CN108280738A - Method of Commodity Recommendation based on image and socialized label - Google Patents
Method of Commodity Recommendation based on image and socialized label Download PDFInfo
- Publication number
- CN108280738A CN108280738A CN201711324898.8A CN201711324898A CN108280738A CN 108280738 A CN108280738 A CN 108280738A CN 201711324898 A CN201711324898 A CN 201711324898A CN 108280738 A CN108280738 A CN 108280738A
- Authority
- CN
- China
- Prior art keywords
- commodity
- matrix
- user
- picture
- label
- 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 29
- 239000011159 matrix material Substances 0.000 claims description 76
- 239000013598 vector Substances 0.000 claims description 47
- 230000006870 function Effects 0.000 claims description 31
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 2
- 239000000047 product Substances 0.000 claims 3
- 239000012467 final product Substances 0.000 claims 1
- 230000000875 corresponding effect Effects 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 238000007418 data mining Methods 0.000 description 2
- 238000009795 derivation Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 235000009508 confectionery Nutrition 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000012163 sequencing technique Methods 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
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Accounting & Taxation (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Finance (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of Method of Commodity Recommendation based on image and socialized label, mainly solves the problems, such as that existing recommendation method cannot capture the preference of user well.Its implementation is：1. extracting color characteristic, textural characteristics and the shape feature of commodity picture；2. designing optimal semantic space：According to user, the contact of label and commodity picture between any two builds bipartite graph and neighbor relationships figure, and by user, label and commodity picture, these three data are mapped in the same space, i.e., optimal semantic space；3. in optimal semantic space according to the Euclidean distance between user and commodity picture to commodity picture according to from being closely ranked up to remote sequence, the top n commodity picture nearest apart from user is recommended into user.The present invention can more accurately calculate the similarity between picture, and the effective complicated relational structure handled between three class object of label data obtains and preferably recommends performance, can be used for the personalized recommendation system of e-commerce.
Description
Technical Field
The invention belongs to the technical field of data mining, and particularly relates to a commodity recommendation method which can be used for an electronic commerce personalized recommendation system.
Background
In recent years, with the increasing scale of electronic commerce, the number and types of products are rapidly increasing, and customers need to spend a lot of time to find the products they want to buy. This process of browsing through large amounts of unrelated information and products will undoubtedly result in a constant loss of consumers who are overwhelmed with a huge amount of information. To solve this problem, personalized recommendation systems have been developed. The personalized recommendation system is an active information service system established on the basis of mass data mining, and is used for helping an e-commerce website to provide completely personalized decision support and information service for shopping of customers.
Most of the traditional recommendation systems utilize user scoring data, use a collaborative filtering algorithm to perform calculation screening, and then recommend the user. However, the rating data cannot intuitively reflect the preference of the user, such as rating of a movie, and only whether the user likes the movie or not but not the specific attention point of the user to the movie can be known by using the rating data of the user. Therefore, recommendation systems based on user scoring data do not capture user specific preferences well.
Disclosure of Invention
The invention aims to provide a commodity recommendation method based on images and social labels to better capture specific preferences of a user aiming at the defects of the prior art.
The technical scheme of the invention is realized as follows:
technical principle
The social label can reflect the preference of the user more directly, because the label provides a meaningful description for the resource, namely the commodity picture, for example, for the picture of the clothes, whether the clothes is 'rural wind' or 'sweet fair maiden' can be known through the label, therefore, the label embodies the semantic understanding of the user for the commodity, and implies the preference of the user. In addition, since tag data has three types of objects: the user, the label and the commodity picture have complex association structures, so that the commodity recommendation cannot be performed by using the traditional collaborative filtering algorithm. Therefore, the invention designs a multi-class associated object dimension reduction algorithm based on a graph, which constructs a bipartite graph for the relationship among a user, a label and a commodity picture, constructs a neighbor relationship graph for a commodity picture set, and maps the three graphs into the same space on the premise of keeping the structures of the two graphs, wherein the space is called as a semantic space because the semantic association relationship of label data is kept. After the optimal semantic space is obtained, the Euclidean distance between the user and the commodity picture is calculated in the semantic space, the commodity corresponding to the picture closest to the user is recommended to the user, and the recommended commodity is displayed in a picture form.
Second, the technical scheme
According to the principle, the technical scheme of the invention comprises the following steps:
(1) extracting content features, namely color features, texture features and shape features, of the commodity picture;
(2) designing an optimal semantic space:
2a) learning a weight parameter vector between the three characteristics of the commodity pictures by using a multi-core learning algorithm to obtain the final similarity between the two commodity pictures, and fusing the weight parameter vector and the calculated similarity between the commodity pictures to obtain a similarity matrix;
2b) defining the tag data as (u)i,dj,tk) U is a user set, D is a commodity picture set, T is a label set, UiIndicating the ith in UDoor, djRepresents the jth commodity picture, t, in DkDenotes the kth tag in T, tag data (u)i,dj,tk) Representing user uiUse of the tag tkTo mark picture dj；
2c) Constructing a weighted bipartite graph for the relation between the user and the label in the label data, constructing a weighted bipartite graph for the relation between the label and the commodity picture in the label data, constructing a non-weighted bipartite graph for the relation between the user and the commodity picture in the label data, and constructing a neighbor relation graph for the commodity picture set;
2d) on the premise of keeping two graph structures of a bipartite graph and a neighbor relation graph in the step 2c), mapping three data of a user, a label and a commodity picture into the same space, wherein the space is the optimal semantic space;
(3) in the optimal semantic space obtained in the step (2), the commodity pictures are sorted from near to far according to the Euclidean distance between the user and the commodity pictures, the first N commodity pictures closest to the user are recommended to the user, and the range of N is (1, + ∞).
