CN114117142A - Label perception recommendation method based on attention mechanism and hypergraph convolution - Google Patents

Label perception recommendation method based on attention mechanism and hypergraph convolution Download PDF

Info

Publication number
CN114117142A
CN114117142A CN202111461962.3A CN202111461962A CN114117142A CN 114117142 A CN114117142 A CN 114117142A CN 202111461962 A CN202111461962 A CN 202111461962A CN 114117142 A CN114117142 A CN 114117142A
Authority
CN
China
Prior art keywords
user
hypergraph
item
label
tag
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
Application number
CN202111461962.3A
Other languages
Chinese (zh)
Inventor
王海艳
尤恺翔
骆健
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202111461962.3A priority Critical patent/CN114117142A/en
Publication of CN114117142A publication Critical patent/CN114117142A/en
Priority to PCT/CN2022/106913 priority patent/WO2023098098A1/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/901—Indexing; Data structures therefor; Storage structures
    • G06F16/9024—Graphs; Linked lists
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/906—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a label perception recommendation method based on attention mechanism and hypergraph convolution. The invention introduces hypergraph convolution to excavate high-order relations for feature extraction, adopts an attention mechanism to carry out weight distribution on features obtained by a hypergraph constructed by user-item direct interaction and label perception, can better distinguish information with different importance degrees, and can fully extract features in user-item direct interaction relations and label interaction relations by skillfully combining the hypergraph convolution and the attention mechanism, thereby effectively improving the performance of a recommendation method.

Description

Label perception recommendation method based on attention mechanism and hypergraph convolution
Technical Field
The invention belongs to the field of information service and computer software technology application, and particularly relates to a label perception recommendation method based on attention mechanism and hypergraph convolution.
Background
With the rapid increase of the amount of various information resources in the network, how to recommend resources or commodities meeting the requirements from huge data volume to users is more and more concerned by the industry and academia, so that the demand service provider needs to have a proper recommendation method. To improve recommendation accuracy, many internet service providers employ User Generated Content (UGC) tagging systems that allow users to actively tag goods, videos, information, etc. The traditional recommendation method such as collaborative filtering is difficult to embody the complex, various and high-order interaction relation between the user and the project, so the effect is not good. With the introduction of deep learning methods, the recommendation method based on the graph neural network draws attention because it can reflect the topology. However, the common graph structure only reflects the pairwise relationship between nodes through edges, but cannot reflect the relationship between three or more nodes.
Disclosure of Invention
In order to solve the problem that complex high-order relationships among users, items and labels in label perception recommendation cannot be embodied by using a traditional graph neural network, the invention introduces a hypergraph to model the high-order relationships among nodes, completes information transmission in the neural network by utilizing hypergraph convolution, and reflects the importance degree of different information by using an attention mechanism, thereby effectively relieving the problem of high-order information loss in the traditional recommendation method.
In order to achieve the purpose, the invention is realized by the following technical scheme:
the invention relates to a label perception recommendation method based on attention mechanism and hypergraph convolution, which constructs different hypergraphs on a user side and an item side respectively through a direct interaction relation between a user and an item and a brief introduction interaction relation between the user, the item and a label, extracts information of high-order relation reaction through hypergraph convolution, distinguishes information of different importance degrees by adopting the attention mechanism, and recommends through obtained characteristic representation, and specifically comprises the following steps:
step 1: initializing the feature representation of a user and a project to obtain u and i;
step 2: obtaining user characteristics u represented by the labels according to the interactive relations between the user and the items and the labels respectivelytagAnd item characteristics itag；
And step 3: constructing three interactive bipartite graphs G according to the interactive relations between users and projects, between users and labels and between projects and labelsuser-item、Guser-tag、Gitem-tag；
And 4, step 4: respectively constructing hypergraph structures of a user side and a project side represented by direct interaction relation according to the bipartite graph of the user and the project obtained in the step 3
Figure BDA0003388114130000021
And 5: respectively constructing hypergraph structures of a user side and a project side represented by label relations according to the bipartite graph of the user and the label and the project and the label obtained in the step 3
Figure BDA0003388114130000022
Step 6: using the two hypergraphs obtained in the step 4 for the characteristics u and i obtained in the step 1
Figure BDA0003388114130000023
Figure BDA0003388114130000024
Respectively carrying out hypergraph convolution to obtain neighborhood feature representation u of direct relationship representation after updating of a user side and an item side1、i1；
And 7: for the characteristics u and i obtained in the step 1, the two hypergraphs from the step 5
Figure BDA0003388114130000025
Respectively carrying out hypergraph convolution to obtain neighborhood feature representation u represented by updated label relationship at user side and item side2、i2；
And 8: processing the features obtained in the steps 2, 6 and 7 by using an attention mechanism to obtain weights of different feature representations to obtain a final user and item feature representation u*、i*；
And step 9: according toAnd the eighth step of splicing the user and the item feature representation to obtain z ═ u ═*；i*]And inputting the data into a full-connection layer, obtaining a prediction probability by using a Sigmoid function, and recommending according to the score.
The invention is further improved in that: in step 2, according to the number of times the user marks and the number of times the item is marked, the label representation characteristics of the user and the item are initialized, and normalization processing is performed.
The invention is further improved in that: in step 4: and respectively constructing a hypergraph of the user side and a hypergraph of the project side represented by the direct interaction relationship according to the user and project bipartite graph. Taking the user side as an example: if a path exists between the two items m and n and the number of users passing through the path is less than k, the two items are k-order reachable neighbors; for item m, if it has k-th order reachable neighbors n and user u interacts directly with m, then user u is a k-th order reachable user for item n. For each item, the k-order reachable users are regarded as a set, the users on the set are regarded as nodes, and the set can be regarded as a hyper-edge, so that the hyper-graph is constructed
Figure BDA0003388114130000031
The project side and the user side are in the same way to construct the hypergraph
Figure BDA0003388114130000032
The invention is further improved in that: in step 5, a hypergraph of the user side and the project side represented by the label relationship is respectively constructed according to the bipartite graph of the user and the label and the project and the label. The idea is similar to the fourth step, but instead of using a user-item relationship, a user-tag and an item-tag relationship are used, respectively. Taking the user side as an example: if a path exists between the two labels m and n and the number of users passing through the path is less than k, the two labels are k-order reachable neighbors; for the label m, if there is k-order reachable neighbor n and the user u interacts with m directly, the user u is a k-order reachable user of the label n. For each label, the k-order reachable users are regarded as a set, the users on the set are regarded as nodes, and the set can be regarded as a hyper-edge, so that the hyper-graph is constructed
Figure BDA0003388114130000033
The project side and the user side are in the same way to construct the hypergraph
Figure BDA0003388114130000034
The invention has the beneficial effects that:
(1) the invention uses the characteristics of the hypergraph data structure to represent the high-order relation between users and items, and uses the hypergraph convolution to update information, compared with the graph neural network, the invention can reduce the loss of information in the information transmission process, and fully uses the high-order interactive relation to obtain the field characteristic representation.
(2) The invention combines the attention mechanism, effectively uses a plurality of characteristics to represent the characteristics of different importance degrees, and avoids the situation that high-value information is bushed by low-value information.
(3) In the label perception recommendation, the relationship among a user, an item and a label is utilized to carry out feature representation and relationship modeling, and various information is fully utilized to carry out feature representation.
(4) In result prediction, a method of splicing the characteristics of the user side and the project side and then processing the characteristics by using a full connection layer and a Sigmoid activation function is adopted, so that the influence of directly adding the characteristics to cause an information loss problem is reduced.
Drawings
FIG. 1 is a user, item, tag relationship diagram.
FIG. 2 is a plurality of hypergraph convolutions representing a hypergraph.
FIG. 3 is a multi-feature process based on an attention mechanism.
Fig. 4 is a score prediction process.
Detailed Description
In the following description, for purposes of explanation, numerous implementation details are set forth in order to provide a thorough understanding of the embodiments of the invention. It should be understood, however, that these implementation details are not to be interpreted as limiting the invention. That is, in some embodiments of the invention, such implementation details are not necessary. In addition, some conventional structures and components are shown in simplified schematic form in the drawings.
In order to solve the problem in label perception recommendation, the invention introduces the hypergraph to model the high-order relation between nodes, completes the information transmission in the neural network by utilizing the hypergraph convolution, and reflects the importance degree of different information by using an attention mechanism, thereby effectively relieving the problem of high-order information loss in the traditional recommendation method. Fig. 1 shows the interaction between users, items and tags, where there is both a direct interaction between users and items, and a user-tag and item-tag relationship. In order to extract features from complex relationships for recommendation, the present invention introduces an attention mechanism and hypergraph convolution.
2-3, the invention is a label perception recommendation method based on attention mechanism and hypergraph convolution, the method constructs different hypergraphs on user side and item side respectively through direct interactive relation of user and item and brief introduction interactive relation of user, item and label, extracts information of high order relation reaction through hypergraph convolution, and distinguishes information of different importance degrees by adopting attention mechanism, and recommends through obtained characteristic representation, specifically comprising the following steps:
step 1: a feature is initialized. Here, the user side establishes an initial feature u based on personal information such as a user ID, sex, age, and the like, and the project side establishes an initial feature i based on commodity information. In this step, no part of the label perception is involved.
Step 2: the user set U, the item set I, the tag set T, and the three-way relationship representation set a are represented by a tuple F ═ (U, I, T, a). In the user-tag feature, the number of times a user marks a tag is represented as a feature
Figure BDA0003388114130000051
Similarly, for item-tag features
Figure BDA0003388114130000052
Wherein
Figure BDA0003388114130000053
And
Figure BDA0003388114130000054
representing the number of times user u and item i are marked with label p, respectively, and σ represents the normalization operation.
And step 3: constructing three interactive bipartite graphs G according to the interactive relations between users and projects, between users and labels and between projects and labelsuser-item、Guser-tag、Gitem-tag. FIG. 1 can be viewed as a combined representation of three bipartite graphs that are built to build a hypergraph structure, where the bipartite graph of users and items uses direct user-item interactions, such as a scoring matrix.
And 4, step 4: and respectively constructing a hypergraph of the user side and a hypergraph of the project side represented by the direct interaction relationship according to the user and project bipartite graph. Taking the user side as an example: if a path exists between the two items m and n and the number of users passing through the path is less than k, the two items are k-order reachable neighbors; for item m, if it has k-th order reachable neighbors n and user u interacts directly with m, then user u is a k-th order reachable user for item n. For each item, the k-order reachable users are regarded as a set, the users on the set are regarded as nodes, and the set can be regarded as a hyper-edge, so that the hyper-graph is constructed
Figure BDA0003388114130000055
The project side and the user side are in the same way to construct the hypergraph
Figure BDA0003388114130000056
The method specifically comprises the following steps: if a path exists between the two users u and v and the number of items passing through the path is less than k, the two users are k-order reachable neighbors; for user u, if the user u has k-order reachable neighbors v and the item m directly interacts with u, the item m is a k-order reachable item of the user u, for each user, the k-order reachable item is regarded as a set, the items on the set are regarded as nodes, and the set can be regarded as a superedge, so that the supergraph is constructed
Figure BDA0003388114130000057
And 5: and respectively constructing a hypergraph of the user side and the project side represented by the label relationship according to the user and label and the project and label bipartite graph. Taking the user side as an example: if a path exists between the two labels m and n and the number of users passing through the path is less than k, the two labels are k-order reachable neighbors; for the label m, if there is k-order reachable neighbor n and the user u interacts with m directly, the user u is a k-order reachable user of the label n. For each label, the k-order reachable users are regarded as a set, the users on the set are regarded as nodes, and the set can be regarded as a hyper-edge, so that the hyper-graph is constructed
Figure BDA0003388114130000061
The project side and the user side are in the same way to construct the hypergraph
Figure BDA0003388114130000062
The method specifically comprises the following steps: if a path exists between the two labels m and n and the number of items passing through the path is less than k, the two labels are k-order reachable neighbors; for the label m, if k-order reachable neighbors n exist and the item p directly interacts with the label m, the item p is a k-order reachable item of the label n, for each label, the k-order reachable item is regarded as a set, the items on the set are regarded as nodes, and the set can be regarded as a hyper-edge, so that the hyper-graph is constructed
Figure BDA0003388114130000063
Step 6: expressing the hypergraph structure in the form of an incidence matrix as H ═ V, E, wherein V is a node set, E is a hyperedge set, and expressing whether a node V is on a hyperedge E by using the following method:
Figure BDA0003388114130000064
thus, a hypergraph is obtained
Figure BDA0003388114130000065
Is expressed by the correlation matrix ofd-user、Hd-item。
Taking the user side as an example, the hypergraph convolution can be expressed as:
Figure BDA0003388114130000066
here, a method of spectrum is used for hypergraph convolution. Wherein Θ is(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a degree matrix of the overcide. Hd-userIs a hypergraph
Figure BDA0003388114130000067
H ═ V, E, where V is the set of nodes and E is the set of super edges. And
Figure BDA0003388114130000068
the multiplication operation of (a) represents a user-side direct interaction relation representation hypergraph
Figure BDA0003388114130000069
Aggregation of the above from node features to hyper-edge features, and Hd-userIs represented by the multiplication operation of
Figure BDA00033881141300000610
From the hyper-edge feature to the node feature. Meanwhile, according to the idea of ResNet, the influence of the previous features is retained by adding the original features in each layer of the hypergraph convolution, so that the situation that the original features cannot be embodied due to overlarge influence of the neighbor features is prevented.
The project side is similar to the user side, and the hypergraph convolution can be expressed as:
Figure BDA0003388114130000071
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEMoment of excess edgeAnd (5) arraying. Hd-itemIs a hypergraph
Figure BDA0003388114130000072
H ═ V, E, where V is the set of nodes and E is the set of super edges. And
Figure BDA0003388114130000073
the multiplication operation of (2) represents the item side direct interactive relation representation hypergraph
Figure BDA0003388114130000074
Aggregation of the above from node features to hyper-edge features, and Hd-itemIs represented by the multiplication operation of
Figure BDA0003388114130000075
From the hyper-edge feature to the node feature.
Therefore, information is updated in a node-super edge-node mode on each layer of the hypergraph convolution, and information is extracted from high-order relations in a hypergraph neural network mode.
And 7: similar to step 6, a hypergraph is obtained
Figure BDA0003388114130000076
Is expressed by the correlation matrix oft-user、Ht-item。
The user-side hypergraph convolution can be expressed as:
Figure BDA0003388114130000077
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a degree matrix of the overcide.
Figure BDA0003388114130000078
Is a hypergraph
Figure BDA0003388114130000079
In a given context of the correlationAnd (V, E), wherein V is a node set and E is a super edge set. And
Figure BDA00033881141300000710
the multiplication operation of (a) represents a user-side label representation hypergraph
Figure BDA00033881141300000711
Aggregation of the above from node features to hyper-edge features, and Ht-userIs represented by the multiplication operation of
Figure BDA00033881141300000712
From the hyper-edge feature to the node feature.
The project-side hypergraph convolution can be expressed as:
Figure BDA00033881141300000713
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a degree matrix of the overcide. Ht-itemIs a hypergraph
Figure BDA0003388114130000081
H ═ V, E, where V is the set of nodes and E is the set of super edges. And
Figure BDA0003388114130000082
the multiplication operation of (2) represents an item-side label representation hypergraph
Figure BDA0003388114130000083
The aggregation of the above node features to the above super edge features, and
Figure BDA0003388114130000084
is represented by the multiplication operation of
Figure BDA0003388114130000085
From the hyper-edge feature to the node feature.
And 8: as shown in FIG. 3, the feature u obtained in step 2, step 6, and step 7tag、itag、u1、i1、u2、i2Obtaining final feature representation u on user side and project side using attention mechanism*、i*。
Take the user side as an example, since utagAnd u1、u2Are not suitable for direct addition operations due to different dimensions. Thus, they are respectively spliced to obtain u1-tagAnd u2-tag。
To obtain u1-tagAnd u2-tagThereafter, the two feature representations are processed into a final feature representation using an attention mechanism:
a(u，k)＝WTtanh(Wuk-tag+b2)
Figure BDA0003388114130000086
obtaining the final feature representation u at the user side according to the attention mechanism*
u*＝α(u，1)u1-tag+α(u，2)u2-tag
Will itagAnd i1、i2Respectively carrying out splicing operation to obtain i1-tagAnd i2-tagThen, the attention mechanism is used to process the two feature representations into one feature representation:
a(i，k)＝WTtanh(Wik-tag+b2)
Figure BDA0003388114130000087
obtaining item side final feature representation i according to attention mechanism and weight*：
i*＝α(i，1)i1-tag+α(i，2)u2-tag。
And step 9: as shown in FIG. 4, according to u*And i*Splicing to obtain z ═ u*；i*]And adopting a Sigmoid function as an activation function to obtain user-item probability prediction so as to recommend.
Figure BDA0003388114130000091
And according to the obtained probability prediction, Top-K sequencing is carried out on the items and the items are recommended to the user.
Model training was performed using the cross entropy function as a loss function:
Figure BDA0003388114130000092
where X is a set of training samples.
The invention introduces hypergraph convolution to mine high-order relations for feature extraction. Meanwhile, the attention mechanism is adopted to carry out weight distribution on the characteristics obtained by the hypergraph of the direct interaction of the user and the item and the hypergraph constructed by label perception, so that the information with different importance degrees can be better distinguished. By skillfully combining hypergraph convolution and attention mechanism, the method provided by the invention can fully extract the characteristics in the direct interaction relationship between the user and the item and the interaction relationship between the user and the label, and effectively improve the performance of the recommendation method.
The above description is only an embodiment of the present invention, and is not intended to limit the present invention. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims (8)

1. A label perception recommendation method based on attention mechanism and hypergraph convolution is characterized by comprising the following steps: the label perception recommendation method comprises the following steps of constructing different hypergraphs on a user side and an item side respectively through direct interactive relations between users and items and brief introduction interactive relations between the users, the items and labels, extracting information of high-order relation reaction through hypergraph convolution, distinguishing information with different importance degrees by adopting an attention system, and recommending through obtained feature representation, and specifically comprises the following steps:
step 1: initializing feature representation of a user u and a project i to obtain an initial feature u of the user u and an initial feature i of the project i;
step 2: obtaining user characteristics u represented by the labels according to the interactive relations between the user and the items and the labels respectivelytagAnd item characteristics itag；
And step 3: constructing three interactive bipartite graphs G according to the interactive relations between users and projects, between users and labels and between projects and labelsuser-item、Guser-tag、Gitem-tag；
And 4, step 4: respectively constructing hypergraph structures of a user side and a project side represented by direct interaction relation according to the bipartite graph of the user and the project obtained in the step 3
Figure FDA0003388114120000011
And 5: respectively constructing hypergraph structures of a user side and a project side represented by label relations according to the bipartite graph of the user and the label and the project and the label obtained in the step 3
Figure FDA0003388114120000012
Step 6: using the two hypergraphs obtained in the step 4 for the characteristics u and i obtained in the step 1
Figure FDA0003388114120000013
Figure FDA0003388114120000014
Respectively carrying out hypergraph convolution to obtain neighborhood feature representation u of direct relationship representation after updating of a user side and an item side1、i1；
And 7: for the characteristics u and i obtained in the step 1, the two hypergraphs from the step 5
Figure FDA0003388114120000015
Respectively carrying out hypergraph convolution to obtain neighborhood feature representation u represented by updated label relationship at user side and item side2、i2；
And 8: processing the features obtained in the steps 2, 6 and 7 by using an attention mechanism to obtain weights of different feature representations to obtain a final user and item feature representation u*、i*；
And step 9: and splicing the user and the item feature representation obtained in the eighth step to obtain z ═ u*；i*]And inputting the data into a full-connection layer, obtaining a prediction probability by using a Sigmoid function, and recommending according to the score.
2. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that: the hypergraph structure of the user side in the step 4
Figure FDA0003388114120000021
The construction method specifically comprises the following steps: if a path exists between the two items m and n and the number of users passing through the path is less than k, the two items are k-order reachable neighbors; for item m, if k-order reachable neighbors n exist and users u and m directly interact, the users u are k-order reachable users of the item n, the k-order reachable users of each item are regarded as a set, the users on the set are regarded as nodes, the set can be regarded as a super edge, and therefore the hypergraph is constructed
Figure FDA0003388114120000022
3. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that: the hypergraph structure of the project side in the step 4
Figure FDA0003388114120000023
The construction method specifically comprises the following steps:
if a path exists between the two users u and v and the number of items passing through the path is less than k, the two users are k-order reachable neighbors; for user u, if the user u has k-order reachable neighbors v and the item m directly interacts with u, the item m is a k-order reachable item of the user u, for each user, the k-order reachable item is regarded as a set, the items on the set are regarded as nodes, and the set can be regarded as a superedge, so that the supergraph is constructed
Figure FDA0003388114120000024
4. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that: the hypergraph structure of the user side in the step 5
Figure FDA0003388114120000025
The construction method specifically comprises the following steps: if a path exists between the two labels m and n and the number of users passing through the path is less than k, the two labels are k-order reachable neighbors; for a label m, if k-order reachable neighbors n exist and users u and m directly interact, the users u are k-order reachable users of the label n, for each label, the k-order reachable users are regarded as a set, the users on the set are regarded as nodes, and the set can be regarded as a super edge, so that a hypergraph is constructed
Figure FDA0003388114120000026
5. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that: the hypergraph structure of the project side in the step 5
Figure FDA0003388114120000031
The construction method specifically comprises the following steps: if a path exists between the two labels m and n and the number of items passing through the path is less than k, the two labels are k-order reachable neighbors; for tag m, if it has k-th orderReaching the neighbor n and the item p directly interacting with m, the item p is a k-order reachable item of the label n, for each label, the k-order reachable item is regarded as a set, the items on the set are regarded as nodes, and the set is regarded as a hyper-edge, thereby constructing the hyper-graph
Figure FDA0003388114120000032
6. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that:
the convolution of the user side hypergraph in the step 6 is represented as:
Figure FDA0003388114120000033
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a degree matrix of the overcide. Hd-userIs a hypergraph
Figure FDA0003388114120000034
H ═ V, E, where V is the set of nodes and E is the set of super edges, and
Figure FDA0003388114120000035
the multiplication operation of (a) represents a user-side direct interaction relation representation hypergraph
Figure FDA0003388114120000036
Aggregation of the above from node features to hyper-edge features, and Hd-userIs represented by the multiplication operation of
Figure FDA0003388114120000037
The aggregation of the above from the hyper-edge features to the node features;
the convolution of the project side hypergraph in the step 6 is represented as:
Figure FDA0003388114120000038
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEDegree matrix of overcrowding, Hd-itemIs a hypergraph
Figure FDA0003388114120000039
H ═ V, E, where V is the set of nodes and E is the set of super edges, and
Figure FDA00033881141200000310
the multiplication operation of (2) represents the item side direct interactive relation representation hypergraph
Figure FDA00033881141200000311
Aggregation of the above from node features to hyper-edge features, and Hd-itemIs represented by the multiplication operation of
Figure FDA0003388114120000041
From the hyper-edge feature to the node feature.
7. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that:
the user-side hypergraph convolution in step 7 is represented as:
Figure FDA0003388114120000042
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a matrix of degrees of the overcrowding,
Figure FDA0003388114120000043
is a hypergraph
Figure FDA0003388114120000044
H ═ V, E, where V is the set of nodes and E is the set of super edges, and
Figure FDA0003388114120000045
the multiplication operation of (a) represents a user-side label representation hypergraph
Figure FDA0003388114120000046
Aggregation of the above from node features to hyper-edge features, and Ht-userIs represented by the multiplication operation of
Figure FDA0003388114120000047
The aggregation of the above from the hyper-edge features to the node features;
the convolution of the project side hypergraph in the step 7 is represented as:
Figure FDA0003388114120000048
wherein: theta(l)Representing a learnable parameter matrix at level I, σ being the activation function, DvDegree matrix of nodes, DEIs a degree matrix of the overcide. Ht-itemIs a hypergraph
Figure FDA0003388114120000049
H ═ V, E, where V is the set of nodes and E is the set of super edges, and
Figure FDA00033881141200000410
the multiplication operation of (2) represents an item-side label representation hypergraph
Figure FDA00033881141200000411
Aggregation of the above from node features to hyper-edge features, and Ht-itemIs represented by the multiplication operation of
Figure FDA00033881141200000412
From the hyper-edge feature to the node feature.
8. The label perception recommendation method based on the attention mechanism and the hypergraph convolution is characterized in that: the user side and project side feature processing in step 8 specifically comprises:
the characteristics u obtained according to the steps 2, 6 and 7tag、itag、u1、i1、u2、i2Obtaining final feature representation u on user side and project side using attention mechanism*、i*Wherein:
will utagAnd u1、u2Respectively carrying out splicing operation to obtain u1-tagAnd u2-tagThen, the attention mechanism is used to process the two feature representations into one feature representation:
a(u，k)＝WTtanh(Wuk-tag+b2)
Figure FDA0003388114120000051
obtaining a user-side final feature representation u by obtaining weights according to an attention mechanism*：
u*＝α(u，1)u1-tag+α(u，2)u2-tag；
Will itagAnd i1、i2Respectively carrying out splicing operation to obtain i1-tagAnd i2-tagThen, the attention mechanism is used to process the two feature representations into one feature representation:
a(i，k)＝WTtanh(Wik-tag+b2)
Figure FDA0003388114120000052
obtaining item side final feature representation i according to attention mechanism and weight*：
i*＝α(i，1)i1-tag+α(i，2)u2-tag。
CN202111461962.3A 2021-12-02 2021-12-02 Label perception recommendation method based on attention mechanism and hypergraph convolution Pending CN114117142A (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202111461962.3A CN114117142A (en) 2021-12-02 2021-12-02 Label perception recommendation method based on attention mechanism and hypergraph convolution
PCT/CN2022/106913 WO2023098098A1 (en) 2021-12-02 2022-07-21 Tag-aware recommendation method based on attention mechanism and hypergraph convolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111461962.3A CN114117142A (en) 2021-12-02 2021-12-02 Label perception recommendation method based on attention mechanism and hypergraph convolution

Publications (1)

Publication Number Publication Date
CN114117142A true CN114117142A (en) 2022-03-01

Family

ID=80365627

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111461962.3A Pending CN114117142A (en) 2021-12-02 2021-12-02 Label perception recommendation method based on attention mechanism and hypergraph convolution

Country Status (2)

Country Link
CN (1) CN114117142A (en)
WO (1) WO2023098098A1 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114817663A (en) * 2022-05-05 2022-07-29 杭州电子科技大学 Service modeling and recommendation method based on class perception graph neural network
WO2023098098A1 (en) * 2021-12-02 2023-06-08 南京邮电大学 Tag-aware recommendation method based on attention mechanism and hypergraph convolution

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117034778A (en) * 2023-08-28 2023-11-10 黑龙江省网络空间研究中心(黑龙江省信息安全测评中心、黑龙江省国防科学技术研究院) Method for inverting aboveground biomass based on hypershaper-transducer structure
CN116845889B (en) * 2023-09-01 2023-12-22 东海实验室 Hierarchical hypergraph neural network-based power load prediction method
CN116894097B (en) * 2023-09-04 2023-12-22 中南大学 Knowledge graph label prediction method based on hypergraph modeling
CN117788122B (en) * 2024-02-23 2024-05-10 山东科技大学 Goods recommendation method based on heterogeneous graph neural network

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109063163B (en) * 2018-08-14 2022-12-02 腾讯科技（深圳）有限公司 Music recommendation method, device, terminal equipment and medium
CN112487200B (en) * 2020-11-25 2022-06-07 吉林大学 Improved deep recommendation method containing multi-side information and multi-task learning
CN113434778B (en) * 2021-07-20 2023-03-24 陕西师范大学 Recommendation method based on regularization framework and attention mechanism
CN113672811B (en) * 2021-08-24 2023-04-18 广东工业大学 Hypergraph convolution collaborative filtering recommendation method and system based on topology information embedding and computer readable storage medium
CN114117142A (en) * 2021-12-02 2022-03-01 南京邮电大学 Label perception recommendation method based on attention mechanism and hypergraph convolution

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023098098A1 (en) * 2021-12-02 2023-06-08 南京邮电大学 Tag-aware recommendation method based on attention mechanism and hypergraph convolution
CN114817663A (en) * 2022-05-05 2022-07-29 杭州电子科技大学 Service modeling and recommendation method based on class perception graph neural network
CN114817663B (en) * 2022-05-05 2023-02-17 杭州电子科技大学 Service modeling and recommendation method based on class perception graph neural network

Also Published As

Publication number Publication date
WO2023098098A1 (en) 2023-06-08

Similar Documents

Publication Publication Date Title
CN114117142A (en) Label perception recommendation method based on attention mechanism and hypergraph convolution
US11803730B2 (en) Webinterface presentation using artificial neural networks
CN111382309B (en) Short video recommendation method based on graph model, intelligent terminal and storage medium
Tan et al. Deep representation learning for social network analysis
WO2023065859A1 (en) Item recommendation method and apparatus, and storage medium
CN111143705B (en) Recommendation method based on graph convolution network
CN113378048B (en) Individualized recommendation method based on multi-view knowledge graph attention network
WO2020147409A1 (en) Text classification method and apparatus, computer device, and storage medium
Guan et al. Bi-directional heterogeneous graph hashing towards efficient outfit recommendation
CN112559764A (en) Content recommendation method based on domain knowledge graph
Xiao et al. Link prediction based on feature representation and fusion
CN115221413B (en) Sequence recommendation method and system based on interactive graph attention network
Sun et al. Graph force learning
Amara et al. Cross-network representation learning for anchor users on multiplex heterogeneous social network
CN114925268A (en) Recommendation method and system based on graph neural network, electronic device and computer readable medium
Zhou et al. Rank2vec: learning node embeddings with local structure and global ranking
CN114579852A (en) Recommendation method based on graph convolution network and attention mechanism
CN117216281A (en) Knowledge graph-based user interest diffusion recommendation method and system
Chen et al. Integrating User-Group relationships under interest similarity constraints for social recommendation
CN112765490A (en) Information recommendation method and system based on knowledge graph and graph convolution network
CN114756768B (en) Data processing method, device, equipment, readable storage medium and program product
Xing et al. AI‐driven user aesthetics preference prediction for UI layouts via deep convolutional neural networks
CN112734519B (en) Commodity recommendation method based on convolution self-encoder network
CN114332472A (en) Data processing method and device based on graph neural network
Di et al. Structural plan of indoor scenes with personalized preferences

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