Binary social contact mode based on machine learning and implementation method

Abstract

The invention discloses a machine learning-based binary social contact mode and an implementation method, and relates to the technical field of social networks, the mode comprises an on-campus social contact domain module, an off-campus self-media domain module, a hybrid processing architecture and cross-domain information fusion module, and a social contact enhancement recommendation engine. Through the on-campus social domain module and the off-campus self-media domain module, on-campus and off-campus information can be comprehensively utilized, more comprehensive social portraits are further provided, and the fusion of the double-domain information is beneficial to more accurate understanding of social behaviors and interest preferences of users. Differentiated modeling of an intra-school strong relationship and an extra-school weak relationship is realized through a domain separation architecture, a cross-domain attention layer effectively captures feature interaction of different social scenes by adopting a domain specific parameter matrix, and semantic-level feature fusion is realized while the independence of a dual-domain topology is kept in cooperation with a feature migration mechanism of a bridging matrix. And the node characterization simultaneously comprises stable campus social features and dynamic cross-school interest features.

Description

Technical Field

[0001] The present invention relates to the technical field of social network technology, and in particular to a dual social model based on machine learning and an implementation method thereof. Background Art

[0002] In the field of social networking technology, as user social activities become increasingly diverse and complex, traditional single-source social models are no longer able to meet users' demands for precise social recommendations and personalized services. Currently, most existing social models focus on capturing and processing information in a single area, such as focusing solely on on-campus social interactions or external social media content. This single-dimensional approach leads to the formation of information silos, limiting the accuracy and comprehensiveness of social recommendation systems.

[0003] Specifically, some existing social models may analyze user behavior solely based on strong on-campus ties. While these models can capture users' social activities and interests within the school, they overlook the rich off-campus weak-ties network resources, such as cross-campus content collaboration and geographic co-occurrence, which also contain users' potential interests and social needs. Conversely, some social models that focus solely on off-campus self-media content may fail to fully understand users' social behavior and true interests due to a lack of in-depth exploration of strong on-campus ties.

[0004] Based on this, a dual social model and implementation method based on machine learning is now provided to eliminate the drawbacks of the existing model. Summary of the Invention

[0005] The purpose of the present invention is to provide a dual social model based on machine learning and an implementation method to solve the problems in the background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A dual social model based on machine learning, including an on-campus social domain module, an off-campus self-media domain module, a hybrid processing architecture, a cross-domain information fusion module, and a social enhancement recommendation engine;

[0008] The campus social domain module is responsible for building a strong relationship social network including student entities, community entities and course entities, and establishing a connection with the hybrid processing architecture through the strong relationship network;

[0009] The off-campus self-media domain module is used to build a weak relationship network based on cross-school content collaboration and geographical location co-occurrence, and is connected to the hybrid processing architecture through the weak relationship network;

[0010] The hybrid processing architecture adopts a dual-domain graph convolution kernel architecture, including a parallel graph neural network branch and a dynamic interest evolution network branch. The graph neural network branch generates a graph neural network contrast loss term, and the dynamic interest evolution network branch also has a dual loss.

[0011] The hybrid processing architecture performs feature space alignment processing on the information and transmits the processed information to the cross-domain information fusion module;

[0012] The cross-domain information fusion module fuses the information to generate fusion features, and then inputs the fusion features into the social enhancement recommendation engine;

[0013] The recommendation engine is used to dynamically generate recommendation results based on user characteristics and user interests, and establish optimization objectives for the dual loss terms of the graph neural network branch comparison loss term and the dynamic interest evolution network branch.

[0014] Preferably, the graph neural network branch is constructed using a hypergraph structure, and its hypergraph order k=3;

[0015] The graph neural network branch is configured with 8 attention heads (h g =8) nodes, whose representation dimension is d e =64;

[0016] The graph neural network branch is connected through a cross-domain bridge matrix Realize dual-domain feature transfer.

[0017] Preferably, the dual-domain graph convolution kernel includes a domain separation architecture, and the domain separation architecture is the campus social graph G intra =(V,E strong ) and off-campus relationship map G inter =(V,E weak ) constituted by a parallel input structure;

[0018] Among them: the campus map node V contains students, clubs, and course entities, E strong Represents strong relationships between class affiliation and activity participation; weak Based on cross-school collaborative creation and weak connection construction of geographical location co-occurrence;

[0019] The dual-domain graph convolution kernel also includes a cross-domain attention layer, which is obtained by the domain-specific parameter matrix W d And the LeakyReLU activation function σ, the attention coefficient of node i in domain d is calculated as follows:

[0020]

[0021] Among them: a i,dRefers to the attention coefficient of node i in domain d; exp exponential function, used to calculate attention weight; σ refers to the LeakyReLU activation function; refers to the transpose of the attention vector of domain d; W d Refers to the weight matrix of domain d, which is used to perform linear transformation on node features; h i The feature vector of node i; || represents the feature concatenation operation, which concatenates two feature vectors together; N d (i) is the neighbor set of node i in domain d; h j and h k represents the feature vector of nodes j and k, h j and h k They represent the neighbor nodes of node i respectively;

[0022] The dual-domain graph convolution kernel also includes a feature migration component, which is connected to the bridging matrix Linearly combine the cross-domain attention output with the intra-school features to generate the final node representation:

[0023]

[0024] Where: h′ represents the processed node feature; σ refers to the activation function; α i,d Refers to the attention coefficient; W d Refers to the weight matrix; h i refers to the feature vector of node i; B refers to the bridging matrix; Φ refers to the cross-domain attention output mapping; Refers to the feature vector of node j after itr iteration.

[0025] Preferably, the dynamic interest evolution network branch includes a time window T w = 15-day temporal feature extraction layer;

[0026] The dynamic interest evolution network branch also includes an interest evolution network composed of 3 layers of GRU, and the hidden layer dimension of the interest evolution network is d h =128;

[0027] The dynamic interest evolution network branch includes a social influence injection unit, which integrates the group behavior trends output by the graph neural network branch through multi-head attention pooling.

[0028] Preferably, the dynamic interest evolution network branch configures a time-aware social influence attenuation factor

[0029] Where: λ(t) is the time-dependent decay factor; e is the base of the natural logarithm; B is a learnable parameter; t is the current time node; and t0 is the time node of special campus events (such as exam week and club recruitment).

[0030] Preferably, the final node representation generation process of the cross-domain information fusion module is

[0031] Where: α is the dynamically adjusted domain weight coefficient, which is used to balance the contribution of intra and inter features; B is the cross-domain bridging matrix; The final feature representation of node i is the result of cross-domain information fusion; Feature representation of node i in the campus domain; The feature representation of node i in the off-campus domain; B is the cross-domain bridging matrix, which is used to realize feature migration and cross-domain feature fusion; (1-α) is the weight coefficient of the off-campus domain feature, which together with α determines the ratio of on-campus and off-campus features in the final feature representation.

[0032] Preferably, the optimization objective of the recommendation engine includes a graph neural network branch contrast loss term:

[0033]

[0034] Set the number of negative sampling nodes K = 512 and the regularization coefficient α = 0.8;

[0035] Where: L GNN Refers to the contrast loss of graph neural network; V refers to the set of all nodes; y v The true label of node v; The predicted label of node v; α is the regularization coefficient used to balance the weight of the loss term Refers to the distribution of positive samples p pos The node pairs sampled in (v i ,v j ) is the expected value of ; f(v) is the feature representation function of node v; τ is the temperature parameter used to control the sharpness of the contrast loss; k is the number of negative sampling nodes, that is, for each positive sample, k negative samples are sampled.

[0036] Preferably, the optimization objective of the recommendation engine further includes a dual loss term of the dynamic interest evolution network branch;

[0037]

[0038] Configure cross entropy loss weight β = 1.2;

[0039] Where: L DIEN Refers to the dual loss of interest evolution network; T wRefers to the size of the time window; l ce Refers to the cross entropy loss function; refers to the target feature at time t; Refers to the predicted features at time t; β refers to the cross entropy loss weight, which is used to balance the weights of the two loss terms; h t refers to the hidden state at time t; h t-1 refers to the hidden state at time t-1; |·|2 refers to the L2 norm, which is used to calculate the difference between hidden states.

[0040] A method for implementing the dual social model based on the above machine learning includes the following steps:

[0041] Step 1: The on-campus social domain module transfers on-campus social related data to the hybrid processing architecture through the strong relationship network; the off-campus self-media domain module transfers the information in the off-campus self-media field to the hybrid processing architecture through the weak relationship network;

[0042] Step 2: The hybrid processing architecture receives information from the on-campus social domain module and the off-campus self-media domain module. After aligning the feature space of this information using the parallel graph neural network branch and the dynamic interest evolution network branch, the processing results are passed to the cross-domain information fusion module.

[0043] Step 3: The cross-domain information fusion module fuses the information transmitted by the hybrid processing architecture to generate fusion features, which are then input into the social enhanced recommendation engine.

[0044] Step 4: The social enhanced recommendation engine generates recommendation results based on the received fusion features.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention uses the on-campus social domain module and the off-campus self-media domain module to comprehensively utilize on-campus and off-campus information to further provide a more comprehensive social portrait. The fusion of this dual-domain information helps to more accurately understand the user's social behavior and interest preferences.

[0047] 2. The present invention realizes differentiated modeling of strong relationships within the school and weak relationships outside the school through a domain separation architecture. The cross-domain attention layer adopts a domain-specific parameter matrix to effectively capture the feature interactions of different social scenarios. Combined with the feature migration mechanism of the bridging matrix, semantic-level feature fusion is achieved while maintaining the topological independence of the two domains, so that the node representation contains both stable campus social characteristics and dynamic cross-school interest characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flowchart of the present invention.

[0049] Figure 2 Schematic diagram of the hybrid processing architecture of the present invention.

[0050] Figure 3 This is the optimization processing diagram of the present invention.

[0051] Figure 4 It is an operation flow chart of the present invention.

[0052] Figure 5 Schematic diagram of the dual-domain graph convolution kernel of the present invention.

[0053] Figure 6 Schematic diagram of the dynamic interest evolution network branches of the present invention. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0055] Example 1

[0056] In one embodiment, Figures 1 to 6 As shown, a dual social model based on machine learning includes an on-campus social domain module, an off-campus self-media domain module, a hybrid processing architecture, a cross-domain information fusion module, and a social enhancement recommendation engine;

[0057] The campus social domain module is responsible for building a strong relationship social network including student entities, community entities and course entities, and establishing a connection with the hybrid processing architecture through the strong relationship network;

[0058] The off-campus self-media domain module is used to build a weak relationship network based on cross-school content collaboration and geographical location co-occurrence, and is connected to the hybrid processing architecture through the weak relationship network;

[0059] The hybrid processing architecture adopts a dual-domain graph convolution kernel architecture, including a parallel graph neural network branch and a dynamic interest evolution network branch. The graph neural network branch generates a graph neural network contrast loss term, and the dynamic interest evolution network branch also has a dual loss.

[0060] The hybrid processing architecture performs feature space alignment processing on the information and transmits the processed information to the cross-domain information fusion module;

[0061] The cross-domain information fusion module fuses the information to generate fusion features, and then inputs the fusion features into the social enhancement recommendation engine;

[0062] The recommendation engine is used to dynamically generate recommendation results based on user characteristics and user interests, and establish optimization objectives for the dual loss terms of the graph neural network branch comparison loss term and the dynamic interest evolution network branch.

[0063] In this embodiment, the hybrid processing architecture performs feature space alignment on the information transmitted from the on-campus social domain module and the off-campus self-media domain module. The purpose of feature space alignment is to unify information from different sources and different feature dimensions into the same feature space to facilitate subsequent fusion and analysis. The processed information is then passed to the cross-domain information fusion module.

[0064] The cross-domain information fusion module fuses the information processed by the hybrid processing architecture to generate fusion features. These fusion features integrate multiple information from strong-relationship social networks within the school and weak-relationship networks outside the school, and can more comprehensively reflect users' social behaviors and interest preferences.

[0065] The social enhancement recommendation engine provides users with personalized social recommendation services based on the fusion features input by the cross-domain information fusion module. For example, it can recommend classmates, clubs, activities, or related self-media content that may be of interest to enhance the user's social experience and participation.

[0066] In one embodiment, the graph neural network branch is constructed using a hypergraph structure, and the hypergraph order k=3;

[0067] The graph neural network branch is configured with 8 attention heads (h g =8) nodes, whose representation dimension is d e =64;

[0068] The graph neural network branch is connected through a cross-domain bridge matrix Realize dual-domain feature transfer.

[0069] It should be noted that the graph neural network branch is constructed using a hypergraph structure, configured with multiple attention heads and node representation dimensions, and realizes dual-domain feature migration through a cross-domain bridging matrix. This branch can capture complex relationships in social networks.

[0070] In one embodiment, the dual-domain graph convolution kernel includes a domain separation architecture, which is a campus social graph G intra =(V,E strong ) and off-campus relationship map G inter =(V,E weak ) constituted by a parallel input structure;

[0071] Among them: the campus map node V contains students, clubs, and course entities, E strong Represents strong relationships between class affiliation and activity participation; weak Based on cross-school collaborative creation and weak connection construction of geographical location co-occurrence;

[0072] The dual-domain graph convolution kernel also includes a cross-domain attention layer, which is obtained by the domain-specific parameter matrix Wd And the LeakyReLU activation function σ, the attention coefficient of node i in domain d is calculated as follows:

[0073]

[0074] Among them: a i,d Refers to the attention coefficient of node i in domain d; exp exponential function, used to calculate attention weight; σ refers to the LeakyReLU activation function; refers to the transpose of the attention vector of domain d; W d Refers to the weight matrix of domain d, which is used to perform linear transformation on node features; h is the feature vector of node i; || represents the feature concatenation operation, which concatenates two feature vectors together; N d (i) is the neighbor set of node i in domain d; h j and h k represents the feature vector of nodes j and k, h j and h k They represent the neighbor nodes of node i respectively;

[0075] The dual-domain graph convolution kernel also includes a feature migration component, which is connected to the bridging matrix Linearly combine the cross-domain attention output with the intra-school features to generate the final node representation:

[0076]

[0077] Among them: represents the processed node features; refers to the activation function; refers to the attention coefficient; refers to the weight matrix; refers to the eigenvector of the node; refers to the bridging matrix; refers to the cross-domain attention output mapping; refers to the iterated eigenvector of node j.

[0078] It should be noted that the dual-domain graph convolution kernel uses a domain separation architecture, with a parallel input structure consisting of the on-campus social graph Gintra = (V, Estrong) and the off-campus relationship graph Ginter = (V, Eweak). This architecture allows the model to process strong on-campus relationships and weak off-campus relationships separately, thereby achieving more refined modeling.

[0079] The nodes of the campus graph include student, club, and course entities, and the edge Estrong represents strong relationships such as class affiliation and activity participation. These relationships reflect the close social connections of students on campus.

[0080] The off-campus graph edge Eweak is constructed based on weak connections such as cross-school collaborative creation and geographic location co-occurrence. These connections reflect students' extensive social activities outside the campus.

[0081] The cross-domain attention layer is a key component of the dual-domain graph convolution kernel. It calculates the attention coefficient of the node within the domain through the domain-specific parameter matrix and activation function. By calculating the attention coefficient, the model can focus on the neighboring nodes that are most important to the current node, thereby more effectively capturing the complex relationships in the social network.

[0082] Feature migration components through the bridge matrix The cross-domain attention output is linearly combined with the on-campus features to generate the final node representation. The migration and fusion of cross-domain features are achieved through the bridging matrix, so that the final node representation contains both on-campus and off-campus information, thereby more comprehensively reflecting the user's social behavior and interest preferences.

[0083] In one embodiment, the dynamic interest evolution network branch includes a time window T w = 15-day temporal feature extraction layer;

[0084] The dynamic interest evolution network branch also includes an interest evolution network composed of 3 layers of GRU, and the hidden layer dimension of the interest evolution network is d h =128;

[0085] The dynamic interest evolution network branch includes a social influence injection unit, which integrates the group behavior trends output by the graph neural network branch through multi-head attention pooling.

[0086] It should be noted that the dynamic interest evolution network branch: including the time window, GRU layer and social influence injection unit, integrates group behavior trends through multi-head attention pooling. This branch can capture the dynamic changes of user interests.

[0087] In one embodiment, the dynamic interest evolution network branch configures a time-aware social influence attenuation factor

[0088] Where: λ(t) is the time-dependent decay factor; e is the base of the natural logarithm; B is a learnable parameter; t is the current time node; and t0 is the time node of special campus events (such as exam week and club recruitment).

[0089] It should be noted that parameter B is learnable, which means that during the model training process, the system can automatically adjust the value of B based on the data to better fit the actual situation of user interests changing over time. This adaptive capability enables the model to more flexibly respond to the interest change patterns of different user groups.

[0090] t0 represents the time of special campus events, such as exam weeks, holidays, and campus events, which often have a significant impact on students' interests and behaviors. By incorporating t0 into the calculation of the decay factor, the model can more accurately capture the impact of these events on user interests, thereby making more reasonable adjustments in the recommendation process.

[0091] The time-aware social influence attenuation factor plays a key role in the dynamic interest evolution network branch. It helps the model understand that user interests are not static, but will change over time and under the influence of external events. By considering this dynamic nature, the model can generate recommendation results that are more in line with the user's current interest status, thereby improving the accuracy and personalization of recommendations.

[0092] In one embodiment, the final node representation generation process of the cross-domain information fusion module is as follows:

[0093] Where: α is the dynamically adjusted domain weight coefficient, which is used to balance the contribution of intra and inter features; B is the cross-domain bridging matrix; The final feature representation of node i is the result of cross-domain information fusion; Feature representation of node i in the campus domain; The feature representation of node i in the off-campus domain; B is the cross-domain bridging matrix, which is used to realize feature migration and cross-domain feature fusion; (1-α) is the weight coefficient of the off-campus domain feature, which together with α determines the ratio of on-campus and off-campus features in the final feature representation.

[0094] It should be noted that the cross-domain information fusion module generates the final node representation through dynamically adjusted domain weight coefficients and cross-domain bridging matrix to achieve effective fusion of dual-domain information.

[0095] In one embodiment, the optimization objective of the recommendation engine includes a graph neural network branch contrast loss term:

[0096]

[0097] Set the number of negative sampling nodes K = 512 and the regularization coefficient α = 0.8;

[0098] Where: L GNN Refers to the contrast loss of graph neural network; V refers to the set of all nodes; y v The true label of node v; The predicted label of node v; α is the regularization coefficient used to balance the weight of the loss term Refers to the distribution of positive samples p pos The node pairs sampled in (v i ,vj ) is the expected value of ; f(v) is the feature representation function of node v; τ is the temperature parameter used to control the sharpness of the contrast loss; k is the number of negative sampling nodes, that is, for each positive sample, k negative samples are sampled.

[0099] It’s important to note that the contrastive loss term in graph neural networks is designed to enhance the model’s understanding of user-item interactions through contrastive learning. Contrastive learning is an unsupervised learning method that learns effective feature representations by comparing similar pairs of samples (positive samples) with dissimilar pairs of samples (negative samples).

[0100] In a specific implementation, the graph neural network contrast loss term may optimize the model by maximizing the similarity between positive samples while minimizing the similarity between negative samples. This loss function helps the model capture the potential associations between users and items, improving the accuracy of recommendations.

[0101] K is set to 512, which means that during the training process, for each positive sample, 512 negative samples are randomly sampled for comparative learning to improve the model's discrimination ability.

[0102] The regularization coefficient α is set to 0.8 to control the model complexity and ensure that the performance of the model on the training data can be generalized to unseen data.

[0103] In one embodiment, the optimization objective of the recommendation engine further includes a dual loss term of the dynamic interest evolution network branch;

[0104]

[0105] Configure cross entropy loss weight β = 1.2;

[0106] Where: L DIEN Refers to the dual loss of interest evolution network; T w Refers to the size of the time window; l ce Refers to the cross entropy loss function; refers to the target feature at time t; Refers to the predicted features at time t; β refers to the cross entropy loss weight, which is used to balance the weights of the two loss terms; h t refers to the hidden state at time t; h t-1 refers to the hidden state at time t-1; |·|2 refers to the L2 norm, which is used to calculate the difference between hidden states.

[0107] It should be noted that the recommendation engine also considers the dual loss of the dynamic interest evolution network branch. This is because the DIEN branch is designed to capture the dynamic changes of user interests, and considering only the basic optimization objective may not be enough to fully train this branch.

[0108] Dual loss typically consists of a main loss (such as cross-entropy loss) and an auxiliary loss (such as a smoothness constraint); by introducing dual loss, the recommendation engine can more comprehensively consider the dynamic changes of user interests and ensure that the DIEN branch is fully optimized during training. This helps improve the accuracy and personalization of recommendation results, making it easier for users to discover content they are interested in.

[0109] Example 2

[0110] like Figure 1 As shown, a method for implementing the dual social mode based on the above machine learning includes the following steps:

[0111] Step 1: The on-campus social domain module transfers on-campus social related data to the hybrid processing architecture through the strong relationship network; the off-campus self-media domain module transfers the information in the off-campus self-media field to the hybrid processing architecture through the weak relationship network;

[0112] Step 2: The hybrid processing architecture receives information from the on-campus social domain module and the off-campus self-media domain module. After aligning the feature space of this information using the parallel graph neural network branch and the dynamic interest evolution network branch, the processing results are passed to the cross-domain information fusion module.

[0113] Step 3: The cross-domain information fusion module fuses the information transmitted by the hybrid processing architecture to generate fusion features, which are then input into the social enhanced recommendation engine.

[0114] Step 4: The social enhanced recommendation engine generates recommendation results based on the received fusion features.

[0115] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A dual social model based on machine learning, characterized by: It includes an on-campus social domain module, an off-campus self-media domain module, a hybrid processing architecture, a cross-domain information fusion module, and a social enhanced recommendation engine; The campus social domain module is responsible for building a strong relationship social network including student entities, community entities and course entities, and establishing a connection with the hybrid processing architecture through the strong relationship network; The off-campus self-media domain module is used to build a weak relationship network based on cross-school content collaboration and geographical location co-occurrence, and is connected to the hybrid processing architecture through the weak relationship network; The hybrid processing architecture adopts a dual-domain graph convolution kernel architecture, including a parallel graph neural network branch and a dynamic interest evolution network branch. The graph neural network branch generates a graph neural network contrast loss term, and the dynamic interest evolution network branch also has a dual loss. The hybrid processing architecture performs feature space alignment processing on the information and transmits the processed information to the cross-domain information fusion module; The cross-domain information fusion module fuses the information to generate fusion features, and then inputs the fusion features into the social enhancement recommendation engine; The recommendation engine is used to dynamically generate recommendation results based on user characteristics and user interests, and establish optimization objectives for the dual loss terms of the graph neural network branch comparison loss term and the dynamic interest evolution network branch.

2. A dual social model based on machine learning according to claim 1, characterized in that: The graph neural network branch is constructed using a hypergraph structure, and its hypergraph order is k=3; The graph neural network branch is configured with 8 attention heads (h g =8) nodes, whose representation dimension is d e =64; The graph neural network branch is connected through a cross-domain bridge matrix Realize dual-domain feature transfer.

3. The dual social model based on machine learning according to claim 1, characterized in that: The dual-domain graph convolution kernel includes a domain separation architecture, which is the campus social graph G intra =(V,E strong ) and off-campus relationship map G inter =(V,E weak ) constituted by a parallel input structure; Among them: the campus map node V contains students, clubs, and course entities, E strong Represents strong relationships between class affiliation and activity participation; weak Based on cross-school collaborative creation and weak connection construction based on geographical location co-occurrence; The dual-domain graph convolution kernel also includes a cross-domain attention layer, which is obtained by the domain-specific parameter matrix W d And the LeakyReLU activation function σ, the attention coefficient of node i in domain d is calculated as follows: Among them: a i,d Refers to the attention coefficient of node i in domain d; exp exponential function, used to calculate attention weight; σ refers to the LeakyReLU activation function; refers to the transpose of the attention vector of domain d; W d Refers to the weight matrix of domain d, which is used to perform linear transformation on node features; h i The feature vector of node i; || represents the feature concatenation operation, which concatenates two feature vectors together; N d (i) is the neighbor set of node i in domain d; h j and h k represents the feature vector of nodes j and k, h j and h k They represent the neighbor nodes of node i respectively; The dual-domain graph convolution kernel also includes a feature migration component, which is connected to the bridging matrix Linearly combine the cross-domain attention output with the intra-school features to generate the final node representation: Where: h′ represents the processed node feature; σ refers to the activation function; α i,d Refers to the attention coefficient; W d Refers to the weight matrix; h i refers to the feature vector of node i; B refers to the bridging matrix; Φ refers to the cross-domain attention output mapping; Refers to the feature vector of node j after itr iteration.

4. The dual social model based on machine learning according to claim 1, characterized in that: The dynamic interest evolution network branch includes the time window T w = 15-day temporal feature extraction layer; The dynamic interest evolution network branch also includes an interest evolution network composed of 3 layers of GRU, and the hidden layer dimension of the interest evolution network is d h =128; The dynamic interest evolution network branch includes a social influence injection unit, which integrates the group behavior trends output by the graph neural network branch through multi-head attention pooling.

5. A dual social model based on machine learning according to claim 4, characterized in that: The dynamic interest evolution network branch configures a time-aware social influence attenuation factor Where: λ(t) is the time-dependent decay factor; e is the base of the natural logarithm; B is a learnable parameter; t is the current time node; and t0 is the time node of special campus events (such as exam week and club recruitment).

6. The dual social model based on machine learning according to claim 1, characterized in that: The final node representation generation process of the cross-domain information fusion module is: Where: α is the dynamically adjusted domain weight coefficient, which is used to balance the contribution of intra and inter features; B is the cross-domain bridging matrix; The final feature representation of node i is the result of cross-domain information fusion; Feature representation of node i in the campus domain; The feature representation of node i in the off-campus domain; B is the cross-domain bridging matrix, which is used to realize feature migration and cross-domain feature fusion; (1-α) is the weight coefficient of the off-campus domain feature, which together with α determines the ratio of on-campus and off-campus features in the final feature representation.

7. The dual social model based on machine learning according to claim 1, characterized in that: The optimization objective of the recommendation engine includes the graph neural network branch contrast loss term: Set the number of negative sampling nodes K = 512 and the regularization coefficient α = 0.8; Where: L GNN Refers to the contrast loss of graph neural network; V refers to the set of all nodes; y v The true label of node v; The predicted label of node v; α is the regularization coefficient used to balance the weight of the loss term Refers to the distribution of positive samples p pos The node pairs sampled in (v i ,v j ) is the expected value of ; f(v) is the feature representation function of node v; τ is the temperature parameter used to control the sharpness of the contrast loss; k is the number of negative sampling nodes, that is, for each positive sample, k negative samples are sampled.

8. The dual social model based on machine learning according to claim 7, characterized in that: The optimization objective of the recommendation engine also includes a dual loss term of the dynamic interest evolution network branch; Configure cross entropy loss weight β = 1.2; Where: L DIEN Refers to the dual loss of interest evolution network; T w Refers to the size of the time window; l ce Refers to the cross entropy loss function; refers to the target feature at time t; Refers to the predicted features at time t; β refers to the cross entropy loss weight, which is used to balance the weights of the two loss terms; h t refers to the hidden state at time t; h t-1 refers to the hidden state at time t-1; |·|2 refers to the L2 norm, which is used to calculate the difference between hidden states.

9. A method for implementing a dual social model based on machine learning according to any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: The campus social domain module uses the strong relationship network to transfer campus social related data to the hybrid processing architecture; The off-campus self-media domain module transmits the information in the off-campus self-media domain to the hybrid processing architecture through the weak relationship network; Step 2: The hybrid processing architecture receives information from the on-campus social domain module and the off-campus self-media domain module. After aligning the feature space of this information using the parallel graph neural network branch and the dynamic interest evolution network branch, the processing results are passed to the cross-domain information fusion module. Step 3: The cross-domain information fusion module fuses the information transmitted by the hybrid processing architecture to generate fusion features, which are then input into the social enhanced recommendation engine. Step 4: The social enhanced recommendation engine generates recommendation results based on the received fusion features.

Images