WO2019196259A1

WO2019196259A1 - Method for identifying false message and device thereof

Info

Publication number: WO2019196259A1
Application number: PCT/CN2018/097540
Authority: WO
Inventors: 王健宗; 黄章成; 吴天博; 肖京
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-04-09
Filing date: 2018-07-27
Publication date: 2019-10-17
Also published as: CN108830630B; CN108830630A

Abstract

The present application applies to the technical field of information processing, and provides a method for identifying a false message and a device thereof. The method comprises: obtaining multiple carrier texts comprising a target message, and propagation paths of the carrier texts; obtaining text matrices of the carrier texts on the basis of the carrier texts and an identifier of a propagation user; importing the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message; generating a user propagation matrix with respect to the target message according to the propagation paths of all the carrier texts; importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message; calculating an authenticity index of the target message according to the user propagation feature vector and the text feature vector; and if the authenticity index falls within a preset authenticity index range, identifying the target message as a false message. According to the present application, there is no need for manual forensic investigation, thereby reducing labor costs and an investigation duration and improving the accuracy of false message identification.

Description

Method for identifying false message and device thereof

The present application claims the priority of the Chinese Patent Application entitled "A Method for Identifying False Messages and Its Equipment", which is filed on Apr. 09, 2018, the entire contents of which are incorporated by reference. In this application.

Technical field

The present application belongs to the field of information processing technologies, and in particular, to a method for identifying a fake message and a device thereof.

Background technique

False news, or "rumor", refers to a message that is fabricated without facts. False news can erroneously influence the public's point of view and lead people to make wrong choices. Especially in the field of financial investment, false news may cause investors to make wrong investment choices, and even cause investors to panic, resulting in investment chaos in the economic market and increasing the risk of loss of economic property of users. Therefore, how to accurately identify whether the target message is a false message is of great significance.

The existing identification technology of the fake message needs to be related to the target message to determine whether the target message is a false message. However, the above method requires a lot of manpower for thread surveying, especially when there are multiple destinations and the same area as the investigator's location, which requires a lot of time and labor costs, and the recognition efficiency is low.

technical problem

In view of this, the embodiment of the present application provides a method for identifying a false message and a device thereof, to solve the existing method for identifying a fake message, which requires a large amount of time and labor cost, and has a low recognition efficiency.

Technical solution

A first aspect of the embodiment of the present application provides a method for identifying a fake message, including:

Acquiring a plurality of carrier texts containing the target message, and a propagation path of each of the carrier texts; the propagation path comprising an identifier of a propagation user propagating the carrier text;

Obtaining a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

Importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

Generating, according to a propagation path of all the carrier texts, a user propagation matrix about the target message; each element included in the user propagation matrix is specifically a number of carrier texts propagated by each of the propagation users;

Importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message;

Calculating an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

If the authenticity index is within a preset false index range, the target message is identified as a false message.

Beneficial effect

The embodiment of the present application obtains a text matrix of each carrier text by acquiring all carrier texts including the target message and the propagation path of each carrier text, and by using the carrier text and the identifier of the propagating user included in the propagation path, and adopting multiple text matrices. Obtaining the text feature vector of the target message; at the same time, obtaining the user propagation matrix through the propagation path of each carrier text, and then calculating the user propagation feature vector of the target message; finally, calculating based on the user propagation feature vector and the text feature vector The authenticity index of the target message identifies whether the target message is a false message by using an authenticity index. Compared with the existing false message recognition technology, the present embodiment does not require manual research and forensics, thereby reducing the labor cost and the time required for the investigation, and can collect the text features of the carrier text of the target message and the dissemination. The user characteristics of each of the propagating users of the target message are analyzed, wherein the text feature vector can indicate whether the target message has a swaying characteristic, and the user eigenvector can indicate whether the target message has burst propagation in the process of propagation. Through the above two feature vectors, the false index of the target message can be obtained, thereby identifying whether the target message is a false message, and improving the recognition accuracy of the fake message.

DRAWINGS

1 is a flowchart of an implementation of a method for identifying a fake message according to a first embodiment of the present application;

2 is a flowchart of a specific implementation method for identifying a false message S103 according to the second embodiment of the present application;

FIG. 3 is a flowchart of a specific implementation method for identifying a false message S105 according to the third embodiment of the present application;

FIG. 4 is a flowchart of a specific implementation method of a method for identifying a fake message according to a fourth embodiment of the present application;

4b is a computational block diagram of a true and false index calculation model provided by an embodiment of the present application;

FIG. 5 is a flowchart of a specific implementation method of a method for identifying a fake message according to a fourth embodiment of the present application;

FIG. 6 is a structural block diagram of a device for identifying a fake message according to an embodiment of the present application;

FIG. 7 is a schematic diagram of an apparatus for identifying a fake message according to another embodiment of the present application.

Embodiments of the invention

In the embodiment of the present application, the execution subject of the process is an identification device of a fake message. The identification device of the fake message includes, but is not limited to, a recognition device for a fake message such as a notebook computer, a computer, a server, a tablet computer, and a smart phone. In particular, the method for identifying the fake message may be a server of a network platform, so that various propagation parameters such as the forwarding amount, the propagation speed, and the propagation path of each piece of the broadcast text on the network platform can be obtained. FIG. 1 is a flowchart of an implementation of a method for identifying a fake message according to a first embodiment of the present application, which is described in detail as follows:

In S101, a plurality of carrier texts including a target message, and a propagation path of each of the carrier texts are acquired; the propagation path includes an identifier of a propagation user that propagates the carrier text.

In this embodiment, the target message may be set by the user, that is, when the user needs to determine the authenticity of a certain message, the content of the target message may be input into the identification device of the fake message provided in this embodiment, or The message carrier, such as an article, link, etc., containing the message is sent to the identification device, which in turn determines the target message from the message carrier. Optionally, the identification device may further set a detection period to periodically detect the authenticity of the message propagated in the network platform where the identification device is located. In this case, the identification device collects the carrier text contained in the network platform by using a preset detection period, and extracts the target message from each carrier text propagated by the network platform based on the preset target message extraction condition, and performs the correlation of S101. operating.

Optionally, the preset target message extraction condition may be: extracting a text keyword from each carrier text based on a semantic recognition algorithm, and counting the number of occurrences of the same text keyword in each carrier text; if a certain text keyword If the number of occurrences is greater than the preset number of times threshold, it is determined that the message corresponding to the text keyword is the target message.

In this embodiment, the propagation of the message depends on various carriers, for example, by texts such as articles, comments, chat records, etc., and the text carrying the target message is the above-mentioned carrier text. After the target device determines the target message, the identification device may query whether the target message is included in each text. If a certain text in the network platform includes the target message, the text is identified as the carrier text. Preferably, since the false message has a certain time limit, that is, the burst period of the false message propagation may be in a short time range ranging from one week or ten days without a long time, for example, even one year. False messages have been spread earlier without being discovered. In order to reduce the number of processing of the carrier text by the identification device, a valid time range is set, that is, the text whose creation time is within the valid time range and contains the target message is recognized as the carrier text, and the text whose creation time is outside the valid time range is It is not recognized, which improves processing efficiency and effectively filters out a large number of invalid texts.

In this embodiment, the identification device obtains a propagation path of the carrier text, where the propagation path is specifically a path in which the carrier text flows between the various propagation users in the network platform, so the propagation path may include a propagation user that propagates the carrier text. Logo. The identifier of the propagating user may be a user name of the propagating user, a user account, or user information of the propagating user. Preferably, in this embodiment, the user information of the user is used. Since the same entity user can register a plurality of different user accounts in the network platform, and multiple user names exist, different user names or user accounts may correspond. The entity is the same, but the user information can avoid the above situation, because the user information, such as the ID number, is unique, so that the entity corresponding to the same user information is the same. Improve the efficiency of false message recognition.

In S102, a text matrix of each of the carrier texts is obtained based on the carrier text and the identifier of the propagating user.

In this embodiment, the same propagating user can propagate a plurality of carrier texts about the target message, and the same carrier text can be propagated by a plurality of different propagating users. Therefore, in order to accurately determine the propagation of the target message, the identifying device Based on the propagation path of the carrier text, the user identifiers of all the propagating users that propagate the carrier text are determined, and a text matrix is constructed for each carrier text based on the user ID of the propagating user.

Preferably, the text matrix may include text content information of the carrier text in addition to the user identification information of the carrier text propagation user. In this case, the identification device performs a keyword extraction operation on the carrier text to determine keywords included in the carrier text. It should be noted that the extracted keyword is a keyword associated with the target message, and after identifying the target message, the identification device determines candidate keywords associated with the target message, and determines which ones are included from the carrier text. The candidate keyword determines a content feature parameter of the carrier text based on the identified candidate keyword, and then constructs a text matrix of the carrier text based on the content feature parameter and the identifier of the propagation user.

In S103, each of the text matrices is imported into a preset feature vector calculation model to obtain a text feature vector of the target message.

In this embodiment, since an important feature of the spurious message is the explosiveness and extensiveness of the propagation speed, the carrier text containing the fake message also has the above two features. The text matrix generated by each carrier text according to the identifier of the corresponding propagating user can represent the relevant feature of the carrier text in the perspective of user propagation, and determine whether there is explosiveness and extensiveness. If it exists, it indicates that the carrier text is very It may carry a false message. Since a carrier text may contain a plurality of different messages, in order to determine whether the burst propagation is caused by the target message, the text matrix of each carrier text needs to be identified. Therefore, after the text matrix of each carrier text is generated, the identification device needs to import each text matrix into a preset feature vector calculation model, and determine a text feature vector of the target message as a reference parameter for identifying the authenticity of the target message. one.

It should be noted that, since S102 and S103 are text feature vectors for calculating a target message, and S104 and S105 are user feature vectors for calculating a target message, that is, there is no order between the above two types of steps, and false The identification device of the message may first execute S102 and S103, and then execute S104 and S105; or execute S104 and S105 first, and then perform S102 and S103. Preferably, if the identification device can concurrently perform dual-thread calculation, the operations of S102 and S104 can be performed simultaneously.

In S104, a user propagation matrix about the target message is generated according to a propagation path of all the carrier texts; each element included in the user propagation matrix is specifically a number of carrier texts propagated by each of the propagation users. .

In this embodiment, as described above, one propagating user can simultaneously transmit a plurality of carrier texts including the target message. Therefore, in order to determine the number of propagating carrier texts of each propagating user, it is necessary to count each according to the propagation path of each carrier text. The number of texts transmitted by the user is propagated, and the user propagation matrix corresponding to the target message is obtained. For false messages, the user of the false message is generally generated by the user, that is, the creator, consciously and continuously disseminating the carrier text of the false message, that is, for the rumor, the number of the carrier files will account for a large proportion of the total amount of the carrier text. The general propagating user of the non-executor has a limited number of propagating carrier texts, which is a scattered propagating behavior. Therefore, through the user propagation matrix, it can better reflect whether the rumor maliciously spreads the target message, thereby judging whether the target message is For false news.

Optionally, the identification device may create a propagation user network map, and draw a propagation path of each carrier text on the propagation user network map according to a propagation path of each carrier text, if the propagation path passes through the network diagram A propagating user adds 1 to the number of propagating texts of the propagating user, so that after all the propagating paths are drawn, the number of carrier texts propagated by each propagating user can be determined, and the user propagating matrix is generated.

Preferably, in the present embodiment, the order of the respective propagating users in the matrix in the user propagation matrix is consistent with the order of propagation on the propagation path. That is, if a propagating user is the author of the carrier file, that is, the first communicator, the order in which the user matrix is propagated is 1, and so on. If multiple users are in the same propagation order, the propagation users of the same propagation order may be sorted again based on the number of propagation carrier texts, and an array of the number of carrier texts propagated by the propagation users of the same propagation order may be used as an array. The user propagates the elements of the order in the matrix.

In S105, the user propagation matrix is imported into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message.

In this embodiment, the propagation rule of the target message between the propagation users can be determined by the user propagation matrix. In order to extract the user propagation characteristics about the propagation user, the identification device imports the user propagation matrix of the target message into the user feature calculation. The model determines a user propagation feature vector for the target message, and whether the feature vector conforms to the user propagation feature of the fake message by the user, and thus can be used as one of the reference parameters for the subsequent calculation of the authenticity index.

In S106, an authenticity index of the target message is calculated according to the user propagation feature vector and the text feature vector.

In this embodiment, after determining the user propagation feature vector of the target message and the text feature vector, the identification device may calculate the authenticity index of the target message. The specific manner of the calculation may be: importing the user propagation feature vector and the text feature vector into a preset authenticity index calculation model, and obtaining the authenticity index of the target message by converting the authenticity index calculation model. Preferably, the authenticity index calculation model can be a neural network. The administrator generates a corresponding user propagation feature vector and a text feature vector through the training message, and imports it into the neural network for calculating the authenticity index, and adjusts each parameter in the neural network to minimize the value of the loss function of the neural network. Then, the adjusted neural network is used as the calculation model of the authenticity index. Specifically, the expression of the loss function of the neural network is specifically:

Where L _j is the actual authenticity index of the training message.

Is the default regular item.

After the feature vector of the training message is propagated and the text feature vector is imported into the authenticity index calculation model, the obtained authenticity index is calculated. N is the total number of training messages.

Optionally, in addition to determining the authenticity index of the target message by using the authenticity index calculation model, the text feature vector and each parameter value included in the user feature vector may be compared with a preset false parameter range, and the statistical parameter is The value of the parameter falls within the range of the false parameter, and the number of the parameter is used as the authenticity index of the target message. The authenticity index can be used to characterize the similarity between the target message and the fake message.

In S107, if the authenticity index is within a preset false index range, the target message is identified as a fake message.

In this embodiment, the identification device has only a false index range. If the authenticity index calculated by a target message is within the false index range, it indicates that the target message conforms to the false message in both the text feature and the propagation user feature. The feature, therefore, identifies the target message as a false message; conversely, if the authenticity index of the target message is outside the range of the false index, it indicates that the target message does not match the feature of the fake message, and the target message is identified as a real message.

