Sub-graph-based information cascade prediction method and system

Abstract

The invention provides an information cascading prediction method and system based on a subgraph, and belongs to the technical field of social network analysis and neural networks, and the method comprises the steps: S1, constructing a deep learning information cascading prediction model CasSubTS, enabling collected user published information to pass through an input layer, and constructing an information cascading graph G; s2, inputting the G into a sub-graph sampling layer, dividing the G into a plurality of information cascade sub-graphs according to different time steps, converting the information cascade sub-graphs into adjacency matrixes, and performing node feature aggregation on the adjacency matrixes to obtain a feature representation matrix B; s3, inputting the B into a feature learning layer to obtain a feature vector # imgabs0 # with a structural feature and a time sequence feature; s4, inputting the # imgabs1 # into a feature weighting layer, and performing weighted fusion on the # imgabs2 # by using a channel attention mechanism to obtain a weighted feature vector # imgabs3 #; and S5, inputting # imgabs4 # into a prediction layer to predict a final macroscopic cascade increment. According to the method, the information cascading in the social network is effectively predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of social network analysis and neural network, and in particular relates to a subgraph-based information cascade prediction method and system. Background Art

[0002] With the continuous innovation of Internet technology and the widespread popularization of mobile smart terminals, people's social behavior has begun to be closely integrated with digital technology, and society has entered the era of digital social interaction. Online social platforms represented by Weibo, WeChat, Douyin and Facebook have broken through the time and space limitations of information interaction, and are gradually replacing traditional information media and becoming the mainstream channel for the public to release and receive information. The social network built on the Internet has unprecedentedly improved the breadth and depth of the spread of various types of information, and has become the main carrier of information dissemination in today's society, profoundly affecting the changes in social culture and public psychology.

[0003] A social network is defined as a network system formed by social relationships between individual participants. In a social network, individuals are regarded as nodes, representing individuals, organizations or other entities participating in social behaviors, and the edges connecting nodes represent the social relationships between individuals. The widespread application of information technology and social platforms has changed the model and rules of traditional social interaction, extending the way of establishing social relationships between individuals to online, such as following, liking and forwarding. 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 features of online social networks is user-generated content (UGC), which usually includes multiple modalities, such as text, pictures and videos. UGC is the mainstream way of information production and reception at present. It can reduce the cost of content creation for users, increase the diversity of content and the immediacy of interaction. Because of this feature, online social networks have attracted a large number of user groups to participate in them, and the scale of the network is also constantly growing.

[0004] Information cascade is an effect caused by the spread of information through social networks. It is generally composed of the information being spread, the participants in the information spread, and the path trajectory of the information spread. The information cascade effect can accelerate the spread of information in the network and reshape the views and attitudes of individual participants, thereby affecting the overall public opinion trend of society. Given its importance, academia and the business community have been paying close attention to research related to information cascade prediction. In practice, information cascades can be predicted from both macro and micro perspectives. Macro information cascade prediction focuses on the overall trend of information evolution, and in specific tasks is usually defined as the prediction of the information cascade increment of a certain information in a social network within a specific time period. Micro information cascade prediction focuses more on exploring the local information propagation path and rules, and predicting the user individuals who will join the information cascade process at the next moment. However, there are many difficulties and challenges in predicting information cascades in social networks:

[0005] (1) The information propagation path in social networks is easily affected by user interactions and has strong randomness, which further aggravates 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 may be affected by time factors and may experience large fluctuations, such as explosive growth or linear decline.

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

[0008] In order to solve the above technical problems, the present invention provides a subgraph-based information cascade 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; the collected user-published information passes through the input layer to construct an information cascade graph G;

[0010] Step S2: input the information cascade graph G into the subgraph sampling layer, divide it into several information cascade subgraphs by setting time steps of different sizes, and convert it into an adjacency matrix, perform node feature aggregation on the adjacency matrix, and obtain a feature representation matrix B;

[0011] Step S3: Input the feature representation matrix B into the feature learning layer, learn the structural features and timing features at the same time, and obtain a feature vector with structural features and timing features ;

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

[0013] Step S5: Input to the prediction layer to predict the final macro 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, and incorporates the node's out-degree and in-degree as directional information into the structural features, in order to learn the structure of the cascade graph more comprehensively. At the same time, the introduction of the multi-head attention mechanism also balances the robustness and flexibility of the model.

