CN115455302A - Knowledge graph recommendation method based on optimized graph attention network - Google Patents
Knowledge graph recommendation method based on optimized graph attention network Download PDFInfo
- Publication number
- CN115455302A CN115455302A CN202211211810.2A CN202211211810A CN115455302A CN 115455302 A CN115455302 A CN 115455302A CN 202211211810 A CN202211211810 A CN 202211211810A CN 115455302 A CN115455302 A CN 115455302A
- Authority
- CN
- China
- Prior art keywords
- information
- recommended
- user
- knowledge graph
- preset
- 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 27
- 241000612182 Rexea solandri Species 0.000 claims abstract description 14
- 230000008447 perception Effects 0.000 claims abstract description 11
- 238000005457 optimization Methods 0.000 claims abstract description 10
- 239000013598 vector Substances 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 13
- 238000012549 training Methods 0.000 claims description 7
- 230000002159 abnormal effect Effects 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000008030 elimination Effects 0.000 claims description 4
- 238000003379 elimination reaction Methods 0.000 claims description 4
- 238000012546 transfer Methods 0.000 claims description 4
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 238000013461 design Methods 0.000 abstract description 7
- 230000003993 interaction Effects 0.000 description 6
- 230000008569 process Effects 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000000470 constituent Substances 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- 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)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Animal Behavior & Ethology (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to a knowledge graph recommendation method based on an optimization graph attention network, which comprises the steps of firstly, constructing a knowledge graph corresponding to an information set to be recommended, then adding users into the knowledge graph, then discovering additional relations among information to be recommended in the knowledge graph except for the difference of preset information correlation attributes, updating the knowledge graph, obtaining interest priorities of the information to be recommended corresponding to the users in the knowledge graph, finally obtaining target information to be recommended corresponding to the users, and recommending information to the users in a user group; according to the scheme design, a knowledge graph embedding model (HAKE) based on hierarchical perception is used for identifying the high-order relations which are not found in the knowledge graph, and an improved IMG-GCN model is used for effectively identifying the users with common interests by utilizing the characteristics of the users and the graph structure, so that negative information can be prevented from being transmitted from high-order neighbors to embedding learning, and the accuracy and the reliability of recommendation results are improved.
Description
Technical Field
The invention relates to a knowledge graph recommendation method based on an optimization graph attention network, and belongs to the technical field of machine learning.
Background
Recommendation systems have become one of the most important technologies for various online platforms, and not only can provide personalized information for a specific user from massive information, but also can increase revenue of service providers. Wherein Collaborative Filtering (CF) based models make substantial progress in learning user and project representations by modeling historical user-project interactions. For example, matrix Factorization (MF) may embed user terms directly as feature vectors and model the interaction of user terms with inner products. The neural collaborative filtering model replaces the MF interaction function of inner products with a nonlinear neural network to learn better user and item representations.
The knowledge graph is a part of knowledge engineering technology, is a huge heterogeneous information network, and has the basic constituent elements of a triple, such as (h, r, t) representing a triple, and h, r and t respectively representing a head node, a relation and a tail node. The knowledge graph establishes a deep semantic relation among the articles, so that more associated information among the articles can be mined, the knowledge graph is applied to a recommendation system, and the problems of data sparseness, cold start and the like can be effectively relieved.
Graph convolutional networks have great potential in recommendations due to their ability to learn good user and item embedding by utilizing cooperative signals from higher-order neighbors, but as layers are stacked more, node embedding becomes more similar and eventually indistinguishable, resulting in performance degradation.
Disclosure of Invention
The invention aims to solve the technical problem of providing a knowledge graph recommendation method based on an optimization graph attention network, which adopts a brand-new structural design, improves the recommendation efficiency and simultaneously improves the recommendation accuracy.
