CN117216281A - Knowledge graph-based user interest diffusion recommendation method and system - Google Patents
Knowledge graph-based user interest diffusion recommendation method and system Download PDFInfo
- Publication number
- CN117216281A CN117216281A CN202210601284.4A CN202210601284A CN117216281A CN 117216281 A CN117216281 A CN 117216281A CN 202210601284 A CN202210601284 A CN 202210601284A CN 117216281 A CN117216281 A CN 117216281A
- Authority
- CN
- China
- Prior art keywords
- user
- item
- entity
- knowledge graph
- project
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000009792 diffusion process Methods 0.000 title claims abstract description 19
- 230000003993 interaction Effects 0.000 claims abstract description 52
- 239000013598 vector Substances 0.000 claims abstract description 33
- 238000013528 artificial neural network Methods 0.000 claims abstract description 15
- 230000000306 recurrent effect Effects 0.000 claims abstract description 13
- 230000005540 biological transmission Effects 0.000 claims abstract description 9
- 230000004931 aggregating effect Effects 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 16
- 230000006399 behavior Effects 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 238000012360 testing method Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 4
- 238000004590 computer program Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a knowledge graph-based user interest diffusion recommendation method and system, wherein an improved recurrent neural network is used for extracting local user interest features from a user history interaction sequence, and extracting project local potential features from the collaborative neighborhood of candidate projects; then, the historical interaction sequence of the user and the collaboration neighbors of the project are outwards diffused into the multi-layer triples on the knowledge graph in an iterative mode, the original interests of the user are utilized to control the transmission of the information of each layer of triples, and the global interest characteristics of the user and the global potential characteristics of the project are adaptively extracted; and finally, aggregating local features and global features in a splicing mode to obtain embedded vectors of the user and the project, and performing inner product operation to predict the interaction probability of the user on the project. The invention can efficiently control the transmission of entity information by utilizing the interaction information of the user and the project, thereby learning more accurate user interest characteristics and obtaining better recommending effect.
Description
Technical Field
The invention relates to the technical field of recommendation systems, in particular to a knowledge graph-based user interest diffusion recommendation method and system.
Background
With the rapid development of mobile internet application technology, various activities of people are closely related to internet application products, the user scale and the project scale of each internet platform are exploded, the data information is far beyond the human brain information processing range, and people are difficult to rapidly select interesting contents, so that the problem of information overload is generated. In order to avoid that users spend a great deal of time searching for favorite contents, a recommendation system is an object of important research in the current academia and industry.
The recommendation system adopts data mining and machine learning technology to find the interests and intentions of the user from the multi-source data such as the demands, interests, historical behaviors and the like of the user, and finally generates a personalized item ordering list for the user. The collaborative filtering algorithm is the most mainstream recommendation method applied in the industry, and only utilizes the historical feedback data of the user and the project to mine the relevance between the user and the project, so as to recommend the content which is most possibly interested for the user. The algorithm does not need to additionally acquire attribute information of the user and the project, has strong expandability and is very easy to realize in engineering, but has the problem of sparsity of interaction data of the user and the project. In order to alleviate the problems, researchers fuse the knowledge graph as auxiliary information into a recommendation system, so that external attribute characteristics of users and projects can be enhanced, further deeper and finer-granularity connection between the users and the projects can be found, and recommendation accuracy is improved.
The existing recommendation method based on the knowledge graph mainly focuses on how to effectively encode external entity information into recommended users or projects, and neglects the importance of interaction between the users and the projects, so that the learned embedded representation cannot effectively express the interests of the users.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a knowledge graph-based user interest diffusion recommendation method and system. On one hand, the local interest features and the project local potential features of the user are extracted from the historical interaction sequence of the user through the recurrent neural network, and on the other hand, when the entity information of different distances is expanded on the knowledge graph through the project to expand the features of the user and the project, the original interests of the user are utilized to control the transmission of each layer of triplet information, the user preference is learned more accurately from the global angle, the problem of data sparsity is finally relieved, and the recommendation performance is improved.