It can be seen that the method for identifying a fake message provided by the embodiment of the present application obtains all the carrier texts including the target message, and the propagation path of each carrier text, and the carrier text and the identifier of the propagating user included in the propagation path. Obtaining a text matrix of each carrier text, and obtaining a text feature vector of the target message through a plurality of text matrices; at the same time, obtaining a user propagation matrix through the propagation path of each carrier text, and then calculating a user propagation characteristic of the target message Finally, the authenticity index of the target message is calculated based on the user propagation feature vector and the text feature vector, and the authenticity index is used to identify whether the target message is a false message. Compared with the existing false message recognition technology, the present embodiment does not require manual research and forensics, thereby reducing the labor cost and the time required for the investigation, and can collect the text features of the carrier text of the target message and the dissemination. The user characteristics of each of the propagating users of the target message are analyzed, wherein the text feature vector can indicate whether the target message has a swaying characteristic, and the user eigenvector can indicate whether the target message has burst propagation in the process of propagation. Through the above two feature vectors, the false index of the target message can be obtained, thereby identifying whether the target message is a false message, and improving the recognition accuracy of the fake message.

FIG. 2 is a flowchart showing a specific implementation of the method S103 for identifying a fake message according to the second embodiment of the present application. Referring to FIG. 2, in the method for identifying a fake message provided by the embodiment, S103 includes S1031 to S1034, and the details are as follows:

In S1031, the number of times of propagation of each of the carrier texts, content feature parameters, and propagation time parameters are respectively acquired.

In this embodiment, in order to improve the accuracy of the text feature vector, the recognition device of the fake message acquires the propagation times of the carrier text, the content feature parameters, and the propagation time parameters, in addition to the propagation path of each carrier text. Multiple aspects of the carrier text are discriminated against authenticity attributes.

Specifically, the number of times of propagation includes, in addition to the number of times the user forwards the carrier text, the number of times the user comments the carrier text and the number of times the carrier text is liked, that is, the number of times various behaviors contributing to the propagation of the vector text. The content feature parameter is specifically used to represent the content information that is required to be expressed by the carrier text, and the manner of extracting may be determined by determining the keyword included in the carrier text, and then determining the carrier text by using the extracted keyword, as described in S102. Content feature parameters. The propagation time parameters include, but are not limited to, at least one of the following: carrier text creation time, average propagation interval, total propagation duration, and the like.

In S1032, each of the carrier texts is sorted based on the propagation time parameter, and an order of introduction of each of the carrier texts is determined.

Since the multi-layer feedback loop neural network used in this embodiment determines the text feature vector of the target message, it is necessary to preset to determine the import order of each carrier text to the multi-layer loop neural network, that is, the loop level at which it is located. Wherein, if the level of the multi-layer feedback loop neural network is greater than the number of the carrier text, the number of layers of the multi-layer loop neural network is reduced during the import operation to match the number of the carrier text.

In this embodiment, the identification device determines the import order of the respective carrier texts according to the propagation time parameters, wherein the manner of determining the import order differs based on the types of parameters included in the propagation time parameters. For example, if the propagation time parameter is the creation time of the carrier text, the import order of each carrier text may be determined according to the order of creation time; if the propagation time parameter is the total propagation duration, the size may be based on the length of the total propagation duration. To determine the order in which the individual carrier texts are imported.

In S1033, the number of propagation times, the content feature parameter, the propagation time parameter, and the text matrix are imported into a text time series vector conversion model to obtain a text timing vector of each of the carrier texts; The vector conversion model is specifically:

among them,

To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x _u is the text matrix; x _t is the fusion of the carrier text with the import order t a matrix; x _τ is the content feature parameter; W _a and b _a are preset adjustment coefficients of the text time series vector conversion model.

In this embodiment, the identification device first constructs a text feature matrix of the carrier text according to the number of propagations, the content feature parameter, the propagation time parameter, and the text matrix, that is, the above-mentioned x _t , and the manner of constructing may be based on the text matrix. Three matrix rows are added to store the three sets of feature quantities: the number of propagation times, the content feature parameters and the propagation time parameters. If the text matrix is an n-dimensional matrix, the corresponding text feature matrix is an n+3 dimensional matrix.

Since the multi-layer cyclic neural network is a neural network with timing relationship, it is necessary to perform time-series conversion between the text feature matrices, that is, to determine the text timing vector of the carrier text. The tanh function is used in this embodiment because the function has good nonlinearity and is matched with the timing characteristics. Therefore, the recognition device will import the text feature matrix into the tanh function to determine the text timing vector corresponding to each carrier text.

In S1034, based on the importing order, text timing vectors of each of the carrier texts are imported into each layer of the multi-layer feedback loop neural network to obtain a text feature vector of the target message; the multi-layer feedback loop neural network Specifically:

Where h ₀ is a preset initial text vector;

a text timing vector for each of the carrier texts; h ₁ , h ₂ ... h _t-1 are text feature iteration intermediate values output by each level of the multi-layer feedback loop neural network; h _t is a text feature of the target message Vector; W, U, b are adjustment coefficients.

In this embodiment, the identification device sequentially introduces the text timing vectors of the respective carrier texts into each level in the multi-layer feedback loop neural network based on the introduction order of the carrier text, and the output of each level is used as the next level. Input, the time series characteristics of each carrier text are continuously superimposed, and the calculated text feature vector is a vector based on the influence of each carrier text superposition, and fully integrates the text features of each text.

In this embodiment, the identification device uses the output of the last layer of the cyclic neural network as the text feature vector of the target message. It should be noted that before the multi-layer loop neural network is extracted, the identification device adjusts the level of each multi-layer cyclic neural network according to the number of carrier texts of the target message, so that the level matches the number of the carrier text.

In the embodiment of the present application, the text time vector of each carrier text is determined by collecting multiple parameter values of the carrier text, and the text feature vector of the target message is calculated based on the multi-layer cyclic neural network, thereby improving the text feature vector for the text characteristic. The richness of the ones, thus improving the accuracy of false message recognition.

FIG. 3 is a flowchart showing a specific implementation of a method for identifying a fake message S105 according to the third embodiment of the present application. As shown in FIG. 3, with respect to the embodiment shown in FIG. 1, the method S105 for identifying a fake message provided by this embodiment further includes S1051 to S1055, which are specifically described as follows:

In S1051, the user propagation matrix is subjected to singular value decomposition to obtain user propagation coefficients of each of the propagation users.

In this embodiment, since the user propagation matrix is a global matrix for all propagating users, if it is necessary to determine the user propagation coefficient of each propagating user, it is necessary to perform singular value decomposition on the user propagation matrix, thereby being able to determine different propagations. The contribution of the user to the dissemination of the target message. Specifically, if the user propagation matrix is a matrix of 1*N, the diagonal matrix of the singular decomposition is a regular matrix of 1*1, which can be decomposed into N 1*1 matrices and identified as each of the propagation users. User propagation factor.

In S1052, each of the user propagation coefficients is respectively imported into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

Where s _i is the user feature vector of the i-th propagation user; y _i is the user propagation coefficient of the i-th propagation user;

The user timing vector for the i-th propagation user; W _u , b _u ,

And b _s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm.

In this embodiment, the identification device first performs time domain transformation on the calculated user propagation coefficients of the respective propagation users, thereby obtaining a user timing vector of each propagation user, that is,

As described above, due to the nonlinearity of the tanh function, it has a good matching degree with the timing characteristics. Therefore, when the user propagation coefficient is time-domain converted in S1052, the tanh function is also used, in order to meet the needs of the user feature vector, wherein the predetermined coefficient for adjusting, i.e. W _u and b _u.

In this embodiment, after the identification device determines the user timing vector of each propagation user, the identification device passes the signal function, that is,

Determining a user feature vector corresponding to each user timing vector, wherein

And b _s is a preset parameter value.

In S1053, a user feature matrix is generated based on user feature vectors of each of the propagation users.

In this embodiment, after the identification device determines the user feature vector of each propagation user, the user feature vector of each propagation user may be determined, for example, by using the user feature vector to identify whether the user is an rumor user or a general communication user, Through the user feature matrix formed by each user feature vector, the user property of all users of the propagation target message can be intuitively determined, thereby improving the efficiency of identifying whether the target message is a fake message.

Specifically, if the user characteristic vector of the multiple propagation users of the propagation target message matches the characteristic vector of the rumor user, it may be determined that the target message is mainly propagated by the creator, indicating that the target message is a false message. high.

In S1054, a mask vector of each of the carrier texts is obtained according to a text matrix, and the mask vector and the user feature matrix are imported into a user propagation feature value calculation model to determine user propagation characteristics of each of the carrier texts. Value; the user propagation feature value calculation model is specifically:

Where [s _i ] is the user feature matrix; m _j is the mask vector of the carrier text described in the jth article; p _j is the user propagation feature value of the carrier text described in the jth article; d([s _i ] *m _j ) is a non-empty element statistical function.

In this embodiment, since the text matrix is generated based on the identifier of the propagating user, if the i-th element in the text matrix is non-empty, it indicates that the i-th user has propagated the carrier text. Therefore, in order to determine the user-propagation feature values of the individual carrier texts, it is first necessary to determine which users have propagated the carrier file, ie to generate the mask vector described above. For example, if the text matrix of a certain carrier text is [5, 0, 0, 5, 0, 7, 5, 6], it means that five propagating users have propagated the carrier text, so the corresponding The mask vector is: [1,0,0,1,0,1,1,1], so that the user characteristics of each propagating user associated with the carrier text can be extracted from the user feature matrix by the mask vector. Vector, that is, get [s _i ]*m _j .

After the present embodiment, the identification device identified by a propagated contribution to support text propagating user, calculates the respective users spread of the mean vector, and thus statistics [s _i] by _{d ([s i] * m} j) functions * The number of non-empty elements in m _j , so that the calculated user propagation feature value is the mean of each user characteristic vector.

In S1055, a user propagation feature vector of the target message is generated according to each of the user propagation feature values.

In this embodiment, after determining the user propagation feature values of all the carrier texts, the identification device aggregates all the user propagation feature values to form a user propagation feature vector corresponding to the target message.

