceae Species 0.000 claims description 3
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 3
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 3
- 235000008434 ginseng Nutrition 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 239000000463 material Substances 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 3
- 238000005295 random walk Methods 0.000 claims description 3
- 238000012549 training Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 description 6
- 230000003542 behavioural effect Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000003475 lamination Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 210000003813 thumb Anatomy 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Information Transfer Between Computers (AREA)
Abstract
The invention discloses a kind of social image recommended methods based on level attention mechanism, comprising: 1. stereo isomers data: user is to the rating matrix of image, user to the social networks matrix between the upload information matrix of image, user and user;2. pair isomeric data and image set handle to obtain social embeded matrix, content embeded matrix, style embeded matrix;3. embeded matrix is input in bottom attention network and obtains social semantic information;4. passing through three aspects that top layer attention network obtains social semantic information weight shared when influencing user and selecting image;5. according to eigenmatrix, three aspect social semantic information matrix and the weight computing shared by them obtain score in predicting value, to carry out image recommendation.The present invention does not solve data sparsity problem effectively merely with the social information between the upload information of user and user, more explains user preferences well by level attention mechanism, realizes accurately social image recommendation.
Description
Technical field
The present invention relates to image recommendation fields, are specifically a kind of social image recommendation sides based on level attention
Method.
Technical background
Social networks based on image has become most popular social networks in recent years.With mobile phone user
It quicklys increase, many users, which shoot image and upload in social network-i i-platform, shares their lives.The image largely uploaded is led
Image overload is caused, how understanding the hobby of different user and carrying out accurately image recommendation to it becomes urgent need.
The problem of collaborative filtering can be with effective solution image recommendation, it passes through the digging to user's history behavioral data
Pick, finds the preference of user and recommends.Although collaborative filtering is widely used, but user-image interbehavior square
The sparsity of battle array limits its recommendation performance.In order to solve the data sparsity problem of collaborative filtering, RuiningHe in 2016
Et al. propose VBPR algorithm, the visual signature of image is extracted by deep neural network CNN, the visual information of image is embedded in
Into existing collaborative filtering, the inherent attribute of article has been excavated sufficiently to improve recommendation precision.
But the inherent attribute that the collaborative filtering method based on Image Visual Feature only accounts for article does not use
The inherent attribute of user.One important feature of social image platform is exactly that there are social networks between different user, therefore
When user shows the preference to image in social platform, social networks can be embedded into collaborative filtering and mitigate number
System recommendation performance is improved according to sparsity.
Although people, which increasingly pay close attention to, combines carry out image recommendation for various auxiliary embedding datas, how more preferable
Ground is improved using the feature of social image platform in comprehensive manner recommends performance still in exploration.In addition to improving social image
Recommend how to excavate the hobby of different user except precision, so that making reasonable explain to recommendation results is also to need to solve
The problem of.
Summary of the invention
The present invention to solve the deficiencies in the prior art place, propose that a kind of social image based on level attention mechanism pushes away
Recommend method, to can social networks sufficiently between user, user to the behavioral data of image and the insertion of Image Visual Feature
Effective solution data sparsity problem, to improve social recommendation precision;User is influenced to image selection by definition simultaneously
Three aspects semantic information, and illustrate user to the influence power of image selection, thus very using level attention mechanism
The good interpretation for solving the problems, such as social image recommendation.
The present invention adopts the following technical scheme that in order to solve the technical problem
A kind of the characteristics of social image recommended method based on level attention mechanism of the invention is to carry out as follows:
Step 1, stereo isomers data, comprising: user is to the rating matrix R of image, user to the upload information square of image
Social networks matrix S between battle array L, user and user:
U is enabled to indicate user's collection, and U={ u1,...,ua,...,ub,...,uM, uaIndicate a-th of user, ubIndicate b
A user, M indicate total number of users, 1≤a, b≤M;V is enabled to indicate image set, and V={ v1,...,vi,...,vj,...,vN, vi
Indicate i-th of image, vjIndicate that j-th of image, N indicate total number of images, 1≤i, j≤N;Enable RaiIndicate a-th of user uaTo
I image viScore value, then rating matrix R={ R of the user to imageai}M×N;Enable LaiIndicate a-th of user uaTo i-th
Image viUpload information, then upload information matrix L={ L of the user to imageai}M×N;Enable SabIndicate a-th of user uaTo b
The u of a userbSocial networks, then the social networks matrix S={ S between user and userab}M×M;
Step 2 handles the isomeric data and image set V to obtain social embeded matrix E, content embeded matrix FcAnd wind
Lattice embeded matrix Fs:
Step 2.1 is handled social networks matrix S by Random Walk Algorithm, obtains social embeded matrix E;
Step 2.2, the visual signature that image is extracted by depth convolutional neural networks model VGG:
The output Fc of the last one full articulamentum of the depth convolutional neural networks model VGG is chosen as i-th of figure
As viContent be embedded in vector Fi c, to obtain the content embeded matrix F of all imagesc;
Assuming that m-th of convolutional layer has NmA filter, each filter size are Mm, then depth convolutional neural networks
The Gram matrix of m-th of convolutional layer feature of model VGGM=1~5;
Feature Mapping relationship between m-th of convolutional layer, p-th of filter and q-th of filter is calculated using formula (1)
In formula (1),Indicate the excitation value of p-th of m-th of convolutional layer, k-th of filter position,Indicate m-th volume
The excitation value of q-th of lamination, k-th of filter position, 1≤p, q≤Nm, 1≤k≤Mm;
Vector F is embedded in using the style that formula (2) obtains i-th of imagei s, to obtain the style embeded matrix of all images
Fs:
Fi s=[v (G1),...,v(Gl),...,v(G5)] (2)
In formula (2), v (Gm) indicate m-th of convolutional layer feature Gram matrix GlVector quantization;
Step 3, by social embeded matrix E, content embeded matrix FcWith style embeded matrix FsIt is input to bottom attention net
Obtain influencing the social semantic information for three aspects that user selects picture in network:
Step 3.1, the upload consistency for calculating user
Step 3.1.1, L is enabledaIndicate a-th of user uaThe image set of upload then obtains j-th of image v using formula (3)j?
Calculate a-th of user uaUpload consistency when shared weight αaj:
In formula (3), σ () indicates sigmoid function;PaAnd QaIt is a-th of user u respectivelyaFoundation characteristic vector sum it is auxiliary
Help feature vector, WjAnd XjIt is j-th of image v respectivelyjFoundation characteristic vector sum supplemental characteristic vector;EaIt is a-th of user ua
Social insertion vector;WithIt is j-th of image v respectivelyjContent insertion vector sum style be embedded in vector;θu=[w2,
W1,Wc,Ws] be the first bottom attention network B otNet1 parameter, wherein w2It is the parameter of first sigmoid function, W1It is
First matrix parameter, WcIt is to content embeded matrix FcCarry out the parameter of dimensionality reduction, WsIt is to style embeded matrix FsCarry out dimensionality reduction
Parameter;It is a-th of user uaThe picture material embeded matrix liked, and obtained by formula (4);It is a-th of user ua
The image style embeded matrix liked, and obtained by formula (5);
Step 3.1.2, to the jth image vjCalculating a-th of user uaUpload consistency when shared weight
αajIt is normalized, the weight after being normalized by formula (6)
Step 3.1.3, a-th of user u is obtained using formula (7)aUpload consistency
Step 3.2 calculates a-th of user uaSocial influence power
Step 3.2.1, S is enabledaIt indicates and a-th of user uaThere is the user of social networks to collect, is then obtained b-th using formula (8)
User ubCalculating a-th of user uaSocial influence power when shared weight βab:
βab=σ (w4×(W3×[Pa,Pb,Qa,Qb,Ea,Eb,Fa c,Fa s])) (8)
In formula (7), PbAnd QbIt is b-th of user u respectivelybFoundation characteristic vector sum supplemental characteristic vector;EbIt is b-th
User ubSocial insertion vector;WithIt is a-th of user u respectivelyaThe image set liked content embeded matrix and wind
Lattice embeded matrix;θs=[w4,W3] be the second bottom attention network B otNet2 parameter, wherein w4It is second sigmoid
The parameter of function, W3It is second matrix parameter;
Step 3.2.2, to b-th of user ubCalculating a-th of user uaSocial influence power when shared weight
βabIt is normalized, the weight after being normalized by formula (9)
Step 3.2.3, a-th of user u is obtained according to formula (10)aSocial influence power
Step 3.3 enables CiIndicate i-th of image viUploader, and each image only one uploader, thus by i-th
A image viUploader influence power be expressed as i-th of image viUploader CiAuxiliary be embedded in vector QCi;
Using the upload consistency, social influence power and uploader influence power as the social semantic information of three aspects;
Step 4, three input A that top layer attention network TopNet is respectively obtained using formula (11)λ, λ=1,2,3:
A-th of user u is being influenced respectively using the social semantic information that formula (12) obtains three aspectsaSelect image when institute
The weight γ accounted foraλ:
γaλ=σ (w6×(W5×Aλ)) (12)
In formula (12), θa=[w6,W5] be top layer attention network parameter, w6It is the ginseng of third sigmoid function
Number, W5It is third matrix parameter;
Step 5, according to three aspect social semantic information and its shared weight computing user it is pre- to the scoring of image
Measured value, to carry out image recommendation to user:
Step 5.1 obtains a-th of user u according to formula (13)aWith i-th of image viCorresponding prediction scoringTo
To user to the score in predicting matrix of article
In formula (13),It is i-th of image viFoundation characteristic vector transposed vector;
Step 5.2 establishes the loss function L (θ) as shown in formula (14):
In formula (13), θ=[θ1,θ2] it is parameter to be optimized, θ1=[P, Q, W, X] is eigenmatrix, θ2=[θu,θs,
θa] be attention network parameter, λ is regularization term, Da=(i, j) | i ∈ Ra∧j∈V-RaIt is a-th of user uaTraining
Data, RaIt is a-th of user uaThe image set liked;
Step 5.3 optimizes the loss function L (θ) by stochastic gradient descent method, so that L (θ) reaches
Minimum to obtain optimum prediction rating matrix, and carries out image recommendation to user according to institute's optimum prediction rating matrix.
Compared with the prior art, the invention has the advantages that:
1, the present invention devise a level attention network by between Image Visual Feature, user social networks and user
Behavioral data (upload behavior and thumb up behavior) fusion, alleviate data sparsity problem, effectively increase image and push away
It recommends precision and explains the preference of user well.
2, the present invention constructs the isomeric data of social image platform by step 1, to greatly alleviate recommendation system
Sparse Problem in system.
3, the present invention is not only extracted the content information of image by depth convolutional neural networks, is also extracted the wind of image
Lattice information, because the style of image is also to influence a key factor of user's selection, the insertion of image style matrix is significantly mentioned
The high precision of social image recommendation.
4, invention defines influence users to the social semantic information of three aspects of image selection, and as one
A important feature is embedded into the inherent attribute for going sufficiently to have excavated user in user interest vector, improves recommendation precision.
5, the present invention devises a level attention network to model to social image proposed algorithm;Bottom pays attention to
The power network struction social semantic information of user, top layer attention network obtain these three in terms of social semantic information in shadow
Weight shared when different user selection image is rung, to solve the problems, such as the interpretation of social platform image very well.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of the social image recommended method of level attention mechanism;
Fig. 2 a is the HR experimental result picture of the method for the present invention and the 6 kinds of control methods selected on F_S data set;
Fig. 2 b is the NDCG experimental result picture of the method for the present invention and the 6 kinds of control methods selected on F_S data set;
Fig. 3 a is the HR experimental result picture of the method for the present invention and the 6 kinds of control methods selected on F_L data set;
Fig. 3 b is the NDCG experimental result picture of the method for the present invention and the 6 kinds of control methods selected on F_L data set.
Specific embodiment
Referring to Fig. 1, in the present embodiment, a kind of social image recommended method based on level attention mechanism is by following step
It is rapid to carry out:
Step 1, stereo isomers data, comprising: user is to the rating matrix R of image, user to the upload information square of image
Social networks matrix S between battle array L, user and user:
U is enabled to indicate user's collection, and U={ u1,...,ua,...,ub,...,uM, uaIndicate a-th of user, ubIndicate b
A user, M indicate total number of users, 1≤a, b≤M;V is enabled to indicate image set, and V={ v1,...,vi,...,vj,...,vN, vi
Indicate i-th of image, vjIndicate that j-th of image, N indicate total number of images, 1≤i, j≤N;Enable RaiIndicate a-th of user uaTo
The v of i imageiScore value, then rating matrix R={ R of the user to imageai}M×N;If a-th of user uaTo i-th image
viThere is scoring, then Rai=1, otherwise Rai=0.Enable LaiIndicate a-th of user uaTo the v of i-th of imageiUpload information, then use
Upload information matrix L={ L of the family to imageai}M×N;If a-th of user uaUpload i-th image vi, then Lai=1, otherwise
Lai=0.Enable SabIndicate a-th of user uaTo the u of b-th of userbSocial networks, then social networks between user and user
Matrix S={ Sab}M×M.If a-th of user uaTo the u of b-th of userbSocial networks, then Sab=1, otherwise Sab=0.
