Article recommendation method and device, storage medium and intelligent equipment

Abstract

The invention is suitable for the technical field of intelligent recommendation, and provides an article recommendation method and device, a storage medium and intelligent equipment, and the method comprises the steps: obtaining the identification information of a target user; determining a user representation corresponding to the identification information of the target user in a trained target hypergraph; based on the user representation and the article representation in the target hypergraph, calculating the degree of interest of the target user in the article; and based on a sorting result of the interesting degrees, performing article recommendation for the target user. According to the method, efficient and personalized recommendation can be provided for the user, and the recommendation accuracy and the satisfaction degree of the user are improved.

Description

Technical Field

[0001] The present application belongs to the field of intelligent recommendation technology, and in particular, relates to an item recommendation method, apparatus, storage medium and intelligent device. Background Art

[0002] The rapid development of Internet social media and network information technology has generated massive amounts of data, including a large amount of interactive information, such as academic paper-citation networks and e-commerce platform user-product interaction networks. These interactive information are also widely present in short video platforms, social software, news websites, music and movies and other fields. Through the recommendation system, merchants can recommend popular products, videos, music, etc. to potential users to increase sales and user stickiness.

[0003] However, with the exponential growth of network resources, the problem of information overload is becoming increasingly serious. When faced with massive amounts of information, it is difficult for users to find content that they are really interested in. How to effectively extract key information resources from interactive information, provide users with efficient and personalized recommendations, and improve the accuracy of recommendations and user satisfaction are issues that need to be considered. Summary of the invention

[0004] The embodiments of the present application provide an item recommendation method, apparatus, storage medium and intelligent device, which can provide users with efficient and personalized recommendations, improve the accuracy of recommendations and user satisfaction.

[0005] In a first aspect, an embodiment of the present application provides an item recommendation method, comprising:

[0006] Obtain identification information of the target user;

[0007] In the trained target hypergraph, determining a user representation corresponding to the identification information of the target user;

[0008] Calculating the interest level of the target user in the item based on the user representation and the item representation in the target hypergraph;

[0009] Based on the ranking results of the interest levels, item recommendations are made for the target user.

[0010] In a possible implementation manner of the first aspect, before the step of determining, in the trained target hypergraph, a user representation corresponding to the identification information of the target user, the step includes:

[0011] Constructing a multimodal dataset, wherein the multimodal dataset includes multimodal features of items and multimodal features of users as well as interaction information between users and items;

[0012] Constructing a hypergraph based on the multimodal features of the item, the multimodal features of the user, and the interaction information;

[0013] Using hypergraph convolution and multi-head attention mechanism to train the hypergraph;

[0014] When the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed.

[0015] In a possible implementation manner of the first aspect, the step of training the hypergraph using hypergraph convolution and multi-head attention mechanism includes:

[0016] Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the user in the hypergraph to obtain user aggregated features of the user;

[0017] Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the objects in the hypergraph to obtain the item aggregation features of the items.

[0018] In a possible implementation manner of the first aspect, the step of training the hypergraph using hypergraph convolution and multi-head attention mechanism includes:

[0019] A multi-head attention mechanism is used to fuse the user aggregation features between multiple modalities in the hypergraph to obtain user representation;

[0020] A multi-head attention mechanism is used to fuse the aggregated features of multi-modal items in the hypergraph to obtain item representation.

[0021] In a possible implementation manner of the first aspect, the step of training the hypergraph using hypergraph convolution and multi-head attention mechanism further includes:

[0022] Inputting the item multimodal features, the user multimodal features and the interaction information into the hypergraph for forward propagation to obtain a predicted value of the user's interest in the item;

[0023] Calculate the error between the predicted value and the true value using the target loss function;

[0024] Model parameters of the hypergraph are adjusted based on the error.

[0025] In a possible implementation manner of the first aspect, the step of calculating the target user's interest in the item based on the user representation and the item representation in the target hypergraph includes:

[0026] The user's interest in the item is calculated according to the following formula:

[0027]

[0028] in, represents the predicted value of user u’s interest in item i, Z u represents the user representation, Z i Represents an item.

[0029] In a second aspect, an embodiment of the present application provides an item recommendation device, comprising:

[0030] An identification information acquisition unit, used to acquire identification information of a target user;

[0031] A user representation determination unit, configured to determine a user representation corresponding to the identification information of the target user in a trained target hypergraph;

[0032] an interest calculation unit, configured to calculate the interest degree of the target user in the item based on the user representation and the item representation in the target hypergraph;

[0033] An item recommendation unit is used to recommend items to the target user based on the ranking result of the interest level.

[0034] In a third aspect, an embodiment of the present application provides an intelligent device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the item recommendation method as described in the first aspect above is implemented.

[0035] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the item recommendation method as described in the first aspect above is implemented.

[0036] In a fifth aspect, an embodiment of the present application provides a computer program product. When the computer program product is run on a smart device, the smart device executes the item recommendation method as described in the first aspect above.

[0037] In an embodiment of the present application, by obtaining the identification information of the target user, the user representation corresponding to the identification information of the target user is determined in the trained target hypergraph, and then the interest level of the target user in the item is calculated based on the user representation and the item representation in the target hypergraph, and then the item recommendation is made to the target user based on the ranking result of the interest level. The present application mines richer high-order semantic information about users and items based on the hypergraph, which can improve the accuracy of the relationship between users and items, and provide users with efficient and personalized item recommendations, thereby improving the accuracy of item recommendations and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 is a flowchart of an implementation of the item recommendation method provided in an embodiment of the present application;

[0040] Figure 2 It is a specific implementation flow chart of constructing and training a target hypergraph in the item recommendation method provided in the embodiment of the present application;

[0041] Figure 3 This is a specific implementation flowchart of using hypergraph convolution to train a hypergraph in the item recommendation method provided in the embodiment of the present application;

[0042] Figure 4 This is a specific implementation flowchart of using a multi-head attention mechanism to train a hypergraph in the item recommendation method provided in an embodiment of the present application;

[0043] Figure 5 is a structural block diagram of an item recommendation device provided in an embodiment of the present application;

[0044] Figure 6 It is a schematic diagram of a smart device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0046] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0047] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0048] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0049] In order to improve the accuracy of item recommendations and user satisfaction, in the embodiments of the present application, the smart device obtains the identification information of the target user, determines the user representation corresponding to the identification information of the target user in the trained target hypergraph, uses the hypergraph to fully mine the high-order semantic information of the user and the item, and then calculates the target user's interest level in the item based on the user representation and the item representation in the target hypergraph. Finally, based on the ranking results of the interest level, item recommendations are made to the user, thereby achieving efficient and personalized item recommendations.

[0050] The embodiments of the present application provide an item recommendation method, which is applicable to various types of smart devices or servers that need to perform item recommendation. The smart devices may specifically include mobile phones, tablet computers, wearable devices, laptop computers, vehicle-mounted devices, augmented reality (AR) devices, virtual reality (VR) devices, personal digital assistants (PDA), digital televisions, etc.

[0051] The item recommendation method provided in this application is exemplarily described below in conjunction with specific embodiments.

[0052] Figure 1 The implementation process of the item recommendation method provided in an embodiment of the present application is shown, and the method process may include the following steps S101 to S104.

[0053] Step S101: Obtain identification information of a target user.

[0054] The target user is the user for whom the item recommendation is to be made. The identification information refers to information that can be used to identify the identity of the user. The identification information is unique. For example, the identification information can be a user ID.

[0055] In this embodiment, when it is necessary to recommend items to a target user, identification information of the target user is obtained.

[0056] Step S102: Determine the user representation corresponding to the identification information of the target user in the trained target hypergraph.

[0057] The target hypergraph can be used to recommend personalized items to users. The target hypergraph is a hypergraph model pre-trained based on a multimodal dataset and can be used to mine high-order semantic information between users and items and implement item recommendations.

[0058] A hypergraph consists of vertices and hyperedges. A hypergraph is a graph model used to construct complex relationships between users and items. In a hypergraph, an edge (called a hyperedge) can connect any number of vertices, unlike a simple graph that only allows edges to connect two vertices. In a hypergraph, a vertex represents an item, a hyperedge represents a user, and a hyperedge can contain any number of vertices, indicating that the user is interested in or has interacted with the items (also called projects) represented by these vertices. A vertex may also be connected by multiple hyperedges.

[0059] As an embodiment of the present application, Figure 2 A specific implementation process of constructing and training a target hypergraph in the item recommendation method provided in an embodiment of the present application is shown, and is described in detail as follows:

[0060] Step S201: construct a multimodal data set, which includes multimodal features of items and multimodal features of users, as well as interaction information between users and items.

[0061] In this embodiment, the data sets used for training include Tiktok data sets, Allrecipes data sets, Amazon's baby and sports data sets, etc. User multimodal features and item multimodal features are obtained. The user multimodal features are multimodal features of the user, and the item multimodal features are multimodal features of the item.

[0062] The multimodal features of an item are the features of multiple models of an item, and the multimodal features of a user are the features of multiple modes of a user. In one embodiment, the operation record of the user is obtained, and based on the operation record, the item with which the user interacts is determined, and the multimodal features of the user and the item are extracted respectively. The operation record includes but is not limited to records of browsing, collecting, commenting, liking, and purchasing. The user refers to a potential user for whom the item recommendation is to be executed. It is worth noting that the target user is one of the users.

[0063] The multimodal features of an item usually include text, images, audio and other modalities, which can be obtained by preprocessing such as BERT and VGG. The user's features include their occupation, age, gender, etc. The user's multimodal features can be obtained by the encoder. If the user's multimodal features are missing, the user's features can be obtained by convolution of the corresponding item's features in the subsequent step.

[0064] In the embodiment of the present application, by acquiring the multimodal features of users and the multimodal features of items, the user features and the item features are fully extracted, which is beneficial to improving the effectiveness of the hypergraph training, and can help to achieve personalized recommendations and improve the accuracy of item recommendations.

[0065] In this embodiment, the interaction information between the user and the item is obtained. The interaction information between the user and the item includes but is not limited to the user's browsing, collection, comment, like and purchase records, which can be regarded as user-item interaction information. The user connects the items associated with the user to form a hyperedge in the hypergraph.

[0066] In some implementations, the image features of the item are extracted using computer vision algorithms such as VGG, and the text description features or short video audio features of the item are extracted using natural language processing methods such as BERT to obtain the multimodal features of the item. Where m represents the modal type, i represents the item, and n represents the v represents the number of items, d is the embedding vector, and R represents a real number set. User multimodal features can be obtained by embedding identity information, browsing, favorites, comments, likes, and purchases. For example, the encoder is trained to encode gender, age, work, and operation context into an embedding vector. If this information is missing, the corresponding information of the item with which the user interacts can be used to convolve and generate the corresponding user features. Similarly, m represents the modal type, n u Indicates the number of users. The interaction information between the user and the item can be determined based on the interactions between the user and the item. For example, which items the user has interacted with is recorded as (u i ,i1,i2,i3,...i n ), each user has an indefinite number of item interactions, a total of n u There are records like this, where n represents the number of users.

[0067] In some embodiments, the multimodal data set includes a positive sample data subset and a negative sample data subset, wherein the positive sample data subset includes the user's multimodal features and the multimodal features of the item that has a sample interaction relationship with the user, and the negative sample data subset includes the user's multimodal features and the multimodal features of the item that has no sample interaction relationship with the user.

[0068] Step S202: construct a hypergraph based on the multimodal features of the item, the multimodal features of the user, and the interaction information.

[0069] The multimodal feature is specifically manifested in that each vertex (items) and each hyperedge (user) in the hypergraph contains multiple modal attribute information, which can be used for intra-modal and inter-modal information transfer in subsequent processing. Intra-modal information here specifically refers to the communication and update of information within a certain modality, such as within the text modality. Inter-modal information transfer refers to the cross-modal information transfer between multiple modalities, such as text, image, and audio modalities.

[0070] In this embodiment, with item items as vertices and users as hyperedges, a typical hypergraph can be recorded as G m = {V G ,ε G , W, H G}. V G represents the vertex set, ε G represents the hyperedge set, W represents the model parameter set, and H represents the association matrix set of the interaction relationship. The attribute features of each vertex v include the multimodal features of historical items, including features of multiple modalities such as text, video, and audio. Each hyperedge ε i (u,i1,i2,i3,...i n ) represents the complete interaction of the user, including the user's multimodal features, which connects all types of item vertices involved in the interaction. For example, a user collects 5 items and builds a hyperedge to connect these 5 item objects. The combination of these item objects i1, i2, i3, i4, and i5 is the complete interaction ε1. In addition, you can also choose to attach the interaction environment, such as the spatiotemporal context, to the hyperedge features. Therefore, the features of such a hyperedge and its connected vertices together can represent a complete interaction instance of a user. A hypergraph G m Contains |V G |=n v vertices, |ε G |=n ε A super edge.

[0071] Representing the incidence matrix of a hypergraph, it is defined as follows:

[0072]

[0073] Among them, h tk is the incidence matrix H G An element in, v and ε are the vertices and hyperedges of the hypergraph G, respectively, <nv And k <n ε , representing the numbers of vertices and hyperedges respectively. represents the trainable information aggregation weight during convolution in the hypergraph. It corresponds one-to-one to the non-zero elements in the association matrix and is initialized to 1, indicating that each vertex is equally important. i The set of vertices belonging to the hypergraph G, namely v i ∈V G , then the vertex v i The degree is defined as the sum of the i The number of hyperedges; hyperedge ε j The degree of D is defined as the number of vertices contained in the hyperedge. V and D ε They represent the degree matrices of vertices and hyperedges respectively, and the degree matrices are used to calculate the Laplacian matrix of the hypergraph.

[0074] The relationships in the real world are often intricate and cross-influenced, and multiple vertices play a role together. For example, in the citation relationship of a paper, multiple authors jointly create an article, or there is a many-to-many relationship between a movie and an actor. These high-order relationships form a whole. The original simple graph structure is difficult to fit this relationship, and the original graph recommendation algorithm also performs poorly. In this embodiment, a hypergraph is constructed. The hypergraph extends the definition of the graph, and a hyperedge in it can contain any number of vertices, so high-order associations can be directly stored, and each hyperedge can define a weight w. Hypergraphs have a more powerful expression ability for complex relationships, so the recommendation modeling based on hypergraphs is more in line with the real network environment. Compared with the common bipartite graphs before, it can better ensure the integrity of information, can use richer high-order interactive semantic information, and can use its variable degree hyperedges to encode high-order data correlations (beyond pairwise connections), which can be easily extended to multimodal and heterogeneous data representation.

[0075] Step S203: using hypergraph convolution and multi-head attention mechanism to train the hypergraph.

[0076] After constructing the hypergraph, the next task is to fully learn the high-level information of user-item interactions, obtain the latent embedding vector representations of users and items respectively, and then use the latent vectors of both to calculate which items the user may be interested in. Therefore, choosing an appropriate information transmission method is crucial, as it has an important impact on the quality of vector generation and subsequent recommendation effects.

[0077] As an embodiment of the present application, Figure 3 A specific implementation process of using hypergraph convolution to train a hypergraph in the item recommendation method provided in an embodiment of the present application is shown, and is described in detail as follows:

[0078] A1: Perform intra-modal feature aggregation on the multi-modal features of the user in the hypergraph based on the hypergraph convolution to obtain user aggregated features of the user.

[0079] A2: Perform intra-modal feature aggregation on the multi-modal features of the objects in the hypergraph based on hypergraph convolution to obtain item aggregation features of the items.

[0080] Specifically, the Hypergraph Neural Networks (HGNN) method is used to transmit user-item information exchange.

[0081] Graph learning usually uses spectral domain convolution to transfer information X, which is necessary when learning the final representation of users and items in the recommendation model. The early convolution theorem spectral convolution operator has high computational complexity. Later, with continuous research and improvement, the k-order Chebyshev polynomial approximation that only considers k-order neighbors and the graph convolution GCN that only considers 1-order neighbors were proposed. The convolution formula can be expressed as follows (2):

[0082]

[0083] Where X l Represents the vertex feature matrix after each layer of convolution, (A is the adjacency matrix of the graph, I is the identity matrix), D is the degree matrix, W is the trainable parameter, and σ is the activation function. Similarly, the hypergraph convolution can be simplified by methods such as truncated Chebyshev to obtain the following formula (3):

[0084]

[0085] The calculation of signal x and filter g can be simplified as above, where θ is the parameter, k is the polynomial order, W is the trainable hyperedge weight, and D v and D ε They represent the degree matrix of the vertex and the degree matrix of the hyperedge respectively, and H is the association matrix in the hypergraph. So the hypergraph convolution formula of the lth layer is:

[0086]

[0087] Where X l represents the vertex feature matrix after the lth layer of convolution, σ represents the activation function, and the filter Θ is the parameter matrix to be learned during training.

[0088] In this embodiment, hypergraph convolution is used to transfer information within the modality and mine high-order information. The hypergraph convolution operation aggregates the information of neighboring vertices to update the representation of the current vertex, which can be regarded as the process of transferring feature X. The output of the previous layer is used as the input of the next layer, and multi-layer convolution mines high-order information.

[0089] The constructed hypergraph includes the incidence matrix, degree matrix, and vertex feature matrix. There are multiple modal attributes, each of which has its own characteristics. Perform hypergraph convolution. Through this intra-modal information transfer, accumulate the outputs of each layer to obtain new feature representations of vertices (items) and hyperedges (users) in each modality. Taking mode m1 as an example, assuming there are p convolution layers, X m Representing the feature representation of the vertex or hyperedge, the total feature matrix after p-layer convolution of the vertex or hyperedge under the final mode m1 can be calculated as follows (5):

[0090]

[0091] Convolution operations are often used to aggregate information from neighboring vertices to update the representation of the current vertex. This aggregation process can be viewed as a process of feature X transfer, as it integrates the feature information of neighboring vertices into the current vertex. In multi-layer convolution operations, this feature transfer process becomes more complicated. For example, the first layer of convolution may only aggregate the information of first-order neighbors, while multi-layer convolutions can extract higher-order potential connections by combining information from multi-order neighbors. This feature of multi-layer convolution enables the network to learn more complex interaction patterns between vertices, thereby improving the expressiveness and generalization capabilities of the model.

[0092] As an embodiment of the present application, Figure 4 A specific implementation process of using a multi-head attention mechanism to train a hypergraph in the item recommendation method provided in an embodiment of the present application is shown, and is described in detail as follows:

[0093] B1: A multi-head attention mechanism is used to fuse the user aggregation features between multiple modalities in the hypergraph to obtain user representation.

[0094] B2: A multi-head attention mechanism is used to fuse the aggregated features of multi-modal items in the hypergraph to obtain item representation.

[0095] User representation and item representation are new hyperedge representation features and new vertex representation features of inter-modal feature fusion.

[0096] In this embodiment, the multi-head attention mechanism is used for information interaction and fusion between modalities. Information transmission within the modality alone is not sufficient. For example, some users particularly like the combination of background music and short videos. Then, the single recommendation of audio or video alone may not be able to discover this preference. Therefore, the multi-head attention mechanism is used to interact and fuse information between modalities. Specifically, the multi-head attention network is used to explore the potential correlation between multi-modal matrices, learn weights to aggregate multi-modal feature matrices, and obtain the final embedding matrix of users and items.

[0097] In a possible implementation, W i Q ,W i K ,W i V is a trainable weight matrix that maps input to query (Q), key (K), and value (V). Q, K, and V are general matrices in the attention mechanism. k Is the dimension of the query or key. Z is the representation.

[0098] The multi-head attention mechanism can be concatenated as follows (6):

[0099] MultiHead(X)=Concat(head1,...head h )W o (6)

[0100] Taking the vertex (items) as an example, the hypergraph outputs the text modality representation X of the vertex after the hypergraph convolution. text With visual modality representation X image , X text With X image The two feature representations are input into the multi-head attention network, and the fusion output is the final vertex feature representation Z i , which is the item representation. Similarly, the user representation Z can be obtained for the hyperedge u .

[0101] In one possible implementation, the multimodal features of the item, the multimodal features of the user, and the interaction information are input into the hypergraph for forward propagation to obtain a predicted value of the user's interest in the item; an objective loss function is used to calculate the error between the predicted value and the true value; and the model parameters of the hypergraph are adjusted based on the error.

[0102] In a possible implementation, the user's interest in the item is predicted according to the following formula (7):

[0103]

[0104] in, represents the predicted value of user u’s interest in item i, Z u represents the user representation, Z i Represents an item.

[0105] In this embodiment, the target loss function adopts the BPR (Bayesian Personalized Ranking) loss function, which uses the user-item potential feature matrix to perform sorting preferences. Given a user and multiple items, BPR ranks items that the user prefers before items that the user dislikes less. It optimizes through the Bayesian maximum a posteriori probability to maximize the difference in scores between positive samples and negative samples, thereby learning the user's personalized preferences.

[0106] In one possible implementation, positive and negative samples are sampled during training, where the positive sample is item i with which the user interacts, and the negative sample is item j with which the user does not interact. The data set is a plurality of triplets (u, i, j). The purpose is to require user u to prioritize item i over item j. The target loss function is calculated as follows:

[0107]

[0108] in Represents the predicted value of user u's interest in item i. Similarly, represents the predicted value of user u’s interest in item j, and ε represents the hyperedge. Based on the above formula (8), when the prediction is correct, that is, the user’s interest in positive sample i is greater than that in negative sample j, the loss L BPR Will be smaller.

[0109] Step S204: When the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed.

[0110] Repeat the steps of training the hypergraph using hypergraph convolution and multi-head attention mechanism. When the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed to obtain the target hypergraph.

[0111] In one possible implementation, each batch of the Hypergraph model input will optimize the user representation and item representation. When the improvement value of the Hypergraph iterative training effect is less than the set threshold, or the model training effect has not improved for a specified number of consecutive times, the model converges and the training is terminated. For example, if the evaluation effect has not improved for 7 consecutive times or the improvement is less than the set threshold, the training is terminated.

[0112] In one possible implementation, it is determined whether the batch training is terminated. If the set number of rounds is reached, the current batch cycle is terminated, otherwise the batch training is continued.

[0113] In an embodiment of the present application, a multimodal dataset is constructed by acquiring multimodal features of users and multimodal features of items, and interaction information between users and the items. The multimodal data is used as a training set for hypergraph training, and the hypergraph is trained using hypergraph convolution and multi-head attention mechanisms to acquire user representations of users and item representations of items, thereby mining richer high-order semantic information about users and items. The target loss function is used to effectively calculate the difference between the predicted value and the true value during the training process, thereby effectively optimizing the model parameters.

[0114] The effect of the hypergraph model in the embodiment of the present application can be further illustrated by the following simulation:

[0115] This example is tested on an Intel(R) Xeon(R) Silver 4215R CPU@3.20GHz Linux system, equipped with a GPU chip model of NVIDIA GeForce RTX 3090, and the experimental platform framework pytorch_1.13.1. The other major environment versions are configured as Python-3.9, CUDA-11.1, and dgl-1.1.2.

[0116] In this example, we select Tiktok dataset, Allrecipes dataset, Amazon's baby and sports datasets. The specific dataset information dimensions can be seen in Table 1:

[0117] Table 1

[0118]

[0119] The Tiktok dataset records short videos watched by users. Multimodal features include the visual, acoustic, and title text embedding features of the video; the Allrecipes dataset comes from a food-based social networking platform and includes more than 50,000 recipes in 27 different categories. Each recipe image and ingredient can be abstractly encoded as visual and text features; Amazon's baby and sports project category datasets, multimodal data includes the embedding features of product images and product titles, brand classifications, and other text.

[0120] The dataset is divided into batches of 1024 interactive training data, and the maximum iteration threshold is set to 1000 times. The user-item interaction relationship data is used to build a hypergraph structure, that is, the connection relationship between vertices and hyperedges. Multimodal features are used to build vertex attribute information, and hyperedge attribute information can be obtained by convolution of associated vertices. The convolution layer parameters are set to 2 layers, and the dimensions of the hidden layer and output layer are set to 64. The number of attention head parameters of the multi-head attention network is 4. The learning rate is initialized to 0.00055, and as the training progresses, the learning rate decays by 95% every 50 epochs.

[0121] In this embodiment, four recommendation system evaluation indicators are used, namely Precision, Recall, Hit Ratio and NDCG (Normalized Discounted Cumulative Gain).

[0122] Precision refers to the proportion of recommended items that users are actually interested in, that is, the ratio of the number of predicted positive examples to the number of all predicted positive examples. In the context of recommendation systems, it is usually defined as the proportion of items recommended to users that users actually like or have interacted with.

[0123] TP (True Positive) refers to the number of correctly recommended items, and FP (False Positive) refers to the number of incorrectly recommended items. A high precision rate means that the system is more likely to accurately recommend items that the user is interested in, but may miss some potential areas of interest.

[0124] The recommendation effect comparison experiment of the embodiment of the present application was conducted with two graph convolution methods, namely LightGCN and MMGCN. The training set and the test set were divided on four data sets. The model training was completed on the training set, and the recommendation effect was tested on the test set. Four evaluation indicators were used for comparison, as shown in Table 2 and Table 3. It can be seen that the Precision evaluation index of the model of the present application on the four data sets is not worse than that of the comparison algorithm. Among them, the improvement effect is more obvious on the other three data sets except Allrecipes. The recommendation effect on the Allrecipes data set is close to that of the comparison algorithm, indicating that compared with the current baseline method, the probability that the items recommended by the present application meet the personalized interests of users is higher.

[0125] Table 2

[0126]

[0127] Table 3

[0128]

[0129] Recall corresponds to Precision, which measures the proportion of items that the system can successfully find out from all the items that should be recommended.

[0130] Among them, FN (False Negative) refers to the number of missed recommended items, that is, those items that the user actually likes but the system does not recommend. A high recall rate means that the system is more able to cover the user's interests, but may include some items that the user is not interested in.

[0131] According to Table 2 and Table 3, we can see that the Recall evaluation index of the model of this application is better than the comparison algorithm in most cases on the four data sets. In the other three data sets except Allrecipes, the recommendation effect is good. On the Allrecipes data set, the Recall is slightly worse than the MMGCN method, but the difference is very small. Therefore, overall, this application performs better in the Recall evaluation index, and the recommended items cover a wider range.

[0132] Hit Ratio is often used to measure the effectiveness of Top-K recommendations, that is, how many of the top K items in the recommendation list are actually of interest to the user. The numerator is usually the sum of the number of items in the test set in each user's Top-K recommendations, and the denominator is the total number of items in all test sets.

[0133] According to Table 2 and Table 3, we can see that the performance of the hit ratio evaluation index of the model in the four datasets is similar to the recall rate. In most cases, it is better than the comparison algorithm, and is slightly worse than the MMGCN comparison method on the Allrecipes dataset. This shows that the top k items recommended by the application can cover user interests more widely.

[0134] Normalized Discounted Cumulative Gain (NDCG) is a metric used to evaluate the ranking quality of a recommendation system. It compares the ranking position of each item in the recommendation system with the ranking position of the items that the user is actually interested in, calculates the Discounted Cumulative Gain of each item, and then normalizes it.

[0135] According to Table 2 and Table 3, it can be seen that the NDCG evaluation index of the model of this application is better than that of the comparison algorithm on the four data sets, and the advantages are obvious in most data sets. The comprehensive recommendation effect of this application is better.

[0136] In summary, this application uses hypergraph modeling, fully utilizes multimodal attribute information, adopts hypergraph convolutional neural network as the intra-modal information transmission method, uses multi-head attention mechanism for inter-modal information transmission, and deeply explores and integrates high-order potential connections under each modality. According to the experimental results, our model recommends a wide range of items in a variety of real data sets, with high user interest, and its overall performance is better than the comparison algorithm.

[0137] Step S103: Calculate the interest level of the target user in the item based on the user representation and the item representation in the target hypergraph.

[0138] The degree of interest can be reflected by the recommendation score, which can be specifically calculated according to the above calculation formula (7). The higher the degree of interest, the higher the recommendation score.

[0139] Step S104: Based on the ranking results of the interest levels, recommend items to the target user.

[0140] Exemplarily, the items are sorted from high to low according to the above-mentioned interest levels, a number of items at the top of the sorting are selected to generate a recommendation list, and item recommendations are made to the user based on the recommendation list.

[0141] As can be seen from the above, in the embodiment of the present application, by obtaining the identification information of the target user, the user representation corresponding to the identification information of the target user is determined in the trained target hypergraph, and then the interest level of the target user in the item is calculated based on the user representation and the item representation in the target hypergraph, and then the item recommendation is made to the target user based on the ranking result of the interest level. The present application mines richer high-order semantic information about users and items based on the hypergraph, which can improve the accuracy of the relationship between users and items, and provide users with efficient and personalized item recommendations, thereby improving the accuracy of item recommendations and user satisfaction.

[0142] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0143] Corresponding to the item recommendation method described in the above embodiment, Figure 5 A structural block diagram of an item recommendation device provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0144] Reference Figure 5 The item recommendation device includes: an identification information acquisition unit 51, a user characterization determination unit 52, an interest calculation unit 53, and an item recommendation unit 54, wherein:

[0145] An identification information acquisition unit 51 is used to acquire identification information of a target user;

[0146] A user representation determination unit 52, configured to determine the user representation corresponding to the identification information of the target user in the trained target hypergraph;

[0147] An interest calculation unit 53, configured to calculate the interest level of the target user in the item based on the user representation and the item representation in the target hypergraph;

[0148] The item recommendation unit 54 is used to recommend items to the target user based on the ranking result of the interest level.

[0149] As a possible implementation of the present application, the item recommendation device further includes:

[0150] A data set construction unit, used to construct a multimodal data set, wherein the multimodal data set includes multimodal features of items and multimodal features of users as well as interaction information between users and items;

[0151] A hypergraph construction unit, configured to construct a hypergraph based on the multimodal features of the item, the multimodal features of the user, and the interaction information;

[0152] A hypergraph training unit is used to train the hypergraph using hypergraph convolution and multi-head attention mechanism; when the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed.

[0153] As a possible implementation manner of the present application, the hypergraph training unit includes a first training prediction module, which is used to:

[0154] Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the user in the hypergraph to obtain user aggregated features of the user;

[0155] Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the objects in the hypergraph to obtain the item aggregation features of the items.

[0156] As a possible implementation manner of the present application, the hypergraph training unit includes a second training prediction module, which is used to:

[0157] A multi-head attention mechanism is used to fuse the user aggregation features between multiple modalities in the hypergraph to obtain user representation;

[0158] A multi-head attention mechanism is used to fuse the aggregated features of multi-modal items in the hypergraph to obtain item representation.

[0159] As a possible implementation of the present application, the hypergraph training unit further includes a parameter adjustment module for:

[0160] Inputting the item multimodal features, the user multimodal features and the interaction information into the hypergraph for forward propagation to obtain a predicted value of the user's interest in the item;

[0161] Calculate the error between the predicted value and the true value using the target loss function;

[0162] Model parameters of the hypergraph are adjusted based on the error.

[0163] As a possible implementation of the present application, the above-mentioned interest calculation unit 53 is specifically used for:

[0164] The user's interest in the item is calculated according to the following formula:

[0165]

[0166] in, represents the predicted value of user u’s interest in item i, Z u represents the user representation, Z i Represents an item.

[0167] As can be seen from the above, in the embodiment of the present application, by obtaining the identification information of the target user, the user representation corresponding to the identification information of the target user is determined in the trained target hypergraph, and then the interest level of the target user in the item is calculated based on the user representation and the item representation in the target hypergraph, and then the item recommendation is made to the target user based on the ranking result of the interest level. The present application mines richer high-order semantic information about users and items based on the hypergraph, which can improve the accuracy of the relationship between users and items, and provide users with efficient and personalized item recommendations, thereby improving the accuracy of item recommendations and user satisfaction.

[0168] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0169] The present application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, Figures 1 to 4 The steps of any item recommendation method represented.

[0170] The embodiment of the present application also provides an intelligent device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, Figures 1 to 4 The steps of any item recommendation method represented.

[0171] The embodiment of the present application also provides a computer program product, when the computer program product is run on a smart device, the smart device executes the following Figures 1 to 4 The steps of any item recommendation method represented.

[0172] Figure 6 Schematic diagram of a smart device provided by an embodiment of the present application. Figure 6 As shown, the smart device 6 of this embodiment includes: a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60. When the processor 60 executes the computer program 62, the steps in the above-mentioned various item recommendation method embodiments are implemented, for example Figure 1Alternatively, when the processor 60 executes the computer program 62, the functions of each module / unit in the above-mentioned device embodiments are realized, for example Figure 5 The functions of units 51 to 54 are shown.

[0173] Exemplarily, the computer program 62 may be divided into one or more modules / units, which are stored in the memory 61 and executed by the processor 60 to complete the present application. The one or more modules / units may be a series of computer-readable instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 62 in the smart device 6.

[0174] The smart device 6 may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art will appreciate that Figure 6 It is only an example of the smart device 6 and does not constitute a limitation of the smart device 6. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the smart device 6 may also include input and output devices, network access devices, buses, etc.

[0175] The processor 60 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0176] The memory 61 may be an internal storage unit of the smart device 6, such as a hard disk or memory of the smart device 6. The memory 61 may also be an external storage device of the smart device 6, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the smart device 6. Further, the memory 61 may also include both an internal storage unit and an external storage device of the smart device 6. The memory 61 is used to store the computer program and other programs and data required by the smart device. The memory 61 may also be used to temporarily store data that has been output or is to be output.

[0177] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0178] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0179] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0180] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0181] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An item recommendation method, characterized in that: include: Obtain identification information of the target user; In the trained target hypergraph, determining a user representation corresponding to the identification information of the target user; Calculating the interest level of the target user in the item based on the user representation and the item representation in the target hypergraph; Based on the ranking results of the interest levels, item recommendations are made for the target user.

2. The item recommendation method according to claim 1, characterized in that: Before the step of determining the user representation corresponding to the identification information of the target user in the trained target hypergraph, the method includes: Constructing a multimodal dataset, wherein the multimodal dataset includes multimodal features of items and multimodal features of users as well as interaction information between users and items; Constructing a hypergraph based on the multimodal features of the item, the multimodal features of the user, and the interaction information; Using hypergraph convolution and multi-head attention mechanism to train the hypergraph; When the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed.

3. The item recommendation method according to claim 2, characterized in that: The step of training the hypergraph using hypergraph convolution and multi-head attention mechanism includes: Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the user in the hypergraph to obtain user aggregated features of the user; Based on the hypergraph convolution, intra-modal feature aggregation is performed on the multi-modal features of the objects in the hypergraph to obtain the item aggregation features of the items.

4. The item recommendation method according to claim 3, characterized in that: The step of training the hypergraph using hypergraph convolution and multi-head attention mechanism includes: A multi-head attention mechanism is used to fuse the user aggregation features between multiple modalities in the hypergraph to obtain user representation; A multi-head attention mechanism is used to fuse the aggregated features of multi-modal items in the hypergraph to obtain item representation.

5. The item recommendation method according to claim 2, characterized in that: The step of training the hypergraph using hypergraph convolution and multi-head attention mechanism also includes: Inputting the item multimodal features, the user multimodal features and the interaction information into the hypergraph for forward propagation to obtain a predicted value of the user's interest in the item; Calculate the error between the predicted value and the true value using the target loss function; Model parameters of the hypergraph are adjusted based on the error.

6. The item recommendation method according to any one of claims 1 to 5, characterized in that: The step of calculating the interest level of the target user in the item based on the user representation and the item representation in the target hypergraph comprises: The user's interest in the item is calculated according to the following formula: in, represents the predicted value of user u’s interest in item i, Z u represents the user representation, Z i Represents an item.

7. An item recommendation device, characterized in that: include: An identification information acquisition unit, used to acquire identification information of a target user; A user representation determination unit, configured to determine a user representation corresponding to the identification information of the target user in a trained target hypergraph; an interest calculation unit, configured to calculate the interest degree of the target user in the item based on the user representation and the item representation in the target hypergraph; An item recommendation unit is used to recommend items to the target user based on the ranking result of the interest level.

8. The item recommendation device according to claim 7, characterized in that: The item recommendation device further includes: A data set construction unit, used to construct a multimodal data set, wherein the multimodal data set includes multimodal features of items and multimodal features of users as well as interaction information between users and items; A hypergraph construction unit, configured to construct a hypergraph based on the multimodal features of the item, the multimodal features of the user, and the interaction information; A hypergraph training unit is used to train the hypergraph using hypergraph convolution and multi-head attention mechanism; when the training result of the hypergraph meets the preset training conditions, the hypergraph training is completed.

9. An intelligent device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the item recommendation method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the item recommendation method according to any one of claims 1 to 6 is implemented.

Images