Topic propagation prediction method and system based on multi-dimensional feature fusion

Abstract

The invention relates to the technical field of network information propagation prediction, and discloses a topic propagation prediction method and system based on multi-dimensional feature fusion, and the method comprises the steps: extracting multi-dimensional features, inputting the multi-dimensional features to a decoder layer of Transform for fusion, and constructing a prediction model for topic propagation prediction; the multi-dimensional features comprise user comprehensive influence, text emotion features, time dynamic features and user interaction behaviors. According to the method, through collaborative optimization of the dynamic user modeling module, the multi-modal feature fusion module and the intelligent time sequence analysis module, the social network propagation prediction accuracy is remarkably improved, and accurate prediction of the topic propagation trend in the social network is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of network information dissemination prediction, and in particular to a topic dissemination prediction method and system based on multi-dimensional feature fusion. Background Art

[0002] The rapid development of internet technology and the widespread use of social networking platforms (such as Weibo, Twitter, and Facebook) have significantly increased the efficiency of information dissemination. While this phenomenon promotes efficient information sharing, it also accelerates the spread of false information and rumors, posing new challenges to personal privacy protection, corporate brand management, and social governance. For example, false information during emergencies can cause public panic, while malicious public opinion in the business sector can severely damage a company's reputation. Therefore, effectively predicting the dissemination trends of topic information on social networks has become crucial.

[0003] Although the information diffusion prediction technology currently used has made some progress, it still has many limitations: 1. Existing methods overly rely on superficial metrics like follower count and repost volume, while neglecting underlying factors like user expertise and content quality. For example, the actual influence of experts in certain fields often far exceeds their follower base, yet this crucial characteristic is difficult to accurately quantify in existing models, resulting in a single dimensional approach to user influence assessment.

[0004] 2. While existing sentiment analysis models can initially assess the emotional inclination of text content, their understanding of textual semantics remains superficial, making it difficult to effectively identify the deeper meaning within complex language contexts such as irony and metaphor. This can lead to misjudgments when faced with potentially controversial content, as they struggle to capture effective emotional characteristics.

[0005] 3. Information dissemination exhibits significant temporal characteristics, with distinct variations in dissemination patterns across time periods and scenarios. However, existing technologies are limited in their ability to capture these dynamic characteristics, making it difficult to accurately predict the trajectory of sudden events.

[0006] 4. Existing communication indicators mainly focus on surface interaction data such as users’ likes, comments, and reposts, and lack analysis of the deep motivations and group characteristics behind user behavior. This makes the existing communication characteristic indicators too simplistic and unable to fully reflect the real interaction of users.

[0007] Furthermore, while existing technologies offer methods for integrating and analyzing multi-dimensional features, they generally rely on simple weighted summation or mechanical concatenation, failing to fully tap into the inherent correlations between features. This fragmented analysis model makes it difficult for the system to fully grasp the complex mechanisms of information dissemination, thus affecting the accuracy and timeliness of predictions. As social media continues to evolve, these technical bottlenecks are becoming increasingly prominent, exacerbating the inaccuracy and risk of topic spread predictions. Summary of the Invention

[0008] The present invention aims to provide a topic propagation prediction method and system based on multi-dimensional feature fusion to solve the problem that the existing prediction model fails to deeply explore the correlation between topic features and cannot achieve comprehensive feature fusion, resulting in low accuracy of topic propagation prediction and difficulty in controlling propagation risks.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions: Solution 1: A topic propagation prediction method based on multi-dimensional feature fusion, which includes extracting multi-dimensional features, inputting the multi-dimensional features into the decoder layer of Transformer for fusion, and constructing a prediction model for topic propagation prediction; the multi-dimensional features include Comprehensive user influence: Using the number of followers, authentication status, quality score of historical posts, and number of citations as key indicators, a user relationship matrix is ​​constructed. The breadth of user relationships and user authority are calculated separately, and these are used as weight factors to iteratively calculate the user's comprehensive influence. Text sentiment features: Extract text sentiment features, classify them into positive and negative categories, and calculate the probability distribution of each category as the text sentiment feature vector; Time dynamic features: Combine the timestamps of published information to construct time series data, perform bidirectional processing on the time series data, and obtain the changing status of nodes at different time points; then concatenate the results of the two directions to generate a time dynamic feature vector; User interaction behavior: Calculates propagation characteristic measurement indicators based on user behavior parameters; and dynamically adjusts the weight coefficients of each behavior parameter based on scenario requirements.

