CN112613602A - Recommendation method and system based on knowledge-aware hypergraph neural network - Google Patents
Recommendation method and system based on knowledge-aware hypergraph neural network Download PDFInfo
- Publication number
- CN112613602A CN112613602A CN202011567183.7A CN202011567183A CN112613602A CN 112613602 A CN112613602 A CN 112613602A CN 202011567183 A CN202011567183 A CN 202011567183A CN 112613602 A CN112613602 A CN 112613602A
- Authority
- CN
- China
- Prior art keywords
- hypergraph
- user
- knowledge
- article
- super
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a recommendation method based on a knowledge-aware hypergraph neural network, which comprises the following steps of: step one, constructing a user hypergraph, wherein an initial hypergraph takes an article which has an interactive relation with a user as an entity node; constructing an article hypergraph, wherein the initial hyperedges take the articles which have an interactive relationship with any user of the articles as entity nodes; step two, carrying out convolution calculation on the user hypergraph and each article hypergraph; and step three, calculating the inner product of the user and all the articles to obtain the interaction scores of the user and all the articles. The method has the advantages that the auxiliary information of the articles is integrated into the vector representation of the user, so that the articles which are more likely to have interaction relation with the user can be selected from a plurality of articles. Disclosed is a recommendation system based on a knowledge-aware hypergraph neural network, comprising: a data capture module; a knowledge perception hypergraph construction module; a domain convolution module; a super-edge convolution module; and a prediction module. The method has the advantages of being simple to operate and accurate in prediction.
Description
Technical Field
The invention relates to the technical field of knowledge perception application. More specifically, the invention relates to a recommendation method and a recommendation system based on a knowledge-aware hypergraph neural network.
Background
With the rapid development of the internet, recommendation systems are widely deployed to mitigate the effects of information overload. A conventional recommendation technique is collaborative filtering, which assigns a representation vector based on user and item identities and then models their interactions through specific operations (such as inner products or neural networks). However, these collaborative filtering based approaches often suffer from sparsity and cold start problems in the absence of auxiliary information. To address these issues, various types of ancillary information have been explored to improve recommendation performance, such as item attributes, item reviews, and the user's social network.
In order to alleviate the problems of sparsity and cold start of the traditional recommendation system, the knowledge-sensing map is widely researched and applied in the field of recommendation systems as auxiliary information. However, most existing knowledge-aware graph-based recommendation methods suffer from the disadvantages that higher-order correlations between users, items and entities are modeled insufficiently, and that simple aggregation strategies cannot preserve relationship information in the neighborhood.
Knowledge-graphs have received increasing attention in recommendation system research in recent years due to their flexibility in modeling integrated assistance data. The key to the fusion of knowledge-graph and recommendation systems is how to efficiently integrate ancillary information into the vector representations of users and items. Existing knowledge-graph based recommendation systems can be divided into two categories, path-based and graph-based neural network approaches, depending on the way this challenge is addressed. Path-based approaches infer user preferences by exploring multiple meta-paths between a target user and an item on a knowledge graph, which typically requires domain knowledge. More importantly, this type of approach ignores the rich structural information that is implicit in the knowledge-graph and does not adequately describe the underlying relationship between a given target user and the item. Inspired by the recent emergence of graph neural networks, existing graph neural network-based approaches provide significant performance by explicitly modeling higher-order connectivity in a knowledge graph. However, these methods still suffer from two limitations, (1) high-order correlations between users, items and entities in the knowledge-graph are essential for data modeling. These methods mainly apply graph neural networks to enrich the representation of the target node by recursively aggregating their original neighbors in the knowledge-graph, thus limiting the modeling of higher order correlations between the target node and non-original neighbors. Furthermore, the graph structure employed by existing methods places constraints on higher order correlation modeling and processing, because graphs can only represent pairs of connections. (2) In neighborhood aggregation, simple aggregation functions (such as mean and max pooling) are typically performed as aggregation functions to generate neighborhood embedding. This aggregation destroys rich relationship information between nodes in the neighborhood. Neighborhood aggregation therefore ignores the fine structure of relationships between nodes in the neighborhood.
Disclosure of Invention
An object of the present invention is to solve at least the above problems and to provide at least the advantages described later.
It is still another object of the present invention to provide a recommendation method based on knowledge-aware hypergraph neural network, which can overcome the sparsity and cold start problems of the conventional recommendation system, and the knowledge-aware hypergraph convolution method can effectively aggregate different neighbors in the neighborhood while preserving the relationship information in the neighborhood.
The recommendation system based on the knowledge-aware hypergraph neural network can overcome the sparsity and cold start problems of the traditional recommendation system, effectively aggregates different neighbors in a neighborhood and simultaneously retains relationship information in the neighborhood.
To achieve these objects and other advantages in accordance with the purpose of the invention, there is provided a recommendation method based on a knowledge-aware hypergraph neural network, comprising the steps of:
step one, constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a user, wherein l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge sensing hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge sensing hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraph respectively correspond to the article and the auxiliary information of the article;
constructing an article hypergraph, wherein the article hypergraph comprises an initial hyperedge of an article and 2-l-order knowledge perception hyperedges, l is an integer larger than 1, the initial hyperedge takes the article which has an interactive relationship with any user of the article as an entity node, the 2 nd-order knowledge perception hyperedge takes the auxiliary information of the article corresponding to the entity node of the initial hyperedge as an entity node, the article or the auxiliary information of the article corresponding to the entity node of the previous-order knowledge perception hyperedge of the next-order knowledge perception hyperedge as an entity node, and the entity nodes of the two adjacent-order knowledge perception hyperedges respectively correspond to the auxiliary information of the article and the auxiliary information of the article;
performing convolution calculation on the user hypergraph and each article hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each article hypergraph;
and step three, calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the interaction between the user and the articles.
Preferably, each initial hyper-edge and each knowledge perception hyper-edge of the user hyper-graph are processed by adopting a domain convolution method to obtain a single vector of each hyper-edge of the user hyper-graph;
and processing each knowledge perception hyper-edge of the object hyper-graph by adopting a domain convolution method to obtain a single vector of each hyper-edge of the object hyper-graph.
Preferably, the single vectors of all the super edges of the user hypergraph are processed by adopting a super edge convolution method to obtain the unique single vector of the user hypergraph;
and processing the single vectors of all knowledge perception super edges of the hypergraph of the article by adopting a super edge convolution method to obtain the unique single vector of the article.
Preferably, the domain convolution method specifically adopts the following method to calculate:
putting the entity vectors in the order l of the super edge and the order l-1 of the super edge into one-dimensional convolution conv1Generating a transformation matrix T, then using the transformation matrix T to perform permutation and weighting on entity vectors in the l-order super-edge to obtain a plurality of transformed super-edge vectors, and then using one-dimensional convolution conv2Polymerizing the transformed super-edge vector to obtain a single vector of a 1-order super-edge, wherein l is an integer greater than 1, and the initial super-edge is a 1-order super-edge;
for the domain convolution of the initial super edge, only the entity vector in the initial super edge is put into the one-dimensional convolution conv1A transformation matrix T is generated. Preferably, for a user hypergraph: the super-edge convolution method adopts a connection aggregator to aggregate all super-edge single vectors of the user super-graph to unique single vectors corresponding to users and articles;
hypergraph for any item: a connection aggregator is employed to aggregate all knowledge-aware hyperedge univariate vectors of the item hypergraph, as well as the initial univariate vector of the item itself.
Preferably, the method further comprises the following steps: of all the items used for the calculation, the same number of positive and negative examples were chosen for each user, the negative examples being randomly drawn from the items that have not interacted with the user, the positive examples being the items that have interacted with the user.
Preferably, the method further comprises sorting the interaction scores from large to small, and outputting the items corresponding to the top interaction score or scores.
A recommendation system based on a knowledge-aware hypergraph neural network is provided, including:
the data capture module is used for acquiring information of the user and the object interacted with the user;
the knowledge perception hypergraph construction module is used for constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge perception hypergraph, l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge perception hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge perception hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge perception hypergraph as the entity node, and the entity nodes of two adjacent order knowledge perception hypergraph respectively correspond to the article and the auxiliary information of the article;
and the method is used for constructing the commodity hypergraph, wherein the commodity hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a commodity, l is an integer larger than 1, the initial hypergraph takes the commodity which has an interactive relation with any user of the commodity as an entity node, the 2 order knowledge sensing hypergraph takes the auxiliary information of the commodity corresponding to the entity node of the initial hypergraph as the entity node, the auxiliary information of the commodity or the auxiliary information of the commodity corresponding to the entity node of the previous order knowledge sensing hypergraph of the next order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraphs respectively correspond to the auxiliary information of the commodity and the auxiliary information of the commodity;
the domain convolution module is used for processing each initial hyper-edge and each knowledge perception hyper-edge of the user hyper-graph by adopting a domain convolution method to obtain a single vector of each hyper-edge of the user hyper-graph;
and processing each knowledge perception hyperedge of the object hypergraph by adopting a domain convolution method to obtain a single vector of each hyperedge of the object hypergraph;
the super-edge convolution module is used for processing single vectors of all super edges of the user super-graph by adopting a super-edge convolution method to obtain a unique single vector of the user super-graph;
processing the single vectors of all knowledge perception hyperedges of the hypergraph of the article by adopting a hyper-edge convolution method to obtain initial single vectors of the article;
and the prediction module is used for calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the user interacting with the articles.
The invention at least comprises the following beneficial effects:
first, the invention provides a knowledge-aware hypergraph neural network framework to solve the sparsity and cold start problems of the traditional recommendation system. First, a knowledge-aware hypergraph structure composed of hyperedges is used to explicitly model the higher-order relevance of users, items, and entities in a knowledgegraph. Secondly, a new knowledge-aware hypergraph convolution method is proposed to effectively aggregate different neighbors in the neighborhood and simultaneously retain the relationship information in the neighborhood.
Second, the present invention proposes to use a knowledge-aware hypergraph to explicitly model higher-order correlations between users, items and entities in a knowledge-aware graph to achieve more accurate and efficient item recommendations.
Thirdly, in order to solve the limitation of the existing recommendation method based on the knowledge graph so as to recommend suitable articles to the user, the invention provides an end-to-end model, namely a knowledge-aware hypergraph neural network. In the model, we first construct initial hyperedges of users and items based on historical interaction records (such as purchases, clicks, comments) of the users, respectively, and then derive subsequent hyperedges of knowledge based on the previous initial hyperedges and knowledge maps related to the items. These hyper-edges associated with the user and the item constitute a hyper-graph of the user and the item, respectively. Then in the hypergraph convolution, a plurality of nodes (represented by vectors) in the hyperedges are firstly aggregated into a single vector through domain convolution, and then the represented vector of each hyperedge obtained before is aggregated into a single vector through the hyperedge convolution. Thus we get a single vector representation of the user and item corresponding hypergraphs, respectively. Finally, vectors of the hypergraphs corresponding to the user and the articles are placed in a prediction layer to obtain the probability of clicking the articles by the user, so that the user can select the articles with high user clicking probability and recommend the articles to the user.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Drawings
FIG. 1 is a block diagram of the knowledge-aware hypergraph neural network model according to one embodiment of the present invention;
fig. 2 is a schematic diagram of the domain convolution module according to one embodiment of the present invention.
Detailed Description
The present invention is further described in detail below with reference to the attached drawings so that those skilled in the art can implement the invention by referring to the description text.
It is to be noted that the experimental methods described in the following embodiments are all conventional methods unless otherwise specified, and the reagents and materials, if not otherwise specified, are commercially available; in the description of the present invention, the terms indicating orientation or positional relationship are based on the orientation or positional relationship shown in the drawings only for the convenience of description and simplification of description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus, should not be construed as limiting the present invention.
As shown in fig. 1-2, the invention provides a recommendation method based on a knowledge-aware hypergraph neural network, comprising the following steps:
step one, constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a user, wherein l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge sensing hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge sensing hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraph respectively correspond to the article and the auxiliary information of the article;
constructing an article hypergraph, wherein the article hypergraph comprises an initial hyperedge of an article and 2-l-order knowledge perception hyperedges, l is an integer larger than 1, the initial hyperedge takes the article which has an interactive relationship with any user of the article as an entity node, the 2-order knowledge perception hyperedge takes the auxiliary information of the article corresponding to the entity node of the initial hyperedge as an entity node, the article or the auxiliary information of the article corresponding to the entity node of the previous-order knowledge perception hyperedge of the next-order knowledge perception hyperedge as an entity node, and the entity nodes of the two adjacent-order knowledge perception hyperedges respectively correspond to the article and the auxiliary information of the article;
performing convolution calculation on the user hypergraph and each article hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each article hypergraph;
and step three, calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the interaction between the user and the articles.
In the above technical solution, the knowledge-aware hypergraph neural network. In the model, we first construct initial super edges of users and articles based on historical interaction records (such as purchases, clicks and comments) of the users, and then obtain subsequent knowledge-aware super edges based on the previous initial super edges and knowledge-aware maps related to the articles. These hyper-edges associated with the user and the item constitute a hyper-graph of the user and the item, respectively. Then in the hypergraph convolution, a plurality of nodes (represented by vectors) in the hyperedges are firstly aggregated into a single vector through domain convolution, and then the represented vector of each hyperedge obtained before is aggregated into a single vector through the hyperedge convolution. Thus we get a single vector representation of the user and item corresponding hypergraphs, respectively. Finally, vectors of hypergraphs corresponding to the user and the articles are put into a prediction layer to obtain the probability of clicking the articles by the user, namely, the inner product of a single vector of the user and single vectors of all articles is calculated, then nonlinear transformation is carried out to obtain the interaction scores of the user and all articles, and the interaction scores represent the probability of interaction between the user and the articles. Therefore, the user can select the items which are more likely to have interaction relation with the user from the plurality of items and recommend the items to the user. By the method, the auxiliary information of the articles can be effectively integrated into the vector representation of the user, so that the articles which are more likely to have interaction relation with the user can be selected from a plurality of articles.
In order to better learn the expression vectors of the target user and the target item to obtain a more accurate recommendation result, an end-to-end model KHNN (knowledge-aware hyper-graph neural network) is provided to explicitly model the high-order correlation among the user, the item and the entity in the knowledge-aware map so as to realize more accurate and efficient item recommendation for the user. In the model we first build initial hyperedges for the user and the item, respectively, based on the user's historical interaction records (e.g., purchases, clicks, comments).
One implementation method comprises the following steps: let U be { U ═ U1,u2,…,uMIs a set of all users and V ═ V1,v2,…,vNIs a collection of owned items. M and N are the numbers of users and articles respectively. We define the user-item interaction matrix as Y ∈ RM×NWherein y isuv1 means that user u interacted with item v, such as clicking, browsing, or purchasing; otherwise yuv0. In addition, we define the knowledge-aware graph as GK{ (h, R, t) | h, t ∈ E, R ∈ R }, where each triplet (h, R, t) represents the existence of a relationship R between the head entity h and the tail entity t, and ε and R are the set of entities and relationships in the knowledge-aware graph. For example, a triplet (Back to the Future, direcdby, Robert Zemeckis) indicates that Robert Zemeckis is the director of the movie Back to the Future. In a conventional graph structure, an edge connects only two vertices, whereas in a hypergraph, a hyperedge may connect two or more vertices. We define the knowledge-aware hypergraph as GHAnd { epsilon, H }, wherein epsilon represents the set of entities and H represents the set of super-edges. We use the set A { (V, E) | V ∈ V, E ∈ E } to align item V with entity E in the knowledge-aware hypergraph. Given a user-item interaction matrix Y and a knowledge-aware hypergraph GHWe aim to predict whether user u has potential interest in items v that he has not interacted with before. Our goal is to learn a prediction function
Wherein
Representing the probability of user u interacting with item v, and theta represents the model parameters of function F.
Establishing an initial excess edge:
generally, the items that the user has interacted with can represent the user's preferences to some extent. Thus, we build the initial hyperedge of the user by the interacted-with items of user u. The related item set of the user u obtained through the historical interaction of the user u and the items in the set are aligned with the entities in the knowledge-aware graph through the set A (namely, the entities with the same ID in the knowledge-aware graph are found through the ID of the items), and the entities can be constructed as the initial super edge (namely, the set of the entities) of the user. The initial hyperedge for user u is defined as follows:
similarly, an initial superedge of an item, a user who has interacted with the same item, may also contribute to the vector representation of the item. We define the items that have been interacted with by these users as the initial neighbors of item v, and formulate the initial set of neighbors for item v as follows:
then, the items in the initial neighbor set of the item v are aligned with the entities in the knowledge-aware graph through the set a, and then these entities can form the initial super edge of the item v, which is defined as follows:
inputting in the model: historical interaction records for a user
Generating: initial overcenter for users and items
The implementation process comprises the following steps: a normal distribution can be used to initialize a random vector for all items separately to contain the intrinsic characteristics of each item. Then, through the historical interaction records of the users, initial super edges (1-step super edges) can be constructed for the target users and the target objects respectively.
Establishing knowledge perception super-edge:
neighboring entities in the knowledge-aware graph always have strong associations. Propagating this association along the links between entities in the knowledge-aware graph (relationships r in triples), we can obtain an extended hyperedge, i.e., a knowledge-aware hyperedge, that is at a different distance from the original hyperedge, which can effectively extend the vector representation of users and items. The definition of the knowledge super-edge (i order super-edge) for user u and item v is recursively expressed as:
where l represents the distance from the initial overcide and the subscript o is a uniform placeholder for the symbol u or v. Building our model using knowledge-aware maps as side information is very meaningful, as neighboring entities can be viewed as extensions of user preferences and item features. Based on these generated initial and knowledge-aware hyperedges, we can build their corresponding hypergraphs (consisting of the hyperedges) for the target user and target item, respectively, as shown in FIG. 1.
Inputting in the model: initial overcenter for users and items
Generating: knowledge overcrowding of users and items
The implementation process comprises the following steps: the entity nodes in the initial super-edge obtained before are used as starting points, the subsequent entity nodes connected with the entity nodes in the initial super-edge are obtained on the knowledge perception map through the links among the entities, the subsequent entity nodes form the knowledge super-edge with the distance of one from the initial super-edge, then the knowledge perception super-edge with the distance of two from the initial super-edge is obtained through the links in the knowledge perception map by using the previously obtained knowledge perception super-edge as the starting points, a plurality of super-edges can be respectively constructed for the user and the article by repeating the construction process, and the super-edges respectively form the super-maps of the user and the article.
In another technical scheme, the domain convolution method specifically adopts the following method to calculate:
putting the entity vectors in the order l of the super edge and the order l-1 of the super edge into one-dimensional convolution conv1GeneratingTransforming the matrix T, then using the transformation matrix T to perform permutation and weighting on the entity vectors in the l-order super-edge to obtain a plurality of transformed super-edge vectors, and then using one-dimensional convolution conv2Polymerizing the transformed super-edge vector to obtain a single vector of a 1-order super-edge, wherein l is an integer greater than 1, and the initial super-edge is a 1-order super-edge;
for the domain convolution of the initial super edge, only the entity vector in the initial super edge is put into the one-dimensional convolution conv1A transformation matrix T is generated.
A neighborhood convolution is applied to cluster several entity vectors in the super edge into a single vector representation. The scheme provides a novel neighborhood convolution method. In particular, in triplets of knowledge-aware atlases, each tail entity and a different head entity have different meanings and potential vector representations when connected by a relationship. Therefore, we put the entity vectors in order l and order l-1 superconcentration into one-dimensional convolution conv1And generating a transformation matrix T, and then using the transformation matrix T to perform permutation and weighting on vectors in the l-order super-edge to obtain a plurality of transformed super-edge vectors. Finally using a one-dimensional convolution conv2The transformed super-edge vectors are aggregated to obtain a single vector representation of the l-order super-edge, as shown in the following formula.
Where the subscript o is a uniform placeholder for the symbol u or v, | | is a concatenation operation. conv1And conv2Are all one-dimensional convolutions but have different output channels.
Since the initial hyper-edge has a strong connection with the target user and the item. Thus, a single vector representation of the initial superedge is added for target user u and item v, as described by the following equation.
The item v has its associated entities originally represented, but user u does not, the initial superedge of the item supergraph. The original related entity is the node in the latent semantic space that is closest to the item v itself. Therefore, we add it to the set of representations for item v and define as follows:
after neighborhood convolution, we formulate the set of vectors for the super-edge representation and the original representation of the target item v and the set of super-edge representation vectors for the target user u as follows:
inputting: all the superedges of the user and the item.
Generating: a single vector representation of all superedges of the user and the item.
The implementation process comprises the following steps: taking the node vectors in two adjacent super edges in the hypergraph as one-dimensional convolution conv1Obtaining a transformation matrix T, multiplying the vector matrix of the super edge to be transformed by the transformation matrix T to obtain a transformed vector matrix of the super edge, and finally taking the transformed vector matrix of the super edge as a one-dimensional convolution conv2The input of (a) results in a single vector representation of the final hyper-edge.
In another technical scheme, for a user hypergraph: the super-edge convolution method adopts a connection aggregator to aggregate all super-edge single vectors of the user super-graph to unique single vectors corresponding to users and articles;
hypergraph for any item: a connection aggregator is employed to aggregate all knowledge-aware hyperedge univariate vectors of the item hypergraph, as well as the initial univariate vector of the item itself. And for the target article, all the entities in the knowledge-perception map have an initial vector, the initial vector of the article is the expression vector of the entity, and the initial single vector of the article is added during aggregation.
The super-edge convolution aggregates the representation vectors of all super-edges in the super-graph for user u and item v, respectively, into one single representation vector, as shown in FIG. 1. We use a connection aggregator to aggregate
And
multiple super edges in (a) represent a vector to a single vector. The connection aggregator formula is as follows:
wherein
And | is a concatenation operation.
Inputting: all the superedges of the user and the item represent vectors.
Generating: final single representation vector of user and article
The implementation process comprises the following steps: and aggregating all vectors in the hypergraphs of the user u and the article v into a single vector in a connection mode respectively, and finally forming the single vector obtained after aggregation into the expression vectors of the user and the article.
In another technical solution, the method further comprises: of all the items used for the calculation, the same number of positive and negative examples were chosen for each user, the negative examples being randomly drawn from the items that have not interacted with the user, the positive examples being the items that have interacted with the user.
With euTo represent a representation vector of the user, evA representation vector representing the item. Finally, the inner product of the user and the item vector is calculated, and then a non-linear transformation is carried out to predict the preference score of the user for the item:
in order to balance the number of positive and negative samples and ensure the effect of model training, the same number of negative samples as the positive samples are extracted for each user, and the negative samples are randomly drawn from the non-interacted articles of the user. Finally, the loss function of the model is as follows:
where J is the cross entropy loss, p+Denotes a positive sample, p-Representing negative examples. The theta is a parameter of the model,
is the L2 regularization term parameterized by λ.
And (3) candidate set recommendation:
inputting: user and item representation vectors by hypergraph convolution
Generating: the user ultimately scores a preference for the candidate item.
The implementation process comprises the following steps: after the expression vectors of the user and the item are obtained, the inner product of the user and the item vector is calculated, and then nonlinear transformation is carried out to predict the preference score of the user for the item. And then, calculating scores for different candidate items in the candidate set by using corresponding user representations, and sequencing the calculated scores to obtain a final result. Model prediction can be made more accurate.
And the method also comprises the steps of sorting the interaction scores from large to small and outputting the items corresponding to the interaction score or scores at the top.
A recommendation system based on a knowledge-aware hypergraph neural network is provided, including:
the data capture module is used for acquiring information of the user and the object interacted with the user;
the knowledge perception hypergraph construction module is used for constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge perception hypergraph, l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge perception hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge perception hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge perception hypergraph as the entity node, and the entity nodes of two adjacent order knowledge perception hypergraph respectively correspond to the article and the auxiliary information of the article;
and the method is used for constructing the commodity hypergraph, wherein the commodity hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a commodity, l is an integer larger than 1, the initial hypergraph takes the commodity which has an interactive relation with any user of the commodity as an entity node, the 2 order knowledge sensing hypergraph takes the auxiliary information of the commodity corresponding to the entity node of the initial hypergraph as the entity node, the auxiliary information of the commodity or the auxiliary information of the commodity corresponding to the entity node of the previous order knowledge sensing hypergraph of the next order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraphs respectively correspond to the auxiliary information of the commodity and the auxiliary information of the commodity;
the domain convolution module is used for processing each initial hyper-edge and each knowledge perception hyper-edge of the user hyper-graph by adopting a domain convolution method to obtain a single vector of each hyper-edge of the user hyper-graph;
and processing each knowledge perception hyperedge of the object hypergraph by adopting a domain convolution method to obtain a single vector of each hyperedge of the object hypergraph;
the super-edge convolution module is used for processing single vectors of all super edges of the user super-graph by adopting a super-edge convolution method to obtain a unique single vector of the user super-graph;
processing all knowledge perception super-edge single vectors of the object super graph and the initial single vector of the object by adopting a super-edge convolution method to obtain a unique single vector of the object; and for the target article, all the entities in the knowledge-perception map have an initial vector, the initial vector of the article is the expression vector of the entity, and the initial single vector of the article is added during aggregation.
And the prediction module is used for calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the user interacting with the articles.
In the technical scheme, a user hypergraph and an article hypergraph are constructed according to historical interaction relations between users and articles, in the construction process, auxiliary information of the articles and the articles is used as links to form links between knowledge perception maps, then hypergraph convolution (domain convolution and super-edge convolution) is carried out to enable each user and each article to have single vector representations, the vector representations can accurately reverse interaction possibilities between the users and the articles (such as preferences and needs of the users on article types, preferences of article characteristics and the like), then the inner products of the single vectors of the users and the single vectors of all the articles are calculated, nonlinear transformation is carried out, interaction scores of the users and all the articles are obtained, and therefore the articles with high interaction scores are conveniently recommended to the users to meet the needs of the users, and prediction accuracy is improved.
The data capture module, the knowledge perception hypergraph construction module, the field convolution module, the super-edge convolution module and the prediction module form a knowledge perception hypergraph neural network model together, and after the model is trained, only historical interaction information of a user and an article needs to be input into the model, and the article with the top rank can be output to the user.
While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable in various fields of endeavor to which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their equivalents.
Claims (8)
1. The recommendation method based on the knowledge-aware hypergraph neural network is characterized by comprising the following steps of:
step one, constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a user, wherein l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge sensing hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge sensing hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraph respectively correspond to the article and the auxiliary information of the article;
constructing an article hypergraph, wherein the article hypergraph comprises an initial hyperedge of an article and 2-l-order knowledge perception hyperedges, l is an integer larger than 1, the initial hyperedge takes the article which has an interactive relationship with any user of the article as an entity node, the 2-order knowledge perception hyperedge takes the auxiliary information of the article corresponding to the entity node of the initial hyperedge as an entity node, the article or the auxiliary information of the article corresponding to the entity node of the previous-order knowledge perception hyperedge of the next-order knowledge perception hyperedge as an entity node, and the entity nodes of the two adjacent-order knowledge perception hyperedges respectively correspond to the article and the auxiliary information of the article;
performing convolution calculation on the user hypergraph and each article hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each article hypergraph;
and step three, calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the interaction between the user and the articles.
2. The knowledge-aware hypergraph neural network-based recommendation method of claim 1, wherein each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph are processed by a domain convolution method to obtain a single vector of each hyperedge of the user hypergraph;
and processing each knowledge perception hyper-edge of the object hyper-graph by adopting a domain convolution method to obtain a single vector of each hyper-edge of the object hyper-graph.
3. The recommendation method based on the knowledge-aware hypergraph neural network of claim 2,
processing single vectors of all super edges of the user hypergraph by adopting a super edge convolution method to obtain a unique single vector of the user hypergraph;
and processing the single vectors of all knowledge perception super edges of the hypergraph of the article by adopting a super edge convolution method to obtain the unique single vector of the article.
4. The recommendation method based on the knowledge-aware hypergraph neural network of claim 2, wherein the domain convolution method is specifically calculated by adopting the following method:
putting the entity vectors in the order l of the super edge and the order l-1 of the super edge into one-dimensional convolution conv1Generating a transformation matrix T, then using the transformation matrix T to perform permutation and weighting on entity vectors in the l-order super-edge to obtain a plurality of transformed super-edge vectors, and then using one-dimensional convolution conv2Polymerizing the transformed super-edge vector to obtain a single vector of a 1-order super-edge, wherein l is an integer greater than 1, and the initial super-edge is a 1-order super-edge;
for the domain convolution of the initial super edge, only the entity vector in the initial super edge is put into the one-dimensional convolution conv1A transformation matrix T is generated.
5. The knowledge-aware hypergraph neural network-based recommendation method of claim 1, wherein for a user hypergraph: the super-edge convolution method adopts a connection aggregator to aggregate all super-edge single vectors of the user super-graph to unique single vectors corresponding to users and articles;
hypergraph for any item: a connection aggregator is employed to aggregate all knowledge-aware hyperedge univariate vectors of the item hypergraph, as well as the initial univariate vector of the item itself.
6. The knowledge-aware hypergraph neural network-based recommendation method of claim 1, further comprising: of all the items used for the calculation, the same number of positive and negative examples were chosen for each user, the negative examples being randomly drawn from the items that have not interacted with the user, the positive examples being the items that have interacted with the user.
7. The knowledge-aware hypergraph neural network-based recommendation method of claim 1, further comprising sorting the interaction scores from large to small and outputting items corresponding to the top-ranked one or more interaction scores.
8. Recommendation system based on knowledge-aware hypergraph neural network, characterized by comprising:
the data capture module is used for acquiring information of the user and the object interacted with the user;
the knowledge perception hypergraph construction module is used for constructing a user hypergraph, wherein the user hypergraph comprises an initial hypergraph and 2-l order knowledge perception hypergraph, l is an integer larger than 1, the initial hypergraph takes an article which has an interactive relation with the user as an entity node, the 2 order knowledge perception hypergraph takes auxiliary information of the article corresponding to the entity node of the initial hypergraph as the entity node, the knowledge perception hypergraph of the next order is based on the article or the auxiliary information of the article corresponding to the entity node of the previous order knowledge perception hypergraph as the entity node, and the entity nodes of two adjacent order knowledge perception hypergraph respectively correspond to the article and the auxiliary information of the article;
and the method is used for constructing the commodity hypergraph, wherein the commodity hypergraph comprises an initial hypergraph and 2-l order knowledge sensing hypergraph of a commodity, l is an integer larger than 1, the initial hypergraph takes the commodity which has an interactive relation with any user of the commodity as an entity node, the 2 order knowledge sensing hypergraph takes the auxiliary information of the commodity corresponding to the entity node of the initial hypergraph as the entity node, the auxiliary information of the commodity or the auxiliary information of the commodity corresponding to the entity node of the previous order knowledge sensing hypergraph of the next order knowledge sensing hypergraph as the entity node, and the entity nodes of two adjacent order knowledge sensing hypergraphs respectively correspond to the auxiliary information of the commodity and the auxiliary information of the commodity;
the domain convolution module is used for processing each initial hyper-edge and each knowledge perception hyper-edge of the user hyper-graph by adopting a domain convolution method to obtain a single vector of each hyper-edge of the user hyper-graph;
and processing each knowledge perception hyperedge of the object hypergraph by adopting a domain convolution method to obtain a single vector of each hyperedge of the object hypergraph;
the super-edge convolution module is used for processing single vectors of all super edges of the user super-graph by adopting a super-edge convolution method to obtain a unique single vector of the user super-graph;
processing all knowledge perception super-edge single vectors of the object super graph and the initial single vector of the object by adopting a super-edge convolution method to obtain a unique single vector of the object;
and the prediction module is used for calculating the inner product of the single vector of the user and the single vectors of all the articles, and then carrying out nonlinear transformation to obtain the interaction scores of the user and all the articles, wherein the interaction scores represent the probability of the user interacting with the articles.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011567183.7A CN112613602A (en) | 2020-12-25 | 2020-12-25 | Recommendation method and system based on knowledge-aware hypergraph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011567183.7A CN112613602A (en) | 2020-12-25 | 2020-12-25 | Recommendation method and system based on knowledge-aware hypergraph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112613602A true CN112613602A (en) | 2021-04-06 |
Family
ID=75247867
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011567183.7A Pending CN112613602A (en) | 2020-12-25 | 2020-12-25 | Recommendation method and system based on knowledge-aware hypergraph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112613602A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113177159A (en) * | 2021-05-11 | 2021-07-27 | 清华大学 | Binding recommendation method based on multichannel hypergraph neural network |
| CN113408706A (en) * | 2021-07-01 | 2021-09-17 | 支付宝(杭州)信息技术有限公司 | Method and device for training user interest mining model and user interest mining |
| CN113434692A (en) * | 2021-06-22 | 2021-09-24 | 上海交通大学医学院附属仁济医院 | Method, system and equipment for constructing graph neural network model and recommending diagnosis and treatment scheme |
| CN113434778A (en) * | 2021-07-20 | 2021-09-24 | 陕西师范大学 | Recommendation method based on regularization framework and attention mechanism |
| CN113505305A (en) * | 2021-05-11 | 2021-10-15 | 清华大学 | Collaborative filtering recommendation method and system based on decoupling type two-channel hypergraph neural network |
| CN113590946A (en) * | 2021-07-26 | 2021-11-02 | 西安工程大学 | Network resource recommendation method and device based on knowledge graph entity anchoring |
| CN114491055A (en) * | 2021-12-10 | 2022-05-13 | 浙江辰时科技集团有限公司 | Recommendation algorithm based on knowledge graph |
| CN114547325A (en) * | 2022-01-14 | 2022-05-27 | 北京帝测科技股份有限公司 | Probabilistic hypergraph-driven geoscience knowledge graph reasoning optimization system and method |
| CN114817663A (en) * | 2022-05-05 | 2022-07-29 | 杭州电子科技大学 | Service modeling and recommendation method based on class perception graph neural network |
| CN114817568A (en) * | 2022-04-29 | 2022-07-29 | 武汉科技大学 | Knowledge hypergraph link prediction method combining attention mechanism and convolutional neural network |
| WO2023024017A1 (en) * | 2021-08-26 | 2023-03-02 | Ebay Inc. | Multi-modal hypergraph-based click prediction |
| CN116664240A (en) * | 2023-06-05 | 2023-08-29 | 山东睿政信息科技有限公司 | Shopping behavior prediction method, system, storage medium and terminal based on graphic neural network |
-
2020
- 2020-12-25 CN CN202011567183.7A patent/CN112613602A/en active Pending
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113505305A (en) * | 2021-05-11 | 2021-10-15 | 清华大学 | Collaborative filtering recommendation method and system based on decoupling type two-channel hypergraph neural network |
| CN113177159A (en) * | 2021-05-11 | 2021-07-27 | 清华大学 | Binding recommendation method based on multichannel hypergraph neural network |
| CN113434692A (en) * | 2021-06-22 | 2021-09-24 | 上海交通大学医学院附属仁济医院 | Method, system and equipment for constructing graph neural network model and recommending diagnosis and treatment scheme |
| CN113434692B (en) * | 2021-06-22 | 2023-08-01 | 上海交通大学医学院附属仁济医院 | Method, system and equipment for constructing graphic neural network model and recommending diagnosis and treatment scheme |
| CN113408706A (en) * | 2021-07-01 | 2021-09-17 | 支付宝(杭州)信息技术有限公司 | Method and device for training user interest mining model and user interest mining |
| CN113408706B (en) * | 2021-07-01 | 2022-04-12 | 支付宝(杭州)信息技术有限公司 | Method and device for training user interest mining model and user interest mining |
| CN113434778A (en) * | 2021-07-20 | 2021-09-24 | 陕西师范大学 | Recommendation method based on regularization framework and attention mechanism |
| CN113590946A (en) * | 2021-07-26 | 2021-11-02 | 西安工程大学 | Network resource recommendation method and device based on knowledge graph entity anchoring |
| WO2023024017A1 (en) * | 2021-08-26 | 2023-03-02 | Ebay Inc. | Multi-modal hypergraph-based click prediction |
| CN114491055B (en) * | 2021-12-10 | 2022-11-08 | 浙江辰时科技集团有限公司 | Recommendation method based on knowledge graph |
| CN114491055A (en) * | 2021-12-10 | 2022-05-13 | 浙江辰时科技集团有限公司 | Recommendation algorithm based on knowledge graph |
| CN114547325A (en) * | 2022-01-14 | 2022-05-27 | 北京帝测科技股份有限公司 | Probabilistic hypergraph-driven geoscience knowledge graph reasoning optimization system and method |
| CN114817568A (en) * | 2022-04-29 | 2022-07-29 | 武汉科技大学 | Knowledge hypergraph link prediction method combining attention mechanism and convolutional neural network |
| CN114817568B (en) * | 2022-04-29 | 2024-05-10 | 武汉科技大学 | Knowledge hypergraph link prediction method combining attention mechanism and convolutional neural network |
| CN114817663B (en) * | 2022-05-05 | 2023-02-17 | 杭州电子科技大学 | Service modeling and recommendation method based on class perception graph neural network |
| CN114817663A (en) * | 2022-05-05 | 2022-07-29 | 杭州电子科技大学 | Service modeling and recommendation method based on class perception graph neural network |
| CN116664240A (en) * | 2023-06-05 | 2023-08-29 | 山东睿政信息科技有限公司 | Shopping behavior prediction method, system, storage medium and terminal based on graphic neural network |
| CN116664240B (en) * | 2023-06-05 | 2024-02-02 | 谢馥聪 | Shopping behavior prediction method, system, storage medium and terminal based on graphic neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112613602A (en) | Recommendation method and system based on knowledge-aware hypergraph neural network | |
| CN110955834B (en) | Knowledge graph driven personalized accurate recommendation method | |
| CN112989064B (en) | Recommendation method for aggregating knowledge graph neural network and self-adaptive attention | |
| Pan et al. | Study on convolutional neural network and its application in data mining and sales forecasting for E-commerce | |
| CN104731962B (en) | Friend recommendation method and system based on similar corporations in a kind of social networks | |
| CN112232925A (en) | Method for carrying out personalized recommendation on commodities by fusing knowledge maps | |
| CN111737535B (en) | Network characterization learning method based on element structure and graph neural network | |
| CN103353872B (en) | A kind of teaching resource personalized recommendation method based on neutral net | |
| Zhang et al. | Depth-based subgraph convolutional auto-encoder for network representation learning | |
| CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
| Wang et al. | Link prediction in heterogeneous information networks: An improved deep graph convolution approach | |
| CN114639483A (en) | Electronic medical record retrieval method and device based on graph neural network | |
| CN111949886B (en) | Sample data generation method and related device for information recommendation | |
| Badhon et al. | A survey on association rule mining based on evolutionary algorithms | |
| CN113806630A (en) | Attention-based multi-view feature fusion cross-domain recommendation method and device | |
| CN116386899A (en) | Graph learning-based medicine disease association relation prediction method and related equipment | |
| Hassan et al. | Performance analysis of neural networks-based multi-criteria recommender systems | |
| CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system | |
| CN114417161A (en) | Virtual article time sequence recommendation method, device, medium and equipment based on special-purpose map | |
| Li et al. | Multi-behavior enhanced heterogeneous graph convolutional networks recommendation algorithm based on feature-interaction | |
| Shirgahi et al. | Trust estimation of the semantic web using semantic web clustering | |
| Chang et al. | KGTN: Knowledge Graph Transformer Network for explainable multi-category item recommendation | |
| Mall et al. | A novel Chebyshev neural network approach for solving singular arbitrary order Lane-Emden equation arising in astrophysics | |
| CN115545833A (en) | Recommendation method and system based on user social information | |
| CN115757897A (en) | Intelligent culture resource recommendation method based on knowledge graph convolution network |
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 |