Step 2 handles isomeric data and image set V to obtain social embeded matrix E, content embeded matrix FcIt is embedding with style
Enter matrix Fs:
Step 2.1 passes through Random Walk Algorithm to social networks matrix SM×MIt is handled, obtains social embeded matrix Ed ×M, d=20;
Step 2.2, the visual signature that image is extracted by depth convolutional neural networks model VGG:
4096 dimensional vectors of the last one full articulamentum output of depth convolutional neural networks model VGG19 are chosen as the
I image viContent be embedded in vector Fi c, to obtain the content embeded matrix F of all imagesc;
Assuming that m-th of convolutional layer has NmA filter, each filter size are Mm, then depth convolutional neural networks model
The Gram matrix of m-th of convolutional layer feature of VGGM=1~5;
Feature Mapping relationship between m-th of convolutional layer, p-th of filter and q-th of filter is calculated using formula (1)
In formula (1),Indicate the excitation value of p-th of m-th of convolutional layer, k-th of filter position,Indicate m-th volume
The excitation value of q-th of lamination, k-th of filter position, 1≤p, q≤Nm, 1≤k≤Mm;
Vector is embedded in using the style that formula (2) obtains i-th of imageTo obtain the style embeded matrix of all images
Fs:
In formula (2), v (Gm) indicate m-th of convolutional layer feature Gram matrix GlVector quantization;
Step 3, by social embeded matrix E, content embeded matrix FcWith style embeded matrix FsIt is input to bottom attention net
Obtain influencing the social semantic information for three aspects that user selects picture in network:
Step 3.1, the upload consistency for calculating user
Step 3.1.1, L is enabledaIndicate a-th of user uaThe image set of upload then obtains j-th of image v using formula (3)j?
Calculate a-th of user uaUpload consistency when shared weight αaj:
In formula (3), σ () indicates sigmoid function;PaAnd QaIt is all the d dimensional vector of Gaussian distributed, difference table
Show a-th of user uaFoundation characteristic vector sum supplemental characteristic vector, WjAnd XjIt is all the d dimensional vector of Gaussian distributed, point
J-th of image v is not indicatedjFoundation characteristic vector sum supplemental characteristic vector;EaIt is a column of social embeded matrix E, indicates a
A user uaSocial insertion vector;It is content embeded matrix FcJth column, indicate j-th of image vjContent be embedded in
Amount,It is style embeded matrix FsJth column, indicate j-th of image vjStyle be embedded in vector;θu=[w2,W1,Wc,Ws] be
The parameter of first bottom attention network B otNet1, wherein w2It is the parameter of first sigmoid function, W1It is first matrix
Parameter, WcIt is to content embeded matrix FcCarry out the parameter of dimensionality reduction, WsIt is to style embeded matrix FsCarry out the parameter of dimensionality reduction;
It is a-th of user uaThe picture material embeded matrix liked, and obtained by formula (4);It is a-th of user uaThe image liked
Style embeded matrix, and obtained by formula (5);
Step 3.1.2, to jth image vjCalculating a-th of user uaUpload consistency when shared weight αajInto
Row normalized, the weight after being normalized by formula (6)
Step 3.1.3, a-th of user u is obtained using formula (7)aUpload consistency
Step 3.2 calculates a-th of user uaSocial influence power
Step 3.2.1, S is enabledaIt indicates and a-th of user uaThere is the user of social networks to collect, is then obtained b-th using formula (8)
User ubCalculating a-th of user uaSocial influence power when shared weight βab:
In formula (7), PbAnd QbAll it is the d dimensional vector of Gaussian distributed, respectively indicates b-th of user ubBasis it is special
Levy vector sum supplemental characteristic vector;EbIt is the b column of social embeded matrix E, indicates b-th of user ubSocial insertion vector;θs
=[w4,W3] be the second bottom attention network B otNet2 parameter, wherein w4It is the parameter of second sigmoid function, W3
It is second matrix parameter;
Step 3.2.2, to b-th of user ubCalculating a-th of user uaSocial influence power when shared weight βabInto
Row normalized, the weight after being normalized by formula (9)
Step 3.2.3, a-th of user u is obtained according to formula (10)aSocial influence power
Step 3.3 enables CiIndicate i-th of image viUploader, and each image only one uploader, thus by i-th
A image viUploader influence power be expressed as i-th of image viUploader CiAuxiliary embeded matrix QCi, QCiIt is to obey height
The d dimensional vector of this distribution;
To upload consistency, social influence power and uploader influence power as the social semantic information of three aspects;
Step 4, three input A that top layer attention network TopNet is respectively obtained using formula (11)λ, λ=1,2,3:
A-th of user u is being influenced respectively using the social semantic information that formula (12) obtains three aspectsaSelect image when institute
The weight γ accounted foraλ:
γaλ=σ (w6×(W5×Aλ)) (12)
In formula (12), θa=[w6,W5] be top layer attention network parameter, w6It is the ginseng of third sigmoid function
Number, W5It is third matrix parameter;
Step 5, according to three aspect social semantic information and its shared weight computing user it is pre- to the scoring of image
Measured value, to carry out image recommendation to user:
Step 5.1 obtains a-th of user u according to formula (13)aWith i-th of image viCorresponding prediction scoringTo
To user to the score in predicting matrix of article
In formula (13),It is the d dimension row vector of Gaussian distributed, indicates i-th of image viFoundation characteristic vector Wi
Transposed vector;
Step 5.2 establishes the loss function L (θ) as shown in formula (14):
In formula (13), θ=[θ1,θ2] it is parameter to be optimized, θ1=[P, Q, W, X] is eigenmatrix, θ2=[θu,θs,
θa] be attention network parameter, λ is regularization term, Da=(i, j) | i ∈ Ra∧j∈V-RaIt is a-th of user uaTraining
Data, RaIt is a-th of user uaThe image set liked;
Step 5.3 optimizes loss function L (θ) by stochastic gradient descent method, so that L (θ) reaches most
It is small, to obtain optimum prediction rating matrix, and image recommendation is carried out to user according to institute's optimum prediction rating matrix.
Embodiment:
In order to verify the validity of this method, the present invention has grabbed a large amount of figure from social image sharing platform Flickr
As being used as data set, it is from widely used NUS-WIDE data set extension.The social link of screening is less than 2, scoring
Record is less than 2 user, and filtered data set is known as F_L data set.Filter F_L data set further to ensure each use
Family and every image at least 10 records, to obtain F_S data set.
The present invention uses Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG)
As evaluation criterion.Selected 6 kinds of methods come with method proposed in this paper carry out effect comparison, be respectively BPR, VBPR,
ACF, SR, ContextMF and VPOI.Specifically, it can be obtained a result as shown in Fig. 2 a, Fig. 2 b according to experimental result, experimental result
Being shown in method proposed by the present invention on data set F_L will be good than 6 kinds of method effects of selection.As shown in Figure 3a, 3b, real
It tests result and is again shown in method proposed by the present invention on data set F_L and be superior to other on two evaluation criterions HR and NDCG
Method, to demonstrate the feasibility of proposition method of the present invention.
Claims (1)
1. a kind of social image recommended method based on level attention mechanism, it is characterized in that carrying out as follows:
Step 1, stereo isomers data, comprising: user to the rating matrix R of image, user to the upload information matrix L of image,
Social networks matrix S between user and user:
U is enabled to indicate user's collection, and U={ u1,...,ua,...,ub,...,uM, uaIndicate a-th of user, ubIndicate b-th of use
Family, M indicate total number of users, 1≤a, b≤M;V is enabled to indicate image set, and V={ v1,...,vi,...,vj,...,vN, viIt indicates
I-th of image, vjIndicate that j-th of image, N indicate total number of images, 1≤i, j≤N;Enable RaiIndicate a-th of user uaTo i-th
Image viScore value, then rating matrix R={ R of the user to imageai}M×N;Enable LaiIndicate a-th of user uaTo i-th of image
viUpload information, then upload information matrix L={ L of the user to imageai}M×N;Enable SabIndicate a-th of user uaB-th is used
The u at familybSocial networks, then the social networks matrix S={ S between user and userab}M×M;
Step 2 handles the isomeric data and image set V to obtain social embeded matrix E, content embeded matrix FcIt is embedding with style
Enter matrix Fs:
Step 2.1 is handled social networks matrix S by Random Walk Algorithm, obtains social embeded matrix E;
Step 2.2, the visual signature that image is extracted by depth convolutional neural networks model VGG:
The output Fc of the last one full articulamentum of the depth convolutional neural networks model VGG is chosen as i-th of image vi's
Content is embedded in vector Fi c, to obtain the content embeded matrix F of all imagesc;
Assuming that m-th of convolutional layer has NmA filter, each filter size are Mm, then the depth convolutional neural networks model
The Gram matrix of m-th of convolutional layer feature of VGG
Feature Mapping relationship between m-th of convolutional layer, p-th of filter and q-th of filter is calculated using formula (1)
In formula (1),Indicate the excitation value of p-th of m-th of convolutional layer, k-th of filter position,Indicate m-th of convolutional layer
The excitation value of q-th of filter, k-th of position, 1≤p, q≤Nm, 1≤k≤Mm;
Vector F is embedded in using the style that formula (2) obtains i-th of imagei s, to obtain the style embeded matrix F of all imagess:
Fi s=[v (G1),...,v(Gl),...,v(G5)] (2)
In formula (2), v (Gm) indicate m-th of convolutional layer feature Gram matrix GlVector quantization;
Step 3, by social embeded matrix E, content embeded matrix FcWith style embeded matrix FsIt is input in bottom attention network
Obtain influencing the social semantic information for three aspects that user selects picture:
Step 3.1, the upload consistency for calculating user
Step 3.1.1, L is enabledaIndicate a-th of user uaThe image set of upload then obtains j-th of image v using formula (3)jIt is calculating
A-th of user uaUpload consistency when shared weight αaj:
In formula (3), σ () indicates sigmoid function;PaAnd QaIt is a-th of user u respectivelyaFoundation characteristic vector sum assist it is special
Levy vector, WjAnd XjIt is j-th of image v respectivelyjFoundation characteristic vector sum supplemental characteristic vector;EaIt is a-th of user uaSociety