In order to achieve the above purpose, the present invention provides a knowledge graph-based user interest diffusion recommendation method and system, comprising the following steps:
s1, extracting characteristics of a local interest layer: extracting local interest features of a user from a user history interaction sequence by using an improved recurrent neural network according to a user-item interaction matrix, and extracting item local potential features from the collaborative neighborhood of candidate items;
s2, extracting features of a global interest layer: the method comprises the steps that a historical interaction sequence of a user and a collaboration neighbor of a project are subjected to out-diffusion of a plurality of layers of triples on a knowledge graph in an iterative mode, the original interests of the user are utilized to control the transmission of information of each layer of triples, and the global interest characteristics of the user and the global potential characteristics of the project are adaptively extracted;
s3, model prediction: and aggregating the characteristics of the local interest layer and the global interest layer by adopting a splicing mode, and performing inner product operation on the obtained user embedded vector and the project embedded vector so as to predict the interaction probability of the user on the project and finish the recommendation task.
The invention provides a knowledge graph-based user interest diffusion recommendation method. Compared with the prior art, the method has the following beneficial effects:
according to the invention, the interests of the user are learned from two angles, on one hand, in the recommendation system, for a certain user, the unique interests of the user can be reflected to a certain extent by the items interacted with the user in the past, and the truest preference characteristics of the user can be extracted from the historical interaction sequence of the user by using the recurrent neural network. On the other hand, in the knowledge graph, the entity associated with the item can provide rich external attribute information for the item, the item interacted by the user is diffused on the knowledge graph, more accurate user preference characteristics can be learned from each layer of triad diffused, and finally, a better recommendation effect can be obtained.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a knowledge-graph-based user interest diffusion recommendation method in an embodiment of the invention.
Fig. 2 is an overall framework diagram of a knowledge-graph-based user interest diffusion recommendation method in an embodiment of the invention.
FIG. 3 is a schematic diagram of the RNN structure according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
The invention will be described in detail with reference to the drawings and the detailed description. As shown in the flowchart of fig. 1, a knowledge graph-based user interest diffusion recommendation method includes steps S1 to S3:
s1, extracting characteristics of a local interest layer: extracting local interest features of a user from a user history interaction sequence by using an improved recurrent neural network according to a user-item interaction matrix, and extracting item local potential features from the collaborative neighborhood of candidate items;
s2, extracting features of a global interest layer: the method comprises the steps that a historical interaction sequence of a user and a collaboration neighbor of a project are subjected to out-diffusion of a plurality of layers of triples on a knowledge graph in an iterative mode, the original interests of the user are utilized to control the transmission of information of each layer of triples, and the global interest characteristics of the user and the global potential characteristics of the project are adaptively extracted;
s3, model prediction: and aggregating the characteristics of the local interest layer and the global interest layer by adopting a splicing mode, and performing inner product operation on the obtained user embedded vector and the project embedded vector so as to predict the interaction probability of the user on the project and finish the recommendation task.
The respective steps are described in detail below.
In step S1, as shown in the overall frame diagram of fig. 2, features of the local interest layer are extracted: according to the user-item interaction matrix, the improved recurrent neural network is used for extracting the local interest characteristics of the user from the user history interaction sequence, and simultaneously extracting the local potential characteristics of the items from the collaboration neighborhood of the candidate items. The method comprises the following steps:
s101, introducing a recommended data set which comprises M user numbers and N project numbers, wherein the user set is defined as follows: u= { U 1 ,u 2 ,…,u M The item set is defined as: v= { V 1 ,v 2 ,…v N }。
S102, defining a user-project interaction matrix Y E R according to implicit feedback information of the user M×N Expressed by the expression:
wherein U is U and V is V.
S103, storing different types of entities and relations in a triplet form by the knowledge graph G, wherein the knowledge graph G is defined as follows:
G＝{(h,r,t)|h,t∈E,r∈R} (2)
wherein, (h, R, t) represents the relationship between the head entity h and the tail entity t described by the relationship R, and E and R represent the entity set and the relationship set in the knowledge graph respectively.
S104, in order to link the recommendation data with the knowledge graph, a matching function is defined, and for each item v in the user-item interaction matrix Y, a corresponding entity e can be found in the knowledge graph G, which is expressed as follows:
A＝{(v,e)|v∈V,e∈E} (3)
s105, matching the historical interaction items of the user with the entities in the knowledge graph to obtain an initial entity set of the user u, namely a historical interaction item set of the user, wherein the initial entity set is expressed as follows:
s106, the RNN structure diagram shown in fig. 3, for a certain user, the content of interest of him is often affected by his past interaction behavior, so that very real preference information can be captured from his history interaction sequence. Meanwhile, the user may carry partial information, the original embedding of the user is input into the recurrent neural network structure in the form of offset vector, and the local interest vector u of the user u is output 0 ：
Wherein,is the user initial entity set->Where the entity represents the user's historical interaction terms, U represents the original embedding of user U, and U and W represent weight parameters shared in the neural network. When i=0, _f->In recurrent neural networks, the last output data contains all the previous input information, here chosen +.>As a local interest vector for the user.
S107, for a certain item, interaction behaviors possibly exist with a plurality of users, which means that the users have similar behavior preference, and partial attribute features can be provided for the item, so that the dependency relationship between the item and more users can be found when the deeper feature information is learned. Treating items interacted with by the same user as collaborative neighbors, e.g. user u and item v 1 、v 2 、v 3 There is interactive behavior, then item v 2 、v 3 Can be regarded as item v 1 Is a cooperative neighbor of (a). First, a user set u interacted with a candidate item v is obtained v Then obtain the collaborative neighbor set V of the candidate item V v It is represented as follows:
u v ＝{u|y uv ＝1} (8)
where u represents the user set u v V of the group u Representing the item interacted with by the user u, so there is
S108, each cooperative neighbor has very strong relevance with the candidate item v, and potential interests of a user group can be reflected. Meanwhile, the original embedded representation v of the candidate item v can also reflect the intrinsic characteristics of the item. In order to obtain a local potential feature vector of a candidate item v, firstly matching an item in a collaborative neighbor set with an entity in a knowledge graph to obtain an initial entity set of the item v, wherein the initial entity set is expressed as follows:
thenPair aggregationThe embedded vectors of all the entities in the candidate item and the original embedded v of the candidate item are added equally to obtain the local feature vector of the candidate item, which is expressed as follows:
in step S2, features of the global interest layer are extracted: and (3) outwards diffusing the multi-layer triples on the knowledge graph in an iterative mode by using the historical interaction sequence of the user and the collaboration neighbors of the project, controlling the transmission of the information of each layer of triples by using the original interests of the user, and adaptively extracting the global interest characteristics of the user and the global potential characteristics of the project. The method comprises the following steps:
s201, the historical interaction sequence of the user can directly reflect preference information of the user, but careful user preference is ignored, for example, the user is particularly interested in which attribute of the item is caused, and the knowledge graph can provide rich external attribute information for the item. The global interest layer is to diffuse the original interests of the user on the knowledge graph, and learn more accurate user preference characteristics from each layer of triads diffused. And taking the entities contained in the initial entity set of the user and the candidate item as head entities, and finding out tail entities associated with the head entities according to the triples (h, r, t), namely first-layer neighbor entities. Similarly, a first tier neighbor entity is associated as a header entity to a second tier neighbor entity. The k-layer neighbor entity set defining user u or item v herein is:
where the subscript o represents the user u or the candidate item v and k represents the distance of the neighbor entity from the initial entity. The k-layer triplet set of user u and item v is constructed in a recursive manner based on the neighbor entity set, which is expressed as follows:
s202, each triplet (h, r, t) contains special semantic information, and can express finer user interests. When the triad information of different layers is transferred on the knowledge graph, the user's own behavior preference needs to be held. Therefore, the collaboration information of the user and the history item is encoded into the triples, and the influence of different triples of different layers in expressing user preferences is obtained by analyzing the association degree of the head entity of the triples and the user interests, so that the global interest feature of the user and the global potential feature of the item are finally obtained. Assuming (h, r, t) is the ith triplet in the kth level triplet set, learning a tail entity perceives the embedded attention coefficient p (·) to aggregate the information passed by each triplet.
Wherein z is 0 Representing collaboration signal x 0 Association with head entity, x when learning user global interest feature 0 ＝v 0 X when learning global features of items 0 ＝u 0 ，Representation header entity embeds representation, r i Representation relation embedding representation->The representation tail entity embeds the representation, the representation is a concatenation operation, W and b are trainable weight parameters and bias vectors, σ (·) represents a nonlinear activation function, here a relu function is employed.
S203, normalizing the attention coefficients of all triples in the k layer by adopting a softmax function, wherein the normalization operation is expressed as follows:
wherein,a k-layer triplet set representing user u or item v.
S204, multiplying the attention coefficient with the tail entity to obtain a new tail entity embedded representation, which is represented as follows:
and S205, the attention coefficient can give different attention according to the association degree of the initial interest of the user and different triples, so that tail entity information in the triples is transferred more effectively, and finally the embedded vector of the k-layer triplet set is obtained.
In step S3, model prediction: and aggregating the characteristics of the local interest layer and the global interest layer by adopting a splicing mode, and performing inner product operation on the obtained user embedded vector and the project embedded vector so as to predict the interaction probability of the user on the project and finish the recommendation task. The method comprises the following steps:
s301, after diffusing the K-layer triplet information, a plurality of embedded representations of the user u and the item v can be obtained, and an embedded set of the user u and the item v is defined:
wherein,
s302, each layer of embedded representation may be interpreted as a potential impact of a set of triples of different distances on a user or item, with the outputs of the different layers emphasizing connectivity and preference information in different orders. To achieve better recommended accuracy, the embedded representations of the different layers are connected into a single vector using a stitching operation, which is expressed as follows:
wherein,this can be done not only by embedding the propagation operations to obtain more global characteristic information of users and items, but also by adjusting K to control propagation strength
S303, finally, carrying out inner product operation on the user embedded vector and the project embedded vector, predicting the probability of interaction of the user on the project, and controlling the probability value between 0 and 1 by using a nonlinear activation function sigmoid, wherein the probability value is expressed as follows:
s304, in order to optimize the recommendation model, better learn user embedded representations and project embedded representations, taking the projects interacted (clicked, evaluated, purchased and the like) by each user as positive samples of training, taking other non-interacted projects as negative samples of training, collecting the same number of positive samples and negative samples in a training set, and designing a loss function as follows:
for Loss function Loss, we train with cross entropy Loss function J (·), pos is the positive set of samples, neg is the negative set of samples. Wherein y is uv Indicating whether the user clicks on the true value of the item,representing the predicted value of the model output, θ contains all the trainable parameters in the model, λ is the hyper-parameter that adjusts the L2 regularization strength to prevent the model from overfitting. In the model training process, an Adam optimizer is used for adjusting the learning rate and updating and adjusting model parameters. In the click rate prediction task, interactions of users on each item in the test set are predicted by using a trained model, and a probability value of predicted click is output. When probability value->When the item is recommended to the user, the recommendation is not made otherwise.
According to the embodiment of the invention, knowledge graph information is added on a traditional collaborative filtering algorithm, firstly, a recurrent neural network is used for mining original interest features and project intrinsic features of a user from a historical interaction sequence of the user, then, an attention coefficient embedded by a tail entity is designed, the original interest of the user controls the transmission of each layer of triplet information, the global interest features of the user and the global potential features of the project are adaptively extracted, the problem of data sparsity in a recommendation system is effectively solved, the recommendation accuracy is improved, and a more efficient framework is provided for industrial recommendation and other works in practical application.
The invention also provides a knowledge graph-based user interest diffusion recommendation system, which comprises:
feature extraction of a local interest layer: extracting local interest features of a user from a user history interaction sequence by using an improved recurrent neural network according to a user-item interaction matrix, and extracting item local potential features from the collaborative neighborhood of candidate items;
feature extraction of a global interest layer: the method comprises the steps that a historical interaction sequence of a user and a collaboration neighbor of a project are subjected to out-diffusion of a plurality of layers of triples on a knowledge graph in an iterative mode, the original interests of the user are utilized to control the transmission of information of each layer of triples, and the global interest characteristics of the user and the global potential characteristics of the project are adaptively extracted;
model prediction: and aggregating the characteristics of the local interest layer and the global interest layer by adopting a splicing mode, and performing inner product operation on the obtained user embedded vector and the project embedded vector so as to predict the interaction probability of the user on the project and finish the recommendation task.
The invention also proposes a computer device comprising a memory storing a computer program and a processor implementing the steps of the method described above when executing the computer program.
The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the invention, and all equivalent structural changes made by the description of the present invention and the accompanying drawings or direct/indirect application in other related technical fields are included in the scope of the invention.
Claims (4)
1. A knowledge graph-based user interest diffusion recommendation method, characterized in that the method is executed by a computer and comprises the following steps:
s1, extracting characteristics of a local interest layer: extracting local interest features of a user from a user history interaction sequence by using an improved recurrent neural network according to a user-item interaction matrix, and extracting item local potential features from the collaborative neighborhood of candidate items;
s2, extracting features of a global interest layer: the method comprises the steps that a historical interaction sequence of a user and a collaboration neighbor of a project are subjected to out-diffusion of a plurality of layers of triples on a knowledge graph in an iterative mode, the original interests of the user are utilized to control the transmission of information of each layer of triples, and the global interest characteristics of the user and the global potential characteristics of the project are adaptively extracted;
s3, model prediction: and aggregating the characteristics of the local interest layer and the global interest layer by adopting a splicing mode, and performing inner product operation on the obtained user embedded vector and the project embedded vector so as to predict the interaction probability of the user on the project and finish the recommendation task.
2. The knowledge graph-based user interest diffusion recommendation method as claimed in claim 1, wherein the specific implementation process of S1 is as follows:
s101, introducing a recommended data set which comprises M user numbers and N project numbers, wherein the user set is defined as follows: u= { U 1 ,u 2 ,…,u M The item set is defined as: v= { V 1 ,v 2 ,…v N }。
S102, defining a user-project interaction matrix Y E R according to implicit feedback information of the user M×N Expressed by the expression:
wherein U is U and V is V.
S103, storing different types of entities and relations in a triplet form by the knowledge graph G, wherein the knowledge graph G is defined as follows:
G＝{(h,r,t)|h,t∈E,r∈R} (2)
wherein, (h, R, t) represents the relationship between the head entity h and the tail entity t described by the relationship R, and E and R represent the entity set and the relationship set in the knowledge graph respectively.
S104, in order to link the recommendation data with the knowledge graph, a matching function is defined, and for each item v in the user-item interaction matrix Y, a corresponding entity e can be found in the knowledge graph G, which is expressed as follows:
A＝{(v,e)|v∈V,e∈E} (3)
s105, matching the historical interaction sequence of the user with the entities in the knowledge graph to obtain an initial entity set of the user u, namely a historical interaction item set of the user, wherein the initial entity set is expressed as follows:
s106, for a certain user, the content that he is interested in now is often influenced by his past interaction behavior, so that very real preference information can be captured from his historical interaction sequence. Meanwhile, the user may carry partial information, the original embedding of the user is input into the recurrent neural network structure in the form of offset vector, and the local interest vector u of the user is output 0 ：
Wherein,is the user initial entity set->Where the entity represents the user's historical interaction item, U represents the user's U original embedding, U and W represent weight parameters shared in the neural network, and +_ when i=0>In recurrent neural networks, the last output data contains all of the previous input information, selected hereAs a local interest vector for the user.
S107, for a certain item, interaction behaviors possibly exist with a plurality of users, which means that the users have similar behavior preference, and partial attribute features can be provided for the item, so that the dependency relationship between the item and more users can be found when the deeper feature information is learned. Items interacted with by the same user are considered to be co-neighbors, e.g. user u has an interaction with items v1, v2, v3, then items v2, v3 can be considered to be co-neighbors of item v 1. First, a user set u interacted with a candidate item v is obtained v A set of co-neighbors of candidate item v is then obtained, which is represented as follows:
u v ＝{u|y uv ＝1} (8)
where u represents the user set u v V of the group u Representing the item interacted with by the user u, so there is
S108, each cooperative neighbor has very strong relevance with the item v, and can reflect the potential interests of a user group. Meanwhile, the original embedded representation v of the candidate item v can also reflect the intrinsic characteristics of the item. In order to obtain a local potential feature vector of a candidate item v, firstly matching an item in a collaborative neighbor set with an entity in a knowledge graph to obtain an initial entity set of the item v, wherein the initial entity set is expressed as follows:
then to the collectionThe embedded vectors of all the entities in the candidate item and the original embedded v of the candidate item are added equally to obtain the local feature vector of the candidate item, which is expressed as follows:
。
3. the knowledge graph-based user interest diffusion recommendation method as claimed in claim 1, wherein the specific implementation process of S2 is as follows:
s201, the historical interaction sequence of the user can directly reflect preference information of the user, but careful user preference is ignored, for example, the user is particularly interested in which attribute of the item is caused, and the knowledge graph can provide rich external attribute information for the item. The global interest layer is to diffuse the original interests of the user on the knowledge graph, and learn more accurate user preference characteristics from each layer of triads diffused. And taking the entities contained in the initial entity set of the user and the candidate item as head entities, and finding out tail entities associated with the head entities according to the triples (h, r, t), namely first-layer neighbor entities. Similarly, a first tier neighbor entity is associated as a header entity to a second tier neighbor entity. The k-layer neighbor entity set defining user u or item v herein is:
where the subscript o represents the user u or the candidate item v and k represents the distance of the neighbor entity from the initial entity. The k-layer triplet set of user u and item v is constructed in a recursive manner based on the neighbor entity set, which is expressed as follows:
s202, each triplet (h, r, t) contains special semantic information, and can express finer user interests. When the triad information of different layers is transferred on the knowledge graph, the user's own behavior preference needs to be held. Therefore, the collaboration information of the user and the history item is encoded into the triples, and the influence of different triples of different layers in expressing user preferences is obtained by analyzing the association degree of the head entity of the triples and the user interests, so that the global interest feature of the user and the global potential feature of the item are finally obtained. Assuming (h, r, t) is the ith triplet in the kth level triplet set, learning a tail entity perceives the embedded attention coefficient p (·) to aggregate the information passed by each triplet.
Wherein z is 0 Representing collaboration signal x 0 Association with head entity, x when learning user global interest feature 0 ＝v 0 X when learning global features of items 0 ＝u 0 ，Representing an embedded representation of a header entity, r i Representation relation embedding representation->The representation tail entity embeds the representation. The I represents a concatenation operation, WAnd b is a trainable weight parameter and bias vector, σ (·) represents a nonlinear activation function, here a relu function is employed.
S203, normalizing the attention coefficients of all triples in the k layer by adopting a softmax function, wherein the normalization operation is expressed as follows:
wherein,is the k-level triplet set for user u or item v.
S204, multiplying the attention coefficient with the tail entity to obtain a new tail entity embedded representation, which is represented as follows:
and S205, the attention coefficient can give different attention according to the association degree of the initial interest of the user and different triples, so that tail entity information in the triples is transferred more effectively, and finally the embedded vector of the k-layer triplet set is obtained.
。
4. The knowledge graph-based user interest diffusion recommendation method as claimed in claim 1, wherein the specific implementation process of S3 is as follows:
s301, after diffusing the K-layer triplet information, a plurality of embedded representations of the user u and the item v can be obtained, and an embedded set of the user u and the item v is defined:
wherein,
s302, each layer of embedded representation may be interpreted as a potential impact of a set of triples of different distances on a user or item, with the outputs of the different layers emphasizing connectivity and preference information in different orders. To achieve better recommended accuracy, the embedded representations of the different layers are connected into a single vector using a stitching operation, which is represented as follows.
Wherein,this can be done not only by embedding the propagation operations to obtain more global characteristic information of users and items, but also by adjusting K to control propagation strength
S303, finally, carrying out inner product operation on the user embedded vector and the project embedded vector, predicting the probability of interaction of the user on the project, and controlling the probability value between 0 and 1 by using a nonlinear activation function sigmoid, wherein the probability value is expressed as follows:
s304, in order to optimize the recommendation model, better learn user embedded representations and project embedded representations, taking the projects interacted (clicked, evaluated, purchased and the like) by each user as positive samples of training, taking other non-interacted projects as negative samples of training, collecting the same number of positive samples and negative samples in a training set, and designing a loss function as follows:
for Loss function Loss, we train with cross entropy Loss function J (·), pos is the positive set of samples, neg is the negative set of samples. Wherein y is uv The true value is represented by a value that is true,representing the predicted values, θ contains all the trainable parameters in the model, λ is the hyper-parameter that adjusts the L2 regularization strength to prevent the model from overfitting. In the model training process, an Adam optimizer is used for adjusting the learning rate and updating and adjusting model parameters. In the click rate prediction task, interactions of users on each item in the test set are predicted by using a trained model, and a probability value of predicted click is output. When probability value->When the item is recommended to the user, the recommendation is not made otherwise.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210601284.4A CN117216281A (en) | 2022-05-30 | 2022-05-30 | Knowledge graph-based user interest diffusion recommendation method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210601284.4A CN117216281A (en) | 2022-05-30 | 2022-05-30 | Knowledge graph-based user interest diffusion recommendation method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117216281A true CN117216281A (en) | 2023-12-12 |
Family
ID=89033930
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210601284.4A Pending CN117216281A (en) | 2022-05-30 | 2022-05-30 | Knowledge graph-based user interest diffusion recommendation method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117216281A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117520406A (en) * | 2023-12-29 | 2024-02-06 | 安徽思高智能科技有限公司 | RPA flow recommendation method, device and storage medium |
| CN117555425A (en) * | 2024-01-10 | 2024-02-13 | 东莞莱姆森科技建材有限公司 | Intelligent mirror control method and system for information display |
-
2022
- 2022-05-30 CN CN202210601284.4A patent/CN117216281A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117520406A (en) * | 2023-12-29 | 2024-02-06 | 安徽思高智能科技有限公司 | RPA flow recommendation method, device and storage medium |
| CN117520406B (en) * | 2023-12-29 | 2024-03-26 | 安徽思高智能科技有限公司 | RPA flow recommendation method, device and storage medium |
| CN117555425A (en) * | 2024-01-10 | 2024-02-13 | 东莞莱姆森科技建材有限公司 | Intelligent mirror control method and system for information display |
| CN117555425B (en) * | 2024-01-10 | 2024-04-30 | 东莞莱姆森科技建材有限公司 | Intelligent mirror control method and system for information display |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111061856B (en) | Knowledge perception-based news recommendation method | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
| CN111966914B (en) | Content recommendation method and device based on artificial intelligence and computer equipment | |
| Li et al. | Image sentiment prediction based on textual descriptions with adjective noun pairs | |
| CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system | |
| CN114780831A (en) | Sequence recommendation method and system based on Transformer | |
| CN113051468B (en) | Movie recommendation method and system based on knowledge graph and reinforcement learning | |
| CN113343100B (en) | Smart city resource recommendation method and system based on knowledge graph | |
| CN111949885A (en) | Personalized recommendation method for scenic spots | |
| CN112085525A (en) | User network purchasing behavior prediction research method based on hybrid model | |
| CN112257841A (en) | Data processing method, device and equipment in graph neural network and storage medium | |
| CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
| CN114298851A (en) | Network user social behavior analysis method and device based on graph sign learning and storage medium | |
| Huang et al. | On the improvement of reinforcement active learning with the involvement of cross entropy to address one-shot learning problem | |
| Zhang et al. | Multi-scale and multi-channel neural network for click-through rate prediction | |
| Liu et al. | TCD-CF: Triple cross-domain collaborative filtering recommendation | |
| Zhang et al. | Hybrid recommendation system with graph neural collaborative filtering and local self-attention mechanism | |
| Wang et al. | Color trend prediction method based on genetic algorithm and extreme learning machine | |
| CN116452293A (en) | Deep learning recommendation method and system integrating audience characteristics of articles | |
| CN116401450A (en) | Recommendation method and device based on knowledge graph and computer equipment | |
| Wang et al. | CHSR: Cross-view Learning from Heterogeneous Graph for Session-Based Recommendation | |
| Bi et al. | An approach for item recommendation using deep neural network combined with the Bayesian personalized ranking | |
| Hmaidi et al. | Anime Link Prediction Using Improved Graph Convolutional Networks | |
| Wang et al. | Feature Fusion Based Parallel Graph Convolutional Neural Network for Image Annotation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication |