Subgraph-based information cascading prediction method and system

Abstract

The application provides a subgraph-based information cascade prediction method and system, belonging to the technical field of social network analysis and neural network, comprising the following steps: S1, constructing a deep learning information cascade prediction model CasSubTS, inputting collected user published information through an input layer, and constructing an information cascade graph G; S2, inputting G into a subgraph sampling layer, dividing G into a plurality of information cascade subgraphs according to different time steps, and converting the information cascade subgraphs into adjacency matrices; aggregating node features of the adjacency matrices to obtain a feature representation matrix B; S3, inputting B into a feature learning layer to obtain a characteristic vector with structural features and time sequence features; S4, inputting an input feature weighting layer, and utilizing a channel attention mechanism to perform weighted fusion to obtain a weighted characteristic vector; and S5, inputting an input prediction layer to predict a final macro cascade increment. The method can effectively predict information cascade in a social network.

Description

Technical Field

[0001] This invention belongs to the field of social network analysis and neural network technology, and specifically relates to a method and system for predicting information cascades based on subgraphs. Background Technology

[0002] With the continuous innovation of internet technology and the widespread adoption of mobile smart terminals, people's social behavior has become closely integrated with digital technology, ushering in a digital social era. Online social platforms such as Weibo, WeChat, Douyin, and Facebook have broken through the time and space limitations of information interaction, gradually replacing traditional information media and becoming the mainstream channels for the public to publish and receive information. Social networks built on the internet have unprecedentedly enhanced the breadth and depth of information dissemination, becoming the main carrier of information dissemination in today's society, profoundly influencing changes in social culture and public psychology.

[0003] Social networks are defined as a network system formed by individual participants through social relationships. In a social network, individuals are considered nodes, representing individuals, organizations, or other entities participating in social activities, while the edges connecting nodes represent the social relationships between individuals. The widespread application of information technology and social platforms has changed the traditional models and rules of social interaction, extending the ways individuals establish social relationships online, such as following, liking, and sharing. This has also expanded the scope of the concept of social networks; now, social networks not only refer to offline relationship networks but also include online social networks. One of the core characteristics of online social networks is user-generated content (UGC), which typically includes multiple modalities such as text, images, and videos. UGC is currently the mainstream method of information production and reception, reducing the cost of content creation for users and increasing content diversity and the immediacy of interaction. Because of this characteristic, online social networks attract a huge user base, and the network continues to grow in scale.

[0004] Information cascading is an effect caused by the spread of information through social networks. It generally consists of the information being spread, the participants in the information dissemination, and the path of information transmission. The information cascading effect can accelerate the spread of information within a network and reshape the opinions and attitudes of individual participants, thereby influencing overall public opinion trends. Given its significance, academia and industry have been closely monitoring research related to information cascading prediction. In practice, information cascading can be predicted from both macro and micro perspectives. Macro-level information cascading prediction focuses on the overall trend of information evolution and is typically defined as predicting the increment of information cascading within a specific time period for a particular piece of information in a social network. Micro-level information cascading prediction focuses more on exploring local information transmission paths and patterns, predicting individual users joining the information cascading process at the next moment. However, predicting information cascading in social networks faces many difficulties and challenges:

[0005] (1) The information propagation path in social networks is easily affected by user interaction and has a strong randomness, which further exacerbates the uncertainty of the network structure and makes the information cascade prediction task more difficult.

[0006] (2) The evolution of online social networks is a dynamic process that is affected by time factors and can fluctuate greatly, such as explosive growth or linear decline.

[0007] (3) The information on social platforms is complex and multi-dimensional. The influence of different factors on the prediction results is not linearly superimposed, and researchers find it difficult to fully capture all the factors that affect the information cascade. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a subgraph-based information cascading prediction method, comprising the following steps:

[0009] Step S1: Construct a deep learning information cascade prediction model CasSubTS, including: an input layer, a subgraph sampling layer, a feature learning layer, a feature weighting layer, and a prediction layer; pass the collected user-posted information through the input layer to construct an information cascade graph G;

[0010] Step S2: Input the information concatenated graph G into the subgraph sampling layer, divide it into several information concatenated subgraphs by setting different time steps, and convert them into adjacency matrices. Aggregate the node features of the adjacency matrices to obtain the feature representation matrix B.

[0011] Step S3: Input the feature representation matrix B into the feature learning layer to learn structural features and temporal features simultaneously, thereby obtaining a feature vector that possesses both structural and temporal features. ;

[0012] Step S4: Input the feature weighting layer and use the channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. ;

[0013] Step S5: ... Input the prediction layer to predict the final macroscopic cascade increment.

[0014] Beneficial effects:

[0015] 1. The subgraph-based information cascade prediction method proposed in this invention considers multi-scale structural information in the learning of structural features, incorporating the out-degree and in-degree of nodes as directional information into the structural features to achieve a more comprehensive learning of the structure of the cascaded graph. Simultaneously, the introduction of a multi-head attention mechanism balances the robustness and flexibility of the model.

[0016] 2. In the learning of temporal features, the method of this invention does not adopt the approach of directly using recurrent neural networks to extract temporal information as in the prior art. Instead, considering the differences in the temporal information implied by each node in different time periods, a node attention mechanism is designed. After aggregating the nodes containing more important information, a recurrent neural network is then introduced to learn temporal dependencies.

[0017] 3. Regarding the weighted fusion of structural and temporal features, this invention uses a channel attention mechanism with stronger feature focus capabilities to perform weighted fusion of features. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the information cascading prediction method based on subgraphs according to the present invention;

[0019] Figure 2 This is a schematic diagram of the CasSubTS model.

[0020] Figure 3 This is a schematic diagram of the node attention mechanism.

[0021] Figure 4 This is a structural block diagram of an information cascade prediction system based on subgraphs according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1

[0024] like Figure 1 As shown in the figure, an information cascading prediction method based on subgraphs provided by an embodiment of the present invention includes the following steps:

[0025] Step S1: Construct a deep learning information cascade prediction model CasSubTS, including: an input layer, a subgraph sampling layer, a feature learning layer, a feature weighting layer, and a prediction layer; the collected user-posted information is passed through the input layer to construct an information cascade graph G;

[0026] Step S2: Input the information concatenated graph G into the subgraph sampling layer, divide it into several information concatenated subgraphs by setting different time steps, and convert them into adjacency matrices. Aggregate the node features of the adjacency matrices to obtain the feature representation matrix B.

[0027] Step S3: Input the feature representation matrix B into the feature learning layer to learn both structural and temporal features, thereby obtaining a feature vector that possesses both structural and temporal features. ;

[0028] Step S4: Input feature weighting layer, using channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. ;

[0029] Step S5: ... Input the prediction layer to predict the final macroscopic cascade increment.

[0030] In one embodiment, step S1 above: constructing a deep learning information cascade prediction model CasSubTS, including: an input layer, a subgraph sampling layer, a feature learning layer, a feature weighting layer, and a prediction layer; passing the collected user-posted information through the input layer to construct an information cascade graph G, specifically including:

[0031] The CasSubTS proposed in this invention is an end-to-end deep learning information cascade prediction model, and its structure diagram is as follows. Figure 2 As shown, it includes five parts: input layer, subgraph sampling layer, feature learning layer, feature weighting layer and prediction layer.

[0032] The collected user-posted information is processed through the input layer to construct an information cascade graph G, which specifically includes:

[0033] Step S11: User u publishes an initial message m at time t0. During the observation time, n other users v interact with this message. Then the message cascade C k The formal definition is;

[0034] ;

[0035] Step S12: Cascade information C k The corresponding C k The definition of an information cascade graph G is:

[0036] ;

[0037] in, This represents a node in an information cascade diagram, i.e., a user who participates in the dissemination of information; This represents the edges in the information cascade graph, i.e., the interaction relationships between users.

[0038] The time factor profoundly influences the scale and propagation path of information cascading. On one hand, an information cascading graph formed by an initial piece of information *m* may attract a large number of nodes to join over time, triggering a large-scale cascading effect and generating a large-scale information cascading graph. Large-scale information cascading graphs typically increase the inconvenience of graph storage and computation, limiting the data range applicable to the model. On the other hand, the order and time intervals at which different nodes join the information propagation process in an information cascading graph are different. Current methods mostly use the aggregation of node sequences to represent the cascading, but this approach ignores the dynamic changes in the temporal sequence of nodes during information propagation, failing to represent the true propagation phenomenon and causing fitting errors in the model.

[0039] To mitigate the impact of the aforementioned two problems on prediction models, a series of sampling methods have been proposed. Among them, two influential methods are random walk-based sampling and propagation path-based sampling. The classic model using random walk sampling is DeepCas, which constructs a subgraph by randomly walking across the network to obtain different node sequences. However, this method fundamentally ignores the global and local dynamics of cascading changes, resulting in significant limitations. Propagation path-based sampling, on the other hand, collects sequences along the propagation path of each node in the cascading graph. This can lead to inconsistent sequence lengths and data sparsity.

[0040] This invention, to fully consider the temporal nature of cascade evolution in the information cascade subgraph sampling stage, proposes a time-step-based information cascade subgraph sampling method, which samples the subgraph based on the temporal information of each node in the observed information cascade graph. Furthermore, to ensure the efficiency of the training process, this invention uses a partial sampling method, setting different time step sizes T to decompose a certain number of cascade subgraphs and obtain the sequence of each subgraph. The final number of sampled subgraphs N can be obtained by the following formula:

[0041] ;

[0042] in, t represents the size of the information cascaded subgraph sequence, and t is the set time step.

[0043] In one embodiment, step S2 above involves: inputting the information concatenated graph G into the subgraph sampling layer, dividing it into several information concatenated subgraphs by setting different time steps, converting them into adjacency matrices, and aggregating node features of the adjacency matrices to obtain the feature representation matrix B, specifically including:

[0044] Step S21: Using a partial sampling method, by setting different time steps T, the information concatenation graph G is decomposed into a certain number of information concatenation subgraphs, and a set of information concatenation subgraph sequences is obtained:

[0045] ;

[0046] in, This represents a sequence of cascaded subgraphs obtained by sampling at different time steps;

[0047] Step S22: For After concatenation, the corresponding adjacency matrix can be represented as:

[0048] ;

[0049] in, The adjacency matrix representation of a single subgraph sequence;

[0050] Step S23: Perform node feature aggregation on the adjacency matrix to obtain the feature representation matrix, which can be represented as:

[0051] ;

[0052] in, is a learnable parameter; j is the column index of the adjacency matrix.

[0053] Structural and temporal features are the two most important features in the information cascade prediction process. To obtain effective representations of structural and temporal features, the CasSubTS model of this invention employs a parallel feature learning approach, using MH-GAT and Bi-GRU to simultaneously learn the structural and temporal features of the cascade graph. This parallel feature learning process helps reduce the time consumed in data processing and model training, improves the efficiency of data information utilization, and enhances the model's generalization ability.

[0054] In one embodiment, step S3 above involves inputting the feature representation matrix B into the feature learning layer to simultaneously learn structural and temporal features, thereby obtaining a feature vector that possesses both structural and temporal features. Specifically, it includes:

[0055] Steps S31-S35 below, from the perspective of spatial convolution, employ the MH-GAT model to convolve the feature representation matrix of the input information cascade graph, dynamically assigning different attention weights to each node to achieve weighted aggregation of node features, ultimately obtaining an updated vector representation containing node structural information, i.e., structural features. Before training the MH-GAT model, considering the multi-scale characteristics of node structural information, and to more comprehensively learn the structural information of the cascade graph and improve the model's predictive performance, this invention incorporates node orientation information into the structural features. This information can be represented by the node's degree. Generally, in information cascade graphs, nodes with higher degree values ​​tend to have more social connections, thus playing a more crucial role in information propagation and having a more significant impact on the model's prediction results.

[0056] With node v i The number of edges that are endpoints is called v. i The degree is expressed by the following formula:

[0057] .

[0058] Step S31: Perform a linear transformation on the node features of the feature representation matrix using a linear transformation matrix, mapping the original node features to different representation spaces to obtain the query vector for each attention head. Key vector Sum value vector :

[0059] ;

[0060] ;

[0061] ;

[0062] Where X is the node feature vector and H is the linear transformation matrix;

[0063] Step S32: In order to evaluate node v i Its neighbor node v j The correlation and importance between nodes are calculated using an attention mechanism for each attention head k. i Its neighbor node v j Attention score e ij :

[0064] ;

[0065] in, It is a linear function. It is the parameter vector of the attention head k;

[0066] Step S33: Calculate node v i Its neighbor node v j Attention weights:

[0067] ;

[0068] in, It is node v i The set of neighboring nodes, Represents any neighbor node and v i The correlation coefficient;

[0069] Step S34: Use attention weights to perform weighted aggregation of the features of each node's neighbor nodes, and dynamically adjust the aggregated representation of each node. The calculation process is as follows:

[0070] ;

[0071] To address the interference of noisy data, avoid overfitting, and make the training process more robust, a multi-head attention mechanism is introduced into GAT to construct the MH-GAT model.

[0072] Step S35: The MH-GAT model uses k attention heads to independently learn the feature representation of each node, and then sums the learned k sets of node representations and takes the average to obtain the structural features. :

[0073] ;

[0074] In the structural feature learning steps described above, the CasSubTS model uses a graph attention network that incorporates a multi-head attention mechanism to extract features from the complex structure of the cascaded graph, thereby obtaining node structural features that contain information about the cascaded topology. .

[0075] To enhance the learning ability of implicit temporal information in information cascade graphs and reduce information loss during the temporal feature learning process, this invention first uses a designed node attention mechanism in the CasSubTS model to adaptively adjust node weights while aggregating representations from different nodes. Then, a variant of the recurrent neural network, the Bi-GRU model, is used to model the temporal dependencies in the information propagation process.

[0076] Step S36: Learn its temporal features based on structural features. Attention mechanism for computation nodes:

[0077] ;

[0078] ;

[0079] in, For projection parameters, These are the dimensions of structural features and temporal features, respectively; These are the node weights learned by the fully connected layer.

[0080] like Figure 3 The diagram illustrates the structure of the node attention mechanism. The node attention mechanism of this invention first uses a fully connected layer to obtain adaptive weights for each node and then normalizes these weights. Simultaneously, to achieve more efficient node aggregation, it further assigns the learned weights to the feature vector of each node. Finally, the feature vectors of all nodes are summed and averaged to complete node aggregation. The node attention mechanism can automatically adjust node weights according to different levels of importance, making the model focus more on node features with high weights. This provides more accurate data input containing more temporal information for the subsequent Bi-GRU model, improving the Bi-GRU's ability to express temporal features from both micro and global perspectives.

[0081] Step S37: Using the Bi-GRU model, perform cascaded calculations through its forward gating units and inverse gating units respectively to obtain the final hidden output containing implicit timing information, expressed by the formula:

[0082] ;

[0083] ;

[0084] ;

[0085] Among them, GRU ( ) is a nonlinear transformation function. For the current input, This represents the output vector of the forward hidden state at time t-1. This represents the output vector of the reverse hidden state at time t-1.

[0086] To overcome the shortcomings of traditional recurrent neural networks in sequence learning, this step uses Bi-GRU to capture the forward and backward information of sequence data, thereby learning the patterns hidden in the sequence more deeply. Specifically, the node representations aggregated by the node attention mechanism are used as the input of Bi-GRU, and are cascaded through its forward gating units and backward gating units to obtain the hidden output containing the implicit temporal information.

[0087] In one embodiment, step S4 above: Input feature weighting layer, using channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. Specifically, it includes:

[0088] right A channel attention mechanism is used, employing Squeeze and Excitation operations. The Squeeze operation uses global average pooling to transform the feature maps of each channel into scalar representations, thereby capturing global information about the channel features. The Excitation operation uses a fully connected layer and an activation function to transform the global feature vector into a channel attention vector, enhancing the feature responses of important channels. The computation process is as follows:

[0089] ;

[0090] ;

[0091] in, For the Sigmoid function, MLP( () is a fully connected neural network. and This represents the MLP weights.

[0092] In the feature weighting layer, the CasSubTS model of this invention introduces a channel attention mechanism (Squeeze and Excitation Networks, SENet) to perform weighted feature fusion. The core idea of ​​SENet is to insert Squeeze and Excitation operations between the layers of the neural network. The Squeeze operation uses global average pooling to effectively transform the feature maps of each channel into scalar representations, thereby capturing the global information of the channel features. The Excitation operation, through a fully connected layer and an activation function, transforms the global feature vector into a channel attention vector. This attention vector is then used on the original feature map to enhance the feature responses of important channels. SENet can adaptively adjust the importance of each channel, making the network pay more attention to channel features with higher information content, while reducing the dependence on redundant and irrelevant features during feature fusion, further improving the expressive power of features and the model's predictive performance.

[0093] In one embodiment, step S5 above: The input prediction layer is used to predict the final macroscopic cascade increment, specifically including:

[0094] Step S51: ... Input the MLP to predict the final macroscopic cascade increment. :

[0095] ;

[0096] In this embodiment of the invention, the prediction layer selects an MLP as the predictor.

[0097] Step S52: Construct the loss function:

[0098] ;

[0099] Where N is the total number of information cascades, This represents the actual cascading increment. This represents the predicted cascading increment. It is the L2 regularization norm.

[0100] To verify the prediction performance and generalization ability of the CasSubTS model of this invention, a large social network dataset and a small graph dataset were selected for comparative experiments. During the experiments, the training set, validation set, and test set were divided into 70%, 10%, and 20% sets, respectively. Relevant statistical information for the two datasets is shown in Table 1.

[0101] Weibo Dataset: This dataset comes from Sina Weibo, China's largest social media platform. It contains all original Weibo posts published on the platform on June 1, 2016, and tracks all reposts for each post within 24 hours of its publication, totaling 119,313 posts. The dataset includes anonymized user data, repost paths, and timestamps. In this chapter's experiments, the observation time window T was set to 1 hour, 2 hours, and 3 hours, and Weibo posts published before 8 AM and after 6 PM were removed based on the cascading growth of information.

[0102] The Synthetic dataset was used to validate the transferability and adaptability of the model in this chapter on graph data of varying sizes. It is a scale-free network built using the Barabasi-Albert model, containing 880 nodes and 1992 edges. Initial nodes were randomly selected, and an independent cascade model and a linear threshold model were used to simulate information propagation paths. During data processing, noisy nodes with fewer than 10 cascades were removed, and the final prediction was made for the information cascade growth at the second step.

[0103] Table 1. Dataset Introduction

[0104]

[0105] The experiment selected MSLE, a commonly used indicator for macroscopic information cascade prediction tasks, as the model evaluation metric. A smaller MSLE value indicates more accurate prediction performance. Specifically, MSLE is calculated as follows:

[0106] ;

[0107] Where N is the total number of information cascades, This represents the actual cascading increment. This represents the predicted cascading increment.

[0108] To evaluate the performance of the CasSubTS model, this invention selected six classic and relatively new models in the field of information cascade prediction in comparative experiments. These models are described below:

[0109] Feature-Linear: The feature-linear model uses a linear regression model with L2 regularization for prediction. The input features of this model include the structural and temporal features of the cascade graph.

[0110] Feature-Deep: The deep feature model uses a two-layer fully connected neural network to complete the prediction, and its input features are the same as those of Feature-Linear.

[0111] DeepCas: The first end-to-end model to apply deep learning techniques to the cascade prediction problem. It performs graph sampling by generating multiple paths through random walks and uses GRU and attention mechanisms to learn representations of the cascade graph.

[0112] DeepHawkes: A classic model combining deep learning with point processes. This model obtains node sequences based on information diffusion paths, feeds the node vectors into a GRU to obtain sequence representations, and then uses a neural network to obtain prediction results after weighting and pooling operations. The forwarding contribution of these representation vectors is calculated using Hawkes' process, which incorporates user influence, self-stimulation, and time decay.

[0113] CasCN is a graph convolutional neural network-based model. It decomposes a cascaded graph into a series of subgraphs based on timestamps, resulting in a sequence of subgraphs, each containing structural and temporal information about the cascaded graph. CasCN uses dynamic multi-directional graph convolution to encode the features of the subgraphs and then uses LSTM to learn the temporal dependencies of the cascade.

[0114] CasSeqGCN: This model uses a graph convolutional neural network to process each subgraph independently, allowing parameters to be shared between different subgraphs. Then, it uses LSTM to learn the temporal features of the subgraph sequence, performs vector aggregation based on a dynamic routing algorithm, and finally uses MLP as the predictor.

[0115] The comparative experimental results are shown in Table 2, where the results represent the MSLE (Minimum Scale Leakage) evaluation metric. The table shows that the CasSubTS model achieves lower MSLEs on both datasets than other benchmark methods, demonstrating the superiority of the CasSubTS model. Compared to the best-performing benchmark method, CasSeqGCN, the CasSubTS model improves MSLE by approximately 3.4%, 3.7%, and 2.8% in prediction tasks with 1-hour, 2-hour, and 3-hour observation windows, respectively, on the Weibo dataset, and by approximately 30% on the Synthetic dataset, exhibiting superior predictive performance.

[0116] Table 2 Comparison of experimental results

[0117]

[0118] In the experimental results, the prediction performance of models based on manual feature engineering was generally inferior to that of deep learning models. This may be because manually constructed features cannot express the potential nonlinear relationships in cascade graphs and lack the ability to model features at different levels. Furthermore, these models heavily rely on prior knowledge, which further limits their generalization ability in prediction tasks involving large-scale cascade graphs. It is worth noting that although the prediction model based on feature engineering is relatively simple in its architecture, it outperformed the Feature-Deep model when predicting a 1-hour time period on the Weibo dataset. This indicates that, in some cases, high-quality feature engineering can still help improve the model's prediction performance.

[0119] The DeepCas model primarily learns node representations through random walks. This random approach may cause the model to overlook some important nodes or paths, and it also suffers from sampling bias, resulting in an imbalanced sequence of sampled nodes. Both of these issues degrade the model's predictive performance. The DeepHawkes model combines deep learning with Hawkes processes, fully considering user radiation, self-excitation effects, and time decay during information diffusion. However, its performance on the Synthetic dataset is poor, likely because it doesn't adequately consider other relevant features during prediction.

[0120] CasCN samples the subgraph based on the activation state of nodes, generating a cascade snapshot whenever a new node is activated. While this method can model temporal variations and improve the fine-grained nature of the prediction task, it may neglect the global structural information of the cascaded graph. The CasSeqGCN model, on the other hand, considers a combination of temporal and structural information, resulting in improved prediction performance to some extent, but its node aggregation stage still has room for improvement. The CasSubTS model, while using a graph attention network based on multi-head attention and Bi-GRU to extract structural and temporal features, employs a more effective node attention mechanism and feature fusion strategy, significantly improving prediction performance.

[0121] Example 2

[0122] like Figure 4 As shown, this embodiment of the invention provides an information cascading prediction system based on subgraphs, comprising the following modules:

[0123] Input module 61 is used to process the collected user-published information through the input layer to construct an information cascade graph G;

[0124] The subgraph sampling module 62 is used to input the information concatenated graph G into the subgraph sampling layer, divide it into several information concatenated subgraphs by setting different time steps, convert them into adjacency matrices, and aggregate the node features of the adjacency matrices to obtain the feature representation matrix B.

[0125] Feature learning module 63 is used to input the feature representation matrix B into the feature learning layer, and simultaneously learn structural features and temporal features to obtain a feature vector containing both structural and temporal features. ;

[0126] Feature weighting module 64 is used to... Input feature weighting layer, using channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. ;

[0127] Prediction module 65 is used to... Input the prediction layer to predict the final macroscopic cascade increment.

[0128] A subgraph-based information cascading prediction device includes one or more electronic devices, wherein the one or more electronic devices are used to implement the subgraph-based information cascading prediction method, system and device.

[0129] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors enable the one or more processors to implement a subgraph-based information cascading prediction method, system, and apparatus.

[0130] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement a subgraph-based information cascading prediction method, system, and apparatus.

[0131] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A subgraph-based information cascading prediction method, characterized in that, include: Step S1: Construct a deep learning information cascade prediction model CasSubTS, including: an input layer, a subgraph sampling layer, a feature learning layer, a feature weighting layer, and a prediction layer; the collected user-posted information is passed through the input layer to construct an information cascade graph G, specifically including: Step S11: User u publishes an initial message m at time t0. During the observation time, n other users v interact with this message. Then the message cascade C k The formal definition is; ； Step S12: Cascade information C k The corresponding C k The definition of an information cascade graph G is: ； in, This represents a node in an information cascade diagram, i.e., a user who participates in the dissemination of information; This represents the edges in an information cascade graph, i.e., the interaction relationships between users; Step S2: Input the information concatenated graph G into the subgraph sampling layer, divide it into several information concatenated subgraphs by setting different time steps, and convert them into adjacency matrices. Aggregate the node features of the adjacency matrix to obtain the feature representation matrix B. Step S3: Input the feature representation matrix B into the feature learning layer to learn structural features and temporal features simultaneously, thereby obtaining a feature vector that possesses both structural and temporal features. Specifically, it includes: Step S31: Perform a linear transformation on the node features of the feature representation matrix using a linear transformation matrix, mapping the original node features to different representation spaces to obtain the query vector for each attention head. Key vector Sum value vector : ； ； ； Where X is the node feature vector and H is the linear transformation matrix; Step S32: In order to evaluate node v i Its neighbor node v j The correlation and importance between nodes are calculated using an attention mechanism for each attention head k. i Its neighbor node v j Attention score e ij : ； in, It is a linear function. It is the parameter vector of the attention head k; Step S33: Calculate node v i Its neighbor node v j Attention weights: ； in, It is node v i The set of neighboring nodes, Represents any neighbor node and v i The correlation coefficient; Step S34: Use attention weights to perform weighted aggregation of the features of each node's neighbor nodes, and dynamically adjust the aggregated representation of each node. The calculation process is as follows: ； Step S35: Use k attention heads to independently learn the feature representation of each node, and sum the learned k sets of node representations and take the average to obtain the structural features. : ； Step S36: Based on the structural features, learn its temporal features. Attention mechanism for computation nodes: ； ； in, For projection parameters, These are the dimensions of structural features and temporal features, respectively; These are the node weights learned by the fully connected layer. Step S37: Using the Bi-GRU model, perform cascaded calculations through its forward gating units and inverse gating units respectively to obtain the final hidden output containing the implicit timing information, expressed by the formula: ； ； ； Among them, GRU ( ) is a nonlinear transformation function. For the current input, This represents the output vector of the forward hidden state at time t-1. This represents the output vector of the inverse hidden state at time t-1; Step S4: Input the feature weighting layer and use the channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. ; Step S5: ... Input the prediction layer to predict the final macroscopic cascade increment.

2. The information cascading prediction method based on subgraphs according to claim 1, characterized in that, Step S2: The information concatenated graph G is input into the subgraph sampling layer. By setting different time steps, it is divided into several information concatenated subgraphs and converted into an adjacency matrix. The adjacency matrix is ​​then subjected to node feature aggregation. The feature representation matrix B is obtained, specifically including: Step S21: Using a partial sampling method, by setting different time steps T, the information concatenation graph G is decomposed into a certain number of information concatenation subgraphs, and a set of information concatenation subgraph sequences is obtained: ； in, This represents a sequence of cascaded subgraphs obtained by sampling at different time steps; Step S22: For After concatenation, the corresponding adjacency matrix can be represented as: ； in, The adjacency matrix representation of a single subgraph sequence; Step S23: Perform node feature aggregation on the adjacency matrix to obtain the feature representation matrix, which can be represented as: ； in, is a learnable parameter; j is the column index of the adjacency matrix.

3. The information cascading prediction method based on subgraphs according to claim 2, characterized in that, Step S4: ... Input the feature weighting layer and use the channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. Specifically, it includes: right A channel attention mechanism is used, employing Squeeze and Excitation operations. The Squeeze operation uses global average pooling to transform the feature maps of each channel into scalar representations, thereby capturing global information about the channel features. The Excitation operation uses a fully connected layer and an activation function to transform the global feature vector into a channel attention vector, enhancing the feature responses of important channels. The computation process is as follows: ； ； in, For the Sigmoid function, MLP( () is a fully connected neural network. and This represents the MLP weights.

4. The information cascading prediction method based on subgraphs according to claim 3, characterized in that, Step S5: ... The input to the prediction layer to predict the final macroscopic cascade increment specifically includes: Step S51: ... Input the MLP to predict the final macroscopic cascade increment. : ； Step S52: Construct the loss function: ； Where N is the total number of information cascades, This represents the actual cascading increment. This represents the predicted cascading increment. It is the L2 regularization norm.

5. A subgraph-based information cascading prediction system, characterized in that, Includes the following modules: The input module is used to construct the deep learning information cascade prediction model CasSubTS, including: an input layer, a subgraph sampling layer, a feature learning layer, a feature weighting layer, and a prediction layer; the collected user-posted information is processed through the input layer to construct an information cascade graph G, specifically including: Step S11: User u publishes an initial message m at time t0. During the observation time, n other users v interact with this message. Then the message cascade C k The formal definition is; ； Step S12: Cascade information C k The corresponding C k The definition of an information cascade graph G is: ； in, This represents a node in an information cascade diagram, i.e., a user who participates in the dissemination of information; This represents the edges in an information cascade graph, i.e., the interaction relationships between users; The subgraph sampling module is used to input the information concatenated graph G into the subgraph sampling layer, divide it into several information concatenated subgraphs by setting different time steps, convert them into adjacency matrices, and aggregate the node features of the adjacency matrix to obtain the feature representation matrix B. The feature learning module is used to input the feature representation matrix B into the feature learning layer, and simultaneously learn structural features and temporal features to obtain a feature vector containing both structural and temporal features. Specifically, it includes: Step S31: Perform a linear transformation on the node features of the feature representation matrix using a linear transformation matrix, mapping the original node features to different representation spaces to obtain the query vector for each attention head. Key vector Sum value vector : ； ； ； Where X is the node feature vector and H is the linear transformation matrix; Step S32: In order to evaluate node v i Its neighbor node v j The correlation and importance between nodes are calculated using an attention mechanism for each attention head k. i Its neighbor node v j Attention score e ij : ； in, It is a linear function. It is the parameter vector of the attention head k; Step S33: Calculate node v i Its neighbor node v j Attention weights: ； in, It is node v i The set of neighboring nodes, Represents any neighbor node and v i The correlation coefficient; Step S34: Use attention weights to perform weighted aggregation of the features of each node's neighbor nodes, and dynamically adjust the aggregated representation of each node. The calculation process is as follows: ； Step S35: Use k attention heads to independently learn the feature representation of each node, and sum the learned k sets of node representations and take the average to obtain the structural features. : ； Step S36: Based on the structural features, learn its temporal features. Attention mechanism for computation nodes: ； ； in, For projection parameters, These are the dimensions of structural features and temporal features, respectively; These are the node weights learned by the fully connected layer. Step S37: Using the Bi-GRU model, perform cascaded calculations through its forward gating units and inverse gating units respectively to obtain the final hidden output containing the implicit timing information, expressed by the formula: ； ； ； Among them, GRU ( ) is a nonlinear transformation function. For the current input, This represents the output vector of the forward hidden state at time t-1. This represents the output vector of the inverse hidden state at time t-1; Feature weighting module, used to... Input the feature weighting layer and use the channel attention mechanism to... Perform weighted fusion to obtain the weighted feature vector. ; The prediction module is used to... Input the prediction layer to predict the final macroscopic cascade increment.

6. A subgraph-based information cascade prediction device, characterized in that, It includes one or more electronic devices, wherein the one or more electronic devices are used to implement the method of any one of claims 1 to 4.

7. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to implement the method described in any one of claims 1 to 4.

Images