Information propagation scale prediction method based on multi-feature fusion

Abstract

The invention discloses an information propagation scale prediction method based on multi-feature fusion. The method comprises the following steps: firstly, preprocessing collected information propagation link data and constructing an attention network and a forwarding network; then, extracting multi-dimensional features from three dimensions of a space structure, a time evolution process and a text attribute of information diffusion; and finally, fusing the multi-dimensional features, and inputting the fused multi-dimensional features into a machine learning prediction model for training and prediction so as to realize automatic and accurate prediction of the propagation scale of the information on the social network. According to the method, through multi-feature fusion, the problems of incomplete feature mining and poor prediction precision in the prior art are solved, the information propagation process can be flexibly and effectively quantified, the prediction accuracy and interpretability are remarkably improved, and the method can be widely applied to the fields of network public opinion analysis, hotspot discovery and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of social network analysis, and relates to an information propagation scale prediction method based on multi-feature fusion. BACKGROUND

[0002] With the rapid development of Internet technology and the rapid growth of mobile user scale, the role of users in information propagation gradually changes from passive receivers to active producers and propagators, which complicates the network information environment and brings new challenges to the governance of network public opinion in the new era. Information propagation prediction, as an important part of social network analysis, has high research significance and application value.

[0003] Existing information propagation prediction methods can be mainly divided into two categories: one is to build a machine learning model based on a feature set, and the other is to build an infectious disease model based on a complex network. In the direction of machine learning, some researches build models from information quantity, emotion and user influence, or analyze forwarding behavior from the perspectives of post publishers and text content, or aggregate time processes and content features for prediction. However, these existing researches have the following limitations: first, they are often limited to a single public opinion event, and lack quantitative analysis of the information propagation environment and propagation continuity; second, the construction of model input features lacks comprehensiveness, and most of them only seek features from the single perspective of publisher's own elements or network structure, and do not fully mine network features and user characteristics in the propagation process, resulting in poor prediction accuracy. SUMMARY

[0004] The present application aims to provide an information propagation scale prediction method based on multi-feature fusion, which aims to solve the problems of insufficient mining of network features and user characteristics in the propagation process and poor prediction accuracy in the existing information propagation scale prediction method based on machine learning.

[0005] The technical solution for achieving the purpose of the present application is: an information propagation scale prediction method based on multi-feature fusion, comprising the following steps:

[0006] Step S1: obtaining information propagation data in a social network, constructing an attention relationship network and a forwarding relationship network, the attention relationship network and the forwarding relationship network being topological representations of user relationships, wherein the attention relationship network is constructed based on user static social relationships, the forwarding relationship network is constructed based on dynamic propagation paths, and the node and edge attributes of the two are different;

[0007] Step S2: Based on the network constructed in step S1, extract information diffusion space structure features, the information diffusion space structure features of the attention relationship network include at least one of node fan number, node attention number, local network structure centrality, node reciprocity coefficient, and forwarding network clustering coefficient, the information diffusion space structure features of the forwarding relationship network include at least one of forwarding network density, hidden fan feature, forwarding scale, forwarding depth, forwarding breadth and boundary network feature;

[0008] Step S3: Based on the time information of the information propagation data, extract information diffusion time sequence features, the information diffusion time sequence features include at least one of information diffusion period, fluctuation frequency, coefficient of variation and information propagation continuity features;

[0009] Step S4: Based on the text content in the information propagation data, extract information text attribute features, the information text attribute features include at least one of text topic category, topic label feature, text length and whether to contain specific keywords or media elements;

[0010] Step S5: Fusion of multi-dimensional features extracted in steps S2, S3 and S4 to form a multi-feature fusion vector, and input the multi-feature fusion vector into a pre-trained machine learning model to output a prediction result of the information propagation scale.

[0011] Further, step S1: After obtaining the information propagation data in the social network, the information propagation data is preprocessed, including invalid data cleaning, missing data processing and redundant data removal, and based on the preprocessed data, the attention relationship network and the forwarding relationship network are constructed, wherein:

[0012] Invalid data cleaning, if there are records in the data set that are contrary to the actual logic of information propagation, such as information forwarding time earlier than publishing time, forwarding chain self-closing loop, multi-level time inversion, user ID missing or abnormal, cross-platform timestamp disorder, etc., they are excluded to ensure the validity and consistency of the data;

[0013] Missing data processing, for information content with missing propagation link data, if the missing data affects the connectivity of the information diffusion network, it is excluded, otherwise it is retained;

[0014] Redundant data removal, first sort each propagation data in time order, sort each relationship data according to node id, then traverse and delete information points with repeated time points or repeated node relationships in turn, and delete data items.

[0015] Further, step S2: Extract information diffusion space structure features, wherein:

[0016] I. Information diffusion space structure features of the attention relationship network

[0017] (1) Node Fans

[0018] Node Fans is the in-degree of the node in the attention relationship network, which is:

[0019] ;

[0020] In the above formula, is the number of fans of node user , is the in-degree of node u to node v, is the in-degree of node u to node v with distance 1;

[0021] (2) Node Followers

[0022] Node Followers is the out-degree of the node in the attention relationship network, which is:

[0023] ;

[0024] In the above formula, is the number of followers of node user , is the total number of nodes in the local network, is the out-degree of node u to node v, is the out-degree of node u to node v with distance 1;

[0025] (3) Local Network Structural Centrality

[0026] Local Network Structural Centrality is obtained by calculating the PageRank value of the local network formed by all nodes within two layers of the node to be measured, which is:

[0027] ;

[0028] In the above formula, is the local network structural centrality of node , is the damping factor, is the total number of nodes in the local network, is the set of neighbor nodes of node u with distance 1, is the number of neighbor nodes of node , is the structural centrality of node u;

[0029] (4) Reciprocity Coefficient ​​​​

[0030] The node reciprocity coefficient is used to measure the degree of forming bidirectional connections between node pairs, and the calculation formula is:

[0031] ;

[0032] wherein, represents the reciprocity coefficient of the node, represents the number of bidirectional edges of the node in the directed network, is the total number of edges of the node ;

[0033] (5) Forwarding network clustering coefficient (clustering coefficient)

[0034] The forwarding network clustering coefficient is the average of the local clustering coefficients of all nodes in the forwarding relationship network structure, and is used to measure the degree of node aggregation in the forwarding network. Denoted as , the formula is:

[0035] ;

[0036] wherein, is the forwarding network clustering coefficient, is the forwarding user set, is the local clustering coefficient of the node , which is calculated based on the adjacency matrix of the forwarding network. The calculation of the local clustering coefficient is the same as above;

[0037] II. Information diffusion space structure characteristics of the forwarding relationship network

[0038] (6) Forwarding network density (density)

[0039] The forwarding network density is used to measure the sparsity of the forwarding network structure, and the calculation formula is:

[0040] ;

[0041] wherein, is the forwarding network density, is the total number of directed edges actually participating in forwarding, is the forwarding user set;

[0042] (7) Hidden fan feature (HFF)

[0043] The hidden fan feature includes two sub-features:

[0044] Hidden fan size (HS): refers to the number of users that exist in the attention network but do not appear in the forwarding network, and the formula is:

[0045] ;

[0046] wherein, is a set of concerned network nodes, is a set of forwarding network nodes;

[0047] Structured Forwarding Coefficient (SFC): used to quantify the degree of information dependence on the "follow" and "fan" relationship of social networks, and its calculation formula is:

[0048]

[0049] wherein, is the structured forwarding coefficient, is the forwarding network, is the structured forwarding number, is the total number of information forwarding;

[0050] (8) Forwarding scale (S)

[0051] Forwarding scale refers to the number of forwarding within a certain period of time after the information is published;

[0052] (9) Forwarding depth (D)

[0053] Forwarding depth represents the deepest level that information can reach from the source of propagation;

[0054] (10) Forwarding breadth (B)

[0055] Forwarding breadth represents the maximum value of the number of forwarding between each layer from the deepest level that information can reach from the source of propagation;

[0056] (11) Boundary network feature (BNF)

[0057] The boundary network feature is used to quantify the degree of information propagation to the edge of the network, including the following sub-features:

[0058] Boundary ratio: calculate the proportion of nodes with degree 1 (leaf nodes) in the total nodes in the forwarding network. In the forwarding network, the degree refers to the out-degree of the node, and the degree of 1 indicates that the node has no subsequent forwarding, i.e. the end of the propagation chain. The formula is:

[0059]

[0060] wherein, is a set of forwarding network nodes, denotes the out-degree of node v;

[0061] End Node Depth: The average depth of end nodes (i.e., nodes with no further forwarding, out-degree is 0) in the information propagation chain is calculated. The depth is defined as the shortest path length from the propagation source node (root node) to the end node. This feature reflects the average distance of information penetration to the network edge, and the formula is:

[0062]

[0063] where, denotes the set of end nodes (out-degree is 0), is the shortest path length from the propagation source to node v (calculated by breadth-first search BFS).

[0064] Further, step S3: Based on the time information of the information propagation data, extract the information diffusion time series features, where:

[0065] (1) Information diffusion period

[0066] The information diffusion period starts from the time of information release and ends with the last user's forwarding time , as a measure of the length of information propagation:

[0067]

[0068] where, denotes the forwarding time series of a certain information in its life cycle, denotes the time of the last forwarding in the information propagation process, the last forwarding refers to the forwarding within 72 hours without further forwarding, denotes the time of information release;

[0069] (2) Number of fluctuations

[0070] The number of fluctuations is obtained by detecting the information forwarding time series using the Pettitt mutation point test method, and has:

[0071]

[0072] where, The calculation formula is as follows:

[0073]

[0074] where, is a statistical variable, and the calculation formula of the significance test probability is:

[0075]

[0076] wherein, denoted as the significance test probability; is the maximum mutation intensity; T represents the length of the sample sequence, for standardization, only related to the length of the sample;

[0077] (3) Coefficient of variation (CV) )

[0078] The coefficient of variation is used to measure the variation of each observation value in the retweet size time series, and has:

[0079]

[0080] wherein, and are the standard deviation and mean of the time series, respectively;

[0081] (4) Information propagation continuity feature

[0082] The information propagation continuity feature is used to quantify the time difference between the user receiving information and forwarding information. For any information propagation path , wherein represents the time at which the user propagates the information, and the information propagation continuity feature is calculated as follows:

[0083] .

[0084] Further, step S4: based on the text content in the information propagation data, extracting information text attribute features, wherein:

[0085] The text topic category uses the TF-IDF algorithm to extract the feature vector of the text content, and uses the LDA topic model to divide the text content into topic categories.

[0086] Further, step S5: fusion of the multi-dimensional features extracted in steps S2, S3 and S4 to form a multi-feature fusion vector, and inputting the multi-feature fusion vector into a pre-trained machine learning model to output a prediction result of the information propagation size, wherein:

[0087] The construction of the multi-feature fusion vector is realized by weighted splicing, and the weights are dynamically allocated based on the feature importance. For the information diffusion space structure feature, the time series feature and the information text attribute feature, their information gain ratios are calculated respectively, and the weight formula is: wherein, is the weight of the i-th feature, and is its information gain ratio, which is calculated through training data. The dimension of the fused vector is fixed at 100, and if it is insufficient, it is filled with zero values, and if it exceeds, it is reduced by PCA. ​​​

[0088] The prediction stage includes at least one of the following tasks: an early prediction task (PS1H) based on propagation data within 1 hour after the information is released, or a phased prediction task (PSnR) based on propagation data before the information is forwarded n times, where n is adaptively selected from 10 to 100 according to the information type, as follows:

[0089] PS1H task: early popularity prediction based on fixed time window

[0090] Obtain all propagation data of the target information within the time window [T_start, T_start+1 hour] within 1 hour after the information is released;

[0091] Based on the data in the time window, extract the spatial structure features, time series features and text attribute features of information diffusion;

[0092] After feature fusion, input the prediction model and output the final propagation scale prediction value of the information;

[0093] When the final propagation scale prediction value is output in the form of a classification label, the classification threshold is dynamically set based on the quantile of historical propagation scale data;

[0094] PSnR task: phased popularity prediction based on fixed forwarding times

[0095] Monitor the real-time forwarding times of the target information, and when the forwarding times reach the preset value n, obtain all propagation data before this forwarding; wherein the value of n ranges from 10 to 100 and is adaptively selected according to the topic type of the target information;

[0096] Based on the data before the forwarding times reach n times, extract the spatial structure features, time series features and text attribute features of information diffusion;

[0097] After feature fusion, input the prediction model and output a binary classification decision result of whether the final propagation scale of the information will exceed the preset threshold λ;

[0098] Wherein, the preset threshold λ is determined by grid search method, and the probability output by the model is obtained by Sigmoid function, and the final threshold of classification decision is optimized by maximizing the Youden index under the ROC curve.

[0099] Further, the machine learning model in step S5 is a LightGBM model, and the construction and training of the LightGBM model includes the following specific steps:

[0100] (a) Optimal hyperparameter combination is selected for the model by Bayesian optimization method, with the basic configuration: learning rate learning_rate=0.05, number of leaf nodes num_leaves=31, maximum depth of tree max_depth=8, and number of iterators n_estimators=1000;

[0101] (b) 5-fold time series cross-validation is used for model training to avoid data leakage; mean square error is used as the loss function for PS1H task, and cross-entropy is used as the loss function for PSnR task;

[0102] (c) During the training process, the early stopping round is set to 50 to prevent overfitting by monitoring the validation set loss;

[0103] (d) For the text topic category feature, the model directly processes it without one-hot encoding;

[0104] (e) Before inputting the fused feature vector into the model, it is standardized by Z-score.

[0105] Further, in step S5, before inputting the features into the model, feature selection is also performed according to the feature importance, and the feature selection specifically includes the following steps:

[0106] Before inputting the features into the model, feature selection is also performed according to the feature importance, and the feature selection specifically includes the following steps:

[0107] (a) Importance evaluation step: the features are evaluated using the feature importance scores obtained after training the LightGBM model, and the feature importance scores are calculated based on information gain, and the calculation formula is:

[0108]

[0109] Wherein, gain represents the split point gain, and coverage represents the sample coverage;

[0110] (b) Selection criterion determination step: based on the evaluation results, the top K features in the importance score are retained, wherein the value of K is determined by the performance of the validation set; at the same time, the median of the importance score is taken as the threshold, and the features below the threshold are removed;

[0111] (c) Iterative optimization step: the recursive feature elimination method is used to remove the feature with the lowest importance in each round until the model performance decreases by more than a preset threshold.

[0112] An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0113] A computer readable storage medium having stored thereon a computer program, wherein the program, when executed by a processor, implements any of the methods.

[0114] Compared with the prior art, the present application has the following advantages: an information propagation scale prediction method based on multi-feature fusion is adopted to realize automatic prediction of information propagation scale on a social network, the propagation characteristics are mined and the propagation scale is predicted flexibly in combination with information diffusion space structure, propagation time evolution process and information text attribute, the importance of information propagation space-time evolution characteristics to information propagation process quantization is comprehensively considered, and the interpretability of information propagation scale prediction problem is maximized, so that the information propagation scale prediction is better assisted and the prediction accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0115] Figure 1 Attention relationship network topology diagram.

[0116] Figure 2 Forward relationship network topology diagram.

[0117] Figure 3 Structured forwarding (left) and unstructured forwarding (right) diagram.

[0118] Figure 4 Propagation scale prediction algorithm architecture diagram. DETAILED DESCRIPTION

[0119] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0120] The present application provides a multi-feature fusion based information propagation scale prediction method, which comprises the following steps:

[0121] Step 1: Preprocessing of collected information propagation link data

[0122] (1) Data preprocessing

[0123] In a real environment, due to system recording errors and other reasons, invalid data, abnormal points, redundant points and other abnormal data that affect model recognition effect will appear in the information propagation data. Therefore, the first step of the process needs to remove similar abnormal data, and then perform necessary conversion processing on the data for subsequent process. The information propagation data preprocessing part mainly processes invalid data, missing data and redundant data. The preprocessing process is as follows:

[0124] Firstly, invalid data is excluded. If there is data in the data set that does not conform to the actual situation of information dissemination, such as information forwarding time earlier than information publishing time, self-closing loop of forwarding chain, multi-level time inversion, missing or abnormal user ID, and cross-platform timestamp disorder, the data is excluded.

[0125] Secondly, missing data point processing. For information content with missing data in the propagation link, if the missing data affects the connectivity of the information diffusion network, it is excluded, otherwise it is retained.

[0126] Finally, redundant data points are removed. Due to insufficient stability of detection equipment and other problems, information dissemination data has data item duplication phenomenon, which can be manifested in two aspects: timestamp duplication and node relationship (edge) duplication. The repeated data items will affect the later feature extraction and further reduce the model recognition efficiency. Therefore, it is necessary to sort each piece of propagation data in time sequence, sort each relationship data according to node id, and then traverse and delete the information points with repeated data items in turn. The timestamp duplication is judged by whether the time difference of adjacent records after sorting is zero, and the edge duplication is judged by whether the number of edges in the same position in the adjacency matrix is greater than 1.

[0127] In specific implementation, invalid data cleaning is realized by automatic script. For example, a timestamp verification rule is set: if the forwarding time is earlier than the publishing time and the difference exceeds the system fault tolerance threshold, it is recorded as invalid. The redundant data removal uses hash deduplication algorithm.

[0128] (2) Network structure construction

[0129] The information diffusion space feature extraction method of the application constructs the spatial structure of information forwarding from the perspective of network structure, quantitatively represents the information propagation process, and quantitatively describes the effect of information propagation and its influence. In subsequent information diffusion space feature extraction, it is necessary to propose a spatial feature measurement method from the aspects of attention relationship network and forwarding relationship network. Therefore, in the data processing stage, it is necessary to complete the construction of the attention relationship network and the forwarding relationship network.

[0130] The attention relationship network is a topological representation of the relationship between users and also a carrier of information propagation. Given the attention relationship network structure , which is composed of a user node set and a user attention edge set , wherein the number of nodes and the total number of edges of the network are defined as and . The node and edge relationship of the network can also be expressed as an adjacency matrix . At this time, when there is an edge between two nodes, it corresponds to , otherwise it is 0. Similarly, the forwarding relationship network is constructed by the node set and edge set composition.

[0131] For the complete network construction, it is necessary to pay attention to the differences in the attributes of nodes and edges in the attention relationship network and the forwarding relationship network to ensure that the program can correctly substitute the relevant information:

[0132] 1. Node attributes:

[0133] In both networks, nodes represent users, but the focus of attribute extraction is different:

[0134] Attention relationship network: node attributes include user ID, number of fans, number of attention, and other static social indicators. When the program is constructed, the node set V comes from user social relationship data.

[0135] Forwarding relationship network: node attributes focus on dynamic indicators of transmission behavior, such as the user's role in the information transmission chain (such as the first publisher, the forwarder). The node set U_r only contains users who actually participate in forwarding.

[0136] 2. Edge attributes:

[0137] Edge in attention relationship network: represents the attention relationship between users, has directionality (such as user A follows user B), and the edge weight is usually 1 (indicates the existence of the relationship). The edge set E is constructed based on social graph data, and a_{uv}=1 in the adjacency matrix A indicates that u follows v.

[0138] Edge in forwarding relationship network: represents the information forwarding path, has timestamp and propagation level attributes. The edge set E_r is constructed based on forwarding chain data, and each edge contains the forwarding timestamp; a_{uv}=1 in the adjacency matrix indicates that user u forwarded the information published by user v, and the weight can be added (such as the number of forwarding times).

[0139] 3. Program substitution example:

[0140] When constructing the network, the edges of the attention network and the forwarding network need to be distinguished: the edges of the attention network represent the attention relationship between users, and the edges of the forwarding network represent the information forwarding path. When constructing the adjacency matrix, the attention network uses social relationship data, and the forwarding network uses forwarding chain data. Specifically as follows:

[0141] When constructing the attention network, the program obtains the attention relationship list from the social platform API, maps each user to a node, and maps the attention relationship to a directed edge.

[0142] When constructing the forwarding network, the program parses the forwarding chain data: the nodes are forwarding users, the edges are forwarding relationships, and the timestamps are recorded for time feature extraction.

[0143] Step 2: Construct information diffusion space structure features

[0144] For those who focus on relational network structure Constructing the spatial structural characteristics of information diffusion, including:

[0145] (1) Number of fans at each node

[0146] In online social networks, users with a large number of followers often have greater social influence and the ability to influence the behavior of other users than ordinary users. For the same information, the reposting by authoritative users is more likely to trigger the growth of the reposting chain. Therefore, this paper first uses the number of followers of the information publisher as the main quantification method of the blogger's influence. In the network structure, the number of followers of a user is represented by the in-degree of the nodes in the attention relationship network, which is as follows:

[0147]

[0148] In the above formula, Defined as a node user Number of fans Let it be denoted as distance node In-degree with a distance of 1. To expand the measure of influence, this invention introduces a second-order follower count.

[0149]

[0150] in, By traversing the first-order fans of node v ( The second-order neighbors of node v are obtained. In practice, breadth-first search (BFS) is used to traverse two levels starting from node v and count all unique nodes. This expansion helps to capture a wider range of propagation potential.

[0151] (2) Number of followers

[0152] In a network structure, a user's number of follows is represented by the in-degree of nodes in the follower relationship network:

[0153] ;

[0154] In the above formula, For node users Number of followers This represents the total number of nodes in the local network. Let denot be the out-degree of node u from node v. Let it be denoted as distance node The out-degree with a distance of 1.

[0155] (3) Local network structural centrality )

[0156] The local network structure centrality measure comprehensively considers two types of information: the topological information of the local network surrounding a node and the influence feedback from its nearest neighbors. When calculating a node's influence, the local network formed by all nodes within two layers (nearest and second-nearest neighbors) of the node under test is selected. And calculate its PageRank value, denoted as . For local networks Nodes in ,have:

[0157]

[0158] In the above formula, For nodes Local network structure centrality, The damping factor is set to 0.85. This represents the total number of nodes in the local network. Distance node The set of neighboring nodes at a distance of 1 For nodes The number of neighboring nodes.

[0159] (4) Node reciprocity coefficient

[0160] Besides the one-way "followers" and "followers," mutual following is also common in social networks. When a user follows another user, that user may also follow them back. This increases the likelihood of information being exposed to both users simultaneously, thus increasing the frequency of interaction. The reciprocity coefficient of a network is a commonly used metric for directed networks. The reciprocity coefficient of nodes typically measures the degree of bidirectional connection between node pairs, and the formula is:

[0161]

[0162] in, The reciprocity coefficient of a node is represented by its coefficient. Represents nodes in a directed network The number of bidirectional edges, For nodes The total number of edges. Bidirectional edges in information forwarding networks and attention networks usually represent mutual forwarding behavior and mutual attention behavior.

[0163] For forwarding relationship network structure Constructing the spatial structural characteristics of information diffusion, including:

[0164] (1) Clustering coefficient of forwarding network

[0165] The forwarding network clustering coefficient is the average of the local clustering coefficients of all nodes in the forwarding relationship network structure, and is used to measure the degree of node aggregation in the forwarding network. Denoted as , the formula is:

[0166] ;

[0167] The forwarding network clustering coefficient is the average of the local clustering coefficients of all nodes in the forwarding relationship network structure, and is used to measure the degree of node aggregation in the forwarding network. Denoted as . The formula is:

[0168]

[0169] Wherein, is the forwarding network clustering coefficient, is the forwarding user set, is the local clustering coefficient of node , which is calculated based on the adjacency matrix of the forwarding network. The calculation of the local clustering coefficient is the same as above;

[0170] (2) Forwarding network density (density)

[0171] Network density represents the proportion of the number of edges that already exist in the network to the number of possible edges in the network. This paper uses this index to measure the sparsity of the forwarding network structure, and the calculation formula is:

[0172]

[0173] Wherein, is the forwarding network density, is the total number of directed edges actually participating in forwarding, is the forwarding user set;

[0174] (3) Structured forwarding coefficient (Structured Forwarding Coefficient, SFC)

[0175] The structured forwarding coefficient in the forwarding network is used to quantify the degree to which information depends on the "attention" and "fan" relationships of social networks. The larger the structured forwarding coefficient, the more information transmission depends on the attention and fan relationship.

[0176]

[0177] In the formula, is the forwarding network, is the number of structured forwarding (i.e. the number of forwarding through the attention relationship), is the total number of information forwarding.

[0178] (4) Forwarding scale (S)

[0179] Forwarding scale (S): The number of forwarding in a certain time period after the information is published.

[0180] (5) Forwarding depth (D)

[0181] Forwarding depth (D): The depth of forwarding represents the deepest level that information can reach from the source of propagation.

[0182] (6) Forwarding breadth (B)

[0183] Forwarding breadth (B): The breadth of forwarding represents the maximum value of the number of forwarding between each level from the deepest level that information can reach from the source of propagation.

[0184] Step 3: Constructing information diffusion time series features

[0185] (1) Information diffusion period

[0186] In the process of information propagation, the propagation life cycle of an information is defined in this paper from the time of information publishing , and ends with the time of the last user's forwarding . This paper takes the "life cycle length" as a measure of the length of information propagation:

[0187]

[0188] where, represents the forwarding time series of a certain information in its life cycle. represents the time of the last forwarding in the process of information propagation, and the last forwarding refers to the forwarding that is not forwarded again within 72 hours, represents the time of information publishing.

[0189] (2) Number of fluctuations

[0190] For the information forwarding scale time series, based on the Pettitt mutation point test method, this study proposes the concept of "number of fluctuations" . Specifically, for the information forwarding time forwarding sequence , we have:

[0191]

[0192] In the formula, represents the forwarding time series of a certain information in its life cycle, the calculation formula is as follows:

[0193]

[0194] Among them, record It is a statistical variable, and the probability of significance testing is... The formula for calculating the value is:

[0195]

[0196] in, Let this be the significance test probability; The maximum mutation intensity; T represents the sample sequence length. For The standardization is only related to the sample length.

[0197] (2) Coefficient of variation ( )

[0198] Using the coefficient of variation ( The coefficient of variation (CV) is another statistic used to measure the variation and fluctuation of individual observations in a time series of forwarding data. It is calculated as the ratio of the standard deviation to the mean of the time series, denoted as CV. Compared to other statistics, the coefficient of variation can, to some extent, reduce the impact of different means on the comparison of the degree of variation of two or more test data. In the problem of extracting the characteristics of forwarding volume fluctuations, the coefficient of variation measures the stability of the entire diffusion curve, and its calculation method is as follows:

[0199]

[0200] in, and These are the standard deviation and mean of the time series, respectively.

[0201] (3) The continuous nature of information dissemination

[0202] The continuity characteristic of information propagation is used to quantify the time difference between a user receiving information and forwarding it, for any information propagation path. ,in On behalf of users The calculation method for the continuity characteristic of information dissemination is as follows:

[0203]

[0204] Step 4: Construct information text attribute features

[0205] (1) Text topic categories

[0206] Considering the influence of text topic on user attention and transmission intention, this paper considers topic features as important characteristics of quantifying information transmission attributes. In order to mine the topic distribution contained in different information contents, this paper uses TF-IDF algorithm to extract the feature vector of text content, and uses LDA topic model to divide the topic categories of text content.

[0207] (2) Topic label feature (topic)

[0208] Topic labels are often in the form of "# keyword #". Information with topic labels is often a discussion about a certain topic. If the relevant topic is very popular, information with the topic label is more likely to attract more user attention.

[0209] (3) Text length

[0210] The habit of "fast food" information consumption puts forward new requirements for text length. For example, microblogging is a simple and fast social networking site with the nature of sharing life, mainly supporting short text (140 characters) information publishing.

[0211] (3) Whether to mention "forward" and "lottery" words (Lottery)

[0212] Identify whether the text content contains "forward" and "lottery" words. In addition, whether the text contains pictures, whether it mentions others, whether it contains videos, whether it contains red packets, and whether it contains link address (URL) text features.

[0213] Step 5: Build information transmission scale prediction algorithm

[0214] (1) Feature fusion

[0215] Multi-feature fusion vector is constructed by weighted splicing, and the weight is dynamically allocated based on feature importance. For spatial structure features (step S2), time series features (step S3) and text attribute features (step S4), their information gain ratio (Information Gain Ratio) is calculated respectively. The weight formula is: Where, is the weight of the ith feature, and is its information gain ratio, which is calculated through training data. The dimension of the fused vector is fixed at 100, and the insufficient is filled with zero value, and the excess is reduced by PCA.

[0216] (2) Information transmission scale prediction algorithm

[0217] Generally, the information heat or popularity prediction on online social media platforms can be regarded as a regression problem of message forwarding scale or a multi-classification problem of forwarding scale interval. In addition to the forwarding scale, the heat of information transmission is closely related to the information penetration and the influence of information at a certain forwarding stage. Therefore, the information transmission scale prediction problem can be summarized as follows: by mining the transmission paradigm of information on the network social platform, combining the transmission characteristics of information in the early stage of publication, and predicting the transmission range of information at different time periods after publication. Among them, the present application has two different definitions for "early stage", setting two types of experimental tasks:

[0218] 1) PS1H task (early prediction based on fixed time window): This task aims to predict the final transmission scale based on the transmission within 1 hour of information publication.

[0219] Input: All transmission data in the fixed time window [T_start, T_start+1 hour] within 1 hour from the information publication time T_start. Based on the data in this window, extract multi-dimensional features such as spatial structure, time series, and text attributes of information diffusion, and fuse them to form a feature vector (for example, the dimension is fixed to 100).

[0220] Output: Regression value (predict the final forwarding scale) or classification label (such as low, medium, and high heat).

[0221] Regression output: Directly predict the final forwarding times of information.

[0222] Classification output: Divide the final transmission scale into discrete heat levels (such as low, medium, and high). The division of heat levels is based on an adjustable heat benchmark value φ (such as φ=600), which can be determined according to the quantile of historical data (for example, the threshold of the top 20% heat) or optimized by grid search. The threshold space of classification can be defined as: [0, ϕ / 2], [ϕ / 2, ϕ], [ϕ, 3ϕ / 2], [3ϕ / 2, 2ϕ], [2ϕ, +∞]. For example, the final forwarding scale exceeding 1000 times (i.e. a specific threshold instance) can be defined as "high heat". The specific threshold for classification can be dynamically adjusted based on the overall data distribution.

[0223] Model structure: Use LightGBM model, the key optimization includes: optimize the hyperparameters for regression task (such as learning_rate=0.05, max_depth=8); Z-score standardization of input features; and selection of Top K (such as K=30) key features based on feature importance for training and prediction.

[0224] 2) PSnR task: based on the information being forwarded The previous propagation situation builds feature data, and the model predicts the information propagation scale in the later stage.

[0225] PSnR task (phase prediction based on fixed number of forwarding times): This task aims to predict whether the final propagation scale will exceed a preset threshold based on the state of information propagation before reaching a specific number of forwarding n times.

[0226] Input: Real-time monitoring of the number of forwarding times, when the number of forwarding times reaches the preset value n, all propagation data before this forwarding occurs is immediately intercepted. Wherein, n is an integer value between 10 and 100, which can be adaptively selected according to the topic type of the information (for example, n=10 for ordinary topics and n=50 for hot topics). Based on this data intercept point, multi-dimensional features such as spatial structure, time series and text attributes of information diffusion are extracted, and all feature values are calculated and counted based on the state before n times of forwarding.

[0227] Output: The output of the task is a binary classification result, that is, whether the final propagation scale of the information exceeds the preset threshold λ (for example, λ=500 or λ=1000, the specific value can be optimized and determined on the training data by grid search method). The model output is a probability value between 0 and 1, obtained by Sigmoid function. The probability threshold for final classification decision can be optimized by maximizing the Youden index under the ROC curve.

[0228] Model structure: To complete this binary classification task, a dedicated prediction model (such as LightGBM) is constructed. The key optimization includes: setting the objective function of the model as a binary classification loss.

[0229] (3) Model training includes:

[0230] Hyperparameter setting: Optimize the optimal parameters through Bayesian optimization, the basic configuration is: learning_rate=0.05 (adapt to the fast convergence of early prediction), num_leaves=31 (balance complexity and overfitting), max_depth=8 (prevent over-deep tree structure), n_estimators=1000 (early stopping strategy).

[0231] Training process: Use 5-fold time series cross-validation (TimeSeriesSplit) to avoid data leakage. The loss function is mean square error (MSE) for regression task (PS1H) and cross-entropy for classification task (PSnR).

[0232] Model optimization: Introduce class features for direct processing (such as text topic categories), without one-hot encoding, which improves efficiency. Set the early stopping round number (early_stopping_rounds=50) during training to monitor the validation set loss.

[0233] Fusion with features: Standardization (Z-score) of fusion vectors before model input to ensure numerical stability. Built-in feature importance score (gain-based) of LightGBM is used for subsequent feature selection.

[0234] (4) Feature selection includes:

[0235] Importance evaluation: Use the feature importance score (based on information gain) of LightGBM after training, the formula is: Where gain is the split point gain, and coverage is the sample coverage.

[0236] Selection criteria: Keep the Top-K features with importance scores, and the value of K is determined by the performance of the validation set (e.g., K=30). The threshold is set to the median of the importance scores, and features below the threshold are removed.

[0237] Iterative optimization: Use the recursive feature elimination (RFE) method, remove the lowest importance feature in each round until the model performance (such as RMSE) decreases by more than 5%.

[0238] Embodiment

[0239] In order to verify the effectiveness of the scheme, the experiment is designed based on the microblog heat prediction data set. The experiment uses a public microblog data set, covering popular topics from January 2020 to December 2022, a total of 100,000 information dissemination records. Data fields include: user ID, forwarding timestamp, text content, and attention relationship list. In the preprocessing stage, about 12% of invalid data (such as time stamp disorder records) are removed, and the network is constructed according to the method of step S1. The data set is divided into training set (70%), validation set (10%) and test set (20%), ensuring the continuity of time sequence.

[0240] This embodiment uses the propagation characteristics of information published early to predict the propagation of information in the future, and constructs the effective feature set of information propagation spatio-temporal evolution characteristics.

[0241] Table 1 Information propagation scale prediction feature set

[0242]

[0243] For the PS1H prediction task, Table 2 shows the prediction performance of different models on different feature sets. Specifically, for different built-in models, group experiments were conducted, using attention network features (FNF), forwarding network features (RNF), boundary network features (BNF), hidden fan features (HFF), time series features (TF), and information text features (WCF) as input features to observe the prediction accuracy of different models. At the same time, all features (ALLF) and the top 30 features by importance (T30F) were set as model feature inputs to observe the overall performance of the feature set.

[0244] Table 2. Performance Evaluation (RMSE) of the FEPS-LGB Algorithm in Prediction Task PS1H

[0245]

[0246] Similarly, this section sets a popularity index. Five information popularity value spaces were defined: , , , , The forwarding scale was relaxed, and Table 3 shows the prediction performance of different models on different feature sets.

[0247] Table 3. Performance evaluation (ACC) of the FEPS-LGB algorithm in the prediction task PS1H.

[0248]

[0249] As can be seen:

[0250] (1) Random and DisBias, two benchmark methods that do not use any input features, have the worst performance in terms of root mean square error (RMSE) and accuracy (ACC).

[0251] (2) Overall, the FEPS-M algorithm (FEPS-LGB) with built-in LightGBM can accurately predict the scale of message forwarding, with a mean square error (RMSE) of only 92.11 and a classification accuracy (ACC) as high as 0.812.

[0252] (4) From the perspective of features, the performance of the retweet network feature (RNF) and the hidden fan feature (HFF) on the PS1H prediction task is better than that of other types of features, which shows that the diffusion mode of early information has an important influence on the final heat of information, and confirms the importance of potential users in the message forwarding process. To verify the significance of the results, the t-test is used to compare the difference in RMSE between FEPS-LGB and the benchmark method (such as Random), and the p-value is less than 0.01, indicating that the improvement has statistical significance. The ablation experiment further shows that if only a single feature (such as only time series features) is used, the RMSE rises to 275.52, while the multi-feature fusion (ALLF) reduces the RMSE to 94.88, confirming the necessity of the fusion strategy.

[0253] In the feature importance analysis, the retweet network density (Density) and the local centrality (Local PR) rank in the top five, which shows the key role of network structure in early prediction.

[0254] In the PSnR task, the propagation scale prediction is converted into a binary classification problem through relaxation processing, and whether the final propagation scale of the information exceeds the preset threshold is predicted . Similarly, this paper uses random benchmark (Random) and distribution bias benchmark (DisBias) as performance comparison benchmarks, and based on the label distribution imbalance characteristics of the to-be-predicted variable, this paper uses common binary classification indicators such as precision Pr (precision), recall Re (recall), F1 score, and AUC value (area under the ROC curve) to measure the prediction performance of the model.

[0255] It can be found from Table 4 that:

[0256] (1) For the Random and DisBias benchmark methods that do not use any input features and the model based on WCF features, the prediction results have poor performance in each evaluation;

[0257] (2) Overall, the FEPS-M algorithm (FEPS-LGB) proposed in this application can accurately identify whether a message can become popular by inputting a small amount of retweet data, with an AUC of 0.822, a classification accuracy (ACC) of 0.857, and a positive sample recall rate of 0.733.

[0258] (3) From the perspective of features, the performance of the retweet network feature (RNF) and the hidden fan feature (HFF) on the PS1H prediction task is better than that of other types of features, which shows that the diffusion mode of early information has an important influence on the final heat of information, and confirms the importance of potential users in the message forwarding process. To verify the significance of the results, the t-test is used to compare the difference in RMSE between FEPS-LGB and the benchmark method (such as Random), and the p-value is less than 0.01, indicating that the improvement has statistical significance. The ablation experiment further shows that if only a single feature (such as only time series features) is used, the RMSE rises to 275.52, while the multi-feature fusion (ALLF) reduces the RMSE to 94.88, confirming the necessity of the fusion strategy.

[0259] In the definition of information propagation scale prediction problem, this paper predicts the propagation of information in the future based on the propagation characteristics in the early stage of information release (such as the features used in PS1H and PSnR tasks). By fusing the multi-dimensional features extracted in steps S2-S4 (including spatial structure, time series and text attribute features), the information propagation spatio-temporal evolution feature set is constructed to form an effective input vector. The information propagation scale prediction algorithm proposed in this paper optimizes the model performance through feature enhancement strategies, including multi-feature fusion and importance-based feature selection (such as T30F), to achieve accurate prediction.

[0260] Table 4 Classification performance of FEPS-LGB algorithm in prediction task PSnR (n = 10)

[0261]

[0262] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.

[0263] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the present application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for predicting the scale of information dissemination based on multi-feature fusion, characterized in that, Includes the following steps: Step S1: Obtain information propagation data in social networks, and construct attention relationship network and forwarding relationship network. The attention relationship network and forwarding relationship network are topological representations of relationships between users. The attention relationship network is constructed based on static social relationships of users, and the forwarding relationship network is constructed based on dynamic propagation paths. The node and edge attributes of the two are different. Step S2: Based on the network constructed in Step S1, extract the spatial structure features of information diffusion. The spatial structure features of information diffusion in the attention relationship network include at least one of the following: number of node followers, number of node followers, local network structure centrality, node reciprocity coefficient, and forwarding network clustering coefficient. The spatial structure features of information diffusion in the forwarding relationship network include at least one of the following: forwarding network density, hidden follower features, forwarding scale, forwarding depth, forwarding breadth, and boundary network features. Step S3: Based on the time information of the information dissemination data, extract the information diffusion time series features, which include at least one of the following: information diffusion cycle, number of fluctuations, coefficient of variation, and information dissemination continuity features; Step S4: Based on the text content in the information dissemination data, extract information text attribute features, which include at least one of the following: text topic category, topic tag features, text length, and whether it contains specific keywords or media elements; Step S5: Integrate the multi-dimensional features extracted in steps S2, S3 and S4 to form a multi-feature fusion vector, and input the multi-feature fusion vector into a pre-trained machine learning model to output the prediction result of the information propagation scale.

2. The information propagation scale prediction method based on multi-feature fusion according to claim 1, characterized in that, Step S1: After obtaining information dissemination data from social networks, preprocess the data, including removing invalid data, handling missing data, and removing redundant data. Based on the preprocessed data, construct a follower relationship network and a forwarding relationship network, wherein: Invalid data removal: If there are records in the dataset that contradict the actual logic of information dissemination, such as information forwarding time earlier than release time, self-closed forwarding chain, multi-level time inversion, missing or abnormal user ID, or cross-platform timestamp disorder, they will be excluded to ensure the validity and consistency of the data. For missing data processing, if the missing data affects the connectivity of the information dissemination network, it will be excluded; otherwise, it will be retained. Redundant data removal involves first sorting each piece of propagation data in chronological order, then sorting each piece of relational data by node ID, and finally iterating through and deleting information points with duplicate time points or duplicate node relationships, thus deleting data items.

3. The information propagation scale prediction method based on multi-feature fusion according to claim 1, characterized in that, Step S2: Extract the spatial structure features of information diffusion, where: I. Focus on the spatial structural characteristics of information diffusion in relational networks (1) Number of fans at each node The number of followers of a node is represented by the in-degree of the node in the follower network: ； In the above formula, For node users Number of fans Let be the in-degree of node u from node v. Let it be denoted as distance node In-degree with a distance of 1; (2) Number of followers The number of nodes followed is represented by the out-degree of nodes in the follower relationship network: ； In the above formula, For node users Number of followers This represents the total number of nodes in the local network. Let denot be the out-degree of node u from node v. Let it be denoted as distance node Out-degree with a distance of 1; (3) Local network structure centrality ) Local network structure centrality is obtained by calculating the PageRank value of the local network formed by all nodes within two layers of the node under test, as follows: ； In the above formula, For nodes Local network structure centrality, The damping factor, This represents the total number of nodes in the local network. Distance node The set of neighboring nodes at a distance of 1 For nodes The number of neighboring nodes, The structural centrality of node u; (4) Node reciprocity coefficient The node reciprocity coefficient measures the degree to which bidirectional connections are formed between node pairs, and is calculated using the following formula: ； in, The reciprocity coefficient of a node is represented by its coefficient. Represents nodes in a directed network The number of bidirectional edges, For nodes The total number of connected edges; (5) Forwarding network clustering coefficient The forwarding network clustering coefficient is the mean of the local clustering coefficients of all nodes in the forwarding relationship network structure, used to measure the degree of node clustering in the forwarding network. It is denoted as... The formula is: ； in, For forwarding network clustering coefficients, For forwarding user set, For nodes The local clustering coefficients are calculated based on the adjacency matrix of the forwarding network. The calculation is the same as above; II. Spatial Structural Characteristics of Information Diffusion in Forwarding Relationship Networks (6) Forwarding network density Forwarding network density is used to measure the sparsity of the forwarding network structure, and its calculation formula is as follows: ； in, For forwarding network density, This represents the total number of directed edges that actually participated in forwarding. Forwarding user set; (7) Hide Fan Features (HFF) The hidden fan feature contains two sub-features: Hidden follower size (HS): refers to the number of users who exist in the following network but do not appear in the retweet network, calculated as follows: ； in, To focus on the set of network nodes, Forwarding network node set; Structured Forwarding Coefficient (SFC): Used to quantify the degree to which information depends on the relationship between "followers" and "followers" on a social network. Its calculation formula is as follows: ； In the formula, To convert the structured data into coefficients, For forwarding on the network, For structured forwarding count, This represents the total number of messages forwarded. (8) Forwarding scale (S) Forwarding scale refers to the number of times a piece of information is forwarded within a certain period of time after it is published; (9) Forwarding depth (D) Forwarding depth represents the deepest level that information can reach from its source. (10) Forwarding breadth (B) Forward breadth represents the maximum number of forwards per level between the deepest levels that information can reach from its source. (11) Boundary Network Characteristics (BNF) Boundary network features are used to quantify the extent to which information propagates to the network edge, and include the following sub-features: Boundary Ratio: Calculates the proportion of nodes with a degree of 1 in the forwarding network (out-degree). In a forwarding network, degree refers to the out-degree of a node. A degree of 1 indicates that the node has no further forwarding, meaning it is at the end of the propagation chain. The formula is: ； in, For the set of forwarding network nodes, This represents the out-degree of node v; End Node Depth: Calculates the average depth of the end nodes in the information propagation chain. Depth is defined as the shortest path length from the source node to the end node. This feature reflects the average distance information penetrates to the network edge. The formula is: ； in, Represents the set of end nodes. It is the shortest path length from the propagation source to node v.

4. The information propagation scale prediction method based on multi-feature fusion according to claim 1, characterized in that, Step S3: Based on the time information of the information dissemination data, extract the time series features of information diffusion, where: (1) Information diffusion cycle The information dissemination cycle begins when the information is released. It begins, accompanied by the forwarding time of the last user. End, as a measure of the duration of information dissemination: ； in, This represents the time sequence of how a piece of information was forwarded during its lifecycle. This refers to the time of the last forwarding of information during its dissemination process. The last forwarding means that the information has not been forwarded again within 72 hours. This refers to the time when the information was published; (2) Number of fluctuations The fluctuation count was analyzed using the Pettitt mutation point test to evaluate the information forwarding time series. The results obtained from the tests are as follows: ； In the formula, The calculation formula is as follows: ； Among them, record It is a statistical variable, and the probability of significance testing is... The formula for calculating the value is: ； in, This is denoted as the significance test probability; The maximum mutation intensity; T represents the sample sequence length. For The standardization is only related to the sample length; (3) Coefficient of variation ( ) The coefficient of variation is used to measure the variation and fluctuation of each observation in a time series of forwarding scale, and includes: ； in, and These are the standard deviation and mean of the time series, respectively. (4) The continuous nature of information dissemination The continuity characteristic of information propagation is used to quantify the time difference between a user receiving information and forwarding it, for any information propagation path. ,in On behalf of users The calculation method for the continuity characteristic of information dissemination is as follows: 。 5. The information propagation scale prediction method based on multi-feature fusion according to claim 1, characterized in that, Step S4: Based on the text content in the information dissemination data, extract the text attribute features of the information, where: The text topic categories are determined by extracting feature vectors from the text content using the TF-IDF algorithm and then using the LDA topic model to classify the text content into topic categories.

6. The information propagation scale prediction method based on multi-feature fusion according to claim 1, characterized in that, Step S5: Fuse the multi-dimensional features extracted in steps S2, S3, and S4 to form a multi-feature fusion vector, and input the multi-feature fusion vector into a pre-trained machine learning model to output the prediction result of the information propagation scale, wherein: The construction of the multi-feature fusion vector is achieved through weighted concatenation. The weights are dynamically allocated based on the importance of the features. For information diffusion spatial structure features, time series features, and information text attribute features, their information gain ratios are calculated separately. The weight formula is as follows: in, The weight of the i-th feature is denoted by , and its information gain ratio is calculated from the training data. The dimension of the fused vector is fixed at 100 dimensions. If the dimension is insufficient, zero values ​​are filled in. If the dimension is excessive, dimensionality reduction is performed using PCA. The prediction phase includes at least one of the following tasks: an early prediction task based on data disseminated within one hour of information release (PS1H), or a phased prediction task based on data disseminated n times before the information is forwarded (PSnR), where n is adaptively selected between 10 and 100 depending on the type of information, as detailed below: PS1H Mission: Early Heat Prediction Based on Fixed Time Window Obtain all propagation data of the target information within a time window [T_start, T_start+1 hour] within the first hour after its release; Based on data within a time window, spatial structural features, time series features, and text attribute features of information diffusion are extracted. After feature fusion, the data is input into the prediction model, and the final propagation scale of the information is predicted as the output. When the final propagation scale prediction is output in the form of classification labels, the classification threshold is dynamically set based on the quantiles of historical propagation scale data. PSnR Task: Phased Popularity Prediction Based on a Fixed Number of Forwards Monitor the real-time number of forwards of the target information. When the number of forwards reaches a preset value n, obtain all the propagation data before this forward. The value of n ranges from 10 to 100 and is adaptively selected according to the topic type of the target information. Based on the data before the number of forwards reaches n, we extract the spatial structure features, time series features, and text attribute features of information diffusion. After feature fusion, the input is used to predict the model, and the final propagation scale of the output information is determined by whether the binary classification result exceeds the preset threshold λ. The preset threshold λ is determined by grid search, the probability of the model output is obtained by the Sigmoid function, and the final threshold for classification decision is optimized by maximizing the Yoden exponent under the ROC curve.

7. The information propagation scale prediction method based on multi-feature fusion according to claim 6, characterized in that, The machine learning model in step S5 is the LightGBM model. The construction and training of the LightGBM model includes the following specific steps: (a) Select the optimal combination of hyperparameters for the model using the Bayesian optimization method. Its basic configuration is: learning rate = 0.05, number of leaf nodes = 31, maximum tree depth = 8, and number of iterators = 1000. (b) Five-fold time series cross-validation is used for model training to avoid data leakage; mean squared error is used as the loss function for the PS1H task and cross-entropy is used as the loss function for the PSnR task. (c) During training, the number of early stopping rounds was set to 50, and overfitting was prevented by monitoring the validation set loss; (d) For text topic category features, the model processes them directly without one-hot encoding; (e) Before inputting the fused feature vectors into the model, Z-score normalization is performed on them.

8. The information propagation scale prediction method based on multi-feature fusion according to claim 7, characterized in that, In step S5, before inputting features into the model, feature selection is performed based on feature importance. The specific steps for feature selection are as follows: Before inputting features into the model, feature selection is performed based on feature importance. The feature selection specifically includes the following steps: (a) Importance assessment steps: The features are evaluated using the feature importance scores obtained after training the LightGBM model. The feature importance scores are calculated based on information gain, and the calculation formula is as follows: ； Where gain represents the split point gain and coverage represents the sample coverage; (b) Selection criteria determination steps: Based on the evaluation results, retain the top K features in terms of importance score, where the K value is determined by the validation set performance; at the same time, use the median importance score as a threshold to remove features below the threshold. (c) Iterative optimization steps: The recursive feature elimination method is used to remove the least important features in each round until the model performance declines by more than a preset threshold.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images