Recommendation method and system based on knowledge-aware hypergraph neural network

Abstract

The present invention discloses a recommendation method based on a knowledge-aware hypergraph neural network, comprising the following steps: step 1, constructing a user hypergraph, with the initial hyperedges taking items that have an interactive relationship with the user as entity nodes; constructing an item hypergraph, with the initial hyperedges taking items that have an interactive relationship with any user that has an interactive relationship with the item as entity nodes; step 2, performing convolution calculations on the user hypergraph and each item hypergraph; step 3, calculating the inner product of the user and all items to obtain the interaction score between the user and all items. It has the beneficial effect of integrating the auxiliary information of the item into the vector representation of the user, so that the items with which the user is more likely to have an interactive relationship can be selected from a large number of items. A recommendation system based on a knowledge-aware hypergraph neural network is disclosed, comprising: a data capture module; a knowledge-aware hypergraph construction module; a domain convolution module; a hyperedge convolution module; and a prediction module. It has the beneficial effects of simple operation and accurate prediction.

Description

Technical Field

[0001] The present invention relates to the field of knowledge-aware application technology, and more specifically, to a recommendation method and system based on a knowledge-aware hypergraph neural network. Background Art

[0002] With the rapid development of the Internet, recommender systems have been widely deployed to alleviate the impact of information overload. The traditional recommendation technique is collaborative filtering, which assigns representation vectors based on user and item identities and then models their interactions through specific operations such as inner products or neural networks. However, in the absence of auxiliary information, these collaborative filtering-based methods usually suffer from sparsity and cold start problems. To address these problems, various types of auxiliary information have been explored to improve recommendation performance, such as item attributes, item reviews, and users' social networks.

[0003] In order to alleviate the sparsity and cold start problems of traditional recommendation systems, knowledge-aware graphs have been widely studied and applied in the field of recommendation systems as auxiliary information. However, most existing recommendation methods based on knowledge-aware graphs have the following shortcomings, namely, insufficient modeling of high-order correlations between users, items, and entities, and simple aggregation strategies that cannot preserve relationship information in the neighborhood.

[0004] In recent years, knowledge graphs (KGs) have attracted increasing attention in recommender system research due to their flexibility in modeling comprehensive auxiliary data. The key to the fusion of KGs and recommender systems is how to effectively integrate auxiliary information into the vector representations of users and items. According to the way to deal with this challenge, existing KG-based recommender systems can be divided into two categories: path-based and graph neural network-based methods. Path-based methods infer user preferences by exploring multiple meta-paths between target users and items on the KG, which usually requires domain knowledge. More importantly, this type of method ignores the rich structural information implicit in the KG and cannot fully describe the potential relationship between a given target user and item. Inspired by the recent emergence of graph neural networks, existing graph neural network-based methods provide remarkable performance by explicitly modeling high-order connectivity in the KG. However, these methods are still subject to two limitations: (1) High-order correlations between users, items, and entities in the KG are essential for data modeling. These methods mainly apply graph neural networks to enrich the representation of target nodes by recursively aggregating their original neighbors in the KG, and therefore have limitations in modeling high-order correlations between target nodes and non-original neighbors. In addition, the graph structure adopted by existing methods has constraints on modeling and processing high-order correlations, because the graph can only represent pairwise connections. (2) When aggregating neighborhoods, simple aggregation functions (such as mean and max pooling) are usually performed as aggregation functions to generate neighborhood embeddings. This aggregation method destroys the rich relationship information between nodes in the neighborhood. Therefore, neighborhood aggregation ignores the fine structure of the relationships between nodes in the neighborhood. Summary of the invention

[0005] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0006] Another object of the present invention is to provide a recommendation method based on a knowledge-aware hypergraph neural network, which can overcome the sparsity and cold start problems of traditional recommendation systems. The knowledge-aware hypergraph convolution method can effectively aggregate different neighbors in the neighborhood and simultaneously retain the relationship information in the neighborhood.

[0007] A recommendation system based on knowledge-aware hypergraph neural network is proposed, which can overcome the sparsity and cold start problems of traditional recommendation systems, effectively aggregate different neighbors in the neighborhood and retain the relationship information in the neighborhood at the same time.

[0008] In order to achieve these purposes and other advantages according to the present invention, a recommendation method based on a knowledge-aware hypergraph neural network is provided, comprising the following steps:

[0009] Step 1: construct a user hypergraph, wherein the user hypergraph includes the user's initial hyperedge and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1. The initial hyperedge uses the item that has an interactive relationship with the user as the entity node, the 2-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively;

[0010] Construct an item hypergraph, wherein the item hypergraph includes an initial hyperedge of the item and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, and the initial hyperedge has an item with which any user has an interactive relationship with the item as an entity node, and the second order knowledge-aware hyperedge has the auxiliary information of the item corresponding to the entity node of the initial hyperedge as an entity node, and the next order knowledge-aware hyperedge has the item or the auxiliary information of the item corresponding to the entity node of the previous order knowledge-aware hyperedge as an entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item, respectively;

[0011] Step 2: Perform convolution calculation on the user hypergraph and each item hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each item hypergraph;

[0012] Step 3: Calculate the inner product of the user's single vector and the single vector of all items, and then perform a nonlinear transformation to obtain the interaction score between the user and all items. The interaction score represents the probability of the user interacting with the item.

[0013] Preferably, each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph is processed by a domain convolution method to obtain a single vector of each hyperedge of the user hypergraph;

[0014] The domain convolution method is used to process each knowledge-aware hyperedge of the item hypergraph to obtain a single vector for each hyperedge of the item hypergraph.

[0015] Preferably, the single vectors of all hyperedges of the user hypergraph are processed by a hyperedge convolution method to obtain a unique single vector of the user hypergraph;

[0016] The hyperedge convolution method is used to process the single vectors of all knowledge-aware hyperedges of the hypergraph of the object to obtain the unique single vector of the object.

[0017] Preferably, the domain convolution method is specifically calculated using the following method:

[0018] Put the entity vectors in the l-order hyperedge and the l-1-order hyperedge into the one-dimensional convolution conv1 to generate the transformation matrix T, then use the transformation matrix T to permute and weight the entity vectors in the l-order hyperedge to obtain several transformed hyperedge vectors, and then use the one-dimensional convolution conv2 to aggregate the transformed hyperedge vectors to obtain a single vector of the l-order hyperedge, where l is an integer greater than 1 and the initial hyperedge is a 1-order hyperedge;

[0019] For the domain convolution of the initial hyperedge, only the entity vector in the initial hyperedge is put into the one-dimensional convolution conv1 to generate the transformation matrix T. Preferably, for the user hypergraph: the hyperedge convolution method uses a connection aggregator to aggregate all the single vectors of the hyperedges of the user hypergraph into a unique single vector corresponding to the user and the item;

[0020] For any item hypergraph: use a connection aggregator to aggregate all knowledge-aware hyperedge single vectors of the item hypergraph and the initial single vector of the item itself.

[0021] Preferably, the method further includes: selecting the same number of positive samples and negative samples for each user from all items used for calculation, wherein the negative samples are randomly selected from items that have not been interacted with the user, and the positive samples are items that have been interacted with the user.

[0022] Preferably, the method further includes sorting the interaction scores from large to small, and outputting the items corresponding to one or several interaction scores that are ranked high.

[0023] A recommendation system based on a knowledge-aware hypergraph neural network is provided, comprising:

[0024] A data capture module, which is used to obtain information about users and items that users have interacted with;

[0025] A knowledge-aware hypergraph construction module is used to construct a user hypergraph, wherein the user hypergraph includes the user's initial hyperedge and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, the initial hyperedge uses the item that has an interactive relationship with the user as the entity node, the 2-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively;

[0026] And, used to construct an item hypergraph, wherein the item hypergraph includes an initial hyperedge of the item and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, the initial hyperedge has an item with which any user has an interactive relationship with the item as an entity node, the 2-order knowledge-aware hyperedge has the auxiliary information of the item corresponding to the entity node of the initial hyperedge as an entity node, the next-order knowledge-aware hyperedge has the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as an entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item, respectively;

[0027] A domain convolution module, which is used to process each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the user hypergraph;

[0028] and, for processing each knowledge-aware hyperedge of the item hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the item hypergraph;

[0029] A hyperedge convolution module, which is used to process the single vectors of all hyperedges of the user hypergraph using the hyperedge convolution method to obtain a unique single vector of the user hypergraph;

[0030] And, using a hyperedge convolution method to process the single vectors of all knowledge-aware hyperedges of the hypergraph of the object to obtain an initial single vector of the object;

[0031] The prediction module is used to calculate the inner product of the user's single vector and the single vector of all items, and then perform a nonlinear transformation to obtain the interaction score between the user and all items. The interaction score represents the probability of interaction between the user and the item.

[0032] The present invention has at least the following beneficial effects:

[0033] First, this paper proposes a knowledge-aware hypergraph neural network framework to solve the sparsity and cold start problems of traditional recommender systems. First, a knowledge-aware hypergraph structure composed of hyperedges is used to explicitly model the high-order correlations of users, items, and entities in the knowledge graph. Second, a new knowledge-aware hypergraph convolution method is proposed to effectively aggregate different neighbors in the neighborhood while preserving the relationship information in the neighborhood.

[0034] Second, the present invention proposes to use a knowledge-aware hypergraph to explicitly model the high-order correlations between users, items, and entities in the knowledge-aware graph to achieve more accurate and efficient item recommendations.

[0035] Third, in order to solve the limitations of existing knowledge graph-based recommendation methods in order to recommend suitable items to users, the present invention proposes an end-to-end model - knowledge-aware hypergraph neural network. In the model, we first construct the initial hyperedges of users and items based on the user's historical interaction records (such as purchases, clicks, and comments), and then obtain subsequent knowledge hyperedges based on the previous initial hyperedges and the knowledge graph related to the items. These hyperedges associated with users and items constitute the hypergraphs of users and items respectively. After that, in the hypergraph convolution, we first aggregate several nodes in the hyperedge (represented by vectors) into a single vector through domain convolution, and then aggregate the representation vectors of each hyperedge obtained previously into a single vector through hyperedge convolution. In this way, we obtain single vector representations of the hypergraphs corresponding to users and items respectively. Finally, we put the vectors of the hypergraphs corresponding to users and items into the prediction layer to obtain the probability of users clicking on the item, so that we can select items with high probability of user clicks from many items and recommend them to the user.

[0036] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A framework diagram of the knowledge-aware hypergraph neural network model of one of the technical solutions of the present invention;

[0038] Figure 2 This is a schematic diagram of the domain convolution module of one of the technical solutions of the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0040] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified; in the description of the present invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, and does 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 therefore cannot be understood as a limitation on the present invention.

[0041] like Figures 1-2 As shown, the present invention provides a recommendation method based on knowledge-aware hypergraph neural network, comprising the following steps:

[0042] Step 1: construct a user hypergraph, wherein the user hypergraph includes the user's initial hyperedge and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1. The initial hyperedge uses the item that has an interactive relationship with the user as the entity node, the 2-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively;

[0043] Constructing an item hypergraph, wherein the item hypergraph includes an initial hyperedge of the item and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, and the initial hyperedge has an item with which any user has an interactive relationship with the item as an entity node, and the 2-order knowledge-aware hyperedge has the auxiliary information of the item corresponding to the entity node of the initial hyperedge as an entity node, and the next-order knowledge-aware hyperedge has the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as an entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item, respectively;

[0044] Step 2: Perform convolution calculation on the user hypergraph and each item hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each item hypergraph;

[0045] Step 3: Calculate the inner product of the user's single vector and the single vector of all items, and then perform a nonlinear transformation to obtain the interaction score between the user and all items. The interaction score represents the probability of the user interacting with the item.

[0046] In the above technical solution, the knowledge-aware hypergraph neural network. In the model, we first construct the initial hyperedges of users and items based on the user's historical interaction records (such as purchases, clicks, and comments), and then obtain the subsequent knowledge-aware hyperedges based on the previous initial hyperedges and the knowledge-aware graphs related to the items. These hyperedges associated with users and items constitute the hypergraphs of users and items respectively. Then, in the hypergraph convolution, we first aggregate several nodes (represented by vectors) in the hyperedge into a single vector through domain convolution, and then aggregate the representation vectors of each hyperedge obtained previously into a single vector through hyperedge convolution. In this way, we obtain single vector representations of the hypergraphs corresponding to users and items respectively. Finally, we put the vectors of the hypergraphs corresponding to users and items into the prediction layer to obtain the probability of the user clicking on the item, that is, calculate the inner product of the user's single vector and the single vectors of all items, and then perform nonlinear transformation to obtain the interaction score between the user and all items, which represents the probability of the user interacting with the item. In this way, we can select items that are more likely to interact with the user from among many items and recommend them to the user. Through the above method, the auxiliary information of the item can be effectively integrated into the vector representation of the user, so that the items with which the user is more likely to interact can be selected from a large number of items.

[0047] In order to better learn the representation vectors of target users and target items to obtain more accurate recommendation results, we proposed an end-to-end model KHNN (Knowledge-aware Hypergraph Neural Network) to explicitly model the high-order correlations between users, items, and entities in the knowledge-aware graph to achieve more accurate and efficient item recommendations for users. In the model, we first construct the initial hyperedges of users and items based on the user's historical interaction records (such as purchases, clicks, and comments).

[0048] One of the implementation methods is: We let U = {u1,u2,…,u M} is a set of all users and V = {v1,v2,…,v N} is a set of owned items. M, N are the number of users and items respectively. We define the user-item interaction matrix as Y∈R M×N , where y uv =1 means user u has interacted with item v, such as clicking, browsing or purchasing; otherwise y uv = 0. In addition, we define the knowledge perception graph as G K={(h,r,t)|h,t∈E,r∈R}, where each triple (h,r,t) indicates that there is a relationship r between the head entity h and the tail entity t, and ε and R are the sets of entities and relationships in the knowledge-aware graph. For example, the triple (Back to the Future, directed by, Robert Zemeckis) means that Robert Zemeckis is the director of the movie Back to the Future. In traditional graph structures, edges only connect two vertices, while in hypergraphs, a hyperedge can connect two or more vertices. We define the knowledge-aware hypergraph as G H ={ε,H}, where ε represents the entity set and H represents the hyperedge set. We use the set A={(v,e)|v∈V,e∈E} to align item v with entity e in the knowledge-aware hypergraph. Given the user-item interaction matrix Y and the knowledge-aware hypergraph G H , we aim to predict whether user u is potentially interested in item v that he has not interacted with before. Our goal is to learn a prediction function in represents the probability of user u interacting with item v, and θ represents the model parameter of function F.

[0049] The establishment of the initial hyperedge:

[0050] Generally speaking, items that users have interacted with can represent the user's preferences to a certain extent. Therefore, we construct the user's initial hyperedge through the items that user u has interacted with. The set of related items of user u is obtained through the historical interactions of user u, and the items in the set are aligned with the entities in the knowledge-aware graph through set A (that is, find entities with the same ID in the knowledge-aware graph through the ID of the item). These entities can be constructed as the user's initial hyperedge (that is, the set of entities). The initial hyperedge of user u is defined as follows:

[0051]

[0052] Similarly, for the initial hyperedges of an item, users who have interacted with the same item can 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 neighbor set of item v as follows:

[0053]

[0054] Then, the items in the initial neighbor set of item v are aligned with the entities in the knowledge-aware graph through set A. Then, these entities can form the initial hyperedge of item v, which is defined as follows:

[0055]

[0056] Input into the model: historical interaction records of users

[0057] Generation: Initial hyperedges of users and items

[0058] Implementation process: Normal distribution can be used to initialize random vectors for all items to contain the intrinsic characteristics of each item. Then, through the user's historical interaction records, the initial hyperedges (first-order hyperedges) of the target user and target item can be constructed respectively.

[0059] Establishment of knowledge-aware hyperedge:

[0060] Adjacent entities in the knowledge-aware graph are always strongly associated. By propagating this association along the links between entities in the knowledge-aware graph (relation r in triples), we can obtain extended hyperedges with different distances from the initial hyperedges, namely knowledge-aware hyperedges, which can effectively expand the vector representation of users and items. The definition of the knowledge hyperedge (l-order hyperedge) between user u and item v is recursively expressed as:

[0061]

[0062] Where l represents the distance from the initial hyperedge, and the subscript symbol o is a uniform placeholder for the symbol u or v. It is very meaningful to use the knowledge-aware graph as auxiliary information to build our model, because neighboring entities can be regarded as extensions of user preferences and item features. Figure 1 As shown in Figure 3, based on these generated initial hyperedges and knowledge-aware hyperedges, we can construct the corresponding hypergraphs (composed of hyperedges) for the target users and target items respectively.

[0063] Input to the model: initial hyperedges of users and items

[0064] Generation: Knowledge Hyperedges of Users and Items

[0065] Implementation process: We take the entity node in the initial hyperedge obtained previously as the starting point, and obtain subsequent entity nodes connected to the entity node in the initial hyperedge through the links between entities on the knowledge-aware graph. These subsequent entity nodes constitute a knowledge hyperedge with a distance of one from the initial hyperedge. Then, we take the knowledge-aware hyperedge obtained previously as the starting point and obtain a knowledge-aware hyperedge with a distance of two from the initial hyperedge through the links in the knowledge-aware graph. By repeating this construction process, we can construct multiple hyperedges for users and items respectively. These hyperedges constitute the hypergraphs of users and items respectively.

[0066] In another technical solution, the domain convolution method is specifically calculated using the following method:

[0067] Put the entity vectors in the l-order hyperedge and the l-1-order hyperedge into the one-dimensional convolution conv1 to generate the transformation matrix T, then use the transformation matrix T to permute and weight the entity vectors in the l-order hyperedge to obtain several transformed hyperedge vectors, and then use the one-dimensional convolution conv2 to aggregate the transformed hyperedge vectors to obtain a single vector of the l-order hyperedge, where l is an integer greater than 1 and the initial hyperedge is a 1-order hyperedge;

[0068] For the domain convolution of the initial hyperedge, only the entity vector in the initial hyperedge is put into the one-dimensional convolution conv1 to generate the transformation matrix T.

[0069] Neighborhood convolution is applied to aggregate several entity vectors in a hyperedge into a single vector representation. This scheme proposes a new neighborhood convolution method. In particular, in the triples of the knowledge-aware graph, when each tail entity is connected to different head entities through relationships, it has different meanings and potential vector representations. Therefore, we put the entity vectors in the l-order and l-1-order hyperedges into a one-dimensional convolution conv1 to generate a transformation matrix T, and then use the transformation matrix T to permute and weight the vectors in the l-order hyperedge to obtain several transformed hyperedge vectors. Finally, a one-dimensional convolution conv2 is used to aggregate the transformed hyperedge vectors to obtain a single vector representation of the l-order hyperedge, as described in the following formula.

[0070]

[0071]

[0072] Where subscript o is a uniform placeholder for the symbol u or v, and || is the concatenation operation. Both conv1 and conv2 are one-dimensional convolutions, but with different output channels.

[0073] Because the initial hyperedge has a strong connection with the target user and item. Therefore, a single vector representation of the initial hyperedge is added for the target user u and item v, as described in the following formula.

[0074]

[0075]

[0076] The initial hyperedge of the item hypergraph, item v has related entities of its original representation, but user u does not. The original related entity is the node closest to item v itself in the latent semantic space. Therefore, we add it to the set of representations of item v and define it as follows:

[0077]

[0078] After the neighborhood convolution, we formulate the vector set of the hyperedge representation and original representation of the target item v and the vector set of the hyperedge representation of the target user u as follows:

[0079]

[0080]

[0081] Input: All hyperedges of users and items.

[0082] Generates: A single vector representation of all hyperedges for users and items.

[0083] Implementation process: The node vectors in two adjacent hyperedges in the hypergraph are used as the input of the one-dimensional convolution conv1 to obtain the transformation matrix T, and then the vector matrix of the hyperedge to be transformed is multiplied by the transformation matrix T to obtain the vector matrix of the transformed hyperedge. Finally, the vector matrix of the transformed hyperedge is used as the input of the one-dimensional convolution conv2 to obtain a single vector representation of the final hyperedge.

[0084] In another technical solution, for the user hypergraph: the hyperedge convolution method uses a connection aggregator to aggregate all hyperedge single vectors of the user hypergraph into a unique single vector corresponding to the user and the item;

[0085] For any item hypergraph: use the connection aggregator to aggregate all knowledge-aware hyperedge single vectors of the item hypergraph and the initial single vector of the item itself. Among them, for the target item, the entity in the knowledge-aware graph has an initial vector, and the initial vector of this item is the representation vector of its entity. The initial single vector of the item itself is added during aggregation.

[0086] Hyperedge convolution aggregates the representation vectors of all hyperedges in the hypergraph of user u and item v into a single representation vector, such as Figure 1 As shown. We use the connection aggregator to aggregate and Multiple hyperedges in represent vectors to a single vector. The connection aggregator formula is as follows:

[0087]

[0088] in || is a concatenation operation.

[0089] Input: All hyperedge representation vectors of users and items.

[0090] Generation: Final single representation vector for users and items

[0091] Implementation process: All vectors in the hypergraph of user u and item v are aggregated into a single vector by connecting them. The single vector obtained after aggregation eventually becomes the representation vector of the user and item.

[0092] In another technical solution, it also includes: among all the items used for calculation, the same number of positive samples and negative samples are selected for each user, the negative samples are randomly selected from the items that have not been interacted with by the user, and the positive samples are the items that have been interacted with by the user.

[0093] Use e u To represent the user's representation vector, e v Represents the representation vector of the item. Finally, the inner product of the user and item vectors is calculated, and then a nonlinear transformation is performed to predict the user's preference score for the item:

[0094]

[0095] 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 positive samples are extracted for each user. These negative samples are randomly selected from items that the user has not interacted with. Finally, the loss function of the model is as follows:

[0096]

[0097] Where J is the cross entropy loss, p + represents a positive sample, p - represents negative samples. Θ is the model parameter, is the L2 regularization term parameterized by λ.

[0098] Recommended candidate sets:

[0099] Input: representation vectors of users and items obtained through hypergraph convolution

[0100] Generation: The user's final preference score for the candidate items.

[0101] Implementation process: After obtaining the representation vectors of the user and the item, the inner product of the user and item vectors is calculated, and then a nonlinear transformation is performed to predict the user's preference score for the item. After that, the scores of different candidate items in the candidate set are calculated using the corresponding user representations, and the calculated scores are sorted to obtain the final result. This can make the model prediction more accurate.

[0102] It also includes sorting the interaction scores from large to small, and outputting the items corresponding to one or several interaction scores that are ranked high.

[0103] A recommendation system based on a knowledge-aware hypergraph neural network is provided, comprising:

[0104] A data capture module, which is used to obtain information about users and items that users have interacted with;

[0105] A knowledge-aware hypergraph construction module is used to construct a user hypergraph, wherein the user hypergraph includes the user's initial hyperedge and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, the initial hyperedge uses the item that has an interactive relationship with the user as the entity node, the 2-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively;

[0106] And, used to construct an item hypergraph, wherein the item hypergraph includes an initial hyperedge of the item and 2-1 order knowledge-aware hyperedges, where 1 is an integer greater than 1, the initial hyperedge has an item with which any user has an interactive relationship with the item as an entity node, the 2-order knowledge-aware hyperedge has the auxiliary information of the item corresponding to the entity node of the initial hyperedge as an entity node, the next-order knowledge-aware hyperedge has the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as an entity node, and the entity nodes of two adjacent order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item, respectively;

[0107] A domain convolution module, which is used to process each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the user hypergraph;

[0108] and, for processing each knowledge-aware hyperedge of the item hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the item hypergraph;

[0109] A hyperedge convolution module, which is used to process the single vectors of all hyperedges of the user hypergraph using the hyperedge convolution method to obtain a unique single vector of the user hypergraph;

[0110] Furthermore, the hyperedge convolution method is used to process the single vectors of all knowledge-aware hyperedges of the hypergraph of the object and the initial single vector of the object itself to obtain the unique single vector of the object; wherein, for the target object, the entity in the knowledge-aware graph has an initial vector, and the initial vector of this object is the representation vector of its entity, and the initial single vector of the object itself is added during aggregation.

[0111] The prediction module is used to calculate the inner product of the user's single vector and the single vector of all items, and then perform a nonlinear transformation to obtain the interaction score between the user and all items. The interaction score represents the probability of interaction between the user and the item.

[0112] In the above technical solution, the user hypergraph and the item hypergraph are constructed based on the historical interaction relationship between users and items. During the construction process, the auxiliary information of items is used as a link to form a connection between knowledge perception graphs. Then, hypergraph convolution (domain convolution and hyperedge convolution) is performed to make each user and each item have a single vector representation. These vector representations can accurately reflect the interaction possibility between users and items (such as users' preferences, needs, and preferences for item types, etc.). Then, by calculating the inner product of the user's single vector and the single vector of all items, a nonlinear transformation is performed to obtain the interaction score between the user and all items. This makes it easier to recommend items with high interaction scores to users to meet their needs and improve prediction accuracy.

[0113] The data capture module, knowledge-aware hypergraph construction module, domain convolution module, hyperedge convolution module, and prediction module together form a knowledge-aware hypergraph neural network model. Once the model is trained, you only need to input the historical interaction information between the user and the item into the model to output the top-ranked items to the user.

[0114] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A recommendation method based on knowledge-aware hypergraph neural network, characterized in that: The following steps are involved: Step 1: Construct a user hypergraph, which includes the user's initial hyperedges and l- 1~ l Level knowledge-aware hyperedge, l is an integer greater than 1. The initial hyperedge uses the items that have an interactive relationship with the user as the entity node. l -The first-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node. The next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node. The entity nodes of the two adjacent-order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively. Construct an item hypergraph, where the item hypergraph includes the initial hyperedges and l -1~ l Level knowledge-aware hyperedge, l is an integer greater than 1. The initial hyperedge takes the item that any user has an interactive relationship with the item as the entity node. l -The first-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, and the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node. The entity nodes of the two adjacent-order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively. Step 2: Perform convolution calculation on the user hypergraph and each item hypergraph to obtain a single vector of the user hypergraph and a single vector corresponding to each item hypergraph; Step 3: Calculate the inner product of the user's single vector and the single vector of all items, and then perform nonlinear transformation to obtain the interaction score between the user and all items. The interaction score indicates the probability of the user interacting with the item. Using the domain convolution method to process each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph, a single vector of each hyperedge of the user hypergraph is obtained; Each knowledge-aware hyperedge of the item hypergraph is processed by the domain convolution method to obtain a single vector of each hyperedge of the item hypergraph; Among them, the domain convolution method is calculated using the following method: Will l Super-edge sum l -Entity vector in the 1st order hyperedge is put into a 1D convolution Generate a transformation matrix T, and then use the transformation matrix T to l The entity vectors in the hyperedge of order are permuted and weighted to obtain several transformed hyperedge vectors, and then one-dimensional convolution is used Aggregate the transformed hyperedge vectors to get l A single vector of a hyperedge of order , where l is an integer greater than 1, and the initial hyperedge is a first-order hyperedge; For the domain convolution of the initial hyperedge, only the entity vector in the initial hyperedge is put into the one-dimensional convolution. Generate transformation matrix T; The hyperedge convolution method is used to process the single vectors of all hyperedges of the user hypergraph to obtain the unique single vector of the user hypergraph; The hyperedge convolution method is used to process the single vectors of all knowledge-aware hyperedges of the hypergraph of the object to obtain the unique single vector of the object; For user hypergraph: the hyperedge convolution method uses a connection aggregator to aggregate all hyperedge single vectors of the user hypergraph into a unique single vector corresponding to the user and item; For any item hypergraph: use the connection aggregator to aggregate all knowledge-aware hyperedge single vectors of the item hypergraph and the initial single vector of the item itself; It also includes sorting the interaction scores from large to small, and outputting the items corresponding to one or several interaction scores that are ranked high.

2. The recommendation method based on knowledge-aware hypergraph neural network according to claim 1, characterized in that: Also includes: Among all the items used for calculation, the same number of positive samples and negative samples are selected for each user. Negative samples are randomly selected from items that have not been interacted with by the user, and positive samples are items that have been interacted with by the user.

3. The recommendation system based on knowledge-aware hypergraph neural network is characterized by: include: A data capture module, which is used to obtain information about users and items that users have interacted with; The knowledge-aware hypergraph construction module is used to construct a user hypergraph module, wherein the user hypergraph module includes the user's initial hyperedges and l -1~ l Level knowledge-aware hyperedge, l is an integer greater than 1. The initial hyperedge uses the items that have an interactive relationship with the user as the entity node. l -The first-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node. The next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node. The entity nodes of the two adjacent-order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively. And, for constructing an item hypergraph module, wherein the item hypergraph module includes an initial hyperedge of an item and l -1~ l Level knowledge-aware hyperedge, l is an integer greater than 1. The initial hyperedge takes the item that any user has an interactive relationship with the item as the entity node. l -The first-order knowledge-aware hyperedge uses the auxiliary information of the item corresponding to the entity node of the initial hyperedge as the entity node, and the next-order knowledge-aware hyperedge uses the item or the auxiliary information of the item corresponding to the entity node of the previous-order knowledge-aware hyperedge as the entity node. The entity nodes of the two adjacent-order knowledge-aware hyperedges correspond to the item and the auxiliary information of the item respectively. A domain convolution module, which is used to process each initial hyperedge and each knowledge-aware hyperedge of the user hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the user hypergraph; and, for processing each knowledge-aware hyperedge of the item hypergraph using a domain convolution method to obtain a single vector of each hyperedge of the item hypergraph; A hyperedge convolution module, which is used to process the single vectors of all hyperedges of the user hypergraph using the hyperedge convolution method to obtain a unique single vector of the user hypergraph; And, using a hyperedge convolution method to process the single vectors of all knowledge-aware hyperedges of the hypergraph of the object and the initial single vector of the object itself to obtain a unique single vector of the object; The prediction module is used to calculate the inner product of the user's single vector and the single vector of all items, and then perform a nonlinear transformation to obtain the interaction score between the user and all items. The interaction score represents the probability of the user interacting with the item.

Images