A Deep Learning-Based Intelligent Detection Method for Misinformation on Social Networks

Abstract

This invention discloses an intelligent detection method for disinformation on social networks based on deep learning. The method includes: collecting text information of all messages on a social network over a given time period, processing reply text information and topological information, and performing text preprocessing; inputting the message text information into a local information encoder to output message text feature vectors; inputting the topological information of the social network topology graph formed by the messages into a global information encoder to output message topological information feature vectors; concatenating the message text feature vectors and topological information feature vectors to obtain a fused feature vector; and inputting the fused feature vector into a linear network to output the final prediction vector. This invention, using fixed-length samples while ensuring training efficiency, significantly enhances the model's feature extraction capability by fusing network message content, reply message content, and message topological information in the propagation graph, overcoming the problem of insufficient model feature extraction ability.

Description

Technical Field

[0001] This invention belongs to the field of network information analysis, specifically relating to an intelligent detection method for false statements on social networks based on deep learning. Background Technology

[0002] With the rapid development of the digital society, social networks have long been a primary source of information for people. Amidst this massive amount of data, the intentional or unintentional dissemination of misinformation can mislead and influence people's behavior and thoughts, even seriously endangering public safety and affecting social stability. Therefore, the detection of misinformation has always been an important research topic. To date, many misinformation detection technologies have been proposed, which can be categorized into several main types, including those based on traditional machine learning methods, text features, and dissemination characteristics. However, in today's era, 5G technology has significantly improved network capacity, transmission speed, and network throughput, leading to the further vigorous development of various network technologies. In particular, the emergence of various artificial intelligence applications, such as Chat-GPT and Wenxin Yiyan, has resulted in more difficult-to-distinguish, untrue, and fabricated information appearing online, posing an even greater challenge to misinformation detection.

[0003] Traditional machine learning methods mostly rely on feature engineering to extract manually generated features from event-related information and then use traditional machine learning models for judgment. However, this method involves a huge workload for manually extracting features, and the features tend to be superficial, failing to meet the accuracy and cost requirements for detecting false information. Text feature-based methods primarily extract semantically representative textual features from the text content. Due to the rapid development of various media processing methods and AI tools, many carefully crafted fake news messages are now difficult to distinguish from genuine information based solely on textual features.

[0004] In summary, current research on the classification of large-scale encrypted network traffic still has the following shortcomings: First, with the popularization of artificial intelligence technology, synthetic misinformation on the internet is more deceptive than manually created false information, closely resembling real news in terms of text format, significantly reducing the accuracy of conventional text-based false information detection methods. Second, based on traditional machine learning models, manually designing features is difficult, and obtaining more representative features usually requires time-consuming offline algorithms, resulting in poor real-time performance. Third, to minimize the impact of false information, it should be identified and processed in its early stages. However, because the early propagation patterns are not obvious, propagation-based false information detection methods suffer from a certain time lag in application. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] In view of the problems existing in the above and / or existing intelligent detection methods for social network misinformation based on deep learning, this invention is proposed.

[0007] Therefore, the problem this invention aims to solve is how to provide an intelligent detection method for social network misinformation based on deep learning. To address the above technical problem, this invention provides the following technical solution: an intelligent detection method for social network misinformation based on deep learning, comprising collecting text information, reply information, and topological information of all messages on a social network within a time period and performing text preprocessing; inputting the text information of the messages into a local information encoder to output the text information feature vector of the messages; inputting the reply information of the messages into the local information encoder to output the message-reply fused information feature vector; concatenating the text information feature vector of the messages and the reply-fused information feature vector with different weights to obtain the message local information feature vector; inputting the topological information of the social network topology graph composed of all messages into a global information encoder to obtain the topological information feature vector; concatenating the text information feature vector of the messages and the topological information feature vector together to obtain a fused feature vector; inputting the fused feature vector into a linear network to output the final prediction vector; each value in the prediction vector is a predicted value for various labels, and the label corresponding to the maximum value among the predicted values ​​is taken as the final predicted label.

