CN-122022023-A - Information diffusion prediction method based on time weighted diffusion supergraph and neural ordinary differential equation

CN122022023ACN 122022023 ACN122022023 ACN 122022023ACN-122022023-A

Abstract

The invention relates to an information diffusion prediction method based on a time weighted diffusion hypergraph and a neural ordinary differential equation, and belongs to the technical field of modeling and prediction of complex diffusion processes. The method comprises the steps of extracting a static node structure representation of a social relation diagram and a user-cascade bipartite diagram, constructing an abnormal diagram to obtain a user static correlation embedding matrix, constructing a time weighted superdiagram correlation matrix, modeling dynamic evolution of user node states by using a coupled neural ordinary differential equation system to obtain user dynamic preference embedding, projecting the user static correlation embedding matrix and the user dynamic preference embedding to a hyperbolic space to carry out gating weighted fusion, obtaining a fused feature representation, and obtaining infection probability of a user by combining a defined causal mask to predict the next user propagated based on infection probability. The method aims to solve the technical problems of time splitting, propagation simplification of a low-order structure and geometric mismatch of a feature space of discrete modeling in the prior art.

Inventors

ZHANG ZHIJIAN
WU TAO
LI ZHENGMI
JIANG LIN

Assignees

昆明理工大学

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (8)

1. An information diffusion prediction method based on a time weighted diffusion hypergraph and a neural ordinary differential equation, which is characterized by comprising the following steps: Step1, extracting a social relation diagram of a social data set and a static node structure representation of a user-cascade bipartite diagram, constructing an abnormal diagram and learning user static correlation embedding to obtain a user static correlation embedding matrix; step2, constructing a time weighted hypergraph incidence matrix according to the historical diffusion sequence; Step3, modeling the dynamic evolution of the user node state in a continuous time domain by using a coupled neural ordinary differential equation system based on the time weighted hypergraph incidence matrix, designing a bidirectional hypergraph message transmission mechanism, capturing high-order diffusion characteristics through two-stage aggregation from node to superside and superside to node, and obtaining user dynamic preference embedding by adopting a residual structure; Step4, embedding and projecting the user static correlation embedding matrix and the user dynamic preference into a hyperbolic space for gating weighted fusion to obtain a fused characteristic representation; Step5, defining a causal mask, adding the fused feature representation and the position codes to obtain an enhanced representation, and carrying out time sequence modeling on the enhanced representation by combining the causal mask to obtain a space-time feature; step6, obtaining the infection probability of the user based on the space-time characteristics and the causal mask so as to predict the next user of information transmission based on the infection probability and realize information diffusion prediction.
2. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 1, wherein Step1 specifically comprises: step1.1 define social relationship graph as directed graph , wherein, For a set of user nodes, For directed edge sets, for any user pair If user u focuses on user v, there is a directed edge And (2) and Conversely, the method can be used for controlling the temperature of the liquid crystal display device, Defining the user-cascade bipartite graph as , wherein, A set of cascading nodes is represented and, A side set for the participation relation of the user and the cascade; Step1.2 social relationship diagram With user-cascade bipartite graph Fused into an isomerism map , wherein, Comprises all user nodes, cascade nodes and edge sets Integrating two types of relations; Step1.3 use of two-layer graph neural network in heterogeneous graph Information transmission and feature learning are carried out on the information to obtain a user static correlation embedded matrix The method specifically comprises the following steps: ; Wherein, the For an initial node feature matrix initialized by a normal distribution, Is that The node characteristic matrix obtained by the two-layer graph neural network, For the embedding dimension of the space, And To map the neural network layer, from And extracting the embedding corresponding to the user node to obtain the user static correlation embedding matrix.
3. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 2, wherein Step2 is specifically: step2.1 time stamp division of the original information propagation sequence into Constructing a time sequence diffusion hypergraph for a plurality of continuous time periods Wherein the diffusion hypergraph of each period is defined as , Is the first A set of users whose propagation behavior occurs within a period of time, Is the first A set of hyperedges within a time period, each hyperedge representing a first A primary information dissemination event occurring within a time period; step2.2 constructing a time weighted hypergraph correlation matrix Wherein the element is Representing a user In the event of diffusion The participation intensity of the formula is as follows: ; Wherein, the Representing user engagement time And diffusion event start time Is used for the difference in (a), Is that The value after the normalization is carried out, A learnable time decay factor for controlling the influence of time on the propagation intensity; is the lower weight limit.
4. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 3, wherein Step3 specifically comprises: step3.1 introducing a system of coupled neural ordinary differential equations whose state vectors are embedded representations of all user nodes The system evolution is controlled by the following expression: ; Wherein, the Representing the spread field of the hypergraph, The representation is an embedded representation of all the nodes, Is composed of a learning parameter A parameterized neural ordinary differential equation vector field function; step3.2 based on time weighted hypergraph correlation matrix Defining a weighting matrix as: ; ; Wherein, the The degree of the user node is indicated, The degree of overrun of the user is indicated, And Are all weighted sums of the values that are to be added, An all 1 vector representing the user dimension, Full 1 vector representing the superside dimension; Step3.3 hypergraph diffusion field Implemented by a two-way hypergraph messaging mechanism, comprising the following 3 phases: stage 1, node-superside aggregation, for any superside Superb (superb) Representation of (2) By participating in superb Is formed by feature weighting aggregation of all user node sets, and has the expression: ; Wherein, the Is a time-aware weight reflecting the user In cascade connection Is determined by the time-dependent effect of (a), Is a node degree normalization factor; Is a node-side feature nonlinear mapping function, Is the user Is characterized by; Representing participation in a cascade Is a set of users of the (a), Representing a user Participate in cascade ; Stage 2, superside-node aggregation, namely, for any user User(s) A kind of electronic device The new representation is obtained by feature weighted aggregation of all the participating superside sets, and the expression is: ; Wherein, the Is a normalization factor of the superlimit; is a nonlinear feature transformation function on the superside and is used for capturing complex semantics of cascade content; Representation and user All the sets of hyperedges that are associated are, Denoted as the ith user After the superside-node information aggregation, obtaining a new feature embedded vector integrated with superside high-order interaction information; stage 3 outputting residual error as hypergraph diffusion field The expression is: ; Finally, the user node embedding is obtained The dynamic equation over time is: ; Wherein, the An inverse matrix representing the over-edge weighting matrix, An inverse matrix representing a weighting matrix of the user node; The dynamics equation models the continuous dynamics of the user node representation, then time of day User state of (2) Directly obtained by an ODE solver, the expression is: ; Wherein, the Representing a numerical integrator; Representing an initial user embedding; Fetch on interval [0, T ] At each time point To do the solver And integrating forward times to obtain a user hidden state sequence of each step, wherein the expression is as follows: ;; Wherein, the Is the user node at the first Embedding the dynamic characteristics of each time step into a vector; The introduction channel attention mechanism adaptively fuses dynamic information of different time steps, specifically: For any user The method comprises the following steps: ; Wherein, the And In order for the parameters to be able to be learned, Importance weights representing the i-th time step user dynamic feature embedding, final user dynamic preference embedding A weighted fusion of all time-step states.
5. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 4, wherein Step4 is specifically: step4.1 embedding user dynamic preferences Embedding matrix with static relevance to user Projection to poincare sphere model by exponential mapping The expression is: ; Wherein, the Is user dynamic preference embedding The representation in hyperbolic space after exponential mapping, Is a user static correlation embedding matrix An exponentially mapped representation projected to hyperbolic space, Representing from the origin An index mapping for starting maps the Euclidean vector to a hyperbolic manifold; step4.2, obtaining a fused characteristic representation through gating weighting and exponential mapping, wherein the expression is: ; ; Wherein, the In order to gate the vector to be controlled, 、 In order for the parameters to be able to be learned, For the representation of the features after the fusion, The representation is taken as a logarithm, Is a Sigmoid activation function.
6. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 5, wherein Step5 specifically comprises: Step5.1 representing the fused features Adding the position codes to obtain enhanced representation, wherein the expression is: ; Wherein, the Is the embedded vector of the kth position of each message; step5.2, adopting a multi-head attention decoder to carry out time sequence modeling on the enhanced representation to obtain space-time characteristics, wherein the expression is as follows: ; ; ; Wherein, attention (-) represents the Attention function, Information representing the need for current prediction, Information tags provided on behalf of the history propagation user, Representing the actual characteristic content of the historically propagated user, For the causal mask, Is the embedding dimension of the space and, , Representing the number of heads in the multi-head attention, As the weight of the material to be weighed, And In order for the weight parameters to be learned, Is a message The time-space characteristics of each user after the fusion of time and the structural characteristics.
7. The method of claim 6, wherein the causal mask is used to ensure that only historical moments are of interest, expressed as: ; If the user Time stamp of (a) Earlier than the current predicted time The causal mask is 0, otherwise set to So that the softmax weight is 0.
8. The method for predicting information diffusion based on time-weighted diffusion hypergraph and neural ordinary differential equation according to claim 7, wherein Step6 is specifically: Cascade characterization with attention mechanism through two-layer fully connected neural network And calculating the infection probability of all users, wherein the expression is as follows: ; ; Wherein, the For the probability of infection, the next user to predict information dissemination, In order for the parameters to be able to be learned, And In order for the offset to be a function of, Is a learnable parameter for calculating the infection probability of each user.

Description

Information diffusion prediction method based on time weighted diffusion supergraph and neural ordinary differential equation Technical Field The invention relates to an information diffusion prediction method based on a time weighted diffusion hypergraph and a neural ordinary differential equation, and belongs to the technical field of modeling and prediction of complex diffusion processes. Background With the rapid development of multimedia social networks, information diffusion modeling becomes a key technology for understanding user behaviors and predicting propagation paths. The existing information diffusion prediction method is mainly divided into two types, namely macroscopic prediction and microscopic prediction. Macroscopic prediction focuses on estimating the overall scale and trend of information propagation, while microscopic prediction focuses on individual behavior, aiming at identifying the next most likely affected user node in the cascade. In recent years, deep learning methods have made remarkable progress in this field, enabling automatic extraction of high-dimensional features from data. However, the information diffusion is essentially a high-order dynamic process which continuously evolves along with time, and the existing method has three limitations that firstly, a time splitting problem of discrete modeling is that most dynamic diffusion models are based on discrete time slicing or sequence modeling, smooth change of propagation intensity is difficult to describe in a continuous time domain, secondly, a propagation simplification problem of a low-order structure is that a traditional graph model usually describes a user relationship by a binary edge, a high-order interaction mode that a plurality of users participate in the same propagation event cooperatively is ignored, and thirdly, a geometric mismatch problem of a feature space is that an information diffusion network has obvious layering and radiation structures, and an Euclidean space is difficult to accurately describe exponential growth of a propagation range and influence. To solve the above problems, researchers have proposed methods based on a graph neural network, an attention mechanism, and a dynamic modeling framework. For example, dyHGCN model extracts neighbor influence and user preference by means of heterogeneous graph convolution network, MS-HGAT model introduces sequence hypergraph structure to capture user interaction preference, hyperID model uniformly processes macro and micro prediction. However, these methods are still mostly based on discrete time modeling, and cannot fully characterize the continuous evolution of user preferences in the time dimension, and the computational overhead is large, which restricts the application of the methods to large-scale data. Disclosure of Invention The invention aims to provide an information diffusion prediction method based on a time weighted diffusion hypergraph and a neural ordinary differential equation, which aims to solve the technical problems of discrete modeling of time fracture, propagation simplification of a low-order structure and geometric mismatch of a feature space in the prior art and realize continuous time dynamic modeling and efficient prediction of an information diffusion process. In order to achieve the purpose, the technical scheme of the invention is that an information diffusion prediction method based on a time weighted diffusion hypergraph and a nerve ordinary differential equation comprises the following specific steps: Step1, extracting a social relation diagram of a social data set and a static node structure representation of a user-cascade bipartite diagram, constructing an abnormal diagram and learning user static correlation embedding to obtain a user static correlation embedding matrix; step2, constructing a time weighted hypergraph incidence matrix according to the historical diffusion sequence; Step3, modeling the dynamic evolution of the user node state in a continuous time domain by using a coupled neural ordinary differential equation system based on the time weighted hypergraph incidence matrix, designing a bidirectional hypergraph message transmission mechanism, capturing high-order diffusion characteristics through two-stage aggregation from node to superside and superside to node, and obtaining user dynamic preference embedding by adopting a residual structure; Step4, embedding and projecting the user static correlation embedding matrix and the user dynamic preference into a hyperbolic space for gating weighted fusion to obtain a fused characteristic representation; Step5, defining a causal mask, adding the fused feature representation and the position codes to obtain an enhanced representation, and carrying out time sequence modeling on the enhanced representation by combining the causal mask to obtain a space-time feature; step6, obtaining the infection probability of the user based on the space-time characteristics and the