A recommendation method, apparatus and device

Abstract

The application provides a recommendation method, comprising: in response to a recommendation request, obtaining historical interaction data of a request user corresponding to the recommendation request, and determining a to-be-recommended object according to the historical interaction data; for each to-be-recommended object, constructing an interaction pair; for each interaction pair, taking a feature corresponding to the interaction pair as a node, and constructing a default hypergraph corresponding to the interaction pair; inputting the default hypergraph corresponding to the interaction pair into a hypergraph learning model, to predict a hyperedge set of the default hypergraph by a hyperedge generation module, and to model the default hypergraph based on the hyperedge set by a hypergraph learning module, to obtain a complete hypergraph corresponding to the interaction pair, and to predict a probability that a recommended object corresponding to the interaction pair is liked by the request user based on the complete hypergraph corresponding to the interaction pair by a prediction module; and selecting candidate recommended objects from the recommended objects corresponding to each interaction pair according to the probability that the recommended objects corresponding to each interaction pair are liked by the request user, to perform recommendation.

Description

Technical Field

[0001] This application relates to the field of recommendation technology, and more particularly to a recommendation method, apparatus, and device. Background Technology

[0002] In today's information-saturated world, recommender systems play a crucial role in e-commerce, social media, and content platforms. Recommender systems aim to provide personalized recommendations based on users' interests and behavioral patterns, helping them discover and access information, products, or services that may be of interest.

[0003] In traditional recommendation scenarios, collaborative filtering is typically used to generate recommendations. This method leverages the similarity between users or between users and items to make recommendations. While this method is applicable to most recommendation scenarios, the prevalence of noise and sparsity in internet data significantly reduces the accuracy of recommendations. Summary of the Invention

[0004] In view of this, this application provides a recommendation method, apparatus, and device for making accurate recommendations to users.

[0005] Specifically, this application is implemented through the following technical solution:

[0006] A first aspect of this application provides a recommended method, the method comprising:

[0007] In response to a recommendation request, the system obtains the historical interaction data of the requesting user corresponding to the recommendation request, and determines the recommended items that have interacted with the requesting user based on the historical interaction data.

[0008] For each item to be recommended, an interaction pair is constructed to characterize the interaction behavior between the requesting user and the item to be recommended; wherein, the interaction pair is a data pair consisting of the requesting user and the item to be recommended;

[0009] For each interaction pair, a default hypergraph without hyperedges is constructed using the features corresponding to the interaction pair as nodes.

[0010] The default hypergraph corresponding to the interaction pair is input into the hypergraph learning model, so that the hyperedge generation module of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair. Based on the complete hypergraph containing the hyperedges corresponding to the interaction pair, the prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user.

[0011] Based on the probability that the requested user likes the recommended items corresponding to each interaction pair, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requested user based on the candidate recommended items.

[0012] A second aspect of this application provides a recommendation apparatus, the apparatus comprising an acquisition module, a generation module, a processing module, and a recommendation module; wherein...

[0013] The acquisition module is used to respond to a recommendation request, acquire the historical interaction data of the requesting user corresponding to the recommendation request, and determine the items to be recommended that have interacted with the requesting user based on the historical interaction data.

[0014] The generation module is configured to construct an interaction pair for each item to be recommended, representing the interaction behavior between the requesting user and the item to be recommended; wherein the interaction pair is a data pair consisting of the requesting user and the item to be recommended;

[0015] The generation module is further configured to construct a default hypergraph without hyperedges for each interaction pair, using the features corresponding to the interaction pair as nodes.

[0016] The processing module is used to input the default hypergraph corresponding to the interaction pair into the hypergraph learning model, so that the hyperedge generation module of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair, and the prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user based on the complete hypergraph containing the hyperedges corresponding to the interaction pair.

[0017] The recommendation module is used to select candidate recommendations from the recommendations corresponding to each interaction pair based on the probability that the requesting user likes the recommendations corresponding to each interaction pair, and to make recommendations to the requesting user based on the candidate recommendations.

[0018] A third aspect of this application provides a recommended device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods provided in the first aspect of this application.

[0019] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in the first aspect of this application.

[0020] The recommendation method, apparatus, and device provided in this application, in response to a recommendation request, acquire historical interaction data of the requesting user corresponding to the recommendation request, determine the items to be recommended that have interacted with the requesting user based on the historical interaction data, construct an interaction pair for each item to be recommended to characterize the interaction between the requesting user and the item, then construct a default hypergraph without hyperedges for each interaction pair using the features corresponding to the interaction pair as nodes, and input the default hypergraph corresponding to the interaction pair into a hypergraph learning model so that the hypergraph learning model can... The hyperedge generation module predicts the set of hyperedges of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the set of hyperedges to obtain the complete hypergraph containing hyperedges corresponding to the interaction pair. The prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair will be liked by the requesting user based on the complete hypergraph containing hyperedges corresponding to the interaction pair. Finally, based on the probability that the recommended item corresponding to each interaction pair will be liked by the requesting user, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requesting user based on the candidate recommended items. In this way, a default hypergraph is constructed based on the interaction pair consisting of the requesting user and the item to be recommended. Then, the complete hypergraph is learned through the hypergraph learning model. Since the complete hypergraph generated by the hypergraph learning model contains the complete relationship between the user and the recommended item, that is, the complete hypergraph includes the high-order feature interaction between the user and the recommended item, it can not only capture the direct interaction relationship between the user and the recommended item, but also capture the deeper interaction relationship between the user and the recommended item. In this way, subsequent recommendation based on the complete hypergraph can alleviate the problems of noise and data sparsity, and greatly improve the accuracy of the recommendation. Attached Figure Description

[0021] Figure 1 A flowchart of the recommended method embodiment provided in this application;

[0022] Figure 2 This is a schematic diagram illustrating a default hypergraph as an exemplary embodiment of this application;

[0023] Figure 3 A schematic diagram illustrating the structure of a hypergraph learning model as shown in an exemplary embodiment of this application;

[0024] Figure 4 A flowchart of the recommended method embodiment two provided in this application;

[0025] Figure 5 This is a schematic diagram illustrating an association matrix as an exemplary embodiment of this application;

[0026] Figure 6 A flowchart of the recommended method embodiment three provided in this application;

[0027] Figure 7 This is a schematic diagram illustrating a hypergraph learning module as shown in an exemplary embodiment of this application;

[0028] Figure 8 A flowchart of the recommended method embodiment four provided in this application;

[0029] Figure 9 This is a schematic diagram illustrating the structure of a knowledge graph, as shown in an exemplary embodiment of this application.

[0030] Figure 10 This is a schematic diagram illustrating the implementation principle of generating prompt words, as shown in an exemplary embodiment of this application.

[0031] Figure 11 A schematic diagram illustrating the implementation principle of obtaining target recommendations is shown in an exemplary embodiment of this application.

[0032] Figure 12 This is a hardware structure diagram of the recommended device in this application;

[0033] Figure 13 This is a schematic diagram of the structure of the recommended device embodiment provided in this application. Detailed Implementation

[0034] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0035] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0036] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0037] This application provides a recommendation method, apparatus, and device for making accurate recommendations to users.

[0038] The recommendation method, apparatus, and device provided in this application, in response to a recommendation request, acquire historical interaction data of the requesting user corresponding to the recommendation request, determine the items to be recommended that have interacted with the requesting user based on the historical interaction data, construct an interaction pair for each item to be recommended to characterize the interaction between the requesting user and the item, then construct a default hypergraph without hyperedges for each interaction pair using the features corresponding to the interaction pair as nodes, and input the default hypergraph corresponding to the interaction pair into a hypergraph learning model so that the hypergraph learning model can... The hyperedge generation module predicts the set of hyperedges of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the set of hyperedges to obtain the complete hypergraph containing hyperedges corresponding to the interaction pair. The prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair will be liked by the requesting user based on the complete hypergraph containing hyperedges corresponding to the interaction pair. Finally, based on the probability that the recommended item corresponding to each interaction pair will be liked by the requesting user, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requesting user based on the candidate recommended items. In this way, a default hypergraph is constructed based on the interaction pair consisting of the requesting user and the item to be recommended. Then, the complete hypergraph is learned through the hypergraph learning model. Since the complete hypergraph generated by the hypergraph learning model contains the complete relationship between the user and the recommended item, that is, the complete hypergraph includes the high-order feature interaction between the user and the recommended item, it can not only capture the direct interaction relationship between the user and the recommended item, but also capture the deeper interaction relationship between the user and the recommended item. In this way, subsequent recommendation based on the complete hypergraph can alleviate the noise and data sparsity problems and greatly improve the accuracy of the recommendation.

[0039] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0040] Figure 1 The flowchart is for Embodiment 1 of the recommended method provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0041] S101. In response to the recommendation request, obtain the historical interaction data of the requesting user corresponding to the recommendation request, and determine the items to be recommended based on the historical interaction data that have interacted with the requesting user.

[0042] Specifically, a recommendation request can be initiated by the requesting user. When a user needs to obtain recommendations that interest them, they can initiate a recommendation request.

[0043] It should be noted that the requesting user's historical interaction data includes data on the requesting user's interactions with interactive objects. For example, when recommending products to the requesting user, the requesting user's historical interaction data may include historical interaction data between the requesting user and products, including but not limited to the requesting user's actions such as purchasing product 1, adding product 2 to favorites, and rating product 3. As another example, when recommending social users to the requesting user, the requesting user's historical interaction data may include the requesting user's actions such as liking social user 1, following social user 2, and liking social user 3.

[0044] Furthermore, after obtaining the historical interaction data of the requesting user, various methods can be used to determine the recommended items based on this data. For example, in one embodiment, all objects that have interacted with the requesting user in the historical interaction data can be directly selected as recommended items. In another embodiment, all objects that have interacted with the requesting user in the historical interaction data can be sorted according to the number of interactions from highest to lowest, and the top 30 objects can be selected as recommended items.

[0045] S102. For each of the items to be recommended, construct an interaction pair to characterize the interaction behavior between the requesting user and the item to be recommended; wherein, the interaction pair is a data pair consisting of the requesting user and the item to be recommended.

[0046] Specifically, interaction pairs can represent the interactive behavior between the requesting user and the item to be recommended. For example, in one possible implementation, each interaction pair can consist of three parts: the requesting user, the item to be recommended, and the interaction relationship. For example, an interaction pair can be represented as: User 1, Item 1, Like, Favorite, Buy. As another possible implementation, each interaction pair can consist of two parts: the requesting user and the item to be recommended. For example, an interaction pair can be represented as: User 1, Item 2.

[0047] It should be noted that, for example, in one embodiment, the requesting user is denoted as U1, and the item to be recommended is denoted as V = {v1, v2, ..., v...}. n} where v n Let Y represent the nth recommendation. Further, let the interaction pair be denoted as Y = {y...} 11 y 12 , ..., y 1n}, where y 1n ={U1,v n} indicates that the requesting user U1 and the recommended item v are... n The interaction pairs formed.

[0048] S103. For each interaction pair, construct a default hypergraph without hyperedges corresponding to the interaction pair, using the features corresponding to the interaction pair as nodes.

[0049] In practical implementation, feature extraction can be performed on the requesting user and the recommended item to obtain the features of the requesting user and the recommended item. For example, traditional feature extraction methods or neural network-based feature extraction can be used. Furthermore, the features corresponding to the interaction pair refer to the features of the requesting user and the features of the recommended item contained in the interaction pair.

[0050] Specifically, a hypergraph consists of hyperedges and nodes. Each hyperedge in a hypergraph can connect multiple nodes, and the hypergraph can represent the complex high-dimensional relationships between nodes in a more detailed and specific way. It should be noted that in this embodiment, nodes are the features corresponding to interaction pairs, and hyperedges describe the higher-order relationships between nodes.

[0051] Furthermore, the default hypergraph is constructed based on interaction pairs. The nodes in the default hypergraph are the features corresponding to the interaction pairs. The default hypergraph only contains nodes and needs to be further processed to obtain the hyperedges of the default hypergraph and generate a complete hypergraph containing the hyperedges.

[0052] For example, in one embodiment, for the interaction pair y 1j ={U1,v j The characteristics of the requesting user U1 include i1 and i2, and the recommended item v j Features include i3 and i4. At this point, the constructed interaction corresponds to the default hypergraph as follows: Figure 2 As shown. (Refer to...) Figure 2 The hypergraph consists of four nodes: i1, i2, i3, and i4.

[0053] Specifically, for ease of explanation, the interaction with y will be... 1j The corresponding hypergraph (default hypergraph and complete hypergraph) is denoted as G. j =(N j E j (See the preceding description, hypergraph G) j The nodes in the text are interactive pairs y 1j The corresponding features (interaction pairs y) 1j The characteristics of the corresponding requesting user U1, and the interaction with y 1j The corresponding recommended item v j (characteristics), for example, for ease of explanation, Nj is represented as: Nj={i j1 i j2 , ..., i (jo) Furthermore, the hyperedge set is represented as Ej={e j1 ej2 , ..., e jp}, where o represents the interaction pair y 1j The corresponding hypergraph contains the number of nodes, where p is the predefined number of hyperedges. At this point, the hyperedges are unknown, and the hypergraph is a default hypergraph. A step of predicting the hyperedges is required to obtain the complete hypergraph.

[0054] S104. The default hypergraph corresponding to the interaction pair is input into the hypergraph learning model, so that the hyperedge generation module of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair. The prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user based on the complete hypergraph containing the hyperedges corresponding to the interaction pair.

[0055] Specifically, Figure 3 This is a schematic diagram illustrating the structure of a hypergraph learning model as an exemplary embodiment of this application. Please refer to... Figure 3 The hypergraph learning model shown in this embodiment includes a hyperedge generation module, a hypergraph learning module, and a prediction module. The hypergraph learning model can obtain the probability that the recommended item corresponding to an interaction pair is liked by the requested user based on the default hypergraph of an interaction pair.

[0056] In specific implementation, the hyperedge generation module in the hypergraph learning model first predicts the set of hyperedges of the default hypergraph corresponding to an interaction pair. This set of hyperedges contains multiple hyperedges. It should be noted that a hyperedge represents a higher-order interaction relationship between the requesting user and the item to be recommended in the interaction pair. It should also be noted that the number of hyperedges in the set is pre-defined; in this embodiment, the specific number of hyperedges in the set is not limited.

[0057] It should be noted that hyperedges and nodes can completely preserve the original information of the interaction pair, improving the completeness of the information. At the same time, by describing in detail the high-order relationship between the requesting user and the item to be recommended, the recommendations can be made to better meet the user's needs.

[0058] Furthermore, the hypergraph learning module models the default hypergraph based on the set of hyperedges of the default hypergraph output by the hyperedge generation module, and obtains the complete hypergraph, which contains the features corresponding to the interaction pairs and the hyperedges connecting these features.

[0059] Furthermore, the prediction module of the hypergraph prediction model predicts the probability that the requesting user likes the recommended item in the interaction pair based on the complete hypergraph corresponding to the interaction pair output by the hypergraph learning module. For example, in one embodiment, combining the above example, the probability of the interaction pair y is predicted to be... 1jAfter the corresponding default hypergraph inputs the hypergraph learning model, the hypergraph learning model outputs the request user U1 to like the recommended item v. j The probability is 90%.

[0060] It should be noted that the training process of this hypergraph learning model can be based on the supervised mutual information maximization criterion to train the embedding representations of hyperedges and nodes, and on the mutual information minimization criterion to train the embedding representations of hyperedges, so as to maximize the mutual information between the embedding representations of hyperedges and nodes and minimize the mutual information between the embedding representations of each hyperedge.

[0061] It should be noted that, based on the supervised mutual information maximization criterion, when training the embedding representations of hyperedges and nodes, the training samples with the same labels are used, with hi and random hypergraph representation h. + The combination of these is sampled as the joint distribution of mutual information M(hi; y). Then, the side distribution hi and the random hypergraph representation h are sampled from the training samples with labels opposite to y. - And the following objective function is used to achieve optimization:

[0062] ,

[0063] Where D1 is the discriminator that distinguishes between joint distributions and marginal distributions, h+ represents the hypergraph representation that has the same interaction label value y as hi, and h- is a randomly selected hypergraph representation.

[0064] Furthermore, when training the embedding representation of the hyperedge based on the mutual information minimization criterion, the joint distribution of the samples is set to (hi, hj), where The marginal distribution is set to (hi, hi). To avoid overfitting, dropout is used to prevent each pair of input vectors from being identical. Furthermore, one sample is drawn from the joint distribution and one from the marginal distribution of each node. The specific objective function is defined as follows:

[0065] ,

[0066] D2 is a discriminator that distinguishes between the joint distribution and the marginal distribution. In its implementation, binary cross-entropy (BCE) loss is used to optimize these two discriminators. f is the dropout function, and hi and hj represent the hypergraph representations of samples obtained from the joint distribution and the marginal distribution of each node, respectively.

[0067] Furthermore, in one possible implementation of this application, the loss function of the hypergraph learning model is:

[0068] ,

[0069] Where λ1, λ2, and λ3 are weighting factors; N is the number of nodes in the hypergraph; For predictions by the hypergraph learning model; y n For a node; D1 and D2 are the L1 norms of the model parameters; D1 and D2 are discriminators that distinguish between joint and marginal distributions. ω is the cross-entropy loss function for binary classification; ω is a learnable parameter. To perform regularization on the hypergraph learning model;

[0070] ,

[0071] Among them, V ij Node representation; Here, α represents the temperature value, and β represents the interval parameters.

[0072] S105. Based on the probability that the recommended items corresponding to each interaction pair are liked by the requesting user, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requesting user based on the candidate recommended items.

[0073] Specifically, after obtaining the probability that the requested user will like the items corresponding to all interaction pairs, recommendations can be made to the requesting user based on this probability. For example, in one embodiment, items with a probability exceeding a preset value can be selected as candidate items and then recommended to the requesting user. In another embodiment, a preset number of items can be selected as candidate items in descending order of probability and then recommended to the requesting user.

[0074] The recommendation method provided in this embodiment, in response to a recommendation request, obtains the historical interaction data of the requesting user corresponding to the recommendation request, and determines the items to be recommended that have interacted with the requesting user based on the historical interaction data. Then, for each item to be recommended, an interaction pair is constructed to represent the interaction between the requesting user and the item. Then, for each interaction pair, a default hypergraph without hyperedges is constructed using the features corresponding to the interaction pair as nodes. The default hypergraph corresponding to the interaction pair is then input into a hypergraph learning model to learn the hyperedges of the hypergraph model. The generation module predicts the set of hyperedges of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the set of hyperedges to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair. The prediction module of the hypergraph learning model, based on the complete hypergraph containing the hyperedges corresponding to the interaction pair, predicts the probability that the recommended item corresponding to the interaction pair will be liked by the requesting user. Finally, based on the probability that the recommended item corresponding to each interaction pair will be liked by the requesting user, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requesting user based on these candidate recommended items. In this way, a default hypergraph is constructed based on the interaction pair consisting of the requesting user and the recommended item, and then the complete hypergraph is learned through the hypergraph learning model. Since the complete hypergraph generated by the hypergraph learning model contains the complete relationship between the user and the recommended item, that is, the complete hypergraph includes high-order feature interactions between the user and the recommended item, it can capture not only the direct interaction relationship between the user and the recommended item, but also deeper interaction relationships. Therefore, subsequent recommendations based on this complete hypergraph can alleviate noise and data sparsity problems, greatly improving the accuracy of recommendations.

[0075] Figure 4 The flowchart for the second embodiment of the recommended method provided in this application is shown below. Please refer to... Figure 4 Based on the above embodiments, the step of the hyperedge generation module of the hypergraph learning model predicting the set of hyperedges of the default hypergraph includes:

[0076] S401. For each node in the default hypergraph, obtain the initial feature vector of the node, and obtain the fused feature vector of the node based on the initial feature vector of the node and the initial feature vectors of other nodes in the default hypergraph besides the node; wherein the dimension of the fused feature vector is the same as the number of hyperedges contained in the hyperedge set to be predicted.

[0077] Specifically, the initial feature vector of each node is a d-dimensional vector. For example, combining the previous example, for node i... jx (where x ranges from 1 to 0), and its initial eigenvector can be represented as: Vi jx , which is a d-dimensional feature vector. Further, for node i jxThe process of obtaining the fused feature vector of this node can be represented as:

[0078] ,

[0079] Among them, Ri jx Indicates node i jx The fused feature vector; Vi jx Indicates node i jx The initial feature vector; Vi jy Indicates node i jy The initial feature vector; y takes values ​​from 1 to 0, and y ≠ x; concat represents the concatenation operation; MLP represents MLP processing.

[0080] In other words, for each node, firstly, the initial feature vector of the node can be concatenated with the initial feature vectors of other nodes in the hypergraph to obtain the concatenated feature vector of the node. Then, MLP processing is performed on the concatenated feature vector of the node to obtain the fused feature vector of the node.

[0081] It should be noted that the dimension of the fused feature vector of a node is the same as the number of hyperedges contained in the set of hyperedges to be predicted. Referring back to the example above, the number of hyperedges contained in the set of hyperedges to be predicted is... Therefore, the dimension of the fused feature vector of each node is p-dimensional.

[0082] S402. Concatenate the fused feature vectors of all nodes in the default hypergraph into a node representation matrix.

[0083] Specifically, in one embodiment, the fused feature vector of each node can be treated as a row, and the fused feature vectors of all nodes can be concatenated together to obtain a node representation matrix. Thus, each row in the node representation matrix represents the fused feature vector of one node. Referring to the preceding description, the node representation matrix is ​​a... The matrix.

[0084] S403. An activation function is used to fuse the uniform random variable and each element in the node representation matrix and map them to binary values ​​to obtain an incidence matrix for representing the hyperedge set; wherein, each column of the incidence matrix represents a predicted hyperedge, and each element in the column indicates whether the hyperedge is connected to each node in the default hypergraph.

[0085] Specifically, for example, in one embodiment, the default hypergraph contains 8 nodes, which are V1 to V8, and e1 to e4 represent the 4 hyperedges to be predicted. In this step, the obtained correlation matrix is ​​as follows: Figure 5 As shown. Among them, the element e in the correlation matrix E... ij When =1, it indicates that the hyperedge e iCan be connected to node v j See also Figure 5 It can be seen that superedge e1 can be connected to nodes V1, V2, and V3; superedge e2 can be connected to nodes V2, V3, V4, and V5; superedge e3 can be connected to nodes V5, V6, and V7; and superedge e4 can be connected to node V8.

[0086] In practice, the specific implementation process of this step can be represented by the following formula:

[0087] ,

[0088] Where z represents a uniform distribution, and sigmoid represents the activation function. This represents the temperature value, where α and β are interval parameters, and β < 0 < α. Here, the range of the Sigmoid function is first stretched to [β, α], and then e is adjusted using MAX and MAX operations. ij The value of e is mapped to 0 or 1. ij V is the element in the i-th row and j-th column of the correlation matrix; ij Let be a node representation (i.e., a node represents the element in the i-th row and j-th column of the matrix).

[0089] The recommended method provided in this embodiment, for each node in the default hypergraph, obtains the node's initial feature vector, and based on the node's initial feature vector and the initial feature vectors of other nodes in the default hypergraph, obtains the node's fused feature vector. Then, the fused feature vectors of all nodes in the default hypergraph are concatenated to form a node representation matrix. Finally, a hard-constrained distribution is used to map each element in the node representation matrix to a binary value, resulting in an association matrix representing the hyperedge set. In this way, the hyperedges of the default hypergraph can be obtained through the hyperedge generation module, and then the complete hypergraph can be obtained based on the hyperedges, allowing the high-order interactions between the corresponding features of the interaction pairs to be characterized through the complete hypergraph.

[0090] Figure 6 The flowchart is for the third embodiment of the recommended method provided in this application. Figure 7 This is a schematic diagram illustrating a hypergraph learning module as shown in an exemplary embodiment of this application. Please refer to... Figure 7 ,exist Figure 7 In the example shown, the hypergraph learning module of the hypergraph learning model includes a predetermined number of sequentially connected learning units and a fusion unit. The following section combines... Figure 7 This section details how to obtain a complete hypergraph. For specifics, please also refer to... Figure 6 and Figure 7Based on the above embodiments, the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing hyperedges corresponding to the interaction pair, including the following steps:

[0091] S601. The first learning unit in the preset number of learning units generates an embedding representation of the hyperedge in the learning unit based on the fusion feature vectors of the nodes connected to the hyperedge in the hyperedge set. For each node in the default hypergraph, the fusion feature vector of the node is updated based on the embedding representations of the hyperedges connected to the node, so as to obtain the embedding representation of the node in the learning unit.

[0092] Specifically, the hypergraph learning module includes a preset number of learning units and a fusion unit connected sequentially. The specific value of the preset number is set according to actual needs; in this embodiment, the specific value of the preset number is not limited. For example, in... Figure 7 In the example shown, the specified quantity can be 2.

[0093] Furthermore, the first learning unit, for each hyperedge in the hyperedge set, fuses the fused feature vectors of each node connected to the hyperedge to generate an embedding representation of the hyperedge on the first learning unit.

[0094] Specifically, the implementation principle of generating the embedding representation of the hyperedge on the first learning unit can be expressed as follows:

[0095] ,

[0096] in, Let represent the embedding representation of the j-th superedge; , and These are the query vector, key vector, and value vector in the attention mechanism, which are calculated through linear transformations. h and s h+1 Here, represents the start and end indices of the (h+1)th block, and is formalized as follows:

[0097] ,

[0098] Where d represents the hidden layer dimension; h represents the current number of heads in the multi-head attention mechanism; and H represents the total number of heads used.

[0099] Furthermore, after obtaining the embedding representations of all hyperedges, for each node in the default hypergraph, the fused feature vector of the node is updated based on the embedding representations of each hyperedge connected to the node, thus obtaining the embedding representation of the node on the learning unit.

[0100] In practice, the embedding representation of the hyperedge can be propagated to the node. That is, for each node, the embedding representations of the hyperedges connected to the node are merged to generate the embedding representation of the node.

[0101] For example, in one possible implementation, the principle is as follows:

[0102] ,

[0103] in, Let represent the embedding representation of the j-th superedge; , and These are the query vector, key vector, and value vector in the attention mechanism, which are calculated through linear transformations. h and s h+1 This represents the start and end indices of the (h+1)th block.

[0104] S602. Among the preset number of learning units, excluding the first learning unit, for each hyperedge in the hyperedge set, the other learning units generate the embedding representation of the hyperedge in the learning unit based on the embedding representation of each node connected to the hyperedge in the previous learning unit before the learning unit. For each node in the default hypergraph, based on the embedding representation of each hyperedge connected to the node in the previous learning unit before the learning unit, the fused feature vector of the node is updated to obtain the embedding representation of the node in the learning unit.

[0105] Specifically, starting from the second learning unit, the embedding representations of each hyperedge and each node output from the previous learning unit are taken as input and processed by the current learning unit. The output is the embedding representation of the hyperedge and the embedding representation of the node in that learning unit. In other words, the hypergraph learning module continues to process until a preset number of learning units have completed their processing.

[0106] S603. The fusion unit performs weighted processing on the embedding representation of each hyperedge in the hyperedge set on each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each hyperedge.

[0107] Specifically, the weights pre-set for each learning unit are determined according to actual needs. In this embodiment, the specific values ​​of the weights pre-set for each learning unit are not limited.

[0108] Furthermore, for each hyperedge, the fusion unit performs weighted processing on the embedding representation of the hyperedge output by each learning unit based on the weight of the learning unit, and uses the weighted processing result as the weighted embedding representation of the hyperedge.

[0109] Specifically, the process can be represented as:

[0110] ,

[0111] Where H represents the weighted embedding representation; L represents the preset number; W l H represents the weight of the l-th pre-learned unit; l This represents the embedding representation of the hyperedge on the l-th learning unit.

[0112] S604. The fusion unit performs weighted processing on the embedding representation of each node of the default hypergraph in each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each node.

[0113] Specifically, the weights pre-set for each learning unit are determined according to actual needs. In this embodiment, the specific values ​​of the weights pre-set for each learning unit are not limited.

[0114] Furthermore, for each node, the fusion unit performs weighted processing on the embedding representation of the node output by each learning unit based on the weight of the learning unit, and uses the weighted processing result as the weighted embedding representation of the node.

[0115] S605. The fusion unit uses the weighted embedding representations of each hyperedge and each node to aggregate and generate the complete hypergraph.

[0116] Specifically, the weighted embedding representations are aggregated to generate the final complete hypergraph.

[0117] The recommended method provided in this embodiment, based on the first learning unit out of a preset number of learning units, for the hyperedges in the hyperedge set, generates an embedding representation of the hyperedge in the learning unit based on the fused feature vectors of each node connected to the hyperedge. Then, based on the embedding representations of each hyperedge connected to the node, the fused feature vector of the node is updated to obtain the embedding representation of the node in the learning unit. Furthermore, based on the other learning units out of the preset number of learning units (excluding the first learning unit), for the hyperedges in the hyperedge set, an embedding representation of the hyperedge in the learning unit is generated based on the embedding representations of each node connected to the hyperedge in the previous learning unit. Finally, based on the node... The embedding representations of each connected hyperedge in the previous learning unit are used to update the fused feature vector of the node, thus obtaining the node's embedding representation in the current learning unit. Then, based on the weights pre-set for each learning unit, the fusion unit weights the embedding representations of each hyperedge in the hyperedge set, resulting in a weighted embedding representation for each hyperedge. Simultaneously, the fusion unit weights the embedding representations of each node in the default hypergraph, also pre-set for each learning unit, resulting in a weighted embedding representation for each node. Finally, the fusion unit aggregates the weighted embedding representations of each hyperedge and each node to generate a complete hypergraph. This allows for the characterization of higher-order interactions between features corresponding to interaction pairs based on the complete hypergraph, thereby improving recommendation accuracy.

[0118] Figure 8 The flowchart for Embodiment 4 of the recommended method provided in this application is shown below. Please refer to... Figure 8 Based on the above embodiments, the steps following the selection of candidate recommendations include:

[0119] S801. Select a preset number of features from the features corresponding to the requesting user and the features corresponding to the item to be recommended.

[0120] Specifically, a predetermined number of features can be selected from the features corresponding to the requesting user and the features corresponding to the item to be recommended, using a random selection method.

[0121] It should be noted that the specific value of the preset quantity is set according to actual needs. In this embodiment, the specific value of the preset quantity is not limited. For example, half of the total number of features can be selected.

[0122] S802. Based on the relationship between the preset number of features and the requesting user or the item to be recommended, construct a knowledge graph in the form of triples.

[0123] Specifically, a triple can be an ordered set containing three items. Triples can represent the relationship between a feature and the item to be recommended, or the relationship between a feature and the requesting user. For example, in one embodiment, the triple consisting of user 2 and item 2 is <user 2, likes, item 2>.

[0124] Furthermore, the triples corresponding to all requesting users are combined into a knowledge graph. This knowledge graph contains the interaction relationship between the requesting user and the item to be recommended. The structured information in the knowledge graph in the form of triples helps to retrieve and process it. Figure 9 This is a schematic diagram illustrating the structure of a knowledge graph, as shown in an exemplary embodiment of this application. Please refer to... Figure 9 This triplet-form knowledge graph contains user A, user B, category A, item A, item B, and item C, and the relationships between them are shown in the connection relationships in the triplet-form knowledge graph.

[0125] S803. Convert the candidate recommendations into first text and the knowledge graph into second text.

[0126] S804. Generate prompt words that correspond to both the first text and the second text.

[0127] Specifically, Figure 10 This is a schematic diagram illustrating the implementation principle of generating prompt words, as shown in an exemplary embodiment of this application. Please refer to... Figure 10 The first text is used to characterize the features of the candidate recommendations, and information about the corresponding candidate recommendations can be obtained from the first text. Furthermore, the second text is used to characterize the features in the knowledge graph, and the features in the knowledge graph in the form of corresponding triples can be obtained from the second text.

[0128] Specifically, prompt words can be generated according to a preset template, corresponding to both the first and second texts. Each prompt word uniquely corresponds to a candidate recommendation and a knowledge graph. In other words, referring to the example above, please continue to refer to... Figure 10 The prompts corresponding to the first and second texts are shown in the figure.

[0129] S805. Input the prompt words into the pre-trained large model, so that the pre-trained large model can re-rank the candidate recommendations based on the prompt words to obtain the ranking result, and select the target recommendation from the candidate recommendations based on the ranking result.

[0130] Specifically, Figure 11 This is a schematic diagram illustrating the implementation principle of obtaining a target recommendation, as shown in an exemplary embodiment of this application. Please refer to... Figure 11The pre-trained large model re-ranks candidate recommendations based on prompts and ultimately selects the target recommendation. Several methods can be used to derive the target recommendation from the ranking results. For example, the top 20 items in the ranking can be directly selected as the target recommendation.

[0131] Furthermore, the pre-trained large model is a model fine-tuned based on the GPT language model, and the final target recommendation is obtained through this large model.

[0132] S806. Recommend the target product to the requesting user.

[0133] Specifically, after selecting the target recommendation item, the target recommendation item is recommended to the requesting user according to the ranking of the pre-trained large model.

[0134] The recommendation method provided in this embodiment first selects a preset number of features from the features corresponding to the requesting user and the features corresponding to the item to be recommended. Then, based on the relationship between the preset number of features and the requesting user or the item to be recommended, a knowledge graph in the form of triplets is constructed. Next, the candidate recommended items are converted into first text, the knowledge graph is converted into second text, and prompt words corresponding to both the first and second texts are generated. Finally, the prompt words are input into a pre-trained large model, which re-ranks the candidate recommended items based on the prompt words, obtains a ranking result, selects the target recommended item from the candidate recommended items based on the ranking result, and recommends the target recommended item to the requesting user. In this way, re-ranking the candidate recommended items based on the knowledge graph can further improve the accuracy of the recommendation.

[0135] Corresponding to the aforementioned embodiment of the recommendation method, this application also provides an embodiment of the recommendation apparatus.

[0136] An embodiment of the recommendation device proposed in this application can be applied to a recommendation device. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the recommendation device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 12 The diagram shown is a hardware structure diagram of the recommended device in this application, except that... Figure 12 In addition to the processor, memory, network interface, and non-volatile memory shown, the recommended device in the embodiment may also include other hardware depending on the actual function of the recommended device, which will not be described in detail here.

[0137] Figure 13 This is a schematic diagram of the structure of the recommended device embodiment provided in this application. Please refer to... Figure 13The apparatus provided in this embodiment includes an acquisition module 1310, a generation module 1320, a processing module 1330, and a recommendation module 1340; wherein,

[0138] The acquisition module 1310 is used to respond to a recommendation request, acquire the historical interaction data of the requesting user corresponding to the recommendation request, and determine the items to be recommended that have interacted with the requesting user based on the historical interaction data.

[0139] The generation module 1320 is configured to construct an interaction pair for each item to be recommended, representing the interaction behavior between the requesting user and the item to be recommended; wherein the interaction pair is a data pair consisting of the requesting user and the item to be recommended;

[0140] The generation module 1320 is further configured to construct a default hypergraph without hyperedges for each interaction pair, using the features corresponding to the interaction pair as nodes.

[0141] The processing module 1330 is used to input the default hypergraph corresponding to the interaction pair into the hypergraph learning model, so that the hyperedge generation module 1320 of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair, and the prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user based on the complete hypergraph containing the hyperedges corresponding to the interaction pair.

[0142] The recommendation module 1340 is used to select candidate recommendations from the recommendations corresponding to each interaction pair based on the probability that the requesting user likes the recommendations corresponding to each interaction pair, and to make recommendations to the requesting user based on the candidate recommendations.

[0143] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0144] Optionally, the process by which the hyperedge generation module of the hypergraph learning model predicts the set of hyperedges of the default hypergraph includes:

[0145] For each node in the default hypergraph, the initial feature vector of the node is obtained, and the fused feature vector of the node is obtained based on the initial feature vector of the node and the initial feature vectors of other nodes in the default hypergraph except for the node; wherein, the dimension of the fused feature vector is the same as the number of hyperedges contained in the hyperedge set to be predicted.

[0146] The fused feature vectors of all nodes in the default hypergraph are concatenated into a node representation matrix;

[0147] An activation function is used to fuse the uniform random variable and each element in the node representation matrix and map them to binary values ​​to obtain an incidence matrix for representing the hyperedge set; wherein each column of the incidence matrix represents a predicted hyperedge, and each element in the column indicates whether the hyperedge is connected to each node in the default hypergraph.

[0148] Optionally, the hypergraph learning module of the hypergraph learning model includes a predetermined number of sequentially connected learning units and a fusion unit; the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing hyperedges corresponding to the interaction pair, including:

[0149] The first learning unit in the preset number of learning units generates an embedding representation of the hyperedge in the learning unit based on the fusion feature vectors of the nodes connected to the hyperedge in the hyperedge set, and updates the fusion feature vector of the node based on the embedding representations of the hyperedges connected to the node, thereby obtaining the embedding representation of the node in the learning unit.

[0150] In the preset number of learning units, excluding the first learning unit, the other learning units generate the embedding representation of the hyperedge in the learning unit based on the embedding representation of each node connected to the hyperedge in the previous learning unit before the current learning unit, and update the fused feature vector of the node based on the embedding representation of each hyperedge connected to the node in the previous learning unit before the current learning unit, thus obtaining the embedding representation of the node in the current learning unit.

[0151] The fusion unit performs weighted processing on the embedding representation of each hyperedge in the hyperedge set on each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each hyperedge.

[0152] The fusion unit performs weighted processing on the embedding representation of each node of the default hypergraph in each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each node.

[0153] The fusion unit uses the weighted embedding representations of each hyperedge and each node to aggregate and generate the complete hypergraph.

[0154] Optionally, the recommendation module 1340 is further configured to select a preset number of features from the features corresponding to the requesting user and the features corresponding to the item to be recommended;

[0155] The recommendation module 1340 is also used to construct a knowledge graph in the form of triples based on the relationship between the preset number of features and the requesting user or the item to be recommended;

[0156] The recommendation module 1340 is further configured to convert the candidate recommendations into first text, convert the knowledge graph into second text, and generate prompt words that correspond to both the first text and the second text.

[0157] The recommendation module 1340 is further configured to input the prompt words into a pre-trained large model, so that the pre-trained large model can re-rank the candidate recommendations based on the prompt words to obtain a ranking result, and select a target recommendation from the candidate recommendations based on the ranking result; and recommend the target recommendation to the requesting user.

[0158] Optionally, the training process of the hypergraph learning model includes: training the embedding representations of hyperedges and nodes based on the supervised mutual information maximization criterion, and training the embedding representations of hyperedges based on the mutual information minimization criterion, so as to maximize the mutual information between the embedding representations of hyperedges and nodes and minimize the mutual information between the embedding representations of each hyperedge.

[0159] Please continue to refer to Figure 12 This application also provides a recommended device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods provided in the first aspect of this application.

[0160] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in this application.

[0161] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0162] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0163] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A recommendation method, characterized in that, The method includes: In response to a recommendation request, the system obtains the historical interaction data of the requesting user corresponding to the recommendation request, and determines the recommended items that have interacted with the requesting user based on the historical interaction data. For each item to be recommended, an interaction pair is constructed to characterize the interaction behavior between the requesting user and the item to be recommended; wherein, the interaction pair is a data pair consisting of the requesting user and the item to be recommended; For each interaction pair, a default hypergraph without hyperedges is constructed using the features corresponding to the interaction pair as nodes. The default hypergraph corresponding to the interaction pair is input into the hypergraph learning model, so that the hyperedge generation module of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair. Based on the complete hypergraph containing the hyperedges corresponding to the interaction pair, the prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user. Based on the probability that the requested user likes the recommended items corresponding to each interaction pair, candidate recommended items are selected from the recommended items corresponding to each interaction pair, and recommendations are made to the requested user based on the candidate recommended items.

2. The method according to claim 1, characterized in that, The process by which the hyperedge generation module of the hypergraph learning model predicts the set of hyperedges of the default hypergraph includes: For each node in the default hypergraph, the initial feature vector of the node is obtained, and the fused feature vector of the node is obtained based on the initial feature vector of the node and the initial feature vectors of other nodes in the default hypergraph except for the node; wherein, the dimension of the fused feature vector is the same as the number of hyperedges contained in the hyperedge set to be predicted. The fused feature vectors of all nodes in the default hypergraph are concatenated into a node representation matrix; An activation function is used to fuse the uniform random variable and each element in the node representation matrix and map them to binary values ​​to obtain an incidence matrix for representing the hyperedge set; wherein each column of the incidence matrix represents a predicted hyperedge, and each element in the column indicates whether the hyperedge is connected to each node in the default hypergraph.

3. The method according to claim 1, characterized in that, The hypergraph learning module of the hypergraph learning model includes a predetermined number of sequentially connected learning units and a fusion unit; the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing hyperedges corresponding to the interaction pair, including: The first learning unit in the preset number of learning units generates an embedding representation of the hyperedge in the learning unit based on the fusion feature vectors of the nodes connected to the hyperedge in the hyperedge set, and updates the fusion feature vector of the node based on the embedding representations of the hyperedges connected to the node, thereby obtaining the embedding representation of the node in the learning unit. In the preset number of learning units, excluding the first learning unit, the other learning units generate the embedding representation of the hyperedge in the learning unit based on the embedding representation of each node connected to the hyperedge in the previous learning unit before the current learning unit, and update the fused feature vector of the node based on the embedding representation of each hyperedge connected to the node in the previous learning unit before the current learning unit, thus obtaining the embedding representation of the node in the current learning unit. The fusion unit performs weighted processing on the embedding representation of each hyperedge in the hyperedge set on each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each hyperedge. The fusion unit performs weighted processing on the embedding representation of each node of the default hypergraph in each learning unit based on the weights pre-set for each learning unit, to obtain the weighted embedding representation of each node. The fusion unit uses the weighted embedding representations of each hyperedge and each node to aggregate and generate the complete hypergraph.

4. The method according to claim 1, characterized in that, After selecting candidate recommendations, the method further includes: Select a preset number of features from the features corresponding to the requesting user and the features corresponding to the item to be recommended; Based on the relationship between the preset number of features and the requesting user or the item to be recommended, a knowledge graph in the form of triples is constructed. The candidate recommendations are converted into first text, and the knowledge graph is converted into second text; Generate prompt words that correspond to both the first and second texts; The prompt words are input into a pre-trained large model, which then re-ranks the candidate recommendations based on the prompt words to obtain a ranking result, and selects the target recommendation from the candidate recommendations based on the ranking result. The target recommendation is then presented to the requesting user.

5. The method according to claim 1, characterized in that, The loss function of the hypergraph learning model is: ， Where λ1, λ2, and λ3 are weighting factors; N is the number of nodes in the hypergraph; For predictions by the hypergraph learning model; y n For a node; D1 and D2 are the L1 norms of the model parameters; D1 and D2 are discriminators that distinguish between joint and marginal distributions. ω is the cross-entropy loss function for binary classification; ω is a learnable parameter. To perform regularization on the hypergraph learning model; ， in, Represented as nodes; This represents the temperature value; α and β are interval parameters.

6. The method according to claim 3, characterized in that, The training process of the hypergraph learning model includes: Based on the supervised mutual information maximization criterion, the embedding representations of hyperedges and nodes are trained, and based on the mutual information minimization criterion, the embedding representations of hyperedges are trained to maximize the mutual information between the embedding representations of hyperedges and nodes and minimize the mutual information between the embedding representations of each hyperedge.

7. A recommended device, characterized in that, The device includes an acquisition module, a generation module, a processing module, and a recommendation module; wherein, The acquisition module is used to respond to a recommendation request, acquire the historical interaction data of the requesting user corresponding to the recommendation request, and determine the items to be recommended based on the historical interaction data that have interacted with the requesting user. The generation module is used to construct an interaction pair for each item to be recommended, representing the interaction behavior between the requesting user and the item to be recommended; wherein, the interaction pair is a data pair consisting of the requesting user and the item to be recommended; The generation module is further configured to construct a default hypergraph without hyperedges for each interaction pair, using the features corresponding to the interaction pair as nodes. The processing module is used to input the default hypergraph corresponding to the interaction pair into the hypergraph learning model, so that the hyperedge generation module of the hypergraph learning model predicts the hyperedge set of the default hypergraph, and the hypergraph learning module of the hypergraph learning model models the default hypergraph based on the hyperedge set to obtain the complete hypergraph containing the hyperedges corresponding to the interaction pair, and the prediction module of the hypergraph learning model predicts the probability that the recommended item corresponding to the interaction pair is liked by the requesting user based on the complete hypergraph containing the hyperedges corresponding to the interaction pair. The recommendation module is used to select candidate recommendations from the recommendations corresponding to each interaction pair based on the probability that the requesting user likes the recommendations corresponding to each interaction pair, and to make recommendations to the requesting user based on the candidate recommendations.

8. The apparatus according to claim 7, characterized in that, The process by which the hyperedge generation module of the hypergraph learning model predicts the set of hyperedges of the default hypergraph includes: For each node in the default hypergraph, the initial feature vector of the node is obtained, and the fused feature vector of the node is obtained based on the initial feature vector of the node and the initial feature vectors of other nodes in the default hypergraph except for the node; wherein, the dimension of the fused feature vector is the same as the number of hyperedges contained in the hyperedge set to be predicted. The fused feature vectors of all nodes in the default hypergraph are concatenated into a node representation matrix; An activation function is used to map the uniform random variable and each element in the node representation matrix to binary values ​​to obtain an incidence matrix for representing the hyperedge set; wherein each column of the incidence matrix represents a predicted hyperedge, and each element in the column indicates whether the hyperedge is connected to each node in the default hypergraph.

9. A recommended device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the steps of the method according to any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images