In order to solve the technical problems, the invention adopts the following technical scheme: the invention designs a knowledge graph recommendation method based on an optimization graph attention network, which is based on an information set to be recommended and comprises the following steps of recommending information to each user in a user group;
step A, taking each piece of information to be recommended in an information set to be recommended as a node, establishing a connection line aiming at the nodes with corresponding preset information associated attribute relations, constructing a knowledge graph corresponding to the information set to be recommended, and then entering step B;
b, respectively aiming at each user in the user group, respectively establishing a connection between the user and each piece of information to be recommended with the user as a node based on whether the user has a corresponding relation preset with the associated attribute of each piece of user information with each piece of information to be recommended, adding the connection into a knowledge graph corresponding to the information set to be recommended, updating the knowledge graph, and then entering the step C;
step C, discovering additional relations among all information to be recommended in the knowledge graph corresponding to the information set to be recommended and except for distinguishing preset information correlation attributes, establishing a connection line between nodes with the additional relations, updating the knowledge graph corresponding to the information set to be recommended, and then entering the step D;
d, obtaining the interest priority of each user corresponding to each information to be recommended in the knowledge graph according to the knowledge graph corresponding to the information set to be recommended, and entering the step E;
step E, respectively aiming at each user in the user group, and obtaining each target information to be recommended corresponding to the user based on the interest priority of each information to be recommended corresponding to the user; and further acquiring target information to be recommended corresponding to each user, and recommending information to each user in the user group.
As a preferred technical scheme of the invention: in the step A, firstly, information contents corresponding to preset information attributes in the information to be recommended are obtained respectively for each piece of information to be recommended in the information set to be recommended, the information to be recommended is updated in a combined manner, and then each piece of information to be recommended in the information set to be recommended is updated; and then establishing a connection line between the nodes with the relation based on whether the relation corresponding to the preset information correlation attributes exists between the information to be recommended or not, and constructing a knowledge graph corresponding to the information set to be recommended.
As a preferred technical scheme of the invention: in the step A, according to a preset abnormal information content library, aiming at each piece of information to be recommended, which is obtained based on the updating of each preset information attribute, abnormal information to be recommended is removed, then, each piece of remaining information to be recommended is used as a node, a connection line is established among the nodes which are in contact, and a knowledge graph corresponding to an information set to be recommended is constructed.
As a preferred technical scheme of the invention: in the step C, firstly, based on the fact that all information samples which are in the same information category as all information to be recommended are input and combined with the classification of all information samples corresponding to all preset information classification categories, the information samples except the information correlation attributes are distinguished and preset among all the information samples to form an extra information relation model which is output, all the information to be recommended in the knowledge graph is processed, and the extra relation among all the information to be recommended in the knowledge graph is obtained;
and then, establishing a connection line aiming at the nodes which are additionally connected with each other according to the additional connection among the information to be recommended in the knowledge graph, and updating the knowledge graph corresponding to the information set to be recommended.
As a preferred technical scheme of the invention: and C, the information additional relation model in the step C is obtained by training a hierarchical perception knowledge graph embedding model HAKE based on information samples which belong to the same information category as the information to be recommended, the information samples are correspondingly preset with the classification of the information classification categories, and the information samples are additionally related with each other except for the information correlation attributes which are preset among the information samples.
As a preferred technical scheme of the invention: the step D comprises the following steps D1 to D2;
d1, establishing a user-information two-dimensional matrix corresponding to the information set to be recommended based on whether connecting lines exist between each user node and each information node to be recommended in a knowledge graph corresponding to the information set to be recommended, taking each user and each information to be recommended as a horizontal coordinate and a vertical coordinate respectively, wherein the connecting lines between the users and the information to be recommended correspond to 1, the connecting lines between the users and the information to be recommended do not correspond to 0, and then entering the step D2;
and D2, based on the fact that the knowledge graphs and the user-information two-dimensional matrix obtained in the steps A, B, C and D1 are input by using an information sample set which belongs to the same information category as the information set to be recommended, an interest level model is output by using the interest level of each information sample in the information sample set corresponding to each user, and the knowledge graphs and the user-information two-dimensional matrix corresponding to the information set to be recommended are processed to obtain the interest level of each information sample in the information set to be recommended corresponding to each user.
As a preferred technical scheme of the invention: the interest level model in the step D2 is obtained by training the interest level model aiming at the IMG-graph convolution neural network GCN based on the interest perception message transfer based on the information sample set which belongs to the same information category as the information set to be recommended, the knowledge graph obtained in the steps A, B, C and D1 and the user-information two-dimensional matrix, wherein each user respectively corresponds to the interest priority of each information sample in the information sample set.
