A self-explanatory information cascade popularity prediction method

Abstract

A self-explanatory information cascade popularity prediction method belongs to the field of information dissemination and prediction technology. First, the bottleneck subgraph is extracted through the information bottleneck principle to filter out noise and retain key user information; then, the propagation perception subgraph is extracted based on user distance and depth relationship to capture the key propagation pattern; finally, the cascade perception graph Transformer is used to adaptively select the features of different subgraphs to generate rich cascade representations and complete popularity prediction. The present invention effectively solves the noise problem in information dissemination through the bottleneck subgraph extraction and propagation perception subgraph extraction modules, and significantly improves the prediction accuracy. At the same time, the cascade perception graph Transformer enhances the interpretability of the model and can provide a reasonable explanation for the prediction results. The present invention is suitable for a variety of scenarios such as social media content management, false news detection, and crime source location, and has wide versatility and adaptability.

Description

Technical Field

[0001] The present invention belongs to the technical field of information dissemination and prediction, and specifically relates to a self-explanatory information cascade popularity prediction method, which can be applied to content management and marketing strategy optimization of social media platforms. Background Art

[0002] Cascading popularity prediction aims to predict users' future engagement with online content and is crucial for the effective marketing and management of social media platforms.

[0003] Current methods for predicting the popularity of information cascades primarily rely on modeling the structure and temporal characteristics of cascade graphs, but often overlook the noise associated with information dissemination and lack interpretability. This leads to inaccurate predictions, particularly in the presence of noise such as fake followers and spam accounts, raising questions about the reliability of the models.

[0004] Patent CN118863161A proposes a method for predicting information popularity based on heterogeneous cascade graphs and public opinion field modeling. This method learns representations by constructing conditional intensity functions and public opinion field paradigms, but the model lacks good interpretability. Patent CN119172573A proposes a method for predicting short video popularity based on multimodal retrieval enhancement. This method predicts popularity by extracting keyframes, obtaining image descriptions, and retrieving related videos. However, it does not address the noise in information dissemination, and the model lacks interpretability. Patent CN119364121A obtains various data sets to determine various video features and integrates them to predict popularity. However, this method lacks data processing diversity and scalability in application scenarios, and it also fails to address noise and lacks interpretability. In contrast, patent CN119692524A first uses a hypergraph neural network and Transformer to obtain embedded representations of user-generated content (UGC) before performing popularity prediction. However, this patent focuses on multimodal social media popularity prediction and has a relatively narrow application scenario. It lacks versatility and adaptability when faced with other information dissemination scenarios such as online video popularity prediction and news dissemination heat analysis. Summary of the Invention

[0005] This paper proposes a self-explanatory information cascade popularity prediction framework, ExCas, which aims to solve the following problems: filtering out noise in information dissemination and improving prediction accuracy; enhancing the interpretability of the model, providing a reasonable explanation of the prediction results, and improving the applicability of the model in risk-sensitive scenarios, such as fake news detection and crime source location.

[0006] To solve the above technical problems, the specific technical solution of a self-explanatory information cascade popularity prediction method of the present invention is as follows:

[0007] Step 1: Bottleneck subgraph extraction: Obtain user network data to construct a cascade graph, filter out noise information based on the information bottleneck principle, retain nodes and edges related to the prediction task, and form a bottleneck subgraph;

[0008] Step 2: Extracting the propagation-aware subgraph: Calculate the propagation correlation between users based on the cascade graph, use the propagation correlation as the edge weight, and form a propagation-aware subgraph.

[0009] Step 3: Cascaded Perceptual Graph Transformer prediction: The multi-head attention mechanism based on the cascaded perceptual graph transformer adaptively selects the most significant features from the bottleneck subgraph and the propagation perceptual subgraph, and generates a cascade representation through a mask function and a feedforward network.

[0010] Step 4: Prediction and optimization; input the cascade representation into the multilayer perceptron for popularity prediction; perform joint optimization through mean squared logarithmic error loss and KL divergence loss.

[0011] The step 1 is specifically as follows:

[0012] First, a cascade graph Gc=(Vc,Ec) is constructed based on the user network data, where Vc represents the user node, which serves as the node of the cascade graph, and Ec represents the forwarding relationship between users, which serves as the edge of the cascade graph;

[0013] Then, the weight of each node and edge in the cascade graph is perturbed, and the weight of the edge (u, v)∈Ec between user u and user v is It is calculated by multi-layer perceptron MLP and expressed as:

[0014] ;

[0015] in, and are the feature representations of users u and v obtained by encoding the cascade graph Gc through the graph neural network, and σ is the sigmoid function;

[0016] Finally, according to the calculated weight , perform Bernoulli sampling on each edge to decide whether to keep the edge.

[0017] The step 2 is specifically as follows:

[0018] The propagation correlation is calculated using the following formula:

[0019] ;

[0020] in is the propagation correlation between user u and user v, is the shortest path length between user u and user v, is the shortest path length from user u to the source user, is the shortest path length from user v to the source user, and e is a natural constant;

[0021] Then the correlation will be propagated The weight information of the edge is input into the graph neural network to obtain the propagation-aware subgraph.

[0022] The step 3 is specifically as follows:

[0023] The attention heads of the multi-head attention mechanism in the graph transformer are divided into two groups, processing the bottleneck subgraph and the propagation-aware subgraph respectively; for each attention head h, the attention score is calculated :

[0024] ;

[0025] where Q h , K h and V h represent query, key, and value respectively, X represents the input features of the attention head, is the adjacency matrix of the bottleneck subgraph or propagation-aware subgraph; M is the mask function; d represents the dimension of the input features, and H represents the number of attention heads;

[0026] The outputs of all attention heads are then fused, layer normalized, and fed point-by-point into a feed-forward network to obtain the updated cascade representation Z.

[0027] The joint optimization through mean square logarithmic error loss and KL divergence loss is as follows:

[0028]

[0029]

[0030]

[0031] Among them, L represents the total loss function, is a hyperparameter; Indicates that the model is good at the kth sample at time The popularity prediction value of Indicates the kth sample at time The real popularity value.

[0032] The beneficial effects of the present invention are as follows:

[0033] This method, through bottleneck subgraph extraction and propagation-aware subgraph extraction modules, effectively filters out noise in information propagation, significantly improving prediction accuracy. Furthermore, through cascading perceptual graph transformers and a joint optimization strategy, the model's interpretability is enhanced.

[0034] The method of the present invention utilizes the cascaded perceptual graph Transformer to adaptively select features of different subgraphs to generate richer cascade representations.

[0035] The method of the present invention is more comprehensive in data processing, performs more in-depth mining and processing of data, and improves the diversity of data processing.

[0036] The method of the present invention has a wider range of application scenarios and is not limited to multimodal social media popularity prediction. It also has good versatility and adaptability in other information dissemination scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the flow of the information cascade popularity prediction method of the present invention;

[0038] Figure 2 Extract a flow diagram for the bottleneck subgraph;

[0039] Figure 3 Extracting a flow chart for the propagation-aware subgraph;

[0040] Figure 4 This is a schematic diagram of the cascaded perception graph Transformer structure; DETAILED DESCRIPTION

[0041] In order to better understand the purpose, structure and function of the present invention, a self-explanatory information cascade popularity prediction method of the present invention is further described in detail below with reference to the accompanying drawings.

[0042] This embodiment proposes a self-explanatory information cascade popularity prediction framework ExCas, whose core technical solution includes three main modules: bottleneck subgraph extraction module, propagation-aware subgraph extraction module and cascade-aware graph transformer module. Figure 1 As shown, first input the cascade graph G CUsing the information bottleneck (IB) principle, we identify key users in the cascade graph, filter out noise information, and retain task-related nodes and edges to form a bottleneck subgraph. We then use the distance and depth relationships between users to calculate propagation relevance and extract a propagation-aware subgraph. We adaptively select the most significant features from the bottleneck and propagation-aware subgraphs to generate a cascade representation for the final popularity prediction. Joint optimization is then performed to improve prediction accuracy and model reliability. The following sections describe each module in detail.

[0043] Step 1: Bottleneck subgraph extraction: Using the information bottleneck (IB) principle, identify key users in the cascade graph, filter out noise information, retain task-related nodes and edges, and form a bottleneck subgraph.

[0044] The information bottleneck principle seeks a balance between information compression and information retention. In this module, it aims to improve the interpretability and prediction accuracy of the model by removing noise information that is useless for predicting popularity, highlighting key users and key structural features.

[0045] A cascade graph Gc=(Vc, Ec) is constructed based on user network data. The cascade graph is a mathematical representation of the information propagation structure in the user network. Vc represents the user node as a node in the cascade graph, and Ec represents the forwarding relationship between users as an edge in the cascade graph.

[0046] A random attention mechanism is designed to perturb the weight of each node and edge in the cascade graph. The weight of the edge (u,v)∈Ec between user u and user v is It is calculated through a multi-layer perceptron (MLP) and expressed as:

[0047]

[0048] in, and is the feature representation of user u and v obtained by encoding the cascade graph Gc through the graph transformer. σ is the sigmoid function, which maps the output of MLP to the interval [0,1] so that It can express the probability that the edge (u,v) is retained. The process is as follows Figure 2 shown.

[0049] According to the calculated weight , perform Bernoulli sampling on each edge to decide whether to retain the edge. Specifically, the probability of edge (u,v) being retained is , the probability of being discarded is 1− In this way, edges that are valuable for popularity prediction are screened out from numerous edges, thereby generating a bottleneck subgraph in which only nodes and edges related to the popularity prediction task are retained. In this process, in order to determine the optimal bottleneck subgraph, the following formula is optimized:

[0050]

[0051] The constraints are ,β>0; where is a function determined by the parameter ϕ, which generates a bottleneck subgraph based on the original cascade graph Gc , is the mapping relationship from the original graph space to the bottleneck subgraph space, where the superscript c indicates the information cascade and is used to identify the graph structure involved in the current information cascade scenario. The parameter ϕ is adjusted through the optimization algorithm during model training to make the generated bottleneck subgraph meet the conditions of tasks such as popularity prediction; β is a hyperparameter used to weigh and The relative importance in the optimization objective; β>0, when the value is large, the model pays more attention to the bottleneck subgraph retaining the original cascade graph information; when the value is small, it pays more attention to the extraction of useful information for popularity prediction. Represents the bottleneck subgraph The mutual information between the bottleneck subgraph and the popularity label Y, which measures the useful information of the bottleneck subgraph for popularity prediction. The present invention uses regression loss as an alternative optimization objective function to minimize this term to improve the accuracy of prediction; Represents the bottleneck subgraph The mutual information between the bottleneck subgraph and the original cascade graph Gc reflects the degree of information retention of the original graph. Since it is difficult to directly estimate these two mutual information, the variational method is used to optimize the second By variational approximation To approximate the marginal distribution:

[0052]

[0053] The upper bound is obtained by deduction: ,in is the bottleneck subgraph The probability distribution of Given the original cascade graph Under the condition of φ, the bottleneck subgraph obtained by the model with parameter φ The conditional probability distribution of ; is the original cascade graph The probability distribution of Represents the original cascade graph Seek expectations.

[0054] Therefore, minimizing is equivalent to minimizing the KL divergence. Finally, the KL divergence is as follows:

[0055]

[0056] where is a hyperparameter with values in the range [0, 1].

[0057] Step 2: Propagation-aware Subgraph Extraction (PSE); The propagation pattern of information among users has an important impact on popularity prediction. This module extracts propagation-aware subgraphs from two dimensions, distance and depth, using the relationships between users to capture key propagation patterns.

[0058] For each edge (u, v) of user pairs in the cascade graph, calculate the shortest path length between them , and for each user u, calculate the shortest path length from it to the source user . The source user refers to the user who initially initiates the information propagation during the information cascade process. It is the starting point of information propagation. Calculating the shortest path length from a user to the source user can measure the distance of the user from the information source in the propagation path and provide an important reference for analyzing the propagation pattern. According to the distance and depth relationships between users, calculate the propagation correlation between users :

[0059]

[0060] where represents the distance relationship between users, represents the depth relationship between users, e is the natural constant, is the shortest path length from user v to the source user.

[0061] To improve computational efficiency, this module only calculates the shortest paths for user pairs whose shortest paths do not exceed a predetermined threshold D, that is < D. Such a design enables the Propagation-aware Subgraph Extractor (PSE) to focus on those user relationships that have a greater impact on information propagation, capture the relationships between closely related users, and at the same time ignore the relatively weak and potentially interfering relationships between users with a greater distance. Finally, PSE learns the graph . The specific process is as follows: Input the calculated propagation correlation between users as the edge weight information into the graph neural network model. During training, the model continuously updates the node feature representations through a multi-layer message passing mechanism based on this weight information and the initial node features (such as user attribute features). After training, determine the adjacency matrix A according to the set rules (such as comparing the edge weights with the threshold, or based on the degree of association of the learned node features) P, retain the user pairs with higher communication correlation in the graph, and then form a communication-aware subgraph, such as Figure 3 As shown in Figure 3, position encoding is designed to better represent node position information in the propagation-aware subgraph.

[0062] Step 3: Cascaded Graph Transformer; This is a graph neural network model based on the Transformer architecture for processing graph-structured data. Its core is to capture the dependencies between nodes in the graph through a self-attention mechanism, thereby learning feature representations for the nodes. The overall architecture of the Graph Transformer consists of an input layer, a multi-head attention layer, layer normalization, a point-by-point feedforward network, and an output layer. It can adaptively select the most salient features from the bottleneck subgraph and the propagation-aware subgraph to generate a rich cascade representation for the final popularity prediction.

[0063] like Figure 4 As shown, the bottleneck subgraph G IB and propagation-aware subgraph G P Enter the Cascaded Graph-Aware Transformer. To fully exploit the key features in the two subgraphs, a Cascaded-Aware multi-head attention mechanism is designed. The attention heads are divided into two groups, one for the bottleneck subgraph and the other for the propagation-aware subgraph. This grouping allows the model to focus on the unique information of each subgraph, avoiding interference between different subgraphs and more accurately capturing features that are important for popularity prediction.

[0064] Specifically, for each attention head h, the attention score is calculated :

[0065]

[0066] where Q h , K h and V h represent query, key, and value respectively, X represents the input features of the attention head, d represents the dimension of the input features, and H represents the number of attention heads. is the adjacency matrix of the bottleneck subgraph or propagation-aware subgraph, which determines the subgraph structure that the attention mechanism focuses on during computation. M is a masking function that adjusts the attention score so that the model can focus more on the relevant information of a specific subgraph when calculating attention, enhancing the model's ability to distinguish between different subgraph features and thus more effectively extracting key information from different subgraphs.

[0067] The outputs of different attention heads are fused, then processed through layer normalization and a point-by-point feedforward network (FFN) to introduce nonlinearity into the generated representation, further enhancing the model's expressive power. Layer normalization accelerates model training and improves model stability. The point-by-point feedforward network uses nonlinear transformations to further explore the complex relationships between features, generating an updated concatenated representation Z. This cascaded representation Z is then output for subsequent popularity prediction.

[0068] Step 4: Prediction and optimization; through joint optimization, the model can effectively filter noise and enhance interpretability while predicting popularity, thereby improving the accuracy of prediction and the reliability of the model.

[0069] The concatenated representation Z is input into a multilayer perceptron (MLP) for popularity prediction:

[0070]

[0071] Use mean squared logarithmic error (MSLE) as the loss function , calculate the error between the predicted value and the true value:

[0072]

[0073] in Indicates that the model is good at the kth sample at time The popularity prediction value of The predicted value is converted to logarithmic space, which makes the loss function more sensitive to the relative error of the predicted value. Indicates the kth sample at time The real popularity value of is to transform the real value into logarithmic space, Together they are used to calculate the error in logarithmic space.

[0074] The mean squared logarithmic error loss function sums the squares of the logarithmic differences between the predicted and true values ​​and then takes the average. Compared with other loss functions, it is more sensitive to the relative error of the predicted value and can better reflect the accuracy of the model's popularity prediction. It is particularly suitable for popularity prediction tasks with a large data range and a large relative error impact.

[0075] The MSLE loss is combined with L for optimizing the prediction popularity model. KL Loss functions are combined for joint optimization:

[0076]

[0077] Among them, L represents the total loss function, is a hyperparameter used to control the auxiliary task (with LKL The degree of influence of correlation) in the overall optimization goal. During the model training process, The value of determines the emphasis on noise filtering and model interpretability. If If the value is too large, the model may focus too much on noise filtering and interpretability, sacrificing a certain degree of prediction accuracy; if If the value is too small, it may not be able to fully play the role of filtering noise and enhancing interpretability. In different data sets and task scenarios, the appropriate The values ​​will vary and need to be tuned and determined through a large number of experiments to balance the model's performance between popularity prediction accuracy, noise filtering, and model interpretability, and to find the optimal model parameter configuration so that the model can effectively filter noise when predicting popularity, provide interpretable prediction results, and ensure high prediction accuracy.

[0078] This paper implements self-explanatory information cascade popularity prediction through three modules: bottleneck subgraph extraction, propagation-aware subgraph extraction, and cascade-aware graph transformer. This framework not only effectively filters out noise in information propagation, but also captures key propagation patterns and generates rich cascade representations, thereby improving prediction accuracy and model interpretability.

[0079] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A self-explanatory information cascade popularity prediction method, characterized in that: The following steps are involved: Step 1: Bottleneck subgraph extraction: Obtain user network data to construct a cascade graph, filter out noise information based on the information bottleneck principle, retain nodes and edges related to the prediction task, and form a bottleneck subgraph; Step 2: Extracting the propagation-aware subgraph: Calculate the propagation correlation between users based on the cascade graph, use the propagation correlation as the edge weight, and form a propagation-aware subgraph. Step 3: Cascade Perceptual Graph Transformer Prediction; The multi-head attention mechanism based on the cascaded perceptual graph transformer adaptively selects the most significant features from the bottleneck subgraph and the propagation perceptual subgraph, and generates a cascade representation through a mask function and a feedforward network; Step 4: Prediction and optimization: Input the cascade representation into the multilayer perceptron for popularity prediction; perform joint optimization using mean squared logarithmic error loss and KL divergence loss; The step 1 is specifically as follows: First, a cascade graph Gc = (Vc, Ec) is constructed based on the user network data, where Vc represents the user node, which serves as the node of the cascade graph, and Ec represents the forwarding relationship between users, which serves as the edge of the cascade graph; Then, the weight of each node and edge in the cascade graph is perturbed, and the weight p of the edge (u, v)∈Ec between user u and user v is u,v It is calculated by multi-layer perceptron MLP and expressed as: p u,v =σ(MLP(X u ,X v )); Among them, X u and X v are the feature representations of users u and v obtained by encoding the cascade graph Gc through the graph neural network, and σ is the sigmoid function; Finally, according to the calculated weight p u,v , perform Bernoulli sampling on each edge to decide whether to keep the edge; The step 3 is specifically as follows: The attention heads of the multi-head attention mechanism in the graph transformer are divided into two groups, processing the bottleneck subgraph and the propagation perception subgraph respectively; for each attention head h, the attention score Attn is calculated h (X): where Q h , K h and V h represent query, key, and value respectively, X represents the input feature of the attention head, and A i is the adjacency matrix of the bottleneck subgraph or propagation-aware subgraph; M is the mask function; d represents the dimension of the input features, and H represents the number of attention heads; The outputs of all attention heads are then fused, layer normalized, and fed point-by-point into a feed-forward network to obtain the updated cascade representation Z.

2. A self-explanatory information cascade popularity prediction method according to claim 1, characterized in that: The step 2 is specifically as follows: The propagation relevance is calculated using the following formula: Among them, P uv is the communication correlation between user u and user v, d uv is the shortest path length between user u and user v, d u is the shortest path length from user u to the source user, d v is the shortest path length from user v to the source user, and e is a natural constant; Then propagate the correlation P uv The weight information of the edge is input into the graph neural network to obtain the propagation-aware subgraph.

3. A self-explanatory information cascade popularity prediction method according to claim 2, characterized in that: The joint optimization using mean square logarithmic error loss and KL divergence loss is specifically as follows: L=L CPP +γL KL ; Among them, L represents the total loss function, γ and ρ are hyperparameters; Indicates that the model is correct for the kth sample at time t p The popularity prediction value, P k (t p ) represents the kth sample at time t p The real popularity value.

4. A self-explanatory information cascade popularity prediction method according to claim 3, characterized in that: The graph neural network in step 1 is specifically a graph transformer.

5. A self-explanatory information cascade popularity prediction method according to claim 4, characterized in that: In step 2, only the shortest paths of user pairs whose shortest paths do not exceed a predetermined threshold D are calculated.

Images