In the embodiment of the present application, the user feature vector of each propagation user is calculated, and the average user feature vector of each carrier text, that is, the user propagation feature value, is determined based on the user feature vector, so that the user propagation feature vector not only has the user The feature also includes the propagation characteristics of the carrier text, thereby improving the accuracy of false message recognition.

FIG. 4a is a flowchart showing a specific implementation of a method for identifying a fake message S106 according to the fourth embodiment of the present application. Referring to FIG. 4a, the method for identifying a fake message according to the embodiment of the present invention provides the method for calculating the false message according to the user propagation feature vector and the text feature vector. The authenticity index of the target tag, including S1061~S1062, is as follows:

Further, the calculating the authenticity index of the target tag according to the user propagation feature vector and the text feature vector, including:

In S1061, the user propagation feature vector and the text feature vector are aggregated to obtain an authenticity recognition matrix of the target message.

In this embodiment, after the user propagation vector and the text feature vector are determined, the identification device performs an aggregation operation on the two vectors to form an authenticity recognition matrix including the above two types of features. Specifically, if the user propagation vector is a matrix of n ₁ *m ₁ and the text feature vector is a matrix of n ₂ *m ₂ , the authenticity recognition matrix obtained by the aggregation is (n ₁ +n ₂ )*max( m ₁ , m ₂ ), wherein if the authenticity recognition matrix after the aggregation has a blank element, the preset character can be used for filling, and preferably, the preset character is 0.

In S1062, the authenticity recognition matrix is introduced into the authenticity index calculation model to obtain an authenticity index of the target message; the authenticity index calculation model is specifically:

among them,

The authenticity index; [c _j ] is the authenticity recognition matrix;

And b _c is a preset coefficient of the authenticity index calculation model; e is a natural logarithm.

In this embodiment, after the identification device of the fake message determines the real recognition matrix, the matrix is imported into the fingerprint index calculation model, and the authenticity index calculation model is specifically a signal function, that is, a function.

among them,

And b _c is a preset coefficient of the authenticity index calculation model, which can be determined through training and learning, and can also be manually adjusted according to the needs of the administrator.

For example, FIG. 4b shows a calculation block diagram of a true and false index calculation model provided by an embodiment of the present application. Wherein, [v _j ] text feature vector, [c _j ] is the authenticity recognition matrix, and [p _j ] is a user propagation feature vector.

In the embodiment of the present application, the text feature vector and the user propagation feature vector are aggregated to obtain an authenticity recognition matrix, so that two parameters can be integrated into one parameter, the number of calculations is reduced, and the calculation of the authenticity index is improved. effectiveness.

FIG. 5 is a flowchart showing a specific implementation of a method for identifying a fake message S102 according to the fifth embodiment of the present application. As shown in FIG. 5, with respect to the embodiment shown in FIG. 1, the method S102 for identifying a fake message provided by this embodiment further includes: S1021 and S1022, which are specifically described as follows:

Further, after the clustering interface is invoked based on each of the clustering interfaces, the method further includes:

In S1021, a global propagation matrix [a _ij ] _{n×m of} the target message is constructed based on the carrier text and the identifier of the propagation user; wherein the a _ij is the i-th propagation user for the jth a propagation tag value of the carrier text; the n is the number of the propagating users; the m is the number of the carrier texts;

In this embodiment, after acquiring the propagation path of each carrier text, the identification device may determine each of the propagation users that propagate the carrier text, and generate a sequence based on the user numbers of the respective propagation users. For each carrier text, the statistical operation of the user is propagated in the above manner, so that it is critical to the global propagation matrix about the target message. Wherein, the set of elements of the i-th row in the global propagation matrix indicates which carrier texts the i-th propagation user propagated; and the set of elements of the j-th column in the global propagation matrix indicates which propagation users of the j-th carrier text are propagated, Therefore, the global propagation matrix can be used to determine the propagation of the target message in the network platform, and the propagation information of each carrier text can be obtained by dividing the columns, and the propagation information of each propagation user can be obtained by dividing the rows.

In this embodiment, a _ij is the propagation flag value of the i-th propagation user for the j-th carrier text, specifically, if the i-th propagation user propagates the j-th carrier text, the propagation flag value is 1; On the other hand, if the i-th propagation user does not propagate the j-th carrier text, the propagation flag value is 0, thereby constituting a global propagation matrix [a _ij ] _n×m composed of 1 and 0, through which the global propagation The matrix can look up the propagation contribution of any propagating user to individual carrier files.

In S1022, a submatrix composed of each of the global propagation matrices [a _ij ] _{n × m} is used as a text matrix of each of the carrier texts.

In this embodiment, in the global propagation matrix [a _ij ] _n×m , the set of elements of the i-th column is the propagation of the i-th carrier text by which users, so the global propagation matrix [a _ij ] can be _{n×m is} divided into m sub-matrices, and each sub-matrix is a text matrix corresponding to the carrier text.

In the embodiment of the present application, by constructing a global propagation matrix, it is convenient to determine the propagation of each carrier text and each propagation user, and the text matrix of each carrier text can be divided based on the global propagation matrix, thereby improving the generation efficiency of the text matrix. .