[0016] 2. In the learning of time series features, the method of the present invention does not adopt the approach of directly using recurrent neural networks to extract time series information in the prior art, but takes into account the differences in the time series information implied by each node in different time periods, and designs a node attention mechanism. After aggregating nodes containing more important information, a recurrent neural network is introduced to perform time series dependency learning.

[0017] 3. Regarding the weighted fusion problem of the weights of structural and temporal features, the present invention uses a channel attention mechanism with stronger feature-focusing capabilities to perform weighted fusion of features. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a subgraph-based information cascade prediction method flow chart of the present invention;

[0019] Figure 2 It is a structural diagram of the CasSubTS model;

[0020] Figure 3 It is a schematic diagram of the structure of the node attention mechanism;

[0021] Figure 4 It is a structural block diagram of a subgraph-based information cascade prediction system of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Embodiment 1

[0024] like Figure 1 As shown, an information cascade prediction method based on a subgraph provided by an embodiment of the present invention comprises 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-published information passes through the input layer to construct an information cascade graph G;

[0026] Step S2: Input the information cascade graph G into the subgraph sampling layer, divide it into several information cascade subgraphs by setting time steps of different sizes, and convert it into an adjacency matrix, aggregate node features of the adjacency matrix, and obtain the feature representation matrix B;

[0027] Step S3: Input the feature representation matrix B into the feature learning layer, learn the structural features and timing features at the same time, and obtain the feature vector with structural features and timing features ;

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

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

[0030] In one embodiment, the above step S1: 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-published 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, sub-graph sampling layer, feature learning layer, feature weighting layer and prediction layer.

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

[0033] Step S11: User u is in t 0 An initial message m is released at time t. During the observation time, there are n other users v who interact with this message. Then the information cascade C k The formal definition of is;

[0034] ;

[0035] Step S12: Cascade information C k The corresponding C k The information cascade graph G is defined as:

[0036] ;

[0037] in, Represents the nodes in the information cascade graph, i.e., the users who participate in information diffusion; Represents the edges in the information cascade graph, that is, the interaction relationship between users.

[0038] Time factors profoundly affect the scale and propagation path of information cascades. On the one hand, the information cascade graph formed by a certain initial information m may attract a large number of nodes to join continuously over time, triggering a large-scale cascade effect and generating a large-scale information cascade graph. Large-scale information cascade graphs usually increase the inconvenience of graph storage and calculation, and limit the data range applicable to the model. On the other hand, the order and time intervals of different nodes joining the information propagation process in the information cascade graph are different. Most current methods use the method of aggregating node sequences to represent the cascade, but this method ignores the dynamic changes in the temporal nature of nodes during information propagation, cannot represent the real propagation phenomenon, and also causes model fitting errors.

[0039] In order to reduce the impact of the above two problems on the prediction model, a series of sampling methods have been proposed. Among them, the two most influential methods are the random walk-based sampling method and the propagation path-based sampling method. The classic model using random walk sampling is DeepCas, which obtains different node sequences by random walks on the network and then completes the construction of the subgraph. However, this method fundamentally ignores the global and local dynamics of cascade changes and has great limitations. The propagation path-based sampling method collects sequences for the propagation path of each node in the cascade graph. In this way, the lengths of the collected sequences may not be unified, which leads to data sparsity.

[0040] In order to fully consider the temporal nature of cascade evolution in the information cascade subgraph sampling link, the present invention proposes the following steps of information cascade subgraph sampling method based on time step, and performs subgraph sampling according to the time information of each node in the observed information cascade graph. Furthermore, in order to ensure the efficiency of the training process, the present invention uses a partial sampling method to decompose a certain number of cascade subgraphs by setting time steps T of different sizes, and obtain the sequence of each subgraph. The number of subgraphs N finally sampled can be obtained by the following formula:

[0041] ;

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

[0043] In one embodiment, the above step S2: inputting the information cascade graph G into the subgraph sampling layer, dividing it into a number of information cascade subgraphs by setting time steps of different sizes, and converting it into an adjacency matrix, performing node feature aggregation on the adjacency matrix, and obtaining a feature representation matrix B, specifically includes:

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

[0045] ;

[0046] in, Represents a sequence of information cascade subgraphs obtained by sampling at different time steps;

[0047] Step S22: After cascade representation, the corresponding adjacency matrix can be expressed as:

[0048] ;

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

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

[0051] ;

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

[0053] Structural features and temporal features are the two most important features in the information cascade prediction process. In order to obtain effective structural and temporal feature representations, the CasSubTS model of the present invention adopts a parallel feature learning method, using MH-GAT and Bi-GRU to simultaneously learn the structural features and temporal features of the cascade graph. The parallel feature learning process helps to reduce the time consumption of data processing and model training, improves the utilization efficiency of data information, and enhances the generalization ability of the model.

[0054] In one embodiment, the above step S3: input the feature representation matrix B into the feature learning layer, and learn the structural features and the timing features at the same time to obtain the characteristic vector with the structural features and the timing features , specifically including:

[0055] The following steps S31~S35 start from the perspective of spatial convolution, and for the feature representation matrix of the input information cascade graph, the MH-GAT model is used to perform convolution operations on the feature representation matrix, dynamically assign different attention weights to each node, and realize weighted aggregation of node features, and finally obtain an updated vector representation containing node structure information, that is, structural features. Before the MH-GAT model is trained, considering the multi-scale characteristics of node structure information, in order to more comprehensively learn the structural information of the cascade graph and improve the prediction performance of the model, the present invention adds the direction information of the node to the structural features. This information can be represented by the degree of the node. Generally speaking, in the information cascade graph, nodes with higher height values ​​often have more social relationship connections, so their role in information dissemination is more critical, and their impact on the model prediction results is more significant.

[0056] Take node v i The number of edges with endpoints is called v i The degree is expressed as follows:

[0057] .

[0058] Step S31: Use the linear transformation matrix to linearly transform the node features of the feature representation matrix, map the original node features to different representation spaces, and obtain the query vector of each attention head , key vector Sum value vector :

[0059] ;

[0060] ;

[0061] ;

[0062] Among them, X is the node feature vector, H is the linear transformation matrix;

[0063] Step S32: To evaluate the node v i Its neighbor node v j For each attention head k, the attention mechanism is used to calculate the correlation and importance between nodes v i Its neighbor node v j The attention score e ij :

[0064] ;

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

[0066] Step S33: Calculate node v i Its neighbor node v j The attention weights are:

[0067] ;

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

[0069] Step S34: Use the attention weight to perform weighted aggregation on the neighbor node features of each node, and dynamically adjust the aggregate representation of each node. The calculation process is as follows:

[0070] ;

[0071] In order to cope with the interference of noisy data, avoid overfitting, and make the training process more robust, the 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 sums and averages the learned k groups of node representations to obtain the structural features :

[0073] ;

[0074] In the structural feature learning of the above steps, the CasSubTS model uses a graph attention network that integrates a multi-head attention mechanism to extract features from the complex structure of the cascade graph, thereby obtaining node structural features containing cascade topological structure information. .

[0075] For learning of temporal features, in order to improve the learning ability of implicit temporal information in information cascade graphs and reduce information loss in the process of learning temporal features, the CasSubTS model first uses the designed node attention mechanism to adaptively adjust node weights and aggregate representations of different nodes. Then, the Bi-GRU model, a variant of the recurrent neural network, is used to model the temporal dependencies in the process of information propagation.

[0076] Step S36: Based on the structural features, learn its temporal features , calculation node attention mechanism:

[0077] ;

[0078] ;

[0079] in, is the projection parameter, They are the dimensions of structural features and temporal features respectively; The node weights learned by the fully connected layer;

[0080] like Figure 3 As shown, a schematic diagram of the structure of the node attention mechanism. The node attention mechanism of the present invention first uses a fully connected layer to obtain the adaptive weight of each node and normalizes the weight. At the same time, in order to achieve node aggregation more efficiently, it further assigns the learned weight to the feature vector of each node. Finally, the feature vectors of all nodes are summed and averaged to complete the aggregation of the nodes. The node attention mechanism can automatically adjust the node weights according to different levels of importance, so that the model pays more attention to node features with high weights, and provides more accurate data input containing more time series information for the subsequent Bi-GRU model, thereby improving the Bi-GRU's ability to express time series features from both micro and global levels.

[0081] Step S37: Use the Bi-GRU model to perform serial calculations through the forward gating unit and the reverse gating unit therein to obtain the final hidden output of the implicit timing information, which is expressed as follows:

[0082] ;

[0083] ;

[0084] ;

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

[0086] In order to overcome the shortcomings of traditional recurrent neural networks in the sequence learning process, this step uses Bi-GRU to capture the forward and backward information of the sequence data, so as to learn the patterns hidden in the sequence more deeply. Specifically, the node representation after the node attention mechanism is aggregated as the input of Bi-GRU, and the forward gating unit and the reverse gating unit are respectively connected in series to obtain the hidden output with implicit timing information.

[0087] In one embodiment, the above step S4: Input feature weighted layer, use channel attention mechanism to Perform weighted fusion to obtain the weighted feature vector , specifically including:

[0088] right Use the channel attention mechanism to perform Squeeze and Excitation operations on it. The Squeeze operation uses global average pooling to convert the feature maps of each channel into a scalar representation to capture the global information of the channel features. The Excitation operation converts the global feature vector into a channel attention vector through a fully connected layer and an activation function to enhance the feature response of important channels. The calculation process is as follows:

[0089] ;

[0090] ;

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

[0092] In the feature weighting layer, the CasSubTS model of the present invention introduces a channel attention mechanism (Squeeze and Excitation Networks, SENet) to perform weighted fusion of features. 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 convert the feature maps of each channel into a scalar representation, thereby capturing the global information of the channel features. The Excitation operation converts the global feature vector into a channel attention vector through a fully connected layer and an activation function. This attention vector will be used on the original feature map to enhance the feature response of important channels. SE-Net can adaptively adjust the importance of each channel, so that the network pays more attention to channel features with higher information content, while reducing the dependence on redundant and irrelevant features during feature fusion, further improving the expressiveness of features and model prediction performance.

[0093] In one embodiment, the above step S5: Input prediction layer to predict the final macro cascade increment, including:

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

[0095] ;

[0096] The prediction layer of the embodiment of the present invention selects MLP as the predictor.

[0097] Step S52: construct a loss function:

[0098] ;

[0099] Where N is the total number of information cascades, represents the real cascade increment, represents the predicted cascade increment, is the L2 regularization norm.

[0100] In order to verify the prediction effect and generalization performance of the CasSubTS model of the present invention, a large social network dataset and a small-scale graph dataset were selected for comparative experiments. During the experiment, the division ratios of the training set, validation set, and test set were 70%, 10%, and 20%, respectively. The relevant data statistics of the two datasets are shown in Table 1.

[0101] Weibo Dataset: This dataset comes from Sina Weibo, the largest social media platform in China. It contains all original Weibo posts published on the platform on June 1, 2016, and tracks all forwarding information of each Weibo post within 24 hours after it is published, totaling 119,313 posts. The dataset covers desensitized user basic data, Weibo forwarding path, and timestamp information. In this chapter, the experiment sets the observation time window T to 1 hour, 2 hours, and 3 hours respectively, and removes Weibo posts published before 8 am and after 6 pm based on the growth of information cascades.

[0102] Synthetic dataset: This dataset is used to verify the transferability and adaptability of the model in this chapter on graphs of different sizes. It is a scale-free network constructed by the Barabasi-Albert model, with 880 nodes and 1992 edges. The initial nodes are randomly selected, and the independent cascade model and linear threshold model are used to simulate the information propagation path. During the data processing, the experiment removed the noise nodes with less than 10 cascades, and finally predicted the information cascade growth size at the second step.

[0103] Table 1 Dataset introduction

[0104] The experiment selected the commonly used indicator MSLE of macro information cascade prediction task as the evaluation indicator of the model. The smaller the value of MSLE, the more accurate the prediction effect of the model. Specifically, the calculation method of MSLE is as follows:

[0105] ;

[0106] Where N is the total number of information cascades, represents the real cascade increment, Represents the predicted cascade increment.

[0107] In order to evaluate the performance of the CasSubTS model, in the comparative experiment, the present invention selected six classic models and newer models in the field of information cascade prediction. The introduction of these models is as follows:

[0108] 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 features and time features of the cascade graph.

[0109] 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 Feature-Linear.

[0110] DeepCas: The first end-to-end model that applies deep learning techniques to the cascade prediction problem. It samples graphs by random walks to generate multiple paths, and uses GRU and attention mechanisms to learn representations of cascade graphs.

[0111] DeepHawkes: A classic model that combines deep learning with point processes. The model obtains node sequences based on the path of information diffusion, and sends node vectors to GRU to obtain sequence representations. After weighting and pooling operations, a neural network is used to obtain prediction results. Among them, the forwarding contribution of these representation vectors is calculated through the Hawkes process of user influence, self-stimulation, and time decay.

[0112] CasCN: It is a model based on graph convolutional neural network. It decomposes the cascade graph into a series of subgraphs based on timestamps to obtain subgraph sequences, each of which contains the structure and time information of the cascade graph. For this information, CasCN uses dynamic multi-directional graph convolution to encode the features of the subgraphs, and then uses LSTM to learn the time-dependent characteristics of the cascade.

[0113] CasSeqGCN: This model uses a graph convolutional neural network to process each subgraph independently, so that parameters are shared between different subgraphs. It then uses LSTM to learn the temporal features of the subgraph sequence and performs vector aggregation based on a dynamic routing algorithm. Finally, MLP is used as a predictor.

[0114] The results of the comparative experiments are shown in Table 2. The results in the table are all the values ​​of the evaluation index MSLE. As can be seen from the table, the MSLE of the CasSubTS model on both datasets is lower than that of other benchmark methods, which proves the superiority of the CasSubTS model. Compared with the CasSeqGCN model with the best performance in the benchmark method, the MSLE of the CasSubTS model on the Weibo dataset is improved by about 3.4%, 3.7% and 2.8% in the 1-hour, 2-hour and 3-hour observation window prediction tasks, respectively, and by about 30% on the Synthetic dataset, showing good prediction performance.

[0115] Table 2 Comparative experimental results

[0116] In the experimental results, the prediction effect of the model based on artificial feature engineering is generally inferior to that of the deep learning model. This may be because the artificially constructed features cannot express the potential nonlinear relationship in the cascade graph and lack the ability to model features at different levels. At the same time, this type of model is highly dependent on artificial prior knowledge. In some large-scale cascade graph prediction tasks, the generalization of the model will be further limited. It is worth noting that although the prediction model based on the feature process is relatively simple in architectural design, its performance surpasses the Feature-Deep model when the prediction time period is 1 hour in the Weibo dataset. This shows that in some cases, high-quality feature engineering still helps to improve the prediction performance of the model.

[0117] The DeepCas model mainly learns node representations through random walks. This random approach may cause the model to ignore some important nodes or paths. There will also be sampling bias problems, which will cause the sampled node sequence to be unbalanced. All of these problems will reduce the prediction performance of the model. The DeepHawkes model combines deep learning with the Hawkes process, and fully considers user radiation, self-excitation effects, and time decay in the process of information diffusion. However, its performance on the Synthetic dataset is poor, and the main reason may be that other relevant features are not fully considered in the prediction process.

[0118] CasCN samples subgraphs based on the activation status of nodes, and a cascade snapshot is generated every time a new node is activated. Although this method can model temporal changes and improve the granularity of prediction tasks, it may ignore the global structural information of the cascade graph. The CasSeqGCN model considers the combination of temporal and structural information, which has improved the prediction performance to a certain extent, but the design of its node aggregation link still has room for improvement. For the CasSubTS model, while using the graph attention network based on the multi-head attention mechanism and Bi-GRU to extract structural and temporal features, a more effective node attention mechanism and feature fusion strategy are adopted, which greatly improves the prediction performance.

[0119] Embodiment 2

[0120] like Figure 4 As shown, an embodiment of the present invention provides a subgraph-based information cascade prediction system, comprising the following modules:

[0121] An input module 61 is used to construct an information cascade graph G by passing the collected user published information through an input layer;

[0122] The subgraph sampling module 62 is used to input the information cascade graph G into the subgraph sampling layer, divide it into a number of information cascade subgraphs by setting time steps of different sizes, and convert it into an adjacency matrix, perform node feature aggregation on the adjacency matrix, and obtain a feature representation matrix B;

[0123] The feature learning module 63 is used to input the feature representation matrix B into the feature learning layer, and learn the structural features and the timing features at the same time to obtain a feature vector having the structural features and the timing features. ;

[0124] The feature weighting module 64 is used to Input feature weighted layer, use channel attention mechanism to Perform weighted fusion to obtain the weighted feature vector ;

[0125] Prediction module 65, used to Input to the prediction layer to predict the final macro cascade increment.

[0126] A subgraph-based information cascade prediction device includes one or more electronic devices, wherein the one or more electronic devices are used to implement a subgraph-based information cascade prediction method, system and device.

[0127] An electronic device includes: 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 implement a subgraph-based information cascade prediction method, system and device.

[0128] A computer-readable storage medium stores executable instructions, which, when executed by a processor, enable the processor to implement a subgraph-based information cascade prediction method, system and device.

[0129] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be 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 the present application. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features applied herein.

Claims

1. A subgraph-based information cascade 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-published information passes through the input layer to construct an information cascade graph G; Step S2: input the information cascade graph G into the subgraph sampling layer, divide it into a number of information cascade subgraphs by setting time steps of different sizes, and convert it into an adjacency matrix, perform node feature aggregation on the adjacency matrix, and obtain a feature representation matrix B; Step S3: Input the feature representation matrix B into the feature learning layer, learn the structural features and timing features at the same time, and obtain a feature vector with structural features and timing features ; Step S4: Input the feature weighted layer and use the channel attention mechanism to Perform weighted fusion to obtain the weighted feature vector ; Step S5: Input to the prediction layer to predict the final macro cascade increment.

2. The information cascade prediction method based on subgraphs according to claim 1, characterized in that: In step S1, the collected user published information is passed through the input layer to construct an information cascade graph G, which specifically includes: Step S11: User u publishes an initial message m at time t0. During the observation time, there are n other users v who interact with this message. Then the information cascade C k The formal definition of is; ； Step S12: Cascade information C k Corresponding C k The information cascade graph G is defined as: ； in, Represents the nodes in the information cascade graph, i.e., the users who participate in information diffusion; Represents the edges in the information cascade graph, that is, the interaction relationship between users.

3. The information cascade prediction method based on subgraphs according to claim 2, characterized in that: The step S2: inputting the information cascade graph G into the subgraph sampling layer, dividing it into a number of information cascade subgraphs by setting time steps of different sizes, and converting it into an adjacency matrix, and performing node feature aggregation on the adjacency matrix, The feature representation matrix B is obtained, which specifically includes: Step S21: Using a partial sampling method, by setting time steps T of different sizes, the information cascade graph G is decomposed into a certain number of information cascade subgraphs, and a set of information cascade subgraph sequences is obtained: ； in, Represents a sequence of information cascade subgraphs obtained by sampling at different time steps; Step S22: After cascade representation, the corresponding adjacency matrix can be expressed as: ； in, is the adjacency matrix representation of a single subgraph sequence; Step S23: Perform node feature aggregation on the adjacency matrix to obtain a feature representation matrix which can be expressed as: ； in, is a learnable parameter; j is the column index of the adjacency matrix.

4. The information cascade prediction method based on subgraphs according to claim 3, characterized in that: Step S3: Input the feature representation matrix B into the feature learning layer, learn the structural features and the time series features at the same time, and obtain a feature vector having the structural features and the time series features. , specifically including: Step S31: Use a linear transformation matrix to linearly transform the node features of the feature representation matrix, map the original node features to different representation spaces, and obtain the query vector of each attention head , key vector Sum value vector : ； ； ； Among them, X is the node feature vector, H is the linear transformation matrix; Step S32: To evaluate the node v i Its neighbor node v j For each attention head k, the attention mechanism is used to calculate the correlation and importance between nodes v i Its neighbor node v j The attention score e ij : ； in, is a linear function, is the parameter vector of attention head k; Step S33: Calculate node v i Its neighbor node v j The attention weights are: ； in, is node v i The set of neighbor nodes of Represents any neighbor node and v i The correlation coefficient of Step S34: Use the attention weight to perform weighted aggregation on the neighbor node features of each node, and dynamically adjust the aggregate 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 and average the learned k groups of node representations to obtain the structural features : ； Step S36: Based on the structural features, learn its timing features , calculation node attention mechanism: ； ； in, is the projection parameter, They are the dimensions of structural features and temporal features respectively; The node weights learned by the fully connected layer; Step S37: Use the Bi-GRU model to perform serial calculations through the forward gating unit and the reverse gating unit therein to obtain the final hidden output of the implicit timing information, which is expressed as follows: ； ； ； Among them, GRU( ) is a nonlinear transformation function, is the current input, represents the forward hidden state output vector at time t-1, Represents the reverse hidden state output vector at time t-1.

5. The subgraph-based information cascade prediction method according to claim 4, characterized in that: Step S4: Input the feature weighted layer and use the channel attention mechanism to Perform weighted fusion to obtain the weighted feature vector , specifically including: right Use the channel attention mechanism to perform Squeeze and Excitation operations on it. The Squeeze operation uses global average pooling to convert the feature maps of each channel into a scalar representation to capture the global information of the channel features. The Excitation operation converts the global feature vector into a channel attention vector through a fully connected layer and an activation function to enhance the feature response of important channels. The calculation process is as follows: ； ； in, is the Sigmoid function, MLP( ) is a fully connected neural network, and Represents the MLP weights.

6. The subgraph-based information cascade prediction method according to claim 5, characterized in that: Step S5: Input the prediction layer to predict the final macro cascade increment, including: Step S51: Input to MLP to predict the final macro cascade increment : ； Step S52: construct a loss function: ； Where N is the total number of information cascades, represents the real cascade increment, represents the predicted cascade increment, is the L2 regularization norm.

7. A subgraph-based information cascade prediction system, characterized in that: Includes the following modules: An input module, used to construct an information cascade graph G by passing the collected user published information through the input layer; A subgraph sampling module is used to input the information cascade graph G into the subgraph sampling layer, divide it into a number of information cascade subgraphs by setting time steps of different sizes, convert it into an adjacency matrix, perform node feature aggregation on the adjacency matrix, and obtain a feature representation matrix B; The feature learning module is used to input the feature representation matrix B into the feature learning layer, and learn the structural features and timing features at the same time to obtain a characteristic vector with structural features and timing features. ; Feature weighting module is used to Input the feature weighted layer and use the channel attention mechanism to Perform weighted fusion to obtain the weighted feature vector ; The prediction module is used to Input to the prediction layer to predict the final macro cascade increment.

8. A subgraph-based information cascade prediction device, characterized in that: The method comprises one or more electronic devices, wherein the one or more electronic devices are used to implement the method according to any one of claims 1 to 6.

9. 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 implement the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, and when the instructions are executed by a processor, the processor implements the method according to any one of claims 1 to 6.

Images