Hand over insertion vector;WithIt is j-th of image v respectivelyjContent insertion vector sum style be embedded in vector;θu=[w2,W1,Wc,
Ws] be the first bottom attention network B otNet1 parameter, wherein w2It is the parameter of first sigmoid function, W1It is first
A matrix parameter, WcIt is to content embeded matrix FcCarry out the parameter of dimensionality reduction, WsIt is to style embeded matrix FsCarry out the ginseng of dimensionality reduction
Number;It is a-th of user uaThe picture material embeded matrix liked, and obtained by formula (4);It is a-th of user uaLike
Image style embeded matrix, and by formula (5) obtain;
Step 3.1.2, to the jth image vjCalculating a-th of user uaUpload consistency when shared weight αajIt carries out
Normalized, the weight after being normalized by formula (6)
Step 3.1.3, a-th of user u is obtained using formula (7)aUpload consistency
Step 3.2 calculates a-th of user uaSocial influence power
Step 3.2.1, S is enabledaIt indicates and a-th of user uaThere is the user of social networks to collect, then obtains b-th of user using formula (8)
ubCalculating a-th of user uaSocial influence power when shared weight βab:
βab=σ (w4×(W3×[Pa,Pb,Qa,Qb,Ea,Eb,Fa c,Fa s])) (8)
In formula (7), PbAnd QbIt is b-th of user u respectivelybFoundation characteristic vector sum supplemental characteristic vector;EbIt is b-th of user ub
Social insertion vector;WithIt is a-th of user u respectivelyaThe image set liked content embeded matrix and style insertion
Matrix;θs=[w4,W3] be the second bottom attention network B otNet2 parameter, wherein w4It is second sigmoid function
Parameter, W3It is second matrix parameter;
Step 3.2.2, to b-th of user ubCalculating a-th of user uaSocial influence power when shared weight βabIt carries out
Normalized, the weight after being normalized by formula (9)
Step 3.2.3, a-th of user u is obtained according to formula (10)aSocial influence power
Step 3.3 enables CiIndicate i-th of image viUploader, and each image only one uploader, thus by i-th of figure
As viUploader influence power be expressed as i-th of image viUploader CiAuxiliary be embedded in vector QCi;
Using the upload consistency, social influence power and uploader influence power as the social semantic information of three aspects;
Step 4, three input A that top layer attention network TopNet is respectively obtained using formula (11)λ, λ=1,2,3:
A-th of user u is being influenced respectively using the social semantic information that formula (12) obtains three aspectsaIt is shared when selection image
Weight γaλ:
γaλ=σ (w6×(W5×Aλ)) (12)
In formula (12), θa=[w6,W5] be top layer attention network parameter, w6It is the parameter of third sigmoid function, W5It is
Third matrix parameter;
Step 5, according to the social semantic informations of three aspects and its shared weight computing user to the score in predicting value of image,
To carry out image recommendation to user:
Step 5.1 obtains a-th of user u according to formula (13)aWith i-th of image viCorresponding prediction scoringTo be used
Score in predicting matrix of the family to article
In formula (13), Wi TIt is i-th of image viFoundation characteristic vector transposed vector;
Step 5.2 establishes the loss function L (θ) as shown in formula (14):
In formula (13), θ=[θ1,θ2] it is parameter to be optimized, θ1=[P, Q, W, X] is eigenmatrix, θ2=[θu,θs,θa] it is note
The parameter of meaning power network, λ is regularization term, Da=(i, j) | i ∈ Ra∧j∈V-RaIt is a-th of user uaTraining data, Ra
It is a-th of user uaThe image set liked;
Step 5.3 optimizes the loss function L (θ) by stochastic gradient descent method, so that L (θ) reaches most
It is small, to obtain optimum prediction rating matrix, and image recommendation is carried out to user according to institute's optimum prediction rating matrix.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811200827.1A CN109410080A (en) | 2018-10-16 | 2018-10-16 | A kind of social image recommended method based on level attention mechanism |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811200827.1A CN109410080A (en) | 2018-10-16 | 2018-10-16 | A kind of social image recommended method based on level attention mechanism |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109410080A true CN109410080A (en) | 2019-03-01 |
Family
ID=65467204
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811200827.1A Pending CN109410080A (en) | 2018-10-16 | 2018-10-16 | A kind of social image recommended method based on level attention mechanism |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109410080A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110148424A (en) * | 2019-05-08 | 2019-08-20 | 北京达佳互联信息技术有限公司 | Method of speech processing, device, electronic equipment and storage medium |
CN110245285A (en) * | 2019-04-30 | 2019-09-17 | 中国科学院信息工程研究所 | A kind of personalized recommendation method based on Heterogeneous Information network |
CN110555132A (en) * | 2019-08-13 | 2019-12-10 | 北京工业大学 | Noise reduction self-encoder recommendation method based on attention model |
CN112700342A (en) * | 2020-12-29 | 2021-04-23 | 四川大学 | Personalized grading prediction model and prediction method for social media information authenticity |
CN112818146A (en) * | 2021-01-26 | 2021-05-18 | 山西三友和智慧信息技术股份有限公司 | Recommendation method based on product image style |
CN112905894A (en) * | 2021-03-24 | 2021-06-04 | 合肥工业大学 | Collaborative filtering recommendation method based on enhanced graph learning |
CN113010796A (en) * | 2021-04-15 | 2021-06-22 | 重庆大学 | Method for item recommendation |
CN113205869A (en) * | 2021-04-29 | 2021-08-03 | 华东师范大学 | Mental health detection and intervention method based on user social software head portrait |
CN113658013A (en) * | 2021-08-16 | 2021-11-16 | 上海花千树信息科技有限公司 | Method and device for recommending self-photographing on love or social software |
CN114677234A (en) * | 2022-04-26 | 2022-06-28 | 河南大学 | Graph convolution neural network social contact recommendation method and system integrating multi-channel attention mechanism |
CN114880583A (en) * | 2022-05-11 | 2022-08-09 | 合肥工业大学 | Cross-domain social recommendation method based on self-supervision learning |
CN115618131A (en) * | 2022-11-11 | 2023-01-17 | 河南大学 | Article recommendation method and system fusing interaction frequency and attention network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021672A (en) * | 2017-12-06 | 2018-05-11 | 北京奇虎科技有限公司 | Social recommendation method, apparatus and computing device based on photograph album |
CN108320187A (en) * | 2018-02-02 | 2018-07-24 | 合肥工业大学 | A kind of recommendation method based on depth social networks |
CN108319659A (en) * | 2018-01-08 | 2018-07-24 | 暨南大学 | A kind of social discovery method based on encrypted image fast search |
CN108509493A (en) * | 2018-02-13 | 2018-09-07 | 北京邮电大学 | A kind of AR visual search recommendation service method |
-
2018
- 2018-10-16 CN CN201811200827.1A patent/CN109410080A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021672A (en) * | 2017-12-06 | 2018-05-11 | 北京奇虎科技有限公司 | Social recommendation method, apparatus and computing device based on photograph album |
CN108319659A (en) * | 2018-01-08 | 2018-07-24 | 暨南大学 | A kind of social discovery method based on encrypted image fast search |
CN108320187A (en) * | 2018-02-02 | 2018-07-24 | 合肥工业大学 | A kind of recommendation method based on depth social networks |
CN108509493A (en) * | 2018-02-13 | 2018-09-07 | 北京邮电大学 | A kind of AR visual search recommendation service method |
Non-Patent Citations (3)
Title |
---|
L. WU: "Modeling the evolution of users’ preferences and social links in social networking services", 《IEEE TRANS. KNOWL. DATA ENG.》 * |
LE WU: "Explainable Social Contextual Image Recommendation with Hierarchical Attention", 《HTTPS://ARXIV.ORG/PDF/1806.00723V1.PDF》 * |
袁寒雪: "基于反卷积网络的图像模式学习及应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110245285B (en) * | 2019-04-30 | 2021-07-13 | 中国科学院信息工程研究所 | Personalized recommendation method based on heterogeneous information network |
CN110245285A (en) * | 2019-04-30 | 2019-09-17 | 中国科学院信息工程研究所 | A kind of personalized recommendation method based on Heterogeneous Information network |
CN110148424A (en) * | 2019-05-08 | 2019-08-20 | 北京达佳互联信息技术有限公司 | Method of speech processing, device, electronic equipment and storage medium |
CN110148424B (en) * | 2019-05-08 | 2021-05-25 | 北京达佳互联信息技术有限公司 | Voice processing method and device, electronic equipment and storage medium |
CN110555132A (en) * | 2019-08-13 | 2019-12-10 | 北京工业大学 | Noise reduction self-encoder recommendation method based on attention model |
CN112700342A (en) * | 2020-12-29 | 2021-04-23 | 四川大学 | Personalized grading prediction model and prediction method for social media information authenticity |
CN112818146A (en) * | 2021-01-26 | 2021-05-18 | 山西三友和智慧信息技术股份有限公司 | Recommendation method based on product image style |
CN112905894A (en) * | 2021-03-24 | 2021-06-04 | 合肥工业大学 | Collaborative filtering recommendation method based on enhanced graph learning |
