CN109446413A - Serializing recommended method based on item associations relationship - Google Patents
Serializing recommended method based on item associations relationship Download PDFInfo
- Publication number
- CN109446413A CN109446413A CN201811116273.7A CN201811116273A CN109446413A CN 109446413 A CN109446413 A CN 109446413A CN 201811116273 A CN201811116273 A CN 201811116273A CN 109446413 A CN109446413 A CN 109446413A
- Authority
- CN
- China
- Prior art keywords
- article
- matrix
- serializing
- user
- recommended
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- Computational Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Molecular Biology (AREA)
Abstract
The present invention provides a kind of serializing recommended method based on item associations relationship, obtains the interaction data between user and article from network-side；The symbiosis figure of interaction data building article is enabled, the symbiosis figure is indicated with incidence relation figure adjacency matrix；It enables incidence relation figure adjacency matrix carry out figure convolution operation, obtains the Relating Characteristic of article；It enables the Relating Characteristic of article input recommended models to be trained；Recommended models output sequenceization is enabled to recommend.Service can be provided for the serializing recommendation of user with serializing recommended models joint training to the excavation for the item associations sexual intercourse implied in user behavior；The incidence relation between article is excavated using user and article interaction data, and the expression of vectorization has been carried out to incidence relation, intuitively and objectively shows the Relating Characteristic of each article, associated article is analyzed using Euclidean distance；In a manner of end to end and recommended models coorinated training is serialized, final serializing recommendation service is provided for user.
Description
Technical field
The present invention relates to information recommendation fields, and in particular, to the serializing recommended method based on item associations relationship, especially
It is to focus on the incidence relation between the article in serializing recommendation according to figure convolution theory, and end-to-end trained item associations are closed
Pastern point and serializing recommended method, final serializing recommendation service is provided for user.
Background technique
Recommend to be used as a kind of information filtering task, have been extended in many applications in the real world, such as commodity
Recommend, video recommendations.Most of recommender system is established in the case where user preference is static constant hypothesis at present.So
And user preference is with time not short dynamic change.Therefore, begin to focus on how to utilize user-there are many work at present
The serialization information of article interbehavior.
According in the way of user-article interbehavior serialization information, current method is broadly divided into two major classes.
Conventional method using serialization information as a kind of contextual feature, this mode cannot model the long-term serializing of high-order according to
Rely.Recent serializing recommended method then utilizes Recognition with Recurrent Neural Network RNN to model the dynamic change in this serializing, with more preferable
Excavation sequence in changing pattern.
Nevertheless, the behavior sequence of real world is usually while closing comprising serializing behavioral characteristics and static article
Gang mould formula.For example, a user shown in Fig. 2 buys the consumer behavior sequence of mobile phone.User has purchased to be had purchased in step1
One iPhone 6s simultaneously has purchased various Cellphone Accessories after step1.This transformation comprising serializing on dynamic change because
The relevance of mobile phone and accessory.However, without obvious between Cellphone Accessories purchase sequence (step2, step3, step4, step5)
Serializing on dependence.In fact, showing stronger stationary article relevance between each Cellphone Accessories rather than sequence
Dynamic change in columnization.Meanwhile the serializing recommended method of this RNN class is difficult to capture this stationary article relevance.
Be currently based on RNN serializing recommended method do not consider it is interrelated between article in customer consumption sequence
Property.How to capture the item associations feature in user behavior sequence and be primarily present two difficult points: how to be filled from user behavior
Divide the incidence relation excavated between article；How to allow article incidence relation and serializing the end-to-end coorinated training of recommended method.
Summary of the invention
For the defects in the prior art, the object of the present invention is to provide a kind of serializings based on item associations relationship to push away
Recommend method.
A kind of serializing recommended method based on item associations relationship provided according to the present invention, includes the following steps, obtains
It takes interaction data step: obtaining the interaction data between user and article from network-side；It constructs symbiosis figure step: enabling interaction
Data construct the symbiosis figure of article, and the symbiosis figure is indicated with incidence relation figure adjacency matrix；Picture scroll accumulates net
Network step: it enables incidence relation figure adjacency matrix carry out figure convolution operation, obtains the Relating Characteristic of article；Recommended models training step
It is rapid: to enable the Relating Characteristic of article input recommended models and be trained；Serializing recommendation step: recommended models output sequence is enabled
Recommend.
Preferably, the building symbiosis figure step mainly includes building frequency matrix step, building incidence relation figure
Adjacency matrix step；Building frequency matrix step: it counts between article two-by-two and occurs jointly in same user's history record
Frequency, the frequency is indicated in the form of frequency matrix；Building incidence relation figure adjacency matrix step: frequency matrix is enabled to carry out
Binary conversion treatment indicates the binary conversion treatment result with incidence relation figure adjacency matrix.
Preferably, the figure convolutional network step includes characteristic value definition step: enabling X ∈ RN×C, X is that the feature of article is defeated
Enter value, if article sets X=I without feature input valueN, wherein INThe unit matrix for being N for size, the row of N representing matrix R
Number, the columns of C representing matrix R；R indicates eigenvalue matrix.
Form Laplacian Matrix step: described according to the Laplacian Matrix of picture scroll product theoretical informatics symbiosis figure
Laplacian Matrix L is as follows:
In formula, D is the degree matrix of diagonalization；A is incidence relation figure adjacency matrix；UΛUTFor the Eigenvalues Decomposition of L；U is
Eigenvectors matrix；Λ=diag ([λ0,...,λn-1])∈RN×NFor the eigenvalue matrix of diagonalization；UTIt is eigenvectors matrix
The transposition of U；λi, i=0,1 ..., n-1 is characterized n characteristic value after value is decomposed.
Form normalization Laplacian Matrix step: normalized Laplacian MatrixThen it is defined by the formula:
In formula, λmaxIt is the maximum value in eigenvalue matrix Λ；
Picture scroll accumulates Operation Definition step: enabling X ∈ RN×CFor the feature input value of article, figure convolution operation is then defined by the formula:
In formula,It indicates last picture scroll product operating result output, is the matrix of a N × d, the matrix
In value be article Relating Characteristic；gθExpression parameter is the convolution function of θ；* convolution operation symbol is indicated；Table
Show Chebyshev polynomials, subscript k is the order in Chebyshev polynomials, and K is the highest order in Chebyshev polynomials；
Θk∈RC×dIt is the parameter to be learned of k rank；D is the article latent space dimension of convolution.
Preferably, the recommended models training step includes obtaining Relating Characteristic step: batch that setting serializing is recommended
Article index, from the Relating Characteristic of articleIn find out the article vector expression for corresponding to batch article index, be denoted as
It carries out as follows:
In formula, flookup() indicates search operation；Index is batch article index；
Recommended models amendment step: serializing model is enabled to be corrected as follows；
(e) it updates door and exports zt:
(f) resetting door exports rt:
(g) candidate state of activation
(h) the output state h of t stept:
Wherein W={ Wz,Wr,Wc}∈R3×d×dWith U={ Uz,Ur,Uc}∈R3×d×dIt is the network ginseng inside GRU unit
Number；σ is sigmoid nonlinear activation function；D is the article latent space dimension of convolution；Indicate batch training picture scroll of t step
Product expression output；Symbol ⊙ representing matrix element corresponds to product；
Recommended models export step: recommended models export consumer articles preference, using following formula；
In formula, u indicates that user, p indicate that positive article, q indicate negative article,Indicate user u to the preference of positive article p compared with
In the preference of negative article q,Indicate user's expression that user u is walked in t,It indicates to walk the relevance expression of positive article in t+1,It indicates to walk the relevance expression of negative article in t+1；
Export amendment step:It is modified as follows:
In formula,Indicate that user u is greater than the probability to the preference of article q to the preference of article p, Ω is recommended models
Parameter,It is the regular terms of recommended models, λΩ>=0 is regularization coefficient.
Preferably, the binary conversion treatment is given threshold, if the value of frequency matrix is more than or equal to threshold value, the two-value
Change processing result and is set as 1；If the value of frequency matrix is less than threshold value, the binary conversion treatment result is set as 0.
Preferably, in the picture scroll product Operation Definition step, Chebyshev polynomialsIt is sparse matrix,
It is obtained by deep learning frame,Loop iteration obtains in such a way that such as following formula defines:
In formula, k takes the integer greater than 2.
Compared with prior art, the present invention have it is following the utility model has the advantages that
1, the present invention can be to the excavation for the item associations sexual intercourse implied in user behavior, with serializing recommended models connection
Training is closed, provides service for the serializing recommendation of user；
2, the present invention excavates the incidence relation between article using user and article interaction data, and carries out to incidence relation
The expression of vectorization intuitively and objectively shows the Relating Characteristic of each article, is carried out using Euclidean distance to associated article
Analysis；
3, the present invention in a manner of end to end and serializes recommended models coorinated training, and final serializing is provided for user
Recommendation service.
Detailed description of the invention
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon:
Fig. 1 is block flow diagram of the present invention；
Fig. 2 is customer consumption sequence chart；
Fig. 3 is the serializing recommended method model structure based on item associations；
Fig. 4 is the serializing recommended method article characteristics analysis chart based on item associations.
Specific embodiment
The present invention is described in detail combined with specific embodiments below.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, several changes and improvements can also be made.These belong to the present invention
Protection scope.
The present invention is first according to user behavior sequence statistic article to the frequency appeared in a user's history jointly time
Number, the figure for then taking threshold value k to indicate its binaryzation at adjacency matrix.The association between article is excavated then in conjunction with figure convolution theory
Relationship is provided most using the graph expression Network Search layer of design with recommended method is end-to-end is trained with serializing for user
Whole serializing recommendation service.
A kind of serializing recommended method based on item associations relationship provided according to the present invention, as shown in Figure 1, include with
Lower step obtains interaction data step: obtaining the interaction data between user and article from network-side, crawls user and contain with article
The interaction data of time；Building symbiosis figure step: the symbiosis figure of interaction data building article, the symbiosis are enabled
Figure is indicated with incidence relation figure adjacency matrix；Figure convolutional network step: incidence relation figure adjacency matrix is enabled to carry out picture scroll product
Operation, obtains the Relating Characteristic of article；Recommended models training step: it enables the Relating Characteristic of article input recommended models and carries out
Training；Serializing recommendation step: recommended models output sequenceization is enabled to recommend.
Specifically, the building symbiosis figure step mainly includes building frequency matrix step, building incidence relation figure
Adjacency matrix step；Building frequency matrix step: it counts between article two-by-two and occurs jointly in same user's history record
Frequency, the frequency is indicated in the form of frequency matrix；Building incidence relation figure adjacency matrix step: frequency matrix is enabled to carry out
Binary conversion treatment indicates the binary conversion treatment result with incidence relation figure adjacency matrix.The binary conversion treatment is setting
Threshold value, if the value of frequency matrix is more than or equal to threshold value, the binary conversion treatment result is set as 1；If the value of frequency matrix is small
In threshold value, then the binary conversion treatment result is set as 0.Preferably, threshold value value is 10.
As shown in figure 3, item associations relational graph adjacency matrix A is obtained according to the user crawled-article interaction data statistics,
Count the frequency occurred jointly in a user's history record between article two-by-two, the frequency square of building article between any two
Battle array, takes threshold value k (k=10) to frequency matrix and the matrix binaryzation (is set as 1 greater than k, less than being set as 0) into adjoining for k
Matrix A.
Specifically, the figure convolutional network step includes characteristic value definition step: enabling X ∈ RN×CIt is inputted for the feature of article
Value, if article sets X=I without feature input valueN, wherein INThe unit matrix for being N for size, the line number of N representing matrix R, C
The columns of representing matrix R；R indicates eigenvalue matrix.It forms Laplacian Matrix step: being closed according to picture scroll product theoretical informatics symbiosis
It is the Laplacian Matrix of figure, the Laplacian Matrix L is as follows:
In formula, D is the degree matrix of diagonalization；A is incidence relation figure adjacency matrix；UΛUTFor the Eigenvalues Decomposition of L；U is
Eigenvectors matrix；Λ=diag ([λ0,...,λn-1])∈RN×NFor the eigenvalue matrix of diagonalization；UTIt is eigenvectors matrix
The transposition of U；λi, i=0,1 ..., n-1 indicates n characteristic value after Eigenvalues Decomposition.
Form normalization Laplacian Matrix step: normalized Laplacian MatrixThen it is defined by the formula:
In formula, λmaxIt is the maximum value in eigenvalue matrix Λ；
Picture scroll accumulates Operation Definition step: enabling X ∈ RN×CFor the feature input value of article, figure convolution operation is then defined by the formula:
In formula,It indicates last picture scroll product operating result output, is the matrix of a N × d, the matrix
In value be article Relating Characteristic；gθExpression parameter is the convolution function of θ；* convolution operation symbol is indicated；It indicates
Chebyshev polynomials, subscript k are the orders in Chebyshev polynomials, and K is the highest order in Chebyshev polynomials；
Θk∈RC×dIt is the parameter to be learned of k rank；D is the article latent space dimension of convolution, and the latent space dimension is hidden variable vector
Dimension.
Specifically, in the picture scroll product Operation Definition step, Chebyshev polynomialsIt is sparse matrix,
It is obtained by deep learning frame,Loop iteration obtains in such a way that such as following formula defines:
In formula, k takes the integer greater than 2.
Specifically, the recommended models training step includes obtaining Relating Characteristic step: batch that setting serializing is recommended
Article index, from the Relating Characteristic of articleIn find out the article vector expression for corresponding to batch article index, be denoted as
It carries out as follows:
In formula, flookup() indicates search operation；Index is batch article index；
Recommended models amendment step: serializing model is enabled to be corrected as follows；
(i) it updates door and exports zt:
(j) resetting door exports rt:
(k) candidate state of activation
(l) the output state h of t stept:
Wherein W={ Wz,Wr,Wc}∈R3×d×dWith U={ Uz,Ur,Uc}∈R3×d×dIt is the network ginseng inside GRU unit
Number；σ is sigmoid nonlinear activation function；The article latent space dimension of d expression convolution；H indicates output state；T indicates t
Step；Indicate batch training picture scroll product expression output of t step；Symbol ⊙ representing matrix element corresponds to product；
Recommended models export step: recommended models export consumer articles preference, using following formula；
In formula, u indicates that user, p indicate that positive article, q indicate negative article,Indicate user u to the preference of positive article p compared with
In the preference of negative article q,Indicate user's expression that user u is walked in t,It indicates to walk the relevance expression of positive article in t+1,It indicates to walk the relevance expression of negative article in t+1；
Export amendment step:It is modified as follows:
In formula,Indicate that user u is greater than the probability to the preference of article q to the preference of article p, Ω is recommended models
Parameter,It is the regular terms of recommended models, λΩ>=0 is regularization coefficient.
A kind of computer readable storage medium for being stored with computer program provided according to the present invention, the computer journey
The step of above method is realized when sequence is executed by processor.
Preferably, recommended models training step is associated with sexual intercourse using existing serializing recommended method and front article
Feature combines, and can use GRU4Rec model, but not limited to this serializing model.It should be noted that most preferred sequence pushes away
It recommends model and supports batch mode of training, but for picture scroll product network portion, the original association of article can be destroyed using batch training
It is relationship.Therefore, both for end-to-end combination, a kind of figure vector expression is devised in the present invention and searches layer, with it is subsequent can
The serializing recommended method for criticizing training combines.As shown in figure 3, concrete operations mode is as follows:
(a) batch article index [1,3,4,5,7,8, N] in a serializing recommendation is given
(b) from the Relating Characteristic of articleIn find out corresponding to article index [1,3,4,5,7,8, N] article to
Amount expression, is denoted as
This can gradient passback operation can be indicated with following formula:
Wherein flookup() indicates search operation defined above.It is the item associations feature of batch processing, s is batch
The size of processing.
After obtaining the item associations feature of batch processing, serializing model may be used and carry out serializing recommendation.With
For GRU4Rec, it is expressed as follows in the information flow changed in model:
(a) it updates door and exports zt:
(b) resetting door exports rt:
(c) candidate state of activation
(d) the output state h of t stept:
Wherein W={ Wz,Wr,Wc}∈R3×d×dWith U={ Uz,Ur,Uc}∈R3×d×dIt is the network ginseng inside GRU unit
Number.σ is sigmoid nonlinear activation function.
Assuming that u indicates user, p, q respectively indicate positive and negative article in implicit feedback.So in step t, user u
Preference to the preference of positive article p compared with negative article qIt can be denoted as following formula:
WhereinIndicate user's expression that user u is walked in t.It is illustrated respectively in the association that t+1 walks positive article
Property expression and negative article relevance expression.The target loss function of last model is as follows:
WhereinIndicate that user u is greater than the probability to the preference of article q to the preference of article p.Ω is entire model
Parameter.It is the regular terms of model.λΩ>=0 is regularization coefficient.
Serialize recommended method step: the step can be using the relevance of existing serializing recommended method and front article
Relationship characteristic combines.Here by taking GRU4Rec model as an example, but not limited to this serializing model.It should be noted that common sequence
Columnization recommended models support batch mode of training, but for picture scroll product network portion, it is original to destroy article using batch training
Association be relationship.So both for end-to-end combination, a kind of figure vector expression is devised in the present invention and searches layer, with it is rear
Continuous trained serializing recommended method of criticizing combines.
Preferably, in serializing recommendation step, after the serializing recommended models based on item associations sexual intercourse train,
A given user is current or history consumes sequence, the model prediction user in next step to the preference-score of all items, according to
This is scored at candidate item and sorts from large to small, k possible goods for consumption before recommending.
For convenience of displaying recommendation gain effect of the invention, by taking the article expression acquired as an example.In TaoBao data set
On, cosine similarity matrix two-by-two is calculated separately between article, remembers that similarity matrix is S two-by-two for the article of the method for the present invention2, institute
The article of the serializing recommended method GRU4Rec used two-by-two similarity matrix for S1.So take out preceding 8 maximum difference article
To S+=S2-S1, corresponding result displaying is in (a)-(h) of Fig. 4.The similarity size of histogram graph representation corresponding method.It is similar
Ground takes out preceding 8 maximum difference article to S-=S1-S2, corresponding result displaying is in (i)-(p) of Fig. 4.
It can be seen from the figure that the method for invention is easier relevance, complementary article in (a)-(h) of Fig. 4
It associates.Such as skirt and slippers in Fig. 4 (b) (d).But as can be seen that the method for GRU4Rec from (i) of Fig. 4-(p)
This relevance can not then be captured, on the contrary, he be more concerned about will be generic in item associations get up such as (i) (j) in Fig. 4,
Or some incoherent articles are associated such as (l) (k) in Fig. 4.This result introduces item associations with us
Imagine consistent.Similarly, the present invention also changes closing to reality and applies because user is after generating one-time-consumption behavior, tends to purchase
Buy relevance, complementary article, rather than same article.
One skilled in the art will appreciate that in addition to realizing system provided by the invention in a manner of pure computer readable program code
It, completely can be by the way that method and step be carried out programming in logic come so that provided by the invention other than system, device and its modules
System, device and its modules are declined with logic gate, switch, specific integrated circuit, programmable logic controller (PLC) and insertion
The form of controller etc. realizes identical program.So system provided by the invention, device and its modules may be considered that
It is a kind of hardware component, and the knot that the module for realizing various programs for including in it can also be considered as in hardware component
Structure；It can also will be considered as realizing the module of various functions either the software program of implementation method can be Hardware Subdivision again
Structure in part.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make a variety of changes or modify within the scope of the claims, this not shadow
Ring substantive content of the invention.In the absence of conflict, the feature in embodiments herein and embodiment can any phase
Mutually combination.
Claims (6)
1. a kind of serializing recommended method based on item associations relationship, which comprises the following steps:
It obtains interaction data step: obtaining the interaction data between user and article from network-side；
Building symbiosis figure step: enabling the symbiosis figure of interaction data building article, and symbiosis figure association is closed
It is that figure adjacency matrix is indicated；
Figure convolutional network step: it enables incidence relation figure adjacency matrix carry out figure convolution operation, obtains the Relating Characteristic of article；
Recommended models training step: it enables the Relating Characteristic of article input recommended models and is trained；
Serializing recommendation step: recommended models output sequenceization is enabled to recommend.
2. the serializing recommended method according to claim 1 based on item associations relationship, which is characterized in that the building
Symbiosis figure step mainly includes building frequency matrix step, building incidence relation figure adjacency matrix step；
Building frequency matrix step: counting the frequency occurred jointly in same user's history record between article two-by-two, will
The frequency is indicated in the form of frequency matrix；
Building incidence relation figure adjacency matrix step: frequency matrix is enabled to carry out binary conversion treatment, by the binary conversion treatment result
It is indicated with incidence relation figure adjacency matrix.
3. the serializing recommended method according to claim 1 based on item associations relationship, which is characterized in that the picture scroll
Accumulating network step includes:
Characteristic value definition step: X ∈ R is enabledN×C, X is the feature input value of article, if article sets X=without feature input value
IN, wherein INThe unit matrix for being N for size, the line number of N representing matrix R, the columns of C representing matrix R；R indicates characteristic value square
Battle array.
Form Laplacian Matrix step: according to the Laplacian Matrix of picture scroll product theoretical informatics symbiosis figure, the drawing is general
Lars matrix L is as follows:
In formula, D is the degree matrix of diagonalization；A is incidence relation figure adjacency matrix；UΛUTFor the Eigenvalues Decomposition of L；U is characterized
Vector matrix；Λ=diag ([λ0,...,λn-1])∈RN×NFor the eigenvalue matrix of diagonalization；UTIt is eigenvectors matrix U
Transposition；λi, i=0,1 ..., n-1 is characterized n characteristic value after value is decomposed.
Form normalization Laplacian Matrix step: normalized Laplacian MatrixThen it is defined by the formula:
In formula, λmaxIt is the maximum value in eigenvalue matrix Λ；
Picture scroll accumulates Operation Definition step: enabling X ∈ RN×CFor the feature input value of article, figure convolution operation is then defined by the formula:
In formula, Xgθ∈RN×dIt indicates last picture scroll product operating result output, is the matrix of a N × d, the value in the matrix
It is the Relating Characteristic of article；gθExpression parameter is the convolution function of θ；* convolution operation symbol is indicated；Indicate Qie Bixue
Husband's multinomial, subscript k are the orders in Chebyshev polynomials, and K is the highest order in Chebyshev polynomials；Θk∈RC×d
It is the parameter to be learned of k rank；D is the article latent space dimension of convolution.
4. the serializing recommended method according to claim 3 based on item associations relationship, which is characterized in that the recommendation
Model training step includes:
Obtain Relating Characteristic step: batch article index that setting serializing is recommended, from the Relating Characteristic X of articlegθMiddle lookup
The article vector expression for corresponding to batch article index out, is denoted as xgθ, it carries out as follows:
xgθ=flookup(Xgθ,xindex)
In formula, flookup() indicates search operation；Index is batch article index；
Recommended models amendment step: serializing model is enabled to be corrected as follows；
(a) it updates door and exports zt:
(b) resetting door exports rt:
(c) candidate state of activation
(d) the output state h of t stept:
Wherein W={ Wz,Wr,Wc}∈R3×d×dWith U={ Uz,Ur,Uc}∈R3×d×dIt is the network parameter inside GRU unit；σ is
Sigmoid nonlinear activation function；D is the article latent space dimension of convolution；Indicate batch training picture scroll product expression of t step
Output；Symbol ⊙ representing matrix element corresponds to product；
Recommended models export step: recommended models export consumer articles preference, using following formula；
In formula, u indicates that user, p indicate that positive article, q indicate negative article,Indicate user u to the preference of positive article p compared with negative object
The preference of product q,Indicate user's expression that user u is walked in t,It indicates to walk the relevance expression of positive article in t+1,Table
Show and walks the relevance expression of negative article in t+1；
Export amendment step:It is modified as follows:
In formula,Indicating that user u is greater than the probability to the preference of article q to the preference of article p, Ω is recommended models parameter,It is the regular terms of recommended models, λΩ>=0 is regularization coefficient.
5. the serializing recommended method according to claim 2 based on item associations relationship, which is characterized in that the two-value
Change processing is given threshold, if the value of frequency matrix is more than or equal to threshold value, the binary conversion treatment result is set as 1；If frequency
The value of rate matrix is less than threshold value, then the binary conversion treatment result is set as 0.
6. the serializing recommended method according to claim 3 based on item associations relationship, which is characterized in that the picture scroll
In product Operation Definition step, Chebyshev polynomialsIt is sparse matrix,It is obtained by deep learning frame,Loop iteration obtains in such a way that such as following formula defines:
In formula, k takes the integer greater than 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811116273.7A CN109446413B (en) | 2018-09-25 | 2018-09-25 | Serialized recommendation method based on article association relation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811116273.7A CN109446413B (en) | 2018-09-25 | 2018-09-25 | Serialized recommendation method based on article association relation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109446413A true CN109446413A (en) | 2019-03-08 |
| CN109446413B CN109446413B (en) | 2021-06-01 |
Family
ID=65530745
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811116273.7A Active CN109446413B (en) | 2018-09-25 | 2018-09-25 | Serialized recommendation method based on article association relation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109446413B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110264277A (en) * | 2019-06-25 | 2019-09-20 | 杭州网易再顾科技有限公司 | Data processing method and device, medium and the calculating equipment executed by calculating equipment |
| CN110443574A (en) * | 2019-07-25 | 2019-11-12 | 昆明理工大学 | Entry convolutional neural networks evaluation expert's recommended method |
| CN110674407A (en) * | 2019-09-30 | 2020-01-10 | 杭州电子科技大学 | Hybrid recommendation method based on graph convolution neural network |
| CN110705613A (en) * | 2019-09-19 | 2020-01-17 | 创新奇智(青岛)科技有限公司 | Object classification method |
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
| CN110991601A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Neural network recommendation method based on multi-user behaviors |
| CN111144976A (en) * | 2019-12-10 | 2020-05-12 | 支付宝(杭州)信息技术有限公司 | Training method and device for recommendation model |
| CN111598644A (en) * | 2020-04-01 | 2020-08-28 | 华瑞新智科技（北京）有限公司 | Article recommendation method, device and medium |
| CN111782925A (en) * | 2019-04-04 | 2020-10-16 | 阿里巴巴集团控股有限公司 | Item recommendation method, device, equipment, system and readable storage medium |
| CN113297490A (en) * | 2021-06-04 | 2021-08-24 | 西南大学 | Bidirectional recommendation method based on graph convolution neural network |
| CN113297839A (en) * | 2021-03-03 | 2021-08-24 | 浪潮云信息技术股份公司 | Innovative subject mining method in patent technical field |
| CN111782925B (en) * | 2019-04-04 | 2024-05-14 | 阿里巴巴集团控股有限公司 | Item recommendation method, device, equipment, system and readable storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE602007001862D1 (en) * | 2006-03-10 | 2009-09-17 | Panasonic Corp | Device for searching in a fixed codebook |
| CN104933143A (en) * | 2015-06-18 | 2015-09-23 | 北京京东尚科信息技术有限公司 | Method and device for acquiring recommended object |
| CN105608219A (en) * | 2016-01-07 | 2016-05-25 | 上海通创信息技术有限公司 | Stream-oriented recommended engine, recommendation system and recommendation method based on clustering |
| CN107622427A (en) * | 2016-07-13 | 2018-01-23 | 阿里巴巴集团控股有限公司 | The method, apparatus and system of deep learning |
-
2018
- 2018-09-25 CN CN201811116273.7A patent/CN109446413B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE602007001862D1 (en) * | 2006-03-10 | 2009-09-17 | Panasonic Corp | Device for searching in a fixed codebook |
| CN104933143A (en) * | 2015-06-18 | 2015-09-23 | 北京京东尚科信息技术有限公司 | Method and device for acquiring recommended object |
| CN105608219A (en) * | 2016-01-07 | 2016-05-25 | 上海通创信息技术有限公司 | Stream-oriented recommended engine, recommendation system and recommendation method based on clustering |
| CN107622427A (en) * | 2016-07-13 | 2018-01-23 | 阿里巴巴集团控股有限公司 | The method, apparatus and system of deep learning |
Non-Patent Citations (2)
| Title |
|---|
| H. ZHAO ET AL.: "Recommendation with Hybrid Interest Model", 《2018 IEEE INTERNATIONAL CONFERENCE ON DATA MINING WORKSHOPS (ICDMW)》 * |
| 游运 等: "考虑对象关联关系的多样化商品推荐方法", 《计算机工程与应用》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111782925B (en) * | 2019-04-04 | 2024-05-14 | 阿里巴巴集团控股有限公司 | Item recommendation method, device, equipment, system and readable storage medium |
| CN111782925A (en) * | 2019-04-04 | 2020-10-16 | 阿里巴巴集团控股有限公司 | Item recommendation method, device, equipment, system and readable storage medium |
| CN110264277A (en) * | 2019-06-25 | 2019-09-20 | 杭州网易再顾科技有限公司 | Data processing method and device, medium and the calculating equipment executed by calculating equipment |
| CN110264277B (en) * | 2019-06-25 | 2022-08-16 | 杭州网易再顾科技有限公司 | Data processing method and device executed by computing equipment, medium and computing equipment |
| CN110443574A (en) * | 2019-07-25 | 2019-11-12 | 昆明理工大学 | Entry convolutional neural networks evaluation expert's recommended method |
| CN110443574B (en) * | 2019-07-25 | 2023-04-07 | 昆明理工大学 | Recommendation method for multi-project convolutional neural network review experts |
| CN110705613A (en) * | 2019-09-19 | 2020-01-17 | 创新奇智(青岛)科技有限公司 | Object classification method |
| CN110705613B (en) * | 2019-09-19 | 2021-06-11 | 创新奇智(青岛)科技有限公司 | Object classification method |
| CN110674407B (en) * | 2019-09-30 | 2021-10-15 | 杭州电子科技大学 | Hybrid recommendation method based on graph convolution neural network |
| CN110674407A (en) * | 2019-09-30 | 2020-01-10 | 杭州电子科技大学 | Hybrid recommendation method based on graph convolution neural network |
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
| CN110879864B (en) * | 2019-10-29 | 2022-06-07 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
| CN110991601A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Neural network recommendation method based on multi-user behaviors |
| CN110991601B (en) * | 2019-11-29 | 2023-04-21 | 中山大学 | Neural network recommendation method based on multi-user behavior |
| CN111144976A (en) * | 2019-12-10 | 2020-05-12 | 支付宝(杭州)信息技术有限公司 | Training method and device for recommendation model |
| CN111598644A (en) * | 2020-04-01 | 2020-08-28 | 华瑞新智科技（北京）有限公司 | Article recommendation method, device and medium |
| CN111598644B (en) * | 2020-04-01 | 2023-05-02 | 华瑞新智科技（北京）有限公司 | Article recommendation method, device and medium |
| CN113297839A (en) * | 2021-03-03 | 2021-08-24 | 浪潮云信息技术股份公司 | Innovative subject mining method in patent technical field |
| CN113297839B (en) * | 2021-03-03 | 2023-04-21 | 浪潮云信息技术股份公司 | Innovative subject mining method in patent technical field |
| CN113297490A (en) * | 2021-06-04 | 2021-08-24 | 西南大学 | Bidirectional recommendation method based on graph convolution neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109446413B (en) | 2021-06-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109446413A (en) | Serializing recommended method based on item associations relationship | |
| Giryes et al. | Deep neural networks with random Gaussian weights: A universal classification strategy? | |
| Shen et al. | Higher order energies for image segmentation | |
| CN108287864B (en) | Interest group dividing method, device, medium and computing equipment | |
| Jia et al. | Label distribution learning with label correlations on local samples | |
| CN111882040B (en) | Convolutional neural network compression method based on channel number search | |
| van Wyk et al. | A pocs-based graph matching algorithm | |
| Guo et al. | Efficient sparse low-rank tensor completion using the Frank-Wolfe algorithm | |
| CN112613602A (en) | Recommendation method and system based on knowledge-aware hypergraph neural network | |
| CN109446420B (en) | Cross-domain collaborative filtering method and system | |
| Pourkamali-Anaraki et al. | Efficient solvers for sparse subspace clustering | |
| US11521133B2 (en) | Method for large-scale distributed machine learning using formal knowledge and training data | |
| Zhang et al. | Depth-based subgraph convolutional auto-encoder for network representation learning | |
| KR102412158B1 (en) | Keyword extraction and analysis method to expand market share in the open market | |
| Xu et al. | A semantic segmentation method with category boundary for Land Use and Land Cover (LULC) mapping of Very-High Resolution (VHR) remote sensing image | |
| Liu et al. | PHD: A probabilistic model of hybrid deep collaborative filtering for recommender systems | |
| Liu et al. | Improving human action recognitionby temporal attention | |
| Zhang | Simultaneous optimization for robust correlation estimation in partially observed social network | |
| Jawanpuria et al. | Efficient robust optimal transport: formulations and algorithms | |
| CN108389113B (en) | Collaborative filtering recommendation method and system | |
| Saboori et al. | Adversarial discriminative active deep learning for domain adaptation in hyperspectral images classification | |
| Shah et al. | Irregular neutrosophic graphs | |
| CN115618098B (en) | Cold-chain logistics recommendation method and device based on knowledge enhancement and cavity convolution | |
| CN114417161B (en) | Virtual article time sequence recommendation method, device, medium and equipment based on special-purpose map | |
| CN116049691A (en) | Model conversion method, device, electronic equipment and storage medium |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |