CN115809339A - Cross-domain recommendation method, system, device and storage medium - Google Patents
Cross-domain recommendation method, system, device and storage medium Download PDFInfo
- Publication number
- CN115809339A CN115809339A CN202210779167.7A CN202210779167A CN115809339A CN 115809339 A CN115809339 A CN 115809339A CN 202210779167 A CN202210779167 A CN 202210779167A CN 115809339 A CN115809339 A CN 115809339A
- Authority
- CN
- China
- Prior art keywords
- user
- domain
- item
- project
- cross
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a cross-domain recommendation method, a cross-domain recommendation system, a cross-domain recommendation device and a storage medium, wherein the method comprises the following steps: constructing a relation graph between users and items based on historical behavior data of the users in the user set and the items in the item set; constructing a knowledge graph between items in the item set based on the item attribute information; fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph; determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map; and determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domains. According to the technical scheme, the accuracy of project recommendation is improved.
Description
Technical Field
The present invention relates to the field of recommendation technologies, and in particular, to a cross-domain recommendation method, system, device, and storage medium.
Background
With the development of the network information age, the available information amount is exponentially increased, and a serious 'information overload' problem occurs. In order to solve the problem of efficiently acquiring effective information from mass data as required, a recommendation system is developed. At present, a single-domain recommendation method or a cross-domain product cross-domain recommendation method mainly exists in a recommendation system. For a cross-domain product cross-domain recommendation method, the fusion of cross-domain knowledge is mostly based on a user item scoring matrix between users and items, but the association of attribute preference between users and between items under a cross-domain recommendation scene is not fully considered, so that the recommendation accuracy is reduced.
Disclosure of Invention
The embodiment of the application aims to solve the problem of low recommendation accuracy of a recommendation system by providing a cross-domain recommendation method, system, device and storage medium.
The embodiment of the application provides a cross-domain recommendation method, which comprises the following steps:
constructing a relation graph between users and items based on historical behavior data of the users in the user set and the items in the item set;
constructing a knowledge graph between items in the item set based on the item attribute information;
fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph;
determining user characteristic vectors and project characteristic vectors between domains and determining user characteristic vectors and project characteristic vectors in domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map;
and determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domains.
Optionally, the cross-domain collaborative knowledge graph includes a plurality of triples composed of head entities, relationships, and tail entities, and in the cross-domain collaborative knowledge graph, the items associated with the user and the user are connected by using directed edges, and the items associated with the user are connected by using directed edges between the corresponding entities in the cross-domain collaborative knowledge graph.
Optionally, the determining the user feature vector and the project feature vector between domains based on the embedded attention mechanism and the cross-domain collaborative knowledge graph includes:
performing high-order contact mining based on the cross-domain collaborative knowledge graph to obtain a plurality of relationship links;
distributing attention weights to the neighbor nodes according to the intimacy of each relationship link to obtain an inter-domain user matrix and a project matrix;
vectorizing the user matrix and the project matrix to obtain the inter-domain user characteristic vectors and the inter-domain project characteristic vectors.
Optionally, the determining the user feature vector and the item feature vector in the domain includes:
acquiring target users associated with the first type of projects in a target field based on the cross-field collaborative knowledge graph, and constructing rows of a user project scoring matrix according to the target users;
constructing a column of the user item scoring matrix according to a second type of item in the target field;
determining a similarity between the first category of items and the second category of items, wherein the similarity characterizes a rating of the second category of items by the target user;
constructing a scoring matrix of the user item according to the scores;
decomposing the user project scoring matrix to obtain a user matrix and a project matrix;
and determining a user characteristic vector in the domain according to the user matrix, and determining a project characteristic vector in the domain according to the project matrix.
Optionally, the step of determining a target recommended item based on the user feature vector and the item feature vector between the domains and the user feature vector and the item feature vector within the domains includes:
determining a scoring result corresponding to each project by the user based on the user characteristic vector and the project characteristic vector between the domains and the user characteristic vector and the project characteristic vector in the domains;
and determining the target recommended item according to the grading result.
Optionally, the step of determining a scoring result corresponding to each project by the user based on the user feature vector and the project feature vector between the domains and the user feature vector and the project feature vector within the domains includes:
fusing the user characteristic vector and the project characteristic vector in the domain and the user characteristic vector and the project characteristic vector between the domains to obtain a prediction scoring result of each project related to the user by the user;
determining a scoring error corresponding to each project according to the prediction scoring result and the actual scoring result;
optimizing the prediction scoring result by adopting the scoring error to obtain a target prediction scoring result corresponding to each project by the user;
the step of determining the target recommended item according to the scoring result includes:
and determining a target recommended item according to a target prediction scoring result corresponding to each item by the user.
Optionally, the step of determining a target recommended item according to the target prediction scoring result corresponding to each item by the user further includes:
sorting each item based on the target prediction scoring result;
acquiring the items of which the target prediction scoring result is greater than a preset scoring threshold value in the sorted items;
determining the item as the target recommended item.
In addition, to achieve the above object, the present invention further provides a cross-domain recommendation system, including:
the construction module is used for constructing a relation graph between users and items based on historical behavior data of the users in the user set and the items in the item set, and constructing a knowledge graph between the items in the item set based on item attribute information;
the fusion module is used for fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph;
the determining module is used for determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map;
and the target recommended item prediction module is used for determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domains.
In addition, to achieve the above object, the present invention further provides a terminal apparatus, which includes: the cross-domain recommendation system comprises a memory, a processor and a cross-domain recommendation program stored on the memory and capable of running on the processor, wherein the cross-domain recommendation program realizes the steps of the cross-domain recommendation method when being executed by the processor.
In addition, to achieve the above object, the present invention further provides a storage medium having a cross-domain recommendation program stored thereon, wherein the cross-domain recommendation program, when executed by a processor, implements the steps of the cross-domain recommendation method.
According to the technical scheme of the cross-domain recommendation method, the cross-domain recommendation system, the cross-domain recommendation equipment and the cross-domain recommendation storage medium, because the users in the user set and the items in the item set are based on the historical behavior data of the users, a relation graph between the users and the items is constructed; constructing a knowledge graph between items in the item set based on the item attribute information; fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph; determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map; according to the technical scheme for determining the target recommended project based on the user feature vector and the project feature vector between the domains and the user feature vector and the project feature vector in the domains, association of attribute preference between users and between projects can be established besides establishing association according to rating data between the users and the projects. The problem of low project recommendation accuracy is solved, and the project recommendation accuracy is improved.
Drawings
FIG. 1 is a schematic diagram of a hardware operating environment according to an embodiment of the present invention;
FIG. 2 is a flowchart illustrating a cross-domain recommendation method according to a first embodiment of the present invention;
FIG. 3 is a functional block diagram of a recommender system in accordance with the present invention;
FIG. 4 is a schematic cross-domain collaborative knowledge map of the present invention;
FIG. 5 is a schematic diagram of a cross-domain recommendation model based on attention knowledge graph embedding in accordance with the present invention;
FIG. 6 is a schematic diagram of an inter-domain model structure of the graph neural network method of the present invention with an embedded attention mechanism;
FIG. 7 is a flow chart illustrating an embodiment of the present invention.
The objectives, features, and advantages of the present application will be further understood by reference to the following description, taken in conjunction with the accompanying drawings, which are included by way of illustration of exemplary embodiments and are not intended to limit the invention to the full extent specified.
Detailed Description
In the application, in order to solve the accuracy of recommendation, a cross-domain knowledge graph is constructed based on project attributes and fused with a user-project behavior interaction graph to form a cross-domain collaborative knowledge graph containing user-user and project-project relationships; processing cross-domain high-order relation by using a collaborative knowledge graph, introducing an attention weight calculation method for relation perception propagation, distributing different attention weights for neighbors according to relation intimacy, introducing an attention mechanism into a ripplenet graph neural network model to realize propagation and calculation of high-order relation, and obtaining vectorization expression of users and items in a cross-domain scene. And simultaneously, in the target domain, obtaining vectorization representation of the users and the items in the single target domain by using a matrix decomposition method according to the user-item scoring matrix. And finally, combining the characteristics of the user and the project under the cross-domain scene and the single target domain, and obtaining the cross-domain recommendation result of the user through a Bayesian posterior sorting algorithm.
Compared with the related art, the main innovation points of the application comprise the following parts:
the invention provides a cross-domain collaborative knowledge graph construction method, which realizes collaborative knowledge graph construction by fusing cross-domain user-project behavior data and a cross-domain project knowledge graph. Meanwhile, a relation propagation-based attention mechanism is introduced and embedded into a ripplenet graph attention network to realize high-order preference propagation and mining, and in a preference prediction stage, user item preference is predicted by fusing characteristics of users and items in a target domain and an inter-domain. The method not only integrates the preference information of the user target domain and the cross-domain, effectively solves the problems of sparsity, cold start and low accuracy, but also efficiently realizes the propagation and mining of high-dimensional and implicit preference information through the cross-domain collaborative knowledge map and the graph attention network.
Secondly, constructing a user-project-entity interconnected collaborative knowledge map in a cross-domain scene, realizing neighbor weight distribution through an attention mechanism for measuring neighbor semantic distance, realizing vectorization expression of user and project characteristics through a ripplenet graph neural network embedded with attention, and finally integrating the characteristics in a user domain and between domains to realize recommendation. Compared with the related art, the method and the device have the advantages that the complementation and the correlation of the cross-domain information are better realized, and meanwhile, the similarity of the implicit potential interest preference among the cross-domain information, different projects and different users is better discovered through the optimized ripplent graph neural network on the basis of the collaborative knowledge graph.
Thirdly, on the basis of constructing the cross-domain collaborative knowledge graph, the cross-domain recommendation is realized on the basis of fully mining potential semantic relation and preference propagation between users, between users and items and between items through a graph neural network with an attention mechanism. The method solves the problems of data sparseness and cold start of single-field recommendation, realizes high-order potential semantic relation mining, and has better performance on accuracy and diversity of recommendation results.
The technical solutions of the present application will be specifically described below by way of examples.
Referring to fig. 7, the present application includes: the system comprises a cross-domain collaborative knowledge map module, a target domain modeling module based on matrix decomposition, an inter-domain modeling module based on a graph neural network method of an embedded attention mechanism and a recommendation result prediction module fusing intra-domain and inter-domain information.
For a better understanding of the above technical solutions, exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
As shown in fig. 1, fig. 1 is a schematic structural diagram of a hardware operating environment according to an embodiment of the present invention.
It should be noted that fig. 1 is a schematic structural diagram of a hardware operating environment of the terminal device.
As shown in fig. 1, the terminal device may include: a processor 1001, e.g. a CPU, a memory 1005, a user interface 1003, a network interface 1004, a communication bus 1002. Wherein a communication bus 1002 is used to enable connective communication between these components. The user interface 1003 may include a Display screen (Display), an input unit such as a Keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface, a wireless interface. The network interface 1004 may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory (e.g., a magnetic disk memory). The memory 1005 may alternatively be a storage device separate from the processor 1001.
Those skilled in the art will appreciate that the terminal device configuration shown in fig. 1 is not meant to be limiting for the terminal device and may include more or fewer components than those shown, or some components may be combined, or a different arrangement of components.
As shown in fig. 1, a memory 1005, which is a storage medium, may include therein an operating system, a network communication module, a user interface module, and a cross-domain recommendation program. The operating system is a program for managing and controlling hardware and software resources of the terminal device, and recommends the running of programs and other software or programs across fields.
In the terminal device shown in fig. 1, the user interface 1003 is mainly used for connecting a terminal, and performing data communication with the terminal; the network interface 1004 is mainly used for the background server and performs data communication with the background server; the processor 1001 may be used to invoke a cross-domain recommender stored in the memory 1005.
As shown in fig. 2, in a first embodiment of the present application, the cross-domain recommendation method of the present application includes the following steps:
step S110, constructing a relation graph between the users and the items based on the historical behavior data of the users in the user set and the items in the item set.
In this embodiment, a cross-domain knowledge graph can be constructed based on project attribute information, and is fused with a user-project behavior interaction graph to form a cross-domain collaborative knowledge graph containing user-user and project-project relationships. In the single domain network graph information, the single domain network graph information includes scoring data from a user node to a project node, and project attribute information. In order to realize the fusion of information of a plurality of fields, firstly, based on the historical behavior data of the user, the behavior data of the user in the plurality of fields is integrated to form a relation graph G between the user and the project under the cross-field scene 1 ，G1＝{(u,y ui ,i|u∈U,i∈I}G1＝{(u,y ui I | U ∈ U, I ∈ I }, where y ui y ui =1 indicates that the user u and the item i have a behavioral interaction.
Specifically, in the cross-domain collaborative knowledge graph, a plurality of nodes exist, and a corresponding edge relationship exists between every two nodes. And constructing a relation graph between the users and the items based on the historical behavior data of the users in the user set and the items in the item set. Specifically, the first project node and the first edge relationship between the user node and the first project node may be determined according to historical behavior data associated with the user node. The historical user behavior data associated with the user nodes comprises information such as time, duration, item names and item types of items such as an art, a movie and a television show watched by the user. The historical behavior data of the user is associated with the user node, for example, when the user logs in the entertainment software to watch a certain project, the platform automatically acquires the behavior data of the user, associates the behavior data with the user account, and stores the behavior data in a local or cloud database. And when user historical behavior data corresponding to the user node is acquired subsequently, the user historical behavior data associated with the user node is crawled from a local database or a cloud database. The items that the user node has viewed in a certain period of time in the past can be determined from the user historical behavior data, that is, the first item node can be determined according to the user historical behavior data associated with the user node.
After determining a user node and a first project node associated with the user node, connecting the user node and the first project node to form a first edge relationship, wherein the first edge relationship is used for representing the relationship between the user node and the first project node. For example, referring to FIG. 4, each user represents a user node and each project represents a project node. Each user node is connected to at least one project node. Where user 1 represents a user node. The projects connected by the user 1 may be "Tang street quest 3" and "happiness trio", with "Tang street quest 3" or "happiness trio" representing the first project node. The connection relationship between the user 1 and the "tang jieying 3" may be referred to as a first edge relationship, or the connection relationship between the user 1 and the "happy trio" may be referred to as a first edge relationship.
And step S120, constructing a knowledge graph among the items in the item set based on the item attribute information.
In the embodiment, secondly, the item-to-item relation among different domain items is established by utilizing the attribute information of the items, and a knowledge graph G between the items under the cross-domain scene is formed 2 ,
Where the triplet (h, r, t) constitutes a representation of a relationship from head entity h to tail entity t, such as the movie tang street quest 3 and the television drama soldier assault being associated by a common actor "Wang Baojiang".
Specifically, the item corresponding to each item node has corresponding item attribute information. For example, in a movie project of "Tang street quest 3", the project attribute information may be actors participating in the project, such as "Wang Baojiang". The item attribute information may also be a type of the item, for example the item type is comedy. The item attribute information may also be the director of the item, for example, the director of the "Tang street visit 3" is Chen Saicheng. The item attribute information may also include other content.
In this embodiment, item attribute information corresponding to a first item node is obtained, and a second item node is further determined according to the item attribute information, that is, the first item node and the second item node have the same item attribute information. For example, referring to fig. 4, the first project node is "down street quest 3", the second project node is "soldier assault", the "down street quest 3" and "soldier assault" are played by a common actor, which is "Wang Baojiang".
Optionally, an entity node may be determined according to the item attribute information corresponding to the first item node; determining a second item node associated with the entity node; determining a third edge relationship of the first project node and the entity node, and determining a fourth edge relationship of the first project node and the second project node; and determining the second edge relation according to the third edge relation and the fourth edge relation. Specifically, referring to fig. 4, an entity node "Wang Baojiang" is determined according to project attribute information, e.g., actor, corresponding to the first project node "tang man street quest 3". And determining a second project node 'soldier assault' associated with 'Wang Baojiang' again according to the project attribute information. After the second item node is determined, the second user node can be obtained, and so on, thereby realizing the propagation and mining of the high-order relationship.
And S130, fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph.
In the embodiment, the relationship graph G between the user and the project is finally integrated 1 And knowledge-graph G between projects 2 And forming a cross-domain collaborative knowledge graph G of user-project-user interconnection.
R' = R utoxy. By constructing a cross-domain collaborative knowledge graph to construct a network of user-project-attribute interrelations, higher-order association mining of graph nodes can be achieved, for example, the 5-order relationship of user 1 is as follows. Through 5-order connection, the user is helped to mine the television series (soldier assault) with the same attribute as the watched movie (Tang street quest 3), and similar user mining (user 3) is realized.
Optionally, referring to fig. 4, the cross-domain collaborative knowledge graph includes a plurality of triples consisting of head entities, relationships, and tail entities. The user in fig. 4 may be a head entity, and the item may be a tail entity. And in the cross-domain collaborative knowledge graph, connecting the user with the project associated with the user by using a directed edge, and connecting the project associated with the user with the project associated with the user by using a directed edge between corresponding entities in the cross-domain collaborative knowledge graph.
And step S140, determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on the embedded attention mechanism and the cross-domain collaborative knowledge graph.
And S150, determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domain.
In this embodiment, after the cross-domain collaborative knowledge graph is determined, since the cross-domain collaborative knowledge graph includes relationships among users, items and entities, the cross-domain collaborative knowledge graph is analyzed, and then the user node corresponding to the target recommendation item is determined according to the cross-domain collaborative knowledge graph. A complex network capable of displaying high-order relation of the items and the users can be constructed through a knowledge graph between the items and a relation graph between the users and the items, and fusion preference propagation of the similar users and the similar items is achieved.
Specifically, a cross-domain high-order relation is processed by using a collaborative knowledge graph, an embedded attention weight calculation method of relation perception propagation is introduced, different attention weights are distributed to neighbors according to relation intimacy, an attention mechanism is introduced into a ripplenet graph neural network model to realize propagation and calculation of a high-order relation, and a user feature vector and a project feature vector under a cross-domain scene are obtained. And simultaneously, in the target domain, according to the user-project user project rating matrix, obtaining the user characteristic vector and the project characteristic vector in the single target domain by using a matrix decomposition method. And finally, combining the user characteristic vector and the project characteristic vector under the cross-domain scene and the user characteristic vector and the project characteristic vector in the single target domain, and obtaining a cross-domain recommendation result of the user through a Bayesian posterior sorting algorithm to obtain the target recommendation project.
In this embodiment, an architecture diagram of the cross-domain recommendation method of the present application is shown in fig. 5, and includes two parts, namely intra-domain modeling and inter-domain modeling, which respectively obtain intra-domain and inter-domain user feature vectors and project feature vectors, and implement fusion prediction in a recommendation result prediction module. Determining that the following conditions exist in the target recommended item corresponding to the user node:
firstly, determining a user characteristic vector and a project characteristic vector in a domain, and further determining a target recommended project according to the user characteristic vector and the project characteristic vector in the domain.
And secondly, determining user characteristic vectors and project characteristic vectors between domains, and determining a target recommendation project based on the user characteristic vectors and the project characteristic vectors between the domains.
And thirdly, determining the user characteristic vector and the project characteristic vector in the domain, and simultaneously determining the user characteristic vector and the project characteristic vector between domains, and further determining the target recommended project according to the user characteristic vector and the project characteristic vector in the domain and the user characteristic vector and the project characteristic vector between domains.
In the technical scheme of the application, the cross-domain collaborative knowledge map is constructed, the graph neural network algorithm embedded with the attention mechanism is introduced, the target domain and cross-domain user and project features are integrated to realize recommendation, and the problems of insufficient user-project relationship mining, sparsity, low accuracy and the like of the existing cross-domain recommendation algorithm are effectively solved.
Optionally, the step of determining the user feature vector and the project feature vector between domains based on the embedded attention mechanism and the cross-domain collaborative knowledge graph includes:
step S141, performing high-order contact mining based on the cross-domain collaborative knowledge graph to obtain a plurality of relationship links;
step S142, distributing attention weight to the neighbor nodes according to the intimacy of each relationship link to obtain an inter-domain user matrix and a project matrix;
step S143, vectorizing the user matrix and the project matrix to obtain the inter-domain user characteristic vector and the project characteristic vector.
In this embodiment, in the cross-domain collaborative knowledge graph, each flow corresponds to an attribute graph, and the attribute graph can be converted into a graph vector, where each flow includes a plurality of edge relationships. On the basis of a cross-domain collaborative knowledge graph, an attention mechanism based on relationship perception is introduced, attention weights are distributed to different neighbor nodes, high-order relationship propagation and calculation are achieved through a ripplenet graph neural network embedded with attention, and a user matrix G (UA) and an item matrix G (IA) under a cross-domain scene are formed. The model mainly comprises an embedding layer, an attention layer based on relation perception and an attention embedding propagation layer, and a schematic diagram of an inter-domain model structure of the graph neural network method of the embedded attention mechanism is shown in FIG. 6.
Specifically, the embedding layer is a process of mapping entities and relationships of the cross-domain collaborative knowledge graph into low-dimensional dense vectors. Firstly, data compression is realized by means of graph embedding, and the attribute graph is converted into a vector or a vector set. By introducing a Graph2vec algorithm, the code of a Graph entity is input as a unique heat vector, all character graphs are sampled and marked again in the Graph, a skipgram hopping Graph model is trained, and a vector set of subgraphs is output. Graph embedding may preserve the graph topology, vertex-to-vertex relationships, and other information of the subgraph in its entirety. Secondly, on the basis of graph embedding, a TransR model is used for realizing vectorization representation of entities and relations in the graph. Specifically, for a triplet (h, r, t), embedding vectors of a head entity and a tail entity formed after graph embedding are mapped to a relationship space r, and a user feature vector and a project feature vector of a project, a user and a project-user relationship are obtained by using SGD gradient descent algorithm training through the following vector operation equation 1 and a scoring function equation 2.
In the above-mentioned formula,
and
respectively representing the embedded vectors of the head entity node and the tail entity node after the transR transformation.
Based on the attention layer of the relation perception, an attention weight calculation method of the relation perception propagation is introduced, and different attention weights are distributed to the neighbors according to the relation intimacy. For an entity h in the cross-domain collaborative knowledge graph, h can be contained in multiple triples to achieve multi-order relationship propagation. Suppose N h Representing all triples, N, headed by entity h h And { (h, r, t) | (h, r, t) ∈ G }. For the first order connectivity of entity h, the linear combination of the h-centric network is
Where pi (h, r, t) is the attenuation coefficient of each propagation of the triplet (h, r, t). An attention mechanism based on relation perception propagation is introduced, neighbor distribution weights are obtained through measuring entity distances, an attenuation coefficient calculation method is shown as a formula 3, and normalization of attenuation coefficients is achieved through a formula 4.
π(h,r,t)＝(W r e t )′tanh((W r e h +e r )). (formula 3)
The attention embedding propagation layer realizes high-order relation propagation and calculation through a ripplenet graph neural network embedding attention on the basis of acquiring neighbor weights based on an attention mechanism, and forms a user matrix G (UA) and an item matrix G (IA) under a cross-domain scene. In cross-domain collaborative knowledge-graph, assumptions
Set of entities representing k-degree of association with entity p
Then higher order preference is implemented in the ripplenet propagation algorithm by the entity associated with entity pAnd propagation, wherein a vector calculation method of the entity p in the k-order propagation scene is shown in formula 5, and an embedded vector calculation method of the entity in the L-order propagation scene of the user p is shown in formula 6. And obtaining a user matrix G (UA) and an item matrix G (IA) under a cross-domain scene through iterative computation.
In the technical scheme of the application, a collaborative knowledge graph is used for processing cross-domain high-order relation, a relation perception propagation attention weight calculation method is introduced, different attention weights are distributed to neighbors according to relation intimacy, an attention mechanism is introduced into a ripplenet graph neural network model to realize propagation and calculation of high-order relation, and vectorization representation of users and items in a cross-domain scene is obtained. Through the optimized ripplent graph neural network on the basis of the collaborative knowledge graph, the similarity of potential interest preference implied among different fields, different projects and different users is better discovered.
Optionally, determining the user feature vector and the item feature vector within the domain comprises:
step S144, acquiring target users associated with the first type of projects in a target field based on the cross-field collaborative knowledge graph, and constructing rows of a user project scoring matrix according to the target users;
step S145, constructing a column of the user item scoring matrix according to the second type of items in the target field;
step S146, determining similarity between the first category of items and the second category of items, wherein the similarity represents the score of the target user on the second category of items;
step S147, constructing a user project scoring matrix according to the scores;
step S148, decomposing the user project scoring matrix to obtain a user matrix and a project matrix;
step S149, determining the user feature vector in the domain according to the user matrix, and determining the item feature vector in the domain according to the item matrix.
In the embodiment, on the basis of constructing the cross-domain knowledge graph, the user-project behavior information of the user in the target domain is collected, a user-project user project rating matrix is constructed, and an SVD matrix decomposition method is introduced to form a user matrix E (UA) and a project matrix E (IA) of the user in the target domain A.
Specifically, from a cross-domain collaborative knowledge graph, there are multiple users viewing the same type of item. Target users viewing a first category of items may be obtained from the cross-domain collaborative knowledge graph, for example, users viewing the movie "Tang street quest 3" including user 1 and user 2. And after determining the target users associated with the first type of items, taking the target users as the rows of the user item scoring matrix. And constructing columns of the user item scoring matrix according to the second type items in the target field. After determining the target users associated with the first type of items and the second type of items, determining the similarity between the first type of items and the second type of items. For example, the similarity between the first category of items and the second category of items may be determined based on the Euclidean distance principle, and the similarity is used as the score of the target user for the second category of items. And after determining the scores of the target users for the second category items, constructing a user item score matrix according to the scores. The user item scoring matrix is a user-item user item scoring matrix.
In this embodiment, after determining the user item scoring matrix, an SVD matrix decomposition method may be introduced, and the user item scoring matrix is decomposed by the SVD matrix decomposition method to obtain a user matrix and an item matrix. And determining a user characteristic vector according to the user matrix, namely determining a project characteristic vector according to the project matrix.
According to the technical scheme, in the target field, the vectorization expression of the users and the items in the single target field is obtained by utilizing a matrix decomposition method according to the user item scoring matrix.
Optionally, the step of determining the target recommended item based on the user feature vector and the item feature vector between the domains and the user feature vector and the item feature vector within the domains includes:
step S151, determining a scoring result corresponding to each project by the user based on the user characteristic vector and the project characteristic vector between the domains and the user characteristic vector and the project characteristic vector in the domains;
and step S152, determining the target recommended item according to the grading result.
Optionally, determining, based on the inter-domain user feature vector and the item feature vector and the intra-domain user feature vector and the item feature vector, a scoring result corresponding to each item by the user specifically includes:
step S1511, fusing the user feature vector and the project feature vector in the domain and the user feature vector and the project feature vector between the domains to obtain the prediction scoring result of the user for each project associated with the user;
step 1512, determining a scoring error corresponding to each project according to the predicted scoring result and the actual scoring result;
and S1513, optimizing the prediction scoring result by using the scoring error to obtain a target prediction scoring result corresponding to each project by the user.
Optionally, the determining the target recommended item according to the scoring result specifically includes the following steps:
s1521, determining target recommended items according to the target prediction scoring results corresponding to the items by the user.
In this embodiment, each item has a corresponding predictive scoring result, which is determined by a bayesian posterior ranking algorithm. And each item has a corresponding actual scoring result, and the actual scoring result is a preset value. And optimizing the model of the prediction scoring result according to the scoring error between the prediction scoring result and the actual scoring result so as to obtain a target prediction scoring result corresponding to each project.
In the embodiment, user feature vectors and project feature vectors obtained by modeling in a target field and modeling between cross-field fields are fused, a Bayesian posterior sorting algorithm is introduced to realize project preference prediction of a user, and N projects which are the best preference of the user are output from high to low according to prediction scores and serve as recommended projects of the user. Firstly, fusing the characteristics of users and items in the domain and between the domains to obtain the prediction scores of the users to the items, wherein the prediction formula is shown as a formula 7. And (3) realizing error measurement of the real preference and the prediction preference of the user by using a Bayesian posterior sorting algorithm (as shown in a formula 8), and outputting the prediction score of the user u on the item i through iterative optimization of a loss function shown in a formula 9.
Optionally, the step of determining a target recommended item according to the target prediction scoring result corresponding to each item by the user further includes:
step 15211, sorting the items based on the target prediction scoring result;
step 15222, acquiring items of which the target prediction scoring result is greater than a preset scoring threshold value in the sorted items;
step S15213, determining the item as the target recommended item.
In this embodiment, given the number M of final recommendation results, the top M contents with the highest prediction scores are taken as the content recommendation results of the user. Or after the items are sorted based on the target prediction scoring result, the items with the target prediction scoring result larger than a preset scoring threshold value in the sorted items are obtained, and the items are determined as target recommended items. And recommending the target recommended item to the user.
In the technical scheme of the application, a cross-domain knowledge graph is constructed based on project attributes and is fused with a user-project behavior interaction graph to form a cross-domain collaborative knowledge graph containing user-user and project-project relationships; processing cross-domain high-order relation by using a collaborative knowledge graph, introducing an attention weight calculation method for relation perception propagation, distributing different attention weights for neighbors according to relation intimacy, introducing an attention mechanism into a ripplenet graph neural network model to realize propagation and calculation of high-order relation, and obtaining vectorization expression of users and items in a cross-domain scene. And simultaneously, in the target domain, according to the user-item user item scoring matrix, obtaining the vectorization representation of the user and the item in the single target domain by using a matrix decomposition method. And finally, combining the characteristics of the user and the project under the cross-domain scene and the single target domain, and obtaining the cross-domain recommendation result of the user through a Bayesian posterior sorting algorithm. The method constructs a cross-domain collaborative knowledge map, introduces a graph neural network algorithm embedded with an attention mechanism, synthesizes target domains and cross-domain user and project characteristics to realize recommendation, and effectively overcomes the problems of insufficient mining of user-project relationship, sparsity problem, low accuracy and the like of the existing cross-domain recommendation algorithm.
While a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in an order different than presented herein.
As shown in fig. 3, the present application provides a cross-domain recommendation system, which includes:
the building module 10 is configured to build a relationship graph between users and items based on historical behavior data of the users in the user set and the items in the item set, and build a knowledge graph between the items in the item set based on item attribute information;
the fusion module 20 is configured to fuse the knowledge graph between the projects and the relationship graph between the users and the projects to obtain a cross-domain collaborative knowledge graph;
a determining module 30, configured to determine user feature vectors and project feature vectors between domains and determine user feature vectors and project feature vectors within domains based on an embedded attention mechanism and the cross-domain collaborative knowledge graph;
and the target recommended item prediction module 40 is used for determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domain.
Optionally, the cross-domain collaborative knowledge graph in the fusion module 20 includes a plurality of triples composed of a head entity, a relationship, and a tail entity, and in the cross-domain collaborative knowledge graph, the user and the item associated with the user are connected by using a directed edge, and the item associated with the user are connected by using a directed edge between the corresponding entities in the cross-domain collaborative knowledge graph.
Optionally, the determining module 30 may be further configured to perform high-order contact mining based on the cross-domain collaborative knowledge graph to obtain a plurality of relationship links; distributing attention weights to the neighbor nodes according to the intimacy of each relationship link to obtain an inter-domain user matrix and a project matrix; vectorizing the user matrix and the project matrix to obtain the inter-domain user characteristic vectors and the inter-domain project characteristic vectors.
Optionally, the determining module 30 may be further configured to obtain a target user associated with the first type of project in the target domain based on the cross-domain collaborative knowledge graph, and construct a row of a user project scoring matrix according to the target user; constructing a column of the user item scoring matrix according to a second type of item in the target field; determining a similarity between the first category of items and the second category of items, wherein the similarity characterizes a rating of the second category of items by the target user; constructing a scoring matrix of the user item according to the scores; decomposing the user project scoring matrix to obtain a user matrix and a project matrix; and determining a user characteristic vector in the domain according to the user matrix, and determining a project characteristic vector in the domain according to the project matrix.
Optionally, the target recommended item prediction module 40 may be further configured to determine, based on the user feature vector and the item feature vector between the domains and the user feature vector and the item feature vector within the domains, a scoring result corresponding to each item by the user; and determining the target recommended item according to the grading result.
Optionally, the target recommended item prediction module 40 may be further configured to fuse the user feature vector and the item feature vector within the domain and the user feature vector and the item feature vector between the domains to obtain a prediction scoring result of each item associated with the user by the user; determining a scoring error corresponding to each project according to the prediction scoring result and the actual scoring result; optimizing the prediction scoring result by adopting the scoring error to obtain a target prediction scoring result corresponding to each project by the user; and determining a target recommended item according to a target prediction scoring result corresponding to each item by the user.
Optionally, the target recommended item prediction module 40 may be further configured to sort each item based on the target prediction scoring result; acquiring the items of which the target prediction scoring result is greater than a preset scoring threshold value in the sorted items; determining the item as the target recommended item.
The specific implementation of the recommendation system of the present invention is substantially the same as the embodiments of the cross-domain recommendation method, and is not described herein again.
Based on the same inventive concept, an embodiment of the present application further provides a computer-readable storage medium, where a cross-domain recommendation program is stored, and when executed by a processor, the cross-domain recommendation program implements the steps of the cross-domain recommendation method described above, and can achieve the same technical effect, and in order to avoid repetition, the description is omitted here.
Since the storage medium provided in the embodiments of the present application is a storage medium used for implementing the method in the embodiments of the present application, based on the method described in the embodiments of the present application, a person skilled in the art can understand a specific structure and a modification of the storage medium, and thus details are not described here. Any storage medium used in the methods of the embodiments of the present application is intended to be within the scope of the present application.
As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the invention.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.
Claims (10)
1. A cross-domain recommendation method is characterized by comprising the following steps:
constructing a relation graph between users and items based on historical behavior data of the users in the user set and the items in the item set;
constructing a knowledge graph between items in the item set based on the item attribute information;
fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph;
determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map;
and determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domains.
2. The cross-domain recommendation method of claim 1, wherein the cross-domain collaborative knowledge graph comprises a plurality of triples consisting of head entities, relations, and tail entities, and wherein in the cross-domain collaborative knowledge graph, a user is connected with an item associated with the user using a directed edge, and an item associated with the user is connected with an item associated with the user using a directed edge between corresponding entities in the cross-domain collaborative knowledge graph.
3. The cross-domain recommendation method of claim 1, wherein the determining user feature vectors and item feature vectors between domains based on an embedded attention mechanism and the cross-domain collaborative knowledge graph comprises:
performing high-order contact mining based on the cross-domain collaborative knowledge graph to obtain a plurality of relationship links;
distributing attention weights to the neighbor nodes according to the intimacy of each relationship link to obtain an inter-domain user matrix and a project matrix;
vectorizing the user matrix and the project matrix to obtain the inter-domain user characteristic vector and the project characteristic vector.
4. The cross-domain recommendation method of claim 3, wherein said determining user feature vectors and item feature vectors within a domain comprises:
acquiring target users associated with the first type of projects in a target field based on the cross-field collaborative knowledge graph, and constructing rows of a user project scoring matrix according to the target users;
constructing a column of the user item scoring matrix according to a second type of item in the target field;
determining a similarity between the first category of items and the second category of items, wherein the similarity characterizes a rating of the second category of items by the target user;
constructing a user project scoring matrix according to the scoring;
decomposing the user project scoring matrix to obtain a user matrix and a project matrix;
and determining a user characteristic vector in the domain according to the user matrix, and determining a project characteristic vector in the domain according to the project matrix.
5. The cross-domain recommendation method according to claim 1 or 4, wherein the step of determining a target recommended item based on the inter-domain user feature vector and the item feature vector and the intra-domain user feature vector and the item feature vector comprises:
determining a scoring result corresponding to each project by the user based on the user characteristic vector and the project characteristic vector between the domains and the user characteristic vector and the project characteristic vector in the domains;
and determining the target recommended item according to the grading result.
6. The cross-domain recommendation method of claim 5, wherein the step of determining the scoring results corresponding to each project by the user based on the inter-domain user feature vectors and project feature vectors and the intra-domain user feature vectors and project feature vectors comprises:
fusing the user characteristic vector and the project characteristic vector in the domain and the user characteristic vector and the project characteristic vector between the domains to obtain a prediction scoring result of each project related to the user by the user;
determining a scoring error corresponding to each project according to the prediction scoring result and the actual scoring result;
optimizing the prediction scoring result by adopting the scoring error to obtain a target prediction scoring result corresponding to each project by the user;
the step of determining the target recommended item according to the scoring result includes:
and determining a target recommended item according to a target prediction scoring result corresponding to each item by the user.
7. The cross-domain recommendation method of claim 6, wherein the step of determining a target recommended item according to the target prediction scoring result corresponding to each item by the user further comprises:
sorting each item based on the target prediction scoring result;
acquiring the items of which the target prediction scoring result is greater than a preset scoring threshold value in the sorted items;
determining the item as the target recommended item.
8. A cross-domain recommendation system, comprising:
the construction module is used for constructing a relation graph between users and items based on historical behavior data of the users in the user set and the items in the item set, and constructing a knowledge graph between the items in the item set based on item attribute information;
the fusion module is used for fusing the knowledge graph between the projects and the relation graph between the users and the projects to obtain a cross-domain collaborative knowledge graph;
the determining module is used for determining user characteristic vectors and project characteristic vectors between domains and determining the user characteristic vectors and the project characteristic vectors in the domains based on an embedded attention mechanism and the cross-domain collaborative knowledge map;
and the target recommended item prediction module is used for determining a target recommended item based on the user characteristic vector and the item characteristic vector between the domains and the user characteristic vector and the item characteristic vector in the domains.
9. A terminal device, characterized in that the terminal device comprises: a memory, a processor, and a cross-domain recommender stored on the memory and operable on the processor, the cross-domain recommender when executed by the processor implementing the steps of the cross-domain recommender method as claimed in any of claims 1 to 7.
10. A computer-readable storage medium, having stored thereon a cross-domain recommendation program which, when executed by a processor, implements the steps of the cross-domain recommendation method of any of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210779167.7A CN115809339A (en) | 2022-07-04 | 2022-07-04 | Cross-domain recommendation method, system, device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210779167.7A CN115809339A (en) | 2022-07-04 | 2022-07-04 | Cross-domain recommendation method, system, device and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115809339A true CN115809339A (en) | 2023-03-17 |
Family
ID=85482253
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210779167.7A Pending CN115809339A (en) | 2022-07-04 | 2022-07-04 | Cross-domain recommendation method, system, device and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115809339A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117033948A (en) * | 2023-10-08 | 2023-11-10 | 江西财经大学 | Project recommendation method based on feature interaction information and time tensor decomposition |
-
2022
- 2022-07-04 CN CN202210779167.7A patent/CN115809339A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117033948A (en) * | 2023-10-08 | 2023-11-10 | 江西财经大学 | Project recommendation method based on feature interaction information and time tensor decomposition |
| CN117033948B (en) * | 2023-10-08 | 2024-01-09 | 江西财经大学 | Project recommendation method based on feature interaction information and time tensor decomposition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7104244B2 (en) | User tag generation method and its devices, computer programs and computer equipment | |
| CN111931062B (en) | Training method and related device of information recommendation model | |
| CN111680219B (en) | Content recommendation method, device, equipment and readable storage medium | |
| CN109190030B (en) | Implicit feedback recommendation method fusing node2vec and deep neural network | |
| CN112732936A (en) | Radio and television program recommendation method based on knowledge graph and user microscopic behaviors | |
| CN111241394B (en) | Data processing method, data processing device, computer readable storage medium and electronic equipment | |
| CN110619081A (en) | News pushing method based on interactive graph neural network | |
| CN114202061A (en) | Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning | |
| CN111611488A (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
| CN112989169B (en) | Target object identification method, information recommendation method, device, equipment and medium | |
| CN111143684A (en) | Artificial intelligence-based generalized model training method and device | |
| CN111949886B (en) | Sample data generation method and related device for information recommendation | |
| CN112699310A (en) | Cold start cross-domain hybrid recommendation method and system based on deep neural network | |
| CN114692007B (en) | Method, device, equipment and storage medium for determining representation information | |
| CN111159570A (en) | Information recommendation method and server | |
| CN112464100A (en) | Information recommendation model training method, information recommendation method, device and equipment | |
| CN113051468B (en) | Movie recommendation method and system based on knowledge graph and reinforcement learning | |
| Luo et al. | HySAGE: A hybrid static and adaptive graph embedding network for context-drifting recommendations | |
| CN115809339A (en) | Cross-domain recommendation method, system, device and storage medium | |
| Zhang et al. | Knowledge graph driven recommendation model of graph neural network | |
| CN112269943B (en) | Information recommendation system and method | |
| CN114817692A (en) | Method, device and equipment for determining recommended object and computer storage medium | |
| CN117251586A (en) | Multimedia resource recommendation method, device and storage medium | |
| CN116956183A (en) | Multimedia resource recommendation method, model training method, device and storage medium | |
| CN113535939A (en) | Text processing method and device, electronic equipment and computer readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |