CN111737591A - Product recommendation method based on heterogeneous heavy-side information network translation model - Google Patents

Product recommendation method based on heterogeneous heavy-side information network translation model Download PDF

Info

Publication number
CN111737591A
CN111737591A CN202010484083.1A CN202010484083A CN111737591A CN 111737591 A CN111737591 A CN 111737591A CN 202010484083 A CN202010484083 A CN 202010484083A CN 111737591 A CN111737591 A CN 111737591A
Authority
CN
China
Prior art keywords
node
heavy
edge
heterogeneous
space
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.)
Granted
Application number
CN202010484083.1A
Other languages
Chinese (zh)
Other versions
CN111737591B (en
Inventor
郑建兴
李德玉
梁吉业
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.)
Shanxi University
Original Assignee
Shanxi University
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 Shanxi University filed Critical Shanxi University
Priority to CN202010484083.1A priority Critical patent/CN111737591B/en
Publication of CN111737591A publication Critical patent/CN111737591A/en
Application granted granted Critical
Publication of CN111737591B publication Critical patent/CN111737591B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/40—Processing or translation of natural language
    • G06F40/58—Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/01—Social networking
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Marketing (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to the technical field of heterogeneous information networks, and discloses a product recommendation method based on a heterogeneous heavy-edge information network translation model. Firstly, performing fusion expression space learning on the multi-element heavy edge relation between node pairs in the heterogeneous heavy edge information network by adopting an attention mechanism; the type information of the heterogeneous nodes is projected to a multivariate heavy-edge relation fusion expression space, and the distributed expression of the heterogeneous nodes and the relation is realized on the multivariate heavy-edge relation fusion expression space by using a translation model; and then carrying out node prediction research based on the representation of the heterogeneous nodes and the fused representation of the multi-element multiple edge relation. The heterogeneous node embedding based on the multivariate repeated edge relation can improve the accuracy of heterogeneous repeated edge information network link prediction, has higher performance than the traditional representation learning method, and has higher practical value in the aspect of commodity recommendation based on the multivariate relation.

Description

Product recommendation method based on heterogeneous heavy-side information network translation model
Technical Field
The invention relates to the technical field of heterogeneous heavy-edge information networks, in particular to a product recommendation method based on a heterogeneous heavy-edge information network translation model.
Background
The development of social media enables users to publish content of interest to the users anytime and anywhere. The social activity of the user is improved by the release of the user interest content, the social relationship of the user is enhanced, and a complex social network is formed. The rapid development of electronic commerce enables the interaction between users and products to be more and more frequent, and the interaction between a huge user group and a plurality of products also forms a complex network. The social network and the E-commerce network are provided with a plurality of types of interactive objects, and the interactive objects have different types of relations, so that a heterogeneous information network is formed. The method for predicting the relationship between objects based on the heterogeneous information network has become a hot research direction of the link prediction and recommendation system of the current social network and electronic commerce platform. In a complex heterogeneous information network, the relationship between any node object may have various kinds, and diversified relationships may be formed between the objects. That is, in a heterogeneous information network, each node object may have multiple types, and any two node objects may have multiple relationship edges, forming a special network, i.e., a heterogeneous heavy-edge information network.
The translation models in the knowledge graph, such as TransE, TransD, TransH, TransR and the like, effectively acquire semantic information of objects and relations by learning distributed representation of head nodes and tail nodes in triple pairs through a translation mechanism, and obtain remarkable performance in the aspect of link prediction. The object types of the node pairs in the heterogeneous heavy-edge information network are different, multiple heavy-edge relations exist among the nodes, the method that the traditional translation model mainly processes the single type relations among the node objects is not applicable any more, and the semantic representation capability of entities in the knowledge graph can be improved by fusing the multiple heavy-edge relations among the node pairs in the heterogeneous heavy-edge information network. Semantic information of entity objects of the heterogeneous nodes can be learned on the basis of a translation model in a multivariate heavy-edge relation fusion expression space of the heterogeneous nodes, and link sequencing recommendation among the nodes is achieved.
Disclosure of Invention
Aiming at the problems, the invention provides a product recommendation method based on a heterogeneous heavy-side information network translation model.
In order to achieve the purpose, the invention adopts the following technical scheme:
a product recommendation method based on a heterogeneous heavy-side information network translation model comprises the following steps:
step S1, performing fusion representation space learning on the multi-element heavy edge relation between the node pairs in the heterogeneous heavy edge information network by adopting an attention mechanism;
step S2, projecting the type information of the heterogeneous nodes to a multivariate heavy-edge relation fusion expression space;
step S3, an embedded model of heterogeneous heavy edges is provided by using a translation mechanism on the multivariate heavy edge relation fusion expression space, and distributed expression of heterogeneous nodes and relations is realized;
in step S4, node prediction is performed based on the fused representation space of the representation of the node and the multiple edge relation.
Further, in the step S1, performing fusion representation space learning on the multivariate heavy-edge relationship between the node pairs in the heterogeneous heavy-edge information network by using an attention mechanism further includes the following steps:
step S1.1, setting the heterogeneous information network G ═ V, E, where V is a set of nodes and E is a set of edges; in heterogeneous information networks, there is a mapping ψ V → T of a set of nodes V to a set of node types T and a mapping E of a set of edges E to a set of edge types A
Figure BDA0002518303760000021
And | T | + | A | > 2; different types of nodes in the heterogeneous information network reflect different semantic information of the nodes, and can be used for describing different semantic features of users or products in the social network; different types of edges in the heterogeneous information network reflect different interactive semantic information, and can be used for modeling different behaviors between users and products and capturing different behavior semantics between the users and the products;
step S1.2, setting a heterogeneous heavy-edge information network G ' ═ V ', E ', where V ' is a set of nodes and E ' is a set of edges; there is a mapping ψ 'of the node set V' to the node type set T V '→ T and a mapping of the edge set E' to the edge type set A
Figure BDA0002518303760000031
And | T | + | AIf the binary group (u, v) ∈ E 'exists at the same time, and the number of edges of the associated node pair (u, v) is more than 1, G' is called as a heterogeneous heavy-edge information network, a plurality of connecting edges of the node pair (u, v) in the heterogeneous heavy-edge information network reflect various interactive behavior relations among the nodes (u, v), and rich social behaviors between learning users and products can be modeled so as to provide accurate product recommendation for the users;
step S1.3, in the heterogeneous heavy-edge information network G '(V', E '), given a bigram (u, V) ∈ E', given a set of multiple heavy-edge relationships between node pairs as R ═ R | R ∈ a }, R representing an arbitrary type of relationship between node pairs (u, V), and R representing a set of different types of relationships between node pairs (u, V), then a set of trigram pairs of multiple heavy-edge relationships are denoted as (u, R, V), and given a behavior relationship R between node pairs (u, V)i∈ R, the mapping process is defined as:
Figure BDA0002518303760000032
wherein, aiRepresents a certain behavioral relationship r of node pairs (u, v)iVector representation of
Figure BDA0002518303760000033
The linear transformation of (a) to (b),
Figure BDA0002518303760000034
the shared parameter is used for learning different importance of the multiple heavy edge relation; a isiBy using the theory of spatial transformation, the invention focuses on the relation riVector representation of
Figure BDA0002518303760000035
The projection is mapped on
Figure BDA0002518303760000036
Transforming the space to obtain new features;
s1.4, calculating the mapping characteristics of any relation for all behavior relations R of the node pairs (u, v), and pairing the relation R through a softmax functioniThe normalization weight coefficients are:
Figure BDA0002518303760000037