The invention has the following beneficial effects:
1. the invention can reflect the preference of the user more directly and obtain better recommendation performance because of using the label data instead of the user scoring data.
2. According to the method, the features of the three pictures are extracted, the weight relation among the features is learned by utilizing a multi-core learning technology, and the picture similarity after weighted fusion among the features is calculated.
3. According to the method, the bipartite graph and the neighbor relation graph are constructed to establish the complex relation among the user, the label and the commodity picture, so that the problem of data sparsity is solved, and the recommendation method is more effective and accurate.
4. The semantic space designed by the invention can effectively process the complex association structure among three types of objects of the label data, namely the user, the label and the commodity picture, and the traditional collaborative filtering algorithm can not process the label data.
Drawings
FIG. 1 is a general flow chart of an implementation of the present invention;
FIG. 2 is a semantic space sub-flow diagram of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
Referring to fig. 1, the implementation steps of the invention are as follows:
step 1, extracting content features, namely color features, texture features and shape features, of a commodity picture;
1.1) extracting color features of the commodity picture:
establishing a color histogram, and for the picture Q, counting the color histogram on a color vector space, wherein the color histogram is a one-dimensional discrete function, namely:
in the formula, nkObtaining a color histogram H (k) of the image Q by taking the number of pixels with a color characteristic value of k, N as the total number of image pixels and l as the number of the color characteristic value;
1.2) extracting the texture features of the commodity picture:
describing texture features of the picture by extracting scale invariant feature sift descriptors;
1.2.1) let the picture be I (x, y), (x, y) is the spatial scale coordinate, then the picture scale space L (x, y, σ) is:
L(x,y,σ)＝G(x,y,σ)*I(x,y)，
wherein G (x, y, σ) is a gaussian function, σ is a scale parameter, and x represents a multiplier;
1.2.2) calculating the difference of gaussians D (x, y, sigma) of adjacent scale images:
D(x,y,σ)＝L(x,y,k1σ)-L(x,y,σ)
wherein k is1GetL (-) is a scale space;
after the Gaussian difference of the images with adjacent scales is calculated, a series of Gaussian difference images are obtained;
1.2.3) solving extreme points in the space of the Gaussian difference image, and respectively comparing a pixel point in each Gaussian difference image with all adjacent points: if a certain pixel point is larger or smaller than all the adjacent points, the pixel point is an extreme point, and the Gaussian difference of the extreme point is as follows:
wherein,the extreme value of Z is represented by (x, y, sigma) which is the offset of the pixel point, and D (·) is a Gaussian difference function;
1.2.4) determining the extreme point, it is necessary to doPerforming curve fitting on the DoG function in the scale space to screen extreme points and remove low-contrast points, namely for each candidate extreme pointAnd (4) judging:
if it is notIf the candidate extreme point is less than 0.03, judging that the candidate extreme point is an extreme point with low contrast, removing the candidate extreme point, and otherwise, keeping the candidate extreme point;
1.2.5) removing the extreme point with low contrast, and then removing the extreme point with abnormal principal curvature, because an extreme value of a poorly defined gaussian difference has a larger principal curvature at a place crossing an edge and a smaller principal curvature at a direction perpendicular to the edge, and ideally, the principal curvatures of the extreme points in any directions are the same, it is necessary to detect whether the principal curvature is below a certain threshold r, and in order to detect whether the principal curvature is below a certain threshold r, only the following inequality needs to be detected:
wherein, tr (H) represents the trace of the matrix H, det (H) represents the determinant of the matrix H, the matrix H is a Hession matrix, and the expression is as follows:
wherein D isxxIs the result of twice derivation of D (x, y, σ) of a scale image in the x direction in DoG space; dyyIs the result of twice derivation of D (x, y, σ) of a scale image in the y direction in DoG space; dxyThe method is characterized in that D (x, y, sigma) of an image with a certain scale is derived once in the x direction of a DoG space and then derived once in the y direction; when the inequalities are satisfied, the extreme point is retained, otherwise, the extreme point is removed and retainedThe remaining extreme points are the characteristic points to be found;
1.2.6) after the positions of the image feature points are determined, a direction needs to be assigned to the image feature points, and the step is realized by solving the gradient of the neighborhood of each feature point, namely defining the gradient amplitude m (x, y) and the gradient direction theta (x, y) as follows:
θ(x,y)＝tan-1(L(x,y+1)-L(x,y-1))/(L(x+1,y)-L(x-1,y))
wherein L (-) is a scale space function;
