A method for predicting information dissemination heat based on dynamic evolution inside and outside a community

Abstract

The application discloses a kind of information propagation heat prediction method based on community inside and outside dynamic evolution, method includes obtaining cascade diagram, constructs community graph structure, constructs community graph structure and information propagation heat prediction.The application belongs to social network data processing field, specifically refers to a kind of information propagation heat prediction method based on community inside and outside dynamic evolution, the scheme is based on the newly added participating node constructs instruction information and forms mask adjacent relationship, carries out community graph structure information propagation update on mask adjacent relationship and is aggregated to obtain community representation;Community graph structure is built, and the edge feature is extracted to the community interrelation and is encoded, and the community representation is updated on community graph structure to obtain intercommunity representation;Multi-head attention aggregation is generated to segmented cascade representation for intercommunity representation, combined with segmented growth statistics to form time series input and carry out sequence coding, output information propagation heat prediction result in future time window.

Description

Technical Field

[0001] This invention relates to the field of social network data processing technology, specifically to a method for predicting the popularity of information dissemination based on the dynamic evolution of information spread within and outside communities. Background Technology

[0002] Information dissemination popularity prediction is the problem of predicting the scale of information dissemination in a network, and it is often used in content recommendation and resource scheduling scenarios. However, general information dissemination popularity prediction methods suffer from several problems: insufficient cascading modeling from a community perspective; insufficient characterization of incremental interaction effects within a community; difficulty in characterizing inter-community dependencies and cross-community dissemination drivers; and insufficient joint modeling of hierarchical information and its temporal evolution within communities and cascades. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method for predicting the popularity of information dissemination based on the dynamic evolution within and outside communities. Addressing the problems of insufficient cascading modeling from a community perspective and inadequate characterization of incremental interactions within communities in general information dissemination popularity prediction methods, this scheme divides the cascading graph into communities and segments it by time within the observation window. Within each time period, it constructs a community-level graph structure based on communities, constructs indicator information based on newly added participating nodes, and forms masked adjacency relationships. It then performs information dissemination updates on the community-level graph structure based on these masked adjacency relationships and aggregates the information to obtain the community representation. This method addresses the limitations of general information dissemination popularity prediction methods. The previous method had the problem of difficulty in characterizing inter-community dependencies and cross-community propagation drivers. This solution constructs an inter-community graph structure, extracts and encodes edge features of inter-community associations, and performs propagation updates on the inter-community graph structure based on an attention update mechanism to obtain inter-community representations. To address the problem that general information propagation heat prediction methods are insufficient in jointly modeling the hierarchical information of communities and cascades and their temporal evolution, this solution performs multi-head attention aggregation on the inter-community representations to generate segmented cascaded representations. Combined with segmented growth statistics, it forms a temporal input and performs sequence encoding to output the information propagation heat prediction results within the future time window.

[0004] The technical solution adopted by this invention is as follows: This invention provides a method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside a community. The method includes the following steps:

[0005] Step S1: Obtain the cascade graph;

[0006] Step S2: Construct the community internal graph structure;

[0007] Step S3: Construct the inter-community graph structure;

[0008] Step S4: Predict the popularity of information dissemination.

[0009] Further, in step S1, obtaining the cascade graph specifically includes:

[0010] Step S11: Obtain the propagation interaction records corresponding to the target information; based on the propagation interaction records, construct an information cascade graph within the observation window;

[0011] Step S12: Perform community detection using the nodes and edges of the cascaded graph within the observation window;

[0012] Step S13: Set the start and end times of the observation window, divide the observation window into K consecutive time periods, and obtain a set of time periods;

[0013] Step S14: Construct a propagation edge set; Distribute the propagation edge set within the observation window to each time period according to the timestamp, obtain the propagation edge subset corresponding to the time period, and determine the set of nodes appearing within the time period; Combine the community set, for any time period and any community, form the set of nodes belonging to the community within the time period, and obtain the propagation edge set within the community within the time period.

[0014] Step S15: Define the cutoff time t k The cumulative set of participating nodes is defined, and the set of newly participating nodes in the k-th time period is defined; the set of newly participating nodes belonging to the community in the k-th time period is obtained based on the community set.

[0015] Furthermore, in step S2, the construction of the community intragraph structure specifically includes:

[0016] Step S21: Based on the time period set and community set, for any time period and community, obtain the set of nodes belonging to the community within the time period and the set of propagation edges within the community, and construct the subgraph within the community of the community within the time period;

[0017] Step S22: For any subgraph within a community, determine the initial representation of each node in the node set;

[0018] Step S23: For any time period and community, construct new participation indication information based on the set of newly added participating nodes and each node in the subgraph within the community;

[0019] Step S24: Within the neighbor range defined by the mask adjacency relationship, the node representation is propagated and updated layer by layer; after the propagation update is completed, the updated representation of the node is obtained;

[0020] Step S25: For any time period and community, use mean pooling to aggregate the update representations of all nodes in the community to obtain the community representation for the corresponding time period.

[0021] Furthermore, in step S3, the construction of the inter-community graph structure specifically includes:

[0022] Step S31: For any time period, based on the community representation set, and stacked according to the community index, a community node feature matrix is ​​formed;

[0023] Step S32: For any given time period, construct an inter-community graph;

[0024] Step S33: For any inter-community edge, extract edge features to characterize the inter-community propagation driving factors; input the edge features into the edge feature encoder to obtain the edge embedding;

[0025] Step S34: Based on the inter-community graph, perform attention propagation update on the community node features and edge embeddings to obtain the inter-community representation set for the time period.

[0026] Furthermore, in step S4, the prediction of information dissemination popularity specifically includes:

[0027] Step S41: For any time period, obtain the inter-community representation set;

[0028] Step S42: Construct segmented concatenated representations; For any time period, perform multi-head attention aggregation on the inter-community representations of each community to generate segmented concatenated representations; Calculate attention weights for each community for each attention head and normalize the weights; Characterize the contribution of each community to the concatenated representation of the current time period based on the attention weights; Perform weighted aggregation on the inter-community representations according to the attention weights to obtain the aggregation results of each attention head, and concatenate the results of each attention head to obtain the segmented concatenated representation of the time period;

[0029] Step S43: For any time period, calculate the segmented growth statistical vector, input the segmented growth statistical vector into the growth statistical encoder to obtain the statistical embedding, and concatenate it with the segmented concatenated representation to form the time series input, thus obtaining the time series input set;

[0030] Step S44: Input the segmented concatenated temporal input set into the sequence encoder to model the temporal evolution within the observation window and obtain the sequence hidden state set; output the information propagation heat prediction result based on the final sequence hidden state.

[0031] The beneficial effects achieved by the present invention using the above solution are as follows:

[0032] (1) In view of the problems of insufficient cascade modeling from the perspective of community and insufficient characterization of incremental interaction effects within the community, the present scheme divides the cascade graph into communities and segments them by time within the observation window. In each time period, the community graph structure is constructed according to the community. Based on the newly added participating nodes, indicator information is constructed and mask adjacency relationship is formed. The community graph structure information propagation and update is performed on the mask adjacency relationship and the community representation is obtained by aggregation.

[0033] (2) In view of the problem that general information dissemination heat prediction methods have difficulty in characterizing inter-community dependencies and cross-community dissemination drivers, this scheme constructs an inter-community graph structure, extracts edge features from inter-community associations and encodes them, and performs dissemination updates on the inter-community graph structure based on the attention update mechanism to obtain inter-community representations.

[0034] (3) In view of the problem that the general information dissemination heat prediction method is insufficient in the joint modeling of the hierarchical information of the community and the cascade and its temporal evolution, this scheme performs multi-head attention aggregation on the community representation to generate segmented cascaded representation, combines segmented growth statistics to form time series input and performs sequence encoding, and outputs the information dissemination heat prediction results within the future time window. Attached Figure Description

[0035] Figure 1 A flowchart illustrating a method for predicting the popularity of information dissemination based on dynamic evolution within and outside a community, provided by this invention;

[0036] Figure 2 This is a flowchart illustrating step S1;

[0037] Figure 3 This is a flowchart illustrating step S2;

[0038] Figure 4 This is a flowchart illustrating step S3;

[0039] Figure 5 This is a flowchart illustrating step S4.

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0043] Example 1, see Figure 1 This invention provides a method for predicting the popularity of information dissemination based on the dynamic evolution within and outside a community. The method includes the following steps:

[0044] Step S1: Obtain the cascaded graph, perform community detection on the cascaded graph to determine the community corresponding to each node, and divide the observation window into multiple consecutive time periods;

[0045] Step S2: Construct the community intragraph structure, construct indication information based on newly added participating nodes within the time period and form masked adjacency relationships, perform community intragraph structure information propagation and update on the masked adjacency relationships and aggregate to obtain the community representation;

[0046] Step S3: Construct an inter-community graph structure, extract and encode edge features of inter-community associations, and propagate and update the community representations on the inter-community graph structure based on an attention update mechanism to obtain the inter-community representations;

[0047] Step S4: Information dissemination heat prediction. Perform multi-head attention aggregation on the inter-community representations. Each attention head calculates the attention weight for each community. The inter-community representations are weighted and aggregated according to the attention weights to obtain the aggregation results of each attention head. The results of each attention head are concatenated to obtain the segmented concatenated representation of the time period. Combine the segmented growth statistics to form the time series input and perform sequence encoding to output the information dissemination heat prediction results within the future time window.

[0048] Example 2, see Figure 2 This embodiment is based on the above embodiment. In step S1, obtaining the cascade graph specifically includes:

[0049] Step S11: Obtain the propagation interaction records corresponding to the target information. These records are derived from social media forwarding logs, comment / reply logs, topic participation logs, or paper citation logs. Each propagation interaction record must contain at least the source user identifier, the recipient user identifier, and the timestamp of the interaction. Based on the propagation interaction records, construct an information cascading graph within the observation window. , is represented as: Among them, V cas (t o ) represents the observation cutoff time t. o The set of participating nodes, E cas (t o Let be the set of propagation edges within the observation window; any propagation edge is represented as: ; where u i For the source user node of the propagation, u j For the user node being propagated, t ij The timestamps of the propagation interactions are used; i and j are user indices, i ≠ j;

[0050] Step S12: To avoid information leakage, community detection is performed only using the nodes and edges of the cascaded graph within the observation window to obtain the community set; the Louvain community detection algorithm is used as the community detection operator; when the cascaded graph is a directed graph, the propagation edges are treated as undirected edges for community detection;

[0051] Step S13: Set the start and end times of the observation window, and divide the observation window into K consecutive time periods using equal durations to obtain a set of time periods;

[0052] Step S14: Construct the propagation edge set; convert the propagation edge set E within the observation window. cas (t o The nodes are assigned to different time periods according to their timestamps to obtain the subset of propagation edges corresponding to the k-th time period, and the set of nodes appearing in the k-th time period is determined. Combined with the community set obtained in step S12, for any time period and any community, a set of nodes belonging to the community within the time period is formed. The propagation edge set within the community during the time period is obtained. (Both endpoints are located within the same community), and m serves as the input index for constructing the community internal graph structure by community in step S2; m is the community index.

[0053] Step S15: Define the cutoff time t k The cumulative set of participating nodes is V cas (t k Define the set of newly added participating nodes for the k-th time period. , is represented as: ; It is the set of participating nodes in the (k-1)th time period; based on the community set, it is the set of nodes belonging to community C in the kth time period. m New participating node set , is represented as: ; This serves as the basis for constructing the indication information for newly added participating nodes and the mask adjacency relationship in step S2.

[0054] Example 3, see Figure 3 This embodiment is based on the above embodiment. In step S2, the construction of the community internal graph structure specifically includes:

[0055] Step S21: Based on the time period set and community set obtained in Example 2, for any time period k and community C m Acquired within the time period belonging to community C m Node set and community-based dissemination of information. And construct community C within the k-th time period. m Community sub-map , is represented as: ;in, and It can be obtained directly from step S14 of Example 2;

[0056] Step S22: For any subgraph within a community , is a set of nodes Each node u in the algorithm determines the initial representation of the node. ; It is obtained by splicing together user attribute features, user structural embedding in social networks, and local propagation structural features within the cascade;

[0057] Step S23: For any time period and community, based on the set of newly added participating nodes obtained in step S15 of Example 2. Based on subgraphs within the community Each node in the process adds participation indication information: when node u belongs to When node u is not in the specified position, set the indicator flag of node u to 1; when node u does not belong to the specified position, set the indicator flag of node u to 1. When the time comes, the indicator flag of node u is set to 0, based on the subgraph within the community. The adjacency relationship is constructed using a masked adjacency relationship. For any edge connecting two nodes in the subgraph within the community, if the indicator flag of at least one endpoint node of the edge is 1, then the adjacency relationship is preserved in the masked adjacency relationship; if the indicator flags of both endpoint nodes of the edge are 0, then the adjacency relationship is masked in the masked adjacency relationship.

[0058] Step S24: Within the neighbor range defined by the mask adjacency relationship, perform layer-by-layer propagation and update of the node representation; for any layer and any node, aggregate the neighbor representations from the set of nodes adjacent to node u under the mask adjacency relationship, and update them in combination with the node's own representation to obtain the representation of the next layer. l is the update layer index, and u is the node index; aggregation uses summation aggregation, and updates are implemented using a combination of linear transformation layers (Linear) and nonlinear activation functions (ReLU); after completing the L-layer propagation update, the updated representation of the node is obtained. ;

[0059] Step S25: For any time period and community, use mean pooling to aggregate the update representations of all nodes within the community to obtain the community representation c for the corresponding time period. k,m , is represented as: The set of nodes is not empty.

[0060] By performing the above operations, this scheme addresses the problems of insufficient cascading modeling from a community perspective and insufficient characterization of incremental interaction effects within communities in general information dissemination heat prediction methods. Within the observation window, the cascading graph is divided into communities and segmented by time. In each time period, a community intragraph structure is constructed according to the community. Indicator information is constructed based on newly added participating nodes, and masked adjacency relationships are formed. The community intragraph structure information is disseminated and updated on the masked adjacency relationships and aggregated to obtain the community representation.

[0061] Example 4, see Figure 4 This embodiment is based on the above embodiment. In step S3, the construction of the inter-community graph structure specifically includes:

[0062] Step S31: For any time period, based on the community representation set output in Example 3, stack the community node feature matrix according to the community index. ;

[0063] Step S32: For any given time period, construct an inter-community graph. , is represented as: ;where V com ={C1,..,C M} represents the set of community nodes, C1 and C2. M These are the 1st and Mth communities, where M is the total number of communities; if there is cross-community propagation interaction within time period k such that the propagation source node belongs to community C. i And the target node for propagation belongs to community C. j (i≠j), then in C i With C j Establish community boundaries and incorporate them ;

[0064] Step S33: For any community edge Edge features are extracted to characterize the driving factors of inter-community propagation. These edge features include: cross-community interaction frequency, outflow / inflow intensity, structural similarity (Jaccard similarity), and time decay factor. k,ij Based on the time difference of cross-community interactions The calculation is expressed as: ; This is the time decay hyperparameter, with a value representing the length of one time interval (t). k -t k-1 1 to 10 times; input the edge features into the edge feature encoder. Get edge embedding ; indicates as: Edge feature encoder It is a multilayer perceptron;

[0065] Step S34: Based on the inter-community graph, perform attention propagation updates on the community node features and edge embeddings: For any community node, calculate the attention weight by combining the representations of neighboring community nodes and the corresponding edge embeddings, and perform weighted aggregation on the neighbor representations to obtain the updated representation of the community node, thus obtaining the set of inter-community representations for time period k. This serves as the input for step S4.

[0066] By performing the above operations, this solution addresses the problem that general information dissemination popularity prediction methods have difficulty in characterizing inter-community dependencies and cross-community propagation drivers. It constructs an inter-community graph structure, extracts and encodes edge features of inter-community associations, and uses an attention update mechanism to propagate and update the community representations on the inter-community graph structure to obtain inter-community representations.

[0067] Example 5, see Figure 5 This embodiment is based on the above embodiment. In step S4, the prediction of information dissemination popularity specifically includes:

[0068] Step S41: For any time period, obtain the inter-community representation set output by Example 4. ,in Indicates community C within time period k. m Inter-community representation;

[0069] Step S42: Construct segmented concatenated representations; for any time period, construct inter-community representations for each community. to Multi-head attention aggregation is performed to generate a segmented cascaded representation z. k Set the number of attention heads to R, which is a positive integer and 1 ≤ R ≤ 16. Calculate the attention weight for each attention head for each community. The weights are normalized; the contribution of each community to the concatenated representation of the current time period is characterized by attention weights; and softmax normalization is used. ;in, The scoring values ​​are obtained from the inter-community representations through linear transformation and nonlinear mapping. Subsequently, the inter-community representations are weighted and aggregated according to their attention weights to obtain the aggregated results of each attention head. The results of each attention head are then concatenated to obtain the segmented concatenated representation z for time period k. k ;

[0070] Step S43: For any time period, calculate the piecewise growth statistical vector g. k g k This includes statistics on newly added communication interactions and cross-community interactions; g k Input Growth Statistical Encoder (MLP) grow Statistical embeddings and concatenated with segmented representations zk Concatenation forms the timing input x k , is represented as: Among them, MLP grow The growth statistical encoder is implemented using a multilayer perceptron. This is a concatenation operation; resulting in a time-series input set.

[0071] Step S44: Input the segmented concatenated temporal input set into the sequence encoder to model the temporal evolution within the observation window, obtaining the sequence hidden state set; select the BiGRU encoder for the sequence encoder; based on the final hidden state s K Output information dissemination popularity prediction results , is represented as: Where SeqEnc is a sequence encoder, and MLP is an MLP. out The growth predictive encoder, implemented using a multilayer perceptron, maps the final hidden state of the sequence encoder to the predicted increment of information propagation heat within the prediction window. To ensure the prediction results are non-negative, for Apply nonnegativity constraints and use the Softplus activation function as the last activation function of the prediction head.

[0072] By performing the above operations, this scheme addresses the problem of insufficient joint modeling of hierarchical information and its temporal evolution between communities and cascades in general information dissemination heat prediction methods. It uses multi-head attention aggregation to generate segmented cascaded representations between communities, combines segmented growth statistics to form time-series inputs and performs sequence encoding, and outputs the information dissemination heat prediction results within the future time window.

[0073] Example 6, based on the above examples, is used to determine the parameters of the prediction model. The information dissemination heat prediction method is implemented by a prediction model containing learnable parameters. Before performing heat prediction, the parameters of the prediction model are learned offline based on historical cascaded samples to obtain the model parameters. During the offline learning process, for any historical cascaded sample, an observation window [t] is taken. s ,t o Using the propagation data within [the specified area] as input, the predicted heat increment value is generated according to the processing flow of Examples 2 to 5. And the sample at prediction time t p The corresponding real popularity increment P(t) p As a supervisory signal, a loss function is constructed to measure the prediction error; mean squared error loss is adopted; based on the mean squared error loss, a gradient descent-type optimization algorithm is used to iteratively update the prediction model parameters until the stopping condition is met; after completing offline learning, the model parameters are used to calculate the cascaded samples to be predicted according to the process described in Examples 2 to 5, thereby outputting the prediction result of the information dissemination heat increment within the prediction window.

[0074] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0076] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting the popularity of information dissemination based on dynamic evolution within and outside a community, characterized in that: The method includes the following steps: Step S1: Obtain the cascaded graph, perform community detection on the cascaded graph to determine the community corresponding to each node, and divide the observation window into multiple consecutive time periods; Step S2: Construct the community intragraph structure, construct indication information based on newly added participating nodes within the time period and form masked adjacency relationships, perform community intragraph structure information propagation and update on the masked adjacency relationships and aggregate to obtain the community representation; Step S3: Construct an inter-community graph structure, extract and encode edge features of inter-community associations, and propagate and update the community representations on the inter-community graph structure based on an attention update mechanism to obtain the inter-community representations; Step S4: Information dissemination heat prediction. Perform multi-head attention aggregation on the inter-community representations. Each attention head calculates the attention weight for each community. The inter-community representations are weighted and aggregated according to the attention weights to obtain the aggregation results of each attention head. The results of each attention head are concatenated to obtain the segmented concatenated representation of the time period. Combine the segmented growth statistics to form the time series input and perform sequence encoding to output the information dissemination heat prediction results within the future time window.

2. The method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside the community, as described in claim 1, is characterized in that: In step S1, obtaining the cascade graph specifically includes: Step S11: Obtain the propagation interaction records corresponding to the target information; construct an information cascade graph within the observation window based on the propagation interaction records; Step S12: Perform community detection using the nodes and edges of the cascaded graph within the observation window; Step S13: Set the start and end times of the observation window, divide the observation window into K consecutive time periods, and obtain a set of time periods; Step S14: Construct the propagation edge set; Step S15: Define the cutoff time t k The cumulative set of participating nodes is defined, and the set of newly participating nodes in the k-th time period is defined; the set of newly participating nodes belonging to the community in the k-th time period is obtained based on the community set.

3. The method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside the community, as described in claim 2, is characterized in that: In step S1, constructing the propagation edge set involves distributing the propagation edge set within the observation window to each time period according to the timestamp, obtaining the propagation edge subset corresponding to each time period, and determining the set of nodes that appear within each time period. By combining the community set, for any time period and any community, a set of nodes belonging to the community within the time period is formed, and the set of propagation edges within the community within the time period is obtained.

4. The method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside the community, as described in claim 3, is characterized in that: In step S2, the construction of the community intragraph structure specifically includes: Step S21: Based on the time period set and community set, for any time period and community, obtain the set of nodes belonging to the community within the time period and the set of propagation edges within the community, and construct the subgraph within the community of the community within the time period; Step S22: For any subgraph within a community, determine the initial representation of each node in the node set; Step S23: For any time period and community, construct new participation indication information based on the set of newly added participating nodes and each node in the subgraph within the community; Step S24: Within the neighbor range defined by the mask adjacency relationship, the node representation is propagated and updated layer by layer; after the propagation update is completed, the updated representation of the node is obtained; Step S25: For any time period and community, use mean pooling to aggregate the update representations of all nodes in the community to obtain the community representation for the corresponding time period.

5. The method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside the community, as described in claim 4, is characterized in that: In step S3, the construction of the inter-community graph structure specifically includes: Step S31: For any time period, based on the community representation set, and stacked according to the community index, a community node feature matrix is ​​formed; Step S32: For any given time period, construct an inter-community graph; Step S33: For any inter-community edge, extract edge features to characterize the inter-community propagation driving factors; input the edge features into the edge feature encoder to obtain the edge embedding; Step S34: Based on the inter-community graph, perform attention propagation update on the community node features and edge embeddings to obtain the inter-community representation set for the time period.

6. The method for predicting the popularity of information dissemination based on the dynamic evolution inside and outside the community, as described in claim 5, is characterized in that: In step S4, the prediction of information dissemination popularity specifically includes: Step S41: For any time period, obtain the inter-community representation set; Step S42: Construct segmented concatenated representations; For any time period, perform multi-head attention aggregation on the inter-community representations of each community to generate segmented concatenated representations; Calculate attention weights for each community for each attention head and normalize the weights; Characterize the contribution of each community to the concatenated representation of the current time period based on the attention weights; Perform weighted aggregation on the inter-community representations according to the attention weights to obtain the aggregation results of each attention head, and concatenate the results of each attention head to obtain the segmented concatenated representation of the time period; Step S43: For any time period, calculate the segmented growth statistical vector, input the segmented growth statistical vector into the growth statistical encoder to obtain the statistical embedding, and concatenate it with the segmented concatenated representation to form the time series input, thus obtaining the time series input set; Step S44: Input the segmented concatenated temporal input set into the sequence encoder to model the temporal evolution within the observation window and obtain the sequence hidden state set; output the information propagation heat prediction result based on the final sequence hidden state.

Images