As a preferred technical scheme of the invention: in the step E, the following steps E1 to E4 are executed for each user in the user group;
step E1, based on the ranking of information to be recommended corresponding to the user from high to low in the priority of interest, deleting the information to be recommended which meets a preset recall elimination rule in the information to be recommended of the user, taking the rest information to be recommended as the primary information to be recommended corresponding to the user, and then entering step E2;
step E2, obtaining information of the user corresponding to preset user characteristic attributes to form a user characteristic vector corresponding to the user, obtaining information of each primary information to be recommended corresponding to the user and corresponding to the preset information characteristic attributes respectively to form an information characteristic vector corresponding to each primary information to be recommended respectively, and entering step E3;
step E3, respectively aiming at each piece of primary information to be recommended corresponding to the user, multiplying the information characteristic vector corresponding to the primary information to be recommended by the user characteristic vector corresponding to the user to form the prediction selection probability of the user about the primary information to be recommended; then the prediction selection probability of the user about each corresponding primary information to be recommended is obtained, and then the step E4 is carried out;
and E4, sequencing the primary information to be recommended corresponding to the user according to the sequence of the predicted selection probability from high to low, and sequentially selecting the information to be recommended before
The primary information to be recommended is used as each piece of target information to be recommended corresponding to the user, and information recommendation is carried out on the user; wherein, L represents the number of the primary information to be recommended corresponding to the user, b represents a preset second percentage,
representing an ceiling function.
As a preferred technical scheme of the invention: the preset recall elimination rule in the step E1 is as follows: eliminating all information to be recommended corresponding to the user from middle to end according to the sequence of the priority of interest from high to low
A piece of information to be recommended, N represents the number of the information to be recommended in the information set to be recommended, a represents a preset first percentage,
representing an rounding-up function.
As a preferred technical scheme of the invention: the information set to be recommended is a commodity information set to be recommended or a search information set to be recommended.
Compared with the prior art, the knowledge graph recommendation method based on the optimization graph attention network has the following technical effects by adopting the technical scheme:
the invention designs a knowledge graph recommendation method based on an optimization graph attention network, which comprises the steps of firstly constructing a knowledge graph corresponding to an information set to be recommended, then adding the knowledge graph with users as nodes, then exploring the information to be recommended in the knowledge graph, distinguishing additional relations among the information to be recommended except for the preset information correlation attributes, updating the knowledge graph, obtaining the interest priority of each user in the knowledge graph corresponding to each information to be recommended, finally obtaining each target information to be recommended corresponding to each user, and recommending information to each user in a user group; the scheme design can avoid the propagation of negative information from high-order neighbors to embedded learning by using a knowledge graph embedding model (HAKE) based on hierarchical perception to identify undiscovered high-order relations in a knowledge graph and embedding the knowledge graph, and using an improved IMG-GCN model to effectively identify users with common interests by using user characteristics and graph structures, thereby improving the accuracy and reliability of recommendation results.
Drawings
FIG. 1 is a flow chart of a knowledge graph recommendation method for designing an optimization graph attention network according to the present invention.
Detailed Description
The following description will explain embodiments of the present invention in further detail with reference to the accompanying drawings.
The invention designs a knowledge graph recommendation method based on an optimization graph attention network, which is used for recommending information to each user in a user group according to the following steps A to E based on an information set to be recommended in practical application and as shown in figure 1.
And step A, taking each piece of information to be recommended in the information set to be recommended as a node, establishing a connection line aiming at the nodes with corresponding preset information associated attribute relations, constructing a knowledge graph corresponding to the information set to be recommended, and then entering step B.
In specific implementation application, firstly, acquiring information contents corresponding to preset information attributes in information to be recommended respectively for each piece of information to be recommended in an information set to be recommended, combining and updating the information to be recommended, and further updating each piece of information to be recommended in the information set to be recommended; then, according to a preset abnormal information content library, aiming at each piece of information to be recommended, which is obtained based on the updating of each preset information attribute, abnormal information to be recommended is removed; and finally, establishing a connection line between the nodes with the connection based on whether the connection corresponding to the preset information correlation attributes exists between the rest pieces of information to be recommended or not, and constructing a knowledge graph corresponding to the information set to be recommended.
