CN114880583A - Cross-domain social recommendation method based on self-supervision learning - Google Patents
Cross-domain social recommendation method based on self-supervision learning Download PDFInfo
- Publication number
- CN114880583A CN114880583A CN202210511621.0A CN202210511621A CN114880583A CN 114880583 A CN114880583 A CN 114880583A CN 202210511621 A CN202210511621 A CN 202210511621A CN 114880583 A CN114880583 A CN 114880583A
- Authority
- CN
- China
- Prior art keywords
- user
- domain
- product
- formula
- social
- 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
- 238000000034 method Methods 0.000 title claims abstract description 36
- 239000011159 matrix material Substances 0.000 claims abstract description 56
- 230000002776 aggregation Effects 0.000 claims abstract description 7
- 238000004220 aggregation Methods 0.000 claims abstract description 7
- 239000013598 vector Substances 0.000 claims description 86
- 230000006870 function Effects 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 6
- 230000003993 interaction Effects 0.000 claims description 5
- 239000000126 substance Substances 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 3
- 230000002159 abnormal effect Effects 0.000 abstract description 2
- 238000012545 processing Methods 0.000 abstract description 2
- 238000010276 construction Methods 0.000 abstract 1
- 238000005457 optimization Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Strategic Management (AREA)
- Biomedical Technology (AREA)
- Development Economics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a cross-domain social product recommendation method based on self-supervision learning, which comprises the following steps: 1. constructing heterogeneous data: a scoring matrix of the product by the user and a social relationship matrix between the user and the user; 2. processing the heterogeneous data through a heterogeneous graph construction layer to obtain a heterogeneous graph network of a user-product; 3. obtaining a user cooperation matrix and a product cooperation matrix through a single hot coding layer; 4. performing feature propagation through the graph convolution layer; 5. constructing a node feature aggregation layer; 6. constructing a node prediction layer to recommend products; 7. and constructing a self-supervision learning layer to carry out mutual information maximization learning on the local characteristics and the global characteristics of the nodes until the product recommendation effect of the social cold start user in the social domain is optimal. The method can fully mine the global information of the whole abnormal graph, reduce the dependence of the model on the number of bridge users and learn better bridge user characteristics, thereby improving the product recommendation performance on social cold-start users.
Description
Technical Field
The invention relates to the field of cross-domain recommendation and social recommendation, in particular to a cross-domain social recommendation method based on self-supervision learning.
Background
Social recommendation effectively alleviates the problems of data sparsity, system cold start and the like. According to social homogeneity theory, users tend to buy products recommended by their friends rather, and there are often more similar interest preferences among friends. Therefore, compared with the traditional collaborative filtering recommendation algorithm, the social recommendation method has more information so as to more effectively model the interest preference of the user and improve the recommendation performance of the recommendation system. However, current social recommendation methods all have a strict constraint that a user must have a social relationship based on historical product interaction records, so that the user can learn the potential interest and preference of the user, which is not in accordance with a real application scenario. In a real scenario, most users often have only social relationships without historical product interaction records, and we define this group as a social cold-start user. Therefore, how to recommend products to a large number of social cold-start users becomes a difficult problem which needs to be solved urgently at present.
One possible approach is to use the idea of cross-domain recommendation to address the problem of product recommendation to a large number of social cold-start users. The main idea of cross-domain recommendation is to use a source domain with richer information content as an auxiliary domain and transfer valuable information to a target domain, so as to improve the performance of a recommendation system on the target domain. However, the current cross-domain recommendation methods are not suitable for the proposed cross-domain social recommendation scenario, because the methods depend on the number of bridge users to a large extent, which may cause that when the number of bridge users is small, the model becomes fragile, and the recommendation effect on the social cold-start user is not accurate. Therefore, how to capture more information from a very limited bridge user is the focus of the present invention.
Disclosure of Invention
The invention provides a cross-domain social recommendation method based on self-supervision learning, aiming at overcoming the defects of the prior art, so that the global information of the whole abnormal graph can be fully mined, the dependence of a model on the number of bridge users is reduced, and the bridge user representation with better learning is realized, thereby improving the product recommendation performance on social cold-start users.
The invention adopts the following technical scheme for solving the technical problems:
the invention relates to a cross-domain social product recommendation method based on self-supervision learning, which is characterized by comprising the following steps of:
step 1, constructing heterogeneous data, comprising: a scoring matrix R of the product by the user and a social relationship matrix S between the users are as follows:
let U I Represents a set of users over an information domain, andwherein the content of the first and second substances,indicating the a-th user in the information field,representing the b-th user in the information domain, M represents the total number of users in the information domain, a is more than or equal to 1, and b is more than or equal to M;
let V denote the product set on the information field, and V ═ V 1 ,…v i ,…v j ,…v L In which v is i Indicating the ith product in the information field, v j Indicating the jth product in the information field, L indicating the letterTotal number of products on the domain;
let R ai Representing the a-th userFor the ith product v i The scoring value of (2) is that the scoring matrix of the product by the user on the information domain is R ═ { R ═ R ai } M×L ;
Let U S Represents a set of users on a social domain,wherein the content of the first and second substances,representing the c-th user in the social domain,representing the d-th user in the social domain, N representing the total number of users in the social domain, c being more than or equal to 1, and d being more than or equal to N;
order S cd Represents the c-th userFor the d-th userIf there is a connection between two users, let S cd If not, let S cd 0, the social relationship matrix between users in the social domain is S ═ S cd } N×N ;
Step 2, utilizing a bridge user set U coexisting in an information domain and a social domain B =U I ∩U S Constructing a heterogeneous graph network G ═ G I ∪G S Wherein G is I Representing the set U of all users in the information domain I Graph network Structure with product set V, G S Representing a set U of all users in a social domain S A graph network structure for collecting U-shaped bridge users B The total number of users in (1) is T;
and 3, obtaining a node cooperation matrix through single hot coding, wherein the method comprises the following steps: user cooperation matrix, product cooperation matrix:
step 3.1, constructing a user cooperation matrix P ═ P in a single hot coding mode for all users in the heterogeneous graph network G 1 ,…,p a ,…,p M-T ,…,p b ,…,p c ,…,p M ,…,p d ,…,p M+N-T And P ∈ R (M+N-T)×d Wherein p is a Representing a d-dimensional cooperation vector of an a-th user in the user cooperation matrix P, wherein the a-th user belongs to a non-bridge user in the information domain; p is a radical of b Representing the d-dimensional cooperation vector of the b-th user in the user cooperation matrix P, wherein the b-th user belongs to the bridge user, P d Representing a d-dimensional cooperation vector of a d-th user in the user cooperation matrix P, wherein the d-th user belongs to a non-bridge user in the social domain;
step 3.2, constructing a product cooperation matrix Q ═ Q in a single-hot coding mode for all products in the heteromorphic graph network G 1 ,…,q i ,…q L And Q ∈ R L×d Wherein q is i A d-dimensional collaborative vector representing an ith product;
and 4, carrying out characteristic propagation through the graph volume layer:
step 4.1, making the graph convolution layers have K convolution layers, and initializing the number K of the current convolution layers to be 0;
step 4.2, obtaining the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain according to the formula (1), the formula (2) and the formula (3) respectively a D-dimensional cooperation vector p of the b-th bridge user b D-dimensional cooperation vector p of the d-th non-bridge user in social domain d After passing through the (k + 1) th graph convolution layer, the corresponding user characteristic vector is obtainedObtaining a d-dimensional cooperative vector q of the ith product according to the formula (4) i The product characteristic vector obtained after passing through the (k + 1) th graph convolution layer
In the formula (1), R a Representing the set of products that the a-th non-bridge user interacted with,is the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain a When k is 0, let the user characteristic vector output from the k-th convolution layerA user feature vector is initialized randomly;
in the formula (2), R b Representing the set of products, S, interacted by the b-th bridge user b Representing a set of users with social connections to the b-th bridge user,is d-dimensional cooperation vector p of the b-th bridge user b When the user characteristic vector output by the k-th convolutional layer is equal to 0,a user feature vector is initialized randomly;
in the formula (3), S d Representing a set of users for which the d-th non-bridge user has social connections,the user feature vector is output by the d-th non-bridge user in the social domain at the kth layer convolutional layer, when k is 0,a user feature vector is initialized randomly;
in the formula (4), R i Representing a set of users who have interacted with the ith product,is the d-dimensional cooperation vector q of the ith product i When the characteristic vector of the product output by the k-th layer convolution layer is 0,the feature vector of the product is initialized randomly;
