Social relationship representation method, information propagation prediction method, system, and storage medium

Abstract

The present application belongs to the field of computer and online social network, and particularly relates to a propagation network reconstruction and prediction method combined with multiple types of dynamic social relationship. A social relationship representation method is provided in the present application, which represents social relationship by using a multi-layer social network, and the multi-layer social network G includes a relationship topology network corresponding to multiple relationship types one by one; the relationship topology network includes network nodes corresponding to users one by one, and edges of the relationship topology network are used to represent relationship strength between users. The social relationship representation method provided in the present application uses a multi-layer graph neural network to describe various different types of social interaction features. The present application comprehensively considers multi-level, multi-type social relationship and information propagation characteristics, and is suitable for various different complex social networks.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer and online social network, in particular, relates to a social relationship representation method, an information propagation prediction method, a system and a storage medium. BACKGROUND

[0002] The rapid development of social networks has changed the way and scale of information propagation. With more and more users joining social media platforms and actively participating in the dissemination and sharing of information, social networks have become one of the main channels for information dissemination. The current information dissemination in the social network environment not only involves the forwarding of information by individual users, but also includes various interactive behaviors such as comments, likes, and forwarding, forming a complex information dissemination network. The characterization and prediction of this network are of great significance in information dissemination, influence dissemination, social recommendation, etc.

[0003] The reconstruction of the propagation network refers to restoring or reconstructing the structure of the original complete propagation network based on the observed information propagation process on the social platform. Existing propagation network construction and analysis techniques mainly include methods based on propagation models and methods based on machine learning. The methods based on propagation models mainly use diffusion models such as Independent Cascade Model (ICM) and Linear Threshold Model (LT) to simulate the propagation of information in the network. These methods assume that information propagation follows a specific propagation model and try to fit the model parameters to restore the network topology and propagation process from known propagation event data. The methods based on machine learning mainly use graph neural networks, time series analysis, etc. to construct models based on known information propagation processes and network structures, etc. to learn the rules of information propagation, thereby reconstructing the topology of the propagation network and restoring unknown or predicting future propagation processes.

[0004] The existing methods have the following shortcomings:

[0005] Firstly, the existing technology is largely limited by the need to obtain complete propagation process data, which requires a large amount of time and resources to collect, and may involve platform restrictions and user privacy issues. More importantly, sometimes, some information propagation may be implicit, for example, Facebook, Twitter platform cannot obtain the propagation process of a user forwarding other users' forwarding (i.e. the second or higher level propagation process of information), which means that it is impossible to obtain complete propagation process data when studying real-world events, making it difficult for most technologies to be applied in practical applications.

[0006] In addition, most of the prior art relies on the explicit relationship in the social network (especially the attention, fan relationship) to build and simulate the possible path of information dissemination. However, the social network in the real world is usually very complex, so the model based on a single relationship may not be able to capture the complex interaction between individuals and the dynamic evolution in the time dimension, and can usually only handle known, static relationship types, making it difficult to provide a dynamic and comprehensive evaluation of the information dissemination process. SUMMARY

[0007] In order to overcome the above-mentioned problems of the prior art which is limited to a single type of social network and requires complete information dissemination process data, the present application proposes a social relationship representation method which can represent a plurality of different types of social relationships, facilitating the reasoning and tracing of data in the information dissemination process.

[0008] The social relationship representation method proposed by the present application represents social relationships using a multi-layer social network, and the multi-layer social network G includes a relationship topology network corresponding to a plurality of relationship types one-to-one; the relationship topology network includes network nodes corresponding to users, and the edges of the relationship topology network are used to represent the relationship strength between users;

[0009] Definition representing the relationship topology network of the lth type of relationship at the time node t, and

[0010] respectively represent the node set and the edge set of M represents the total number of relationship types;

[0011] The node features in the relationship topology network include user metadata, action metadata, content information and topic labels; the user metadata is used to describe the basic information of the user, the action metadata is used to describe the action information of the user, the content information refers to the content of the article corresponding to the action metadata, and the topic label refers to the topic label corresponding to the content information; the node set is used to count the feature representation of each user in the corresponding type of relationship, and the feature representation of the user is obtained by encoding the user's original information through a pre-trained model.