[0008] As a preferred embodiment of the intelligent detection method for false statements on social networks based on deep learning described in this invention, the text preprocessing operation includes the following steps: traversing each line of the dataset, removing symbols that do not belong to the text, and formatting the text; for each line traversed, storing the ID, source message content, and corresponding tag information of each line into three storage lists: message ID, source message content, and tag; using the message ID as an index to index all reply messages corresponding to the source message ID, storing all reply message texts in a list, and storing this message list into the reply information list to which the source message belongs; concatenating all text data lists together to form a sentence list; creating an empty dictionary; sorting all appearing characters by frequency, using the character itself as the key and the sorted sequence number as the value, and storing the key-value pair in the dictionary; traversing all text sentences and serializing all text symbols; limiting the maximum length of the message, the maximum number of replies to a single message, and the maximum length of a single reply; traversing all text sentences one by one, truncating the sequence portion exceeding the limit length, and expanding the sequence portion below the limit length with all zeros; obtaining the sequence number sequence of social messages and their reply messages.

[0009] As a preferred embodiment of the intelligent detection method for false statements on social networks based on deep learning described in this invention, the step of inputting text information into a local information encoder and outputting a text information feature representation vector of the message includes the following steps: vectorizing the sequence number of the message and its reply to obtain a time-sequence-free text vector; inputting the text vector without time-sequence information into a neural network model, wherein the sequence output by the model contains time-sequence information, wherein any sequence is only related to the other sequences input at the current time step, and is independent of all sequences from other steps; each message sample is currently a single-precision floating-point two-dimensional vector in the form of L×D, where L is the maximum number of characters in the sentence and D is the dimension of a single word vector; by convolving this vector with convolution kernels of different sizes, one-dimensional vector results of different sizes are obtained; the maximum value in each vector is extracted and concatenated together to obtain a new one-dimensional vector, which serves as the final one-dimensional vector of the message text representation in the current step.

[0010] As a preferred embodiment of the intelligent detection method for social network misinformation based on deep learning described in this invention, the following steps are included: concatenating the textual information feature vector of the message and the fused information feature vector of the reply with different weights to obtain the local information feature vector of the message: setting a parameter alpha, which represents the proportion of the original message and the reply in the fused feature vector, with the sum of the proportions being 1. This parameter can be trained, and the final fused text vector can be represented as follows:

[0011] feature text =feature source*alpha+feature replies *(1-alpha)

[0012] Among them, feature text feature source feature replies These represent the text fusion feature vector, the original message text feature vector, and the reply message text feature vector, respectively.

[0013] As a preferred embodiment of the intelligent detection method for false statements in social networks based on deep learning described in this invention, the step of inputting the topological information of the social network topology graph composed of all messages into a global information encoder and outputting the topological information feature vector of the messages includes the following steps: performing the same operation on the reply messages to obtain a one-dimensional feature vector of the reply messages; transforming the vector expression obtained based on the reply message text using a self-attention mechanism to obtain a new reply message feature vector; concatenating the corresponding message vector to the front of each transformed reply message vector, passing it through a linear network and activation layers to finally obtain the weight of each message in relation to the message; calculating the sum of the weights of all replies for each message to finally obtain a fused reply message vector with the same dimension as the message; concatenating the two together to obtain the final message text fusion vector.

[0014] As a preferred embodiment of the intelligent detection method for false statements on social networks based on deep learning described in this invention, the preprocessed graph data is input into a graph attention network for training, where each node in the graph represents a message posted in a social network; the model finally outputs the feature vector of each node to represent the information of the node in terms of topological structure.

[0015] As a preferred embodiment of the intelligent detection method for false statements on social networks based on deep learning described in this invention, the preprocessed graph data includes adjacent edges and random initial node features.

[0016] As a preferred embodiment of the intelligent detection method for false statements on social networks based on deep learning described in this invention, the step of concatenating the textual information feature vector and the topological information feature vector of a message to obtain a fused feature vector includes the following steps: concatenating the obtained message text fused feature vector of the form [n,m] with the message topological information feature vector of the form [n,m] according to the dimension dim = -1, where n is the number of messages and m is the representation dimension of a single feature vector, to obtain a message fused feature vector of the form [n,2m].

[0017] As a preferred embodiment of the intelligent detection method for social network misinformation based on deep learning described in this invention, the step of inputting the fused feature vector into a linear network and outputting the final prediction vector includes the following steps: inputting the message fused feature vector of the form [n, 2m] into a linear network with an input channel of 2m and an output channel of dim, where dim is 4, representing the four labels of the misinformation detection task: true rumor, false rumor, non-rumor, and unverified rumor; the preceding network outputs a set of prediction vectors of the form [n, dim], where each one-dimensional prediction vector corresponds to a prediction value of a message for the four labels; the output vector set is then passed through a softmax activation function to convert the vector element values ​​into probability values ​​of a message corresponding to a certain label, and finally outputs a final set of prediction vectors of the form [n, dim].

[0018] As a preferred embodiment of the intelligent detection method for social network misinformation based on deep learning described in this invention, the following steps are included: for each prediction vector, the label with the largest probability value among the four predicted values ​​is selected as the final predicted label of the model; for all the extracted predicted label sets and sample label sets, the cross-entropy loss between them is calculated, and the total loss is calculated; the model parameters are updated using the backpropagation algorithm based on the calculated loss, and finally the optimal network parameters are obtained.

[0019] The beneficial effects of this invention are as follows: The training method proposed in this invention, which combines the proposed model features and preprocessing methods, eliminates the need for manual feature design by experts. In tests on multiple real network datasets, it outperforms recognition methods based on traditional machine learning. Using fixed-length samples while ensuring training efficiency, this invention significantly enhances the model's feature extraction capability by fusing network message content, its reply message content, and the topological information of the message in the propagation graph. This overcomes, to some extent, the problem of insufficient model feature extraction capability. In tests on multiple real network datasets, for the same model, the classifier trained using the proposed method outperforms the traditional classifier trained using traditional methods. The intelligent detection model for social network misinformation based on deep learning, using the training strategy and classification method proposed in this invention, shows in tests on multiple real network datasets that even when the number of replies is significantly less than the number of replies in the training samples, the model's accuracy does not decrease significantly. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0021] Figure 1 This is a flowchart illustrating the training and classification process of the intelligent detection method for social network misinformation based on deep learning in Example 1.

[0022] Figure 2 This is a schematic diagram of the preprocessing sample normalization of the intelligent detection method for social network misinformation based on deep learning in Example 1.

[0023] Figure 3 This is a schematic diagram of the target model structure of the intelligent detection method for social network misinformation based on deep learning in Example 1.

[0024] Figure 4 This is a schematic diagram of the local information extraction network of the intelligent detection method for false statements on social networks based on deep learning in Example 1. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Example 1

[0029] Reference Figures 1-4 This is the first embodiment of the present invention. This embodiment provides a method for constructing a comprehensive radiation model of a transportation hub based on intelligent prediction. The method for constructing a comprehensive radiation model of a transportation hub based on intelligent prediction includes the following steps:

[0030] S1: Collect text information of all messages on the social network within a certain time period, reply with text information and topology information, and perform text preprocessing operations.

[0031] Specifically, the text preprocessing steps are as follows: Traverse each line of the dataset, remove symbols that do not belong to the text, and format the text; for each line, store the ID, source message content, and corresponding tag information of each line into three storage lists: message ID, source message content, and tag; use the message ID as an index to index all reply messages corresponding to the source message ID, store all reply message texts in a list, and store this message list in the reply information list of the source message; concatenate all text data lists together to form a sentence list; create an empty dictionary; sort all appearing characters by frequency, using the character itself as the key and the sorted sequence number as the value, and store this key-value pair in the dictionary; traverse all text sentences and serialize all text symbols; limit the maximum length of the message, the maximum number of replies per message, and the maximum length of a single reply; traverse all text sentences one by one, truncate the sequence portion exceeding the length limit, and padded the sequence portion below the length limit with all zeros; obtain the sequence number of the social message and its reply message.

[0032] Furthermore, in order to conveniently utilize the four data items of the same message—message number, text content, all reply message text content, and message tag—each time a message is accessed, the above four data items corresponding to that message are stored in the corresponding list. In this way, in the model, when you want to access the full content of a message, you can conveniently retrieve all the corresponding data items from the four data lists using the same index.

[0033] S2: Input the message text information into the local information encoder, and output the message text information feature vector. Input the message reply information into the local information encoder, and output the message reply fusion information feature vector.

[0034] Furthermore, the model's input is divided into two parts: text input and graph topology input. The text input is further divided into a message part and a response message part. The message part is [Batch × num]. word ×len dim A vector in the form of [Batch × num], where the reply message part is [Batch × num]. replies ×num word ×len dim The vector is in the form of ], and the input for the graph part of the data consists of the node features and adjacent edge information of the topological graph. The node features are [num nodes A vector of the form [×dim], num nodesThe number of nodes is [number], and each node in the graph represents a message. Adjacency information includes the edge relationships. index , is [2×num edges A vector in the form of ], and the weights of adjacent edges.

[0035] S3: Concatenate the text information feature vector of the message with the fusion information feature vector of the reply with different weights to obtain the local information feature vector of the message; input the topological information of the social network topology graph composed of all messages into the global information encoder to obtain the topological information feature vector.

[0036] Furthermore, extracting topological information from the message propagation graph requires processing non-Euclidean data like graphs using graph neural networks. The chosen model is GATv2. The weights between two nodes in a social network graph change as the graph evolves. Conventional graph neural networks, such as graph attention networks, consider static attention, which might lead to a node consistently focusing on a fixed neighbor. This doesn't align with the dynamic nature of attention in social network topologies. GATv2, however, considers dynamic attention, taking into account these graph characteristics. Therefore, it can better represent the features of nodes in the graph, especially when the interactions between nodes are complex.

[0037] S4: Concatenate the text information feature vector and the topological information feature vector of the message together to obtain the fused feature vector.

[0038] Specifically, the message text fusion feature vector of the form [n,m] is concatenated with the message topology information feature vector of the form [n,m] according to the dimension dim = -1, where n is the number of messages and m is the representation dimension of a single feature vector, to obtain the message fusion feature vector of the form [n,2m].

[0039] S5: Input the fused feature vector into the linear network and output the final prediction vector.

[0040] Specifically, the message fusion feature vector of the form [n, 2m] is input into a linear network with 2m input channels and dim output channels, where dim = 4, representing the four labels of the misinformation detection task: true rumor, false rumor, no rumor, and unverified rumor. The preceding network outputs a set of prediction vectors of the form [n, dim], where each one-dimensional prediction vector corresponds to a prediction value of a message for the four labels. The output vector set is then passed through a softmax activation function to transform the vector element values ​​into probability values ​​of a message corresponding to a certain label, and the final output is a set of prediction vectors of the form [n, dim].

[0041] S6: Each value in the prediction vector is a predicted value for each type of label. The label corresponding to the maximum value among the predicted values ​​is taken as the final predicted label.

[0042] Specifically, for each predicted vector, the label with the largest probability value among the four is taken as the final predicted label of the model; for all the predicted label sets and the sample label sets, the cross-entropy loss between them is calculated, and the total loss is calculated; the model parameters are updated using the backpropagation algorithm based on the calculated loss, and finally the optimal network parameters are obtained.

[0043] S7: As Figure 1 As shown, this illustrates the main steps of the training and classification methods for misinformation detection. Both training and classification require the same preprocessing procedure, which transforms multiple consecutive samples into sequential input samples. The classification method classifies one message sample at a time, while the training method processes a batch of message samples each time. The classification method inputs the sequential samples into the target model and obtains the model's sequential output. The output vector is the final prediction vector, and the class corresponding to the highest confidence value in this vector is taken as the classification result. The training method inputs each sequential sample into the target model and obtains the corresponding sequential output. It calculates the cross-entropy loss based on the label corresponding to the prediction vector output at each step and the true label. Finally, it uses the backpropagation algorithm to update the model weights, and repeats the training until the model converges. The weights of the target model used in the classification method are the model weights obtained by the misinformation detection training method.

[0044] S8: As Figure 3 As shown, this is a hybrid network model that integrates the text information of the message itself and the reply with the topological information corresponding to the message in the message propagation graph. The model processes the local information, i.e., the text information, and the global information, i.e. the topological information, respectively, to obtain the representation vector. After dimensionality reduction through a linear network, the prediction vector is finally obtained.

[0045] S9: As Figure 4 As shown

[0046] Specifically, it demonstrates the main steps of the local information encoding part of the above method. The text sequence number is vectorized by a word embedding model to generate a corresponding representation vector for each symbol, thus vectorizing the text sequence. At this point, the representation vector without temporal features is input into a long short-term memory neural network model to obtain a vector representation with temporal information in the sentence. After performing a one-dimensional convolution on the vector with temporal information in the sentence, the results obtained by the same convolution kernel are pooled and finally concatenated to obtain the dimensionality-reduced feature expression vector. The reply message is processed by an attention mechanism to obtain the weight of the reply to the source message. The weight sum is calculated to obtain the fused information feature expression of all replies corresponding to a message. The feature expression of the message itself and the fused information feature expression of the replies are concatenated to obtain the final local information feature expression of the message, that is, the text information feature expression.

[0047] In summary, this invention, while ensuring training efficiency by using fixed-length samples, greatly enhances the model's feature extraction capability by fusing network message content, its reply message content, and the topological information of the message in the propagation graph, thus overcoming the problem of insufficient model feature extraction capability to a certain extent.

[0048] Example 2

[0049] Referring to Table 1, which is an embodiment of the present invention, a comparative description with conventional solutions is provided based on the above method in order to verify its beneficial effects.

[0050] Table 1 Comparison Table

[0051]

[0052]

[0053] The above explanation clearly demonstrates the advantages of our solution over traditional social network misinformation detection methods in terms of reliability, efficiency, and time lag. It effectively achieves intelligent detection of social network misinformation and enhances the credibility of our solution.

[0054] Through the above description of the implementation method, those skilled in the art can clearly understand that traditional social network misinformation detection typically uses traditional machine learning models, such as decision trees. However, to achieve good detection results, traditional methods require selecting sufficiently representative features, sometimes even necessitating manually designed feature extraction methods by human experts. This not only consumes a large amount of manpower and energy, resulting in low efficiency, but also suffers from high time delays. This is a significant shortcoming for misinformation detection tasks with high real-time requirements. Our invention precisely solves these problems. Due to the complexity of information related to online messages, it is difficult for humans to determine which features can more effectively represent the overall characteristics of online messages. Numerous misinformation classification methods based on text features or topological features have proven that misinformation cannot be reliably detected using only a single feature. Based on this characteristic, our invention uses multi-feature information to detect online misinformation with high accuracy and reliability.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1.A deep learning-based intelligent detection method for false rumors in social networks, characterized in that: The method comprises the following steps: collecting text information, reply information and topology information of all messages on a social network within a time period and performing text preprocessing operations; the text preprocessing operation comprises the following steps: traverse each row of the data set, remove symbols that do not belong to text, and perform text formatting; each time a row is traversed, the ID, source message content and message corresponding label information in each row are stored in the message ID, source message content and belonging label three storage lists respectively; using the message ID as an index, indexing to all reply messages corresponding to the source message ID, storing all reply message texts in the form of a list, and storing the message list in the reply information list to which the source message belongs; splicing all text data lists together to form a sentence list; creating an empty dictionary; sorting the frequency of all appearing characters, taking the characters themselves as keys, taking the sorted sequence numbers as values, and storing the key-value pairs in the dictionary; traversing all text sentences to serialize all text characters; limiting the maximum length of a message, the maximum number of replies of a single message and the maximum length of a single reply; traversing all text sentences one by one, truncating the part of the sequence that exceeds the limit length, and expanding the sequence that does not meet the limit length with all 0s; obtaining the sequence number sequence of the social message and its reply message; inputting the text information of the message into the local information encoder to output the text information feature vector of the message; inputting the reply information of the message into the local information encoder to output the reply fusion information feature vector of the message; the step of inputting the text information of the message into the local information encoder to output the text information feature vector of the message comprises the following steps: vectorizing the sequence number sequence of the message and its reply to obtain a text vector without time sequence; inputting the text vector without time sequence information into a neural network model, and the model outputs a sequence with time sequence information, wherein any sequence is only related to the remaining sequences input at the current time step, and is irrelevant to all sequences at other steps; Each current message sample is in the form of a single-precision floating-point two-dimensional vector, where L is the maximum number of characters in a sentence, and D is the dimension of a single word vector, and different size one-dimensional vector results are obtained by convolution of different size convolution kernels on the vector. ​ taking the maximum value in each vector, splicing them together to obtain a new one-dimensional vector as the final one-dimensional vector of the message text expression at the current step; splicing the text information feature vector of the message and the reply fusion information feature vector with different weights to obtain a message local information feature vector; inputting the topology information of the social network topology graph formed by all messages into a global information encoder to obtain a topology information feature vector; splicing the text information feature vector of the message and the topology information feature vector to obtain a fusion feature vector; the step of splicing the text information feature vector of the message and the reply fusion information feature vector with different weights to obtain a message local information feature vector comprises the following steps: setting a parameter alpha, which represents the proportion of the original message and the reply in the fusion feature vector, the proportion and is 1, and the parameter can be trained, the final fusion text vector representation can be represented as: wherein respectively denote the text fusion feature vector, the original message text feature vector and the reply message text feature vector; the step of inputting the topology information of the social network topology graph formed by all messages into a global information encoder to output a topology information feature vector of the message comprises the following steps: performing the same operation on the reply message to obtain a one-dimensional feature vector of the reply message; The vector representation obtained based on the reply message text is converted by a self-attention mechanism to obtain a new reply message feature vector; The converted reply message vector of each message is spliced in front of the corresponding message vector, and finally the weight of each message to the message is obtained through a linear network and an activation layer; The weight sum of all replies of each message is calculated to finally obtain a fusion reply message vector with the same size as the message dimension; The two are spliced together to obtain the final message text fusion vector; The preprocessed graph data is input into the graph attention network for training, and each node in the graph represents a published message in the social network; The model finally outputs the feature vector of each node to represent the information of the node in the topology structure; The fusion feature vector is input into a linear network to output a final prediction vector; Each value in the prediction vector is a prediction value for each label, and the label corresponding to the maximum prediction value is taken as the final predicted label. 2.The deep learning based intelligent detection method for social network false rumors according to claim 1, wherein: The preprocessed graph data includes adjacent edges and random initial node features. 3.The deep learning based intelligent detection method of social network false rumors according to claim 1, wherein: The step of splicing the text information feature vector of the message and the topology information feature vector to obtain a fusion feature vector includes the following steps: The obtained message text fusion feature vector in the form of [n, m] is spliced with the message topology information feature vector in the form of [n, m] according to the dimension dim=-1, wherein n is the number of messages, m is the representation dimension of a single feature vector, and a message fusion feature vector in the form of [n, 2m] is obtained. 4.The deep learning based intelligent detection method of social network false rumors according to claim 1, wherein: The step of inputting the fusion feature vector into a linear network to output a final prediction vector includes the following steps: The message fusion feature vector in the form of [n, 2m] is input into a linear network with an input channel of 2m and an output channel of dim, wherein dim is 4, representing four labels of rumor detection tasks, true flow, false rumor, non-rumor, and unconfirmed rumor; The first network outputs a prediction vector set in the form of [n, dim], wherein each one-dimensional prediction vector corresponds to a prediction value of four labels for a message; The output vector set is converted into a probability value of a message corresponding to a certain label by a softmax activation function, and finally a final prediction vector set in the form of [n, dim] is output. 5.The deep learning based intelligent detection method of social network false rumors according to claim 3, wherein: The step of taking each value in the prediction vector as a prediction value for each label and taking the label corresponding to the maximum prediction value as the final predicted label includes the following steps: For each prediction vector, the label with the maximum value of the four probability values is taken as the final prediction label of the model; For the obtained prediction label set and the sample label set, the cross-entropy loss between them is calculated, and the total loss is counted; The model parameters are updated using the backpropagation algorithm according to the calculated loss to finally obtain the best network parameters.

Images