CN112905894B (en) * | 2021-03-24 | 2022-08-19 | 合肥工业大学 | Collaborative filtering recommendation method based on enhanced graph learning |
CN113010796A (en) * | 2021-04-15 | 2021-06-22 | 重庆大学 | Method for item recommendation |
CN113205869A (en) * | 2021-04-29 | 2021-08-03 | 华东师范大学 | Mental health detection and intervention method based on user social software head portrait |
CN113205869B (en) * | 2021-04-29 | 2022-10-04 | 华东师范大学 | Mental health detection and intervention method based on user social software head portrait |
CN113658013A (en) * | 2021-08-16 | 2021-11-16 | 上海花千树信息科技有限公司 | Method and device for recommending self-photographing on love or social software |
CN114677234A (en) * | 2022-04-26 | 2022-06-28 | 河南大学 | Graph convolution neural network social contact recommendation method and system integrating multi-channel attention mechanism |
CN114677234B (en) * | 2022-04-26 | 2024-04-30 | 河南大学 | Graph convolution neural network social recommendation method and system integrating multichannel attention mechanisms |
CN114880583A (en) * | 2022-05-11 | 2022-08-09 | 合肥工业大学 | Cross-domain social recommendation method based on self-supervision learning |
CN114880583B (en) * | 2022-05-11 | 2024-03-05 | 合肥工业大学 | Cross-domain social recommendation method based on self-supervision learning |
CN115618131A (en) * | 2022-11-11 | 2023-01-17 | 河南大学 | Article recommendation method and system fusing interaction frequency and attention network |
CN115618131B (en) * | 2022-11-11 | 2023-05-23 | 河南大学 | Article recommendation method and system integrating interaction frequency and attention network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109410080A (en) | A kind of social image recommended method based on level attention mechanism | |
CN111428147B (en) | Social recommendation method of heterogeneous graph volume network combining social and interest information | |
CN110196946A (en) | A kind of personalized recommendation method based on deep learning | |
CN110473592B (en) | Multi-view human synthetic lethal gene prediction method | |
CN103164713B (en) | Image classification method and device | |
CN110472817A (en) | A kind of XGBoost of combination deep neural network integrates credit evaluation system and its method | |
CN107633513A (en) | The measure of 3D rendering quality based on deep learning | |
CN108320187A (en) | A kind of recommendation method based on depth social networks | |
CN109960737B (en) | Remote sensing image content retrieval method for semi-supervised depth confrontation self-coding Hash learning | |
CN110599234A (en) | Product sales prediction method | |
CN108446312B (en) | Optical remote sensing image retrieval method based on deep convolution semantic net | |
CN109496322A (en) | Credit assessment method and device and the progressive decision tree parameter regulation means of gradient and device | |
CN110399819A (en) | A kind of remote sensing image residential block extraction method based on deep learning | |
CN108550077A (en) | A kind of individual credit risk appraisal procedure and assessment system towards extensive non-equilibrium collage-credit data | |
CN111538827A (en) | Case recommendation method and device based on content and graph neural network and storage medium | |
CN109858566A (en) | A method of it being added to the scorecard of mould dimension based on multilayered model building | |
CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
CN110415022A (en) | Handle the method and device of user behavior sequence | |
CN115917577A (en) | System and method for next basket recommendation using dynamic attribute modeling | |
CN108960488A (en) | A kind of accurate prediction technique of saturation loading spatial distribution based on deep learning and Multi-source Information Fusion | |
CN106407477A (en) | Multidimensional interconnection recommendation method and system | |
CN110070116A (en) | Segmented based on the tree-shaped Training strategy of depth selects integrated image classification method | |
CN110503508A (en) | A kind of item recommendation method of the more granularity matrix decompositions of level | |
CN108491864A (en) | Based on the classification hyperspectral imagery for automatically determining convolution kernel size convolutional neural networks | |
CN109272487A (en) | The quantity statistics method of crowd in a kind of public domain based on video |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190301 |