A social constraint-based network video and audio intelligent recommendation method and system

Abstract

The present application relates to the technical field of artificial intelligence, in particular to a network video intelligent recommendation method and system based on social constraints. The recommendation method specifically includes five links of a heterogeneous network fusion stage, a model information extraction stage, a model construction stage, a model training stage and a preference prediction stage. Using model as the unit of interaction modeling and message propagation effectively captures the high-level semantics in the interaction network, significantly improves the accuracy of the recommendation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a network video intelligent recommendation method and system based on social constraints. BACKGROUND

[0002] With the rapid development of recommendation systems and the wide application of various intelligent terminals, the demand for personalized recommendation is becoming more and more prominent, showing a growing trend. In recent years, the rise of smart phones has driven the rapid development and popularization of recommendation systems. For network video platforms, intelligent recommendation technology can not only improve user experience and increase user stickiness, but also bring rich profit returns to enterprises, and has gradually become the core competitiveness of major platforms.

[0003] From the existing research and application situation, common intelligent recommendation technologies mainly include collaborative filtering recommendation and graph neural network recommendation. However, these technologies are usually limited to studying the pairwise relationship between users and items, and cannot meet the high-precision and diverse recommendation needs of old users; and these technologies face the challenge of data sparsity, reducing the user experience of new users. At the same time, more and more network video platforms have online friend interaction functions, focusing on mining the value between platform friends, bringing new opportunities and challenges to intelligent recommendation technology. Therefore, how to effectively use social relationships to achieve more accurate personalized recommendation in the context of big data has very important practical significance for network video platforms. SUMMARY

[0004] The technical problem to be solved by the present application is to design a network video intelligent recommendation method and system based on social constraints, which can effectively use social relationships to achieve high-precision personalized recommendation in the context of big data.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A network video intelligent recommendation method based on social constraints, which comprises four components, step S10 is a heterogeneous network fusion stage, step S20 is a model information extraction stage, step S30 is a model construction stage, step S40 is a model training stage, and step S50 is a preference prediction stage. The specific steps are as follows:

[0007] S10, collect historical data to generate an information database, and store and fuse the behavior interaction network and the social relationship network into a heterogeneous network in the form of graph data;

[0008] S20, extract social patterns, behavior patterns and joint patterns from the heterogeneous network obtained in step S10, and convert the input data into corresponding hypergraphs based on different patterns;

[0009] S30, modeling the hypergraph obtained in step S20 by using a hypergraph neural network, constructing a multi-channel hypergraph neural network model based on the hypergraph neural network, including a user vector representation module, an item vector representation module, a contrast learning module, and a preference prediction module;

[0010] S40, for the multi-channel hypergraph neural network model constructed in step S30, solving the model by an iterative calculation method to determine the optimal parameters of the model;

[0011] S50, applying the multi-channel hypergraph neural network model constructed in step S30 and the optimal parameters of the model determined in step S40 to generate a final recommendation prediction for the user.

[0012] Further, the step S10 specifically includes the following steps:

[0013] S11, collecting historical data from a network video and audio social platform to generate an information database, and taking the behavior interaction matrix as Taking the social relationship matrix as Where U={u1,u2,…,u |U|} represents a user set, I={i1,i2,…,i |I|} represents an item set, R u,i =1 represents that the user u has an interaction with the item i, and vice versa, R u,i =0, S j,k =1 represents that the user j has a social contact with the user k, and vice versa, S j,k =0.

[0014] S12, using a graph G=(V,E) to represent the network, where V and E represent the set of nodes and edges in the network, respectively, and taking the behavior interaction graph G R =(V R ,E R ) and the social relationship graph G S =(V S ,E S ) to be fused into a heterogeneous graph G F =(V F ,E F ), where the behavior interaction graph G R is a bipartite graph, the nodes V R contain both users and network video resources, and the edges V R represent that the user has watched or purchased the network video resource; the nodes V S in the social relationship graph G S only represent users, and the edges E S are directed edges and connect two users who have a directed social relationship on the network video and audio social platform; in the heterogeneous graph G F ,

[0015] Further, the step S20 specifically includes the following steps:

[0016] S21, the heterogeneous graph G obtained from the step S12 F contains rich local structures, which preserve specific social connections and corresponding interaction behaviors among users; identifying and exploiting the features of these local structures can mine high-level semantic information and better understand user preferences; for this purpose, three types of motifs are designed to characterize the local structures in the heterogeneous graph G F : (1) Social motifs describe explicit social relationships in a group of users, whose adjacency matrix can be represented as (2) Behavior motifs simulate interaction patterns that are not constrained by social relationships, in which users are related through common interaction items, whose adjacency matrix can be represented as (3) Joint motifs describe reinforced local connection patterns by integrating social information and interaction information, whose adjacency matrix can be represented as where m1

[0017]

[0018] where denotes the adjacency matrix of the motif M m , and m denotes the number of designed motifs;

[0019] S22, after the step S21, the heterogeneous graph G F is reorganized by different types of motifs from the perspectives of social relationships, interaction behaviors, and joint perspectives; in each perspective, the heterogeneous graph G F is actually transformed into a hypergraph conditioned on the corresponding type of motifs; in the hypergraph, nodes are connected through hyperedges, and each hyperedge can connect multiple nodes, using matrix A∈{0,1} |V|×|E| to record the adjacency relationship between nodes, where each element a(v,e) indicates whether the hyperedge e contains the node v, as shown in equation (2):

[0020]

[0021] Further, the step S30 includes the following specific steps:

[0022] S31, the hypergraph neural network is a specific application of the hypergraph, which is used for modeling and analyzing the hypergraph. Compared with the traditional graph neural network, the hypergraph neural network can process multi-element relationships and high-order relationships. At the same time, the hypergraph neural network can process hypergraphs of different sizes and shapes, and can perform embedding learning and representation learning on nodes and hyperedges in the hypergraph;

[0023] In the user vector representation module, the vector representation of the user in the low-dimensional space is calculated by formula (3):

[0024]

[0025] wherein H k ∈{H S ,H B ,H J} represents the user vector representation learned based on the corresponding motif typology, represents the degree matrix of A k , A k ∈{A S ,A B ,A J}, and l represents the number of layers of the neural network;

[0026] In order to overcome the problem of over-smoothing in the information aggregation process, the average value of the output of all layers in the hypergraph neural network is taken as the vector representation of each layer, as shown in formula (4):

[0027]

[0028] wherein L represents the number of layers of the hypergraph neural network, represents the output of the hypergraph neural network;

[0029] In order to obtain comprehensive user vector representation, a multi-channel attention mechanism is used to adaptively aggregate vector representations learned based on different motifs. For each user u, weight coefficients (α S ,α B ,α J ) are used to describe the weight of the user vector representation under different motifs, as shown in formula (5):

[0030]

[0031] wherein a k represents the attention weight coefficient, and are learning parameters in the attention function, represents the user vector representation learned based on a certain motif, and T represents transposition;

[0032] A new user vector representation is generated by means of the attention mechanism, as shown in formula (6):

[0033]

[0034] S32. In the project vector representation module, the project vector representation is defined using equation (7):

[0035] Q (l+1) =D -1 R Τ H *(l) (7)

[0036] Where R is the user-project interaction matrix, D is the degree matrix of R, and H... * This is the output of the user vector representation module;

[0037] S33. In the contrastive learning module, a self-gating activation function is introduced to replace the activation function with a single scalar as input. Its function expression is y = x * sigmoid(x). To make self-supervised learning more effective, the nonlinear approximation ability and stability of the model are first enhanced by the self-gating activation function designed in equation (8).

[0038]

[0039] in, This represents the output of the hypergraph neural network after passing through the self-gating activation function. g represents the output of the hypergraph neural network. k (·) represents the self-gating activation function. and These are learnable parameters, σ(·) represents the sigmoid activation function, and ⊙ represents the Hadamard product of vectors, which is the new vector obtained by multiplying corresponding elements of two vectors.

[0040] For each type of motif, a two-level comparison task was established. The first level is the comparison between the microscale and the mesoscale, and the second level is the comparison between the mesoscale and the macroscale. The self-supervised loss function is shown in Equation (9):

[0041]

[0042] Among them, f d (·) is a discriminant function that measures the consistency between two representations. This represents a vector representation of a single user at a microscale, to avoid the model becoming overly reliant on the order of data for z. k Shuffle and reorder the samples by disrupting their feature order to generate interfering negative samples. The READOUT function in formula (10) is used to obtain the mesoscale user vector representation z by aggregating node features k The AveragePooling function in formula (11) is used to obtain the macro-scale user vector representation y by retaining the average value on the mesoscale as the pooled value to preserve the features of the overall data k :

[0043]

[0044] In the formula, ak represents the corresponding row vector of the target user in matrix A k sum(a k ) represents the total number of connections of the network on the mesoscale;

[0045]

[0046] wherein, represents the output of the hypergraph neural network after the self-gating activation function;

[0047] In the preference prediction module, the loss function of the multi-channel hypergraph neural network model is shown in formula (12):

[0048] l=l P +βl A (12)

[0049] wherein, β represents the adjustment coefficient of the self-supervised loss, l A represents the self-supervised loss function constructed in formula (9), and l P represents the optimization of recommendation prediction using the BPR loss function, as shown in formula (13):

[0050]

[0051] wherein, I(u) represents the set of items interacting with user u, Θ represents the set of trainable parameters in the model, and λ is the regularization coefficient. In the training process, the optimization triple consists of user u, randomly sampled positive item i∈I(u), and negative item , represents the negative item recommendation result between u and i, represents the positive item recommendation result between u and i, generated by their respective representations, calculated by formula (14):

[0052]

[0053] wherein, and q irepresent the optimal user and item vector representations determined through iterative computation.

[0054] Further, the step S40 is specifically: given the number of iterations N, iteratively execute steps S31-S34 until the number of iterations is reached, complete the training of the multi-channel hypergraph neural network model, and obtain the optimal parameters of the model.

[0055] Further, the step S50 is specifically: importing the user and network video resource data to be predicted obtained from the network video social platform into the multi-channel hypergraph neural network model, and obtaining the Top-10 prediction result from formula (14).

[0056] A network video intelligent recommendation system based on social constraints is used to implement the network video intelligent recommendation method based on social constraints, and includes a computer processor and memory, a heterogeneous network fusion unit, a model information extraction unit, a model construction unit, a model training unit, and a preference prediction unit.

[0057] Further, the heterogeneous network fusion unit executes step S10 by using the collected user historical data through the computer processor and loading it into the computer memory; the model information extraction unit extracts social patterns, behavior patterns, and joint patterns from the heterogeneous network according to the heterogeneous network generated by the heterogeneous network fusion unit, and converts the input data into corresponding hypergraphs based on different patterns; the model construction unit executes step S30, models the hypergraph using a hypergraph neural network, and constructs a multi-channel hypergraph neural network model based on the hypergraph neural network, including a user vector representation module, an item vector representation module, a contrast learning module, and a preference prediction module; the model training unit executes step S40, determines the parameters through iterative computation, updates the vector representation of the user and the item using a multi-scale contrast learning mechanism, improves the effectiveness of representation learning, and obtains the optimal parameters of the model; and the preference prediction unit executes step S50, substitutes the user and item data to be predicted into the model to generate the final prediction result, and outputs the recommendation result to various users to provide intelligent recommendation for various network video platforms.

[0058] Compared with the prior art, the present application has the following advantages:

[0059] 1. A network video intelligent recommendation system based on social constraints extracts a series of local structures described by patterns, learns representations that preserve latent semantic patterns under the constraint of social relationships, effectively alleviates the cold start problem, and accurately understands the preferences of new users even if the data generated by the new users is sparse, and recommends network video resources that interest the new users to the new users.

[0060] 2. A socially constrained intelligent recommendation system for online video and audio content establishes multi-scale self-supervised learning tasks to enhance learning representations, further explore the intrinsic characteristics of the network, improve recommendation performance, thereby increasing user stickiness of the online video and audio platform and bringing more revenue to the platform. Attached Figure Description

[0061] Figure 1 A schematic diagram of the modal design of the network audio-visual intelligent recommendation method based on social constraints provided by the present invention;

[0062] Figure 2 This invention provides a system architecture diagram of a network audio-visual intelligent recommendation system based on social constraints.

[0063] Figure 3 A flowchart of the network audio-visual intelligent recommendation system based on social constraints provided by the present invention. Detailed Implementation

[0064] To further illustrate the technical solution of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0065] This invention discloses a socially constrained intelligent recommendation system for online video and audio content, used to implement a socially constrained intelligent recommendation method for online video and audio content. It is implemented through a computer program, and will be described below. Figure 3 The illustrated process details the specific implementation of the technical solution proposed in this invention. It demonstrates how this technical solution is used for intelligent recommendation on the Douban dataset, which originates from the Chinese online video and audio social platform Douban. This dataset contains 129,490 unique users and 58,541 unique movie entries. This embodiment selects 2,848 users and 39,586 items from this dataset to elaborate on the implementation process of the technical solution.

[0066] The implementation method mainly includes the following key aspects:

[0067] S10. Import historical data from the Douban dataset to generate an information database. Store and merge the behavioral interaction network and social relationship network as graph data into a heterogeneous network. The specific steps are as follows:

[0068] S11. Collect the Douban dataset from Douban.com to generate an information database, and denote the behavioral interaction matrix as follows: Let the social relationship matrix be denoted as Where U={u1,u2,…,u |U| Let} represent the user set, I = {i1, i2, ..., i |I|} represents a set of items, R u,i =1 indicates that user u interacts with item i, otherwise it is represented as R. u,i =0, Sj,k = 1 means that user j has social contact with user k, and vice versa, denoted as S j,k = 0, where |U| ∈ [1, 2848], |I| ∈ [1, 39586];

[0069] S12, represents the network with a graph G = (V, E), where V and E represent the set of nodes and edges in the network, respectively, and the behavior interaction graph G R = (V R , E R ) and the social relationship graph G S = (V S , E S ) are fused into a heterogeneous graph G F = (V F , E F ), where the behavior interaction graph G R is a bipartite graph, the nodes V R contain both users and network video resources, and the edges V R represent that the user has watched or purchased the network video resource; the nodes V S in the social relationship graph G S only represent users, and the edges E S are directed edges and connect two users who have a directed social relationship on the network video social platform; in the heterogeneous graph G F ,

[0070] S20, extracts social motifs, behavior motifs, and joint motifs from the heterogeneous network, and converts the input data into corresponding hypergraphs based on different motifs, as shown in Figure 1 , the specific steps are as follows:

[0071] S21, the heterogeneous graph G F obtained from step S12 contains rich local structures, which retain specific social relationships and corresponding interaction behaviors between users; identifying and utilizing the features of these local structures can mine high-level semantic information and better understand user preferences; for this purpose, three types of motifs are designed to describe the local structure in the heterogeneous graph G F : (1) The social motif describes a group of users with explicit social relationships, and its adjacency matrix can be represented as (2) The behavior motif simulates an interaction mode that is not constrained by social relationships, in which users are associated through common interaction items, and its adjacency matrix can be represented as (3) The joint motif describes a strengthened local connection mode by integrating social information and interaction information, and its adjacency matrix can be represented as For the adjacency matrix induced by the motif, the element in the matrix can be calculated by the co-occurrence of two points i and j in the motif, as shown in equation (1):

[0072]

[0073] wherein, denotes the co-occurrence of two points i and j in the motif M m induced by the motif, m denotes Figure 1 Eight designed motifs, wherein M1-M3 are social motifs, M4-M5 are behavior motifs, and M6-M8 are joint motifs;

[0074] S22, after step S21, the heterogeneous graph G F is reorganized by different types of motifs from the perspective of social relationship, the perspective of interactive behavior and the joint perspective; in each perspective, the heterogeneous graph G F is actually transformed into a hypergraph conditioned on the corresponding type of motif; in the hypergraph, nodes are connected by hyperedges, each hyperedge can connect multiple nodes, and the matrix A∈{0,1} |V|×|E| records the adjacency relationship between nodes, wherein each element a(v,e) represents whether the hyperedge e contains the node v, as shown in equation (2):

[0075]

[0076] S30, modeling the hypergraph obtained in step S20 by using a hypergraph neural network, constructing a multi-channel hypergraph neural network model based on the hypergraph neural network, including a user vector representation module, an item vector representation module, a contrast learning module and a preference prediction module, and the specific steps are as follows:

[0077] S31, the hypergraph neural network is a specific application of the hypergraph, which is used for modeling and analyzing the hypergraph; compared with the traditional graph neural network, the hypergraph neural network can handle multi-element relationships and high-order relationships; at the same time, the hypergraph neural network can handle hypergraphs of different sizes and shapes, and can perform embedding learning and representation learning on nodes and hyperedges in the hypergraph, so a multi-channel hypergraph neural network model is constructed based on the hypergraph neural network;

[0078] In the user vector representation module, the vector representation of the user in the low-dimensional space is calculated by equation (3):

[0079]

[0080] wherein, H k ∈{H S ,H B ,H J} represents the user vector representation learned based on the corresponding motif type, denotes Ak the degree matrix of A k ∈{A S ,A B ,A J}, and l represents the number of layers of the hypergraph neural network.

[0081] To overcome the problem of over-smoothing in the information aggregation process, the average of the output of all layers of the hypergraph neural network is taken as the vector representation of each layer, as shown in equation (4):

[0082]

[0083] where L represents the number of layers of the hypergraph neural network, represents the output of the hypergraph neural network.

[0084] To obtain a comprehensive user vector representation, a multi-channel attention mechanism is used to adaptively aggregate the vector representations learned based on different motifs. For each user u, the weight coefficients (α S ,α B ,α J ) are used to describe the weight of the user vector representation under different motifs, as shown in equation (5):

[0085]

[0086] where a k represents the attention weight coefficient, and are learning parameters in the attention function, represents the user vector representation learned based on a certain motif, and T represents the transpose.

[0087] A new user vector representation is generated with the help of the attention mechanism, as shown in equation (6):

[0088]

[0089] S32, in the item vector representation module, the item vector representation is defined by equation (7):

[0090] Q (l+1) =D -1 R Τ H *(l) (7)

[0091] where R is the interaction matrix of users and items, D is the degree matrix of R, H * is the output of the user vector representation module.

[0092] S33, in the contrast learning module, in order to replace the activation function with a single scalar input self-gating activation function, the function expression is y=x*sigmoid(x); in order to make self-supervised learning more effective, first of all, through the self-gating activation function designed in formula (8), enhance the nonlinear approximation ability and stability of the model:

[0093]

[0094] wherein, represents the output of the hypergraph neural network after self-gating activation function, represents the output of the hypergraph neural network, g k (·) represents self-gating activation function, and is a learnable parameter, sigma (·) represents sigmoid activation function, ⊙ represents vector Hadamard product, and refers to the new vector obtained by multiplying the elements at corresponding positions in two vectors;

[0095] For each type of motif, a two-level contrast task is established, the first level is the contrast between micro scale and meso scale, and the second level is the contrast between meso scale and macro scale, and the self-supervised loss function is shown in formula (9):

[0096]

[0097] wherein, f d (·) is a discriminant function for measuring the consistency between two representations, represents the vector representation of a single user at the micro scale, in order to avoid the model from over-relying on the arrangement order of data, shuffle is used to reorder z k , and interference negative samples are generated by disturbing the sample feature order The READOUT function in formula (10) is used to obtain the user vector representation z k at the meso scale by aggregating node features k The AveragePooling function in formula (11) is used to reserve the features of the whole data by taking the average value at the meso scale as the pooled value, and the user vector representation y k at the macro scale is obtained:

[0098]

[0099] In the formula, a k represents the corresponding row vector of the target user in matrix A k , and sum (a k ) represents the total number of connections of the network at the meso scale;

[0100]

[0101] wherein, represents the output of the hypergraph neural network after the self-gating activation function;

[0102] S34, in the preference prediction module, the loss function of the multi-channel hypergraph neural network model is as shown in formula (12), when formula (12) takes the minimum value, the parameters in the model reach the optimum at this time:

[0103] l=l P +βl A (12)

[0104] wherein, β represents the adjustment coefficient of the self-supervised loss, l A represents the self-supervised loss function constructed by formula (9), l P represents the optimization of recommendation prediction by using BPR loss function, as shown in formula (13):

[0105]

[0106] wherein, I(u) represents the item set interacting with user u, Θ represents the set of trainable parameters in the model, λ is the regularization coefficient, in the training process, the optimization triple consists of user u, randomly sampled positive item i∈I(u) and negative item , represents the negative item recommendation result between u and i, represents the positive item recommendation result between u and i, generated by their respective representations, calculated by formula (14):

[0107]

[0108] wherein, and q i represent the optimal user and item vector representations determined by iterative calculation.

[0109] S40, for the multi-channel hypergraph neural network model constructed in step S30, given the number of iterations N, steps S31-S34 are iteratively executed until the number of iterations is reached, the training of the model is completed, and the optimal parameters of the model are obtained;

[0110] S50, for the multi-channel hypergraph neural network model constructed in step S30, and the optimal parameters of the model determined in step S40, the user and network video resource data to be predicted obtained from the network video social platform are imported into the model, and the top-10 prediction result is obtained by formula (14).

[0111] Technical effect evaluation:

[0112] To verify the effectiveness and advancement of the technical solutions proposed in the present application, the present application is compared with the recommendation methods of DiffNet, LightGCN, S 2 The recommendation results of the above methods are evaluated by the Douban data set in the embodiment, and the results are shown in Table 1, taking the average recommendation precision and recall rate of 10 experiments as evaluation indexes:

[0113] Table 1: Comparison and analysis of results

[0114]

[0115]

[0116] The cold start recommendation results of the above methods are evaluated by the Douban data set in the embodiment, and the results are shown in Table 2:

[0117] Table 2: Comparison and analysis of results

[0118]

[0119] As can be seen from the results in the table, the technical solutions of the present application can obtain high-precision recommendation results when making recommendations, and effectively alleviate the cold start problem.

[0120] As Figure 2 shown, a network video intelligent recommendation system based on social constraints includes a computer processor and memory, a heterogeneous network fusion unit, a model information extraction unit, a model construction unit, a model training unit and a preference prediction unit. The heterogeneous network fusion unit performs step S10 by using the collected user historical data through the computer processor and loading it into the computer memory. The model information extraction unit performs step S20 to extract social patterns, behavior patterns and joint patterns from the heterogeneous network generated by the heterogeneous network fusion unit, and converts the input data into corresponding hypergraphs based on different patterns. The model construction unit performs step S30 to model the hypergraph using a hypergraph neural network, and constructs a multi-channel hypergraph neural network model based on the hypergraph neural network, including a user vector representation module, an item vector representation module, a contrast learning module and a preference prediction module. The model training unit performs step S40 to determine the parameters by iterative calculation, updates the vector representation of users and items using a multi-scale contrast learning mechanism, improves the effectiveness of representation learning, and obtains the optimal parameters of the model. The preference prediction unit performs step S50 to input the user and item data to be predicted into the model to generate the final prediction result, and outputs the recommendation result to various users to provide intelligent recommendation for various network video platforms.

[0121] The details of the application not described herein are considered known to those skilled in the art. Although the foregoing description of the application has been described in some detail for the purposes of clarity and the understanding of the application, it should be appreciated that the application is not limited to the particular embodiments or examples described. It should be readily understood that various changes can be made therein without departing from the spirit and scope of the application defined by the appended claims and that equivalent materials can be substituted for those described.

Claims

1. A method for intelligent recommendation of online audio and video content based on social constraints, characterized in that, Includes the following steps: S10. Collect historical data to generate an information database, and store and merge the behavioral interaction network and social relationship network into a heterogeneous network in the form of graph data. S20. Extract social motifs, behavioral motifs, and joint motifs from the heterogeneous network obtained in step S10, and transform the input data into corresponding hypergraphs based on different motifs. S30. Model the hypergraph obtained in step S20 using a hypergraph neural network. Construct a multi-channel hypergraph neural network model based on the hypergraph neural network, which includes a user vector representation module, an item vector representation module, a contrastive learning module, and a preference prediction module. S40. For the multi-channel hypergraph neural network model constructed in step S30, the model is solved by iterative calculation to determine the optimal parameters of the model. S50. For the multi-channel hypergraph neural network model constructed in step S30 and the optimal parameters of the model determined in step S40, generate the final recommendation prediction. Step S20 specifically includes the following steps: S21, Heterogeneous graph obtained in step S12 It contains rich local structures; therefore, three types of motif pairs were designed for heterogeneous graphs. The local structure is characterized as follows: (1) The social module describes a set of explicit social relationships among users, and its adjacency matrix is ​​represented as follows: (2) The behavioral motif simulates an interaction pattern unconstrained by social relationships, in which users are interconnected through shared interactive items, and its adjacency matrix is ​​represented as follows: (3) The joint motif describes the enhanced local connectivity patterns by integrating social and interactive information, and its adjacency matrix is ​​represented as follows: ,in For the adjacency matrix generated by the phantom, the elements in the matrix are calculated based on the co-occurrence of two points i and j in the phantom, as shown in equation (1): （1） in, motif The adjacency matrix, Indicates the number of designed phantoms; S22. After step S21, the heterogeneous diagram... The graphs are recombined using different types of motifs from the perspectives of social relationships, interactive behavior, and collaboration; within each perspective, the heterogeneous graph... In practice, they are all transformed into a hypergraph conditioned on the corresponding type of motif; in the hypergraph, nodes are connected by hyperedges, and each hyperedge connects multiple nodes, using matrices. Record the adjacency relationships between nodes, where each element... Indicate whether the hyperedge e contains node v, as shown in equation (2): （2）。 2. The method for intelligent recommendation of online audio and video based on social constraints according to claim 1, characterized in that, Step S10 specifically includes the following steps: S11. Collect historical data from online video and social media platforms to generate an information database, and denote the behavioral interaction matrix as follows: The social relationship matrix is ​​denoted as ,in Represents a user set, Indicates a set of items. This indicates that user u interacts with item i, and vice versa. , This indicates that user j has a social connection with user k, and vice versa. ; S12, Using diagrams Let V and E represent the set of nodes and edges in the network, respectively, and let the behavioral interaction graph be used. and social relationship graph Merging into a heterogeneous graph Among them, the behavior interaction graph For a bipartite graph, nodes It includes both users and online video resources, on the side This indicates that the user has viewed or purchased the online video / audio resource; social relationship graph. Nodes in Only representing users, side A directed edge connects two users with a directed social relationship on an online video-sharing social platform; in a heterogeneous graph middle, , .

3. The method for intelligent recommendation of online audio and video based on social constraints according to claim 2, characterized in that, Step S30 specifically includes the following steps: S31. Hypergraph neural networks are a specific application of hypergraphs, used for modeling and analyzing hypergraphs. Hypergraph neural networks can handle multivariate and higher-order relations, as well as hypergraphs of different sizes and shapes, and perform embedding and representation learning on nodes and hyperedges in hypergraphs. In the user vector representation module, the user's vector representation in the low-dimensional space is calculated by equation (3): （3） in, This represents the user vector representation learned based on the corresponding modality type. express The degree matrix, , l represents the number of layers in the hypergraph neural network; To overcome the oversmoothing problem that occurs during information aggregation, the average value of the output of all layers in the hypergraph neural network is taken as the vector representation of each layer, as shown in equation (4): （4） Where L represents the number of layers in the hypergraph neural network, This represents the output of the hypergraph neural network; To obtain comprehensive user vector representations, a multi-channel attention mechanism is used to adaptively aggregate vector representations learned from different motifs. For each user u, attention weight coefficients are used... The weights used to describe the user vector representation under different modalities are shown in Equation (5): （5） in, This represents the attention weight coefficient. and These are the learning parameters in the attention function. This represents the user vector representation learned based on a certain motif. Indicates transpose; A new user vector representation is generated using an attention mechanism, as shown in equation (6): （6） S32. In the project vector representation module, the project vector representation is defined using equation (7): （7） in, A matrix of user-project interactions. for The degree matrix, This is the output of the user vector representation module; S33. In the contrastive learning module, a self-gating activation function is introduced to replace the activation function that takes a single scalar as input. Its function expression is: To make self-supervised learning more effective, the nonlinear approximation capability and stability of the model are first enhanced by the self-gating activation function designed in equation (8): （8） in, This represents the output of the hypergraph neural network after passing through the self-gating activation function. This represents the output of the hypergraph neural network. This represents the self-gating activation function. and These are learning parameters. This represents the sigmoid activation function. The Hadamard product of vectors refers to the new vector obtained by multiplying corresponding elements of two vectors. For each type of motif, a two-level comparison task was established. The first level is the comparison between the microscale and the mesoscale, and the second level is the comparison between the mesoscale and the macroscale. The self-supervised loss function is shown in Equation (9): （9） in, It is a discriminant function that measures the consistency between two representations. This represents a vector representation of a single user at a microscale, to avoid the model becoming overly reliant on the order of the data. Shuffle and reorder the samples by disrupting their feature order to generate interfering negative samples. The READOUT function in equation (10) is used to obtain the user vector representation at the mesoscale by aggregating node features. Using the AveragePooling function in equation (11), the overall data features are preserved by using the average value at the mesoscale as the pooled value, thus obtaining the user vector representation at the macroscale. : （10） in, Indicates the target user in the matrix The corresponding row vector in the middle, This represents the total number of connections in a network at a mesoscale. （11） in, This represents the output of the hypergraph neural network after passing through the self-gating activation function; S34. In the preference prediction module, the loss function of the multi-channel hypergraph neural network model is shown in equation (12): （12） in, This represents the adjustment factor for the self-monitoring loss. This represents the self-supervised loss function constructed by equation (9). The BPR loss function is used to optimize the recommendation prediction, as shown in equation (13): （13） in, This represents the set of items that interact with user u. This represents the set of training parameters in the model. The regularization coefficient is used during training to optimize the triplet, which consists of user u and randomly sampled positive terms. and negative terms composition, This represents the negative recommendation results between u and i. This represents the positive recommendation results between u and i. Generated from their respective representations, and calculated by equation (14): （14） in, and This represents the optimal user and item vector representation determined through iterative calculation.

4. The method for intelligent recommendation of online audio and video based on social constraints according to claim 3, characterized in that, Specifically, step S40 involves: given the number of iterations N, iteratively executing steps S31 to S34 until the number of iterations is reached, completing the training of the multi-channel hypergraph neural network model, and obtaining the optimal parameters of the model.

5. The method for intelligent recommendation of online audio and video based on social constraints according to claim 3, characterized in that, The specific steps of step S50 are as follows: the user and online video resource data to be predicted obtained from the online video social platform are imported into the multi-channel hypergraph neural network model, and the Top-10 prediction results are obtained by equation (14).

6. A network audio-visual intelligent recommendation system based on social constraints, characterized in that, for implementing the network audio-visual intelligent recommendation method based on social constraints according to any one of claims 1 to 5, it includes a computer processor and memory, a heterogeneous network fusion unit, a modality information extraction unit, a model building unit, a model training unit, and a preference prediction unit.

7. A network audio-visual intelligent recommendation system based on social constraints according to claim 6, characterized in that: the heterogeneous network fusion unit executes step S10, using a computer processor to utilize collected user historical data and load it into computer memory; the motif information extraction unit executes step S20 based on the heterogeneous network generated by the heterogeneous network fusion unit, extracting social motifs, behavioral motifs, and joint motifs from the heterogeneous network, and transforming the input data into corresponding hypergraphs based on different motifs; the model construction unit executes step S30, using a hypergraph neural network to model the hypergraph, constructing a multi-channel hypergraph neural network model based on the hypergraph neural network, including a user vector representation module, an item vector representation module, a contrastive learning module, and a preference prediction module; the model training unit executes step S40, determining parameters through iterative calculation, updating the vector representations of users and items using a multi-scale contrastive learning mechanism, improving the effectiveness of representation learning, and obtaining the optimal parameters of the model; the preference prediction unit executes step S50, substituting the user and item data to be predicted into the model, generating the final prediction result, and outputting the recommendation result to various users, providing intelligent recommendations for various network audio-visual platforms.

Images