A social network information propagation influence prediction method based on information cascading

Abstract

The application discloses a social network information propagation influence prediction method based on information cascade, which comprises the following steps: 1, storing user node information by using a directed graph; 2, calculating the structural characteristic information of each user node; 3, extracting an observation sequence and processing data; 4, using diffusion random sampling to model cascade characteristics; 5, using a positive sine position vector to encode time information; 6, converting a time matrix into an input vector of an encoder; 7, performing self-attention transformation in the encoder layer; 8, alternately stacking a GAT layer and the encoder layer while fusing time characteristics and structural characteristics to perform feature transformation; 9, performing hierarchical Drop operation; 10, adding a cascade graph, an aggregation vector and a time matrix; 11, repeating steps 7-9; 12, obtaining a final aggregation vector from the encoder; 13, performing cascade growth prediction; and 14, processing data and reducing prediction error.

Description

Technical Field

[0001] This invention relates to the field of cascading growth prediction, and in particular to a method for predicting the influence of social network information dissemination based on information cascading. Background Technology

[0002] Online social networking platforms provide a wide and rapid channel for information dissemination, and multimedia technologies enable information to carry more content. This also makes online social networking platforms a breeding ground for the large-scale spread of malicious information, significantly amplifying its speed and scope of dissemination. Research shows that rumors, fake news, and other misinformation are more likely to attract attention and spread on social media. If we can identify potentially widespread and influential trending information in advance by predicting the scale of information growth and monitor it in real time, we can take timely countermeasures to prevent and resolve crises, minimizing the negative impact of the rapid spread of harmful information with fewer resources.

[0003] Information in social networks spreads in cascading fashion, with the length of the cascade typically representing its influence; the longer the cascade, the greater the influence. In recent years, deep learning technology has demonstrated significant superiority in end-to-end cascade popularity prediction, automatically extracting useful information from cascaded data. Some research methods represent information cascades as sequences of multiple user nodes and input them into recurrent neural network (RNN) models to better uncover potential diffusion patterns. Furthermore, some researchers represent information cascades as cascaded graphs or social networks and apply graph neural network (GNN) models to extract structural features during the early propagation of information cascades.

[0004] However, current research methods have the following problems:

[0005] (1) RNNs can only learn temporal features from the input order of concatenated data sequences, ignoring more temporal feature details that may exist in the intermediate time series. However, the time intervals and propagation speed per unit time in the information propagation process have been proven to be crucial for the information propagation prediction problem;

[0006] (2) The length of the information concatenation within the same time period follows a power-law distribution, resulting in significant differences in the length of the concatenation. When recurrent neural networks encounter excessively long sequence inputs, they experience long-term dependency problems, and the vanishing and exploding gradient problems also prevent the network from effectively updating its parameters.

[0007] (3) In early cascade prediction tasks, due to the short observation time and the power-law distribution of the number of people affected by the information, the observed cascade sequences were mainly short sequences. However, when constructing cascade graphs based on short cascade sequences, due to the small number of graph nodes, it is difficult to extract features from the graph structure sufficient to distinguish the predicted cascade lengths, resulting in insufficient information in the graph structure to distinguish cascades of different lengths.

[0008] These issues all contribute to the low accuracy of current methods in predicting the impact of information dissemination on social networks. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention provides a method for predicting the influence of information dissemination in social networks based on information cascading. It improves existing methods for extracting temporal and structural features. First, it uses a positional encoding function and the encoder structure in Transformer to replace the traditional RNN structure for temporal feature extraction. At the same time, it uses a data augmentation method based on diffusion random sampling to extract the structural features of the cascading graph. By combining these two feature extraction methods, a model for predicting the influence of information dissemination in social networks is proposed to predict the growth of early information cascading.

[0010] To achieve the aforementioned objectives of the invention, the technical solution adopted to solve its technical problems is as follows:

[0011] A method for predicting the influence of social network information dissemination based on information cascading includes the following steps:

[0012] Step 1: Use a directed graph to store the extracted user nodes and the relationships between users;

[0013] Step 2: Calculate the structural feature information of each user node in the user's global social relationship network;

[0014] Step 3: Extract the observation sequence and process the data according to the definition of the information cascade prediction problem;

[0015] Step 4: Use diffuse random sampling to introduce a portion of global features to model cascaded features;

[0016] Step 5: Encode the time information using sine and cosine position vectors to obtain the time matrix;

[0017] Step 6: Convert the time matrix into the encoder's input vector;

[0018] Step 7: Add the second-order diffused random sampling concatenation graph and the time vector together, and input the result of the addition and the aggregation vector into the encoder layer for self-attention transformation;

[0019] Step 8: Alternately stack the GAT layer and encoder layer, and simultaneously fuse temporal and structural features to perform feature transformation;

[0020] Step 9: Take each multi-head self-attention vector and the aggregate vector as input, and the new aggregate vector as output;

[0021] Step 10: Output the aggregation vector and diffusion cascade graph;

[0022] Step 11: Use GAT to extract structural features to obtain a new diffusion cascade diagram;

[0023] Step 12: Perform a layered Drop operation;

[0024] Step 13: Add the cascaded graph, aggregation vector, and time matrix together;

[0025] Step 14: Repeat steps 7 to 12, using the obtained concatenated graph, aggregation vector, and initial time matrix as input to the next encoder;

[0026] Step 15: Obtain the final aggregated vector from the encoder;

[0027] Step 16: Input the aggregated vector obtained in Step 15 into the prediction module for cascaded growth prediction to obtain the final output;

[0028] Step 17: Process the data and update the learnable parameters in the network to reduce prediction error.

[0029] Furthermore, step 1 includes the following:

[0030] The relationships between user nodes and users are extracted from online social platforms and stored in a directed graph, which is called the user global social relationship network. In this graph, nodes represent users in the social network, edges represent the attention relationships between users, and the direction of the edges represents the direction of information transmission.

[0031] Furthermore, step 2 includes the following:

[0032] The information that needs to be calculated includes the number of cores, PageRank score, hub coefficient, authority coefficient, eigenvector centrality, and clustering coefficient. These features can represent a part of the structural features of the nodes in the graph. These six attributes are used to represent a node. The vector composed of these six attributes is called the user global attribute vector. Since the numerical range of each attribute is different, it is normalized according to the attribute vector of the user. The normalized attribute vector is then used for subsequent numerical calculations.

[0033] Furthermore, step 3 includes the following:

[0034] Information in social networks is stored in the form of message cascades. Each message corresponds to a cascade sequence, and each element in the cascade sequence is a tuple consisting of a user ID and the message sending time. The number of elements in the cascade sequence represents the number of users affected by the message. Arranging the elements in the sequence in chronological order represents the information propagation trajectory. Based on the information in the hours before the message spreads, the final number of newly affected users can be predicted. In other words, the number of subsequent nodes can be predicted based on the first few nodes in the cascade sequence.

[0035] Furthermore, step 4 includes the following:

[0036] For a set of cascaded information data, a data augmentation method using diffuse random sampling is adopted to introduce some global features, thereby facilitating the modeling of cascaded features. First, the order of diffuse random sampling is determined according to the number of stacked layers of the graph neural network. All neighbor nodes of the K-order node are searched in the social relationship network, excluding nodes that have already been sampled. Then, up to 128 nodes are randomly sampled from the remaining nodes as the K+1-order nodes, until the specified order is reached. After sampling twice, a second-order diffuse random sampling cascaded graph of the cascaded sequence is obtained, which is a subgraph containing the node and some of its neighbor nodes.

[0037] Furthermore, step 5 includes the following:

[0038] For time information in a concatenated sequence, a special positional encoding function is used to convert a time number into a vector. Converting time into a vector facilitates the joint calculation of time with other node attributes, while also ensuring that the converted time vector retains the most basic time characteristics. The conversion method is as follows:

[0039] Determine the dimension n of the vector TE to be transformed at each time point;

[0040] Determine the value of the k-th element in vector TE. If k is odd, then... If k is even, then

[0041] Perform the above transformation on all times in the cascaded sequence to obtain the final time matrix.

[0042] Furthermore, step 6 includes the following:

[0043] For the time matrix, add a vector of all zeros at its beginning as the input vector for the spatiotemporal aggregation embedding partial encoder.

[0044] Furthermore, step 7 includes the following:

[0045] Define an N-dimensional all-zero aggregation vector to aggregate the attributes of all information cascade nodes during the neural network transformation process. Add the second-order diffusion random sampling cascade graph obtained in step 4 and the time vector obtained in step 6. The result of the addition and the aggregation vector are input into the encoder layer of the spatiotemporal aggregation embedding module. After performing self-attention transformation, the aggregation vector and the diffusion cascade graph are output.

[0046] In the CasDiffGNN model, the encoder of each layer produces an aggregate vector representing the features of the entire image. This vector has no corresponding node in the image and does not participate in the feature transformation of GAT.

[0047] The process of self-attention transformation is as follows:

[0048] (1) Calculate the self-attention transformation result of the user node: Define three matrices W whose parameters can be updated through backpropagation. Q W K W V Each user node's Q, K, and V vectors are obtained by performing matrix multiplication with each user's input vector. Then, a user's Q vector is multiplied by the K vectors of all other users. The result is divided by the square root of the vector dimension, and the Softmax function is calculated. The Softmax function "compresses" a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), ensuring that each element is within the range (0,1) and the sum of all elements is 1. The correlation between other users and the current user is obtained from the calculation results. This correlation is then used as a weight to perform a weighted summation of the V vectors of all user nodes, resulting in a set of self-attention heads for each user node.

[0049] (2) Define multiple parameter matrices W Q W K W V The system calculates multiple sets of self-attention heads for each user node, concatenates all user self-attention heads to obtain a multi-head attention matrix, defines a linear neural network layer, inputs the multiple attention heads of each user into the linear neural network layer, and obtains the output of the user self-attention transformation.

[0050] Furthermore, step 8 includes the following:

[0051] The diffusion cascade graph obtained in step 7 is used as the input to the GAT layer. In the CasDiffGNN model, the GAT model is selected as the method for structural feature extraction. In the GAT model, neighboring nodes are weighted by attention coefficients, and each node only considers the information of its local neighbors. The CasDiffGNN model uses an alternating stacking method of encoder layer and GAT layer to deeply fuse temporal and structural features. Before the data is input into the GAT layer, a single-layer temporal aggregation transformation is performed to integrate temporal features into the node features for calculation. Through the self-attention mechanism of the encoder layer, global node features and temporal features of the graph are introduced. Because the GAT layer uses an attention mechanism, it can adaptively calculate the contribution weight of each neighboring node to the current node. The attention weight of GAT layer u for node v and the definition of the output are as follows:

[0052]

[0053]

[0054] in, It is the output of the l-th GAT layer. N is the weight vector used for node feature transformation in the l-th layer. v Let α represent the set of neighboring nodes of v. uv This represents the attention weight of u to node v. Let a represent the parameter matrix used for node attention in the l-th layer. (l) These are the weight vector parameters of the l-th layer;

[0055] Finally, a new diffusion cascade diagram was obtained.

[0056] Furthermore, step 9 includes the following:

[0057] Perform a layered drop operation on the nodes in the diffusion cascade graph obtained in step 8, that is, discard the outermost nodes in the network layer by layer in the reverse order of sampling to obtain the cascade graph.

[0058] Furthermore, step 10 includes the following:

[0059] The concatenated graph obtained in step 9 is added to the initial time matrix, and the result of the addition, together with the aggregation vector obtained in step 7, is used as the input of the next spatiotemporal aggregation embedding part of the Encoder layer.

[0060] Furthermore, step 12 includes the following:

[0061] The concatenated graph, aggregation vector, and initial time matrix obtained in step 11 are input into the encoder. The final aggregation vector is obtained from the output of the encoder layer. The final aggregation vector will aggregate the full graph features of GAT layers at different depths, with concatenated graph features as the main feature and neighboring nodes as the secondary feature. During the aggregation process, the nodes that are farther away from the concatenated graph contribute less to the final aggregation vector. This vector is the final embedding vector of the information concatenated by the spatiotemporal aggregation embedding module.

[0062] Furthermore, step 14 includes the following:

[0063] The squared difference between the predicted result and the actual result is calculated, and the average of the squared differences of all data is taken as the average error of the neural network model, which is used to evaluate the performance of the network model. The learnable parameters in the network are updated by gradient descent through backpropagation to reduce the prediction error.

[0064] By employing the above technical solutions, this invention has the following advantages and positive effects compared with the prior art:

[0065] This invention presents a method for predicting the influence of information propagation in social networks based on information cascading. It employs a data augmentation method using diffuse random sampling to introduce global features, thereby better modeling the cascading structural features. It uses sine and cosine position vector encoding of temporal features and employs Gaussian Attraction (GAT) to extract features from the diffuse random sampling cascading graph. Furthermore, it alternately stacks the encoding layers for extracted temporal features and the GAT layer with the aggregated cascading node representations to form a spatiotemporal aggregation embedding module, achieving deep fusion of temporal and structural features. This enhances the influence of temporal and structural features on the prediction results. This method exhibits high prediction accuracy, particularly in scenarios where features are insufficient in the early stages of information cascading propagation. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0067] Figure 1 This is a flowchart illustrating a method for predicting the influence of social network information dissemination based on information cascading, according to the present invention.

[0068] Figure 2 This is a block diagram of a method for predicting the influence of social network information dissemination based on information cascading, according to the present invention.

[0069] Figure 3aThis is a statistical chart of the distribution of observed and predicted cascade lengths in this invention (Weibo);

[0070] Figure 3b This is a statistical chart of the distribution of observed and predicted cascade lengths in this invention (Twitter);

[0071] Figure 3c This is a statistical chart of the distribution of observed and predicted cascade lengths in this invention (Douban).

[0072] Figure 3d This invention provides a statistical diagram of the distribution of observed and predicted cascade lengths (synthetic).

[0073] Figure 4a This is a graph showing the relationship between predicted loss and observation time in this invention (Weibo);

[0074] Figure 4b This is a graph showing the relationship between the predicted loss and observation time in this invention (Twitter);

[0075] Figure 4c This is a graph showing the relationship between the predicted loss and observation time in this invention (Douban).

[0076] Figure 5a This is a graph showing the relationship between the predicted loss and the number of sampling layers in this invention (Weibo);

[0077] Figure 5b This is a graph showing the relationship between the predicted loss and the number of sampling layers in this invention (Twitter);

[0078] Figure 5c This is a graph showing the relationship between the prediction loss and the number of sampling layers in this invention (Douban). Detailed Implementation

[0079] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] This embodiment proposes a method for predicting the influence of social network information dissemination based on information cascading. The method first utilizes the social relationship network, employing a random diffusion method to obtain the neighborhood of nodes in the information cascading sequence. Then, a fixed number of nodes are randomly sampled hierarchically from these neighborhoods to construct a second-order diffused random sampling cascading graph, which serves as the input to a GNN. Subsequently, during the hierarchical extraction of structural features by the GNN, the outermost neighborhood nodes are discarded layer by layer to reduce the impact of the number of graph nodes on computational speed and results. We construct an end-to-end cascading growth prediction model—CasDiffGNN—to evaluate the proposed structural feature extraction method. In CasDiffGNN, a two-layer Graph Attention Network (GAT) is used to extract features from the diffused random sampling cascading graph. Furthermore, the encoder layer, used for extracting temporal features and aggregating cascading node representations, is alternately stacked with the GAT layer to form a spatiotemporal aggregation embedding module, achieving deep fusion of temporal and structural features. Finally, the aggregated vector output from the encoder layer is input into a fully connected layer for cascading growth prediction.

[0081] Specific examples Figure 1 and 2 As shown in the figure, this embodiment discloses a method for predicting the influence of social network information dissemination based on information cascading. The main steps of the prediction are as follows:

[0082] Step 1: Use a directed graph to store the extracted user nodes and the relationships between users;

[0083] Furthermore, step 1 includes the following:

[0084] This method extracts user nodes and relationships between users from online social platforms, storing the information in a directed graph known as the user global social relationship network. In this network, nodes represent users within the social network, edges represent the following relationships between users, and the direction of the edges indicates the direction of information transmission. For example, if node A points to node B, it means user B is a follower of user A, and information sent by user A is transmitted to user B.

[0085] Step 2: Calculate the structural feature information of each user node in the user's global social relationship network;

[0086] Furthermore, step 2 includes the following:

[0087] The information that needs to be calculated includes the number of cores, PageRank score, hub coefficient, authority coefficient, eigenvector centrality, and clustering coefficient. These features can represent a part of the structural features of the nodes in the graph. These six attributes are used to represent a node. The vector composed of these six attributes is called the user global attribute vector. Since the numerical range of each attribute is different, it is normalized according to the attribute vector of the user. The normalized attribute vector is then used for subsequent numerical calculations.

[0088] Step 3: Extract the observation sequence and process the data according to the definition of the information cascade prediction problem;

[0089] Furthermore, step 3 includes the following:

[0090] Information in social networks is stored in a concatenated format, with each message corresponding to a concatenated sequence. Each element in the concatenated sequence is a tuple consisting of a user ID and the message's sending time. The number of elements in the concatenated sequence represents the number of users affected by the message. Arranging the elements in the sequence chronologically represents the information's propagation trajectory. Based on information from the first few hours of propagation, the final number of newly affected users is predicted; that is, the number of subsequent nodes is predicted based on the first few nodes in the concatenated sequence. Assuming the observation period for the concatenated sequence is 3 hours, the elements from the first 3 hours of the information concatenated sequence are extracted to form an observation sequence. The number of remaining elements is predicted using the temporal and structural characteristics of the observation sequence. Since the number of people affected by the information follows a power-law distribution, the actual predicted value is logarithmically reduced to base 2.

[0091] Step 4: Use diffuse random sampling to introduce a portion of global features to model cascaded features;

[0092] Furthermore, step 4 includes the following:

[0093] For a set of cascaded information data, a data augmentation method using diffuse random sampling is adopted to introduce some global features, thereby facilitating the modeling of cascaded features. First, the order of diffuse random sampling is determined according to the number of stacked layers of the graph neural network. All neighbor nodes of the K-order node are searched in the social relationship network, excluding nodes that have already been sampled. Then, up to 128 nodes are randomly sampled from the remaining nodes as the K+1-order nodes, until the specified order is reached. After sampling twice, a second-order diffuse random sampling cascaded graph of the cascaded sequence is obtained, which is a subgraph containing the node and some of its neighbor nodes.

[0094] Step 5: Encode the time information using sine and cosine position vectors to obtain the time matrix;

[0095] Furthermore, step 5 includes the following:

[0096] For time information in a concatenated sequence, a special positional encoding function is used to convert a time number into a vector. Converting time into a vector facilitates the joint calculation of time with other node attributes, while also ensuring that the converted time vector retains the most basic time characteristics, such as the distance relationship between different times. The conversion method is as follows:

[0097] Determine the dimension n of the vector TE to be transformed at each time point;

[0098] Determine the value of the k-th element in vector TE. If k is odd, then... If k is even, then

[0099] Perform the above transformation on all times in the cascaded sequence to obtain the final time matrix.

[0100] Step 6: Convert the time matrix into the encoder's input vector;

[0101] Furthermore, step 6 includes the following:

[0102] For the time matrix, add a vector of all zeros at its beginning as the input vector for the spatiotemporal aggregation embedding partial encoder.

[0103] Step 7: Add the second-order diffused random sampling concatenation graph and the time vector together, and input the result of the addition and the aggregation vector into the encoder layer for self-attention transformation;

[0104] Furthermore, step 7 includes the following:

[0105] Define an N-dimensional all-zero aggregation vector to aggregate the attributes of all information cascade nodes during the neural network transformation process. Add the second-order diffusion random sampling cascade graph obtained in step 4 and the time vector obtained in step 6. The result of the addition and the aggregation vector are input into the encoder layer of the spatiotemporal aggregation embedding module. After performing self-attention transformation, the aggregation vector and the diffusion cascade graph are output.

[0106] In the CasDiffGNN model, the encoder of each layer produces an aggregate vector representing the features of the entire image. This vector has no corresponding node in the image and does not participate in the feature transformation of GAT.

[0107] The process of self-attention transformation is as follows:

[0108] (1) Calculate the self-attention transformation result of the user node: Define three matrices W whose parameters can be updated through backpropagation. Q W K W VEach user node's Q, K, and V vectors are obtained by performing matrix multiplication with each user's input vector. Then, a user's Q vector is multiplied by the K vectors of all other users. The result is divided by the square root of the vector dimension, and the Softmax function (normalized exponential function) is calculated. The Softmax function "compresses" a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), ensuring that each element is within the range (0,1) and the sum of all elements is 1. The correlation between other users and the current user is obtained from the calculation results. This correlation is then used as a weight to perform a weighted summation of the V vectors of all user nodes, resulting in a set of self-attention heads for each user node.

[0109] (2) Define multiple parameter matrices W Q W K W V The system calculates multiple sets of self-attention heads for each user node, concatenates all user self-attention heads to obtain a multi-head attention matrix, defines a linear neural network layer, inputs the multiple attention heads of each user into the linear neural network layer, and obtains the output of the user self-attention transformation.

[0110] Step 8: Alternately stack the GAT layer and encoder layer, and simultaneously fuse temporal and structural features to perform feature transformation;

[0111] Furthermore, step 8 includes the following:

[0112] The diffusion cascade graph obtained in step 7 is used as input to the GAT layer. In the CasDiffGNN model, the GAT model is chosen as the method for structural feature extraction. Compared to the more common GCN model, the GAT model can consider the importance of each individual node. In the GAT model, each neighboring node has its unique weights, which are learned rather than fixed. This means that GAT can more accurately capture the information transmission between nodes. Neighbor nodes are weighted using attention coefficients, with each node considering only its local neighbors. Global node information exchange improves model performance. The Transformer in the encoder layer allows information to be passed between global nodes. The CasDiffGNN model uses an alternating stacking of encoder and GAT layers to deeply fuse temporal and structural features. Before data input to the GAT layer, a single-layer temporal aggregation transformation is performed, integrating temporal features into node features for calculation. The encoder layer's self-attention mechanism introduces global node and temporal features of the graph. The GAT layer, employing an attention mechanism, adaptively calculates the contribution weights of each neighbor node to the current node, thus better handling sparse, large graphs and reducing storage usage. The attention weights of the GAT layer u for node v and the output are defined as follows:

[0113]

[0114]

[0115] in, It is the output of the l-th GAT layer. N is the weight vector used for node feature transformation in the l-th layer. v Let α represent the set of neighboring nodes of v. uv This represents the attention weight of u to node v. Let a represent the parameter matrix used for node attention in the l-th layer. (l) These are the weight vector parameters of the l-th layer;

[0116] Finally, a new diffusion cascade diagram was obtained.

[0117] Step 9: Perform a layered Drop operation;

[0118] Furthermore, step 9 includes the following:

[0119] The nodes in the diffusion cascade graph obtained in step 8 are subjected to a layer-by-layer drop operation, that is, the outermost nodes in the network are discarded layer by layer in the reverse order of sampling, resulting in the cascade graph. The layer-by-layer drop operation can reduce the impact of noisy data in the extra information on the model performance, while speeding up the operation, allowing the model to discard redundant feature information in a timely manner during the feature extraction process.

[0120] Step 10: Add the cascaded graph, aggregation vector, and time matrix together;

[0121] Furthermore, step 10 includes the following:

[0122] The concatenated graph obtained in step 9 is added to the initial time matrix, and the result of the addition, together with the aggregation vector obtained in step 7, is used as the input of the next spatiotemporal aggregation embedding part of the Encoder layer.

[0123] Step 11: Repeat steps 7 to 9, using the obtained concatenated graph, aggregation vector, and initial time matrix as input to the next encoder;

[0124] Step 12: Obtain the final aggregated vector from the encoder;

[0125] Furthermore, step 12 includes the following:

[0126] The concatenated graph, aggregation vector, and initial time matrix obtained in step 11 are input into the encoder. The final aggregation vector is obtained from the output of the encoder layer. The final aggregation vector will aggregate the full-graph features of GAT layers at different depths, with concatenated graph features as the main component and neighboring nodes as the secondary component. During the aggregation process, the nodes that are farther away from the concatenated graph contribute less to the final aggregation vector, which can prevent redundant information in the neighborhood from affecting the prediction and improve the prediction accuracy. This vector is the final embedding vector of the information concatenated by the spatiotemporal aggregation embedding module.

[0127] Step 13: Input the aggregated vector obtained in Step 12 into a three-layer fully connected neural network for cascaded growth prediction to obtain the final output;

[0128] Step 14: Process the data and update the learnable parameters in the network to reduce prediction error.

[0129] Furthermore, step 14 includes the following:

[0130] The squared difference between the predicted result and the actual result is calculated, and the average of the squared differences of all data is taken as the average error of the neural network model, which is used to evaluate the performance of the network model. The learnable parameters in the network are updated by gradient descent through backpropagation to reduce the prediction error.

[0131] Example:

[0132] In the implementation phase, experiments were conducted on synthetic datasets and real datasets from Sina Weibo, Twitter, and Douban. Statistical information for the datasets is shown in Table 1.

[0133] Table 1. Dataset Description

[0134]

[0135] The network generation method in the synthetic dataset adopts the extended-Barabasi-Albert model. The Barabasi-Albert model (BA model for short) is a commonly used social network generation method in network analysis, which can generate network structures that are basically consistent with real-world social networks. The extended BA model adds some randomness to the basic BA model, making the generated network more realistic. Referring to the synthetic network generation methods in previous studies, this embodiment sets the initial number of nodes in the BA model to 1, the number of iterations to 3000, p=0.4, q=0.4, and finally removes isolated nodes in the generated graph and converts the graph into a directed graph, resulting in a generated user network containing 2951 users and 8967 edges.

[0136] The method for generating information concatenation adopts the settings of CoupledGNN, where the seed set size is sampled according to a power-law distribution with a parameter of 2.5, i.e., p(n)∝n. -2.5 The nodes in each seed set are sampled uniformly. The information diffusion model uses an independent cascade model, and the activation probability from node u to node v is 1 / d. v d v It is the in-degree of node v. The iteration is set to 15,000 times. At the same time, cascades with a propagation step size of less than or equal to 3 are excluded. Finally, 3,057 valid cascades are obtained, and the valid cascades are divided into training set, test set and validation set according to the ratio of 0.8:0.1:0.1.

[0137] For the real-world dataset, information cascading from Sina Weibo was initially selected. To obtain the most realistic user relationship network and user cascading information possible, this embodiment tracked and recorded detailed information on nearly 1 billion Weibo posts in one month from a large network of 1.7 million users and 400 million relationships, including 300,000 information cascading messages. This dataset can well demonstrate the cascading propagation situation in reality. Therefore, following the subset extraction method, a small user network and cascading information were extracted from this large-scale dataset.

[0138] The extracted network contains 841,356 edges and 26,720 user nodes, with a total of 2,234 information concatenations. These concatenations are also divided into training, validation, and test sets in a ratio of 0.8:0.1:0.1.

[0139] The Twitter dataset contains tweets from October 2010 and their propagation paths among users. After excluding invalid concatenations that are too short, there are 3,453 message concatenations involving 12,627 users. The friendship relationships between all users participating in the concatenation are considered as social relationships, and a social relationship network is constructed based on this. The social relationship network contains 309,631 edges.

[0140] The Douban dataset was collected from the Douban website, where users can share their book or movie reading status and follow others' statuses. Users participating in the dissemination of the same information form an information cascade, containing 3484 information cascades and 24926 related users. Co-occurrence relationships between users (e.g., reading the same book) are considered their social relationships, with a total of 379155 co-occurrence relationships among all users in the dataset. Information spreads more slowly on Douban than on Weibo and Twitter, therefore we chose to observe cascades on a monthly basis.

[0141] For the Douban dataset, the observation time was defined as 3 months, 6 months, and 9 months; for the Weibo dataset, the observation time was 1 hour, 2 hours, and 3 hours; for the Twitter dataset, the observation time was 0.5 hours, 1 hour, and 1.5 hours; for the synthetic dataset, since there was no concept of propagation time, observations were based on the number of steps in information propagation, and the observation duration was taken as when the propagation took 2 steps. The statistics of the datasets under these observation times are shown in Table 2.

[0142] Table 2. Statistical information of datasets at different observation times.

[0143]

[0144] In addition, the relationship between the observation length and prediction length of information cascade and the corresponding quantity was analyzed, as shown in Figure 3.

[0145] from Figure 3a As shown in ~d, the statistical values ​​of both the observed and predicted lengths of the cascades conform to a power-law distribution. Furthermore, the distributions of observed and predicted lengths within the same dataset are largely consistent, and the distribution of the total length of the cascades also tends to be uniform. This indicates that predicting the growth of the cascades based on the set observation time is reasonable. Within this observation time, the propagation patterns of different cascades have already shown differences, and their future growth can be predicted based on these factors.

[0146] Based on previous work experience in setting evaluation metrics, MLSE will also be selected as the evaluation metric for this experiment:

[0147]

[0148] Where N represents the total number of cascades, Δy′ i Δy represents the predicted increase in cascade i by the model. i This represents the actual increase in the cascade i.

[0149] Regarding the parameter settings in the comparison method, for the feature-based method, the hidden layer dimension is set to 128 in this embodiment. The hyperparameters of the other models are all taken from the hyperparameters of the optimal model given in the original paper and code. At the same time, the limitation on the number of iterations is removed to ensure that a converged loss value can be obtained.

[0150] Since the DeepHawkes model requires information propagation paths, and this data is missing from the dataset used in this embodiment, an algorithm is used to reconstruct the propagation paths. First, the user nodes in the information cascade are arranged chronologically. The first user node in the sequence is taken as the information propagation source user and added to the influencing user set, while the remaining nodes are added to the user set to be influenced. User nodes are then sequentially selected from the user set to be influenced as information receivers, and it is determined whether the user has a neighbor relationship with any user in the influencing user set within the social relationship network. If a neighbor relationship exists with only one user, that user is selected as the information propagator. If a neighbor relationship exists with multiple users, the out-degree, in-degree, and node number of these users in the social relationship network are compared sequentially to select the information propagator. If no neighbor relationship exists, the out-degree, in-degree, and node number of all users in the influencing user set are compared to select the information propagator. Finally, the propagation trajectory from the information propagator to the receiver is constructed, and the information receiver is moved from the user set to be influenced to the influencing user set. Repeat the above steps until there are no more users in the user group to be affected. Use all the propagation trajectories built in this process as the propagation trajectories of this information cascade for use in DeepHawkes.

[0151] The CasDiffGNN algorithm proposed in this embodiment has 2 layers for random diffusion and 2 layers for GAT stacking, and 3 layers for augmented temporal aggregation. The hidden layer dimension in GAT is 32, the number of multi-head attention heads is 2, and the dropout rate is 0.1. The hidden layer dimension in augmented temporal aggregation layer is 32, the number of multi-head attention heads is 2, and the dropout rate is 0.2. The fully connected layer has three hidden layers with dimensions of 32, 32, and 1 respectively, and the dropout rate is 0.1. The Adam optimizer is used with an initial learning rate of 0.005 and a weight decay coefficient of 0.001.

[0152] Following the same data processing and partitioning methods, this embodiment compared all the comparison models with the proposed CasDiffGNN model on the Douban dataset, Weibo dataset, Twitter dataset, and synthetic dataset, respectively. The comparison results are shown in Table 3:

[0153] Table 3 Overall Predictive Performance

[0154]

[0155] The performance comparison is shown in Table 3. The CasDiffGNN model outperforms other models on all datasets, with the prediction error reduced by up to 24.86% and by an average of 9.75% compared to previous models. DeepHawkes uses a time decay function to handle the relationship between time and prediction results, while CCasGNN uses a positional encoding function to encode the order of user participation in the cascade sequence. This embodiment uses a positional encoding function to encode the timestamps of information propagation. In synthetic datasets, both timestamps and sequence information are represented as propagation steps, which cannot accurately represent the order or propagation rate. CasSeqGCN, on the other hand, divides time periods and counts user nodes within the same time period in the same cascade graph. This model is consistent with the idea of ​​propagating information according to the number of iterations in synthetic datasets, so its performance is not affected and it achieves a smaller prediction error. Except for the model in this embodiment, Feature Linear, and DeepCas models, other models all exhibit the problem that the prediction error increases with the increase of observation time.

[0156] To better explore the contribution of each component in CasDiffGNN to the final prediction result and to delve into the true factors influencing the cascade prediction problem, this embodiment designs various variant models of CasDiffGNN. In CasDiffGNN, the main contributions lie in the diffuse random sampling and spatiotemporal aggregation embedding parts. To explore the role of these two parts, the following model variants were designed:

[0157] CasDiffGNN-noDrop: The original CasDiffGNN discards nodes and their corresponding edges in the outermost neighborhood of the diffuse random sampling cascade graph after the graph neural network layers end. In this variant model, no nodes in the diffuse random sampling cascade graph are discarded.

[0158] CasDiffGNN-seqStack: In CasDiffGNN, the encoder layer and GAT layer are stacked alternately to achieve deep fusion of temporal and structural features, aggregating node embedding vectors extracted by GAT layers of different depths. To explore the rationality and effectiveness of this design, in this variant model, the GAT layer and encoder layer are stacked sequentially.

[0159] CasDiffGNN-diff0: Sets the number of sampling iterations in diffuse random sampling to 0, i.e., does not perform diffuse random sampling.

[0160] CasDiffGNN-diff1: Sets the number of sampling iterations in the diffusion random sampling to 1, performs diffusion random sampling only on the first-order neighbor nodes, and then discards the neighbor nodes after the first GAT layer.

[0161] CasDiffGNN-diff2: Sets the number of sampling iterations in the diffuse random sampling to 2, consistent with the original CasDiffGNN.

[0162] CasDiffGNN-diff3: Sets the number of sampling iterations in the diffuse random sampling to 3, samples the nodes in the 3rd order neighbors of the user, and then discards the outermost neighbor nodes after the first GAT layer and the remaining two neighbor nodes after the second GAT layer.

[0163] Table 4 Comparison of CasDiffGNN ablation model results

[0164]

[0165]

[0166] As can be seen from the data in Table 4, the CasDiffGNN model proposed in this embodiment, namely the model variant CasDiffGNN-diff2, achieved the best prediction results on most datasets. Furthermore, the table also shows that the CasDiffGNN-noDrop model without hierarchical drop operations slightly outperformed CasDiffGNN-diff2 in terms of the shortest observation time on the Weibo and Twitter datasets, with prediction errors reduced by 3.25% and 7.72%, respectively. However, after removing the hierarchical drop operation, the stability of the algorithm decreased significantly. The trend of linear function fitting error with observation time shows that the slopes of the linear function of the model without hierarchical drop operations on the three real datasets are 0.143, 0.561, and 0.030, respectively, while the corresponding slopes of the CasDiffGNN model are 0.016, 0.032, and 0.007. As can be seen, the diffuse random sampling graph contains a large amount of noisy data irrelevant to the prediction. Using all this data for prediction without filtering would not only increase computational and storage costs but also negatively impact the prediction results. Therefore, discarding some nodes at appropriate times after diffuse random sampling is essential for improving prediction accuracy and algorithm stability.

[0167] pass Figure 4a Comparing experimental data from CasDiffGNN-diffX and CasDiffGNN-seqStack, we can see that the predictive ability of the model significantly decreases after changing the stacking order of its components. The predictive loss of CasDiffGNN-seqStack is generally lower than that of a series of alternately stacked CasDiffGNN-diffX models. This indicates that there is information in the neighboring nodes after diffusion sampling that is helpful for prediction. However, if we only use GAT to extract this information without relying on the augmented temporal aggregation module, we cannot effectively utilize this information. Therefore, the augmented temporal aggregation module should be used to aggregate the features of the neighboring nodes before discarding them.

[0168] To investigate the impact of the number of diffusion layers on prediction results in diffuse random sampling, this embodiment compares models with 0th, 1st, 2nd (CasDiffGNN), and 3rd order diffusion layers without changing the network structure. The table shows that when the diffusion order is 2, the error between the predicted and actual results is minimized. When the diffusion order is 1, the model's prediction error is smaller than most baseline methods, and even outperforms the CasDiffGNN model proposed in this chapter under long observation times on Twitter. Figure 5a~c shows the relationship between the order of diffusion random sampling and prediction error for different datasets and observation times. We can clearly see that on most datasets, when the order of diffusion random sampling is less than or equal to 2, the prediction error of the model decreases as the order increases. This indicates that diffusion sampling of the neighborhood of the cascade graph is an effective method to improve model performance.

[0169] However, the algorithm's error increases rapidly once the third-order neighborhood of cascaded user nodes is sampled. This phenomenon is related to the relationship between the order of the r-order reachability graph and the number of nodes. Previous research has shown that even in a second-order reachability graph of a cascaded sequence with only 10 user nodes, the number of nodes can reach hundreds of thousands. Diffusion random sampling reaching order 3 is equivalent to randomly sampling user nodes throughout the entire social network graph. Therefore, neighboring nodes within the third-order neighborhood not only introduce beneficial information that enhances prediction accuracy but also bring more noisy data, thus affecting the model's accurate prediction of future cascade growth. Therefore, given the inability to eliminate irrelevant information interference in the graph, only nodes within the second-order neighborhood should be used to predict the future growth of cascaded sequences.

[0170] This embodiment uses a processing method that combines time features and structural features to avoid the loss of accuracy caused by insufficient structural features, and has more advantages in early cascade prediction with shorter observation time.

[0171] In terms of runtime efficiency, CoupledGNN has the highest computational cost and lowest runtime efficiency because it performs graph convolutions on the entire graph. Compared to RNNs, a major advantage of the Encoder is that the computation is parallel, resulting in faster speeds for both forward and backward propagation. Therefore, our model's runtime efficiency is superior to CasSeqGCN, which uses RNN-type networks, and is basically consistent with CCasGNN.

[0172] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the influence of social network information dissemination based on information cascading, characterized in that, Includes the following steps: Step 1: Use a directed graph to store the extracted user nodes and the relationships between users; Step 1 includes the following: The relationships between user nodes and users are extracted from online social platforms and stored in a directed graph called the global social relationship network. In this graph, nodes represent users in the social network, edges represent the following relationships between users, and the direction of the edges represents the direction of information transmission. Step 2: Calculate the structural feature information of each user node in the user's global social relationship network; Step 2 includes the following: The information that needs to be calculated includes the number of cores, PageRank score, hub coefficient, authority coefficient, feature vector centrality and clustering coefficient. These features can represent a part of the structural features of the nodes in the graph. These six attributes are used to represent a node. The vector composed of these six attributes is called the user global attribute vector. Since the numerical range of each attribute is different, it is normalized according to the attribute vector of the user. The normalized attribute vector is then used for subsequent numerical calculations. Step 3: Extract the observation sequence and process the data according to the definition of the information cascade prediction problem; Step 4: Use diffuse random sampling to introduce a portion of global features to model cascaded features; Step 5: Encode the time information using sine and cosine position vectors to obtain the time matrix; Step 6: Convert the time matrix into the encoder's input vector; Step 7: Add the second-order diffused random sampling concatenation graph and the time vector together, and input the result of the addition and the aggregation vector into the encoder layer for self-attention transformation; Step 8: Alternately stack the GAT layer and encoder layer, and simultaneously fuse temporal and structural features to perform feature transformation; Step 9: Perform a layered Drop operation; Step 10: Add the cascaded graph, aggregation vector, and time matrix together; Step 11: Repeat steps 7 to 9, using the obtained concatenated graph, aggregation vector, and initial time matrix as input to the next encoder; Step 12: Obtain the final aggregated vector from the encoder; Step 13: Input the aggregated vector obtained in Step 12 into a three-layer fully connected neural network for cascaded growth prediction to obtain the final output; Step 14: Process the data and update the learnable parameters in the network to reduce prediction error.

2. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 1, characterized in that, Step 3 Includes the following: Information in social networks is stored in the form of message cascades. Each message corresponds to a cascade sequence, and each element in the cascade sequence is a tuple consisting of a user ID and the message sending time. The number of elements in the cascade sequence represents the number of users affected by the message. Arranging the elements in the sequence in chronological order represents the information propagation trajectory. Based on the information in the hours before the message spreads, the final number of newly affected users can be predicted. In other words, the number of subsequent nodes can be predicted based on the first few nodes in the cascade sequence.

3. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 2, characterized in that, Step 4 includes the following: For a set of cascaded information data, a data augmentation method using diffuse random sampling is adopted to introduce some global features, thereby facilitating the modeling of cascaded features. First, the order of diffuse random sampling is determined according to the number of stacked layers of the graph neural network. All neighbor nodes of the K-order node are searched in the social relationship network, excluding nodes that have already been sampled. Then, up to 128 nodes are randomly sampled from the remaining nodes as the K+1-order nodes, until the specified order is reached. After sampling twice, a second-order diffuse random sampling cascaded graph of the cascaded sequence is obtained, which is a subgraph containing the node and some of its neighbor nodes.

4. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 3, characterized in that, Step 5 includes the following: For time information in a concatenated sequence, a special positional encoding function is used to convert a time number into a vector. Converting time into a vector facilitates the joint calculation of time with other node attributes, while also ensuring that the converted time vector retains the most basic time characteristics. The conversion method is as follows: Determine the dimension n of the vector TE to be transformed at each time point; Determine the value of the k-th element in vector TE. If k is odd, then... If k is even, then Perform the above transformation on all times in the cascaded sequence to obtain the final time matrix.

5. The method for predicting the influence of social network information dissemination based on information cascading according to claim 4, characterized in that, Step 6 includes the following: For the time matrix, add a vector of all zeros at its beginning as the input vector for the spatiotemporal aggregation embedding partial encoder.

6. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 5, characterized in that, Step 7 includes the following: Define an N-dimensional all-zero aggregation vector to aggregate the attributes of all information cascade nodes during the neural network transformation process. Add the second-order diffusion random sampling cascade graph obtained in step 4 and the time vector obtained in step 6. The result of the addition and the aggregation vector are input into the encoder layer of the spatiotemporal aggregation embedding module. After performing self-attention transformation, the aggregation vector and the diffusion cascade graph are output. In the CasDiffGNN model, the encoder of each layer produces an aggregate vector representing the features of the entire image. This vector has no corresponding node in the image and does not participate in the feature transformation of GAT. The process of self-attention transformation is as follows: (1) Calculate the self-attention transformation result of the user node: Define three matrices W whose parameters can be updated through backpropagation. Q W K W V Each user node's Q, K, and V vectors are obtained by performing matrix multiplication with each user's input vector. Then, a user's Q vector is multiplied by the K vectors of all other users. The result is divided by the square root of the vector dimension, and the Softmax function is calculated. The Softmax function "compresses" a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), ensuring that each element is within the range (0,1) and the sum of all elements is 1. The correlation between other users and the current user is obtained from the calculation results. This correlation is then used as a weight to perform a weighted summation of the V vectors of all user nodes, resulting in a set of self-attention heads for each user node. (2) Define multiple parameter matrices W Q W K W V The system calculates multiple sets of self-attention heads for each user node, concatenates all user self-attention heads to obtain a multi-head attention matrix, defines a linear neural network layer, inputs the multiple attention heads of each user into the linear neural network layer, and obtains the output of the user self-attention transformation.

7. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 6, characterized in that, Step 8 includes the following: The diffusion cascade graph obtained in step 7 is used as the input to the GAT layer. In the CasDiffGNN model, the GAT model is selected as the method for structural feature extraction. In the GAT model, neighboring nodes are weighted by attention coefficients, and each node only considers the information of its local neighbors. The CasDiffGNN model uses an alternating stacking method of encoder layer and GAT layer to deeply fuse temporal and structural features. Before the data is input into the GAT layer, a single-layer temporal aggregation transformation is performed to integrate temporal features into node features for calculation. Through the self-attention mechanism of the encoder layer, global node features and temporal features of the graph are introduced. Because the GAT layer employs an attention mechanism, it can adaptively calculate the contribution weight of each neighboring node to the current node. The attention weight of GAT layer u for node v and the definition of the output are as follows: in, It is the output of the l-th GAT layer. N is the weight vector used for node feature transformation in the l-th layer. v Let α represent the set of neighboring nodes of v. uv This represents the attention weight of u to node v. Let a represent the parameter matrix used for node attention in the l-th layer. (l) These are the weight vector parameters of the l-th layer; Finally, a new diffusion cascade diagram was obtained.

8. The method for predicting the influence of social network information dissemination based on information cascading as described in claim 7, characterized in that, Step 9 includes the following: Perform a layered drop operation on the nodes in the diffusion cascade graph obtained in step 8, that is, discard the outermost nodes in the network layer by layer in the reverse order of sampling to obtain the cascade graph.

9. A method for predicting the influence of social network information dissemination based on information cascading as described in claim 8, characterized in that, Step 10 includes the following: The concatenated graph obtained in step 9 is added to the initial time matrix, and the result of the addition, together with the aggregation vector obtained in step 7, is used as the input of the next spatiotemporal aggregation embedding part of the Encoder layer.

10. A method for predicting the influence of social network information dissemination based on information cascading as described in claim 9, characterized in that, Step 12 includes the following: The concatenated graph, aggregation vector, and initial time matrix obtained in step 11 are input into the encoder. The final aggregation vector is obtained from the output of the encoder layer. The final aggregation vector will aggregate the full graph features of GAT layers at different depths, with concatenated graph features as the main feature and neighboring nodes as the secondary feature. During the aggregation process, the nodes that are farther away from the concatenated graph contribute less to the final aggregation vector. This vector is the final embedding vector of the information concatenated by the spatiotemporal aggregation embedding module.

11. A method for predicting the influence of social network information dissemination based on information cascading as described in claim 10, characterized in that, Step 14 includes the following: The squared difference between the predicted result and the actual result is calculated, and the average of the squared differences of all data is taken as the average error of the neural network model, which is used to evaluate the performance of the network model. The learnable parameters in the network are updated by gradient descent through backpropagation to reduce the prediction error.

Images