Information diffusion prediction method and device based on dynamic interaction relation perception

Abstract

The invention discloses an information diffusion prediction method and device based on dynamic interaction relation perception, and belongs to the information diffusion prediction direction in the technical field of artificial intelligence. The method is used for solving the problem of poor prediction performance caused by difficulty in accurately representing a dynamic relationship and potential preference of a user in an information propagation environment in an existing method. The dynamic interaction relationship and potential preference change of the user in a global information propagation environment can be effectively perceived and represented by utilizing the potential characteristics of the user in the social relation graph and the information diffusion graph under the continuous time of modeling by using a graph neural differential equation; the problems that an existing prediction method based on graph representation is unreasonable in time granularity selection, and it is difficult to naturally capture user preferences and continuous evolution of behaviors along with time are solved, and the accuracy of information diffusion prediction is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence, and particularly relates to an information diffusion prediction method and device based on dynamic interaction relationship perception. BACKGROUND

[0002] The rise of social media and wireless technology has completely changed the pattern of online information sharing. Unlike traditional communication methods, digital platforms give users the ability to quickly spread content, triggering a ripple effect often referred to as "information cascades." These information cascades describe situations where a large number of individuals make similar decisions in succession, which can influence public discourse, set trends, and even affect real-world behavior. Therefore, how to effectively perceive the situation of the information cascade process and then carry out situation prediction and risk warning has become an important issue in network space governance. In actual research and application, the focus of information diffusion prediction is to perceive and predict future diffusion dynamics based on observable social interaction records, especially to predict information propagation paths at the user level and identify the next user that may be affected in a timely manner, which is of great significance for information recommendation, information interception, risk assessment and other applications.

[0003] Before the development of deep learning technology, the research on information diffusion prediction methods mainly focused on generative diffusion models and feature engineering. Methods based on feature engineering use carefully designed features as input and use matrix decomposition, tensor decomposition, and other techniques to obtain cascade prediction results. Methods based on generative diffusion models follow pre-set propagation rules to let information spread from active users to inactive users through social relationships, thereby simulating the real propagation process. For example, some studies use the heat diffusion equation to model information diffusion and embed users in a low-dimensional space; other studies use independent cascade models to embed users in a hidden space in social networks to obtain more reliable diffusion probabilities. However, these methods usually rely on strong assumptions and fixed parameters, limiting their generalization ability and expression ability. In addition, simplifying the information diffusion mechanism often reduces prediction accuracy and limits the ability to model real dynamics.

[0004] In recent years, deep learning-based methods have become the mainstream in the field of information diffusion prediction due to their strong representation capabilities. Existing mainstream methods can be roughly divided into sequence-based representation methods and graph-based representation methods. Sequence-based representation methods mainly use models such as Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), and Transformer, and combine the unique features of information cascades to improve the representation capabilities of cascade systems. In contrast, graph-based representation methods focus on learning high-quality user representations from graph structures. By constructing various types of graph structures such as social graphs and diffusion graphs, different role attributes of users can be effectively captured. Some methods further divide the cascade into different time intervals, establish multiple snapshots for the dynamic graph or hypergraph of each interval, and capture the dynamic preferences and spatiotemporal characteristics of users through a fusion mechanism.

[0005] However, there are two main defects in existing information diffusion prediction methods based on graph representation: 1) Time discretization bias: Traditional sequence graphs or hypergraph methods need to divide time into discrete intervals, while real-world cascades occur in continuous time. Therefore, this method of discretizing time intervals will lead to unreasonable selection of time granularity, affecting prediction accuracy. 2) Lack of continuity in user node representation: Sequence graphs and hypergraph methods usually construct independent snapshots for each time interval, and the node representations of different snapshots are spliced through a gating fusion module, but this "snapshot splicing" method cannot model the feature transition process between adjacent snapshots, resulting in "jumping" of user preference evolution, which cannot naturally reflect the continuous change of user behavior over time in real propagation. In summary, existing methods cannot accurately represent the dynamic relationships and potential preferences of users in the information propagation environment, resulting in poor prediction performance. SUMMARY

[0006] To solve the problem that existing methods cannot accurately represent the dynamic relationships and potential preferences of users in the information propagation environment, thereby leading to poor prediction performance, embodiments of the present application provide a method and device for predicting information diffusion based on dynamic interaction relationship perception.

[0007] Embodiments of the present application provide a method for predicting information diffusion based on dynamic interaction relationship perception, comprising:

[0008] Obtain a cascade data set and a static social relationship graph, construct a static user-cascade bipartite graph, a static information diffusion graph, a static information diffusion hypergraph, and a new heterogeneous graph;

[0009] An initial node representation of the new heterogeneous graph is obtained through a graph convolution network (GCN), a graph neural differential equation of evolution of a new heterogeneous node over time is constructed, and an edge weight included in a new heterogeneous dynamic graph and a node representation of the new heterogeneous graph at different time points are determined;

[0010] A plurality of intervals and key time points are obtained by normalizing and equally dividing the participation time, a node representation of the new heterogeneous graph corresponding to each key time point is determined based on the node representation of the new heterogeneous graph, and a final dynamic node representation of a first user is determined based on the intervals and the first time.

[0011] A comprehensive node representation of the user is obtained by fusing the final dynamic node representation of the user and the static node representation of the user through an attention gate fusion mechanism, a final feature of the cascade is obtained by concentrated fusion of the comprehensive node representation of the user through multi-head attention, and the final feature of the cascade is used to predict a possibility of future participation of each user in propagation through a multi-layer perception.

[0012] An information diffusion prediction device based on dynamic interaction relationship perception is provided in an embodiment of the application, and the device comprises:

[0013] A construction unit is configured to acquire a cascade data set and a static social relationship graph, and construct a static user-cascade bipartite graph, a static information diffusion graph, a static information diffusion hypergraph, and a new heterogeneous graph.

[0014] A first determination unit is configured to obtain an initial node representation of the new heterogeneous graph through a graph convolution network (GCN), construct a graph neural differential equation of evolution of a new heterogeneous node over time, determine an edge weight included in a new heterogeneous dynamic graph, and determine a node representation of the new heterogeneous graph at different time points.

[0015] A second determination unit is configured to normalize and equally divide the participation time to obtain a plurality of intervals and key time points, determine a node representation of the new heterogeneous graph corresponding to each key time point based on the node representation of the new heterogeneous graph, normalize a first participation time included in a first cascade to obtain a first time, and determine a final dynamic node representation of a first user based on the intervals and the first time.

[0016] A prediction unit is configured to obtain a comprehensive node representation of the user by fusing the final dynamic node representation of the user and the static node representation of the user through an attention gate fusion mechanism, obtain a final feature of the cascade by concentrated fusion of the comprehensive node representation of the user through multi-head attention, and predict a possibility of future participation of each user in propagation through a multi-layer perception.

[0017] The embodiment of the present application provides a computer device, the computer device comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute the information diffusion prediction method based on dynamic interaction relationship perception.

[0018] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to enable the processor to execute the information diffusion prediction method based on dynamic interaction relationship perception.

[0019] The embodiment of the present application provides an information diffusion prediction method and device based on dynamic interaction relationship perception, which is from the perspective of the model, and based on the graph neural differential equation to represent the dynamic interaction relationship between users and the potential preference of the users in continuous time, so as to improve the accuracy of diffusion prediction and the perception ability of risk situation. First, real information propagation event data on a social network and social relationship networks of users are collected, and a user-cascade bipartite graph, an information diffusion relationship graph and an information diffusion hypergraph are constructed. Second, the node representation of the user and the information cascade is initialized by using a gating mechanism, the node state of the user in the information propagation process is represented by using the graph neural differential equation to realize dynamic interaction and global situation awareness of information diffusion, the dynamic node representation of the user at a specific time is retrieved by combining the time information of the user in the information cascade, and real-time perception and analysis of the diffusion situation are realized. Then, the static node representation of the user is obtained by using a graph convolution network and a hypergraph convolution network, and the dynamic and static node representations are fused by using an attention gating fusion mechanism. Finally, the features of the information cascade are obtained by using a multi-head attention mechanism in combination with the user node representation and the time information, and the possibility of future propagation of the user is predicted by using a multi-layer perception. In the training stage, a cross-entropy loss function is used to optimize the model parameters, and in the test stage, the trained model is used to predict the users who may participate in the event. The method provided in the embodiment of the present application models the potential features of the user in the social relationship graph and the information diffusion graph in continuous time by using the graph neural differential equation, can effectively perceive and represent the dynamic interaction relationship and the change of potential preference of the user in the global information propagation environment, solves the problems that the prediction method based on graph representation in the prior art is unreasonable in time granularity selection and is difficult to naturally capture the continuous evolution of user preference and behavior with time, and effectively improves the accuracy of information diffusion prediction. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any creative labor on the basis of these drawings.