[0010] Solution 2: A topic propagation prediction system based on multi-dimensional feature fusion, which is applied to the above-mentioned topic propagation prediction method based on multi-dimensional feature fusion, including Data collection module, used to obtain raw data related to topic dissemination; Feature extraction module, used to extract user relationship breadth, user authority, emotional features, temporal dynamic features, and user interaction behavior from raw data; Model building module, used to build prediction models based on the improved PageRank algorithm, Text-CNN model, Bi-LSTM model and Transformer model; The prediction output module is used to generate prediction results of topic propagation based on input data.

[0011] The principles and advantages of this solution are: This solution innovatively constructs a three-dimensional feature analysis framework based on user, sentiment, and time series, breaking through the limitations of traditional methods that focus solely on the user dimension. In terms of user feature extraction, it improves on the classic PageRank algorithm, not only considering static attributes such as user authority and relationship breadth, but also incorporating dynamic user behavior to construct a more accurate user relationship network model. This combined dynamic and static user influence assessment method effectively addresses the one-sided nature of existing technologies in assessing user value.

[0012] Secondly, we propose a deep feature fusion mechanism based on Transformer. Unlike traditional methods that simply concatenate features, this solution uses the extracted multi-dimensional features as K, Q, and V, which are then fed into the Transformer decoding layer. This approach achieves intelligent feature fusion through a self-attention mechanism. This method deeply explores the potential correlations between features, significantly improving the effectiveness of feature fusion.

[0013] In terms of time series processing, the innovative use of one-dimensional serialization replaces the traditional multidimensional heterogeneous graph approach. Combined with an improved Bi-LSTM model, this not only solves the challenge of long sequence processing but also enhances the ability to extract temporal features. Particularly noteworthy is that this solution, based on the essential structure of the data, cleverly combines serialization processing (Bi-LSTM) with graph structure analysis (improved PageRank), enabling multi-perspective collaborative analysis of social network data.

[0014] The advantages of applying this solution are: 1) By integrating multi-dimensional features, the prediction accuracy is significantly improved compared to the single-dimensional method; 2) Dynamic and static user modeling makes impact assessment closer to real-world scenarios; 3) The Transformer-based intelligent fusion mechanism effectively captures the complex correlations between features; 4) Innovative time series processing methods have greatly improved the model’s ability to predict transmission dynamics.

[0015] This will build a comprehensive and accurate social network information dissemination prediction system, providing strong technical support for addressing real challenges such as the spread of false information and public opinion warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is an overall process framework diagram of Example 1 of the present invention.

[0017] Figure 2 This is a flow chart of the Transformer model in the first embodiment of the present invention.

[0018] Figure 3 This is a flow chart of the improved PageRank algorithm in Example 1 of the present invention.

[0019] Figure 4 This is a diagram showing the results of an experiment conducted using the method in Example 1 of the present invention. DETAILED DESCRIPTION

[0020] The following is further described in detail through specific implementation methods: Example 1 In this embodiment, a topic propagation prediction method and system based on multi-dimensional feature fusion is provided. By introducing multi-dimensional features such as user influence, text sentiment features, temporal dynamic features, and user interaction behavior, and combining an improved PageRank algorithm, Text-CNN, Bi-LSTM, and Transformer models, feature fusion is performed by inputting into the decoder layer of the Transformer to achieve accurate prediction of topic propagation trends in social networks.

[0021] In this embodiment, a topic propagation prediction method based on multi-dimensional feature fusion is provided. Figure 1 The figure shows the complete process from data collection to multi-dimensional feature fusion, including steps such as user influence assessment, text sentiment feature extraction, temporal dynamic feature modeling, user interaction behavior modeling, and multi-dimensional feature fusion. Specifically, it involves extracting multi-dimensional features, inputting them into the decoder layer of the Transformer for fusion, and building a prediction model to predict topic spread.

[0022] In this example, we use crawler technology to obtain raw data related to the target topic from mainstream social platforms, including user information, text content, posting time, number of reposts, number of likes, etc. The obtained raw data is preprocessed through operations such as denoising, word segmentation, and word vectorization to facilitate subsequent multi-dimensional feature extraction.

[0023] Among them, multidimensional features include user comprehensive influence, text sentiment features, time dynamic features and user interaction behavior.

[0024] 1. User comprehensive influence is used to comprehensively evaluate the influence of users. In this embodiment, in order to conduct a more comprehensive evaluation, an improved PageRank algorithm is used for evaluation. Figure 3The working principle of user feature processing and the process of multi-dimensional feature fusion shown in the figure use the number of user followers, authentication status, historical post quality score, and number of citations as key indicators to construct a user relationship matrix, calculate the breadth of user relationships and user authority respectively, and use them as weight factors to iteratively calculate the user's comprehensive influence. The specific calculation steps are as follows: (1) The number of user followers, authentication status, historical post quality score and number of citations are used as key indicators, and the weight coefficients of each key indicator are designed respectively, expressed as , and must meet , ensuring that the importance of each indicator is reasonably reflected.

[0025] (2) Calculate the user authority by calculating the number of users’ followers and the weight coefficient of the authentication status. In this embodiment, the user authority is used to measure the user’s credibility and influence in the social network. It is evaluated by combining the user’s authentication status, the quality of historical posts, and the number of citations, and is expressed as ; ; Where, Indicates the authority score of the user; is the number of citations; is the maximum number of citations of the user in the corresponding dataset; The user's authentication status, that is, whether it has been verified; They are the weight coefficients of the number of user followers and authentication status respectively.

[0026] (3) Construct a user relationship matrix. Based on the user relationship matrix, the number of users’ followers and authentication status are used as weight factors to perform weighted processing on the matrix, and the user relationship breadth is calculated at the same time. In this embodiment, the user relationship breadth is used to describe the connection range of the user in the social network. Specifically, it is quantified by analyzing the number of users’ followers, the number of fans, and the frequency of their interactions, and is expressed as ; Where, Expressed as user relationship breadth; For users Number of fans; The maximum number of followers among the selected users.

[0027] (4) Based on the improved PageRank algorithm, user authority and relationship breadth are introduced as weight factors, and the influence score of each user is calculated by iterative formula update until convergence. The final score reflects the comprehensive influence of the user and is expressed as ; in, The breadth of user relationships; Authoritativeness for users.

[0028] 2. Text sentiment is one of the important driving factors for information dissemination. In this example, we use Text-CNN combined with the Attention mechanism to complete the sentiment classification task and extract sentiment features. The specific steps are as follows: (1) Preprocess the obtained raw data text content, including word segmentation, stop word removal, word vectorization, etc., to ensure that the format suitable for model input is generated.

[0029] (2) Extract text sentiment features, use convolution kernels of different sizes to perform convolution processing on the acquired text information, extract local semantic features, capture key words in the text and their contextual relationships, and express them as ; ; Where, Expressed as the corresponding weight matrix; Represented as input matrix; is the offset matrix; is the sentiment feature vector processed in the middle; is the final sentiment feature vector after maximum pooling.

[0030] (3) The Attention mechanism is used to weight the key text sentiment features output by the convolutional layer, and the sentiment features are classified into positive and negative categories, and the probability distribution of each sentiment type is calculated. This is used as the text sentiment feature vector to highlight the semantic information that has an important impact on sentiment classification.

[0031] In this embodiment, the key semantic information is weighted to generate sentiment features, which are mapped to the sentiment classification space and divided into three sentiment categories: positive sentiment, negative sentiment, and neutral sentiment. Each sentiment category is assigned a different weight to reflect its impact on the spread of the topic, and the probability distribution of each category is calculated separately as the text sentiment feature vector, which is expressed as ; ; in, is the feature vector obtained from the fully connected layer; To obtain the final probability distribution of emotion categories from the output layer; is the weight matrix of the fully connected layer; is the bias vector.

[0032] 3. Temporal dynamic features are crucial for capturing the temporal patterns of information propagation. In this embodiment, a Bi-LSTM model is used to encode the temporal embedding of nodes to capture temporal dynamic features. The specific steps are as follows: (1) Combine the timestamps of user-posted information to form time series data, recording the user's behavior trajectory at different times.

[0033] (2) The Bi-LSTM model processes time series data bidirectionally to obtain the changing states of nodes at different time points, thereby fully reflecting the potential laws in the time series.

[0034] In this embodiment, Bidirectional processing uses forward LSTM units to capture positive time dependencies and backward LSTM units to capture reverse time dependencies, which can be expressed as ; ; Where, For the direction of progress; For the backward direction; is the information input at the current time step. The model calculates the hidden state (forward direction) and (Backward direction). These hidden states are then used as input to the next time step, ensuring that rich contextual information is propagated through the sequence.

[0035] (3) The results of the two directions are spliced ​​together to generate a temporal dynamic feature vector, which is expressed as .

[0036] In this embodiment, the final time dynamic feature vector generated includes information such as the user's active time period, the periodicity of information release, and the time delay of the propagation path.

[0037] 4. User interaction behavior. User interaction behavior is the core driving force of information dissemination. In this embodiment, by analyzing user behaviors such as the number of likes (L), comments (C), and reposts (S), the measurement method of user interaction traces is improved. The traditional topic page views indicator is replaced with a more accurate dissemination characteristic measurement indicator. The dissemination characteristic measurement indicator is generated by combining the depth and breadth of the dissemination path. The specific steps are as follows: (1) Calculate the propagation characteristic measurement index based on the user behavior parameters. In this embodiment, the behavior parameters include the number of likes, comments, and reposts of the user. Based on the user behavior data such as the number of topic reposts and the number of likes, the propagation characteristic measurement index is calculated. The calculation formula of the propagation characteristic measurement index is: ; ; ; Where, For the topic Direct transmission factors; For the topic At different times By counting the number of forwarding of topics in different time periods, the total number of forwarding of topics is obtained as the direct propagation factor of the topic, thereby obtaining the propagation characteristics of the topic. For the topic Total resonance of For the topic At different times By counting the number of likes for different texts under a topic, the total resonance of the topic can be calculated as one of the dissemination characteristics of the topic. For the topic Total exposures; For the topic At different times By counting the number of comments on a topic at different times, we can get the total number of comments on the topic, and ultimately the propagation characteristics of the topic.

[0038] At the same time, in this embodiment, the weight coefficient of each behavior parameter is dynamically adjusted according to the requirements of different scenarios, and the propagation characteristic measurement index is obtained through weighted calculation, so that the propagation characteristic measurement index is more suitable for actual applications.

[0039] Multi-dimensional feature fusion and prediction model construction.

[0040] After extracting the above multi-dimensional features, the self-attention mechanism is combined to perform weighted fusion on the multi-dimensional features. In this embodiment, in order to give full play to the advantages of each feature, as shown in the attached Figure 2 The improved Transformer structure flow chart shown in the figure includes the processes of indicator introduction, weight allocation, matrix construction, and iterative calculation. The Transformer model is used to fuse multi-dimensional features. The specific steps are as follows: (1) Use the self-attention mechanism to perform weighted fusion of multi-dimensional features to enhance the correlation between features, which is expressed as ; ; (2) Construct a fully connected layer, map the fused features to the probability distribution space of topic propagation, build a prediction model to predict topic propagation, and output the probability distribution of topic propagation, which is expressed as ; Where, They are user comprehensive influence, text sentiment characteristics and time dynamic characteristics; is the result of matrix inversion multiplication; is the dimension of the key or query vector for each attention head, , is the number of heads that the multi-head is concerned about. C is the weight used to mask future timestamps, ensuring that the prediction of position j can only depend on sequences less than j; 、 、 and are respectively learnable weight parameters; is the fusion feature of information i; is the i-th processing result in multi-head attention.

[0041] (3) Visualizing the prediction results. In this embodiment, a key node identification diagram of topic propagation, a visualization diagram of the propagation path, and a dynamic simulation diagram of the propagation trend are generated based on the prediction results to facilitate intuitive understanding by users.

[0042] In this embodiment, this solution can be used to quickly identify hot topics and their dissemination trends, helping relevant departments respond to emergencies promptly and monitor public opinion in a timely manner. It can also predict the spread of false information, take targeted measures to curb its spread, and effectively prevent and control false information. Furthermore, it can analyze the interests and preferences of target user groups, develop precise marketing strategies, improve marketing efficiency, and mine hot topics to discover potential hot topics, providing inspiration for content creators.

[0043] Furthermore, this solution was validated on a real social network dataset using evaluation metrics including FI-Score, AUC, and Recall. FI-Score is used to evaluate the overall performance of the model, combining precision and recall; AUC is used to assess the model's classification capabilities, particularly on imbalanced datasets; and Recall is used to assess the model's ability to identify positive samples, reflecting the model's overall performance.

[0044] Some of the verification results are shown in the attached Figure 4 As shown in the figure, the experimental results show that this solution outperforms traditional machine learning models and single-modal deep learning models in all the above indicators, proving its effectiveness. The application scenarios of this invention are wide-ranging, including but not limited to public opinion monitoring, false information prevention and control, hot topic mining, marketing strategy optimization and other fields.

[0045] In this embodiment, the limitations of existing methods are effectively solved through three key technological innovations. First, different from the traditional single-dimensional analysis, a three-dimensional analysis framework is constructed that integrates user dynamic and static features, emotional semantics and temporal dynamics. Among them, the improved PageRank algorithm realizes the unified quantification of user authority, relationship breadth and dynamic behavior for the first time. Secondly, the Transformer decoding layer is used to replace the traditional feature splicing, and the intelligent fusion of multi-dimensional features is realized through the self-attention mechanism, which significantly improves the effectiveness of feature interaction. Finally, the innovatively designed one-dimensional serialized time series processing method, combined with the sequence modeling capability of Bi-LSTM and the graph structure analysis advantage of the improved PageRank, realizes the multi-perspective collaborative modeling of the complex communication patterns of social networks.

[0046] Example 2 In this embodiment, a topic propagation prediction system based on multi-dimensional feature fusion is provided, which is applied to the above-mentioned topic propagation prediction method based on multi-dimensional feature fusion, including: The data acquisition module is used to obtain raw data related to topic dissemination. In this embodiment, the data acquisition module supports access to real-time data streams and multiple data sources, including but not limited to mainstream social platforms such as Twitter, Facebook, and Weibo, and can automatically adapt to the data formats of different platforms. It can also pre-process the accessed data, including denoising, formatting, and normalization, to ensure data standardization and validity.

[0047] The feature extraction module is used to extract user relationship breadth, user authority, emotional features, temporal dynamic features, and user interaction behaviors from the raw data. In this embodiment, the feature extraction module adopts a distributed architecture to support efficient processing of large-scale data.

[0048] The model building module is used to build a prediction model based on the improved PageRank algorithm, the Text-CNN model, the Bi-LSTM model, and the Transformer model. In this embodiment, the model building module provides a configurable parameter interface, allowing users to adjust model parameters according to specific needs.

[0049] Among them, the Text-CNN model structure includes input layer, convolution layer, pooling layer and fully connected layer.

[0050] The input layer receives preprocessed text data, which is then segmented, stop-word removed, and vectorized before being fed into the model. The convolutional layer extracts local semantic features using convolution kernels of varying sizes. The pooling layer performs dimensionality reduction on the convolutional layer output. The fully connected layer maps the pooling layer output to the sentiment classification space and generates the final sentiment feature vector.

[0051] The Bi-LSTM model structure consists of a forward LSTM part and a backward LSTM unit.

[0052] The forward LSTM unit is used to capture the positive dependencies of time series data. The backward LSTM unit is used to capture the negative dependencies of time series data. The outputs of the two directions are concatenated to generate the final temporal dynamic feature vector.

[0053] The prediction output module is used to generate prediction results for topic propagation based on input data. In this embodiment, the prediction output module supports visual display and can present the trend of topic propagation and the influence distribution of key nodes in the form of charts. Specifically, the prediction output module supports multi-dimensional analysis functions, including identification of key nodes in topic propagation; visualization of propagation paths; dynamic simulation of propagation trends; and comparative analysis of propagation effects in different time periods. This improves information display efficiency, facilitates quick viewing and judgment, and enables timely response measures.

[0054] It also includes a model evaluation module for evaluating the performance of the prediction model. The performance evaluation indicators include FI-Score, AUC and Recall to comprehensively evaluate the model prediction performance and ensure the effectiveness and coverage of the model.

[0055] In this embodiment, a multimodal fusion module enables cross-dimensional feature interaction. This modular design not only enables the system to process complex social network data but also significantly improves predictive performance through collaborative optimization between modules, demonstrating superior real-time and accuracy in emergency communication warning scenarios.

[0056] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A topic propagation prediction method based on multi-dimensional feature fusion, characterized in that: This involves extracting multidimensional features, inputting them into the decoder layer of the Transformer for fusion, and building a prediction model to predict topic spread. The multidimensional features include Comprehensive user influence: Using the number of followers, authentication status, quality score of historical posts, and number of citations as key indicators, a user relationship matrix is ​​constructed. The breadth of user relationships and user authority are calculated separately, and these are used as weight factors to iteratively calculate the user's comprehensive influence. Text sentiment features: Extract text sentiment features, classify them into positive and negative categories, and calculate the probability distribution of each sentiment type, which is used as the text sentiment feature vector; Time dynamic features: Combine the timestamps of published information to construct time series data, perform bidirectional processing on the time series data, and obtain the changing status of nodes at different time points; then concatenate the results of the two directions to generate a time dynamic feature vector; User interaction behavior: Calculates propagation characteristic measurement indicators based on user behavior parameters; and dynamically adjusts the weight coefficients of each behavior parameter based on scenario requirements.

2. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The user's comprehensive influence is expressed as ； in, The breadth of user relationships; Authoritativeness for users.

3. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 2 is characterized in that: The user authority is calculated by calculating the weight coefficient of the number of user followers and authentication status, which is expressed as ； ； Where, is the number of citations; is the maximum number of citations of the user in the corresponding dataset; The user's authentication status; They are the weight coefficients of the number of user followers and authentication status respectively.

4. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 2, characterized in that: According to the user relationship matrix, the number of fans and authentication status are used as weight factors to weight the matrix and calculate the user relationship breadth, which is expressed as ； Where, For users Number of fans; The maximum number of followers among the selected users.

5. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: Perform convolution processing on the acquired text information to extract keywords and contextual relationships; generate sentiment features after weighting key semantic information; The emotional features are mapped to the emotional classification space, divided into three categories: positive, negative and neutral, and the probability distribution of each category is calculated separately.

6. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that: The bidirectional processing uses the forward LSTM unit to capture the positive time dependency and the backward LSTM unit to capture the reverse time dependency, which is expressed as ； ； Where, For the direction of progress; For the backward direction; Input for the information at the current time step.

7. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that: The behavior parameters include the number of likes, comments and reposts of users; the calculation formula of the communication characteristic measurement index is: ； ； ； Where, For the topic Direct transmission factors; For the topic At different times The number of single forwarding; For the topic Total resonance of For the topic At different times Number of likes; For the topic Total exposures; For the topic At different times The number of individual comments.

8. The topic propagation prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that: Use the self-attention mechanism to perform weighted fusion of multi-dimensional features, expressed as ； ； Then construct the probability distribution of topic propagation output by the fully connected layer, expressed as ； Where, They are user comprehensive influence, text sentiment characteristics and time dynamic characteristics; is the result of matrix inversion multiplication; is the dimension of the key or query vector for each attention head, , is the number of heads followed by the bulls; C is the weight used to mask future timestamps; 、 、 and are respectively learnable weight parameters; is the fusion feature of information i; is the i-th processing result in multi-head attention.

9. A topic propagation prediction system based on multi-dimensional feature fusion, characterized by: The topic propagation prediction method based on multidimensional feature fusion applied to any one of claims 1-8 comprises: Data collection module, used to obtain raw data related to topic dissemination; Feature extraction module, used to extract user relationship breadth, user authority, emotional features, temporal dynamic features, and user interaction behavior from raw data; Model building module, used to build prediction models based on the improved PageRank algorithm, Text-CNN model, Bi-LSTM model and Transformer model; The prediction output module is used to generate prediction results of topic propagation based on input data.

10. The topic propagation prediction system based on multi-dimensional feature fusion according to claim 9, characterized in that: It also includes a model evaluation module for evaluating the performance of the prediction model. The performance evaluation indicators include FI-Score, AUC and Recall.

Images