And B, respectively aiming at each user in the user group, respectively establishing a connection between the user and each piece of information to be recommended with the user as a node based on whether the user has a corresponding relation with the information to be recommended and preset the associated attribute of each piece of user information, respectively adding the connection between the user and each piece of information to be recommended with the contact, adding the connection into a knowledge graph corresponding to the information set to be recommended, updating the knowledge graph, and then entering the step C.
And C, discovering additional relations among the information to be recommended in the knowledge graph corresponding to the information set to be recommended and other information to be recommended except for distinguishing preset information correlation attributes, establishing a connection line between nodes with the additional relations, updating the knowledge graph corresponding to the information set to be recommended, and entering the step D.
In practical application, the step C is designed based on pre-training, and using each information sample of the same information category as each information to be recommended as input, combining the classification of each information sample corresponding to each preset information classification category, using the additional association among the information samples except the information correlation attribute which is distinguished and preset among the information samples as an output information additional relationship model, processing each information to be recommended in the knowledge graph, and obtaining the additional association among each information to be recommended in the knowledge graph; and then, establishing a connection line between nodes which are additionally connected according to additional connection among the information to be recommended in the knowledge graph, and updating the knowledge graph corresponding to the information set to be recommended.
And the information additional relation model is obtained by training a hierarchical perception knowledge graph embedding model HAKE based on information samples of the same information category as the information to be recommended, the information samples are correspondingly preset with the classification of the information classification categories, and the information samples are additionally related except for the information correlation attributes which are different and preset among the information samples.
The hierarchical perception knowledge graph embedding model HAKE is composed of two parts: a modeling part and a phase part, which are respectively used for modeling aiming at two different classes of entities.
The model part is intended to model entities at different levels in the hierarchy, using e m (e may also be h or t) and r m Representing entity embedding and relationship embedding. We will h m And t m Each term of (1), i.e. [ h ] m ] i And [ t m ] i Regarded as modulus; will r is m Each term of (1), i.e. [ r ] m ] i Considering the scale transformation between two modes, the modulus part can be expressed as:
the corresponding distance function is:
the purpose of the phase part is to model entities in the same hierarchy, using phase information to distinguish different entities in the same hierarchy, we will h p And t p Each term of [ h ] p ] i And [ t p ] i Is regarded as a phase, and r is p Each term of (i), i.e. [ r ] p ] i Considering phase conversion, the phase portion can be expressed as follows:
(h p +r p )mod 2π＝t p ,where h p ,t p ,r p ∈[0,2π) k (3)
the corresponding distance function is:
d r,p (h p ,t p )＝||sin((h p +r p -t p )/2)|| 1 (4)
in conjunction with the mode and phase sections, HAKE maps the entity into a polar coordinate system, where the radial and angular coordinates correspond to the mode and phase sections, respectively. In conjunction with the mode and phase sections, HAKE maps the entity into a polar coordinate system, where the radial and angular coordinates correspond to the mode and phase sections, respectively. That is, HAKE maps an entity h to [ h m ,h p ],h m And h p Are generated separately from the modulo part and the phase part and then spliced together. ([ h) m ] i ,[h p ] i ) Is a 2D point coordinate in the polar coordinate system. The HAKE formula is specifically defined as follows:
the distance function of HAKE is:
d r (h,t)＝d r,m (h m ,t m )+λd r,p (h p ,t p ) (6)
λ is a parameter of model learning. The corresponding score function is:
f r (h,t)＝d r (h,t)＝-d r,m (h,t)-λd r,p (h,t) (7)
by combining the mold part and the phase part, the HAKE can model the entities in class (a) and class (b). γ is a fixed margin, σ is the sigmoid function, (h' i ,r,t' i ) Is the ith negative sample. Thus, the HAKE can model the semantic hierarchy of the knowledge graph, resulting in higher order relationships by using a negative sampling loss function and self-confrontation training.
And D, obtaining the interest priority of each user corresponding to each information to be recommended respectively in the knowledge graph according to the knowledge graph corresponding to the information set to be recommended, and then entering the step E.
In practical applications, the step D specifically performs the following steps D1 to D2.