[0021] Figure 1 A flow chart of an information diffusion prediction method based on dynamic interaction relationship perception is provided for an embodiment of the present application.

[0022] Figure 2 A diagram of retrieving user final dynamic node representation using user time in cascade is provided for an embodiment of the present application.

[0023] Figure 3A A diagram of relationship between training time and performance index of Android dataset is provided for an embodiment of the present application.

[0024] Figure 3B A diagram of relationship between training time and performance index of Douban dataset is provided for an embodiment of the present application.

[0025] Figure 4 A structural diagram of an information diffusion prediction device based on dynamic interaction relationship perception is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0026] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made. These all belong to the protection scope of the present application.

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the present application.

[0028] The dynamic graph structure is essentially a "time-dependent graph structure", that is, the elements (nodes, edges, edge weights) of the graph will be updated according to the change of time or the superscript with time parameter , explicitly expressed in the formula, in contrast to "static graph structure (such as static social relationship graph , initial bipartite graph " — the nodes and edges of the static graph are fixed for a long time and do not change with time.

[0029] Figure 1 A flow chart of an information diffusion prediction method based on dynamic interaction relationship perception is provided for an embodiment of the present application, as shown in Figure 1As shown, the method includes the following steps:

[0030] Step 101: Obtain the cascaded data set and static social relationship graph, and construct the static user-cascaded bipartite graph, static information diffusion graph, static information diffusion hypergraph, and new heterogeneous graph;

[0031] Step 102: Obtain the initial node representation of the new heterogeneous graph through GCN (Graph Convolutional Network), construct the dynamic edge-coupled graph neural differential equation of the new heterogeneous nodes evolving over time, and determine the edge weights and node representations of the new heterogeneous dynamic graph at different times.

[0032] Step 103: Normalize the participation time and divide it equally to obtain multiple intervals and key time points. Determine the dynamic graph node representation corresponding to each key time point based on the node representation of the new heterogeneous graph. Normalize the first participation time included in the first concatenation to obtain the first time. Determine the final dynamic node representation of the first user based on the intervals and the first time.

[0033] Step 104: The user's final dynamic node representation and the user's static node representation are fused through an attention gating fusion mechanism to obtain the user's comprehensive node representation; the user's comprehensive node representation is fused through multi-head attention to obtain the cascaded final features; the cascaded final features are used to predict the probability of each user participating in propagation in the future through a multilayer perceptron.

[0034] In step 101, it is necessary to collect cascaded data sets and static social relationship graphs on social networks.

[0035] Specifically, cascaded data sets include Each cascade records the complete trajectory of information propagation, including users, participation time, and propagation order; that is, each cascade is a combined record of "users, participation time, and propagation order," fully reflecting the user participation sequence and timeline from the starting point of information propagation to its subsequent spread. In this embodiment of the invention, a cascade can be represented as:

[0036] (1)

[0037] in, Indicates the first One user who participated in this cascading propagation, Indicates the order of propagation. This indicates the total number of users participating in the propagation within the cascade (propagation length). Indicates the first The original participation time of the k-th user in each cascade is the discrete user participation time. denotes a user set including the number of all users, , denotes a single cascade. It should be noted that the subsequent "participation time" can refer to the original participation time or the participation time after processing such as normalization, which needs to be judged in combination with the operation scene.

[0038] In the embodiments of the present application, the cascade is the abbreviation of "information cascade", which refers to the complete trajectory of information propagation in a social network, including users, participation time and propagation order; the user is the abbreviation of "user participating in information cascade propagation", that is, an individual participating in the process of information cascade propagation.

[0039] Further, the static social relationship graph is a structured representation of the social association between users, specifically including a user set and a social relationship edge, specifically, the static social relationship graph is represented as:

[0040] (2)

[0041] Wherein, denotes a static social relationship graph, denotes a user set including the number of all users, denotes a social relationship edge set, denotes a user and a user have a social relationship, such as a friend, a follow, etc.

[0042] It should be noted that the static social relationship graph represents , which is an abstract representation of the entire network, wherein represents the overall structure of "static social relationship graph", and the superscript s is used to identify the "social" attribute, which is distinguished from the subsequent static information diffusion graph , static user-cascade bipartite graph , etc. (the superscript d represents diffusion, and the superscript b represents bipartite).

[0043] In actual applications, the data included in the cascade data set and the static social relationship graph is discrete, which cannot be directly used for feature learning of GCN and other models. In order to solve the above problem, the embodiments of the present application propose to construct a static user-cascade bipartite graph, a static information diffusion graph, and a static information diffusion hypergraph based on the above data, so that the unstructured data can be converted into a structured representation of "node-edge", capturing the association patterns between users and cascades, and users and users.

[0044] Specifically, the static user-cascade bipartite graph is constructed based on the cascade data set, specifically, for each cascade , all users In all users With each cascade Establish an edge between them to form a static user-cascaded bipartite graph, as shown below:

[0045] (3)

[0046] in, This represents a static user-cascaded bipartite graph. This represents the set of nodes in a static user-cascaded bipartite graph, which includes two types of nodes: one type is the IDs of all users, such as... Another type is the numbering of all cascades, such as... ; Represents the user-cascaded binary edge set. Indicates user Participated in the Cascade The process of its spread.

[0047] Specifically, a static information diffusion graph is constructed based on the cascaded data set. Specifically, the propagation sequence of each cascade (e.g., ...) is traversed. ), for adjacent user pairs in the propagation sequence An edge is then established between the two to form a static information diffusion graph, as shown below:

[0048] (4)

[0049] in, This represents a static information diffusion diagram. Represents a set of users. Represents the set of edges in a static information diffusion graph. Indicates user and users At least in one cascade, there exists a relationship of "adjacent propagation order" (i.e., After the spread, Then spread, or After the spread, Then it propagates (i.e., bidirectional adjacency). It should be noted that if the same pair of users are adjacent in multiple cascades, only one edge needs to be retained.

[0050] Furthermore, a static information diffusion hypergraph is constructed based on the cascaded data set. Specifically, each cascaded... Corresponding to a hyperedge This will include all users in the cascade (i.e. Included Both are related to the superedge The connections form a static information diffusion hypergraph, i.e., a "single hyperedge - multiple users" association, as shown below:

[0051] (5)

[0052] wherein, represents a static information diffusion hypergraph, represents a user set, represents an information diffusion hyperedge, represents a piece of cascaded information , a hyperedge is a cascade In the static information diffusion hypergraph, each cascade corresponds to a hyperedge in the static information diffusion hypergraph The node set of the hyperedge is all users participating in the cascade.

[0053] It should be noted that the new heterogeneous graph provided by the embodiment of the application includes a dynamic new heterogeneous diffusion graph, a dynamic new heterogeneous social graph and a static new heterogeneous relationship graph.

[0054] Further, based on the static user-cascade bipartite graph, the static information diffusion graph and the static information diffusion hypergraph, a dynamic new heterogeneous diffusion graph, a dynamic new heterogeneous social graph and a static new heterogeneous relationship graph can be constructed, and specific embodiments are as follows:

[0055] (6)

[0056] (7)

[0057] (8)

[0058] wherein, represents a dynamic new heterogeneous diffusion graph, represents a static user-cascade bipartite graph, represents a static information diffusion graph, represents a dynamic new heterogeneous social graph, represents a static social relationship graph, represents a static new heterogeneous relationship graph. It should be noted that the new heterogeneous graph in the embodiment of the application includes a dynamic new heterogeneous diffusion graph, a dynamic new heterogeneous social graph and a static new heterogeneous relationship graph.

