CN114443954B - One-to-many cross-domain recommendation method and system based on high-order graph structure - Google Patents
One-to-many cross-domain recommendation method and system based on high-order graph structure Download PDFInfo
- Publication number
- CN114443954B CN114443954B CN202210018292.6A CN202210018292A CN114443954B CN 114443954 B CN114443954 B CN 114443954B CN 202210018292 A CN202210018292 A CN 202210018292A CN 114443954 B CN114443954 B CN 114443954B
- Authority
- CN
- China
- Prior art keywords
- user
- module
- embedded
- migration
- domain
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 230000005012 migration Effects 0.000 claims abstract description 133
- 238000013508 migration Methods 0.000 claims abstract description 133
- 238000012512 characterization method Methods 0.000 claims abstract description 103
- 238000012549 training Methods 0.000 claims abstract description 43
- 230000008569 process Effects 0.000 claims abstract description 15
- 230000003993 interaction Effects 0.000 claims description 20
- 230000006870 function Effects 0.000 claims description 18
- 230000004913 activation Effects 0.000 claims description 13
- 238000012546 transfer Methods 0.000 claims description 11
- 230000003044 adaptive effect Effects 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 239000010410 layer Substances 0.000 description 40
- 238000010276 construction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 238000007334 copolymerization reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000002344 surface layer Substances 0.000 description 1
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
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- 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
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a pair of multi-cross-domain recommendation method and a system based on a high-order graph structure, wherein in the process of embedding and characterizing users and projects, the high-order graph structure is utilized to establish a plurality of embedded migration modules, multi-level structured information of users and projects of dense source domains is used for guiding users and projects of sparse target domains to characterize, the method comprises the steps of firstly obtaining user and project characterizations of the source domains with context invariants, then training an embedded part of an embedded migration module by using source domain data, and finally fitting a residual layer of the target domain by using the trained migration part of the embedded migration module, so as to achieve the migration target.
Description
Technical Field
The invention belongs to the technical field of cross-domain recommendation, and particularly relates to a one-to-many cross-domain recommendation method and system based on a high-order graph structure.
Background
Cross-domain recommendation is an important method for solving the data sparseness problem. Most of cross-domain recommendation researches learn user characterization containing user features and article characterization containing article features by using shared users and articles in dense and sparse domains through a cross-domain copolymerization method, and then migrate high-efficiency characterization learned in the dense domain into the sparse domain so as to solve the problem of sparse data and improve recommendation quality. However, the scenes with common users or items in the real scenes are rare, so Hao et al propose a one-to-many cross-domain recommendation method based on the context invariants, the common users or items are not required to be owned among domains, the users and the items in the target domain are guided to be embedded by the context invariants extracted from the source domain, high-quality user and item characterization is obtained, and the recommendation quality is improved.
However, this method has two disadvantages: first, structured information in user-project interactions is not considered in the user and project embedding process. Such as user 1-item 1-user 2-item 2; second, the user and project information is underutilized and the depth of mining is insufficient, failing to construct high quality representations. For example, the method only uses embedding after simple one-hot coding to obtain the embedded characterization of the user and the item, and the simple embedding mode only focuses on the surface layer information of the user and the item, so that the information utilization is insufficient, and the generalization of the characterization is reduced.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a one-to-many cross-domain recommendation method and a system based on a high-order graph structure, wherein structured information is added in the process of embedding users and projects, a plurality of embedding migration modules are established, and multi-level structured information in a source domain is migrated to guide the users and projects in a target domain to be embedded, so that the user and project information is fully utilized, and high-quality characterization is constructed.
In order to solve the technical problems, the invention adopts the following technical scheme:
Firstly, the invention provides a pair of multi-cross-domain recommendation method based on a high-order graph structure, which utilizes the high-order graph structure to establish a plurality of embedded migration modules in the process of embedding and characterizing users and projects, and uses multi-level structured information of the users and projects of a dense source domain to guide the users and projects of a sparse target domain to characterize, wherein the method comprises the following steps:
s1, acquiring user and item characterization with context invariants:
extracting context invariants in dense source domain by context module M 1 User embedded module through source domainAnd project embedding moduleObtaining token vectors for users and items with different hierarchically structured informationAndSplicing to obtain user and item characterization with multi-layer structured informationAndClustering module through user and project context conditionsAndExtracting context invariants from context module M 1 And (3) withAndCombining to finally obtain the user and item characterization with the most relevant contextAnd
S2, predicting interaction scoring results:
the end user and item characterization vectors obtained in step S1 in the source domain are processed by a predictive scoring module M 4 AndObtaining predictive score by multiplyingThen, carrying out back propagation on the difference between the predicted result and the real result to carry out training;
S3, migration:
The multi-level structured information of the source domain is used for guiding the user and the project embedding of the target domain through the embedding and migration module, the target domain user and the project embedding characterization with different levels of structured information obtained through the embedding and migration module are spliced, and then the context condition clustering module of the target domain user and the project is used for guiding the user and the project embedding characterization AndObtaining a final target domain user and a final project embedding representation;
s4, outputting a cross-domain recommendation result.
Obtaining the predictive score of the target domain by multiplying the end user obtained in the step S3 by the item characterizationAnd outputting a recommendation result according to the scores.
Further, after the training of the source domain in the step S2 is completed, a user embedded module is obtainedAnd project embedding moduleIn step S3, it is used for training of the embedded migration module; by fixing the parameters of the embedded part in each embedded migration module as the source domain user embedded moduleAnd project embedding moduleTraining the migration part embedded in the migration module according to the trained parameters; the trained embedded migration modules are used for fitting residual layers in the target domain, in the process of fitting the target domain, parameters of migration parts of each embedded migration module are fixed to be trained parameters by using source domain data, and the embedded parts are embedded by using target domain usersAnd project embedding moduleGenerated data and parameters, thereby training user embedded module of target domainAnd project embedding moduleAchieves the purpose of using the multi-level structured information of the source domain to guide the user of the target domain and the embedding of the project
Further, the embedded migration module includes two parts:
First, embedded part:
embedded characterization of users and items with source domains AndAs input of the embedded migration module, the embedded migration module is patterned in an embedded part to obtain a user and a project representation with multi-level structured information, wherein the user representation with the u 1 self-node information and the user representation with the project v n and the user u 1 self-information are defined as the user representations with different hierarchical structured information, the project representation with the v 1 self-node information and the project representation with the u i and the v 1 self-information are defined as the project representations with different hierarchical structured information,
Wherein v n is the first-order neighbor node of user u 1, and u i is the first-order neighbor node of item v 1;
second, migration section:
Only the user and project characterization with different hierarchical structural information obtained by the source domain embedding part are used as input for training a universal embedded migration module; the migration part is used for fitting the residual layer of the target domain to achieve the goal of one-to-many migration.
Further, the resulting output after passing through a layer of embedded migration module embedded portionAs an input of the embedded part of the second layer embedded migration module, and the like, in the high-order propagation process, the output of the first layer-1 is used as the input of the first layer:
Wherein, For the first order user characterization,Is the Laplace norm, where N u is the number of items that have interactions with user u 1 and N v is the number of users that have interactions with item v n; In order for the weight matrix to be trainable, For the l-1 order item characterization obtained by the embedding part of the embedding module,For the l-1 order user characterization obtained by the embedding part of the embedding module,Structured information between user u 1 and item v n; leakyReLU is a function of the activation ReLU,Characterization of the l-1 order itemAnd l-1 order user characterizationThe information of the interaction between the two,Characterization for l-1 order userInformation of the same.
Furthermore, by designing an countermeasure network composed of a generating network GN and a discriminating network DN, learning data distribution of user characterization with structured information, and migrating the structured information to a target domain for guiding user embedding of the target domain, specifically:
By using As input, mapping the input to a normal distribution of low dimensions through a discriminant network DNIn (1) obtaining hidden variablesSatisfy normal distribution and pass hidden variableReconstructing to obtainAt the same time, generating a noise-carrying signal with the generation network GNAnd generating hidden variables through networkIt also satisfies normal distribution and is obtained by reconstruction according to hidden variablesFurther, the generation network GN generates an input adaptive noise
LeakyReLU is the activation function ReLU, W, b is a trainable network parameter,Embedding a token for a first-order user with adaptive noise generated by the network GN; reusing a first-order user embedding characterization generated by the source domain user embedding module parameters for the embedding migration module embedding part;
The distinguishing network DN adopts a variation self-encoder for distinguishing whether the input comes from the original data distribution space or the noisy data distribution space:
where L DN is the discrimination of network loss, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy first-order user characterizations generated for GNs.
Furthermore, the discrimination network and the generation network adopt an alternate training mode during training, and the generation network is used for confusing the discrimination network, so that the discrimination network recognizes the noisy input as a true value:
Where L GN is the loss of the generation network, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userThe loss L DN of the discrimination network DN can distinguish the original data from the noisy data in its coding space, while the first term of the loss L GN of the generation network GN tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
Further, training of the embedded migration module with the l-layer structured information is achieved by:
first, a user embedded representation with l-layer structured information is obtained by the formula (4)
This is then used as input to train the migration section:
Wherein LeakyReLU is the activation function ReLU, W (l)、b(l) is a trainable network parameter, An l-order user-embedded token with adaptive noise generated by the generation network GN,Is a source domain moduleThe generated l-order user characterization; l DN is a decision to determine the network loss,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy l-order user characterizations generated for GNs; l GN is a loss to the generation network, wherein,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userThe first term to generate the loss L GN of the network GN then tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
The embedded migration module of the project is the same as the embedded migration module training of the user.
Further, in step S2, the loss function is as follows:
Wherein, For source domain network training loss, u is user ID, c is context, v is item ID, s uv is user u's true score for item v,The prediction of item v for user u is scored.
Further, in step S3, the residual layer is fitted by using the embedded migration module, so as to achieve the migration effect of guiding the input of the target domain by using the model parameters of the source domain, and when the migration is performed, the parameters of the embedded migration module are fixed, the residual layer of the target domain is trained, and the loss is as follows:
LTransfer＝Lu+Lv (16)
L Transfer is migration loss, wherein L u is loss of user multi-level structured information migration, and L v is loss of item multi-level structured information migration; For user characterization with different hierarchically structured information, Representing the project with different hierarchical structured information; where u is the user set of the source domain and v is the item set;
wherein the residual layer is defined as follows:
Wherein, AndIs a module in the target domainAndThe obtained first-order user and item characterization; Is a trainable parameter; tanh is the activation function; for the l-order user and item characterizations generated after passing through the residual layer that are used as inputs to the embedded migration module.
The invention further provides a one-to-many cross-domain recommendation system based on the high-order graph structure, which comprises the following steps:
A user and item characterization module for obtaining user and item characterizations with context invariants, specifically comprising a context module M 1 for obtaining context invariants c nc from a source domain, an embedded characterization for obtaining user and item with different hierarchical structured information from the source domain AndSplicingAndResulting user and item representations with multi-layer structured informationAndIs embedded in the module of (a)
A context condition clustering module for combining the context invariant c nc extracted by the context module M 1 with the moduleAndResulting user and item representations with multi-layer structured informationAndCombining and obtaining user and item representations with most relevant contexts by multi-layer linear transformationAnd
A predictive scoring module M 4 for characterizing the resulting users and items with the most relevant contextAndVector multiplication is carried out to obtain a prediction score;
an embedded migration module comprising an embedded part and a migration part, the embedded module using a source domain AndEmbedding module for parameter training of migration portions thereof, followed by residual layer training of target domain fitting of migration portions thereof to target domainAndThe parameters of the target domain are achieved, and the target domain user and the project embedding target is guided by the source domain multi-level structured information.
Compared with the prior art, the invention has the advantages that:
(1) The invention fully utilizes the potential information in the user-project interaction to develop the learning of the user/project embedded characterization, and obtains the user and project embedded characterization with higher quality: in the process of representing the user and the project, the multi-level structured information is added by utilizing the high-level graph structure, the potential information of the user-project interaction is added in the process of embedding the user/project in the specific domain, the association relationship between the hidden user and the project in the potential information is fully utilized, and the quality of the embedded representation of the user/project in the specific domain is improved. In cross-domain recommendation, the multi-level structured information learned by the source domain is used for guiding the user and project embedding of the target domain through the embedding migration module, so that the recommendation quality is improved.
(2) The information embedding characterization method is reasonable to select. The invention establishes a plurality of embedded propagation layers for representing information embedding by using the graph structure, fully utilizes the function of capturing neighbor node information of the graph structure, enables a user to obtain information transmitted by nodes, adds interaction potential information in the process of embedding the representation of the user/item, and improves generalization of the representation of the user/item.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a source domain single domain training network in accordance with an embodiment of the present invention;
fig. 3 is a cross-domain flow chart of an embodiment of the present invention.
Detailed Description
The invention will be further described with reference to the accompanying drawings and specific examples.
In the one-to-multi-cross-domain recommendation method based on the high-order graph structure, a plurality of embedded migration modules are established by utilizing the high-order graph structure in the process of embedding and characterizing the users and the items, and multi-level structured information of the users and the items of the dense source domain is used for guiding the users and the items of the sparse target domain to characterize.
The present invention will be described in detail with reference to fig. 1 and 2. First the method comprises the steps of:
s1, acquiring user and item characterization with context invariants:
extracting context invariants in dense source domain by context module M 1 User embedded module through source domainAnd project embedding moduleObtaining token vectors for users and items with different hierarchically structured informationAndSplicing to obtain user and item characterization with multi-layer structured informationAndClustering module through user and project context conditionsAndExtracting context invariants from context module M 1 And (3) withAndCombining to obtain user and item characterization with context features and multi-level structured informationAndFinally, obtaining the user and item representation with the most relevant context through linear transformationAnd
S2, predicting interaction scoring results:
the end user and item characterization vectors obtained in step S1 in the source domain are processed by a predictive scoring module M 4 AndObtaining predictive score by multiplyingAnd then, carrying out back propagation on the difference between the predicted result and the real result to carry out training.
In the above way, the source domain single domain training can be realized.
S3, migration:
The multi-level structured information of the source domain is used for guiding the user and the project embedding of the target domain through the embedding and migration module, the target domain user and the project embedding characterization with different levels of structured information obtained through the embedding and migration module are spliced, and then the context condition clustering module of the target domain user and the project is used for guiding the user and the project embedding characterization AndAnd obtaining the final target domain user and the item embedded characterization. The method specifically comprises the following steps:
obtaining the user embedded module after the source domain training is completed And project embedding moduleIs used for training the embedded migration module; by fixing the parameters of the embedded part in each embedded migration module as the source domain user embedded moduleAnd project embedding moduleTraining the migration part embedded in the migration module according to the trained parameters; the trained embedded migration modules are used for fitting residual layers in the target domain, and in the process of fitting the target domain, parameters of a migration part of each embedded migration module are fixed to be trained parameters by using source domain data, and the embedded part is embedded by using target domain usersAnd project embedding moduleGenerated data and parameters, thereby training user embedded module of target domainAnd project embedding moduleThe multi-level structured information of the source domain is used for guiding the user and the project embedding of the target domain, the target domain user and the project embedding characterization with different levels of structured information through the embedding migration module are spliced, and then the target domain user and the project context condition clustering module is used for the target domainAndAnd obtaining the final target domain user and the item embedded characterization.
S4, outputting a cross-domain recommendation result.
Obtaining the predictive score of the target domain by multiplying the end user obtained in the step S3 by the item characterizationAnd outputting a recommendation result according to the scores.
A source domain single domain training network diagram as exemplified in connection with fig. 2, wherein l=3; the cross-domain flow chart shown in fig. 3, introduces an embedded migration module. Comprising two parts:
First, embedded part:
embedded characterization of users and items with source domains AndAs input of the embedded migration module, the embedded migration module is patterned in an embedded part to obtain a user and a project representation with multi-level structured information, wherein the user representation with the u 1 self-node information and the user representation with the project v n and the user u 1 self-information are defined as the user representations with different hierarchical structured information, the project representation with the v 1 self-node information and the project representation with the u i and the v 1 self-information are defined as the project representations with different hierarchical structured information,
Wherein v n is the first-order neighbor node of user u 1, and u i is the first-order neighbor node of item v 1;
second, migration section:
Only the user and project characterization with different hierarchical structural information obtained by the source domain embedding part are used as input for training a universal embedded migration module; the migration part is used for fitting the residual layer of the target domain to achieve the goal of one-to-many migration.
That is, when the embedded migration module is trained, the source domain single domain network is trained by the source domain data first; then use the trained embedded moduleAndAs parameters of the embedded part of the embedded migration module, training the migration part of the embedded migration module by using the source domain data; and finally, fitting the residual layer of the target domain by using the trained migration part embedded in the migration module to achieve the migration target.
Specifically, first, a layer of embedded migration module acquires a first-order user embedded representation with a layer of structured information, and the method comprises the following steps:
a. Firstly, the interaction information of each item and a user u 1 is obtained through information construction:
Representing interaction information of the item v n with the user u 1, the item v n being a first order neighbor of the user u 1, n=1, 2,3,4; Is the Laplace norm, where N u is the number of items that have interactions with user u 1 and N v is the number of users that have interactions with item v n; w 1,W2 is a trainable weight matrix, Is embedded in the moduleThe zeroth order embedded representation of the resulting item v n, multiplied by the corresponding element,Structured information between user u 1 and item v n;
b. Then, integrating the information obtained in the last step to obtain a final first-order user embedded representation:
A token is embedded for a first-order user with a layer of structured information, leakyReLU is the activation function ReLU, Information for user u 1 itself: Where W 1 is shared with the weights at the time of information construction, Is embedded in the moduleThe resulting zeroth order embedded representation of user u 1.
The resulting output after having passed through a layer of embedded migration module embedded portionsAs an input of the embedded part of the second layer embedded migration module, and the like, in the high-order propagation process, the output of the first layer-1 is used as the input of the first layer:
Wherein, For the first order user characterization,Is the Laplace norm, where N u is the number of items that have interactions with user u 1 and N v is the number of users that have interactions with item v n; In order for the weight matrix to be trainable, For the l-1 order item characterization obtained by the embedding part of the embedding migration module,For the l-1 order user characterization obtained by embedding the migration module embedding portion,Structured information between user u 1 and item v n; leakyReLU is a function of the activation ReLU,Characterization of the l-1 order itemAnd l-1 order user characterizationThe information of the interaction between the two,Characterization for l-1 order userInformation of the same.
In the migration part, through designing an countermeasure network composed of a generation network GN and a discrimination network DN, learning data distribution of user characterization with structured information, and migrating the structured information to a target domain for guiding user embedding of the target domain, specifically:
By using As input, mapping the input to a normal distribution of low dimensions through a discriminant network DNIn (1) obtaining hidden variablesSatisfy normal distribution and pass hidden variableReconstructing to obtainAt the same time, generating a noise-carrying signal with the generation network GNAnd generating hidden variables through networkIt also satisfies normal distribution and is obtained by reconstruction according to hidden variables
Regarding the loss function, the following is introduced:
the generation network GN generates an input adaptive noise
Wherein LeakyReLU is the activation function ReLU, W, b is the trainable network parameters,Embedding a token for a first-order user with adaptive noise generated by the network GN; Reuse of source domain modules for embedded migration modules First-order user embedded characterization of parameter generation;
The distinguishing network DN adopts a variation self-encoder for distinguishing whether the input comes from the original data distribution space or the noisy data distribution space:
where L DN is the discrimination of network loss, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy first-order user characterizations generated for GNs.
The judging network and the generating network adopt an alternate training mode when training, and the generating network is used for confusing the judging network, so that the judging network recognizes noisy input as a true value:
Where L GN is the loss of the generation network, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userThe loss L DN of the discrimination network DN can distinguish the original data from the noisy data in its coding space, while the first term of the loss L GN of the generation network GN tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
As one embodiment of the invention, the training of the embedded migration module with the l-layer structured information is performed by:
first, a user embedded representation with l-layer structured information is obtained by the formula (4)
This is then used as input to train the migration section:
Wherein LeakyReLU is the activation function ReLU, W (l)、b(l) is a trainable network parameter, An l-order user-embedded token with adaptive noise generated by the generation network GN,Is a source domain moduleThe generated l-order user characterization; l DN is a decision to determine the network loss,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy l-order user characterizations generated for GNs; l GN is a loss to the generation network, wherein,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userThe first term to generate the loss L GN of the network GN then tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
It should be noted that, the training of the embedded migration module of the project is the same as that of the embedded migration module of the user, and will not be repeated here.
The following description will be made regarding the loss function, and in step S2, the loss function is as follows:
Wherein, For source domain network training loss, u is user ID, c is context, v is item ID, s uv is user u's true score for item v,The prediction of item v for user u is scored.
In step S3, the residual layer is fitted by using the embedded migration module, so as to achieve the migration effect of guiding the input of the target domain by using the model parameters of the source domain, and when the migration is performed, the parameters of the embedded migration module are fixed, the residual layer of the target domain is trained, and the loss is as follows:
LTransfer＝Lu+Lv (16)
L Transfer is migration loss, wherein L u is loss of user multi-level structured information migration, and L v is loss of item multi-level structured information migration; For user characterization with different hierarchically structured information, Representing the project with different hierarchical structured information; where u is the user set of the source domain and v is the item set;
wherein the residual layer is defined as follows:
Wherein, AndIs a module in the target domainAndThe obtained first-order user and item characterization; Is a trainable parameter; tanh is the activation function; for the l-order user and item characterizations generated after passing through the residual layer that are used as inputs to the embedded migration module.
As another embodiment of the present invention, there is also provided a one-to-many cross-domain recommendation system based on a high-level graph structure, including:
A user and item characterization module for obtaining user and item characterizations with context invariants, specifically comprising a context module M 1 for obtaining context invariants c nc from a source domain, an embedded characterization for obtaining user and item with different hierarchical structured information from the source domain AndSplicingAndResulting user and item representations with multi-layer structured informationAndIs embedded in the module of (a)
A context condition clustering module for combining the context invariant c nc extracted by the context module M 1 with the moduleAndResulting user and item representations with multi-layer structured informationAndCombining and obtaining user and item representations with most relevant contexts by multi-layer linear transformationAnd
A predictive scoring module M 4 for characterizing the resulting users and items with the most relevant contextAndVector multiplication is carried out to obtain a prediction score;
Embedding migration module, embedding module using source domain AndModule parameters train their migration portions, and then use their migration portions to fit the residual layers of the target domain to train the embedded modules of the target domainAndThe parameters of the module reach the aim of guiding the user of the target domain and embedding the project by the source domain multi-level structured information. Because we train the residual layer for each target domain, the embedded transfer module learns the data distribution for the source domain, therefore, we only need to use the embedded transfer module transfer part trained by the source domain data to fit the residual layer of the target domain during transfer, and the rest parameters use the parameters trained by the source domain, thereby achieving the one-to-many transfer target.
When the method is used, the feedback data of online users from different areas are collected as input data of a source domain and a target domain, and the different areas are divided into different domains. The online users are more, the feedback is more dense source domains, and the data is less sparse target domains. The data of the source domain is applied to obtain parameters of a user and a project embedding module through a single domain training frame, the parameters are used for training a migration part of the embedding migration module, then the embedding migration module is used for fitting a target domain residual layer with little data, the parameters of the target domain embedding module are trained, the user and the project embedding of the target domain are guided, the prediction scores of the user and the project are obtained through a target domain network, and whether the product is recommended to the user is judged.
It should be understood that the above description is not intended to limit the invention to the particular embodiments disclosed, but to limit the invention to the particular embodiments disclosed, and that various changes, modifications, additions and substitutions can be made by those skilled in the art without departing from the spirit and scope of the invention.
Claims (9)
1. The one-to-multi-cross-domain recommendation method based on the high-order graph structure is characterized in that in the process of embedding and characterizing users and projects, a plurality of embedded migration modules are established by utilizing the high-order graph structure, and multi-level structured information of the users and projects of a dense source domain is used for guiding the users and projects of a sparse target domain to be characterized in that the method comprises the following steps:
s1, acquiring user and item characterization with context invariants:
extracting context invariants in dense source domain by context module M 1 User embedded module through source domainAnd project embedding moduleObtaining token vectors for users and items with different hierarchically structured informationAndSplicing to obtain user and item characterization with multi-layer structured informationAndClustering module through user and project context conditionsAndExtracting context invariants from context module M 1 And (3) withAndCombining to finally obtain the user and item characterization with the most relevant contextAnd
S2, predicting interaction scoring results:
the end user and item characterization vectors obtained in step S1 in the source domain are processed by a predictive scoring module M 4 AndObtaining predictive score by multiplyingThen, carrying out back propagation on the difference between the predicted result and the real result to carry out training;
S3, migration:
The multi-level structured information of the source domain is used for guiding the user and the project embedding of the target domain through the embedding and migration module, the target domain user and the project embedding characterization with different levels of structured information obtained through the embedding and migration module are spliced, and then the context condition clustering module of the target domain user and the project is used for guiding the user and the project embedding characterization AndObtaining a final target domain user and a final project embedding representation;
the embedded migration module includes two parts:
First, embedded part:
embedded characterization of users and items with source domains AndAs input of the embedded migration module, the embedded migration module is patterned in an embedded part to obtain a user and a project representation with multi-level structured information, wherein the user representation with the u 1 self-node information and the user representation with the project v n and the user u 1 self-information are defined as the user representations with different hierarchical structured information, the project representation with the v 1 self-node information and the project representation with the u i and the v 1 self-information are defined as the project representations with different hierarchical structured information,
Wherein v n is the first-order neighbor node of user u 1, and u i is the first-order neighbor node of item v 1;
second, migration section:
Only the user and project characterization with different hierarchical structural information obtained by the source domain embedding part are used as input for training a universal embedded migration module; the migration part is used for fitting a residual layer of the target domain to achieve a one-to-many migration target;
S4, outputting a cross-domain recommendation result:
obtaining the predictive score of the target domain by multiplying the end user obtained in the step S3 by the item characterization And outputting a recommendation result according to the scores.
2. The one-to-many cross-domain recommendation method based on the high-order graph structure of claim 1, wherein the user embedded module is obtained after the step S2 source domain training is completedAnd project embedding moduleIn step S3, it is used for training of the embedded migration module; by fixing the parameters of the embedded part in each embedded migration module as the source domain user embedded moduleAnd project embedding moduleTraining the migration part embedded in the migration module according to the trained parameters; the trained embedded migration modules are used for fitting residual layers in the target domain, in the process of fitting the target domain, parameters of migration parts of each embedded migration module are fixed to be trained parameters by using source domain data, and the embedded parts are embedded by using target domain usersAnd project embedding moduleGenerated data and parameters, thereby training user embedded module of target domainAnd project embedding moduleThe multi-level structured information of the source domain is used for guiding the user of the target domain and the purpose of embedding the project.
3. The one-to-many cross-domain recommendation method based on a high-level graph structure according to claim 1, wherein in the embedding section, an output obtained after passing through a layer of embedding migration module embedding sectionAs an input of the embedded part of the second layer embedded migration module, and the like, in the high-order propagation process, the output of the first layer-1 is used as the input of the first layer:
Wherein, For the first order user characterization,Is the Laplace norm, where N u is the number of items that have interactions with user u 1 and N v is the number of users that have interactions with item v n; In order for the weight matrix to be trainable, For the l-1 order item characterization obtained by the embedding part of the embedding module,For the l-1 order user characterization obtained by the embedding part of the embedding module,Structured information between user u 1 and item v n; leakyReLU is a function of the activation ReLU,Characterization of the l-1 order itemAnd l-1 order user characterizationThe information of the interaction between the two,Characterization for l-1 order userInformation of the same.
4. A one-to-many domain recommendation method based on a high-level graph structure according to claim 3, characterized in that in the migration section, by designing an countermeasure network composed of a generation network GN and a discrimination network DN, learning a data distribution of user characterization with structured information, the structured information is migrated to a target domain for guiding user embedding of the target domain, specifically:
By using As input, mapping the input to a normal distribution of low dimensions through a discriminant network DNIn (1) obtaining hidden variablesSatisfy normal distribution and pass hidden variableReconstructing to obtainAt the same time, generating a noise-carrying signal with the generation network GNAnd generating hidden variables through networkIt also satisfies normal distribution and is obtained by reconstruction according to hidden variables
5. The one-to-many cross-domain recommendation method based on higher-order graph structure of claim 4, wherein the generation network GN generates an input adaptive noise
LeakyReLU is the activation function ReLU, W, b is a trainable network parameter,Embedding a token for a first-order user with adaptive noise generated by the network GN; reusing a first-order user embedding representation generated by the source domain user embedding module parameters for an embedding part of the embedding migration module;
The distinguishing network DN adopts a variation self-encoder for distinguishing whether the input comes from the original data distribution space or the noisy data distribution space:
where L DN is the discrimination of network loss, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy first-order user characterizations generated for GNs.
6. The one-to-many cross-domain recommendation method based on a high-order graph structure according to claim 5, wherein the discrimination network and the generation network adopt an alternate training mode when training, and the generation network is used for confusing the discrimination network, so that the discrimination network recognizes the noisy input as a true value:
Where L GN is the loss of the generation network, Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for first order userThe loss L DN of the discrimination network DN can distinguish the original data from the noisy data in its coding space, while the first term of the loss L GN of the generation network GN tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
7. A one-to-many cross-domain recommendation method based on a high-level graph structure according to claim 3, wherein the training of the embedded migration module with the l-layer structured information is performed by:
first, a user embedded representation with l-layer structured information is obtained by the formula (4)
This is then used as input to train the migration section:
wherein LeakyReLU is the activation function ReLU, W (l)、b(l) is a trainable network parameter, An l-order user-embedded token with adaptive noise generated by the generation network GN,Is a source domain moduleThe generated l-order user characterization; l DN is a decision to determine the network loss,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userIs provided for the distribution of the data of (a),Is distributed intoAnd standard normal distributionKL divergence between whereNoisy l-order user characterizations generated for GNs; l GN is a loss to the generation network, wherein,Is distributed intoAnd standard normal distributionThe degree of divergence of KL between the two,Characterization for an l-th order userThe first term to generate the loss L GN of the network GN then tries to make the discrimination network DN generate a more confusing noisy input, the second term is to prevent
The embedded migration module of the project is the same as the embedded migration module training of the user.
8. The one-to-many domain recommendation method based on the higher-order graph structure according to claim 1, wherein in step S3, an embedded migration module is used to fit a residual layer, so as to achieve a migration effect of guiding input of a target domain by using model parameters of a source domain, and when the migration is performed, parameters of the embedded migration module are fixed, the target domain residual layer is trained, and the loss is as follows:
LTransfer＝Lu+Lv (16)
L Transfer is migration loss, wherein L u is loss of user multi-level structured information migration, and L v is loss of item multi-level structured information migration; For user characterization with different hierarchically structured information, Representing the project with different hierarchical structured information; where u is the user set of the source domain and v is the item set;
wherein the residual layer is defined as follows:
Wherein, AndIs a module in the target domainAndThe obtained first-order user and item characterization; Is a trainable parameter; tanh is the activation function; for the l-order user and item characterizations generated after passing through the residual layer that are used as inputs to the embedded migration module.
9. A high-level graph structure-based one-to-many cross-domain recommendation system for implementing the high-level graph structure-based one-to-many cross-domain recommendation method as claimed in any one of claims 1 to 8, the high-level graph structure-based one-to-many cross-domain recommendation system comprising:
A user and item characterization module for obtaining user and item characterizations with context invariants, specifically comprising a context module M 1 for obtaining context invariants c nc from a source domain, an embedded characterization for obtaining user and item with different hierarchical structured information from the source domain AndSplicingAndResulting user and item representations with multi-layer structured informationAndIs embedded in the module of (a)
A context condition clustering module for combining the context invariant c nc extracted by the context module M 1 with the moduleAndResulting user and item representations with multi-layer structured informationAndCombining and obtaining user and item representations with most relevant contexts by multi-layer linear transformationAnd
A predictive scoring module M 4 for characterizing the resulting users and items with the most relevant contextAndVector multiplication is carried out to obtain a prediction score;
an embedded migration module comprising an embedded part and a migration part, the embedded module using a source domain AndEmbedding module for parameter training of migration portions thereof, followed by residual layer training of target domain fitting of migration portions thereof to target domainAndThe parameters of the target domain are achieved, and the target domain user and the project embedding target is guided by the source domain multi-level structured information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210018292.6A CN114443954B (en) | 2022-01-07 | 2022-01-07 | One-to-many cross-domain recommendation method and system based on high-order graph structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210018292.6A CN114443954B (en) | 2022-01-07 | 2022-01-07 | One-to-many cross-domain recommendation method and system based on high-order graph structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114443954A CN114443954A (en) | 2022-05-06 |
| CN114443954B true CN114443954B (en) | 2024-09-17 |
Family
ID=81367890
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210018292.6A Active CN114443954B (en) | 2022-01-07 | 2022-01-07 | One-to-many cross-domain recommendation method and system based on high-order graph structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114443954B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117743694B (en) * | 2024-02-06 | 2024-05-14 | 中国海洋大学 | Multi-layer transfer learning cross-domain recommendation method and system based on graph node characteristic enhancement |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111783526A (en) * | 2020-05-21 | 2020-10-16 | 昆明理工大学 | Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment |
| AU2020103905A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Unsupervised cross-domain self-adaptive medical image segmentation method based on deep adversarial learning |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112417298B (en) * | 2020-12-07 | 2021-06-29 | 中山大学 | Cross-domain recommendation method and system based on a small number of overlapped users |
| CN113806630B (en) * | 2021-08-05 | 2024-06-14 | 中国科学院信息工程研究所 | Attention-based multi-view feature fusion cross-domain recommendation method and device |
| CN113449205B (en) * | 2021-08-30 | 2021-11-09 | 四川省人工智能研究院(宜宾) | Recommendation method and system based on metadata enhancement |
-
2022
- 2022-01-07 CN CN202210018292.6A patent/CN114443954B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111783526A (en) * | 2020-05-21 | 2020-10-16 | 昆明理工大学 | Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment |
| AU2020103905A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Unsupervised cross-domain self-adaptive medical image segmentation method based on deep adversarial learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114443954A (en) | 2022-05-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Giles et al. | Learning, invariance, and generalization in high-order neural networks | |
| Zuo et al. | Fuzzy rule-based domain adaptation in homogeneous and heterogeneous spaces | |
| US9159021B2 (en) | Performing multistep prediction using spatial and temporal memory system | |
| KR102203065B1 (en) | Triple verification device and method | |
| Abro et al. | Natural language understanding for argumentative dialogue systems in the opinion building domain | |
| CN111506814B (en) | Sequence recommendation method based on variational self-attention network | |
| Zhou et al. | Learning household task knowledge from WikiHow descriptions | |
| CN113780470B (en) | Graph classification method based on self-adaptive multichannel cross graph convolutional network | |
| Ku et al. | A study of the Lamarckian evolution of recurrent neural networks | |
| CN113722439B (en) | Cross-domain emotion classification method and system based on antagonism class alignment network | |
| CN118296150B (en) | Comment emotion recognition method based on multi-countermeasure network improvement | |
| Rabuñal et al. | A new approach to the extraction of ANN rules and to their generalization capacity through GP | |
| CN114443954B (en) | One-to-many cross-domain recommendation method and system based on high-order graph structure | |
| CN111444328B (en) | Natural language automatic prediction and inference method with interpretation generation | |
| CN111274359B (en) | Query recommendation method and system based on improved VHRED and reinforcement learning | |
| CN113239143A (en) | Power transmission and transformation equipment fault processing method and system fusing power grid fault case base | |
| CN111026852B (en) | Financial event-oriented hybrid causal relationship discovery method | |
| Bouneffouf et al. | Survey on applications of neurosymbolic artificial intelligence | |
| Ishida et al. | Adjust-free adversarial example generation in speech recognition using evolutionary multi-objective optimization under black-box condition | |
| Gao et al. | Chinese short text classification method based on word embedding and Long Short-Term Memory Neural Network | |
| Hamida et al. | Hybrid-MELAu: A Hybrid Mixing Engineered Linguistic Features Framework Based on Autoencoder for Social Bot Detection. | |
| Hsu et al. | An interpretable generative adversarial approach to classification of latent entity relations in unstructured sentences | |
| CN112528015B (en) | Method and device for judging rumor in message interactive transmission | |
| Dasgupta et al. | A Review of Generative AI from Historical Perspectives | |
| Agathangelou et al. | Balancing between holistic and cumulative sentiment classification |
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 |