1.2.7) taking an image feature point as a center, defining a region, constructing a direction histogram by using gradient amplitudes and gradient directions of all points in the region, wherein the gradient amplitude is a vertical axis, the gradient direction is a horizontal axis, selecting a gradient direction corresponding to an item with the maximum gradient amplitude from the direction histogram as a main direction of the feature point, and if the gradient amplitudes corresponding to other gradient directions are more than 80% of the gradient amplitudes corresponding to the main direction, also taking the gradient direction as an auxiliary direction of the image feature point;
1.2.8) after the image characteristic points are detected, determining the descriptors of the characteristic points according to the following steps:
firstly, rotating the neighborhood of the feature point by theta by taking the feature point as a circle center0Where θ is the direction of the feature point;
secondly, taking a 16 multiplied by 16 neighborhood window by taking the feature point as a center in the rotated image, wherein each cell represents one pixel in the neighborhood window of the feature point; uniformly dividing a 16 x 16 rectangular window into 16 sub-regions, increasing the weight value of a neighborhood close to the feature point by adopting a Gaussian blur method, and reducing the weight value of a neighborhood far away from the feature point; then, calculating gradient histograms of 8 directions in each region to obtain a feature vector of the feature point descriptor, wherein the vector is a 128-dimensional vector with 4 multiplied by 8;
thirdly, normalizing the feature vector of the feature point descriptor, and setting S as the feature vector of the feature point descriptor, namely S ═ S (S)1,s2,......s128) Normalizing S to obtain normalized feature point descriptor
To reduce the effect of large gradient values, isSetting a threshold value of 0.2 ifIf the value of one dimension is greater than 0.2, setting the value to 0.2 and then pairing againNormalization processing is carried out, and finally 128-dimensional normalized feature point descriptors are obtained
1.2.9) pairs of normalized feature point descriptorsClustering to obtain sift characteristic vectors;
in order to obtain fixed-length sift feature vectors, normalized feature point descriptors need to be obtainedClustering is performed because the number of descriptors contained in different pictures is different, which results in thatThe sift characteristic vectors extracted from different pictures are lengthened, so the clustering number is set to be 500, k-means clustering is carried out on the characteristic point descriptors contained in each picture, a fixed-length 500-dimensional clustering result can be obtained from each picture, and the result is the final sift characteristic vector and is used for describing the texture characteristics of the commodity picture;
1.3) extracting the shape characteristics of the commodity picture:
1.3.1) utilizing a Gabor filter to perform sampling filtering on the commodity picture according to the following formula:
wherein,
l1is the size of the filter; k is a radical of2Is a positive constant; sigma1Is the standard deviation of the gaussian function;is scale l1The total number of directions of the lower part,
1.3.2) convolving the commodity picture with a Gabor filter to obtain a filtered commodity picture:
dividing the filtered commodity picture into 4 multiplied by 4 grids, taking an average value in each grid, and finally putting the average values obtained in grids of all directions and scales in a vector, wherein the vector is a final universal search tree feature vector and is used for describing the shape features of the commodity picture;
thus, the extraction of the color feature, the texture feature and the shape feature of the commodity picture is completed.
And 2, designing a semantic space.
Referring to FIG. 2, the specific implementation of this step is as follows
2.1) constructing a similarity matrix:
the invention uses the multi-core learning technology to construct the similarity matrix among the commodity pictures, namely three characteristics of color characteristics, texture characteristics and shape characteristics are extracted from the commodity pictures, and when the similarity among the commodity pictures is calculated, the similarity among the three characteristics and the commodity pictures needs to be weighted and fused to obtain the similarity among the weighted and fused commodity pictures, and then the similarity matrix is constructed according to the similarity among the weighted and fused commodity pictures, and the method comprises the following implementation steps:
2.1.1) calculating the similarity between the commodity pictures:
selecting a Gaussian kernel as the kernel function, the kernel function K (x)i,xj) Is represented as follows:
wherein x isiFeature vector, x, representing the ith product picturejThe characteristic vector of the jth commodity picture is represented, i is more than or equal to 1, j is more than or equal to 1, i is not equal to j, and sigma is2Denotes the standard deviation, | xi-xj||2Represents a vector (x)i-xj) 2 norm of, kernel function K (x)i,xj) Representing the similarity between the ith commodity picture and the jth commodity picture if K (x)i,xj) The closer to 1, the more similar the two commodity pictures are;
2.1.2) degree of similarity K (x)i,xj) Normalization is performed to select kernel function K (x)i,xj) All values are [0,1 ]]The normalized similarity K (i, j) is as follows:
2.1.3) fusing the normalized similarity K (i, j) with the weight parameter vector η to obtain the similarity K (i, j, η) between the two weighted and fused commodity pictures:
wherein i represents the ith commodity picture, j represents the jth commodity picture, H1The number of feature vectors indicating a commercial picture is [ η ] as parameter η1,η2,......,ηH]TIs a vector of weight parameters, parameters η, to be learnedvRepresents the weight parameter vector η ═ η1,η2,......,ηH]TThe v-th parameter, Kv(i, j) represents the similarity between the ith commodity picture and the jth commodity picture under the vth feature;