And D1, establishing a user-information two-dimensional matrix corresponding to the information set to be recommended based on whether connecting lines exist between each user node and each information node to be recommended in the knowledge graph corresponding to the information set to be recommended, taking each user and each information to be recommended as a horizontal coordinate and a vertical coordinate respectively, wherein the connecting lines between the users and the information to be recommended correspond to 1, and the connecting lines between the users and the information to be recommended do not correspond to 0, and then entering the step D2.
And D2, based on the information sample set which is pre-trained and belongs to the same information category as the information set to be recommended, executing the knowledge graph and the user-information two-dimensional matrix obtained in the steps A, B, C and D1 as input, and processing the knowledge graph and the user-information two-dimensional matrix corresponding to the information set to be recommended according to the interest priority of each user corresponding to each information sample in the information sample set as an output interest level model to obtain the interest priority of each user corresponding to each information to be recommended in the information set to be recommended.
In the implementation application, the interest level model in the step D2 executes the knowledge graph and the user-information two-dimensional matrix obtained in the steps a, B, C, and D1 based on the information sample set belonging to the same information category as the information set to be recommended, and each user is trained to obtain the interest level of each information sample in the information sample set corresponding to the interest level of each information sample in the interest perception information transfer IMG-graph convolutional neural network GCN.
The invention relates to an IMG-graph convolution neural network (GCN) based on interest perception message transfer, and in application, a bipartite graph containing interaction between users and items is defined
Point set composed of N nodes
The user and the item represented by two types of nodes are included, and an edge set epsilon formed by M edges represents the distance between the user and the itemInteracting, then an adjacency matrix representing the graph structure is
The sub-graph generation can be regarded as a node classification task, aiming at grouping users with similar interests and hobbies, and the fused embedding calculation of the users is as follows:
location vector U representing subscriber group affiliation o The calculation is as follows:
wherein the vector U o The position with the largest value indicates the allocated group, and the latitude size is the number of all the groups, namely the number of the sub-graphs which need to be created, and can be regarded as a hyper-parameter.
A sub-graph is defined as G s Wherein s belongs to {1,2, \8230, N s Represents N contained in the subgraph s And (4) each node. The subgraph is constructed according to user nodes, users with the same interests are divided into the same group, and project nodes with direct interaction behaviors are also divided into the same subgraph. Thus, a user has only one sub-graph, and an item can be associated with multiple sub-graphs.
Since the direct interaction behavior of the user with the project is crucial to obtaining the user's interest, the first-order propagation process of the GCN is set as: each interactive item participates directly in the information dissemination process as a first-order neighbor to the user. Define the first order GCN as:
in order to prevent noise interference caused by high-order information in a propagation process, a node in a user subgraph is defined to acquire information only from other adjacent nodes until the node acquires the informationThe purpose of isolating the information propagation of different interested users is achieved. The embedding of the item i in the subgraph Gs after passing through the kth layer GCN is expressed as
Accordingly, the user including the high-order information can be acquired as
The above formula ensures that the nodes in the subgraph only carry out information propagation in the subgraph, and the attention is paid to
The characteristic information transmitted by users with similar interests in the subgraph is accepted, and users who are helpful to analyze and draw similar interests always select a certain item for a common reason. The embedding for each layer with respect to item i is calculated as the sum of the resulting embeddings for each sub-graph s:
by weighted summation of the user and the item embedding obtained from each layer, by using the averaging method, i.e. setting
Thereby reach multilayer embedding and unite:
average aggregation of the output of each layer is carried out to obtain the final embedded representation of the user and the item, and the obtained user embedding e is utilized u Calculating the similarity with the target item:
step E, aiming at each user in the user group, obtaining each target information to be recommended corresponding to the user based on the interest priority of each information to be recommended corresponding to the user; and further acquiring target information to be recommended corresponding to each user, and recommending information to each user in the user group.
In practical applications, the step E performs the following steps E1 to E4 for each user in the user group.
And E1, based on the sequence of the information to be recommended corresponding to the user from high to low in the interest priority, deleting the information to be recommended which meets a preset recall rejection rule in the information to be recommended of the user, using the rest information to be recommended as the primary information to be recommended corresponding to the user, and entering the step E2.
Wherein the preset recall rejection rule is as follows: eliminating all information to be recommended corresponding to the user from middle to end according to the sequence of the priority of interest from high to low
The information to be recommended is obtained, N represents the number of the information to be recommended in the information set to be recommended, a represents a preset first percentage,
representing an ceiling function.
And E2, obtaining information of the user corresponding to preset user characteristic attributes to form a user characteristic vector corresponding to the user, obtaining information of each primary information to be recommended corresponding to the user and corresponding to the preset information characteristic attributes respectively to form an information characteristic vector corresponding to each primary information to be recommended respectively, and then entering the step E3.
Step E3, respectively aiming at each piece of primary information to be recommended corresponding to the user, multiplying the information characteristic vector corresponding to the primary information to be recommended by the user characteristic vector corresponding to the user to form the prediction selection probability of the user about the primary information to be recommended; and then the prediction selection probability of the user respectively about each piece of primary information to be recommended corresponding to the user is obtained, and then the step E4 is carried out.
And E4, sequencing the primary information to be recommended corresponding to the user according to the sequence of the predicted selection probability from high to low, and sequentially selecting the information to be recommended before
The primary information to be recommended is used as each piece of target information to be recommended corresponding to the user, and information recommendation is carried out on the user; wherein, L represents the number of the primary information to be recommended corresponding to the user, b represents a preset second percentage,
representing an rounding-up function.
The knowledge graph recommendation method based on the optimization graph attention network is applied to practice, such as the information set to be recommended is a to-be-recommended commodity information set or a to-be-recommended search information set, and information recommendation of the to-be-recommended commodity information set or the to-be-recommended search information set is further achieved for each user in a user group.
The method comprises the steps of firstly constructing a knowledge graph corresponding to an information set to be recommended, then adding users into the knowledge graph by taking the users as nodes, then exploring additional relations among information to be recommended in the knowledge graph except for the difference preset information correlation attributes, updating the knowledge graph, obtaining interest priorities of the information to be recommended corresponding to the users in the knowledge graph, finally obtaining target information to be recommended corresponding to the users, and recommending information to the users in a user group; the scheme design can avoid negative information from being propagated from high-order neighbors to embedded learning by using a knowledge graph embedding model (HAKE) based on hierarchical perception to identify undiscovered high-order relations in a knowledge graph and embedding the knowledge graph, and using an improved IMG-GCN model to effectively identify users with common interests by using user characteristics and graph structures, so that the accuracy and reliability of recommendation results are improved.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.
Claims (10)
1. A knowledge graph recommendation method based on an optimization graph attention network is characterized by comprising the following steps: based on the information set to be recommended, information recommendation is carried out on each user in the user group according to the following steps;
step A, taking each piece of information to be recommended in an information set to be recommended as a node, establishing a connection line aiming at the nodes with corresponding preset information associated attribute relations, constructing a knowledge graph corresponding to the information set to be recommended, and then entering step B;
b, respectively aiming at each user in the user group, respectively establishing a connection between the user and each piece of information to be recommended with the user as a node based on whether the user has a corresponding relation preset with the associated attribute of each piece of user information with each piece of information to be recommended, adding the connection into a knowledge graph corresponding to the information set to be recommended, updating the knowledge graph, and then entering the step C;
step C, discovering additional relations among all information to be recommended in the knowledge graph corresponding to the information set to be recommended and except for distinguishing preset information correlation attributes, establishing a connection line between nodes with the additional relations, updating the knowledge graph corresponding to the information set to be recommended, and then entering the step D;
d, obtaining the interest priority of each user corresponding to each information to be recommended in the knowledge graph according to the knowledge graph corresponding to the information set to be recommended, and entering the step E;
step E, respectively aiming at each user in the user group, and obtaining each target information to be recommended corresponding to the user based on the interest priority of each information to be recommended corresponding to the user; and further acquiring target information to be recommended corresponding to each user, and recommending information to each user in the user group.
2. The knowledge graph recommendation method based on the optimized graph attention network according to claim 1, characterized in that: in the step A, firstly, information contents corresponding to preset information attributes in the information to be recommended are respectively obtained for each piece of information to be recommended in the information to be recommended set, the information to be recommended is updated in a combined mode, and then each piece of information to be recommended in the information to be recommended set is updated; and then, based on whether a relation corresponding to preset information correlation attributes exists between the pieces of information to be recommended, taking the pieces of information to be recommended as nodes, establishing a connection line between the nodes with the relation, and constructing a knowledge graph corresponding to the information set to be recommended.
3. The knowledge graph recommendation method based on the optimized graph attention network as claimed in claim 2, wherein: in the step A, according to a preset abnormal information content library, aiming at each piece of information to be recommended, which is obtained by updating based on each preset information attribute, abnormal information to be recommended is removed, then, each piece of residual information to be recommended is used as a node, a connection line is established between the nodes with connection, and a knowledge graph corresponding to an information set to be recommended is constructed.
4. The method for optimizing the knowledge graph recommendation based on the graph attention network as claimed in claim 1, wherein: in the step C, firstly, based on the fact that all information samples which are in the same information category as all information to be recommended are input and combined with the classification of all information samples corresponding to all preset information classification categories, the information samples except the information correlation attributes are distinguished and preset among all the information samples to form an extra information relation model which is output, all the information to be recommended in the knowledge graph is processed, and the extra relation among all the information to be recommended in the knowledge graph is obtained;
and then, establishing a connection line aiming at the nodes which are additionally connected with each other according to the additional connection among the information to be recommended in the knowledge graph, and updating the knowledge graph corresponding to the information set to be recommended.
5. The knowledge graph recommendation method based on the optimized graph attention network according to claim 4, characterized in that: and C, the information additional relation model in the step C is obtained by training a hierarchical perception knowledge graph embedding model HAKE based on information samples which belong to the same information category as the information to be recommended, the information samples are correspondingly preset with the classification of the information classification categories, and the information samples are additionally related with each other except for the information correlation attributes which are preset among the information samples.
6. The method for optimizing the knowledge graph recommendation based on the graph attention network as claimed in claim 1, wherein: the step D comprises the following steps D1 to D2;
step D1, based on whether connection lines exist between each user node and each information node to be recommended in a knowledge graph corresponding to an information set to be recommended, taking each user and each information to be recommended as a horizontal coordinate and a vertical coordinate respectively, enabling the connection lines between the users and the information to be recommended to correspond to 1, enabling the connection lines between the users and the information to be recommended to correspond to 0, constructing a user-information two-dimensional matrix corresponding to the information set to be recommended, and then entering step D2;
and D2, based on the fact that the knowledge graphs and the user-information two-dimensional matrix obtained in the steps A, B, C and D1 are input by using an information sample set which belongs to the same information category as the information set to be recommended, an interest level model is output by using the interest level of each information sample in the information sample set corresponding to each user, and the knowledge graphs and the user-information two-dimensional matrix corresponding to the information set to be recommended are processed to obtain the interest level of each information sample in the information set to be recommended corresponding to each user.
7. The method of claim 6, wherein the method comprises the steps of: the interest level model in the step D2 is obtained by executing the knowledge graph and the user-information two-dimensional matrix obtained in the steps A, B, C and D1 based on the information sample set which belongs to the same information category as the information set to be recommended, wherein each user respectively corresponds to the interest priority of each information sample in the information sample set, and the interest level model is obtained by training the IMG-graph convolution neural network GCN based on interest perception message transfer.
8. The method for optimizing the knowledge graph recommendation based on the graph attention network as claimed in claim 1, wherein: in the step E, the following steps E1 to E4 are executed for each user in the user group;
step E1, based on the ranking of information to be recommended corresponding to the user from high to low in the priority of interest, deleting the information to be recommended which meets a preset recall elimination rule in the information to be recommended of the user, taking the rest information to be recommended as the primary information to be recommended corresponding to the user, and then entering step E2;
step E2, obtaining information of each preset user characteristic attribute corresponding to the user to form a user characteristic vector corresponding to the user, obtaining information of each preset information characteristic attribute corresponding to each primary information to be recommended corresponding to the user to form an information characteristic vector corresponding to each primary information to be recommended, and entering step E3;
step E3, respectively aiming at each piece of primary information to be recommended corresponding to the user, multiplying the information characteristic vector corresponding to the primary information to be recommended by the user characteristic vector corresponding to the user to form the prediction selection probability of the user about the primary information to be recommended; then the prediction selection probability of the user about each corresponding primary information to be recommended is obtained, and then the step E4 is carried out;
and E4, sequencing the primary information to be recommended corresponding to the user according to the sequence of the predicted selection probability from high to low, and sequentially selecting the information to be recommended before
Initial information to be recommendedInformation serving as information to be recommended of each target corresponding to the user is recommended to the user; wherein, L represents the number of the primary information to be recommended corresponding to the user, b represents a preset second percentage,
representing an rounding-up function.
9. The knowledge graph recommendation method based on the optimized graph attention network according to claim 8, characterized in that: the preset recall elimination rule in the step E1 is as follows: eliminating all information to be recommended corresponding to the user from middle to end according to the sequence of the priority of interest from high to low
A piece of information to be recommended, N represents the number of the information to be recommended in the information set to be recommended, a represents a preset first percentage,
representing an rounding-up function.
10. The knowledge graph recommendation method based on the optimized graph attention network according to claim 1, characterized in that: the information set to be recommended is a commodity information set to be recommended or a search information set to be recommended.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211211810.2A CN115455302A (en) | 2022-09-30 | 2022-09-30 | Knowledge graph recommendation method based on optimized graph attention network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211211810.2A CN115455302A (en) | 2022-09-30 | 2022-09-30 | Knowledge graph recommendation method based on optimized graph attention network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115455302A true CN115455302A (en) | 2022-12-09 |
Family
ID=84309269
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211211810.2A Pending CN115455302A (en) | 2022-09-30 | 2022-09-30 | Knowledge graph recommendation method based on optimized graph attention network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115455302A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117057929A (en) * | 2023-10-11 | 2023-11-14 | 中邮消费金融有限公司 | Abnormal user behavior detection method, device, equipment and storage medium |
-
2022
- 2022-09-30 CN CN202211211810.2A patent/CN115455302A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117057929A (en) * | 2023-10-11 | 2023-11-14 | 中邮消费金融有限公司 | Abnormal user behavior detection method, device, equipment and storage medium |
| CN117057929B (en) * | 2023-10-11 | 2024-01-26 | 中邮消费金融有限公司 | Abnormal user behavior detection method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110837602B (en) | User recommendation method based on representation learning and multi-mode convolutional neural network | |
| CN112232925A (en) | Method for carrying out personalized recommendation on commodities by fusing knowledge maps | |
| CN112861967B (en) | Social network abnormal user detection method and device based on heterogeneous graph neural network | |
| CN111881350B (en) | Recommendation method and system based on mixed graph structured modeling | |
| CN112052404B (en) | Group discovery method, system, equipment and medium of multi-source heterogeneous relation network | |
| CN113918832A (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
| CN116601626A (en) | Personal knowledge graph construction method and device and related equipment | |
| CN115422441A (en) | Continuous interest point recommendation method based on social space-time information and user preference | |
| CN113918834A (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
| CN115344698A (en) | Label processing method, label processing device, computer equipment, storage medium and program product | |
| CN116089645A (en) | Hierarchical style-based conditional text-e-commerce picture retrieval method and system | |
| CN115455302A (en) | Knowledge graph recommendation method based on optimized graph attention network | |
| CN113849725B (en) | Socialized recommendation method and system based on graph attention confrontation network | |
| Tang et al. | Cold-start link prediction integrating community information via multi-nonnegative matrix factorization | |
| CN112905906B (en) | Recommendation method and system fusing local collaboration and feature intersection | |
| Maksimov et al. | Addressing cold start in recommender systems with hierarchical graph neural networks | |
| CN110543601B (en) | Method and system for recommending context-aware interest points based on intelligent set | |
| CN116932923A (en) | Project recommendation method combining behavior characteristics and triangular collaboration metrics | |
| Ariza-Casabona et al. | Exploiting graph structured cross-domain representation for multi-domain recommendation | |
| Sangeetha et al. | Predicting personalized recommendations using GNN | |
| CN115545833A (en) | Recommendation method and system based on user social information | |
| CN113744023B (en) | Double-channel collaborative filtering recommendation method based on graph convolution network | |
| CN115391555A (en) | User-perceived knowledge map recommendation system and method | |
| CN115131058A (en) | Account identification method, device, equipment and storage medium | |
| Taranto et al. | Uncertain Graphs meet Collaborative Filtering. |
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 |