4.3, assigning K +1 to K, judging whether K is greater than K, and if so, executing the step 5; otherwise, returning to the step 4.2 for sequential execution;
step 5, constructing a node feature aggregation layer:
performing feature aggregation on the K-layer user feature vector and the product feature vector according to the formula (5) to obtain a feature vector p of any user u in the heterogeneous network G u And the feature vector q of the ith product i :
In the formula (5), AVG represents the average pooling operation,a feature vector representing the user u output by the kth convolutional layer,i-th product q representing the output of the k-th convolutional layer i The feature vector of (2);
step 6, constructing a node prediction layer for product recommendation:
step 6.1, calculating the prediction score value of the user u on the ith product according to the formula (6)Thereby obtaining a prediction scoring matrix of the product by the user
In the formula (6), <, > represents a vector inner product operation;
step 6.2, establishing a supervision loss function L according to the formula (7) R :
In formula (7), σ (z) is a nonlinear activation function, and θ ═ P, Q]Is the parameter to be optimized,. psi. I ={(u,i,j)|u∈U I ,i∈R u ,j∈(V-R u ) Is the training data of all users on the information domain; r u Representing a set of products representing the u-th user interaction;representing the predicted value of credit of the jth product to the jth user;
and 7, constructing an automatic supervision learning layer to carry out mutual information maximization learning on the local characteristics and the global characteristics of the nodes:
step 7.1, constructing the d user in the social domain according to the formula (8) and the formula (9)Local feature h of d And the a-th user in the information domainLocal feature h of ai :
h d =σ(p d ) (8)
h ai =[σ(p a ),σ(q i )] (9)
In the formula (8), sigma is sigmoid nonlinear activation function;
in the formula (9), [, ] is the splicing operation of the vectors;
step 7.2, constructing global features g of all users in the social domain according to the formula (10) and the formula (11) S And global characteristics g of all users in the information domain I :
Step 7.3, constructing a discriminator D on the social domain according to the formula (12) and the formula (13) S And a discriminator D on the information field I :
In the formula (12), h S Local features expressed as users on the social domain; w S ∈R d×d Is a trainable weight matrix, (h) S ,g S ) Is a positive sample pair on the social domain;
in the formula (13), h I Local features expressed as users on the information domain; w I ∈R 2d×2d Is a trainable weight matrix, (h) I ,g I ) Is a positive sample pair over the information field;
and 7.4, randomly adding and deleting the same number of edges of the heterogeneous graph network G to form a disturbing graphAnd obtaining local features on the fictitious social domain according to the processes of the step 4, the step 5 and the step 7.1And local features on fictitious information domainsThus forming negative sample pairs on the social domainAnd negative example pairs on the information field
Step 7.5, constructing a social domain self-supervision loss function L according to the formula (14) and the formula (15) S And information Domain autonomous loss function L I :
And 7.5, constructing a total loss function L according to the formula (16):
L=L R +αL s +λL I (16)
in the formula (16), α and λ are hyper-parameters for balancing two kinds of the auto-supervised loss functions;
7.6, solving all the loss functions L by using a gradient descent method to ensure that L converges to a minimum value, thereby obtaining an optimal parameter theta * =[P * ,Q * ]Further obtain the optimal user pre-predictionEvaluation matrixAnd predicting a scoring matrix according to the optimal userAnd recommending the product to the non-bridge user in the social domain.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides a cross-domain social recommendation method based on self-supervision, which aims at social cold start users, utilizes bridge users to construct a heterogeneous graph network of an information domain and a social domain, learns potential cooperative information between the social cold start users and products through graph convolution operation, and utilizes a self-supervision learning method to construct supervision signals of nodes in a heterogeneous graph, fully excavates global characteristics of the heterogeneous graph, effectively improves the representation precision of the users and the products, and improves the recommendation precision on the social cold start users.
2. The invention processes the user set and the product set in a single hot coding mode, can effectively expand data dimensionality and is convenient for fast matrix operation.
3. The method and the system utilize the bridge user to construct the information domain and the heteromorphic network on the social domain to construct the connection between the social cold start user and the product, thereby relieving the problem of social cold start.
4. The invention adopts an automatic supervision optimization method, captures more overall information of the heterogeneous graph in a mode of maximizing the mutual information of local characteristics and overall characteristics of user nodes in the heterogeneous graph, and improves the node representation precision.
Drawings
FIG. 1 is a flowchart of a cross-domain social recommendation method based on self-supervised learning according to the present invention.
Detailed Description
In the embodiment, the cross-domain social recommendation method based on the self-supervision learning considers the problem that the number of bridge users is excessively depended on in cross-domain recommendation, establishes the connection between the social cold-starting users and products by establishing a heterogeneous graph network, realizes the mutual information maximization of the local features and the global features of heterogeneous graph nodes by an self-supervision optimization method, and captures more global features of heterogeneous graphs so as to realize more accurate product recommendation of the social cold-starting users. Specifically, as shown in fig. 1, the method comprises the following steps:
step 1, constructing heterogeneous data, comprising: a scoring matrix R of the product by the user and a social relationship matrix S between the users are as follows:
let U I Represents a set of users over an information domain, andwherein the content of the first and second substances,indicating the a-th user in the information field,representing the b-th user in the information domain, M represents the total number of users in the information domain, a is more than or equal to 1, and b is more than or equal to M;
let V denote the product set on the information field, and V ═ V 1 ,…v i ,…v j ,…v L In which v is i Indicating the ith product in the information field, v j Indicating the jth product in the information field, and L indicating the total number of products in the information field;
let R ai Representing the a-th userFor the ith product v i The scoring value of (2) is that the scoring matrix of the product by the user on the information domain is R ═ { R ═ R ai } M×L ;
Let U S Represents a set of users on a social domain,wherein, the first and the second end of the pipe are connected with each other,representing the c-th user in the social domain,representing the d-th user in the social domain, N representing the total number of users in the social domain, c being more than or equal to 1, and d being more than or equal to N;
order S cd Represents the c-th userFor the d-th userIf there is a connection between two users, let S cd If not, let S cd 0, the social relationship matrix between users in the social domain is S ═ S cd } N×N ;
Step 2, utilizing a bridge user set U coexisting in an information domain and a social domain B =U I ∩U S Constructing a heterogeneous graph network G ═ G I ∪G S Wherein G is I Representing the set U of all users in the information domain I Graph network structure, G, with product set V S Representing a set U of all users in a social domain S A graph network structure for collecting U-shaped bridge users B The total number of users in (1) is T;
and 3, obtaining a node cooperation matrix through single hot coding, wherein the method comprises the following steps: user cooperation matrix, product cooperation matrix:
step 3.1, constructing a user cooperation matrix P ═ P in a single hot coding mode for all users in the heterogeneous graph network G 1 ,…,p a ,…,p M-T ,…,p b ,…,p c ,…,p M ,…,p d ,…,p M+N-T And P ∈ R (M+N-T)×d Wherein p is a Representing a d-dimensional cooperation vector of an a-th user in the user cooperation matrix P, wherein the user belongs to a non-bridge user on the information domain; p is a radical of b Representing the d-dimensional cooperation vector of the b-th user in the cooperation matrix P of the users belonging to the bridge user, P d Representing the d-dimensional cooperation vector of the d-th user in the user cooperation matrix P, wherein the user belongs to the social domainNon-bridge users;
step 3.2, constructing a product cooperation matrix Q ═ Q in a single-hot coding mode for all products in the heteromorphic graph network G 1 ,…,q i ,…q L And Q ∈ R L×d Wherein q is i A d-dimensional collaborative vector representing an ith product;
and 4, carrying out characteristic propagation through the graph volume layer:
step 4.1, making the graph convolution layers have K convolution layers, and initializing the number K of the current convolution layers to be 0;
step 4.2, obtaining the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain according to the formula (1), the formula (2) and the formula (3) respectively a D-dimensional cooperation vector p of the b-th bridge user b D-dimensional cooperation vector p of the d-th non-bridge user in social domain d After passing through the (k + 1) th graph convolution layer, the corresponding user characteristic vector is obtainedObtaining a d-dimensional cooperative vector q of the ith product according to the formula (4) i The product characteristic vector obtained after passing through the (k + 1) th graph convolution layer
In the formula (1), R a Representing the collection of products that the a-th non-bridge user has interacted with,is the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain a When k is 0, let the user characteristic vector output from the k-th convolution layerA user feature vector is initialized randomly;
in the formula (2), R b Representing the set of products, S, interacted by the b-th bridge user b Representing a set of users with social connections to the b-th bridge user,is d-dimensional cooperation vector p of the b-th bridge user b When the user characteristic vector output by the k-th convolutional layer is equal to 0,a user feature vector is initialized randomly;
in formula (3), S d Representing a set of users for which the d-th non-bridge user has social connections,the user feature vector is output by the d-th non-bridge user in the social domain at the kth layer convolutional layer, when k is 0,a user feature vector is initialized randomly;
in the formula (4), R i Representing a set of users who have interacted with the ith product,is the d-dimensional cooperation vector q of the ith product i When the characteristic vector of the product output by the k-th layer convolution layer is 0,the feature vector of the product is initialized randomly;
4.3, assigning K +1 to K, judging whether K is greater than K, and if so, executing the step 5; otherwise, returning to the step 4.2 for sequential execution;
step 5, constructing a node feature aggregation layer:
performing feature aggregation on the K-layer user feature vector and the product feature vector according to the formula (5) to obtain a feature vector p of any user u in the heterogeneous network G u And the feature vector q of the ith product i :
In the formula (5), AVG represents the average pooling operation,a feature vector representing the user u output by the kth convolutional layer,i-th product q representing the output of the k-th convolutional layer i The feature vector of (2);
step 6, constructing a node prediction layer for product recommendation:
step 6.1, calculating the prediction scoring value of the user u on the ith product according to the formula (6)Thereby obtaining a prediction scoring matrix of the product by the user
In the formula (6), <, > represents a vector inner product operation;
step 6.2, establishing a supervision loss function L according to the formula (7) R :
In formula (7), σ (z) is a nonlinear activation function, and θ ═ P, Q]Is the parameter to be optimized,. psi. I ={(u,i,j)|u∈U I ,i∈R u ,j∈(V-R u ) The training data of all users in the information domain; r u Representing a set of products representing the u-th user interaction;representing the predicted value of credit of the jth product to the jth user;
and 7, constructing an automatic supervision learning layer to carry out mutual information maximization learning on the local characteristics and the global characteristics of the nodes:
step 7.1, constructing the d user in the social domain according to the formula (8) and the formula (9)Local feature h of d And the a-th user in the information domainLocal feature h of ai :
h d =σ(p d ) (8)
h ai =[σ(p a ),σ(q i )] (9)
In the formula (8), sigma is sigmoid nonlinear activation function;
in the formula (9), [, ] is the splicing operation of the vectors;
step 7.2, constructing global features g of all users in the social domain according to the formula (10) and the formula (11) S And global characteristics g of all users in the information domain I :
Step 7.3, constructing a discriminator D on the social domain according to the formula (12) and the formula (13) S And a discriminator D on the information field I :
In the formula (12), h S Local features expressed as users on the social domain; w S ∈R d×d Is a trainable weight matrix, (h) S ,g S ) Is a positive sample pair on the social domain;
in the formula (13), h I Local features expressed as users on the information domain; w I ∈R 2d×2d Is a trainable weight matrix, (h) I ,g I ) Is a positive sample pair over the information field;
and 7.4, randomly adding and deleting the same number of edges of the heterogeneous graph network G to form a disturbing graphAnd obtaining local features on the fictitious social domain according to the processes of the step 4, the step 5 and the step 7.1And local features on the imaginary information domainThus forming negative sample pairs on the social domainAnd negative example pairs on the information field
Step 7.5, constructing a social domain self-supervision loss function L according to the formula (14) and the formula (15) S And information Domain autonomous loss function L I :
And 7.5, constructing a total loss function L according to the formula (16):
L=L R +αL S +λL I (16)
in the formula (16), α and λ are hyper-parameters for balancing two kinds of the auto-supervised loss functions;
7.6, solving all loss functions L by using a gradient descent method to ensure that L converges to a minimum value, thereby obtaining an optimal parameter theta * =[P * ,Q * ]Further obtain the optimal user prediction scoring matrixAnd predicting a scoring matrix according to the optimal userAnd recommending the product to the non-bridge user in the social domain. The method excavates the inter-domain collaborative information in a very limited bridge user, optimizes the model in a self-supervision learning mode, and fully excavates the global characteristics of the heterogeneous map, thereby realizing more accurate product recommendation of the social cold start user.
Example (b):
in order to verify the effectiveness of the method, the invention performs data processing on two public data sets: epionins, Dianping, to meet the experimental conditions of the present invention.
For the product recommendation task, the invention adopts Hit Ratio (HR) and Normalized counted graphical Gain (NDCG) as evaluation criteria. The invention selects 3 methods for effect comparison, namely EMCDR, NSCR and BiTGCF.
TABLE 1 social Cold Start user product recommendation results on Epines dataset by the method and comparison method of the present invention
TABLE 2 social Cold Start user product recommendation results on Dianping datasets by the method of the present invention and the comparison method
Specifically, table 1 and table 2 show the experimental results on the epipons and Dianping data sets, respectively, and it can be seen that on the two data sets, the method proposed by the present invention is superior to the 3 methods of comparison in both HR and NDCG indexes on social cold-start users.
Claims (1)
1. A cross-domain social product recommendation method based on self-supervision learning is characterized by comprising the following steps:
step 1, constructing heterogeneous data, comprising: a scoring matrix R of the product by the user and a social relationship matrix S between the users are as follows:
let U I Represents a set of users over an information domain, andwherein the content of the first and second substances,indicating the a-th user in the information field,representing the b-th user in the information domain, M represents the total number of users in the information domain, a is more than or equal to 1, and b is more than or equal to M;
let V denote the product set on the information field, and V ═ V 1 ,…v i ,…v j ,…v L In which v is i Indicating the ith product in the information field, v j Indicating the jth product in the information field, and L indicating the total number of products in the information field;
let R ai Representing the a-th userFor the ith product v i The scoring value of (2) is that the scoring matrix of the product by the user on the information domain is R ═ { R ═ R ai } M×L ;
Let U S Represents a set of users on a social domain,wherein the content of the first and second substances,representing the c-th user in the social domain,representing the d-th user in the social domain, N representing the total number of users in the social domain, c being more than or equal to 1, and d being more than or equal to N;
order S cd Represents the c-th userFor the d-th userIf there is a connection between two users, let S cd If not, let S cd 0, the social relationship matrix between users in the social domain is S ═ S cd } N×N ;
Step 2, utilizing a bridge user set U coexisting in an information domain and a social domain B =U I ∩U S Constructing a heterogeneous graph network G ═ G I ∪G S Wherein G is I Representing the set U of all users in the information domain I Graph network structure, G, with product set V S Representing a set U of all users in a social domain S A graph network structure for collecting U-shaped bridge users B The total number of users in (1) is T;
and 3, obtaining a node cooperation matrix through single hot coding, wherein the method comprises the following steps: user cooperation matrix, product cooperation matrix:
step 3.1, constructing a user cooperation matrix P ═ P in a single hot coding mode for all users in the heterogeneous graph network G 1 ,…,p a ,…,p M-T ,…,p b ,…,p c ,…,p M ,…,p d ,…,p M+N-T And P ∈ R (M+N-T)×d Wherein p is a Representing a d-dimensional cooperation vector of an a-th user in the user cooperation matrix P, wherein the a-th user belongs to a non-bridge user in the information domain; p is a radical of b Representing the d-dimensional cooperation vector of the b-th user in the user cooperation matrix P, wherein the b-th user belongs to the bridge user, P d Representing a d-dimensional cooperation vector of a d-th user in the user cooperation matrix P, wherein the d-th user belongs to a non-bridge user in the social domain;
step 3.2, constructing a product cooperation matrix Q ═ Q in a single-hot coding mode for all products in the heteromorphic graph network G 1 ,…,q i ,…q L And Q ∈ R L×d Wherein q is i A d-dimensional collaborative vector representing an ith product;
and 4, carrying out characteristic propagation through the graph volume layer:
step 4.1, making the graph convolution layers have K convolution layers, and initializing the number K of the current convolution layers to be 0;
step 4.2, obtaining the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain according to the formula (1), the formula (2) and the formula (3) respectively a D-dimensional cooperation vector p of the b-th bridge user b D-dimensional collaborative direction of the d-th non-bridge user in social domainQuantity p d After passing through the (k + 1) th graph convolution layer, the corresponding user characteristic vector is obtainedObtaining a d-dimensional cooperative vector q of the ith product according to the formula (4) i The product characteristic vector obtained after passing through the (k + 1) th graph convolution layer
In the formula (1), R a Representing the set of products that the a-th non-bridge user interacted with,is the d-dimensional cooperation vector p of the a-th non-bridge user in the information domain a When k is 0, let the user characteristic vector output from the k-th convolution layerA user feature vector is initialized randomly;
in the formula (2), R b Represents the product set interacted by the b-th bridge user, S b Representing the b-th bridge userA collection of users having a social connection,is d-dimensional cooperation vector p of the b-th bridge user b When the user characteristic vector output by the k-th convolutional layer is equal to 0,a user feature vector is initialized randomly;
in formula (3), S d Representing a set of users for which the d-th non-bridge user has social connections,the user feature vector is output by the d-th non-bridge user in the social domain at the kth layer convolutional layer, when k is 0,a user feature vector is initialized randomly;
in the formula (4), R i Representing a set of users who have interacted with the ith product,is the d-dimensional cooperation vector q of the ith product i When the characteristic vector of the product output by the k-th layer convolution layer is 0,a randomly initialized product feature vector;
4.3, assigning K +1 to K, judging whether K is more than K, and if so, executing step 5; otherwise, returning to the step 4.2 for sequential execution;
step 5, constructing a node feature aggregation layer:
performing feature aggregation on the K-layer user feature vector and the product feature vector according to the formula (5) to obtain a feature vector p of any user u in the heterogeneous network G u And the feature vector q of the ith product i :
In the formula (5), AVG represents the average pooling operation,a feature vector representing the user u output by the kth convolutional layer,i-th product q representing the output of the k-th convolutional layer i The feature vector of (2);
step 6, constructing a node prediction layer for product recommendation:
step 6.1, calculating the prediction scoring value of the user u on the ith product according to the formula (6)Thereby obtaining a prediction scoring matrix of the product by the user
In the formula (6), <, > represents a vector inner product operation;
step 6.2, establishing a supervision loss function L according to the formula (7) R :
In formula (7), σ (z) is a nonlinear activation function, and θ ═ P, Q]Is the parameter to be optimized,. psi. I ={(u,i,j)|u∈U I ,i∈R u ,j∈(V-R u ) Is aTraining data of all users on the information domain; r u Representing a set of products representing the u-th user interaction;representing the predicted value of credit of the jth product to the jth user;
and 7, constructing an automatic supervision learning layer to carry out mutual information maximization learning on the local characteristics and the global characteristics of the nodes:
step 7.1, constructing the d user in the social domain according to the formula (8) and the formula (9)Local feature h of d And the a-th user in the information domainLocal feature h of ai :
h d =σ(p d ) (8)
h ai =[σ(p a ),σ(q i )] (9)
In the formula (8), sigma is sigmoid nonlinear activation function;
in the formula (9), [, ] is the splicing operation of the vectors;
step 7.2, constructing global features g of all users in the social domain according to the formula (10) and the formula (11) S And global characteristics g of all users in the information domain I :
Step 7.3, constructing a discriminator D on the social domain according to the formula (12) and the formula (13) S And a discriminator D on the information field I :
In the formula (12), h S Local features expressed as users on the social domain; w S ∈R d×d Is a trainable weight matrix, (h) S ,g S ) Is a positive sample pair on the social domain;
in the formula (13), h I Local features expressed as users on the information domain; w I ∈R 2d×2d Is a trainable weight matrix, (h) I ,g I ) Is a positive sample pair over the information field;
and 7.4, randomly adding and deleting the same number of edges of the heterogeneous graph network G to form a disturbing graphAnd obtaining local features on the fictitious social domain according to the processes of the step 4, the step 5 and the step 7.1And local features on the imaginary information domainThus forming negative sample pairs on the social domainAnd negative example pairs on the information field
Step 7.5, constructing a social domain self-supervision loss function L according to the formula (14) and the formula (15) S And information domain self-monitoringGovernor loss function L I :
And 7.5, constructing a total loss function L according to the formula (16):
L=L R +αL S +λL I (16)
in the formula (16), α and λ are hyper-parameters for balancing two kinds of the auto-supervised loss functions;
7.6, solving all the loss functions L by using a gradient descent method to ensure that L converges to a minimum value, thereby obtaining an optimal parameter theta * =[P * ,Q * ]Further obtain the optimal user prediction scoring matrixAnd predicting a scoring matrix according to the optimal userAnd recommending the product to the non-bridge user in the social domain.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210511621.0A CN114880583B (en) | 2022-05-11 | 2022-05-11 | Cross-domain social recommendation method based on self-supervision learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210511621.0A CN114880583B (en) | 2022-05-11 | 2022-05-11 | Cross-domain social recommendation method based on self-supervision learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114880583A true CN114880583A (en) | 2022-08-09 |
CN114880583B CN114880583B (en) | 2024-03-05 |
Family
ID=82674781
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210511621.0A Active CN114880583B (en) | 2022-05-11 | 2022-05-11 | Cross-domain social recommendation method based on self-supervision learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114880583B (en) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109410080A (en) * | 2018-10-16 | 2019-03-01 | 合肥工业大学 | A kind of social image recommended method based on level attention mechanism |
CN111291270A (en) * | 2020-03-02 | 2020-06-16 | 合肥工业大学 | Attribute reasoning and product recommendation method based on self-adaptive graph convolution network |
CA3094507A1 (en) * | 2019-10-25 | 2021-04-25 | The Governing Council Of The University Of Toronto | Systems, devices and methods for transfer learning with a mixture of experts model |
US20210272217A1 (en) * | 2018-08-02 | 2021-09-02 | Arizona Board Of Regents On Behalf Of Arizona State University | Systems and methods for a cross media joint friend and item recommendation framework |
CN113779380A (en) * | 2021-08-12 | 2021-12-10 | 阿里巴巴(中国)有限公司 | Cross-domain recommendation method, device and equipment, and content recommendation method, device and equipment |
CN114281976A (en) * | 2021-08-27 | 2022-04-05 | 腾讯科技(深圳)有限公司 | Model training method and device, electronic equipment and storage medium |
-
2022
- 2022-05-11 CN CN202210511621.0A patent/CN114880583B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210272217A1 (en) * | 2018-08-02 | 2021-09-02 | Arizona Board Of Regents On Behalf Of Arizona State University | Systems and methods for a cross media joint friend and item recommendation framework |
CN109410080A (en) * | 2018-10-16 | 2019-03-01 | 合肥工业大学 | A kind of social image recommended method based on level attention mechanism |
CA3094507A1 (en) * | 2019-10-25 | 2021-04-25 | The Governing Council Of The University Of Toronto | Systems, devices and methods for transfer learning with a mixture of experts model |
CN111291270A (en) * | 2020-03-02 | 2020-06-16 | 合肥工业大学 | Attribute reasoning and product recommendation method based on self-adaptive graph convolution network |
CN113779380A (en) * | 2021-08-12 | 2021-12-10 | 阿里巴巴(中国)有限公司 | Cross-domain recommendation method, device and equipment, and content recommendation method, device and equipment |
CN114281976A (en) * | 2021-08-27 | 2022-04-05 | 腾讯科技(深圳)有限公司 | Model training method and device, electronic equipment and storage medium |
Non-Patent Citations (2)
Title |
---|
刘华锋;景丽萍;于剑;: "融合社交信息的矩阵分解推荐方法研究综述", 软件学报, no. 02, 9 October 2017 (2017-10-09), pages 120 - 142 * |
王晓阳;郑骁庆;肖仰华;: "智慧搜索中的实体与关联关系建模与挖掘", 通信学报, no. 12, 25 December 2015 (2015-12-25), pages 21 - 31 * |
Also Published As
Publication number | Publication date |
---|---|
CN114880583B (en) | 2024-03-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111309983B (en) | Method and device for processing service based on heterogeneous graph | |
CN112381179B (en) | Heterogeneous graph classification method based on double-layer attention mechanism | |
Ebiaredoh-Mienye et al. | Artificial neural network technique for improving prediction of credit card default: A stacked sparse autoencoder approach | |
CN111723285A (en) | Depth spectrum convolution collaborative filtering recommendation method based on scores | |
CN115270007B (en) | POI recommendation method and system based on mixed graph neural network | |
CN105306296A (en) | Data filter processing method based on LTE (Long Term Evolution) signaling | |
WO2020024444A1 (en) | Group performance grade recognition method and apparatus, and storage medium and computer device | |
CN116383519A (en) | Group recommendation method based on double weighted self-attention | |
Rafie et al. | A multi-objective online streaming multi-label feature selection using mutual information | |
CN116244513A (en) | Random group POI recommendation method, system, equipment and storage medium | |
CN106817390A (en) | A kind of shared method and apparatus of user data | |
CN112364245B (en) | Top-K movie recommendation method based on heterogeneous information network embedding | |
CN114708479A (en) | Self-adaptive defense method based on graph structure and characteristics | |
CN116757262B (en) | Training method, classifying method, device, equipment and medium of graph neural network | |
CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
CN114880583A (en) | Cross-domain social recommendation method based on self-supervision learning | |
Acharya et al. | Using Optimal Embeddings to Learn New Intents with Few Examples: An Application in the Insurance Domain. | |
CN110544124A (en) | waste mobile phone pricing method based on fuzzy neural network | |
CN113887806B (en) | Long-tail cascade popularity prediction model, training method and prediction method | |
CN114254738A (en) | Double-layer evolvable dynamic graph convolution neural network model construction method and application | |
CN113962748A (en) | Method for aligning users of heterogeneous e-commerce platform by using holomorphic information representation based on meta-path | |
CN115481215A (en) | Partner prediction method and prediction system based on temporal partner knowledge graph | |
Gao et al. | Wide & ResNet: An improved network for CTR prediction | |
Cui | Research on the filtering recommendation technology of network information based on big data environment | |
CN116192650B (en) | Link prediction method based on sub-graph features |
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 |