[0012] Preferably, the feature representation of the user is obtained by encoding the original information of the user through a feature encoding model, and the original information includes user metadata, action metadata, content information and topic labels; the original information is labeled with an action type, and the action type includes original, comment, quote and forward;

[0013] ​The feature coding model comprises an action coding attention module and an information coding module corresponding to the action type, and further comprises an attention weighting module; the information coding module comprises a vector representation unit corresponding to the information type; each vector representation unit in the information coding module is configured to convert original information of a corresponding information type under a corresponding action type into a vector representation original embedding;

[0014] The action coding attention module constructs an action representation based on the original embedding of the original information of each information type under the corresponding action output by the corresponding information coding module; and the attention weighting module performs attention feature weighting calculation based on the action representation of each type of action to output a feature representation of the user.

[0015] The training method of the feature coding model comprises the following steps: constructing an original information sample and a decoding reconstruction module, allowing the feature coding model to code the original information sample to obtain a feature representation of the original information sample; allowing the decoding reconstruction module to restore the feature representation to obtain reconstructed information; calculating a loss based on the difference between the reconstructed information and the original information, updating the model parameters according to the loss function until the loss converges, and then outputting the feature coding model.

[0016] Preferably, the working process of the action coding attention module is represented by the following formula:

[0017] x=∑ i α i e i

[0018]

[0019] α′ i =q i tanh(U i ×e i +b i )

[0020] wherein e i represents the original embedding output by the i-th vector representation unit in the action coding attention module, α i is a weight normalized value of the i-th information type, and α′ i is the weight of the i-th information type.

[0021] U i , b i , and q i are parameters corresponding to the i-th information type in the action coding attention module; and x is the action representation output by the action coding attention module.

[0022] Let x n be the action representation output by the n-th action coding attention module, and the working process of the attention weighting module is represented by the following formula:

[0023]

[0024]

[0025]

[0026] wherein, is the weight normalization value of the nth action type under the lth relationship type, is the weight of the nth action type under the lth relationship type; tanh is the hyperbolic tangent function; is the model parameter corresponding to the nth action type in the weighted attention module of the feature encoding model under the lth relationship type; u l represents the feature representation of any user in the relationship topology network of the lth relationship type.

[0027] Preferably, in the feature encoding model training process, the parameters of the action encoding attention module, the attention weighting module and the decoding reconstruction module are updated according to the loss; the vector representation unit related to the text information adopts a pre-trained BERT model.

[0028] Preferably, the relationship types include fan relationship, interaction relationship and real social relationship.

[0029] The information propagation prediction method provided by the application can capture the dynamic evolution process of information propagation and provide strong prediction support for future relationship development, and realizes the reconstruction and prediction of propagation network by combining multiple types of dynamic social relationships.

[0030] The information propagation prediction method provided by the application comprises the following steps:

[0031] S1, obtaining a pre-trained feature encoding model and an information propagation prediction model; the information propagation prediction model is used to combine the multi-layer social network at time nodes t, t+1, t+2, …, t+T-1 to construct a multi-layer social network at time node t+T;

[0032] S2, obtaining the original information of each user at the T time nodes before the predicted time node, and encoding the original information through the feature encoding model to obtain a social relationship snapshot g(t, T) of the multi-layer social network at the T time nodes before the predicted time node;

[0033] S3, inputting the social relationship snapshot g(t, T) into the information propagation prediction model, and outputting the multi-layer social network G(t+T) at the predicted time node by the information propagation prediction model;

[0034]

[0035]

[0036] indicates the social relationship network corresponding to the lth relationship type at the time node t+T'; 0≤T'≤T, 1≤l≤M, and M represents the total number of relationship types.

[0037] Preferably, the information propagation prediction model comprises an encoder and a decoder; the encoder extracts a feature representation based on the social relationship snapshot g(t, T) at the continuous T time nodes, and the decoder constructs a multi-layer social network G(t+T) at the next time node based on the feature representation;

[0038] The encoder comprises a graph embedding module and a recurrent neural network; the input of the graph embedding module is the input of the encoder, the input of the recurrent neural network is the output of the graph embedding module, and the output of the recurrent neural network is the output of the encoder; and the decoder adopts a fully connected neural network.

[0039] The application further provides an information propagation prediction system and a storage medium for carrying the information propagation prediction method, so as to promote the application of the information propagation prediction method.

[0040] The information propagation prediction system provided by the application comprises:

[0041] A data acquisition module is configured to acquire original information of each user at a known time node;

[0042] A data encoding module is configured to store a feature encoding model and convert the original information sent by the data acquisition module into a feature representation of the user;

[0043] A relationship arrangement module is connected to the data encoding module and configured to aggregate the feature representations of each user at the same time node and construct a social relationship network under each type of relationship at the time node;

[0044] A network reconstruction module is connected to the relationship arrangement module; the network reconstruction module internally stores an information propagation prediction model, inputs the social relationship snapshot formed by the social relationship networks under each type of relationship at the multiple time nodes generated by the relationship arrangement module into the information propagation prediction model, and outputs a multi-layer social relationship network at the next time node of the social relationship snapshot by the information propagation prediction model;

[0045] A data processing module is configured to acquire a to-be-predicted time node and take the T time nodes before the to-be-predicted time node as reference time nodes; the data processing module is connected to the data encoding module, the relationship arrangement module and the network reconstruction module; the data processing module instructs the data encoding module and the relationship arrangement module to process the data at the reference time nodes, and instructs the network reconstruction module to construct a multi-layer social relationship network at the to-be-predicted time node based on the social relationship snapshot at the reference time nodes.

[0046] Another information propagation prediction system is provided by the present application, comprising a memory and a processor, the memory storing a computer program, a feature encoding model and an information propagation prediction model; the processor is connected with the memory, and the processor is used for the computer program to realize the information propagation prediction method.

[0047] The present application also provides a storage medium storing a computer program, a feature encoding model and an information propagation prediction model; the computer program is executed to realize the information propagation prediction method.

[0048] The present application has the advantages of:

[0049] (1) The social relationship representation method provided by the present application uses a multi-layer graph neural network to describe various different types of social interaction features. This feature enables the model to comprehensively consider multi-level and multi-type social relationships and information propagation features, and is suitable for a variety of different complex social networks.

[0050] (2) The multi-layer social network G in the present application calculates the features of various types of information on different level networks. In each layer network, information will be embedded through encoders for different action types, and the outputs of different encoders will be processed and integrated by introducing an attention mechanism, so as to determine the weights of various types of information on different levels, and generate a comprehensive representation of the user's participation in the multi-layer network and the content information.

[0051] (3) The information propagation prediction model provided by the present application integrates different types of information propagation methods into a unified model, more comprehensively analyzes and models the propagation process of information, avoids the need for complete propagation process data and explicit relationship type restrictions, and allows the model to analyze the propagation process of information at different levels and nodes in the absence of explicit propagation data. Through the information propagation prediction model, the future social relationship network can be predicted based on the multi-layer social relationship network at a known time node, and the multi-layer social relationship at a known time node can also be reconstructed based on the social relationship snapshot at an earlier time node, thereby recovering and mining the data loss of the multi-layer social relationship network at the known time node.

[0052] (4) The present application uses a recurrent neural network to process the dynamic development of relationships. The model can effectively integrate the time series information of multiple types of social relationships, making it more informative. This not only helps to improve the accuracy and prediction ability of the model, but also captures the dynamic evolution process of information propagation and provides strong prediction support for future relationship development.

[0053] (5) This invention considers multiple types of dynamic social relationships, such as fan relationships, interactive relationships, and real-world social relationships, to provide reconstruction and prediction of information dissemination networks. In fact, it is a method for reconstruction and prediction of dissemination networks that combines multiple types of dynamic social relationships. Attached Figure Description

[0054] Figure 1 A flowchart of a method for representing social relationships;

[0055] Figure 2 This is a structural diagram of an information dissemination prediction model. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Reference Figure 1 This embodiment proposes a method for representing social relationships, which uses a multi-layer social network to represent social relationships. The multi-layer social network G includes a relationship topology network that corresponds one-to-one with multiple relationship types.

[0058] Relationship types include: fan relationships, interactive relationships, and real-world social relationships.

[0059] A relational topology network consists of network nodes that correspond one-to-one with users, and the edges of the relational topology network are used to represent the strength of the relationships between users.

[0060] Define G (l) G represents the relational topology network of type l relations. (l) =(V (l) E (l) V (l) G represents (l) The set of nodes in E (l) G represents (l) The set of edges in;

[0061] G = (G (l) G (2) , ..., G (l) , ..., G (M) M represents the total number of relation types;

[0062] definition G represents the time node t. (l) , and They represent a node set and an edge set;

[0063] a relationship topology network G (l) The node features in the relationship topology network G include user metadata, action metadata, content information, and topic labels. The user metadata includes basic user information such as the number of fans, the number of followings, and the number of posts. The action metadata includes information such as the number of likes, comments, and reposts. The content information refers to the content of the article corresponding to the action metadata, and the topic label refers to the topic label corresponding to the content information.

[0064] a node set V (l) The feature representation of each user in the l-type relationship is used for statistics. The feature representation of the user is obtained by encoding the original information of the user through a feature encoding model.

[0065] The feature encoding model includes an action encoding attention module corresponding to the action type and an information encoding module, and further includes an attention weighting module. The information encoding module includes a vector representation unit corresponding to the information type. The vector representation unit related to the text can use a pre-trained BERT model.

[0066] The information type includes user metadata, action metadata, content information, and topic labels. The user metadata includes information such as the number of fans, the number of followings, and the number of posts. The action metadata includes information such as the number of likes, comments, and reposts. The content information refers to the content of the article corresponding to the action metadata, and the topic label refers to the topic label corresponding to the content information.

[0067] The action type includes original post, comment, quote, and repost. In this embodiment, the information encoding module corresponds to the action type one by one, that is, each information encoding module only encodes each type of information under the corresponding action. For example, the information encoding module corresponding to the original post action only obtains and encodes the user metadata, action metadata, article content, and topic label corresponding to the original post. In each information encoding module, each type of information is converted into a vector by the corresponding vector representation unit, thereby realizing information encoding.

[0068] The i-th vector representation unit is used to convert the original information of the i-th information type into a vector representation of the original embedding e i The original information is user metadata, action metadata, content information, or topic labels.

[0069] The input of each action encoding attention module is connected to the output of each vector representation unit in the corresponding information encoding module. The n-th action encoding attention module is used to combine the original embeddings e i of each type of original information output by the corresponding information encoding module to perform weighted calculation to obtain the action representation x n of the user under the n-th action.

[0070] The input of the attention weighting module is connected to the output of each action encoding attention module, and the attention weighting module outputs the feature representation u of the user n The weighted sum is performed to obtain the feature representation u of the user l and output.

[0071] The training method of the feature encoding model is as follows: constructing an original information sample and a decoding reconstruction module, allowing the feature encoding model to encode the original information sample to obtain the feature representation of the original information sample; allowing the decoding reconstruction module to restore the feature representation to obtain the reconstructed information; calculating the loss according to the difference between the reconstructed information and the original information, updating the parameters of the action encoding attention module and the attention weighting module in the feature encoding model and the decoding reconstruction module according to the loss function, and outputting the feature encoding model until the loss converges. It is worth noting that in the feature encoding model, each vector representation unit can directly use a pre-trained BERT model, or the parameters can be updated directly in the training process of the feature encoding model. That is, the vector representation unit can participate in training or not, and directly use the pre-trained model.

[0072] The feature encoding model corresponds to the relationship type one by one.

[0073] The node set V at a certain time node (l) The feature representation of any user is denoted as u l The feature encoding model can be represented by the following formula:

[0074]

[0075]

[0076]

[0077] wherein x n is the action representation output by the nth action encoding attention module, is the weight normalized value of the nth action type under the lth relationship, is the weight of the nth action type under the lth relationship; tanh is the hyperbolic tangent function; is the model parameter corresponding to the nth action type in the weighted attention module of the feature encoding model under the lth relationship.

[0078] The action encoding attention modules have the same structure but different model parameters. In this embodiment, the output of a certain action encoding attention module is denoted as x, and the formula is represented as:

[0079] x = ∑ i α i e i

[0080]

[0081] α' i = q i tanh(U i × e i + b i )

[0082] wherein the i-th vector in the action encoding attention module represents the original embedding e i outputted by the i-th vector representation unit, a i is the weight normalized value of the i-th information type, a' i is the weight of the i-th information type.

[0083] U i , b i , q i are the parameters corresponding to the i-th information type in the action encoding attention module.

[0084] The information propagation prediction model proposed in the embodiment is used to obtain a multi-layer social network at time node t+T in combination with the multi-layer social network at time nodes t, t+1, t+2, …, t+T-1; t, t+1, t+2, …, t+T-1, t+T are continuous T+1 time nodes.

[0085] Referring to Figure 2 , the information propagation prediction model comprises an encoder and a decoder; the input of the encoder is the input of the information propagation prediction model, the output of the encoder is the input of the decoder, and the output of the decoder is the output of the information propagation prediction model.

[0086] The encoder extracts a feature representation based on the social relationship snapshot g(t, T) at the continuous T time nodes, and the decoder constructs a multi-layer social network G(t+T) at the next time node based on the feature representation.

[0087] The encoder comprises a graph embedding module and a recurrent neural network.

[0088] The input of the graph embedding module is the input of the encoder. The input of the graph embedding module is the social relationship snapshot g(t, T) at the continuous T time nodes;

[0089]

[0090]

[0091] represents the social relationship network corresponding to the l-th relationship type at time node t+T'; 0≤T'≤T.

[0092] The input of the recurrent neural network is the output of the graph embedding module, and the output of the recurrent neural network is the output of the encoder, that is, the feature representation of the multi-layer relationship network; the recurrent neural network is used to learn the time characteristics of the social network and embed the time characteristics into the output features of the graph embedding module, thereby forming the feature representation of the multi-layer relationship network.

[0093] The decoder adopts a fully connected neural network, and the decoder predicts the distribution probability of each relationship network in the multi-layer relationship network at the next time node based on the feature representation extracted by the encoder, obtains the relationship network corresponding to the probability value of each relationship type as the predicted value of the relationship network of the relationship type, and obtains all relationship network predicted values as the predicted value of the multi-layer relationship network.

[0094] In this embodiment, the encoder processes the graph structure through the graph embedding module, learns the network time evolution characteristics through the recurrent neural network, and the decoder adopts a fully connected neural network layer for mapping the extracted features back to the original space. Through the encoder-decoder structure, the model can locate the nodes or edges that have the greatest impact on the propagation process in the multi-layer network and reconstruct the information propagation network based on this. Specifically, the encoder structure allows the model to encode the original multi-layer social network data (i.e., social relationship snapshots) into a latent representation of the information propagation network, capturing the features of each node and the complex relationships and potential connections between nodes; then, the decoder decodes the latent representation of the social relationship network back to the reconstructed network, and in this process, the model will determine the nodes and edges that have the most significant impact on the propagation process in each time step, that is, the most likely information propagation sources and propagation paths.

[0095] In the training process of the information propagation prediction model, first, a learning sample {g(t, T), G(t+T)} is constructed based on historical data, and then the information propagation prediction model is subjected to machine learning. In the learning process, the mean square error is used to update the encoder and the decoder, and the learning process is terminated until the information propagation prediction model converges.

[0096] Of course, for those skilled in the art, the present application is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be considered as limiting the claims involved.

[0097] In addition, it should be understood that, although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand.

[0098] The technologies, shapes, and structural parts not described in detail in the present application are well-known technologies.

Claims

1. A method of representing social relationships, characterized by, The social relationship is represented by a multi-layer social network, and the multi-layer social network G includes a plurality of relationship topology networks corresponding to a plurality of relationship types; the relationship topology network includes network nodes corresponding to users, and edges of the relationship topology network are used to represent relationship strength between the users; Definitions a relation topology network representing the l-th type of relation at time node t, and respectively represent a node set and an edge set of 1≤l≤M, M represents the total number of relation types; The node features in the relationship topology network include user metadata, action metadata, content information, and topic labels; the user metadata is used to describe basic information of the user, the action metadata is used to describe action information of the user, the content information refers to content of an article corresponding to the action metadata, and the topic label refers to a topic label corresponding to the content information; the node set is used to count feature representations of the users in the corresponding type of relationship, and the feature representation of the user is obtained by encoding the user through a pre-trained model; The feature representation of the user is obtained by encoding the original information of the user through a feature encoding model, and the original information includes user metadata, action metadata, content information, and topic labels; the original information is labeled with an action type, and the action type includes original, comment, quote, and forward; The feature encoding model includes an action encoding attention module and an information encoding module corresponding to the action type, and further includes an attention weighting module; the information encoding module includes a vector representation unit corresponding to an information type; each vector representation unit in the information encoding module is used to convert the original information of the corresponding information type under the corresponding action type into a vector representation original embedding; The action encoding attention module constructs an action representation based on the original embedding of the original information of each information type under the corresponding action output by the corresponding information encoding module; and the attention weighting module performs attention feature weighting calculation based on the action representation of each type of action to output the feature representation of the user; The training method of the feature encoding model is as follows: constructing an original information sample and a decoding reconstruction module, allowing the feature encoding model to encode the original information sample to obtain the feature representation of the original information sample; allowing the decoding reconstruction module to restore the feature representation to obtain reconstructed information; combining the difference between the reconstructed information and the original information calculating a loss, updating the model parameters according to the loss function until the loss converges, and then outputting the feature encoding model.

2. The social relationship representation method of claim 1, wherein, The working process of the action encoding attention module is represented by the following formula: α' i = q i tanh(U i × e i + b i ) wherein e i represents the original embedding output by the i-th vector representation unit in the action encoding attention module, a i is the weight normalization value of the i-th information type, a′ i is the weight of the i-th information type; U i , b i , q i are parameters corresponding to the i-th information type in the action encoding attention module; x is the action representation of the output of the action encoding attention module; Let x n The action representation output by the attention module for the nth action is encoded, and the working process of the attention weighting module is represented by the following formula: wherein, is a weight normalization value of the nth action type under the lth relationship type, is a weight of the nth action type under the lth relationship type; tanh is a hyperbolic tangent function; is a model parameter corresponding to the nth action type in the weighted attention module of the feature encoding model under the lth relationship type; u l represents the feature representation of any user in the relationship topology network of the lth relationship type.

3. The social relationship representation method of claim 1, wherein, During the training process of the feature encoding model, the parameters of the action encoding attention module, the attention weighting module, and the decoding reconstruction module are updated according to the loss; the vector representation units related to the text information all use a pre-trained BERT model.

4. The social relationship representation method of claim 1, wherein, The relationship types include a fan relationship, an interaction relationship, and a real social relationship.

5. An information propagation prediction method employing the social relationship representation method according to any one of claims 1 to 4, characterized by, The method comprises the following steps: S1, obtaining a pre-trained feature encoding model and an information propagation prediction model; the information propagation prediction model is used to construct a multi-layer social network at a time node t+T based on the multi-layer social networks at the time nodes t, t+1, t+2, …, t+T-1; S2, obtaining original information of each user at the T time nodes before the predicted time node, and encoding the original information through the feature encoding model to obtain a social relationship snapshot g(t, T) of the multi-layer social network at the T time nodes before the predicted time node; S3, inputting the social relation snapshot g(t, T) into an information propagation prediction model, and outputting a multi-layer social network G(t+T) at a time node to be predicted by the information propagation prediction model; represents the social relation network corresponding to the lth relation type at time node t+T'; 0≤T'≤T, 1≤l≤M, and M represents the total number of relation types.

6. The information propagation prediction method of claim 5, wherein, The information propagation prediction model comprises an encoder and a decoder; the encoder extracts a feature representation based on the social relation snapshot g(t, T) at the continuous T time nodes, and the decoder constructs the multi-layer social network G(t+T) at the next time node based on the feature representation; The encoder comprises a graph embedding module and a recurrent neural network; the input of the graph embedding module is the input of the encoder, the input of the recurrent neural network is the output of the graph embedding module, and the output of the recurrent neural network is the output of the encoder; and the decoder adopts a fully connected neural network.

7. An information propagation prediction system that implements the information propagation prediction method according to claim 5 or 6, characterized by, Comprise: A data acquisition module configured to acquire original information of each user at a known time node; A data encoding module configured to store a feature encoding model and convert the original information sent by the data acquisition module into a feature representation of the user; A relation arrangement module connected with the data encoding module and configured to aggregate the feature representations of the users at the same time node to construct a social relation network under each type of relation at the time node; A network reconstruction module connected with the relation arrangement module; the network reconstruction module internally stores an information propagation prediction model, and inputs a social relation snapshot formed by the social relation networks under each type of relation at the multiple time nodes generated by the relation arrangement module into the information propagation prediction model, and the information propagation prediction model outputs a multi-layer social relation network at the next time node of the social relation snapshot; A data processing module configured to acquire a time node to be predicted and take T time nodes before the time node to be predicted as reference time nodes; the data processing module is connected with the data encoding module, the relation arrangement module and the network reconstruction module; the data processing module instructs the data encoding module and the relation arrangement module to process the data at the reference time nodes, and instructs the network reconstruction module to construct the multi-layer social relation network at the time node to be predicted based on the social relation snapshot at the reference time nodes.

8. An information propagation prediction system characterized by, Comprise a memory and a processor, the memory stores a computer program, a feature encoding model and an information propagation prediction model; the processor is connected with the memory, and the processor is used for the computer program to realize the information propagation prediction method of claim 5.

9. A storage medium, characterized by Store a computer program, a feature encoding model and an information propagation prediction model; the computer program is executed to realize the information propagation prediction method of claim 5.

Images