[0059] Further, the user and the cascade are initialized by using a gating mechanism. The gating mechanism is a function for screening and optimizing the original features of the nodes, and the core function is to selectively highlight key features and suppress secondary features from the original node representation containing the user and the cascade.

[0060] The gating mechanism processes the original representation containing all users, and obtains the node representation set focusing on static social attribute and the node initial representation set focusing on dynamic preference through two different gating functions, respectively. Specifically, the gating mechanism weights the original features through the gating function output) by a "filtering factor", and the formula is:

[0061] (9)

[0062] (10)

[0063] (11)

[0064] wherein, represents the node representation set focusing on a specific attribute after the gating function processing, specifically, the specific attribute obtained after the gating filtering of the initial node representation set , wherein, represents different attributes, such as static social attribute s and dynamic preference attribute o; represents the general abstract representation of all gating functions, represents the trainable weight matrix of the gating mechanism, represents the trainable bias parameter of the gating mechanism, represents the matrix Hadamard product operation, represents the normalization exponential function, represents the node representation set focusing on static social attribute, represents the node representation set focusing on dynamic preference, represents the initial node representation set containing all users and cascades.

[0065] In practical applications, the core role of the gating mechanism is to generate a "gating mask" with a value range of 0-1 through the sigmoid normalization exponential function, and then perform Hadamard product with the initial representation to realize the filtering of different attribute features in , that is, to retain important features and suppress irrelevant features.

[0066] Further, the node representation set focusing on static social attribute is obtained through the gating function for filtering static dynamic preference features , which focuses on features related to social behavior, such as user social activity and user interest relevance. The node initial representation set focusing on dynamic preference is obtained through the gating function for filtering dynamic interaction attribute features ​Get the features related to the propagation behavior, such as the user's propagation frequency and the propagation speed of the cascade.

[0067] In step 102, the dynamic new heterogeneous diffusion graph is merged from the static user-cascade bipartite graph and the static information diffusion graph, that is:

[0068] (6)

[0069] wherein, represents the dynamic new heterogeneous diffusion graph, represents the static user-cascade bipartite graph, represents the static information diffusion graph, and in practical applications, provides the "user-cascade" user-cascade bipartite edge which reflects the participation relationship between users and cascades, provides the "user-user" static information diffusion graph edge which reflects the direct propagation association between users. The merged dynamic new heterogeneous diffusion graph has a node set of users and cascades, and an edge set of , which contains two types of relationships.

[0070] Further, the dynamic preference node initial representation set is filtered to obtain features related to the diffusion relationship using the gating function for the dynamic diffusion relationship; then the GCN aggregates the neighbor features of each stage on the dynamic new heterogeneous diffusion graph to obtain the node representation of the initial moment of the GCN as the starting state of the subsequent dynamic evolution. It should be noted that the dynamic new heterogeneous diffusion graph uses a dynamic edge coupled graph neural differential equation, and the dynamic new heterogeneous social graph uses a highly scalable graph neural differential equation.

[0071] (12-1)

[0072] wherein, represents the node representation set of the initial moment of the dynamic new heterogeneous diffusion graph, represents the GCN operation, represents the gating function for the dynamic diffusion relationship, which is used to filter the diffusion-related features, represents the dynamic preference node initial representation set, represents the dynamic new heterogeneous diffusion graph.

[0073] Further, the update rule of the GCN is as follows:

[0074] (13)

[0075] wherein, represents a node updated by GCN, which integrates the features of the node itself and its neighbor nodes, represents a nonlinear activation function, which is used to perform nonlinear transformation on the aggregated features to enhance the expression ability of the model, represents a graph structure , wherein, represents a node in the graph structure , wherein, represents the weight of an edge in any graph, which is used to adjust the influence degree of the neighbor node on , and is 1 by default, represents the representation of the neighbor node , represents the representation of the node , represents the sum of the edge weights corresponding to all neighbor nodes of the node , which is used as a normalization factor to avoid the inconsistency of feature scales caused by the difference in the number of neighbors, represents a general trainable weight matrix of GCN, which is used to perform linear transformation on the representation of the neighbor node, represents the original representation of the neighbor node

[0076] It should be noted that the graph structure represents any graph structure, which can represent any one of a dynamic new heterogeneous diffusion graph , a dynamic new heterogeneous social graph , and a static new heterogeneous relationship graph , and provides a topological basis for the neighbor relationship of the nodes and . represents the weight of an edge in a network structure , which is 1 by default.

[0077] Further, a graph neural differential equation of a dynamic edge-coupled graph is constructed based on the evolution of the nodes of the dynamic new heterogeneous diffusion graph over time, and the edge weights included in the nodes of the dynamic new heterogeneous diffusion graph and the node representations of the dynamic new heterogeneous diffusion graph at different time points are determined.

[0078] Specifically, the graph neural differential equation of the dynamic edge-coupled graph is a “rule definition” process of dynamic evolution, which is used to describe the change rule of the nodes of the dynamic new heterogeneous diffusion graph over time. The core of the equation is to define rate of change, express The set of node representations of the dynamic new heterogeneous diffusion graph at any given time is driven by two parts, specifically:

[0079] Structures that partially depend on dynamic novel heterogeneous diffusion maps Specifically, the node feature update amount under the current dynamic new heterogeneous diffusion graph can be calculated through GCN. The difference between this update amount and the current feature is the driver of this change, which reflects the influence of neighbor relationships on the feature.

[0080] Another part depends on the node features themselves. Specifically, the update amount of the feature itself is obtained through nonlinear transformation of the feature, and the difference between this update and the current feature is the driving force of this change. The entire dynamically edge-coupled graph neural differential equation combines the dynamic changes of the dynamic new heterogeneous diffusion graph with the evolution of the feature itself, forming a complete time evolution rule.

[0081] In this embodiment of the invention, the graph neural differential equation with dynamic edge coupling can be expressed by the following formula:

[0082] (14)

[0083] (14-1)

[0084] (14-2)

[0085] in, express Time node representation set Time derivative, The weight parameters representing the balance between the two evolution functions control the relative importance of the influence of the graph structure and the influence of the node's own characteristics. This represents the evolution function associated with the dynamic graph structure, and outputs the influence of the graph structure on node evolution. This represents the evolution function related to the node's own characteristics, and outputs the influence of the node's own attributes on the evolution. express The set of node representations for a dynamic, new heterogeneous diffusion graph at any given time. express Dynamic new heterogeneous diffusion diagram at time t, This represents the evolution function associated with the dynamic graph structure. This represents the evolution function related to the node's own characteristics. Indicates based on Time node representation and animated GIFs The graph convolution operation outputs a node representation that incorporates neighbor information. express The set of node representations for a dynamic, new heterogeneous diffusion graph at any given time. express Dynamic new heterogeneous diffusion diagram at time t, This represents the trainable weight matrix in the differential equation of the dynamic novel heterogeneous diffusion graph, used to characterize the internal correlations of node features. This represents the bias vector in the differential equation of the dynamic new heterogeneous diffusion graph, used to adjust the offset of the internal correlation of node features, and can be used to train the weight matrix. Synergistic effect on feature transformation, express The transpose of .

[0086] It should be noted that, The instantaneous evolution rate of node features is characterized by a larger value, indicating that the node features change more drastically at that moment. It is the core indicator for describing the continuous evolution of node representation. Used to adjust the relative importance of "the influence of graph structure on evolution" and "the influence of node characteristics on evolution". The larger the value, the stronger the constraint that the graph structure places on the evolution of nodes.

[0087] Furthermore, regarding the dynamic novel heterogeneous diffusion map For each edge in the array (connecting users or users to cascades), calculate its position in the array. Time weight First, determine the first appearance time of the nodes at both ends of the edge: the default first appearance time of a user node is 0 (always active), and the first appearance time of a cascade node is the participation time of its first user; only when the first appearance times of both nodes are no earlier than 0... Time (i.e.) (Only when all nodes have participated in the propagation) are edges considered "active," at which point the weight is determined by the weight of the two nodes. The features at each time point are concatenated and then calculated using a function; otherwise, the weight is 0. In this way, The dynamic new heterogeneous diffusion map at any given time can dynamically filter out The propagation relationship that is effective at all times is the structural dependency term in the evolution equation. Provides accurate topology.

[0088] Specifically, any edge in the dynamic novel heterogeneous diffusion graph The corresponding weight Determined by the following formula:

[0089] (15-1)

[0090] in, express the weight of an edge in the dynamic new-heterogeneous diffusion graph at time , denotes a normalized exponential function, whose formula is , which maps the output to the interval [0, 1] and represents the intensity of the edge weight, denotes the dynamic diffusion edge weight calculation function, whose input is the concatenation result of the node representations, and whose output is the intermediate value of the edge weight, denotes the operation of concatenating the node and the node at time , and , denotes the representation of the neighbor node in the dynamic new-heterogeneous diffusion graph at time , denotes the representation of the node in the dynamic new-heterogeneous diffusion graph at time , denotes the time at which the node first appears in the entire information propagation environment, if is a user, then , if is a cascade, then denotes the time at which the first user in the cascade participates in the propagation, denotes the filtering of node pairs that have appeared before time and assigning weights only to edges of such node pairs, otherwise the edge weight is 0.

[0091] Further, using the ordinary differential equation numerical solver (ODESolver), taking the node representations of the dynamic new-heterogeneous diffusion graph at the initial time as the starting point, and substituting into the graph neural differential equation defined above, the node representations of the dynamic new-heterogeneous diffusion graph at different times are obtained, as shown below:

[0092] (16-1)

[0093] wherein, denotes the set of node representations of the dynamic new-heterogeneous diffusion graph at time , denotes the ordinary differential equation numerical solver tool, denotes the set of node representations of the dynamic new-heterogeneous diffusion graph at the initial time, denotes the target time point, i.e., the specific time at which the node representation needs to be solved.​

[0094] Furthermore, dynamic new heterogeneous social graphs include: Provides user-cascaded binary edges for "user-cascaded" It reflects the user's participation relationship with the cascading system. It provides a "user-to-user" social relationship edge. The merged dynamic new heterogeneous social graph reflects users' inherent social connections. The node set is users and cascades, and the edge set is... .

[0095] Furthermore, a gating function representing a dynamic new heterogeneous social graph is used. Initial representation set of dynamic preference nodes The system filters features related to social interaction, such as users' long-term social preferences and social activity levels. These filtered features are then input into the GCN (Generative Network Context) based on a dynamic, novel, heterogeneous social graph. The topological structure is used to aggregate the social neighbor features of each node, ultimately obtaining the node representation at the initial time step. This serves as the initial state for the evolution of dynamic, novel, and heterogeneous social graphs.

[0096] (12-2)

[0097] in, This represents the set of node representations in the dynamic new heterogeneous social graph at the initial moment. This represents a graph convolutional network operation. The input is node features and graph structure, and the output is a node representation after aggregating neighbor information. The gate function represents a dynamic, novel, heterogeneous social graph. This represents the initial set of dynamic preference nodes. This represents a dynamic, new, heterogeneous social graph, which is composed of a static user-cascaded bipartite graph. With static social relationship graph The resulting dynamic new heterogeneous social graph (i.e.) ).

[0098] For input feature vectors and topology of The update rules for its node representation are shown in formula (13), and will not be elaborated here.

[0099] Furthermore, a highly scalable graph neural differential equation is constructed to represent the evolution of nodes in a dynamic new heterogeneous social graph over time, and the edge weights of the nodes in the dynamic new heterogeneous social graph and the node representations of the dynamic new heterogeneous social graph at different time points are determined.

[0100] Specifically, the high-scalable graph neural differential equation is a "law definition" process of social relationship dynamic evolution, which is used to describe the change rule of node representation over time, and the core of the equation is to define the change rate of , represents the node representation set of the dynamic new heterogeneous social graph at the moment, which is jointly driven by "user feature similarity" and "social structure dependence". Specifically, the feature change rate of a single user is equal to the "feature difference" between it and all other users multiplied by the sum of "similarity weight" and "structure weight". Among them, reflects the dynamic similarity of user features, and the more similar the features are, the stronger the mutual influence is, reflects the structural constraint of the social network, and the closer the social distance is, the more significant the influence is.

[0101] Specifically, the high-scalable graph neural differential equation is as follows:

[0102] (17)

[0103] (17-1)

[0104] (17-2)

[0105] wherein, represents the time derivative of the representation of the user node in the dynamic new heterogeneous social graph at the moment , characterizing the instantaneous change rate of its features, represents the representation of the user node in the dynamic new heterogeneous social graph at the moment, represents the total number of user and cascade nodes, represents the normalized weight of the social representation cosine similarity between the node and at the moment, reflecting the dynamic similarity between the node and the node , represents the static structure weight, , represents the number of neighbors of the node in the graph , reflecting the static structure constraint, represents the representation difference item between the neighbor node and the current node , represents the user node and The original similarity, with values ​​ranging from [0,2], express User nodes at any time With all user nodes The sum of the original similarities, This indicates the user set excluding user nodes. Any user index outside of, , express A set of node representations in a constantly evolving and heterogeneous social graph, containing the social characteristics of all users. Represents user node exist Representation of dynamic new heterogeneous social graphs in time The result after L2 norm normalization This represents the L2 norm.

[0106] Dynamic New Heterogeneous Social Graph The calculation of edge weights is a "structural selection" process in the dynamic evolution of social relationships, targeting dynamic and heterogeneous social graphs. Each social edge in the process is determined based on the node's "first appearance time". Whether the edge is active at any given time – an edge is only retained (with a weight of 1) if both users at both ends of the edge have participated in the propagation (first appearance time ≤ t), otherwise the weight is 0 (the edge is inactive).

[0107] Specifically, any edge in a dynamic new heterogeneous social graph The corresponding weight Determined by the following formula:

[0108] (15-2)

[0109] in, express The edge of a dynamic and heterogeneous social graph The weight, , Represents a node In dynamic new heterogeneous social graphs The number of neighbors in a dynamic new heterogeneous social graph The edges in the algorithm are filtered based on the occurrence time of each node.

[0110] Furthermore, a numerical solution method for ordinary differential equations (ODESolver) is used, with the node representation of the dynamic new heterogeneous social graph at the initial time. Starting with the previously defined highly scalable graph neural differential equation, we can solve for the results at different time points. Node representation of dynamic new heterogeneous social graph The details are as follows:

[0111] (16-2)

[0112] in, express A set of node representations for a dynamic, heterogeneous social graph. This refers to a numerical solution tool for ordinary differential equations. This represents the set of node representations in the dynamic new heterogeneous social graph at the initial moment. This represents the target time point, i.e., the specific moment represented by the node that needs to be solved.

[0113] In step 103, all participating times included in the cascade set obtained in step 101 are... Normalized to the interval [0, 1], the normalized time is obtained. This method can eliminate the time scale differences between different cascades, making the time features comparable.

[0114] Furthermore, the normalized time interval is then divided into R intervals to obtain R+1 key time points. Further, based on the dynamic novel heterogeneous diffusion graph, the node representations of the dynamic novel heterogeneous diffusion graph at different times corresponding to the R+1 key time points are determined. [ X bd 0 , X bd 1 R ,…, X bd 1 ] Based on the dynamic novel heterogeneous social graph, the node representations of the dynamic novel heterogeneous social graph at different times corresponding to R+1 key time points are determined. [ X bs 0 , X bs 1 R ,…, X bs 1 ] .in, and The "dynamic novel heterogeneous diffusion graph" and "dynamic novel heterogeneous social graph" obtained by solving the graph neural differential equation in the above steps are at time... The node representation aims to discretize continuous time, making it easier to retrieve features corresponding to any point in time.

[0115] Further, an arbitrary cascade is selected from the cascade set, and the selected cascade is referred to as a first cascade. A first participation time included in the first cascade is normalized to obtain a first time. One interval matched with the first time is determined from the R intervals, and the determined matched interval is referred to as a first interval. A start time point and an end time point of the first interval are sequentially determined. Then, a candidate feature set of a first user corresponding to the start time point and the end time point of the first interval is sequentially determined.

[0116] It should be noted that the first participation time, i.e., the original participation time of the user in the first cascade (such as formula (1) in step 101 ), after normalization, the first time obtained belongs to the processed participation time, which is consistent with the meaning of "participation time" in general expression. The candidate feature set of the first user here corresponds to the node representation of the dynamic new heterogeneous social graph at different time points of the first user at the start time point of the first interval, the node representation of the dynamic new heterogeneous diffusion graph at different time points; the node representation of the dynamic new heterogeneous social graph at different time points of the first user at the end time point of the first interval and the node representation of the dynamic new heterogeneous diffusion graph at different time points. Specifically, the candidate feature set of the first user is as follows:

[0117] x u k ms =[ X u k bs r -1 R , X u k bs r R , X u k bd r -1 R , X u k bd r R ] (18)

[0118] wherein, represents the multi-scene candidate feature set of the user in the time interval; represents the i-th user in the cascade, represents the representation of the user at the start time of the interval in the dynamic new heterogeneous social graph; represents the representation of the user at the end time of the interval in the dynamic new heterogeneous social graph; represents the representation of the user at the start time of the interval in the dynamic new heterogeneous diffusion graph; represents the representation of the user at the end time of the interval in the dynamic new heterogeneous diffusion graph; represents the representation of the user at the start time of the interval in the dynamic new heterogeneous diffusion graph; The index of the interval containing the participation time. This represents the hyperparameter, indicating the number of equal divisions the normalized time interval [0,1] is into. This represents the start time of the r-th interval. This represents the end time of the r-th interval.

[0119] It should be noted that the candidate feature set is determined in this step using a cascade as an example. Here, the selected cascade is referred to as the first cascade. However, in actual applications, all cascades included in the cascade data set in step 101 need to be processed. Therefore, when explaining other steps, the term "first" will no longer be added before the term "first" in step 103.

[0120] Because the influence weights of features at different time points within a given interval on user behavior vary (e.g., the end time point may be closer to the user's actual participation time than the start time point), an attention mechanism is needed to automatically learn and fuse these weights. Specifically, for the candidate feature set... Each feature in By calculating their importance weights and then summing them up, the user's score is obtained. The final dynamic node representation The specific formula is as follows:

[0121] (19)

[0122] (19-1)

[0123] in, Indicates user The final dynamic node representation is given by The weighted sum of the elements is obtained. Indicates user Attention weights within a specific time interval Indicates user exist Given any representation in the set, Indicates user The set of representations of the start and end times of a time interval. This represents the natural exponential function, used to amplify the differences in importance between different representations. This is represented as a temperature coefficient and is used to control the sharpness of attention weighting. The smaller the value, the more concentrated the weight distribution. This represents the trainable weight matrix in the attention mechanism. express Any node in the set is represented.

[0124] existFigure 2 In this case, the time interval [0, 1] is evenly divided into four segments, and the normalized time of user A participating in the cascade is 0.4, so its corresponding r in the cascade is 2. Then, the start and end embedding representations of the user node under different roles are retrieved, and the corresponding time points are marked with pentagrams of different colors. Finally, the final dynamic node representation of the user is obtained through the fusion operation.

[0125] Further, a static new heterogeneous relation graph is obtained according to the static user-cascade bipartite graph, the static information diffusion graph and the static social relation graph, as shown in formula (8).

[0126] In practical applications, the static new heterogeneous relation graph contains three types of edges: user-cascade participation relations, user-user propagation adjacent relations and user-user social relations, which can comprehensively capture the multi-dimensional structural associations of users in the "social + propagation" scenario.

[0127] The static node representation set of the user in the static new heterogeneous relation graph is obtained through the GCN, and the static node representation set of the user in the static information diffusion supergraph is obtained through the HGCN

[0128] Further, the static node representation set of the user in the static new heterogeneous relation graph is obtained by aggregating the node representation set of the static social attribute through a 2-layer GCN, as follows:

[0129] (20-1)

[0130] wherein, represents the static node representation set of the static new heterogeneous relation graph, represents the graph convolution network operation, represents the static social attribute node representation set, represents the static new heterogeneous relation graph.

[0131] In practical applications, the first layer GCN captures the local neighbor relations of the node, and the second layer GCN further aggregates the structural information of a longer distance, realizing the fusion of "local-global" structural features.

[0132] Further, the static node representation set of the user in the static information diffusion supergraph is obtained through the HGCN (Hypergraph Convolutional Network, supergraph convolutional network), and specifically, the static node representation set of the user in the static information diffusion supergraph is as follows:

[0133] (20-2)

[0134] wherein, a node representation set representing the static information diffusion hypergraph, a node representation set representing the static information diffusion hypergraph, a node degree matrix of the static information diffusion hypergraph, a node-hyperedge incidence matrix of the static information diffusion hypergraph, a hyperedge degree matrix of the static information diffusion hypergraph, a node representation focusing on the static social relationship graph attribute, a trainable weight matrix of the HGCN.

[0135] the two static node representation sets , reflecting the long-term structural attribute of the user, the dynamic feature reflecting the real-time state of the user at a specific time, and the combination of the two can obtain a more comprehensive node representation.

[0136] In step 104, the comprehensive node representation of the user is obtained by fusing the final dynamic node representation of the user and the static node representation of the user through an attention gate fusion mechanism.

[0137] Specifically, the static node representation of the user in the static new heterogeneous relationship graph and the final dynamic node representation of the user on the dynamic new heterogeneous social graph and the dynamic new heterogeneous diffusion graph are obtained according to the static node representation set of the user in the static new heterogeneous relationship graph. The importance weight of the static node representation of the user in the static new heterogeneous relationship graph and the final dynamic node representation of the user on the dynamic new heterogeneous social graph and the dynamic new heterogeneous diffusion graph is calculated through an attention gate fusion mechanism, and the weighted sum is obtained to obtain the comprehensive node representation of the user, specifically:

[0138] (21)

[0139] (22)

[0140] wherein, denotes the final dynamic node representation of the user fused in the static node representation of the static new heterogeneous relationship graph, in the dynamic new heterogeneous social graph and the dynamic new heterogeneous diffusion graph, denotes the node representation to be evaluated, denotes the static node representation of the user in the static new heterogeneous relationship graph, denotes the final dynamic node representation of the user found and fused in the dynamic new heterogeneous diffusion graph and the dynamic new heterogeneous social graph , ​​attention weights representing static / dynamic node representations, trainable parameters representing attention fusion, with superscript corresponding to Static-Dynamic Fusion, trainable weight matrix representing static-dynamic fusion attention, representing concatenation of the th user, representing natural exponential function, used to amplify the importance difference of different representations, representing the exponential score summation of the static node representation of the user in the static new-heterogeneous relationship graph and the final dynamic node representation of the user in the dynamic new-heterogeneous social graph and the dynamic new-heterogeneous diffusion graph, realizing weight normalization, generically the static node representation of the user in the static new-heterogeneous relationship graph and the final dynamic node representation of the user in the dynamic new-heterogeneous social graph and the dynamic new-heterogeneous diffusion graph.

[0141] The comprehensive node representation of the user determined by the above steps is the representation of each user after the fusion of the static node representation of the user in the static new-heterogeneous relationship graph and the final dynamic node representation of the user, containing the long-term structural relationship and real-time propagation behavior of the user.

[0142] In an embodiment of the present application, the time information of the user includes two types, namely the discrete form of propagation order and the continuous form of propagation time interval. Among them, the propagation order represents the propagation serial number of the user in the cascade ; and the propagation time interval represents the participation time difference between the user and the previous user.

[0143] In an embodiment of the present application, the discrete order and the continuous time interval are converted into calculable feature vectors through embedding mechanism, which are discrete order embedding reflecting the position attribute of the user in the propagation sequence and propagation time interval embedding capturing the time rhythm of the propagation.

[0144] Specifically, for the participation time sequence [ t 1 c ,…, t K c ] of each user in the cascade

[0145] (23-1)

[0146] (23-2)

[0147] wherein, represents the concatenation th user in the cascade user propagation order embedding vector, represents the multi-layer perception, represents the one-hot vector of the th user in the cascade represents the concatenation th user in the cascade user propagation time interval embedding vector, represents the concatenation th user in the cascade user after the correction of the normalized time, which is calculated as follows:

[0148] (23-3)

[0149] wherein, , represents the participation time difference between the th user and the th user, represents the maximum threshold value of the propagation time interval, represents the minimum threshold value of the propagation time interval, represents the discrete granularity after the normalization of the control propagation time interval, which is a positive integer, represents the floor function, which is used to convert the continuous normalized time difference into a discrete integer label.

[0150] It should be noted that in the above steps, the comprehensive node representation of the user is obtained, which only fuses the static node representation and the dynamic node representation, but determines the time correlation information of the user in the propagation sequence. In the embodiment of the application, the time correlation information is the key to determine whether the user will participate in the subsequent propagation, therefore, the discrete order embedding and the continuous time embedding determined above need to be spliced according to the comprehensive node representation of the user to form the complete feature of the user as follows:

[0151] z i c =[ x u i c str ; p o i c ; p t i c ] (24-1)

[0152] Z c =[ z 1 c ,…, z K c ] (24-2)

[0153] wherein, denotes the concatenated th user's full feature, denotes the concatenated th user's integrated node representation, denotes the concatenated set of base features, each element of which fuses the user's integrated node representation, discrete ordinal embedding, and continuous temporal embedding, denotes the propagation length of the concatenated

[0154] Further, the user full feature is concatenated with the user's static node representation in the static information diffusion hypergraph to obtain the static information diffusion hypergraph enhanced feature representation of the user as follows:

[0155] z ¯ i c =[ z i c ; x u i c h ] (25-1)

[0156] Z ¯ c =[ z ¯ 1 c ,…, z ¯ K c ] (25-2)

[0157] wherein, denotes the concatenated th user's static information diffusion hypergraph enhanced feature representation, wherein, denotes the concatenated th user's full feature, denotes the concatenated th user's static node representation in the static information diffusion hypergraph, denotes the static information diffusion hypergraph enhanced feature set of the concatenated It should be noted that after obtaining the set of users' static node representations in the static information diffusion hypergraph

[0158] according to formula (20-2) in step 103, the static node representation of the user in the static information diffusion hypergraph can be obtained according to the set of users' static node representations in the static information diffusion hypergraph .​​​​​

[0159] Furthermore, the cascaded features are aggregated using a multi-head attention mechanism, as follows:

[0160] (26-1)

[0161] (26-2)

[0162] (26-3)

[0163] in, This represents the result of attention calculation. Represents a query, key, and value matrix. express Activation function express Dimensions , , Indicates the first Learnable parameters of head attention Indicates the first The output feature vector of head attention Indicates the first The query matrix for head attention. Indicates the first The key matrix of head attention, Indicates the first The value matrix of head attention, The trainable output weight matrix representing multi-head attention. This indicates a cascade resulting from multi-head attention fusion. Overall characteristics Indicates the first The output feature vector of head attention Indicates the number of attention heads. Represents a matrix mask.

[0164] Furthermore, the multi-head attention output cascades the overall features, which, after layer normalization and residual network optimization, yield the final cascaded features:

[0165] (26-4)

[0166] (26-5)

[0167] in, This indicates a cascade resulting from multi-head attention fusion. The first layer of normalized overall features, This indicates a cascade resulting from multi-head attention fusion. the final feature of the user i, representing layer normalization operation.

[0168] The multi-layer perception is used to predict the possibility of each user participating in the propagation in the future, specifically:

[0169] (27)

[0170] wherein, representing the trainable weight matrix of the multi-layer perception prediction layer, representing the trainable bias parameter of the multi-layer perception prediction layer, representing the probability that the user i is the i-th user in the cascade representing the probability that the user i is the i-th user in the cascade The final feature set is fused through multi-head attention.

[0171] In the embodiment of the application, the cross-entropy loss function is used to optimize the model parameters in the training stage, specifically:

[0172] (28)

[0173] wherein, is the real propagation data label, 1 represents that the user is the i-th user participating in the cascade , All the above trainable parameters are trained through gradient back propagation and parameter optimizer.

[0174] Table 1 shows the size of the data set used in the embodiment:

[0175] Table 1 Size of test data set

[0176]

[0177] ​​​​Table 2 shows the comparison of the method provided by the embodiment of the application (DIIDP, Dynamic Interaction-aware Information Diffusion Prediction) and other comparative methods in the Hits@10 index on all data sets. The bold font is the highest accuracy, and the underlined font is the second highest accuracy. The experimental settings are as follows: all cascade data is randomly divided into a training set (80%), a validation set (10%), and a test set (10%). The maximum cascade length on all data sets is set to 200. The model training uses the Adam optimizer, the learning rate is set to 0.001, and the batch size of the training set is 16. The dimension of the user embedding is set to 64, and the dimension of the time embedding is set to 8. The ordinary differential equation solver uses the Euler method (first-order numerical method), and the step size is 0.05. The number of attention heads is set to 10.

[0178] Table 2: Hits@10 index accuracy on all data sets

[0179]

[0180] Table 3 shows the comparison of the method provided by the embodiment of the application (DIIDP) and other comparative methods in the MAP@10 index on all data sets. The bold font is the highest accuracy, and the underlined font is the second highest accuracy. The experimental settings are as follows: all cascade data is randomly divided into a training set (80%), a validation set (10%), and a test set (10%). The maximum cascade length on all data sets is set to 200. The model training uses the Adam optimizer, the learning rate is set to 0.001, and the batch size of the training set is 16. The dimension of the user embedding is set to 64, and the dimension of the time embedding is set to 8. The ordinary differential equation solver uses the Euler method (first-order numerical method), and the step size is 0.05. The number of attention heads is set to 10.

[0181] Table 3: MAP@10 index accuracy on all data sets

[0182]

[0183] Table 4 shows the ablation experiment of the method provided by the embodiment of the application on the Christianity data set and the Android data set. The experimental settings are as follows: all cascade data is randomly divided into a training set (80%), a validation set (10%), and a test set (10%). The maximum cascade length on all data sets is set to 200. The model training uses the Adam optimizer, the learning rate is set to 0.001, and the batch size of the training set is 16. The dimension of the user embedding is set to 64, and the dimension of the time embedding is set to 8. The ordinary differential equation solver uses the Euler method (first-order numerical method), and the step size is 0.05. The number of attention heads is set to 10.

[0184] Table 4 Ablation experiments for the patent method of the invention

[0185]

[0186] Method DIIDP-DO (Dynamic Interaction-aware Information Diffusion Prediction-Diffusion ODEs) is a graph differential equation representation module for removing dynamic new heterogeneous diffusion graphs, method DIIDP-SO (Dynamic Interaction-aware Information Diffusion Prediction-Social ODEs) is a graph differential equation representation module for removing social networks, method DIIDP-O (Dynamic Interaction-aware Information Diffusion Prediction-ODEs) is a graph differential equation representation module for removing dynamic new heterogeneous social graphs and dynamic new heterogeneous diffusion graphs, method DIIDP-H (Dynamic Interaction-aware Information Diffusion Prediction-Hypergraph) is a diffusion hypergraph representation module, and method DIIDP-E (Dynamic Interaction-aware Information Diffusion Prediction-Enhanced graph) is a GCN module for removing enhanced graphs. As can be seen from Table 4, the complete model DIIDP achieves the best performance in terms of Hits@10 and MAP@10 on two datasets (Christianity and Android), which shows that the synergy of various graph representation modules can significantly improve the overall performance of the model. After removing the graph differential equation module on the diffusion network (DIIDP-DO), the performance of the model decreases significantly, indicating that the diffusion network information is important for the prediction effect. Removing the graph differential equation module on the social network (DIIDP-SO) also leads to a decrease in performance, especially on the Android dataset, where MAP@10 is close to the complete model, but Hits@10 decreases, indicating that social network modeling is also indispensable and complements the diffusion network. When the graph differential equation modules on the diffusion and social networks are removed (DIIDP-O), the model performance further decreases, reflecting the enhanced effect of the combination of the two on information modeling. In addition, removing the hyper-GCN module (DIIDP-H) and the GCN module on the enhanced graph (DIIDP-E) also causes the model to perform worse, indicating that hypergraph structures and enhanced graph structures help capture more rich node relationships and structural information, improving the generalization ability of the model.In summary, each graph structure module has a positive effect on model performance, and the absence of any module will lead to a decrease in effect, verifying the effectiveness and necessity of the multi-graph fusion modeling strategy, and the graph differential equation modules of the diffusion network and social network are particularly critical.

[0187] Figure 3A and Figure 3B The average training time of different models on the Android and Douban data sets is shown, the blue column corresponds to the left Y-axis, representing the average training time of each epoch, and the red dashed line corresponds to the right Y-axis, representing the key performance indicators of the algorithm, and the models are sorted from left to right according to performance. Overall, although the training time of the DIIDP provided by the embodiment of the application increases slightly, it achieves competitive performance within an acceptable training time, demonstrating a good balance between efficiency and effectiveness.

[0188] Based on the same inventive concept, the embodiment of the application provides an information diffusion prediction device based on dynamic interaction relationship perception. Since the principle of the device for solving the technical problem is similar to that of the information diffusion prediction method based on dynamic interaction relationship perception, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described again.

[0189] As shown in Figure 4 , the device includes a construction unit 401, a first determination unit 402, a second determination unit 403, and a prediction unit 404.

[0190] The construction unit 401 is configured to obtain a cascade data set and a static social relationship graph, construct a static user-cascade bipartite graph, a static information diffusion graph, a static information diffusion hypergraph, and a new heterogeneous graph.

[0191] The first determination unit 402 is configured to obtain an initial node representation of the new heterogeneous graph through a graph convolution network GCN, construct a graph neural differential equation of the new heterogeneous node evolving over time, determine an edge weight included in the new heterogeneous dynamic graph, and determine a node representation of the new heterogeneous graph at different time points.

[0192] The second determination unit 403 is configured to normalize the participation time and divide it into multiple intervals and key time points, determine a node representation of the new heterogeneous graph corresponding to each key time point based on the node representation of the new heterogeneous graph, normalize a first participation time included in the first cascade to obtain a first time, and determine a final dynamic node representation of the first user based on the interval, the first time.

[0193] The prediction unit 404 is configured to fuse the final dynamic node representation of the user and the static node representation of the user through an attention gate fusion mechanism to obtain a comprehensive node representation of the user, and the comprehensive node representation of the user is obtained through multi-head attention concentration fusion to obtain concatenated final features, and the concatenated final features are used to predict the possibility of each user participating in future propagation through a multi-layer perception.

[0194] It should be understood that the above information diffusion prediction device based on dynamic interaction relationship perception includes units only for logical division according to the functions implemented by the device, and in actual application, the above units can be stacked or split. And the function implemented by the information diffusion prediction device based on dynamic interaction relationship perception provided in the embodiment corresponds to the information diffusion prediction method based on dynamic interaction relationship perception provided in the above embodiment. For more detailed processing flow of the device, the above method embodiment is described in detail, which will not be described in detail here.

[0195] Another embodiment of the present application also provides a computer device, which comprises a processor and a scene database; the scene database is used to store computer program codes, and the computer program codes comprise computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the information diffusion prediction method based on dynamic interaction relationship perception shown in the above method embodiment.

[0196] Another embodiment of the present application also provides a computer readable storage medium, which stores computer instructions, and when the computer instructions run on a computer device, the computer device executes each step of the information diffusion prediction method based on dynamic interaction relationship perception shown in the above method embodiment.

[0197] Finally, it should be pointed out that the above preferred embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present application.

[0198] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for information diffusion prediction based on dynamic interaction relationship awareness, characterized in that, The method comprises the following steps: obtain a cascade data set and a static social relationship graph, construct a static user-cascade bipartite graph, a static information diffusion graph, a static information diffusion hypergraph and a new heterogeneous graph; obtain an initial node representation of the new heterogeneous graph through a graph convolution network (GCN), construct a graph neural differential equation of the new heterogeneous node evolution over time, determine the edge weight included in the new heterogeneous dynamic graph and the node representation of the new heterogeneous graph at different time points; normalize the participation time and divide it into multiple intervals and key time points, and determine the node representation of the new heterogeneous graph corresponding to each key time point based on the node representation of the new heterogeneous graph; normalize the first participation time included in the first cascade to obtain a first time, and determine the final dynamic node representation of the first user based on the intervals and the first time; fuse the final dynamic node representation of the user and the static node representation of the user through an attention gate fusion mechanism to obtain a comprehensive node representation of the user; the final feature of the cascade is obtained through multi-head attention concentrated fusion, and the final feature of the cascade is predicted through a multi-layer perception to predict the possibility of each user participating in the propagation in the future.

2. The method of claim 1, wherein, The new heterogeneous graph includes a dynamic new heterogeneous diffusion graph, and the initial node representation of the new heterogeneous graph includes a dynamic new heterogeneous diffusion graph initial node representation, and the dynamic new heterogeneous diffusion graph corresponds to a dynamic edge coupled graph neural differential equation; The initial node representation of the new heterogeneous graph is obtained through a GCN, a dynamic edge coupled graph neural differential equation of the new heterogeneous node evolution over time is constructed, the edge weight included in the new heterogeneous dynamic graph and the node representation of the dynamic graph at different time points are determined, and the method comprises the following steps: through a gating function, a dynamic preference node initial representation set , a GCN, and a dynamic new heterogeneous diffusion graph, to obtain a dynamic new heterogeneous diffusion graph initial node representation: The node update rule for the following is shown below: construct a dynamic new heterogeneous diffusion graph node evolution over time dynamic edge coupled graph neural differential equation, and determine the edge weight included in the node of the dynamic new heterogeneous diffusion graph, and the node representation of the dynamic new heterogeneous diffusion graph at different time points: wherein, denotes a set of node representations of a dynamic new heterogeneous diffusion graph at an initial time instant, denotes a GCN operation, denotes a gating function for dynamic diffusion relations, denotes a set of initial representations of dynamic preference nodes, denotes a dynamic new heterogeneous diffusion graph, denotes a node representation after GCN update, denotes a nonlinear activation function, denotes any one node in a graph structure G, denotes a node in a graph structure , denotes a set of neighbor nodes in a graph structure , denotes a weight of an edge in a graph structure , denotes a general trainable weight matrix of a GCN, denotes an original representation of a neighbor node , denotes a time derivative of a set of node representations at a time instant , denotes a weight parameter balancing two evolution functions, denotes a set of node representations of a dynamic new heterogeneous diffusion graph at an initial time instant, denotes a dynamic new heterogeneous diffusion graph at a time instant , denotes a dynamic new heterogeneous diffusion graph at a time instant ,denotes an evolution function related to a dynamic graph structure, denotes an evolution function related to a node's own feature, denotes a weight of an edge in a dynamic new heterogeneous diffusion graph at a time instant , denotes a normalized exponential function, denotes a nonlinear transformation function of a node embedding feature, denotes an operation of concatenating representations of a node and a node at a time instant , , denotes a new heterogeneous dynamic diffusion representation of a neighbor node at a time instant , denotes a new heterogeneous dynamic diffusion representation of a node at a time instant , denotes a new heterogeneous dynamic diffusion representation of a node ​time of first appearance in the overall information propagation environment, representing a filter node pair that has appeared up to the time instant, representing node representation set of the dynamic new heterogeneous diffusion graph at the time instant, representing a numerical ordinary differential equation solver, node representation set of the dynamic new heterogeneous diffusion graph at the initial time instant, representing a target time point.

3. The method of claim 1, wherein, The new heterogeneous graph includes a dynamic new heterogeneous social graph, and the initial node representation of the new heterogeneous graph includes a dynamic new heterogeneous social graph initial node representation, and the dynamic new heterogeneous social graph corresponds to a high expansion graph neural differential equation; The initial node representation of the new heterogeneous graph is obtained through a GCN, a high expansion graph neural differential equation of the new heterogeneous node evolution over time is constructed, the edge weight included in the new heterogeneous dynamic graph and the node representation of the dynamic graph at different time points are determined, and the method comprises the following steps: through a gating function, a dynamic preference node initial representation set , a GCN, and a dynamic new heterogeneous social graph, obtaining an initial node representation of the dynamic new heterogeneous social graph: The node update rules are as follows: construct a dynamic new heterogeneous social graph node evolution over time high expansion graph neural differential equation, and determine the edge weight included in the node of the dynamic new heterogeneous social graph, and the node representation of the dynamic new heterogeneous social graph at different time points: in, This represents the set of node representations in the dynamic new heterogeneous social graph at the initial moment. This represents the operation of a graph convolutional network. The gate function represents a dynamic, novel, heterogeneous social graph. This represents the initial set of dynamic preference nodes. This represents a dynamic, new, and heterogeneous social graph. Represents a node The representation after GCN update. Represents a non-linear activation function. Let G represent any node in the graph structure G. Represents a node In graph structure The set of neighboring nodes in the middle, Representing graph structure Middle The weight, This represents the general trainable weight matrix of GCN. Representing neighboring nodes The original representation, express User nodes at any time Representation in dynamic neo-heterogeneous social graphs The time derivative of represents its characteristic instantaneous rate of change. This represents the total number of users and cascading nodes. express Time Node and The normalized weights of the social representation cosine similarity. express Nodes in a constantly evolving and heterogeneous social graph and Static structure weights, Represents a node exist Representation of dynamic, heterogeneous social graphs in real time. Representing neighboring nodes exist Representation of dynamic, heterogeneous social graphs in real time. express The edge of a dynamic and heterogeneous social graph The weight, Represents a node The time when it first appeared in the entire information dissemination environment, Indicates filtering The time and the node pairs that have appeared before it. express A set of node representations for a dynamic, heterogeneous social graph. represents a numerical solver for ordinary differential equations, represents a set of node representations of a dynamic new heterogeneous social graph at an initial time instant, represents a target time point.

4. The method of claim 1, wherein, normalize the participation time and divide it into multiple intervals and key time points, and determine the node representation of the new heterogeneous graph corresponding to each key time point based on the node representation of the new heterogeneous graph; normalize the first participation time included in the first cascade to obtain a first time, and determine the final dynamic node representation of the first user based on the intervals and the first time, and the method comprises the following steps: normalize the participation time included in the cascade to the [0, 1] interval, and divide the normalized time interval into R intervals and R+1 key time points; determine node representations of the dynamic new heterogeneous diffusion graph at different time points corresponding to the R+1 key time points based on the dynamic new heterogeneous diffusion graph, and determine node representations of the dynamic new heterogeneous social graph at different time points corresponding to the R+1 key time points based on the dynamic new heterogeneous social graph; normalize a first participation time included in a first cascade selected from the cascade set to obtain a first time, determine a first interval matching the first time from the R intervals, and in turn determine a candidate feature set of a first user corresponding to a start time point and an end time point of the first interval and a final dynamic node representation of the first user; the candidate feature set of the first user is as follows: the final dynamic node representation of the first user is as follows: wherein, represents a user a multi-scenario subsequent feature set within a time interval, represents the i-th user in the cascade represents the representation of a user in the dynamic new-heterogeneous social graph at the start time of the interval; represents the representation of a user in the dynamic new-heterogeneous social graph at the end time of the interval; represents the representation of a user in the dynamic new-heterogeneous diffusion graph at the start time of the interval; represents the representation of a user in the dynamic new-heterogeneous diffusion graph at the end time of the interval; represents a user participating time interval index, represents a hyper-parameter, represents the start time point of the r-th interval, represents the end time point of the r-th interval, represents the final dynamic node representation of a user represents the attention weight of a user in a specific time interval, represents any given representation of a user in the set represents a natural exponential function, represents a trainable parameter vector of the attention mechanism, represents a trainable weight matrix in the attention mechanism, represents any node representation in the set MultiScenario is an abbreviation of Multi-Scenario, which means that the feature set contains features of two scenarios of dynamic new-heterogeneous social graph and dynamic new-heterogeneous diffusion graph.​​​​​​​ 5. The method of claim 1, wherein, before the user's comprehensive node representation is obtained by fusing the user's final dynamic node representation and the user's static node representation through the attention gate fusion mechanism, it further includes: obtain a static new heterogeneous relationship graph included in the new heterogeneous graph according to the static user-cascade bipartite graph, the static information diffusion graph and the static social relationship graph, and obtain a static node representation set of the user in the static new heterogeneous relationship graph and a static node representation set of the user in the static information diffusion hypergraph according to the node representation set of the static social attribute, the GCN and the hypergraph convolution network HGCN, respectively; obtain the static node representation of the user in the static new heterogeneous relationship graph and the final dynamic node representation of the user in the dynamic new heterogeneous social graph and the dynamic new heterogeneous diffusion graph according to the static node representation set of the user in the static new heterogeneous relationship graph.

6. The method of claim 1, wherein, before the initial node representation of the new heterogeneous graph is obtained through the GCN, it further includes: Based on the gating mechanism, the user and the cascade are initialized to obtain a static social attribute node representation set and a dynamic preference node initial representation set ; wherein the cascade data set is composed of multiple cascades, and each cascade includes a user, a participation time, and an information propagation order; the gating mechanism is as follows: wherein, represents a set of node representations focusing on a specific attribute after being processed by a gating function, represents a general abstract representation of all gating functions, represents a trainable weight matrix of the gating mechanism, represents a trainable bias parameter of the gating mechanism, represents a matrix Hadamard product operation, represents a normalized exponential function, represents an initial set of node representations containing all users and cascades.

7. The method of claim 1, wherein, the comprehensive node representation of the user is obtained by multi-head attention concentrated fusion to obtain the final feature of the cascade, and the final feature of the cascade is predicted by the multi-layer perception to predict the possibility of each user participating in the propagation in the future, specifically including: convert the propagation order and the propagation time interval into discrete order embedding reflecting the propagation position attribute and time interval embedding capturing the propagation time rhythm through mapping respectively, splice the comprehensive node representation of the user, the discrete order embedding and the time interval embedding to obtain the complete feature of the user, and obtain the overall feature of the cascade and the final feature of the cascade by multi-head attention concentrated fusion of the complete feature of the user and the static information diffusion hypergraph enhanced feature representation; the final feature of the cascade is predicted by the multi-layer perception to predict the possibility of each user participating in the propagation in the future.

8. An information diffusion prediction apparatus based on dynamic interaction relationship awareness, characterized by, including: a building unit for obtaining a cascade data set and a static social relation graph , building a static user-cascade bipartite graph , a static information diffusion graph , a static information diffusion hypergraph and a new heterogeneous graph; a first determination unit configured to obtain an initial node representation of the new heterogeneous graph through a graph convolution network GCN, construct a graph neural differential equation of new heterogeneous nodes evolving over time, and determine edge weights of a new heterogeneous dynamic graph and node representations of the new heterogeneous graph at different time points; a second determination unit configured to normalize participation time and divide it into multiple intervals and key time points, and determine node representations of the new heterogeneous graph corresponding to each key time point based on the node representations of the new heterogeneous graph; The normalized first cascade includes a first participation time to obtain a first time, and a final dynamic node representation of the first user is determined based on the interval and the first time; The prediction unit is configured to fuse the final dynamic node representation of the user and a static node representation of the user through an attention gate fusion mechanism to obtain a comprehensive node representation of the user; The comprehensive node representation of the user is fused through a multi-head attention mechanism to obtain a final feature of the cascade, and the final feature of the cascade is used to predict a possibility of each user participating in future propagation through a multi-layer perception machine.

9. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a computer program. When the computer program is executed by the processor, the processor executes the information diffusion prediction method based on dynamic interaction relationship perception according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer program is stored in the memory and executed by the processor, so that the processor executes the information diffusion prediction method based on dynamic interaction relationship perception according to any one of claims 1-7.

Images