wherein, αiRepresents the relation riWeight coefficient in all behavioral relationships of node pair (u, v), e is a natural constant, akIs the relation r of node pair (u, v)k(k ═ 1, 2., | | R | |), the number of elements in the multiple heavy-edge relation set R |, and a weight coefficient αiDifferent importance of the node pair (u, v) behavior relation is described, and recommendation of products based on the behavior relation with large weight is facilitated;
s1.5, reflecting different importance of different types of relations in the node pair (u, v) connection by weight coefficients; then, the multivariate and heavy-edge relationship fusion expression space of the node pairs is learned through the attention weight coefficient, and is recorded as:
Figure BDA0002518303760000041
wherein the content of the first and second substances,
Figure BDA0002518303760000042
vector weighted fusion representing the multivariate multiple edge relationship of node pairs (u, v),
Figure BDA0002518303760000043
is a relation riA vector representation of (a); fused representation space
Figure BDA0002518303760000044
Linear combinations of different types of relationships of node pairs (u, v) are depicted, which can explain the synergy of various interaction behaviors of social network users and products.
Further, the step S2, projecting the type information of the heterogeneous node to the multivariate heavy-edge relationship fusion expression space further includes the following steps:
s2.1, in the heterogeneous heavy-edge information network, giving a triple heterogeneous node pair (u, R, v) of a multi-element heavy-edge relation, wherein u and v have heterogeneous types; the heterogeneous nodes u and v have a multiple heavy edge relation, and the multiple heavy edge relation is set as R; r describes various interactive behaviors between a user and a product;
step S2.2
Figure BDA0002518303760000045
Respectively mapping a head node u and a tail node v of the node pair (u, v) to a uniform potential semantic space; the representation of the head node u and the tail node v in the unified space can be expressed as:
Figure BDA0002518303760000046
Figure BDA0002518303760000047
wherein the content of the first and second substances,
Figure BDA0002518303760000048
different types of spaces T for head node u and tail node vu、TvThe learning parameter matrix of (2);
Figure BDA0002518303760000049
vector representations of a head node u and a tail node v, respectively; vector representation of heterogeneous type head node u and tail node v
Figure BDA00025183037600000410
From the space T of the typeu、TvRespectively mapping to a uniform space phi to obtain a vector representation of the uniform space phi
Figure BDA00025183037600000411
The vector representation of the uniform space is beneficial to realizing the latent space semantic projection of the user and product type information, and the representation between the user and the product can be learned on a uniform scale;
step S2.3, considering the relevance of the head node and the tail node on the multi-element heavy edge relation fusion semantic space, adopting the concept of TransR to map the potential space representation of the head node and the tail node to the multi-element heavy edge relation fusion semantic space, and establishing the representation of the head node and the tail node, namely:
Figure BDA0002518303760000051
Figure BDA0002518303760000052
wherein the content of the first and second substances,
Figure BDA0002518303760000053
a projection parameter matrix of the expression space is fused for the multivariate heavy-edge relation; representing vectors of a head node u and a tail node v in phi space
Figure BDA0002518303760000054
Vector mapping to multiple heavy edge relation fusion expression space
Figure BDA0002518303760000055
To obtain a fused representation space
Figure BDA0002518303760000056
Representation of vectors in
Figure BDA0002518303760000057
The distributed representation between the users and the products is learned on the multivariate heavy-edge relation fusion representation space, and the establishment of the relation between the users and the products on the multivariate type interactive behavior relation is facilitated.
Further, in the step S3, the method for providing an embedded model of heterogeneous heavy edges by using a translation mechanism in the multivariate heavy edge relationship fusion expression space, and implementing distributed expression of heterogeneous nodes and relationships further includes the following steps:
s3.1, adopting the idea of knowledge representation translation, and establishing the association between the heterogeneous node pairs (u, v) through a translation mechanism on the multivariate and multiple edge relation fusion representation space, wherein the association is formed as follows:
Figure BDA0002518303760000058
based on the idea of translation, word2vec translation invariance is used, distributed representation of head nodes and tail nodes of heterogeneous types in a low-dimensional dense space can be achieved, the semantics of the head nodes and the tail nodes of the heterogeneous types are represented, and rich semantic information of users and products can be described;
step S3.2, for a triple heterogeneous node pair (u, R, v) with a multiple heavy-edge relationship, a distance function based on the translation mechanism may be defined as:
Figure BDA0002518303760000059
wherein f isR(u, v) represents the fusion of the head node u and the tail node v in the multi-element heavy-edge relation representation space
Figure BDA00025183037600000510
A distance function of (a);
Figure BDA0002518303760000061
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure BDA0002518303760000062
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure BDA0002518303760000063
a fusion representation space of a multivariate heavy edge relation set R representing a head node u and a tail node v; the preference distance of the user to the product can be in the fusion representation space of the multiple edge relation set R
Figure BDA0002518303760000064
And calculating, and based on the multivariate heavy edge relation, realizing accurate sequencing recommendation of the product by the user.
Step S3.3 for the positive sample triplet (u, R, v) and the negative sample triplet (u, R, v) according to the distance function of the triplet heterogeneous node pair (u, R, v)Triple unit
Figure BDA0002518303760000065
Defining a translation-relationship-based Hinge Loss function as:
Figure BDA0002518303760000066
wherein S is a set of (u, R, v) positive samples, and a head node u and a tail node v have a multiple-element multiple-edge relation R;
Figure BDA0002518303760000067
is a set of negative examples of the negative samples,
Figure BDA0002518303760000068
is a head node based on (u, R, v) replacement, and
Figure BDA0002518303760000069
there is no connection edge relation or multiple edge relation R with v,
Figure BDA00025183037600000610
is a tail node based on (u, R, v) replacement, and u is
Figure BDA00025183037600000611
No continuous edge relation or multiple edge relation R exists, gamma represents boundary parameters, lambda represents a hyper-parameter, η represents a regularization hyper-parameter;
Figure BDA00025183037600000612