2.1.4) defining the kernel function K in the ideal stateideal(i, j) is:
wherein y isiIndicates the category of the ith product picture, yjA category representing a jth commodity picture;
2.1.5) define the square loss function:
for the ith commodity picture and the jth commodity picture, the similarity K (i, j, η) between the two commodity pictures needs to be close to K as much as possibleideal(i, j) accordinglyThe square loss function l (i, j, η) is defined as:
l(i,j,η)＝(K(i,j,η)-Kideal(i,j))2，
the smaller the value of l (i, j, η), the closer the similarity K (i, j, η) is to Kideal(i,j)；
2.1.6) defining an optimization calculation formula to obtain a weight parameter vector η to minimize the loss function l (i, j, η), wherein the specific calculation formula is as follows:
wherein q represents the total number of the commodity pictures, tijThe weight of the loss function representing the ith and jth product pictures, λ is a compromise parameter,representing a regularization term;
the above equation finally yields a weight parameter vector η, which is added with a regularization termThe formula is a typical quadratic programming problem, can be directly solved by using a common convex optimization software package, and adopts a data set which is a large-scale image set ImageNet of Google in order to ensure that the obtained weight parameter vector η is more accurate;
2.1.7) defining a similarity matrix W based on the obtained weight parameter vector η as:
2.2) setting parameters:
defining the tag data as (u)i,dj,tk) U is a user set, D is a commodity picture set, T is a label set, UiRepresents the ith user in U, djRepresents the jth commodity picture, t, in DkDenotes the kth tag in T, tag data (u)i,dj,tk) Representing user uiUse of the tag tkTo mark picture dj；
2.3) constructing a weighted bipartite graph of the relation between the user and the label in the label data:
2.3.1) define the following first matrix:
wherein u isiRepresents the ith user in U, dkDenotes the kth product picture in D, tjRepresents the jth label in T, and B represents the set of all label data;
2.3.2) for the first matrixAnd normalizing to obtain a normalized first matrix as follows:
then R isutThe weighted bipartite graph between the users and the labels is formed by structure matrixing;
2.4) constructing a weighted bipartite graph for the relation between the labels and the commodity pictures in the label data:
2.4.1) defining the second matrix
Wherein,ukdenotes the kth user in U, djRepresents the jth commodity picture, t, in DiRepresenting the ith label in T, and B representing the set of all label data;
2.4.2) for the second matrixAnd normalizing to obtain a normalized second matrix as follows:
Rtdweighted bipartite graph between labels and commodity pictures in structural matrixing
2.5) establishing an unauthorized bipartite graph for the relation between the user and the commodity picture in the tag data
A third matrix R is defined asud：
Wherein u isiRepresents the ith user in U, djRepresents the jth commodity picture, t, in DkRepresenting the kth label in T, and B representing the set of all label data;
Rudthe method comprises the following steps of (1) forming an unweighted bipartite graph between users and commodity pictures in a structural matrixing manner;
2.6) constructing a neighbor relation graph for the commodity picture set:
defining a fourth matrix as follows:
wherein d isiRepresenting the ith product picture in D, DjTo representJ-th commodity picture in D, sim (D)i,dj) Representing the similarity between the ith commodity picture and the jth commodity picture;
Wija neighbor relation graph of the structural matrixing commodity picture set;
2.7) constructing semantic spaces
In this step, in order to learn an optimal semantic space, a graph-based subspace learning idea is fused, that is, for two strongly correlated objects, they should be very close in the semantic space;
2.7.1) consider the simplest case, the semantic space being one-dimensional, i.e. when k is 1, the following vector and loss function are defined:
define a vector of | U | × 1: f ═ f1,f2,...f|U|}，
Define a vector of | T | × 1: g ═ g1,g2,...g|T|}，
Define a vector of | D | × 1: p ═ p1,p2,...p|D|}，
The loss function is defined as:
wherein f isiE f, representing the coordinate of the ith user in a one-dimensional space; giE g, representing the coordinate of the ith label in a one-dimensional space; gjE g, representing the coordinate of the jth label in a one-dimensional space; p is a radical ofiE.g. p, representing the coordinate of the ith commodity picture in one dimension; p is a radical ofjThe parameter α, gamma and η are four scaling factors with different values and satisfy 0 < α, gamma, η <1, α + β + gamma + η as 1;
loss function as described aboveThe first term on the right of the equation indicates if user uiFrequently used labels tjThey should be mapped very close to each other in space, i.e., fiAnd gjThe squared difference of (c) is to be as small as possible; the second item represents if many users use the tag tiDe-labeling picture dj，tiAnd djShould also be very close, i.e. giAnd pjThe difference is not large; the third term represents each user uiAll should be in contact with user uiMarked picture djVery close, i.e. fiAnd pjThe difference is not large; the last item is a smooth item of the internal structure of the commodity picture, namely for two very similar commodity pictures diAnd djThe influence of the above four terms on the loss function is determined by parameters α, gamma, η;
2.7.2) to solve the above-mentioned loss function Q (f, g, p), it is necessary to define seven diagonal matrices, D respectivelyut,Dtu,Dtd,Ddt,Dud,DduAnd D, wherein:
Dutis a diagonal matrix of size | Ux | U |, DutIs the matrix RutThe sum of the elements of row i of (1);
Dudis a diagonal matrix of size | Ux | U |, DudIs the matrix RudThe sum of the elements of row i of (1);
Dtuis a diagonal matrix of size | T | × | T |, DtuIs the matrix RutThe sum of the ith column element of (1);
Dtdis a diagonal matrix of size | T | × | T |, DtdIs the matrix RtdThe sum of the elements of row i of (1);
Dduis a diagonal matrix of size | D | × | D |, DduIs the matrix RudThe sum of the ith column element of (1);
Ddtis a diagonal matrix of size | D | × | D |, DdtIs the matrix RtdThe sum of the ith column element of (1);
d is a diagonal matrix with size | D | × | D |, the (i, i) -th element in D is the matrix WijThe sum of the elements of row i of (1);
the first term of the loss function Q (f, g, p) can be rewritten as:
the second term of the loss function Q (f, g, p) is also rewritten as:
the third term of the loss function Q (f, g, p) is also rewritten as:
the fourth term for Q (f, g, p) can be rewritten as:
wherein the matrix L ═ D-WijIs a Laplace matrix;
2.7.3) the loss function Q (f, g, p) is represented as a matrix according to equations <1> to <5 >:
Q1(f,g,p)＝α(fTDutf+gTDtug-2fTRutg)
+β(gTDtdg+pTDdtp-2gTRtdp)
+γ(fTDudf+pTDdup-2fTRudp)
+ηpTLp
＝fT(αDut+γDud)f+gT(αDtu+βDtd)g<6>
+pT(βDdt+γDdu+L)p
-2αfTRutg-2βgTRtdp-2γfTRudp
2.7.4) to eliminate the effect of any scaling factors α, γ, η, Q is added1Normalized (f, g, p) to obtain normalized
Formula (II)<7>Is equivalent to fTf+gTg+pTMinimizing loss function Q under constraint of p-11(f,g,p)；
2.7.5) defines an augmented vector h ═ fTgTpT]TSimplified form<7>；
Due to fTRutg，gTRtdp and fTRudp is a scalar, the transpose of the scalar is equal to the scalar itself, so that the normalizedCan be abbreviated as:
where e is a unit vector of the vector,for a semi-positive definite matrix, the following is defined:
2.7.6) maximizing the global variance of the target subspace;
formula (II)<8>Denotes that h is required to minimize the loss function Q (f, g, p)Th maximization, h maximizationTh is a better method to maximize the global variance of the target subspace, and the specific method is as follows:
for a random variable x, the variance is:
var(x)＝∫(x-μ)2dP(x),μ＝∫xdP(x)<9>
where dP (x) represents the probability of x being observed;
in the discrete case, the probability of observing a node on the graph can be estimated by the degree of the node, so the global variance of f, g, and p is:
matrix arrayIs a diagonal matrix defined as follows:
2.7.7) according to formula<8>And formula<10>Will lose a functionIs rewritten as
2.7.8) in practical applications, in order to better capture the relationship between objects, it is often necessary to obtain a semantic space in k dimension (k > 1), and therefore, it is necessary to extend the above calculation process to the k dimension case, where k > 1, defines the following vectors, matrices and loss functions:
defining a matrix F ═ k of | U | ×1f2......fk]，
Defining a matrix G ═ k of | T | ×1g2......gk]，
Defining a matrix P ═ P | D × k1p2......pk]，
Definition matrix H3＝[h1h2......hk]Wherein h isi＝[fi Tgi Tpi T]T，
For each dimension i e {1, 1i,gi,pi) And maximize (f)i Tfi+gi Tgi+pi Tpi) Thus, a loss function in k-dimension is definedThe following were used:
wherein f isie.F represents the coordinates of all users in the user set U in the ith dimension, giE G represents the coordinates of all the labels in the label set T in the ith dimension, piE.g. P represents the coordinates of all the commodity pictures in the commodity picture set D in the ith dimension; tr (-) represents the trace of the matrix;
2.7.9) like the case where k is 1, it is also necessary to maximize the global variance of the target subspace when k > 1, so in the case where k > 1, the equation <12> is rewritten as:
solution formula<13>Obtaining a semi-positive definite matrixAnd diagonal matrixThen, according to Rayleigh theory, solvingAnd removing the eigenvectors with equal components from the eigenvectors corresponding to the first k smallest eigenvalues, wherein lambda is the eigenvalue, and the matrix formed by the remaining eigenvectors is the optimal semantic space M.
And 3, calculating the Euclidean distance between the user and the commodity picture.
Giving a target user, and calculating the Euclidean distance between the target user and the commodity picture through the corresponding row vector of the target user and the commodity picture in the optimal semantic space M:
using n-dimensional vectors A1＝(x11,x12,...x1k...x1n) Indicates that the user is at the mostCorresponding row vectors in the preferred semantic space M, where x1kIs represented by A1The k value is more than or equal to 1 and less than or equal to n,
using n-dimensional vectors A2＝(x21,x22,...x2k...x2n) Representing the corresponding line vector of the commodity picture in the optimal semantic space M, wherein x2kIs represented by A2The k-th value of the number k,
calculating the Euclidean distance d between the user and the commodity picture12Comprises the following steps:
and sequencing all the commodity pictures according to the obtained Euclidean distance in the order from near to far, wherein the larger the Euclidean distance is, the more irrelevant the user is to the commodity pictures, and returning the commodity recommendation corresponding to the first N commodity pictures closest to the user, wherein the range of N is (1, + ∞), so that the personalized commodity recommendation work for the user is completed.
The above description is only one specific example of the present invention and should not be construed as limiting the scope of the invention; it should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention.
Claims (7)
1. A commodity recommendation method based on images and social labels is characterized by comprising the following steps:
(1) extracting content features, namely color features, texture features and shape features, of the commodity picture;
(2) designing an optimal semantic space:
2a) learning a weight parameter vector between the three characteristics of the commodity pictures by using a multi-core learning algorithm to obtain the final similarity between the two commodity pictures, and fusing the weight parameter vector and the calculated similarity between the commodity pictures to obtain a similarity matrix;
2b) defining the tag data as (u)i,dj,tk) U is a user set, D is a commodity picture set, T is a label set, UiRepresents the ith user in U, djRepresents the jth commodity picture, t, in DkDenotes the kth tag in T, tag data (u)i,dj,tk) Representing user uiUse of the tag tkTo mark picture dj；
2c) Constructing a weighted bipartite graph for the relation between the user and the label in the label data, constructing a weighted bipartite graph for the relation between the label and the commodity picture in the label data, constructing a non-weighted bipartite graph for the relation between the user and the commodity picture in the label data, and constructing a neighbor relation graph for the commodity picture set;
2d) on the premise of keeping two graph structures of a bipartite graph and a neighbor relation graph in the step 2c), mapping three data of a user, a label and a commodity picture into the same space, wherein the space is the optimal semantic space;
(3) in the optimal semantic space obtained in the step (2), the commodity pictures are sorted from near to far according to the Euclidean distance between the user and the commodity pictures, the first N commodity pictures closest to the user are recommended to the user, and the range of N is (1, + ∞).
2. The method according to claim 1, wherein the learning of the weight relationship between the three features in the commodity picture by using the multi-kernel learning algorithm in step 2a) is to learn the weight parameter vector between the color feature, the texture feature and the shape feature of the commodity picture by using the following kernel functions:
wherein i represents the ith commodity picture, j represents the jth commodity picture, H1The number of feature vectors indicating a commercial picture is [ η ] as parameter η1,η2,......,ηH]TIs the weight parameter to be learnedVector, parameter ηvRepresents the weight parameter vector η ═ η1,η2,......,ηH]TThe v-th parameter, Kv(i, j) represents the similarity between the two product pictures under the v-th feature, and K (i, j, η) represents the similarity between the two final product pictures.
3. The method of claim 1, wherein the step 2c) of constructing a weighted bipartite graph of the association between the user and the tags is performed by:
2c1) the following matrix is defined:
wherein u isiRepresents the ith user in U, dkDenotes the kth product picture in D, tjRepresents the jth label in T, and B represents the set of all label data;
2c2) for the above matrixNormalization is carried out to obtain a normalized matrix:
Ruta weighted bipartite graph between users and tags matrixed to the structure.
4. The method according to claim 1, wherein the step 2c) of constructing the weighted bipartite graph for the association between the labels and the commodity pictures is performed as follows:
2c3) define the following matrix
Wherein u iskIndicating the k-th use in UDoor, djRepresents the jth commodity picture, t, in DiRepresenting the ith label in T, and B representing the set of all label data;
2c4) for the above matrixNormalization is carried out to obtain a normalized matrix:
Rtdthe method is a weighted bipartite graph between the labels and the commodity pictures which are structured in a matrix mode.
5. The method according to claim 1, wherein the non-weighted bipartite graph is constructed in step 2c) for the association between the user and the commodity picture, and the following matrix is defined:
wherein u isiRepresents the ith user in U, djRepresents the jth commodity picture, t, in DkRepresenting the kth label in T, and B representing the set of all label data;
Rudthe method is an unweighted bipartite graph between users and commodity pictures which are structured in a matrix mode.
6. The method according to claim 1, wherein the constructing of the neighborhood relationship graph for the commodity picture set in step 2c) is performed by defining a matrix as follows:
wherein d isiRepresenting the ith product picture in D, DjRepresenting the jth commodity picture in the D;
Wijproximity between sets of pictures of a commodity matrixed for structureAnd (4) a neighborhood relationship graph.
7. The method according to claim 1, wherein the three data of the user, the label and the commodity picture are mapped into the same space in the step 2d), and the following steps are performed:
2d1) considering the semantic space as one-dimensional, a vector f of | U | × 1 is defined, a vector g of | T | × 1 is defined, a vector p of | D | × 1 is defined, and the following loss function is defined:
wherein f isiRepresents user uiCoordinates in one dimension, giRepresents a label tiCoordinates in one dimension, gjRepresents a label tjCoordinate in one dimension, pjCommodity picture djCoordinates in one dimension, RutWeighted bipartite graph, R, between users and labels that is structured into a matrixtdIs a weighted bipartite graph, R, between structured matrixed labels and commodity picturesudIs an unweighted bipartite graph, W, between users and commodity pictures in a structural matrixingijFor the neighbor relation graph among the structure matrixed commercial picture sets, parameters α, γ, η satisfy α + β + γ + η as 1, 0 < α, γ, η < 1;
2d2) optimizing and solving the loss function, expanding the solving process to the situation of k dimension, and finally obtaining a semi-positive definite matrixAnd a diagonal matrix
Wherein D isutIs a diagonal matrix of size | Ux | U |, DutIs the matrix RutThe sum of the elements of row i of (1); dudIs a diagonal matrix of size | Ux | U |, DudIs the matrix RudThe sum of the elements of row i of (1); dtuIs a diagonal matrix of size | T | × | T |, DtuIs the matrix RutThe sum of the ith column element of (1); dtdIs a diagonal matrix of size | T | × | T |, DtdIs the matrix RtdThe sum of the elements of row i of (1); dduIs a diagonal matrix of size | D | × | D |, DduIs the matrix RudThe sum of the ith column element of (1); ddtIs a diagonal matrix of size | D | × | D |, DdtIs the matrix RtdThe sum of the ith column element of (1); d is a diagonal matrix with size | D | × | D |, the (i, i) -th element in D is the matrix WijThe element of row i.
2d3) Obtaining a semi-positive definite matrixAnd diagonal matrixThen, according to Rayleigh theory, solvingAnd removing the eigenvectors with equal components from the eigenvectors corresponding to the first k smallest eigenvalues, wherein a matrix formed by the remaining eigenvectors is the optimal semantic space.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711324898.8A CN108280738A (en) | 2017-12-13 | 2017-12-13 | Method of Commodity Recommendation based on image and socialized label |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711324898.8A CN108280738A (en) | 2017-12-13 | 2017-12-13 | Method of Commodity Recommendation based on image and socialized label |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108280738A true CN108280738A (en) | 2018-07-13 |
Family
ID=62799846
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711324898.8A Pending CN108280738A (en) | 2017-12-13 | 2017-12-13 | Method of Commodity Recommendation based on image and socialized label |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108280738A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109886310A (en) * | 2019-01-25 | 2019-06-14 | 北京三快在线科技有限公司 | Picture sort method, device, electronic equipment and readable storage medium storing program for executing |
| CN110321473A (en) * | 2019-05-21 | 2019-10-11 | 山东省计算中心（国家超级计算济南中心） | Diversity preference information method for pushing, system, medium and equipment based on multi-modal attention |
| CN110827045A (en) * | 2018-08-07 | 2020-02-21 | 北京京东尚科信息技术有限公司 | Method and device for distinguishing commodity relation |
| CN111523590A (en) * | 2020-04-21 | 2020-08-11 | 莫毓昌 | Multi-data source fusion method for scoring matrix |
| WO2021077989A1 (en) * | 2019-10-25 | 2021-04-29 | 腾讯科技（深圳）有限公司 | Method and device for making recommendation, computer device, and storage medium |
| CN113159891A (en) * | 2021-04-24 | 2021-07-23 | 桂林电子科技大学 | Commodity recommendation method based on fusion of multiple user representations |
| CN113362131A (en) * | 2021-06-02 | 2021-09-07 | 合肥工业大学 | Intelligent commodity recommendation method based on map model and integrating knowledge map and user interaction |
| CN113793182A (en) * | 2021-09-15 | 2021-12-14 | 广州华多网络科技有限公司 | Commodity object recommendation method and device, equipment, medium and product thereof |
| CN116308327A (en) * | 2022-12-28 | 2023-06-23 | 深圳市销邦数据技术有限公司 | Self-service cashing system and method based on RFID technology |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102314614A (en) * | 2011-10-24 | 2012-01-11 | 北京大学 | Image semantics classification method based on class-shared multiple kernel learning (MKL) |
| CN102629257A (en) * | 2012-02-29 | 2012-08-08 | 南京大学 | Commodity recommending method of e-commerce website based on keywords |
| CN103116639A (en) * | 2013-02-20 | 2013-05-22 | 新浪网技术(中国)有限公司 | Item recommendation method and system based on user-item bipartite model |
| CN103309972A (en) * | 2013-06-08 | 2013-09-18 | 清华大学 | Recommend method and system based on link prediction |
| US20160012088A1 (en) * | 2014-07-08 | 2016-01-14 | Palo Alto Research Center Incorporated | Parallel collective matrix factorization framework for big data |
-
2017
- 2017-12-13 CN CN201711324898.8A patent/CN108280738A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102314614A (en) * | 2011-10-24 | 2012-01-11 | 北京大学 | Image semantics classification method based on class-shared multiple kernel learning (MKL) |
| CN102629257A (en) * | 2012-02-29 | 2012-08-08 | 南京大学 | Commodity recommending method of e-commerce website based on keywords |
| CN103116639A (en) * | 2013-02-20 | 2013-05-22 | 新浪网技术(中国)有限公司 | Item recommendation method and system based on user-item bipartite model |
| CN103309972A (en) * | 2013-06-08 | 2013-09-18 | 清华大学 | Recommend method and system based on link prediction |
| US20160012088A1 (en) * | 2014-07-08 | 2016-01-14 | Palo Alto Research Center Incorporated | Parallel collective matrix factorization framework for big data |
Non-Patent Citations (3)
| Title |
|---|
| MINGXIN GAN: "TAFFY: incorporating tag information into a diffusion", 《WORLD WIDE WEB-INTERNET AND WEB INFORMATION SYSTEMS》 * |
| 孔祥迎: "基于社会化标签的个性化推荐技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 管子玉: ""基于图学习的Web信息检索技术研究"", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110827045A (en) * | 2018-08-07 | 2020-02-21 | 北京京东尚科信息技术有限公司 | Method and device for distinguishing commodity relation |
| CN109886310B (en) * | 2019-01-25 | 2020-06-09 | 北京三快在线科技有限公司 | Picture sorting method and device, electronic equipment and readable storage medium |
| CN109886310A (en) * | 2019-01-25 | 2019-06-14 | 北京三快在线科技有限公司 | Picture sort method, device, electronic equipment and readable storage medium storing program for executing |
| CN110321473B (en) * | 2019-05-21 | 2021-05-25 | 山东省计算中心（国家超级计算济南中心） | Multi-modal attention-based diversity preference information pushing method, system, medium and device |
| CN110321473A (en) * | 2019-05-21 | 2019-10-11 | 山东省计算中心（国家超级计算济南中心） | Diversity preference information method for pushing, system, medium and equipment based on multi-modal attention |
| WO2021077989A1 (en) * | 2019-10-25 | 2021-04-29 | 腾讯科技（深圳）有限公司 | Method and device for making recommendation, computer device, and storage medium |
| CN111523590B (en) * | 2020-04-21 | 2023-04-07 | 莫毓昌 | Multi-data source fusion method for scoring matrix |
| CN111523590A (en) * | 2020-04-21 | 2020-08-11 | 莫毓昌 | Multi-data source fusion method for scoring matrix |
| CN113159891A (en) * | 2021-04-24 | 2021-07-23 | 桂林电子科技大学 | Commodity recommendation method based on fusion of multiple user representations |
| CN113159891B (en) * | 2021-04-24 | 2022-05-17 | 桂林电子科技大学 | Commodity recommendation method based on fusion of multiple user representations |
| CN113362131A (en) * | 2021-06-02 | 2021-09-07 | 合肥工业大学 | Intelligent commodity recommendation method based on map model and integrating knowledge map and user interaction |
| CN113362131B (en) * | 2021-06-02 | 2022-09-13 | 合肥工业大学 | Intelligent commodity recommendation method based on map model and integrating knowledge map and user interaction |
| CN113793182A (en) * | 2021-09-15 | 2021-12-14 | 广州华多网络科技有限公司 | Commodity object recommendation method and device, equipment, medium and product thereof |
| CN113793182B (en) * | 2021-09-15 | 2024-05-28 | 广州华多网络科技有限公司 | Commodity object recommendation method and device, equipment, medium and product thereof |
| CN116308327A (en) * | 2022-12-28 | 2023-06-23 | 深圳市销邦数据技术有限公司 | Self-service cashing system and method based on RFID technology |
| CN116308327B (en) * | 2022-12-28 | 2023-10-27 | 深圳市销邦数据技术有限公司 | Self-service cashing system and method based on RFID technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108280738A (en) | Method of Commodity Recommendation based on image and socialized label | |
| EP2955645B1 (en) | System for automated segmentation of images through layout classification | |
| US9330111B2 (en) | Hierarchical ranking of facial attributes | |
| CN104572804B (en) | A kind of method and its system of video object retrieval | |
| US8990199B1 (en) | Content search with category-aware visual similarity | |
| CN111310662B (en) | Flame detection and identification method and system based on integrated deep network | |
| US9424466B2 (en) | Shoe image retrieval apparatus and method using matching pair | |
| WO2017181892A1 (en) | Foreground segmentation method and device | |
| CN104281572B (en) | A kind of target matching method and its system based on mutual information | |
| US20140079316A1 (en) | Segmentation co-clustering | |
| Zhang et al. | Salient object detection via compactness and objectness cues | |
| CN110008844A (en) | A kind of long-term gesture tracking method of KCF merging SLIC algorithm | |
| Kumar et al. | Leaf classification based on shape and edge feature with k-NN classifier | |
| CN111091129A (en) | Image salient region extraction method based on multi-color characteristic manifold sorting | |
| Huo et al. | Semisupervised learning based on a novel iterative optimization model for saliency detection | |
| Ji et al. | Fast segmentation of colour apple image under all-weather natural conditions for vision recognition of picking robots | |
| CN107423771B (en) | Two-time-phase remote sensing image change detection method | |
| CN108388869B (en) | Handwritten data classification method and system based on multiple manifold | |
| CN110647897A (en) | Zero sample image classification and identification method based on multi-part attention mechanism | |
| CN108921872A (en) | A kind of robustness visual target tracking method suitable for long-range tracking | |
| Xia et al. | Face clustering in photo album | |
| CN111814853B (en) | Decorrelation clustering method and device under data selection deviation | |
| Blandon et al. | An enhanced and interpretable feature representation approach to support shape classification from binary images | |
| Sun et al. | Non-negative discriminative collective target nearest-neighbor representation | |
| Zhang et al. | Data-driven saliency region detection based on undirected graph ranking |
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 |
Application publication date: 20180713 |
|
| RJ01 | Rejection of invention patent application after publication |