Social network label propagation prediction method and device based on heterogeneous attention network

Abstract

The application provides a social network label propagation prediction method and device based on a heterogeneous attention network, the method comprising collecting cascade data of label propagation related to a social network event and user interaction relationships in a social network medium; constructing a label-user heterogeneous influence relationship graph and a user time sequence preference graph from the collected data, the node relationship of the heterogeneous influence relationship graph comprising an influence relationship between users, an influence relationship between a user and a label, and an influence relationship between labels, and the node relationship of the user time sequence preference graph comprising a user's preference for different labels over time; establishing a user preference-based heterogeneous time sequence attention network model, comprising a homogeneous influence relationship learning module, a heterogeneous influence relationship learning module, a heterogeneous user time sequence preference learning module, a feature fusion module, and a label propagation prediction module; processing a to-be-predicted user and a label and inputting them into the trained user preference-based heterogeneous time sequence attention network model, and outputting a prediction result of the to-be-predicted user participating in label propagation.

Description

Technical Field

[0001] This invention relates to the field of social network information dissemination, and more specifically to a method and apparatus for predicting the dissemination of social network tags based on heterogeneous attention networks. Background Technology

[0002] Social networking platforms, such as Sina Weibo and Twitter, attract hundreds of millions of users by sharing information and pushing trending topics. The emergence of social network hashtags provides users with a convenient mechanism to spread topics of shared interest. Therefore, real-time trending events on social platforms are often accompanied by the widespread dissemination of related hashtags. Furthermore, hashtags, as a form of information sharing, align with the information dissemination process. Predictive technologies analyzing user participation in hashtag dissemination can be applied to various task scenarios, such as topic analysis, social trend monitoring, and commercial recommendations.

[0003] In information propagation user prediction research, the task of predicting who will become the next user in a cascade has received extensive attention in recent years. The successful application of deep learning technology in various fields has further improved the performance of user prediction tasks. Most tasks focus on traditional information items (such as a tweet, an image, etc.) and use deep learning methods to capture the dependencies between users. However, the propagation of network tags is unobservable; users can directly add tags to tweets to cause propagation, making it impossible to determine the source of the previous user causing the propagation from the data stream. Therefore, implicit propagation relationships exist in the data, making it difficult to capture the true propagation cascade graph. Furthermore, most existing propagation user prediction methods are based on the influence relationships between users, such as using friend relationships as the underlying propagation network. However, in real tag propagation scenarios, when users post / retweet tags, they are not only influenced by previous users but also by their preferences for the tags themselves. Therefore, existing propagation user prediction methods cannot effectively predict the propagation of social network tags. Summary of the Invention

[0004] The present invention aims to provide a method and apparatus for predicting the propagation of social network tags based on heterogeneous attention networks, thereby solving the existing problems.

[0005] The technical solution is as follows: a method for predicting the propagation of social network tags based on heterogeneous attention networks, characterized by the following steps:

[0006] Step 1: In social media, collect cascading data of tag propagation and user interaction relationships related to social media events. The cascading data of tag propagation includes data of all users participating in tag propagation and behavioral time data. The user interaction relationships include the friend relationships and interaction relationships of all users participating in tag propagation.

[0007] Step 2: Construct a label-user heterogeneous influence relationship graph and a user time series preference graph from the collected data. The node relationships in the heterogeneous influence relationship graph include the influence relationship between users, the influence relationship between users and labels, and the influence relationship between labels. The node relationships in the user time series preference graph include the user's preference for different labels over time.

[0008] Step 3: Establish a heterogeneous temporal attention network model based on user preferences. The heterogeneous temporal attention network model includes the following sequential settings:

[0009] The homogeneous influence relationship learning module includes a multi-layer GCN model. The homogeneous influence relationship learning module learns the node representations of the label-user heterogeneous influence relationship graph in different feature spaces, and obtains the embedding representations of user nodes and labels respectively.

[0010] The heterogeneous influence relationship learning module employs a multi-layered GAT network to fuse the embedding representations of heterogeneous user nodes and labels, and then integrates them with the label-user heterogeneous influence relationship. Figure 1 The input is iterated through the heterogeneous influence relationship learning module, and then projected onto the same feature space through the heterogeneous influence relationship learning module to learn the association between the two types of nodes, and obtain the embedding representation of heterogeneous user nodes and the embedding representation of heterogeneous labels;

[0011] Heterogeneous user temporal preference learning module, the heterogeneous user temporal preference learning module includes:

[0012] A time representation unit, wherein the time representation unit uses a global hierarchical timestamp to represent the time feature of the user's behavior on the tag, and the time feature is embedded and represented through the time representation unit;

[0013] The user temporal attention learning unit takes the embedding representations of heterogeneous user nodes obtained through the heterogeneous influence relationship learning module, the embedding representations of heterogeneous labels, and the temporal embedding representations obtained through the temporal representation unit as inputs, and projects them onto an adjustable attention matrix to obtain the embedding representations of user nodes with background knowledge.

[0014] The feature fusion module employs a multi-head attention mechanism to learn the embedded representation of the label cascade. By taking the embedded representation of user nodes with background knowledge as input, it is fed into the label propagation cascade. The interaction characteristics between users are learned through the multi-head attention mechanism, and the representation matrix of the entire label propagation cascade is obtained. Then, the MLP network is used to map the representation matrix of the label propagation cascade into the label cascade vector representation.

[0015] The label propagation prediction module takes the obtained label concatenated vector representation as input and outputs the probability of candidate users participating in label propagation.

[0016] The heterogeneous temporal attention network model based on user preferences is trained using cross-entropy as the loss function to obtain the trained heterogeneous temporal attention network model based on user preferences.

[0017] Step 4: Process the users and labels to be predicted and input them into the trained heterogeneous temporal attention network model based on user preferences. Output the prediction results of the users to be predicted participating in label propagation.

[0018] Furthermore, in step 1, during the collection of cascaded data for tag propagation, network tags are used as the data source to collect data from all users participating in tag propagation. h and behavioral time t h The data is organized into the following formats:

[0019]

[0020] Among them, c h This represents the propagation cascade data of label h. Indicates user u n At time t n It participated in the spread of tag h; during the collection of user interaction relationships, it collected the friend relationships and interaction relationships of all users who participated in the spread of tag h. The friend relationships include online attention and fan relationships, and the interaction relationships include discussions and comments on tag h.

[0021] Furthermore, in step 2, a label-user heterogeneous influence graph is constructed, specifically as follows:

[0022] Constructing a diagram of the heterogeneous influence of labels and users Tag-User Heterogeneous Influence Relationship Diagram Includes user nodes and tag nodes Represented as:

[0023]

[0024] The set of nodes ε = {edge(u, u)} ∪ {edge(u, h)} ∪ {edge(h, h)}, where ∪ represents the union of the edges. When edge(u, u) = 1, it indicates an influence relationship between two user nodes; when edge(u, h) = 1, it indicates an influence relationship between a user and a label; and when edge(h, h) = 1, it indicates an influence relationship between two labels. These can respectively form a label-user heterogeneous influence relationship graph. The edge.

[0025] Furthermore, in step 2, a user time-series preference graph is constructed, specifically as follows:

[0026] Constructing user time-series preference graphs User time-series preference graph It includes a user node u, a time sequence attribute t, and a tag node h, and is represented as:

[0027]

[0028] Each user-time-label triple (u, t, h) represents the historical propagation behavior of user u toward label h over time t.

[0029] Furthermore, in step 3, the homogeneous node network in the label-user heterogeneous influence graph includes: the user-user relationship network. This represents an edge set containing only user-user relationships; a tag-tag relationship network. This represents an edge set that contains only label-label relationships;

[0030] In the homogeneous influence relationship learning module, a multi-layered GCN model is used to study the user-user relationship network G. u Learning node representations in different feature spaces is represented as follows:

[0031]

[0032] Where σ represents the ReLU activation function, express The degree matrix of the nodes, Represents graph G u The adjacency matrix, To initialize the user node embeddings into the first-layer GCN model, the homogeneous influence relationship learning module uses a normal distribution to randomly initialize the nodes. This represents the number of user nodes, and d represents the embedding dimension. The trainable model parameters are obtained by iterative learning through an L-layer GCN model, resulting in the embedding representation of the user node.

[0033] For the label relationship diagram The L-layer GCN model is used to embed the features of the labels into the corresponding feature space to obtain the embedded representation of the labels. Furthermore, in the heterogeneous influence relationship learning module, the embedding representation of heterogeneous user nodes is... Embedded representation of tags The data is then merged and compared with the tag-user heterogeneous influence graph. The heterogeneous influence relationship learning module is input together and iterated to obtain the heterogeneous node embeddings. Represented as:

[0034]

[0035]

[0036] in This represents the embedding representation of heterogeneous user nodes. Embedded representation of tags To splice, This is the initialization vector for the user node embeddings input to the GAT network for the first time, where d is the dimension of the embedding representation. Represents the heterogeneous influence relationship between labels and users. Let x be the set of neighboring nodes of node i in the graph. For node i, j represents all nodes adjacent to node i and with which there is an edge. i and x j Let α represent the representation vectors of nodes i and j, respectively, whose weights need to be learned for mutual influence. i，j It is the attention coefficient, representing the importance of node j to node i. These are trainable weight coefficients, [Wx i |Wx j ]express x in i and x j Feature concatenation is performed after mapping the weight matrix W, [Wx i |Wx k ]express x in i and x k After mapping the weight matrix W, feature concatenation is performed. k represents the summation process of all j-index symbols in the preceding summation symbol.

[0037] After L iterations using the GAT network, we obtain:

[0038]

[0039] in, It is a set of two types of node features. Then, the user nodes and tag nodes are split into matrices to obtain:

[0040]

[0041] and These are the embedding representations of heterogeneous user nodes and heterogeneous labels, respectively.

[0042] Furthermore, the feature is that, in the time representation unit of the heterogeneous user temporal preference learning module, a globally hierarchical timestamp (second, minute, hour, day, week) is used to represent the temporal characteristics of user behavior on tags, expressed as:

[0043] t′ g = [s, m, h, d, w], t′ g ∈R 5

[0044] Here, the five indicators [s, m, h, d, w] represent the second / minute / hour / day / week of the timestamp t, respectively, and these indicators are normalized to a numerical range of [-0.5, 0.5] to represent the offset of different time granularities; then, an MLP network is used to map the time features into a d-dimensional space, represented as...

[0045] t g =σ(Wt′) g +b), t g ∈R d

[0046] Where W represents trainable weights, b represents offset, and σ represents activation function;

[0047] In the user temporal attention learning unit, the embedding representations of heterogeneous users and labels are first concatenated. Then, the projection of temporal features is combined to reflect the temporal attentional relevance between users and labels. Finally, the user's preference representation for labels is learned through an attention mechanism, generating a user embedding representation with background knowledge. The specific formalization is as follows:

[0048]

[0049]

[0050] in, The embedding matrix represents the user node. The embedding matrix represents the labels. It is a user time series preference graph The set of all interest tags h of user u, α(u,t) g h) is the attention coefficient, representing the set of labels for user u. For user u at global time t g Attentional correlation, attention coefficient α(u,t) g (h) encodes the different preferences of user u for label h under different temporal features, x u express The feature representation of user u in the matrix, x h express The feature representation of label h in the matrix, [x u ||x h The symbol ] indicates that two vectors are concatenated. Let W represent the time feature matrix of the transposed (u, h), where (u, h) indicates that user u participated in the propagation of label h, and W ∈ R. d×2d It is represented as an attention matrix with adjustable parameters for users and tags under time importance, and LeakyReLU represents a non-linear activation function.

[0051] Furthermore, the user feature representation vector after mapping by the feature fusion module is represented as follows:

[0052]

[0053] in, Let m be the embedding representation of user m learned in the heterogeneous influence relationship learning module. This represents the user feature representation vector after mapping by the feature fusion module. It is a trainable parameter matrix, where d represents the embedding dimension and H represents the number of attention heads;

[0054] t in the formula p (m) represents the position code of the user in the sequence, where m represents the sequence position. The calculation process is as follows:

[0055] t p (m, 2j)=sin(m / (2|c m |) 2j / d )

[0056] t p (m,2j+1)=cos(m / (2|c m |) 2j / d )

[0057] Where j∈{1,…,|d / 2|},|c m| represents the maximum cascading length, and d represents the embedding dimension.

[0058] In the formula, Mask represents the mask matrix. The mask matrix Mask is added to the formula to shield future nodes, and the formalization is as follows:

[0059]

[0060] The representation matrix of the entire tag propagation cascade is as follows:

[0061]

[0062] Among them, W C It is a trainable parameter matrix;

[0063] The representation matrix C of the entire label propagation cascade h Inputting the MLP network yields the label concatenated vector representation:

[0064] X h =σ(C h W+b1)W′+b′

[0065] Where W and W′ are trainable parameter matrices, b1 and b′ are bias parameters, and the activation function σ uses the ReLU method.

[0066] Furthermore, when making predictions in the tag propagation prediction module, given an observable propagation cascade sequence... hour, This represents the m-th user participating in the propagation in the propagation cascade. Indicates the time of participation in the dissemination, and predicts the next user to participate in the dissemination in the future. By predicting future users Participation in the propagation cascade Implementation, represented as:

[0067]

[0068] For user prediction in the final propagation sequence, the label propagation prediction module calculates the probability of a user participating in label propagation. Formalized as follows:

[0069]

[0070] Where W is a trainable parameter matrix, which will train the sequence X hMapping to the user representation space to fit which user is more likely to participate in the propagation of the cascade sequence in the future, the mask matrix is ​​used to shield users before the prediction position to prevent leakage of prediction data. Finally, cross-entropy is used as the loss function to train the heterogeneous temporal attention network model based on user preferences.

[0071]

[0072] Where θ represents all trainable parameters in a heterogeneous temporal attention network model based on user preferences, if user u i Participating in the propagation of label h, in the propagation sequence c of label h. h If y is at position j, then y ji =1, otherwise y ji =0.

[0073] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that: when the processor executes the computer program, it implements the above-described method for predicting the propagation of social network event labels based on heterogeneous attention networks.

[0074] A computer-readable storage medium having a program stored thereon, characterized in that: when the program is executed by a processor, it implements the social network event label propagation prediction method based on heterogeneous attention networks as described above.

[0075] This invention establishes a heterogeneous temporal attention network model based on user preferences. Specifically, the model first introduces a heterogeneous user-tag influence network and uses a graph neural network to capture the influence relationships between the two types of nodes. Furthermore, to model users' temporal preferences for tags, a heterogeneous temporal attention mechanism that integrates tag preferences and row-time features is designed. By utilizing the historical interaction relationships between users and tags, and analyzing the data characteristics of users' interest preferences and temporal behaviors regarding tags, the potential associations between users and tags can be captured more accurately, and user behavior in the propagation of network tags can be predicted. Attached Figure Description

[0076] Figure 1 This is a schematic diagram illustrating the steps of a label propagation user prediction method using a heterogeneous temporal attention neural network in one embodiment of the present invention.

[0077] Figure 2 This is a flowchart illustrating the steps of a heterogeneous temporal attention neural network for label propagation and user prediction in one embodiment of the present invention.

[0078] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0079] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples, so that the process of how the present invention uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly. It should be noted that, as long as there is no conflict, the various embodiments and features in the various embodiments of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0080] Furthermore, the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0081] As described in the background section, existing methods for predicting user propagation cannot effectively predict the propagation of social network tags. However, modeling the propagation influence relationship between tags and users and the user's historical tag preferences to predict user behavior in the propagation of network tags has significant application value for network monitoring of social network event tags.

[0082] Therefore, see Figure 1 The present invention provides a method for predicting the propagation of social network tags based on heterogeneous attention networks, the method comprising at least the following steps:

[0083] Step 1: In social media, collect cascading data on the spread of tags and user interaction relationships related to social media events. The cascading data on tag spread includes data on all users participating in tag spread and their behavior time data. The user interaction relationships include the friend relationships and interaction relationships of all users participating in tag spread.

[0084] Step 2: Construct a label-user heterogeneous influence relationship graph and a user time series preference graph from the collected data. The node relationships in the heterogeneous influence relationship graph include the influence relationship between users, the influence relationship between users and labels, and the influence relationship between labels. The node relationships in the user time series preference graph include the user's preference for different labels over time.

[0085] Step 3: Establish a heterogeneous temporal attention network model based on user preferences. The heterogeneous temporal attention network model includes the following sequence settings:

[0086] The homogeneous influence relationship learning module includes a multi-layered GCN model. Through the homogeneous influence relationship learning module, the node representations of the label-user heterogeneous influence relationship graph in different feature spaces are learned, and the embedding representations of user nodes and labels are obtained respectively. Embedding refers to the process of projecting discrete data onto continuous vector representation, and the embedding representation is the vector obtained.

[0087] The heterogeneous influence relationship learning module employs a multi-layered GAT network to fuse the embedding representations of heterogeneous user nodes and labels to establish label-user heterogeneous influence relationships. Figure 1 The input heterogeneous influence relationship learning module is iterated, and then projected onto the same feature space through the heterogeneous influence relationship learning module to learn the association between the two types of nodes, and obtain the embedding representation of heterogeneous user nodes and the embedding representation of heterogeneous labels.

[0088] The heterogeneous user time series preference learning module includes:

[0089] The time representation unit uses a global hierarchical timestamp to represent the time characteristics of the user's behavior on the tag, and the time characteristics are embedded and represented through the time representation unit;

[0090] The user temporal attention learning unit takes the embedding representations of heterogeneous user nodes obtained through the heterogeneous influence relationship learning module, the embedding representations of heterogeneous labels, and the temporal embedding representations obtained through the temporal representation unit as inputs, and projects them onto an adjustable attention matrix to obtain the embedding representations of user nodes with background knowledge.

[0091] The feature fusion module uses a multi-head attention mechanism to learn the embedded representation of the label cascade. By taking the embedded representation of user nodes with background knowledge as input, it is fed into the label propagation cascade. The multi-head attention mechanism is used to learn the interaction characteristics between users and obtain the representation matrix of the entire label propagation cascade. Then, the MLP network is used to map the representation matrix of the label propagation cascade into the label cascade vector representation.

[0092] The label propagation prediction module takes the obtained label concatenated vector representation as input and outputs the probability of candidate users participating in label propagation.

[0093] The heterogeneous temporal attention network model based on user preferences is trained using cross-entropy as the loss function to obtain the trained heterogeneous temporal attention network model based on user preferences.

[0094] Step 4: Process the users and labels to be predicted and input them into the trained heterogeneous temporal attention network model based on user preferences. Output the prediction results of the users to be predicted participating in label propagation.

[0095] This invention addresses the task of predicting users in a tag propagation cascade. It utilizes the historical interaction relationship between users and tags and analyzes the data characteristics of users' interest preferences and temporal behavior towards tags to more accurately model and predict users who may participate in the tag propagation path.

[0096] This invention proposes a novel model called Heterogeneous Temporal Attention Network Model Based on User Preferences (UP-HTAN). It learns the mutual influence between users and labels through heterogeneous influence relationships. Traditional research initializes user features through random distribution, which may lead to error problems due to this cold start method. Then, the heterogeneous temporal attention mechanism is applied to learn the user's behavioral preferences and historical patterns of labels in terms of temporal features.

[0097] This invention utilizes the characteristics of time sequence, tags, and users in information dissemination, and has more reliable prediction accuracy for tag dissemination prediction scenarios. Therefore, for different scenarios, different deep learning model parameters can be obtained through training to better predict user behavior in network tag dissemination.

[0098] The method provided by this invention can be used in online public opinion event analysis and data mining, especially for the prediction and supervision of events with relatively concentrated semantics, such as social issues, judicial issues, and livelihood issues. It can also be used for the supervision of network information of enterprises, predicting the possible paths of the spread of tags that enterprises are concerned about, so as to assist in the work of dissemination diversion or blocking.

[0099] Specifically, in one embodiment of the present invention, in step 1, during the collection of cascaded data for tag propagation, network tags are used as the data source to collect data from all users participating in tag propagation. h and behavioral time t h The data is organized into the following formats:

[0100]

[0101] Among them, c h This represents the propagation cascade data of label h. Indicates user u n At time t n It participated in the spread of tag h; during the collection of user interaction relationships, it collected the friend relationships and interaction relationships of all users who participated in the spread of tag h. Friend relationships include online attention and fan relationships, and interaction relationships include discussions and comments on tag h.

[0102] In one embodiment of the present invention, in step 2, a label-user heterogeneous influence relationship graph and a user temporal preference graph are constructed from the collected data. The node relationships of the heterogeneous influence relationship graph include the influence relationship between users, the influence relationship between users and labels, and the influence relationship between labels. The node relationships of the user temporal preference graph include the user's external preference for different labels over time.

[0103] In this embodiment, a label-user heterogeneous influence relationship graph is constructed, specifically as follows:

[0104] Constructing a diagram of the heterogeneous influence of labels and users Tag-User Heterogeneous Influence Relationship Diagram Includes user nodes and tag nodes Represented as:

[0105]

[0106] The set of nodes ε = {edge(u, u)}∪{edge(u, h)}∪{edge(h, h)}, where ∪ represents the union of the two edges. When edge(u, u) = 1, it indicates an influence relationship between two user nodes, which can be determined by whether they follow or interact. When edge(u, h) = 1, it indicates an influence relationship between a user and a tag, which can be determined by whether the user has participated in the spread of the tag in the past. When edge(h, h) = 1, it indicates an influence relationship between two tags, which can be determined by whether the two tags co-occur in tweets or whether the tags are semantically similar. When edge(u, u) = 1, edge(u, h) = 1, and edge(h, h) = 1, a heterogeneous influence relationship graph between tags and users can be constructed. The edge.

[0107] In step 2 of the embodiment, a user time-series preference graph is constructed, specifically as follows:

[0108] Constructing user time-series preference graphs User time-series preference graph It includes a user node u, a time sequence attribute t, and a tag node h, and is represented as:

[0109]

[0110] Each user-time-label triple (u, t, h) represents the historical propagation behavior of user u toward label h over time t.

[0111] In step 3 of the embodiment, a heterogeneous temporal attention network model based on user preferences is constructed. Specifically, the heterogeneous temporal attention network model includes the following modules arranged in sequence: homogeneous influence relationship learning module, heterogeneous influence relationship learning module, heterogeneous user temporal preference learning module, feature fusion module, and label propagation prediction module.

[0112] Firstly, in the homogeneous influence relationship learning module, considering that user nodes and label nodes have different feature spaces, we first learn the embedding representations of user nodes and label nodes based on the relationship between homogeneous nodes. Then, after combining the feature matrices, we project them onto the same feature space using the GAT method.

[0113] The heterogeneous influence graph of the tag-user relationship diagram includes a homogeneous node network: user-user relationship network. This represents an edge set containing only user-user relationships; a tag-tag relationship network. This represents an edge set containing only label-label relationships; to capture the representations of different types of nodes in different feature spaces, a multi-layered GCN model is used in the homogeneous influence relationship learning module to study the user-user relationship network G. u Learning node representations in different feature spaces is represented as follows:

[0114]

[0115] Where σ represents the ReLU activation function, This represents the degree matrix of the nodes of Gu. Represents graph G u The adjacency matrix, To initialize the user node embeddings into the first-layer GCN model, the homogeneous influence relationship learning module uses a normal distribution to randomly initialize the nodes. This represents the number of user nodes, and d represents the embedding dimension. The trainable model parameters are obtained by iterative learning through an L-layer GCN model, resulting in the embedding representation of the user node.

[0116] Similarly, for label relationship graphs The L-layer GCN model is used to embed the features of the labels into the corresponding feature space to obtain the embedded representation of the labels.

[0117] To learn the different influences between heterogeneous nodes, the heterogeneous temporal attention network model based on user preferences of this invention embeds heterogeneous nodes into the same vector space through a heterogeneous influence relationship learning module. The heterogeneous influence relationship learning module uses graph attention network learning (GAT) to calculate the degree of influence between different types of nodes.

[0118] Embedding representation of heterogeneous user nodes Embedded representation of tags Combine and plot the relationship between tag-user heterogeneity. The data is fed into the GAT network for iteration to obtain the embeddings of heterogeneous nodes. Represented as:

[0119]

[0120]

[0121] in This represents the embedding representation of heterogeneous user nodes. Embedded representation of tags To splice, This is the initialization vector for the user node embeddings input to the GAT network for the first time, where d is the dimension of the embedding representation. Represents the heterogeneous influence relationship between labels and users. Let x be the set of neighboring nodes of node i in the graph. For node i, j represents all nodes adjacent to node i and with which there is an edge. i and x j Let α represent the representation vectors of nodes i and j, respectively, whose weights need to be learned for mutual influence. i，j It is the attention coefficient, representing the importance of node j to node i. These are trainable weight coefficients, [Wx i |Wx j ]express x in i and x j Feature concatenation is performed after mapping the weight matrix W, [Wx i |Wx k ]express x in i and x k After mapping the weight matrix W, feature concatenation is performed. k represents the summation process of all j-index symbols in the preceding summation symbol.

[0122] After L iterations using the GAT network, we obtain:

[0123]

[0124] in, It is a set of two types of node features. Then, the user nodes and tag nodes are split into matrices to obtain:

[0125]

[0126] and These are the embedding representations of heterogeneous user nodes and heterogeneous labels, respectively.

[0127] To better capture users' temporal preferences for different tags, inspired by graph attention mechanisms, a deep learning method based on an attention mechanism-aware embedding layer is designed in the user temporal preference learning module of a heterogeneous temporal attention network model. This method learns the attention representation of user preferences for tags in the user-tag temporal relationship graph. Three representations—user embeddings, embeddings of tags of interest, and embeddings of temporal relationships—are projected as background knowledge onto a specific representation with a parameterized attention matrix. Specifically, the model improves upon the traditional GAT algorithm by integrating the attention aggregation relationships of users on temporal edges and tags in a bipartite graph. This mechanism establishes a mechanism for different users' tag interests and preferences, ultimately generating embedded representations of user nodes with background knowledge.

[0128] In the time representation unit of the heterogeneous user time-series preference learning module in this embodiment, considering the time characteristics of user behavior during propagation, traditional sequence position information is insufficient to describe users' posting time preferences. Therefore, the heterogeneous user time-series preference learning module uses a globally hierarchical timestamp (second, minute, hour, day, week) to represent the time characteristics of user behavior related to tags, expressed as:

[0129] t′ g = [s, m, h, d, w], t′ g ∈R 5

[0130] Here, the five indicators [s, m, h, d, w] represent the second / minute / hour / day / week of the timestamp t, respectively. These indicators are normalized to a numerical range of [-0.5, 0.5], representing the offset of different time granularities. It is worth noting that the reason for choosing a week as the coarsest granularity is that the granularity of the longest potential user cycle is "week," thus representing the temporal characteristics of user behavior on the global timestamp. In order to map the global time features to a unified space, an MLP network is used to map the time features to the d-dimensional space, as shown below:

[0131] t g =σ(Wt′) g +b), t g ∈R d

[0132] Where W represents trainable weights, b represents offset, and σ represents activation function.

[0133] To better represent the relevance of users' interests and preferences during the propagation process, the user's background knowledge on labels and time features is projected onto an attention matrix with adjustable parameters in the user time-series attention learning unit of the user time-series preference learning module of the model.

[0134] In the user temporal attention learning unit, the embedding representations of users and labels are first concatenated. Then, the projection of parameter matrices and temporal features is combined to reflect the temporal correlation between users and labels. Finally, the user's preference representation for labels is learned through an attention mechanism, generating the user's embedding representation with background knowledge. The specific formalization is as follows:

[0135]

[0136]

[0137] in, The embedding matrix represents the user node. The embedding matrix represents the labels. It is a user time series preference graph The set of all interest tags h of user u, α(u,t) g h) is the attention coefficient, representing the set of labels for user u. For user u at global time t g Attentional correlation, attention coefficient α(u,t) g (h) encodes the different preferences of user u for label h under different temporal features, x u express The feature representation of user u in the matrix, x h express The feature representation of label h in the matrix, [x u ||x h The symbol ] indicates that two vectors are concatenated. Let W represent the time feature matrix of the transposed (u, h), where (u, h) indicates that user u participated in the propagation of label h, and W ∈ R. d×2d It is represented as an attention matrix with adjustable parameters for users and tags under time importance, and LeakyReLU represents a non-linear activation function.

[0138] In the example, within the feature fusion module, for label propagation cascade c... h User embedding representation in tag propagation cascade The user feature representation vector after mapping by the feature fusion module is represented as follows:

[0139]

[0140] in, Let m be the embedding representation of user m learned in the heterogeneous influence relationship learning module. This represents the user feature representation vector after mapping by the feature fusion module. It is a trainable parameter matrix, where d represents the embedding dimension and H represents the number of attention heads;

[0141] t in the formula p (m) represents the position code of the user in the sequence, where m represents the sequence position. The calculation process is as follows:

[0142] t p (m, 2j) = sin(m / (2|cm|)) 2j / d )

[0143] t p (m,2j+1)=cos(m / (2|c m |) 2j / d )

[0144] Where j∈{1,…,|d / 2|},|c m | represents the maximum cascading length, and d represents the embedding dimension;

[0145] In the formula, Mask represents the mask matrix. The mask matrix Mask is added to the formula to shield future nodes, and the formalization is as follows:

[0146]

[0147] The representation matrix of the entire tag propagation cascade is as follows:

[0148]

[0149] Among them, W C It is a trainable parameter matrix;

[0150] The representation matrix C of the entire label propagation cascade h Inputting the MLP network yields the label concatenated vector representation:

[0151] X h =σ(C h W+b1)W′+b′

[0152] Where W and W′ are trainable parameter matrices, b1 and b′ are bias parameters, and the activation function σ uses the ReLU method.

[0153] In this embodiment, the label propagation prediction module predicts the propagation cascade sequence. Given an observable propagation cascade sequence... hour, This represents the m-th user participating in the propagation in the propagation cascade. Indicates the time of participation in the dissemination, and predicts the next user to participate in the dissemination in the future. By predicting future users Participation in the propagation cascade Implementation, represented as:

[0154]

[0155] For user prediction in the final propagation sequence, the label propagation prediction module calculates the probability of a user participating in label propagation. Formalized as follows:

[0156]

[0157] Where W is a trainable parameter matrix, which will train the sequence X h Mapping to the user representation space to fit which user is more likely to participate in the propagation of the cascade sequence in the future, the mask matrix is ​​used to shield users before the prediction position to prevent leakage of prediction data. Finally, cross-entropy is used as the loss function to train the heterogeneous temporal attention network model based on user preferences.

[0158]

[0159] Where θ represents all trainable parameters in a heterogeneous temporal attention network model based on user preferences, if user u i Participating in the propagation of label h, in the propagation sequence c of label h. h If y is at position j, then y ji =1, otherwise y ji =0, y ji =1, y ji =0 corresponds to the actual result, indicating the situation that occurred, so it is 0 or 1, used for optimization during the training phase.

[0160] In the embodiment, the propagation probability predicted in the tag propagation prediction module It refers to the probability of a user participating in the spread of this tag. Predict the probability for each user in the user set and select the one with the highest probability.

[0161] With the continuous advancement of deep learning technology, the performance of predicting who will become the next user in a cascade of information propagation has been continuously improving in user prediction research. Currently, most tasks focus on traditional information items (such as a tweet, an image, etc.) and use deep learning methods to capture the dependencies between users. However, the propagation of network tags is unobservable; users can directly add tags to tweets to cause propagation, and it is impossible to determine from the data stream who the previous level source is, making it difficult to capture the true propagation cascade graph. Furthermore, most existing propagation user prediction methods are based on the influence relationships between users, such as using friend relationships as the underlying propagation network. However, in real tag propagation scenarios, when users post / forward tags, they are not only influenced by previous level users, but also by their preferences for the tags themselves. Based on this, this implementation proposes a heterogeneous temporal attention network model based on user preferences (UP-HTAN). Specifically, it models the influence relationship between tags and users through a heterogeneous graph neural network and learns representations of the mutual dependencies between tags and users. Secondly, a heterogeneous temporal attention mechanism is proposed to model users' preferences for tags in time sequence, and the user representations obtained by these models are used to drive the real tag propagation cascade to more accurately predict user behavior in network tag propagation.

[0162] This architecture has two advantages:

[0163] (1) More granular modeling and analysis capabilities. This invention analyzes the propagation trend of hot tags, constructs user propagation data that focuses on tags, extracts user preference features for tags, and performs granular modeling of user relationships and user preferences. This allows for a more thorough learning of user characteristics and provides support for more accurate prediction of tag propagation cascade users.

[0164] (1) More accurate model prediction capability. This invention considers two factors: the influence relationship between user-tag nodes and user preferences for tag content. It constructs a heterogeneous influence network where user nodes and tag nodes coexist, and uses a graph neural network to capture the influence relationship between the two types of nodes. Then, it utilizes a heterogeneous temporal attention mechanism to learn user behavioral preferences and historical patterns regarding tags in terms of temporal features, thereby more accurately capturing the potential association between users and tags. Experimental results on four real-world propagation datasets show that the proposed model outperforms the state-of-the-art baseline method MS-HGAT (paper title: MS-HGAT: Memory-Enhanced Sequential HypergraphAttention Network for Information Diffusion Prediction) in the user prediction task of tag propagation. While MS-HGAT is based on homogeneous user analysis and does not consider the heterogeneous interaction between tags and users, the UP-HTAN model of this invention exhibits superior prediction performance, with average improvements of 9.12% and 7.45% in the Hits@100 and MAP100 metrics, respectively. Details are shown in Table 1.

[0165] Table 1

[0166] Comparative Experiment Results of UP-HTAN Prediction Task - Hits@Indicators

[0167]

[0168] Cascaded deep learning models

[0169]

[0170] Comparative Experiment Results of UP-HTAN Prediction Tasks - MAP@Indicators

[0171]

[0172] In an embodiment of the present invention, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for predicting the propagation of social network tags based on heterogeneous attention networks.

[0173] The computer device can be a terminal, and its internal structure diagram can be as follows: Figure 3As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting the propagation of social network tags based on heterogeneous attention networks. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0174] Memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). Memory stores programs, and the processor executes these programs after receiving execution instructions.

[0175] A processor can be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. The processor can also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor can be a microprocessor or any conventional processor.

[0176] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0177] In an embodiment of the present invention, a computer-readable storage medium is also provided, on which a program is stored, characterized in that: when the program is executed by a processor, it implements the above-described method for predicting the propagation of social network tags based on heterogeneous attention networks.

[0178] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, computer devices, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0179] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, computer apparatuses, or computer program products according to embodiments of the invention. These computer program instructions may be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in the flowchart.

[0181] The foregoing has provided a detailed description of the method for predicting the propagation of social network tags based on heterogeneous attention networks, the computer device, and the application of the computer-readable storage medium provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for predicting the propagation of social network tags based on heterogeneous attention networks, characterized in that, Includes the following steps: Step 1: In social media, collect cascading data of tag propagation and user interaction relationships related to social media events. The cascading data of tag propagation includes data of all users participating in tag propagation and behavioral time data. The user interaction relationships include the friend relationships and interaction relationships of all users participating in tag propagation. Step 2: Construct a label-user heterogeneous influence relationship graph and a user time series preference graph from the collected data. The node relationships in the heterogeneous influence relationship graph include the influence relationship between users, the influence relationship between users and labels, and the influence relationship between labels. The node relationships in the user time series preference graph include the user's preference for different labels over time. Step 3: Establish a heterogeneous temporal attention network model based on user preferences. The heterogeneous temporal attention network model includes the following sequential settings: The homogeneous influence relationship learning module includes a multi-layer GCN model. The homogeneous influence relationship learning module learns the node representations of the label-user heterogeneous influence relationship graph in different feature spaces, and obtains the embedding representations of user nodes and labels respectively. The heterogeneous influence relationship learning module employs a multi-layer GAT network to fuse the embedding representations of heterogeneous user nodes and labels. These embedding representations are then input into the heterogeneous influence relationship learning module along with the label-user heterogeneous influence relationship graph for iterative processing. The graph is then projected onto the same feature space to learn the association between the two types of nodes and obtain the embedding representations of heterogeneous user nodes and heterogeneous labels. Heterogeneous user temporal preference learning module, the heterogeneous user temporal preference learning module includes: A time representation unit, wherein the time representation unit uses a global hierarchical timestamp to represent the time feature of the user's behavior on the tag, and the time feature is embedded and represented through the time representation unit; The user temporal attention learning unit takes the embedding representations of heterogeneous user nodes obtained through the heterogeneous influence relationship learning module, the embedding representations of heterogeneous labels, and the temporal embedding representations obtained through the temporal representation unit as inputs, and projects them onto an adjustable attention matrix to obtain the embedding representations of user nodes with background knowledge. The feature fusion module employs a multi-head attention mechanism to learn the embedded representation of the label cascade. By taking the embedded representation of user nodes with background knowledge as input, it is fed into the label propagation cascade. The interaction characteristics between users are learned through the multi-head attention mechanism, and the representation matrix of the entire label propagation cascade is obtained. Then, the MLP network is used to map the representation matrix of the label propagation cascade into the label cascade vector representation. The label propagation prediction module takes the obtained label concatenated vector representation as input and outputs the probability of candidate users participating in label propagation. The heterogeneous temporal attention network model based on user preferences is trained using cross-entropy as the loss function to obtain the trained heterogeneous temporal attention network model based on user preferences. Step 4: Process the users and labels to be predicted and input them into the trained heterogeneous temporal attention network model based on user preferences. Output the prediction results of the users to be predicted participating in label propagation.

2. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 1, characterized in that, In step 1, during the collection of cascading data for tag propagation, network tags are used as the data source to collect data from all participating tags. Disseminating users and the time of the behavior The data is organized into the following formats: ； in, Indicates label Cascaded propagation data, Indicates user In time Participated in the tag The dissemination of tags; during the collection of user interaction relationships, the friend relationships and interaction relationships of all users participating in tag dissemination are collected. The friend relationships include online following and fan relationships, and the interaction relationships include those related to tags. Commentary relationship.

3. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 2, characterized in that, In step 2, a label-user heterogeneous influence graph is constructed, specifically as follows: Constructing a diagram of the heterogeneous influence of labels and users Tag-User Heterogeneous Influence Relationship Diagram Includes user nodes and tag nodes , is represented as: ； The set of nodes , , To represent taking the union of sets, This indicates that there is an influence relationship between two user nodes. This indicates that there is an influence relationship between the user and the tag. This indicates that there is an influence relationship between the two labels, and they can each form a label-user heterogeneous influence relationship graph. The edge.

4. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 3, characterized in that, In step 2, the user time series preference graph is constructed, specifically as follows: Constructing user time-series preference graphs User time series preference graph Includes user nodes Timing attributes and tag nodes , is represented as: ； User-Time-Label Triple Characterizing users For tags In time sequence Historical dissemination behavior.

5. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 4, characterized in that, In step 3, the homogeneous node network in the label-user heterogeneous influence graph includes: the user-user relationship network. , This represents an edge set containing only user-user relationships; a tag-tag relationship network. , This represents an edge set that contains only label-label relationships; In the homogeneous influence relationship learning module, a multi-layered GCN model is used to study the user-user relationship network. Learning node representations in different feature spaces is represented as follows: ； in, This represents the ReLU activation function. express The degree matrix of the nodes, Representation diagram The adjacency matrix, To initialize the user node embeddings into the first-layer GCN model, the homogeneous influence relationship learning module uses a normal distribution to randomly initialize the nodes. , Indicates the number of user nodes. Indicates the embedding dimension. Represents the trainable model parameters, through After iterative learning, the layered GCN model obtains the embedding representation of user nodes. ; For the label relationship diagram An L-layer GCN model is used to embed the features of the labels into the corresponding feature space, thus obtaining the embedded representation of the labels. .

6. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 5, characterized in that, In the heterogeneous influence relationship learning module, the embedding representation of heterogeneous user nodes is used. Embedded representation of tags The data is then merged and compared with the tag-user heterogeneous influence graph. The heterogeneous influence relationship learning module is input together and iterated to obtain the heterogeneous node embeddings. , is represented as: ； ； in , , This represents the embedding representation of heterogeneous user nodes. Embedded representation of tags To splice, This is the initialization vector for the user node embeddings that are first input into the GAT network. It is the dimension of the embedded representation. Represents the heterogeneous influence relationship between labels and users. Nodes in the diagram The set of neighboring nodes, for graph nodes , Let i represent all nodes that are adjacent to node i and have an edge with it. Therefore, and These represent the nodes whose mutual influence weights need to be learned. and nodes The representation vector; It is the attention coefficient, representing the node For nodes The importance of These are trainable weight coefficients. express In and Feature concatenation is performed after mapping on the weight matrix W. express In and Feature concatenation is performed after mapping on the weight matrix W. In the preceding summation symbol, it indicates that all The process of summing subscripts; When using GAT network to complete After the iteration, we get: ； in, It is a set of two types of node features. Then, the user nodes and tag nodes are split into matrices to obtain: ； and These are the embedding representations of heterogeneous user nodes and heterogeneous labels, respectively.

7. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 6, characterized in that, In the time representation unit of the heterogeneous user temporal preference learning module, a global hierarchical timestamp is used to represent the temporal characteristics of user behavior on tags. The global hierarchical timestamp includes seconds, minutes, hours, days, and weeks, and is represented as follows: ； in, The five indicators represent timestamps. The time features are determined by the second / minute / hour / day / week, and these metrics are normalized to a numerical range of [-0.5, 0.5] to represent the offset at different time granularities. Then, a single-layer MLP network is used to process the time features. A mapping of dimensional space is represented as ； in, Indicates trainable weights. Indicates the offset. Indicates the activation function; In the user temporal attention learning unit, the embedding representations of heterogeneous users and labels are first concatenated. Then, the projection of temporal features is combined to reflect the temporal attentional relevance between users and labels. Finally, the user's preference representation for labels is learned through an attention mechanism, generating a user embedding representation with background knowledge. The specific formalization is as follows: ； ； in, The embedding matrix represents the user node. The embedding matrix represents the labels. It is a user time series preference graph Chinese users All interest tags The set, The attention coefficient represents the user's attention level. tag collection For users In global time Attention correlation, attention coefficient Encoded user For tags The different preferences that arise under different time characteristics, express Users in the matrix Feature representation, express Labels in the matrix Feature representation, This indicates that two vectors are concatenated. Indicates the transposed form The time feature matrix, Indicates user Participated in the tag The spread of the behavior This is represented as an attention matrix with adjustable parameters for users and tags under time importance. This represents a non-linear activation function.

8. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 7, characterized in that, The user feature representation vector after mapping by the feature fusion module is represented as follows: ； in, This represents the embedding representation of user m learned in the heterogeneous influence relationship learning module. This represents the user feature representation vector after mapping by the feature fusion module. It is a trainable parameter matrix. Indicates the embedding dimension. Indicates the number of heads of attention; In the formula This indicates the user's position encoding within the sequence. The sequence position is represented by the following calculation process: ； ； in , Indicates the maximum cascade length. Indicates the embedding dimension; In the formula This represents the mask matrix, and the formula includes the mask matrix. Shielding future nodes can be formalized as follows: ； The representation matrix of the entire tag propagation cascade is as follows: ； in, It is a trainable parameter matrix; The representation matrix of the entire label propagation cascade Inputting the MLP network yields the label concatenated vector representation: ； in, It is a trainable parameter matrix. and These are bias parameters and activation functions. Used method.

9. The method for predicting the propagation of social network tags based on heterogeneous attention networks according to claim 8, characterized in that, When making predictions in the tag propagation prediction module, given an observable propagation cascade sequence... hour, Indicating the propagation cascade, the first One user who participated in the dissemination, Indicates the time of participation in the dissemination, and predicts the next user to participate in the dissemination in the future. By predicting future users Participation in the propagation cascade Implementation, represented as: ； For user prediction in the final propagation sequence, the label propagation prediction module calculates the probability of a user participating in label propagation. The formalization is as follows: ； in It is a trainable parameter matrix that stores the sequence Mapping to the user representation space to fit which user is more likely to participate in the propagation of the propagation cascade sequence in the future. The matrix is ​​used to mask users before the predicted position, preventing the leakage of prediction data. Finally, cross-entropy is used as the loss function to train the heterogeneous temporal attention network model based on user preferences. ； in This represents all trainable parameters in a heterogeneous temporal attention network model based on user preferences, if the user... Participation Tags The spread of the label Propagation sequence In the middle The position, then ,otherwise .

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the social network event label propagation prediction method based on heterogeneous attention networks as described in claim 1.

11. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by the processor, it implements the social network event label propagation prediction method based on heterogeneous attention network as described in claim 1.

Images