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure BDA00025183037600000613
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure BDA00025183037600000614
representing head node
Figure BDA00025183037600000615
Vector representation on the multi-element heavy-edge relation fusion representation space;
Figure BDA00025183037600000616
representing tail nodes
Figure BDA00025183037600000617
Vector representation on the multi-element heavy-edge relation fusion representation space; t isu、TvType spaces of a head node u and a tail node v respectively;
Figure BDA00025183037600000618
and
Figure BDA00025183037600000619
orthogonality ensures that different types of heterogeneous head nodes and tail nodes can learn different parameter matrixes; the method comprises the steps that Hinge Loss learns the distance between a head node and a tail node in a space through multivariate heavy edge relation fusion, Hinge Loss minimization requires that the distance of a positive sample is smaller than that of a negative sample and is lower than a boundary parameter gamma, and distributed representation of the head node and the tail node is iteratively learned; the smaller distance between the positive samples is learned to the user, and semantic representation of the product is beneficial to selecting favorite products for the user on the multivariate and repeated edge relation fusion representation space.
Further, in step S4, the method for node prediction based on the fused representation space of the node representation and the multi-element multiple edge relationship is to adopt the new node triple (u, R, v
Figure BDA0002518303760000071
Determining the distance between the node v' and the head node u, and sorting in the candidate set according to the distance; fused representation space in a multi-element heavy edge relation R
Figure BDA0002518303760000072
And calculating a distance function between the projection of the user and the projection of the product, wherein the product with smaller distance from the user is semantically similar to the user, and is the product preferred by the user.
The invention provides a product recommendation method based on a heterogeneous heavy-side information network translation model. Firstly, performing fusion expression space learning on multi-element heavy edge relations between node pairs in a heterogeneous heavy edge information network by adopting an attention mechanism, then projecting and learning type information of heterogeneous nodes to a multi-element heavy edge relation fusion expression space, providing an embedded model of the heterogeneous heavy edges on the multi-element heavy edge relation fusion expression space by utilizing a translation mechanism, realizing distributed expression of the heterogeneous nodes and relations, and then performing node prediction research on the basis of the expression of the nodes and the fusion expression space of the multi-element heavy edge relations. The recommended ordering of the nodes is the final output result of the invention.
Compared with the prior art, the invention has the following advantages:
the product recommendation method based on the heterogeneous heavy-edge information network translation model is different from the existing model and is characterized in that fusion representation space learning is carried out on multi-element heavy-edge relations among node pairs in the heterogeneous heavy-edge information network by adopting an attention mechanism, the type information of heterogeneous nodes is projected and learned to the multi-element heavy-edge relation fusion representation space, an embedded model of heterogeneous heavy edges is provided on the multi-element heavy-edge relation fusion representation space by utilizing a translation mechanism, distributed representation of the heterogeneous nodes and relations is realized, the accuracy of heterogeneous heavy-edge information network link prediction can be improved, the performance is higher than that of the traditional representation learning method, and the product recommendation method has higher practical value in the aspect of commodity recommendation.
Drawings
FIG. 1 is a schematic diagram of an overall model architecture.
Detailed Description
The method for identifying and marking the attention relationship based on the user hierarchical theme preference semantic matrix is implemented by a computer program. The following is a detailed description of the embodiments of the present invention according to the flow chart. FIG. 1 is a schematic diagram of an overall model architecture.
A product recommendation method based on a heterogeneous heavy-side information network translation model comprises the following steps:
step S1, performing fusion representation space learning on the multi-element heavy edge relation between the node pairs in the heterogeneous heavy edge information network by adopting an attention mechanism;
step S1.1, setting the heterogeneous information network G ═ V, E, where V is a set of nodes and E is a set of edges; in heterogeneous information networks, there is a mapping ψ V → T of a set of nodes V to a set of node types T and a mapping E of a set of edges E to a set of edge types A
Figure BDA0002518303760000081
And | T | + | A | > 2;
step S1.2, setting a heterogeneous heavy-edge information network G ' ═ V ', E ', where V ' is a set of nodes and E ' is a set of edges; there is a mapping ψ 'of the node set V' to the node type set T V '→ T and a mapping of the edge set E' to the edge type set A
Figure BDA0002518303760000082
If two tuples (u, v) ∈ E 'exist at the same time and the number of edges of the associated node pair (u, v) is more than 1, G' is called a heterogeneous heavy-edge information network;
step S1.3, in the heterogeneous heavy-edge information network G '(V', E '), given a bigram (u, V) ∈ E', given a set of multiple heavy-edge relationships between node pairs as R ═ R | R ∈ a }, R representing an arbitrary type of relationship between node pairs (u, V), and R representing a set of different types of relationships between node pairs (u, V), then a set of trigram pairs of multiple heavy-edge relationships are denoted as (u, R, V), and given a behavior relationship R between node pairs (u, V)i∈ R, the mapping process is defined as:
Figure BDA0002518303760000083
wherein, aiRepresents a certain behavioral relationship r of node pairs (u, v)iVector representation of
Figure BDA0002518303760000084
The linear transformation of (a) to (b),
Figure BDA0002518303760000085
for sharing parameters, for learning moreDifferent importance of the meta-heavy edge relationship;
s1.4, calculating the mapping characteristics of any relation for all behavior relations R of the node pairs (u, v), and pairing the relation R through a softmax functioniThe normalization weight coefficients are:
Figure BDA0002518303760000091
wherein, αiRepresents the relation riWeight coefficient in all behavioral relationships of node pair (u, v), e is a natural constant, akIs the relation r of node pair (u, v)k(k ═ 1,2, ·, | | R | |), where | R | | is the number of elements in the set of multivariate multiple edge relationships R;
step S1.5 the weighting factors reflect the different importance of different types of relationships in the node pair (u, v) connection. Then, the multivariate and heavy-edge relationship fusion expression space of the node pairs is learned through the attention weight coefficient, and is recorded as:
Figure BDA0002518303760000092
wherein the content of the first and second substances,
Figure BDA0002518303760000093
vector weighted fusion representing the multivariate multiple edge relationship of node pairs (u, v),
Figure BDA0002518303760000094
is a relation riIs represented by a vector of (a).
Step S2, projecting the type information of the heterogeneous nodes to a multivariate heavy-edge relation fusion expression space;
s2.1, in the heterogeneous heavy-edge information network, giving a triple heterogeneous node pair (u, R, v) of a multi-element heavy-edge relation, wherein u and v have heterogeneous types; the heterogeneous nodes u and v have a multiple heavy edge relation, and the multiple heavy edge relation is set as R;
step S2.2
Figure BDA0002518303760000095
Respectively mapping a head node u and a tail node v of the node pair (u, v) to a uniform potential semantic space; the representation of the head node u and the tail node v in the unified space can be expressed as:
Figure BDA0002518303760000096
Figure BDA0002518303760000097
wherein the content of the first and second substances,
Figure BDA0002518303760000101
different types of spaces T for head node u and tail node vu、TvThe learning parameter matrix of (2);
Figure BDA0002518303760000102
vector representations of a head node u and a tail node v, respectively; vector representation of heterogeneous type head node u and tail node v
Figure BDA0002518303760000103
From the space T of the typeu、TvRespectively mapping to a uniform space phi to obtain a vector representation of the uniform space phi
Figure BDA0002518303760000104
Step S2.3, considering the relevance of the head node and the tail node on the multi-element heavy edge relation fusion semantic space, adopting the concept of TransR to map the potential space representation of the head node and the tail node to the multi-element heavy edge relation fusion semantic space, and establishing the representation of the head node and the tail node, namely:
Figure BDA0002518303760000105
Figure BDA0002518303760000106
wherein the content of the first and second substances,
Figure BDA0002518303760000107
a projection parameter matrix of the expression space is fused for the multivariate heavy-edge relation; representing vectors of a head node u and a tail node v in phi space
Figure BDA0002518303760000108
Vector mapping to multiple heavy edge relation fusion expression space
Figure BDA0002518303760000109
To obtain a fused representation space
Figure BDA00025183037600001010
Representation of vectors in
Figure BDA00025183037600001011
Step S3, an embedded model of heterogeneous heavy edges is provided by using a translation mechanism on the multivariate heavy edge relation fusion expression space, and distributed expression of heterogeneous nodes and relations is realized;
s3.1, adopting the idea of knowledge representation translation, and establishing the association between the heterogeneous node pairs (u, v) through a translation mechanism on the multivariate and multiple edge relation fusion representation space, wherein the association is formed as follows:
Figure BDA00025183037600001012
step S3.2, for a triple heterogeneous node pair (u, R, v) with a multiple heavy-edge relationship, a distance function based on the translation mechanism may be defined as:
Figure BDA00025183037600001013
wherein f isR(u, v) represents the fusion of the head node u and the tail node v in the multi-element heavy-edge relation representation space
Figure BDA00025183037600001014
A distance function of (a);
Figure BDA00025183037600001015
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure BDA00025183037600001016
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure BDA0002518303760000111
a fusion representation space of a multivariate heavy edge relation set R representing a head node u and a tail node v;
step S3.3 for the positive sample triplet (u, R, v) and the negative sample triplet (u, R, v) according to the distance function of the triplet heterogeneous node pair (u, R, v)
Figure BDA0002518303760000112
Defining a translation-relationship-based Hinge Loss function as:
Figure BDA0002518303760000113
wherein S is a set of (u, R, v) positive samples, and a head node u and a tail node v have a multiple-element multiple-edge relation R;
Figure BDA0002518303760000114
is a set of negative examples of the negative samples,
Figure BDA0002518303760000115
is a head node based on (u, R, v) replacement, and
Figure BDA0002518303760000116
there is no connection edge relation or multiple edge relation R with v,
Figure BDA0002518303760000117
is a tail node based on (u, R, v) replacement, and u is
Figure BDA0002518303760000118
No continuous edge relation or multiple edge relation R exists, gamma represents boundary parameters, lambda represents a hyper-parameter, η represents a regularization hyper-parameter;
Figure BDA0002518303760000119
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure BDA00025183037600001110
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure BDA00025183037600001111
representing head node
Figure BDA00025183037600001112
Vector representation on the multi-element heavy-edge relation fusion representation space;
Figure BDA00025183037600001113
representing tail nodes
Figure BDA00025183037600001114
Vector representation on the multi-element heavy-edge relation fusion representation space; t isu、TvType spaces of a head node u and a tail node v respectively;
Figure BDA00025183037600001115
and
Figure BDA00025183037600001116
the orthogonality ensures that different types of heterogeneous head nodes and tail nodes can learn different parameter matrixes.
In step S4, node prediction is performed based on the representation of the node and the fused representation of the multi-element multiple edge relationship.
In particular for new node triplets (u, R, v'), adopt
Figure BDA00025183037600001117
And determining the distance between the node v' and the head node u, and sorting in the candidate set according to the distance.
Evaluation of Effect
In order to verify the effectiveness and the advancement of the technical scheme provided by the invention, several existing translation model methods are selected for comparison: TransE and TransR. In addition, we generated new TransE + Multi-edges and TransR + Multi-edges models for comparison by replacing the relationship of the TransE and TransR models with the multivariate heavy-edge relationship fusion representation of the attention mechanism of the present invention. The link prediction results of the method on the partial data set scored by the MovieLens100K movie are evaluated through Hits @ K, and the results are shown in table 1:
TABLE 1
Figure BDA0002518303760000121
The results in the table show that the technical scheme of the invention can obtain the results with better precision and reliability than the existing method when the user carries out the link prediction of the movie recommendation.
Those skilled in the art will appreciate that the invention may be practiced without these specific details. Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (5)

1. A product recommendation method based on a heterogeneous heavy-side information network translation model is characterized by comprising the following steps: the method comprises the following steps:
step S1, performing fusion representation space learning on the multi-element heavy edge relation between the node pairs in the heterogeneous heavy edge information network by adopting an attention mechanism;
step S2, projecting the type information of the heterogeneous nodes to a multivariate heavy-edge relation fusion expression space;
step S3, an embedded model of heterogeneous heavy edges is provided by using a translation mechanism on the multivariate heavy edge relation fusion expression space, and distributed expression of heterogeneous nodes and relations is realized;
in step S4, node prediction is performed based on the fused representation space of the representation of the node and the multiple edge relation.
2. The product recommendation method based on the heterogeneous heavy-side information network translation model according to claim 1, wherein: in the step S1, performing fusion representation space learning on the multivariate heavy edge relationship between the node pairs in the heterogeneous heavy edge information network by using an attention mechanism further includes the following steps:
step S1.1, setting the heterogeneous information network G ═ V, E, where V is a set of nodes and E is a set of edges; in heterogeneous information networks, there is a mapping ψ V → T of a set of nodes V to a set of node types T and a mapping E of a set of edges E to a set of edge types A
Figure FDA0002518303750000011
And | T | + | A | > 2;
step S1.2, setting a heterogeneous heavy-edge information network G ' ═ V ', E ', where V ' is a set of nodes and E ' is a set of edges; there is a mapping ψ 'of the node set V' to the node type set T V '→ T and a mapping of the edge set E' to the edge type set A
Figure FDA0002518303750000012
If two tuples (u, v) ∈ E 'exist at the same time and the number of edges of the associated node pair (u, v) is more than 1, G' is called a heterogeneous heavy-edge information network;
step S1.3, in the heterogeneous heavy-edge information network G '(V', E '), given a bigram (u, V) ∈ E', given a set of multiple heavy-edge relationships between node pairs as R ═ R | R ∈ a }, R representing an arbitrary type of relationship between node pairs (u, V), and R representing a set of different types of relationships between node pairs (u, V), then a set of trigram pairs of multiple heavy-edge relationships is denoted as (u, R, V), and for node pairs (u, V), a set of trigram pairs of multiple heavy-edge relationships is denoted as (u, R, V)Behavioral relationship ri∈ R, the mapping process is defined as:
Figure FDA0002518303750000013
wherein, aiRepresents a certain behavioral relationship r of node pairs (u, v)iVector representation of
Figure FDA0002518303750000021
The linear transformation of (a) to (b),
Figure FDA0002518303750000022
the shared parameter is used for learning different importance of the multiple heavy edge relation;
s1.4, calculating the mapping characteristics of any relation for all behavior relations R of the node pairs (u, v), and pairing the relation R through a softmax functioniThe normalization weight coefficients are:
Figure FDA0002518303750000023
wherein, αiRepresents the relation riWeight coefficient in all behavioral relationships of node pair (u, v), e is a natural constant, akIs the relation r of node pair (u, v)k(k ═ 1,2, ·, | | R | |), where | R | | is the number of elements in the set of multivariate multiple edge relationships R;
s1.5, reflecting different importance of different types of relations in the node pair (u, v) connection by weight coefficients; learning a multivariate heavy-edge relation fusion expression space of the node pairs through the attention weight coefficients, and recording the multivariate heavy-edge relation fusion expression space as:
Figure FDA0002518303750000024
wherein the content of the first and second substances,
Figure FDA0002518303750000025
representing a multiple-heavy-edge relationship of node pairs (u, v)The vector-weighted fusion represents the space of representation,
Figure FDA0002518303750000026
is a relation riIs represented by a vector of (a).
3. The product recommendation method based on the heterogeneous heavy-side information network translation model according to claim 1, wherein: in the step S2, projecting the type information of the heterogeneous node to the multivariate multiple-edge relationship fusion representation space further includes the following steps:
s2.1, in the heterogeneous heavy-edge information network, giving a triple heterogeneous node pair (u, R, v) of a multi-element heavy-edge relation, wherein u and v have heterogeneous types; the heterogeneous nodes u and v have a multiple heavy edge relation, and the multiple heavy edge relation is set as R;
step S2.2
Figure FDA0002518303750000027
Respectively mapping a head node u and a tail node v of the node pair (u, v) to a uniform potential semantic space; the representation of the head node u and the tail node v in the unified space can be expressed as:
Figure FDA0002518303750000028
Figure FDA0002518303750000029
wherein the content of the first and second substances,
Figure FDA0002518303750000031
different types of spaces T for head node u and tail node vu、TvThe learning parameter matrix of (2);
Figure FDA0002518303750000032
vector representations of a head node u and a tail node v, respectively; vector representation of heterogeneous type head node u and tail node v
Figure FDA0002518303750000033
From the space T of the typeu、TvRespectively mapping to a uniform space phi to obtain a vector representation of the uniform space phi
Figure FDA0002518303750000034
Step S2.3, considering the relevance of the head node and the tail node on the multi-element heavy edge relation fusion semantic space, adopting the concept of TransR to map the potential space representation of the head node and the tail node to the multi-element heavy edge relation fusion semantic space, and establishing the representation of the head node and the tail node, namely:
Figure FDA0002518303750000035
Figure FDA0002518303750000036
wherein the content of the first and second substances,
Figure FDA0002518303750000037
a projection parameter matrix of the expression space is fused for the multivariate heavy-edge relation; representing vectors of a head node u and a tail node v in phi space
Figure FDA0002518303750000038
Vector mapping to multiple heavy edge relation fusion expression space
Figure FDA0002518303750000039
To obtain a fused representation space
Figure FDA00025183037500000310
Representation of vectors in
Figure FDA00025183037500000311
4. The product recommendation method based on the heterogeneous heavy-side information network translation model according to claim 1, characterized in that: in the step S3, the method for extracting an embedded model of heterogeneous heavy edges by using a translation mechanism in the multivariate heavy edge relationship fusion expression space to realize the distributed expression of heterogeneous nodes and relationships further includes the following steps:
s3.1, adopting the idea of knowledge representation translation, and establishing the association between the heterogeneous node pairs (u, v) through a translation mechanism on the multivariate and multiple edge relation fusion representation space, wherein the association is formed as follows:
Figure FDA00025183037500000312
step S3.2, for a triple heterogeneous node pair (u, R, v) with a multiple heavy-edge relationship, a distance function based on the translation mechanism may be defined as:
Figure FDA00025183037500000313
wherein f isR(u, v) represents the fusion of the head node u and the tail node v in the multi-element heavy-edge relation representation space
Figure FDA00025183037500000314
A distance function of (a);
Figure FDA00025183037500000315
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure FDA00025183037500000316
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure FDA0002518303750000041
a fusion representation space of a multivariate heavy edge relation set R representing a head node u and a tail node v;
step (ii) ofS3.3 for positive sample triplets (u, R, v) and negative sample triplets according to the distance function of the triple heterogeneous node pair (u, R, v)
Figure FDA0002518303750000042
Defining a translation-relationship-based Hinge Loss function as:
Figure FDA0002518303750000043
wherein S is a set of (u, R, v) positive samples, and a head node u and a tail node v have a multiple-element multiple-edge relation R;
Figure FDA0002518303750000044
is a set of negative examples of the negative samples,
Figure FDA0002518303750000045
is a head node based on (u, R, v) replacement, and
Figure FDA0002518303750000046
there is no connection edge relation or multiple edge relation R with v,
Figure FDA0002518303750000047
is a tail node based on (u, R, v) replacement, and u is
Figure FDA0002518303750000048
No continuous edge relation or multiple edge relation R exists, gamma represents boundary parameters, lambda represents a hyper-parameter, η represents a regularization hyper-parameter;
Figure FDA0002518303750000049
representing the vector representation of the head node u on the multivariate heavy edge relation fusion representation space;
Figure FDA00025183037500000410
representing vector representation of the tail node v on the multivariate heavy edge relation fusion representation space;
Figure FDA00025183037500000411
representing head node
Figure FDA00025183037500000412
Vector representation on the multi-element heavy-edge relation fusion representation space;
Figure FDA00025183037500000413
representing tail nodes
Figure FDA00025183037500000414
Vector representation on the multi-element heavy-edge relation fusion representation space; t isu、TvType spaces of a head node u and a tail node v respectively;
Figure FDA00025183037500000415
and
Figure FDA00025183037500000416
the orthogonality ensures that different types of heterogeneous head nodes and tail nodes can learn different parameter matrixes.
5. The product recommendation method based on the heterogeneous heavy-side information network translation model according to claim 1, characterized in that: in step S4, the method for node prediction based on the fused representation space of the node representation and the multiple heavy edge relation is to adopt the new node triple (u, R, v
Figure FDA00025183037500000417
And determining the distance between the node v' and the head node u, and sorting in the candidate set according to the distance.
CN202010484083.1A 2020-06-01 2020-06-01 Product recommendation method based on heterogeneous heavy side information network translation model Active CN111737591B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010484083.1A CN111737591B (en) 2020-06-01 2020-06-01 Product recommendation method based on heterogeneous heavy side information network translation model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010484083.1A CN111737591B (en) 2020-06-01 2020-06-01 Product recommendation method based on heterogeneous heavy side information network translation model

Publications (2)

Publication Number Publication Date
CN111737591A true CN111737591A (en) 2020-10-02
CN111737591B CN111737591B (en) 2024-03-15

Family

ID=72646617

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010484083.1A Active CN111737591B (en) 2020-06-01 2020-06-01 Product recommendation method based on heterogeneous heavy side information network translation model

Country Status (1)

Country Link
CN (1) CN111737591B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784153A (en) * 2020-12-31 2021-05-11 山西大学 Tourist attraction recommendation method integrating attribute feature attention and heterogeneous type information

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20180092194A (en) * 2017-02-08 2018-08-17 경북대학교 산학협력단 Method and system for embedding knowledge gragh reflecting logical property of relations, recording medium for performing the method
CN110334219A (en) * 2019-07-12 2019-10-15 电子科技大学 The knowledge mapping for incorporating text semantic feature based on attention mechanism indicates learning method
CN110516160A (en) * 2019-08-30 2019-11-29 中国科学院自动化研究所 User modeling method, the sequence of recommendation method of knowledge based map
CN110570869A (en) * 2019-08-09 2019-12-13 科大讯飞股份有限公司 Voiceprint recognition method, device, equipment and storage medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20180092194A (en) * 2017-02-08 2018-08-17 경북대학교 산학협력단 Method and system for embedding knowledge gragh reflecting logical property of relations, recording medium for performing the method
CN110334219A (en) * 2019-07-12 2019-10-15 电子科技大学 The knowledge mapping for incorporating text semantic feature based on attention mechanism indicates learning method
CN110570869A (en) * 2019-08-09 2019-12-13 科大讯飞股份有限公司 Voiceprint recognition method, device, equipment and storage medium
CN110516160A (en) * 2019-08-30 2019-11-29 中国科学院自动化研究所 User modeling method, the sequence of recommendation method of knowledge based map

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
JI-CHAO LI, DAN-LING ZHAO, BING-FENG GE: "A link prediction method for heterogeneous networks based on BP neural network", 《 PHYSICA A: STATISTICAL MECHANICS AND ITS APPLICATIONS》 *
ZIJIAN LI; WENHAO ZHENG; XUELING LIN: "TransN: Heterogeneous Network Representation Learning by Translating Node Embeddings", 《2020 IEEE 36TH INTERNATIONAL CONFERENCE ON DATA ENGINEERING (ICDE)》 *
周俞: "基于用户—词语表示联合学习的情感分析方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *
郑建兴;李沁文;王素格;李德玉: "基于翻译模型的异质重边信息网络链路预测研究", 《南京大学学报(自然科学)》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784153A (en) * 2020-12-31 2021-05-11 山西大学 Tourist attraction recommendation method integrating attribute feature attention and heterogeneous type information

Also Published As

Publication number Publication date
CN111737591B (en) 2024-03-15

Similar Documents

Publication Publication Date Title
Luo et al. Non-negativity constrained missing data estimation for high-dimensional and sparse matrices from industrial applications
CN110162700B (en) Training method, device and equipment for information recommendation and model and storage medium
Luo et al. Personalized recommendation by matrix co-factorization with tags and time information
CN111339415B (en) Click rate prediction method and device based on multi-interactive attention network
CN112529168B (en) GCN-based attribute multilayer network representation learning method
CN112380435B (en) Document recommendation method and system based on heterogeneous graph neural network
CN111541570B (en) Cloud service QoS prediction method based on multi-source feature learning
CN111488734A (en) Emotional feature representation learning system and method based on global interaction and syntactic dependency
CN111079532A (en) Video content description method based on text self-encoder
CN111709518A (en) Method for enhancing network representation learning based on community perception and relationship attention
CN110032679B (en) Dynamic news recommendation method based on hierarchical attention network
CN116542720B (en) Time enhancement information sequence recommendation method and system based on graph convolution network
CN114565053A (en) Deep heterogeneous map embedding model based on feature fusion
CN109348229A (en) Jpeg image mismatch steganalysis method based on the migration of heterogeneous characteristic subspace
CN112784118A (en) Community discovery method and device in graph sensitive to triangle structure
CN115878757A (en) Concept decomposition-based hybrid hypergraph regularization semi-supervised cross-modal hashing method
CN113590965B (en) Video recommendation method integrating knowledge graph and emotion analysis
CN115238191A (en) Object recommendation method and device
Yin et al. An efficient recommendation algorithm based on heterogeneous information network
CN111506832B (en) Heterogeneous object completion method based on block matrix completion
CN111737591A (en) Product recommendation method based on heterogeneous heavy-side information network translation model
CN117194765A (en) Dual-channel graph contrast learning session recommendation method for interest perception
Qi et al. Breaking the barrier to transferring link information across networks
CN111723301B (en) Attention relation identification and labeling method based on hierarchical theme preference semantic matrix
Duan et al. Anonymity can help minority: A novel synthetic data over-sampling strategy on multi-label graphs

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant