A method and system for predicting information diffusion in social networks

Abstract

The present invention relates to a method and system for predicting information diffusion in a social network, and relates to the field of information diffusion prediction. The method comprises: obtaining diffusion messages from social media; constructing a dynamic diffusion tree set and an adjacency matrix of the diffusion process based on the diffusion messages and the diffusion cascades corresponding to the diffusion messages; reconstructing the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph; learning and fusing local diffusion representations based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a user local diffusion representation; learning a global representation based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain a user global diffusion representation; and predicting information diffusion using an attention mechanism based on the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and a diffusion probability. The present invention improves the accuracy of information diffusion prediction in online social media by fusing local and global diffusion representations.

Description

Technical Field

[0001] The present invention relates to the field of information diffusion prediction, and in particular to a method and system for predicting information diffusion in a social network. Background Art

[0002] In recent years, with the rapid development of internet technology and the widespread use of smartphones, online social media has become an integral part of everyone's daily lives. While people can freely and quickly access news, express opinions, and forward information on social platforms, the resulting proliferation of massive amounts of data and information is exploding across these platforms.

[0003] On the one hand, these diffusion processes contain many beneficial factors that can guide business behavior. For example, on the Weibo platform, based on the current diffusion status of a specific Weibo post, its subsequent spread can be estimated, predicting its potential to become a trending search; or, based on the diffusion status of an existing trending search, its subsequent duration on the trending search list can be estimated. Another example is that the platform can estimate each user's forwarding preferences based on their reading, liking, and forwarding habits, thereby providing personalized product advertising recommendations to each user. On the other hand, the diffusion process also harbors malicious information that poses a risk of harm to virtual platforms and even real society. How to detect and prevent such information in the shortest possible time and with the least possible dissemination has become a critical issue in the current governance of the internet environment.

[0004] Information diffusion prediction, also known as cascade prediction, aims to study how information items spread along existing diffusion paths and build models to predict their subsequent diffusion processes. This task can summarize the diffusion process of a specific piece of information or a large amount of information, and provide foundational models and theoretical support for downstream application tasks such as popularity prediction, personalized recommendations, rumor detection, and malicious user detection.

[0005] To meet the needs of different downstream tasks, currently known information diffusion prediction models can be divided into two categories: micro-information diffusion prediction models and macro-information diffusion prediction models. The former aims to accurately predict the diffusion status of specific information in the short term, such as analyzing the commercial value of information dissemination paths of influential users and real-time monitoring and prevention of rumors. The latter, on the other hand, predicts the diffusion status of complete information groups over the subsequent longer term. This method is often used to analyze and predict the joint diffusion behavior of multiple related information, such as predicting the spread of sudden events and identifying spam.

[0006] Because the two prediction methods target different downstream tasks, they also differ in model construction. Micro-prediction typically uses linked lists or tree structures to represent specific information diffusion structures. For example, Wang Jia et al. (ICDM, 2017) proposed a micro-prediction method based on recurrent neural networks (RNNs). This method uses RNNs to simulate the linked list structure of a single specific diffusion cascade, learn the current diffusion state, and predict the subsequent diffusion process. Yang Chen et al. (arXiv, 2018) proposed an improved micro-prediction method based on convolutional neural networks (CNNs) and a self-attention mechanism. This method uses CNNs to represent node attribute information and uses a self-attention mechanism instead of RNNs to address the long-term memory loss problem caused by RNNs, thereby improving prediction performance. Building on this, Yuan Chunyuan et al. (PKDD, 2020) proposed a micro-diffusion prediction method that combines dynamic social networks with diffusion networks. This method incorporates attention relationships between users and encodes real-time temporal information into the model, constructing a dynamic heterogeneous network set to learn real-time changes in users' diffusion intentions. On the other hand, macro-forecasting models typically use a complete graph structure to represent the diffusion state of an information group. For example, Li Cheng et al. (WWW, 2017) designed a macro-diffusion prediction method based on random walking. Through random walking, they converted a cascade graph consisting of multiple diffusion information into multiple node sequences, and combined it with the RNN framework for macro-prediction. Chen Xueqin et al. (ICDE, 2019) further expanded the macro-diffusion prediction method, using a graph convolutional neural network (GCN) to simultaneously capture the diffusion structure information and node attribute information of the entire diffusion graph, and used a long short-term memory network (LSTM) to learn the intrinsic dependencies between user forwarding behaviors in the cascade sequence information. At the same time, Cheng Yang et al. (IJCAI, 2019) also combined social network information between users in macro-diffusion prediction and used reinforcement learning to further improve the accuracy of macro-prediction.

[0007] In reality, while using linked lists or tree structures to represent diffusion processes can accurately depict the local diffusion state of specific information, they are unable to predict, supplement, or correct information through the global diffusion relationships between related information. While using graph structures to represent diffusion processes can provide a macroscopic representation of the global diffusion state of a group of information, it ignores the unique local diffusion properties of each piece of information. Clearly, both of these diffusion relationships are beneficial for improving prediction accuracy, whether for microscopic or macroscopic information diffusion prediction. However, current known methods fail to consider combining local and global diffusion relationships for information diffusion prediction. Furthermore, current known methods lack the ability to meticulously classify and effectively learn representations of multiple diffusion properties within these two distinct diffusion relationships. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for predicting information diffusion in social networks, which improves the accuracy of information diffusion prediction in online social media by fusing local and global diffusion representations.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A method for predicting information diffusion in a social network, comprising:

[0011] Get the spread of news on social media;

[0012] Constructing a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message;

[0013] Reconstructing the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph;

[0014] Performing local diffusion representation learning and fusion based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a local diffusion representation of the user;

[0015] Performing global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation;

[0016] Information diffusion prediction is performed using an attention mechanism method according to the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and a diffusion probability.

[0017] Optionally, constructing a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message specifically includes:

[0018] Constructing a dynamic diffusion tree according to the diffusion message and the diffusion cascade corresponding to the diffusion message;

[0019] Diffusion time is assigned and collected for all diffusion processes of each dynamic diffusion tree to obtain a dynamic diffusion tree set and an adjacency matrix of the diffusion process.

[0020] Optionally, reconstructing the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph specifically includes:

[0021] Deleting all edges in the dynamic diffusion tree, and reconstructing the tree with the diffusion information as the root node and the user node as the first-order leaf node according to a set depth to obtain a reconstructed diffusion tree;

[0022] Collecting the reconstructed diffusion trees according to the serial numbers of the diffusion messages to obtain a reconstructed diffusion tree set;

[0023] All reconstructed diffusion trees in the reconstructed diffusion tree set are connected according to user nodes to obtain a heterogeneous information bipartite graph.

[0024] Optionally, the performing local diffusion representation learning and fusion according to the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain the user local diffusion representation specifically includes:

[0025] Performing local diffusion representation learning on all user nodes of each dynamic diffusion tree in the dynamic diffusion tree set according to the adjacency matrix of the diffusion process to obtain a first local diffusion representation;

[0026] The first local diffusion representations corresponding to different diffusion messages of the same user are fused to obtain a user local diffusion representation.

[0027] Optionally, performing global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation specifically includes:

[0028] Performing representation learning on the user nodes and the message nodes in the heterogeneous information bipartite graph to obtain attention weights of the message nodes and the user nodes;

[0029] A global representation learning is performed using a graph attention network according to the attention weights of the message node and the user node and the user's local diffusion representation to obtain a user's global diffusion representation.

[0030] Optionally, performing information diffusion prediction using an attention mechanism method based on the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and a diffusion probability specifically includes:

[0031] A user diffusion representation is obtained by fusing the user local diffusion representation and the user global diffusion representation using a heuristic algorithm;

[0032] According to the user diffusion representation, the attention mechanism method is used to predict information diffusion to obtain the diffusion representation and diffusion probability.

[0033] A social network information diffusion prediction system, comprising:

[0034] An acquisition module is used to obtain diffusion messages from social media;

[0035] A local diffusion construction module, configured to construct a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message;

[0036] A global diffusion reconstruction module, configured to reconstruct the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph;

[0037] A local diffusion representation learning module, configured to perform local diffusion representation learning and fusion based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a user local diffusion representation;

[0038] A global diffusion representation learning module, configured to perform global representation learning based on the user's local diffusion representation and the heterogeneous information bipartite graph to obtain a user's global diffusion representation;

[0039] The information diffusion prediction module is used to perform information diffusion prediction based on the user local diffusion representation and the user global diffusion representation using an attention mechanism method to obtain a diffusion representation and a diffusion probability.

[0040] Optionally, the local diffusion building module specifically includes:

[0041] A construction unit, configured to construct a dynamic diffusion tree according to the diffusion message and the diffusion cascade corresponding to the diffusion message;

[0042] The diffusion time assignment and collection unit is used to assign and collect diffusion time for all diffusion processes of each dynamic diffusion tree to obtain a dynamic diffusion tree set and an adjacency matrix of the diffusion process.

[0043] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] The present invention constructs a dynamic diffusion tree set and an adjacency matrix of the diffusion process based on diffusion messages and the diffusion cascades corresponding to the diffusion messages. The dynamic diffusion trees in the dynamic diffusion tree set are reconstructed to obtain a heterogeneous information bipartite graph. Local diffusion representations are learned and fused based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a user local diffusion representation. Global representations are learned based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain a user global diffusion representation. Information diffusion is predicted using an attention mechanism based on the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and diffusion probability. By fusing local and global diffusion representations, the accuracy of information diffusion prediction in online social media is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flow chart of the social network information diffusion prediction method provided by the present invention;

[0047] Figure 2 A schematic diagram of the social network information diffusion prediction method provided by the present invention;

[0048] Figure 3 Construct example graphs for the dynamic diffusion tree set of local diffusion relations;

[0049] Figure 4 Construct an example graph for a bipartite graph of heterogeneous information of a global diffusion relation;

[0050] Figure 5 Flowchart for representation learning and diffusion prediction. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The purpose of the present invention is to provide a method and system for predicting information diffusion in social networks, which improves the accuracy of information diffusion prediction in online social media by fusing local and global diffusion representations.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] A method for predicting information diffusion in online social media is proposed. By connecting the local and global relationships of the information diffusion process, the model can autonomously learn the weights of the representations under the two diffusion relationships, and use the fused representation to jointly predict the subsequent diffusion process, which can effectively improve the accuracy of information diffusion prediction in online social media.

[0055] like Figure 1 As shown, the present invention provides a method for predicting information diffusion in a social network, comprising:

[0056] S1: Obtain diffusion information from social media.

[0057] S2: construct a dynamic diffusion tree set and an adjacency matrix of the diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message. The local relationship in the information diffusion process (ie, each specific information diffusion process) is represented using a dynamic diffusion tree set T structure.

[0058] S2, specifically including:

[0059] A dynamic diffusion tree is constructed based on the diffusion message and the diffusion cascade corresponding to the diffusion message. i ∈M, construct the corresponding tree structure r i ∈T.

[0060] Diffusion time is assigned and collected for all diffusion processes of each dynamic diffusion tree to obtain a dynamic diffusion tree set and an adjacency matrix of the diffusion process.

[0061] For each dynamic diffusion tree r i All diffusion processes in <u,v>,u,v∈u are assigned diffusion time t, where u is the user set and u,v are users in the user set. All diffusion trees are collected into a set independently according to the diffusion information number, and the dynamic diffusion tree set T=[r 1 ,r 2 ,...].

[0062] S3: Reconstruct the dynamic diffusion tree in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph, wherein this structure is obtained by extending the information diffusion direction based on the traditional bipartite graph in the present invention. The global relationship in the information diffusion process (i.e., the overall information group diffusion process) is analyzed using the heterogeneous information bipartite graph. Structural representation.

[0063] S3, specifically including:

[0064] Delete all edges in the dynamic diffusion tree, and reconstruct it with the diffusion information as the root node and the user node as the first-order leaf node according to the set depth to obtain a reconstructed diffusion tree. i As a unit, eliminate all edges in the tree and diffuse information m i is the root node, all user nodes u are first-order leaf nodes, and an intermediate reconstruction diffusion tree with a depth of 2 is constructed.

[0065] The reconstructed diffusion trees are collected according to the serial numbers of the diffusion messages to obtain a reconstructed diffusion tree set.

[0066] All reconstructed diffusion trees in the reconstructed diffusion tree set are connected according to user nodes to obtain a heterogeneous information bipartite graph.

[0067] According to the forwarding relationship between user node u and diffusion message m, user node u is used as an intermediary to connect all reconstructed diffusion trees. Constructing a bipartite graph with heterogeneous information where ε r Is a user node and spread the news The set of undirected edges between .

[0068] S4: performing local diffusion representation learning and fusion according to the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain the user local diffusion representation.

[0069] S4, specifically including:

[0070] Local diffusion representation learning is performed on all user nodes of each dynamic diffusion tree in the dynamic diffusion tree set according to the adjacency matrix of the diffusion process to obtain a first local diffusion representation.

[0071] The first local diffusion representations corresponding to different diffusion messages of the same user are fused to obtain a user local diffusion representation.

[0072] The process includes the following steps:

[0073] Using graph convolutional neural networks with temporal features to train dynamic diffusion tree sets Each diffusion tree r i All user nodes u in i Perform local diffusion representation learning, and the calculation formula is:

[0074]

[0075] Among them, X i With A i Indicates that in the diffusion tree r i The user feature embedding in is the diffusion process adjacency matrix, Et is the time feature matrix, which records the time when the user diffuses the message, W i is the parameter matrix; σ(·) is the activation function; This is the first local diffusion characterization.

[0076] Integrate users in all information Diffusion expression in Get the user's local diffusion representation

[0077] S5: Perform global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation.

[0078] S5, specifically including:

[0079] Representation learning is performed based on the user nodes and message nodes in the heterogeneous information bipartite graph to obtain attention weights of the message nodes and the user nodes.

[0080] A global representation learning is performed using a graph attention network according to the attention weights of the message node and the user node and the user's local diffusion representation to obtain a user's global diffusion representation.

[0081] The process includes the following steps:

[0082] Using graph attention network, we can learn heterogeneous information bipartite graphs. All user nodes u perform global diffusion representation learning, and the calculation formula is:

[0083]

[0084] Among them, H m It is the feature embedding of the diffusion message in the bipartite graph. Through multiple iterations, the global diffusion representation H of the user is finally learned. global .W m is the parameter matrix.

[0085] S6: Using the attention mechanism method to predict information diffusion based on the user's local diffusion representation and the user's global diffusion representation, to obtain a diffusion representation and a diffusion probability.

[0086] S6, specifically including:

[0087] The user local diffusion representation and the user global diffusion representation are fused using a heuristic algorithm to obtain a user diffusion representation.

[0088] According to the user diffusion representation, the attention mechanism method is used to predict information diffusion to obtain the diffusion representation and diffusion probability.

[0089] Combine local and global relationships to learn user diffusion representations and perform diffusion prediction.

[0090] The process includes the following steps:

[0091] Fusion local diffusion characterization H local With global diffusion characterization H global , obtain user diffusion representation

[0092] Use the attention mechanism method to perform diffusion prediction and calculate the user's diffusion representation Z and diffusion probability The calculation formula is:

[0093]

[0094]

[0095] Calculate the loss function and return the layout diffusion representation learning step feedback optimization.

[0096] The present invention integrates local and global diffusion representations to predict social network information diffusion. First, models are constructed based on the local and global diffusion relationships in the information diffusion process, and a graph convolutional neural network and a graph attention network with time features improved for social media are used for representation learning. In the learning of the two diffusion relationship representations, multiple diffusion attributes are accurately and effectively integrated. At the same time, the invention can autonomously learn the importance of the two representations for information diffusion prediction, and integrate the local and global representations of diffusion to jointly predict information diffusion.

[0097] The present invention has the following advantages:

[0098] By separately modeling the local and global diffusion relationships of information, user diffusion representations possess both local and global characteristic attributes. The model can autonomously learn the importance of the two different diffusion relationships and autonomously adjust the weights based on different micro or macro diffusion prediction tasks, greatly enhancing the versatility of the present invention. The present invention incorporates specific information diffusion features such as diffusion time attributes and diffusion structure attributes into local diffusion representation learning, while integrating information group diffusion features such as diffusion information text attributes into global diffusion representation learning. This allows the model to more carefully learn the dependencies between different attributes and the two diffusion relationships, thereby effectively improving the accuracy of information diffusion prediction.

[0099] The present invention also provides a social network information diffusion prediction system, comprising:

[0100] The acquisition module is used to obtain diffusion messages from social media.

[0101] A local diffusion construction module is configured to construct a dynamic diffusion tree set and an adjacency matrix for the diffusion process based on the diffusion message and the diffusion cascade corresponding to the diffusion message. The construction module specifically includes: a construction unit configured to construct a dynamic diffusion tree based on the diffusion message and the diffusion cascade corresponding to the diffusion message; and a diffusion time assignment and collection unit configured to assign and collect diffusion times for all diffusion processes in each dynamic diffusion tree to obtain the dynamic diffusion tree set and the adjacency matrix for the diffusion process.

[0102] The global diffusion reconstruction module is used to reconstruct the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph.

[0103] The local diffusion representation learning module is used to perform local diffusion representation learning and fusion according to the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain the user local diffusion representation.

[0104] The global diffusion representation learning module is used to perform global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation.

[0105] The information diffusion prediction module is used to perform information diffusion prediction based on the user local diffusion representation and the user global diffusion representation using an attention mechanism method to obtain a diffusion representation and a diffusion probability.

[0106] like Figure 2 As shown, the present invention also provides a specific workflow of the social network information diffusion prediction method in practical application:

[0107] S101: Dynamic Diffusion Tree Set Construction:

[0108] The dynamic diffusion tree set is constructed based on the local diffusion relationship of information. Figure 3 is an example diagram of the dynamic diffusion tree set construction of the present invention, Figure 3 The left side of the arrow is the S201 dynamic diffusion path construction. Figure 3 The right side of the arrow is the process of constructing the dynamic diffusion tree of S202. Figure 3 As shown, the specific method of constructing the dynamic diffusion tree set in the present invention is:

[0109] S201: Dynamic diffusion path construction:

[0110] Generally speaking, information diffusion in social platforms is spread in a cascade form, and each specific message m i ∈M diffuses in user nodes Form a cascade Where t is the specific time when diffusion occurs. Figure 3 In the cascade c i can be expressed as

[0111] Since the diffusion time nodes between users are different, it is difficult for the model to learn the representation. Therefore, timestamps are used to replace all time information within a specific time interval, such as Figure 3 As shown, the time axis is divided into equally spaced time intervals Δt = T / k. The start time of each interval is used as the timestamp tmp and the diffusion time of all user nodes in the interval is replaced by the timestamp. This technology is a common means of information diffusion prediction. The process will not be repeated here. Please refer to the literature "Chunyuan Yuan, Jiacheng Li, Wei Zhou, Yijun Lu, Xiaodan Zhang, and Songlin Hu. 2020. DyHGCN: A Dynamic Heterogeneous Graph Convolutional Network to LearnUsers Dynamic Preferences for Information Diffusion Prediction. In The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases 2020. Based on the timestamp replacement technology, the cascade is

[0112] S202: Dynamic Diffusion Tree Construction:

[0113] Integrate each diffusion cascade Constructed as a dynamic diffusion tree structure Each dynamic diffusion tree r i Corresponding to a specific message m i and the diffusion cascade c that forwards the message i . Collect all dynamic diffusion trees according to the message number to build a dynamic diffusion tree set Each subtree needs to record the adjacency matrix A of its diffusion process i .

[0114] S102: Heterogeneous Information Bipartite Graph Construction:

[0115] Construction of heterogeneous information bipartite graph based on global diffusion relation of information. Figure 4 This is an example diagram of the construction of a heterogeneous information bipartite graph according to the present invention. Figure 4 The left side shows the S202 dynamic diffusion tree construction process. Figure 4 The middle is the S301 diffusion information tree reconstruction process, Figure 4The right side shows the process of constructing the heterogeneous information bipartite graph of S301. Figure 4 As shown, the specific method for constructing a heterogeneous information bipartite graph in the present invention is:

[0116] S301: Information diffusion tree reconstruction:

[0117] In the dynamic diffusion tree set, each tree r i As a unit, eliminate all edges in the tree and diffuse the message m i is the root node, all user nodes u are first-order leaf nodes, and a reconstruction diffusion tree with a depth of 2 is constructed. Collect all reconstructed diffusion trees into a set according to the message number

[0118] S302: Heterogeneous Information Bipartite Graph Construction:

[0119] According to the forwarding relationship between user node u and message node m, user node u is used as the connection medium to connect and reconstruct all subtrees in the diffusion tree set. Constructing a bipartite graph with heterogeneous information where ε r Is a user node and spread the news The undirected edge set between them is used to record the adjacency relationship between each user node and the message node in the form of a set.

[0120] S103: Representation Learning and Diffusion Prediction:

[0121] According to the dynamic diffusion tree set based on local relations and the heterogeneous information bipartite graph based on global relations constructed in steps S101 and S102, the model will perform diffusion representation learning and information diffusion prediction. Figure 5 This is a flow chart of the present invention characterizing learning and diffusion prediction. Figure 5 As shown in Figure 2, the specific steps in representation learning and diffusion prediction include:

[0122] S401: Learning local representations of temporal feature graph convolutional neural networks:

[0123] According to the dynamic diffusion tree set learned in S202 And the corresponding adjacency matrix A i , for dynamic diffusion tree set Each diffusion tree r i All user nodes u in i Perform local diffusion representation learning, and the calculation formula is:

[0124]

[0125] Among them, this operation uses two graph convolutional neural networks; E tis the time feature matrix, which records the time when the user diffuses the message; is the parameter matrix; σ(·) is the activation function; X i and They are the user representations of all users in each specific message and the learned diffusion representation, namely the first local diffusion representation.

[0126] S402: Fusion of multi-information local representations:

[0127] In response to what we learned in S401 For the diffusion representation of all users in each specific message, it is necessary to fuse the local representations of the same user based on different messages H local , the fusion formula is:

[0128]

[0129] S403: Graph Attention Network Global Representation Learning:

[0130] In order to obtain the local diffusion characterization H local After that, it is necessary to use the heterogeneous information bipartite graph learned in S302 As well as the set of adjacency relationships between its user nodes and message nodes for global representation learning.

[0131] User-based bipartite graph with heterogeneous information With Message For two types of nodes, we design two different meta-paths: 1) user-centered relationships (m→u→m); 2) message-centered relationships (u→m→u). These two types of relationships are encoded into the node representation, and representation learning is performed on the two meta-paths at the same time. The calculation formula is:

[0132] β=softmax(LeakyReLu(β T [H m ||H Local ]))

[0133] γ=softmax(LeakyReLu(γ T [H Local ||H m ]))

[0134] Among them, H m is the feature attribute of the message text; || is the concatenation operation; β and γ are the attention weights of each message node and user node.

[0135] Secondly, in order to capture multiple different attention representations from the global relationship, this paper extends the traditional graph attention network and adopts a multi-head approach to learn global representations. This method is similar to multi-head attention, and its calculation formula is:

[0136]

[0137]

[0138] Among them H m′ It is a new message representation that is iterated through a meta-path centered on the message node; H global is the global diffusion representation of the user through a round of iteration of the meta-path centered on the user node; K is the number of attention heads; W m , W u is the parameter matrix. In each iteration, each node learns similarity from its neighboring nodes. Through multiple iterations, each user node can learn a diffuse global relationship representation.

[0139] S404: Fusion of local and global representations:

[0140] In order to obtain the local diffusion characterization H local With global diffusion characterization H global , a heuristic algorithm is used to fuse the two relationship representations, and the calculation formula is:

[0141] H′=[H local ;H global ;H local ⊙H global ;H local -H global ]

[0142]

[0143] Where ⊙ represents element-by-element multiplication; - represents element-by-element subtraction. This step allows the model to autonomously learn the weights of the two diffusion relationships for diffusion prediction, learn the intermediate process H′, and finally obtain the user diffusion representation

[0144] S405: Information Diffusion Prediction:

[0145] Finally, in order to better capture relevant information in the diffusion process and obtain more accurate prediction results, a prediction module with a multi-head self-attention mechanism is used to perform information diffusion prediction. The prediction process is as follows:

[0146]

[0147]

[0148] Among them is Query, Yes key, is the value; is the corresponding parameter matrix; k is the number of heads in the multi-head self-attention mechanism; d k Diffusion characterization for users Dimension d / k; head i is the diffusion prediction representation learned under a single attention mechanism; the information Z learned by fusing the multi-head attention mechanism is the final diffusion prediction representation learned. The multi-head attention mechanism can more comprehensively utilize multiple feature attributes and is very suitable for learning multiple feature attributes under the two diffusion relationships of the present invention.

[0149] A two-layer fully connected neural network is used to calculate the probability, and the calculation formula is:

[0150]

[0151] in is the cross entropy loss function; is the predicted diffusion probability, y i is the real diffusion state; when y i = 1, the user forwards information within the diffusion prediction time interval; when y i = 0, the user does not forward information; W1, W2, b1, b2 are parameter matrices, |U| is the user set The number of users in the model; θ represents all the parameters that need to be learned in the model.

[0152] This invention is highly versatile and can meet the needs of a variety of downstream tasks across different social platforms. By dynamically controlling the time intervals for different downstream tasks, the model can autonomously learn the impact of local and global diffusion relationships on diffusion prediction, thereby improving the accuracy of information diffusion prediction.

[0153] To better illustrate the technical solutions and effects of the present invention, a real social media dataset is used as an example to illustrate the present invention. As shown in Table 1, this example uses data from the Twitter platform. This dataset was collected in October 2010 and contains 12,627 user nodes. There are 3,442 diffusion messages, or 3,442 diffusion cascades. The average number of diffusion users per cascade is 32.60, with the longest super cascade having 2,368 diffusion users and the shortest micro cascade having only 4 diffusion users.

[0154] Table 1 Data statistics of Twitter dataset

[0155] Dataset Twitter #number of users 12,627 #Number of spread information 3,442 #Average cascade length 32.60 # Maximum length of cascade 2368 # Minimum length of cascade 4

[0156] In this embodiment, a dynamic diffusion tree set is first constructed: the dynamic diffusion tree set contains 3,442 subtrees, each representing the complete diffusion process of a specific message. Each subtree records the local user attributes, the message's diffusion relationship between users, and the diffusion time. Next, a heterogeneous information bipartite graph is constructed: the heterogeneous information bipartite graph contains two types of nodes: user and message, and a "user-message" meta-path constructed by undirected edges. The entire bipartite graph is split and reconstructed based on the dynamic diffusion tree set. The graph contains 12,627 user nodes and 3,442 message nodes. The heterogeneous diffusion graph represents the diffusion process of the entire message group, recording the textual attributes of the messages, the global attributes of the users, and the correlation between users and messages during the entire diffusion process.

[0157] We then use techniques such as graph convolutional neural networks and graph attention networks with temporal information for representation learning, and employ a multi-head self-attention mechanism for diffusion prediction. The model autonomously learns the importance of local and global relationships in the diffusion process for information diffusion prediction, calculates weights, and performs representation fusion. The specific prediction results are shown in Table 2:

[0158] Table 2 Experimental results of all models on Twitter dataset

[0159]

[0160] Among them, LGDP is the design model of the present invention. LGDP-noL is an ablation model for the local representation of the present invention, and LGDP-noG is an ablation model for the global representation of the present invention. SNIDSA (Zhitao Wang, Chengyao Chen, and Wenjie Li. 2018. A sequential neural information diffusion model with structureattention. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management) and DyHGCN (same reference in S201) are currently more advanced prediction models in this field.

[0161] In addition, the dataset was split into three sets in a ratio of 8:1:1. Secondly, through dynamic temporal partitioning of the dataset, the present invention can spontaneously perform multi-scale information diffusion prediction. Therefore, the present invention uses two evaluation methods, Mean Average Precision (MAP) and Hit-To-True Score (HITS), and uses prediction performance on the top 10, top 50, and top 100 scales as experimental evaluation indicators.

[0162] As shown in Table 2, the LGDP model of the present invention has the best prediction results than the other four models.

[0163] Compared to existing information diffusion prediction baseline models SNIDSA and DyHGCN, the proposed method achieves an overall improvement of 8%-11% across three top-k metrics: mean average precision (MAP) and hit score (HITS). This demonstrates its superiority over currently popular diffusion prediction models. Compared to LGDP-noL, which ablates local representations, and LDGP-noG, which ablates global representations, the proposed model achieves an overall improvement of 5%-9% across three top-k metrics: mean average precision (MAP) and hit score (HITS). This demonstrates the importance of the proposed method's representation learning, which bridges local and global relationships and allows the model to autonomously learn both diffusion relationships, for improving information prediction accuracy. Compared to SNIDSA and DyHGCN, LGDP-noL and LDGP-noG achieve superior prediction results, demonstrating that even when predicting diffusion using a single diffusion relationship, the detailed classification of multiple diffusion attributes and effective representation learning are key to improving diffusion prediction accuracy. In summary, the proposed method is effective for predicting information diffusion on online social media platforms.

[0164] The present invention first constructs models based on the local and global diffusion relationships during the information diffusion process. It then employs a graph convolutional neural network (GCNN) and a graph attention network (GAN) with temporal features, both modified for social media, for representation learning. This approach accurately and effectively integrates multiple diffusion attributes into the learning of these two diffusion relationship representations. Furthermore, the invention can autonomously learn the importance of both representations for information diffusion prediction, integrating the local and global diffusion representations to jointly predict information diffusion. This invention improves upon existing information diffusion prediction models, effectively enhancing the versatility and effectiveness of information diffusion prediction for online social media.

[0165] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0166] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for predicting information diffusion in social networks, characterized in that: include: Get the spread of news on social media; Constructing a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message; Reconstruct the dynamic diffusion tree in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph; according to the forwarding relationship between the user node u and the diffusion message m, use the user node u as an intermediary to connect all the reconstructed diffusion trees Constructing a bipartite graph with heterogeneous information where ε r Is a user node and spread the news The set of undirected edges between ; Performing local diffusion representation learning and fusion based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a local diffusion representation of the user; Performing global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation; Information diffusion prediction is performed using an attention mechanism method according to the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and a diffusion probability.

2. The social network information diffusion prediction method according to claim 1, characterized in that: The step of constructing a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message specifically includes: Constructing a dynamic diffusion tree according to the diffusion message and the diffusion cascade corresponding to the diffusion message; Diffusion time is assigned and collected for all diffusion processes of each dynamic diffusion tree to obtain a dynamic diffusion tree set and an adjacency matrix of the diffusion process.

3. The social network information diffusion prediction method according to claim 1, characterized in that: The reconstructing the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph specifically includes: Deleting all edges in the dynamic diffusion tree, and reconstructing the tree with the diffusion message as the root node and the user node as the first-order leaf node according to a set depth to obtain a reconstructed diffusion tree; Collecting the reconstructed diffusion trees according to the serial numbers of the diffusion messages to obtain a reconstructed diffusion tree set; All reconstructed diffusion trees in the reconstructed diffusion tree set are connected according to user nodes to obtain a heterogeneous information bipartite graph.

4. The social network information diffusion prediction method according to claim 1, characterized in that: The performing local diffusion representation learning and fusion based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain the user local diffusion representation specifically includes: Performing local diffusion representation learning on all user nodes of each dynamic diffusion tree in the dynamic diffusion tree set according to the adjacency matrix of the diffusion process to obtain a first local diffusion representation; The first local diffusion representations corresponding to different diffusion messages of the same user are fused to obtain a user local diffusion representation.

5. The social network information diffusion prediction method according to claim 1, characterized in that: The performing global representation learning based on the user local diffusion representation and the heterogeneous information bipartite graph to obtain the user global diffusion representation specifically includes: Performing representation learning on the user nodes and the message nodes in the heterogeneous information bipartite graph to obtain attention weights of the message nodes and the user nodes; A global representation learning is performed using a graph attention network according to the attention weights of the message node and the user node and the user's local diffusion representation to obtain a user's global diffusion representation.

6. The social network information diffusion prediction method according to claim 1, characterized in that: The information diffusion prediction is performed using an attention mechanism method based on the user local diffusion representation and the user global diffusion representation to obtain a diffusion representation and a diffusion probability, specifically including: A user diffusion representation is obtained by fusing the user local diffusion representation and the user global diffusion representation using a heuristic algorithm; According to the user diffusion representation, the attention mechanism method is used to predict information diffusion to obtain the diffusion representation and diffusion probability.

7. A social network information diffusion prediction system, characterized in that: include: An acquisition module is used to obtain diffusion messages from social media; A local diffusion construction module, configured to construct a dynamic diffusion tree set and an adjacency matrix of a diffusion process according to the diffusion message and the diffusion cascade corresponding to the diffusion message; The global diffusion reconstruction module is used to reconstruct the dynamic diffusion trees in the dynamic diffusion tree set to obtain a heterogeneous information bipartite graph; according to the forwarding relationship between the user node u and the diffusion message m, the user node u is used as an intermediary to connect all the reconstructed diffusion trees. Constructing a bipartite graph with heterogeneous information where ε r Is a user node and spread the news The set of undirected edges between ; A local diffusion representation learning module, configured to perform local diffusion representation learning and fusion based on the dynamic diffusion tree set and the adjacency matrix of the diffusion process to obtain a user local diffusion representation; A global diffusion representation learning module, configured to perform global representation learning based on the user's local diffusion representation and the heterogeneous information bipartite graph to obtain a user's global diffusion representation; The information diffusion prediction module is used to perform information diffusion prediction based on the user local diffusion representation and the user global diffusion representation using the attention mechanism method to obtain the diffusion representation and diffusion probability.

8. The social network information diffusion prediction system according to claim 7, characterized in that: The local diffusion building module specifically includes: A construction unit, configured to construct a dynamic diffusion tree according to the diffusion message and the diffusion cascade corresponding to the diffusion message; The diffusion time assignment and collection unit is used to assign and collect diffusion time for all diffusion processes of each dynamic diffusion tree to obtain a dynamic diffusion tree set and an adjacency matrix of the diffusion process.

Images