It should be understood that the size of the sequence of the steps in the above embodiments does not mean that the order of execution is performed. The order of execution of each process should be determined by its function and internal logic, and should not be construed as limiting the implementation process of the embodiments of the present application.

FIG. 6 is a structural block diagram of a device for identifying a fake message according to an embodiment of the present application. The device included in the device for identifying a fake message is used to execute each step in the embodiment corresponding to FIG. 1. For details, please refer to the related description in the embodiment corresponding to FIG. 1 and FIG. 1. For the convenience of explanation, only the parts related to the present embodiment are shown.

Referring to FIG. 6, the device for identifying a fake message includes:

a target message parameter obtaining unit 61, configured to acquire a plurality of carrier texts including a target message, and a propagation path of each of the carrier texts; the propagation path includes an identifier of a propagation user that propagates the carrier text;

a text matrix generating unit 62, configured to obtain a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

a text feature vector generating unit 63, configured to import each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

a user propagation matrix generating unit 64, configured to generate a user propagation matrix about the target message according to a propagation path of all the carrier texts; each element included in the user propagation matrix is specifically propagated for each of the propagation users The number of vector texts;

The user propagation feature vector calculation unit 65 is configured to import the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message.

The authenticity index calculation unit 66 is configured to calculate an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

The false message identifying unit 67 is configured to identify the target message as a fake message if the authenticity index is within a preset false index range.

Optionally, the text feature vector generating unit 63 includes:

a text parameter obtaining unit, configured to respectively acquire a propagation number, a content feature parameter, and a propagation time parameter of each of the carrier texts;

And an import order determining unit, configured to sort each of the carrier texts based on the propagation time parameter, and determine an import order of each of the carrier texts;

a text timing vector calculation unit, configured to import the number of propagation times, the content feature parameter, the propagation time parameter, and the text matrix into a text time series vector conversion model to obtain a text timing vector of each of the carrier texts; The text time series vector conversion model is specifically:

among them,

To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x _u is the text matrix; x _t is the fusion of the carrier text with the import order t a matrix; x _τ is the content feature parameter; W _a and b _a are preset adjustment coefficients of the text time series vector conversion model;

a text feature vector calculation unit, configured to import text timing vectors of each of the carrier texts into each layer of the multi-layer feedback loop neural network based on the import order, to obtain a text feature vector of the target message; The feedback cyclic neural network is specifically:

Where h ₀ is a preset initial text vector;

Optionally, the user propagation feature vector calculation unit 65 includes:

a propagation coefficient determining unit, configured to perform singular value decomposition on the user propagation matrix to obtain a user propagation coefficient of each of the propagation users;

a user feature vector calculation unit, configured to import each of the user propagation coefficients into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

The user timing vector for the i-th propagation user; W _u , b _u ,

And b _s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm;

a user feature matrix generating unit, configured to generate a user feature matrix based on user feature vectors of each of the propagation users;

a user propagation feature value calculation unit, configured to obtain a mask vector of each of the carrier texts according to a text matrix, and import the mask vector and the user feature matrix into a user propagation feature value calculation model to determine each of the carriers The user of the text propagates the feature value; the user propagation feature value calculation model is specifically:

Where [s _i ] is the user feature matrix; m _j is the mask vector of the carrier text described in the jth article; p _j is the user propagation feature value of the carrier text described in the jth article; d([s _i ] *m _j ) is a non-empty element statistical function;

And a user propagation feature vector determining unit, configured to generate a user propagation feature vector of the target message according to each of the user propagation feature values.

Optionally, the authenticity index calculation unit 66 includes:

An authenticity recognition matrix generating unit, configured to aggregate the user propagation feature vector and the text feature vector to obtain an authenticity recognition matrix of the target message;

The authenticity index calculation unit is configured to import the authenticity recognition matrix into the authenticity index calculation model to obtain an authenticity index of the target message; the authenticity index calculation model is specifically:

among them,

The authenticity index; [c _j ] is the authenticity recognition matrix;

Optionally, the text matrix generating unit 62 includes:

a global propagation matrix creating unit, configured to construct a global propagation matrix [a _ij ] _{n×m of} the target message based on the carrier text and the identifier of the propagation user; wherein the a _ij is an ith propagation user a propagation tag value for the jth carrier text; the n is the number of the propagating users; the m is the number of the carrier texts;

And a text matrix dividing unit configured to use a sub-matrix formed by each of the global propagation matrices [a _ij ] _n×m as a text matrix of each of the carrier texts.

Therefore, the identification device of the fake message provided by the embodiment of the present application can also reduce the labor cost and the time required for the investigation by eliminating the need for manual research and forensics, and can collect the text feature of the carrier text of the target message and the propagation of the target text. The user characteristics of each of the propagating users of the target message are analyzed, wherein the text feature vector can indicate whether the target message has a swaying characteristic, and the user eigenvector can indicate whether the target message has burst propagation during the process of propagation. Through the above two feature vectors, the false index of the target message can be obtained, thereby identifying whether the target message is a false message, and improving the recognition accuracy of the fake message.

FIG. 7 is a schematic diagram of an apparatus for identifying a fake message according to another embodiment of the present application. As shown in FIG. 7, the identification device 7 of the fake message of this embodiment includes a processor 70, a memory 71, and computer readable instructions 72 stored in the memory 71 and operable on the processor 70, for example The identification procedure for false messages. The processor 70 executes the computer readable instructions 72 to implement the steps in the foregoing method for identifying the respective fake messages, such as S101 to S107 shown in FIG. 1. Alternatively, the processor 70, when executing the computer readable instructions 72, implements the functions of the various units in the various apparatus embodiments described above, such as the functions of the modules 61 through 67 shown in FIG.

Illustratively, the computer readable instructions 72 may be partitioned into one or more units, the one or more units being stored in the memory 71 and executed by the processor 70 to complete the application. . The one or more units may be a series of computer readable instruction instruction segments capable of performing a particular function for describing the execution of the computer readable instructions 72 in the identification device 7 of the fake message. For example, the computer readable instructions 72 may be segmented into a target message parameter acquisition unit, a text matrix generation unit, a text feature vector generation unit, a user propagation matrix generation unit, a user propagation feature vector calculation unit, an authenticity index calculation unit, and a false The message identification unit has specific functions as described above.

The above-mentioned embodiments are only used to explain the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still implement the foregoing embodiments. The technical solutions described in the examples are modified or equivalently replaced with some of the technical features; and the modifications or substitutions do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in Within the scope of protection of this application.

Claims

A method for identifying a false message, comprising:

Acquiring a plurality of carrier texts containing the target message, and a propagation path of each of the carrier texts; the propagation path comprising an identifier of a propagation user propagating the carrier text;

Obtaining a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

Importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

Generating, according to a propagation path of all the carrier texts, a user propagation matrix about the target message; each element included in the user propagation matrix is specifically a number of carrier texts propagated by each of the propagation users;

Importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message;

Calculating an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

If the authenticity index is within a preset false index range, the target message is identified as a false message.
The identification method according to claim 1, wherein the importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message comprises:

Obtaining, respectively, the number of times of propagation of each of the carrier texts, content feature parameters, and propagation time parameters;

And sorting each of the carrier texts according to the propagation time parameter, and determining an import order of each of the carrier texts;

Importing the number of times of propagation, the content feature parameter, the propagation time parameter, and the text matrix into a text time series vector conversion model to obtain a text time series vector of each of the carrier texts; the text time series vector conversion model is specific for:

among them,
To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x u is the text matrix; x t is the fusion of the carrier text with the import order t a matrix; x τ is the content feature parameter; W a and b a are preset adjustment coefficients of the text time series vector conversion model;

And displaying, according to the importing order, a text time series vector of each of the carrier texts to each level of the multi-layer feedback loop neural network to obtain a text feature vector of the target message; the multi-layer feedback loop neural network is specifically:

Where h 0 is a preset initial text vector;
a text timing vector for each of the carrier texts; h 1 , h 2 ... h t-1 are text feature iteration intermediate values output by each level of the multi-layer feedback loop neural network; h t is a text feature of the target message Vector; W, U, b are adjustment coefficients.
The identification method according to claim 1, wherein the importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message comprises:

Performing singular value decomposition on the user propagation matrix to obtain user propagation coefficients of each of the propagation users;

Introducing each of the user propagation coefficients into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

Where s i is the user feature vector of the i-th propagation user; y i is the user propagation coefficient of the i-th propagation user;
The user timing vector for the i-th propagation user; W u , b u ,
And b s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm;

Generating a user feature matrix based on user feature vectors of each of the propagation users;

Obtaining, according to the text matrix, a mask vector of each of the carrier texts, and importing the mask vector and the user feature matrix into a user propagation feature value calculation model, and determining user propagation feature values of each of the carrier texts; The user propagation feature value calculation model is specifically as follows:

Where [s i ] is the user feature matrix; m j is the mask vector of the carrier text described in the jth article; p j is the user propagation feature value of the carrier text described in the jth article; d([s i ] *m j ) is a non-empty element statistical function;

Generating a user propagation feature vector of the target message according to each of the user propagation feature values.
The identification method according to any one of claims 1-3, wherein the calculating the authenticity index of the target message according to the user propagation feature vector and the text feature vector comprises:

And synthesizing the user propagation feature vector and the text feature vector to obtain an authenticity recognition matrix of the target message;

The authenticity recognition matrix is imported into the authenticity index calculation model to obtain the authenticity index of the target message; the authenticity index calculation model is specifically:

among them,
The authenticity index; [c j ] is the authenticity recognition matrix;
And b c is a preset coefficient of the authenticity index calculation model; e is a natural logarithm.
The identification method according to claim 1, wherein the text matrix of each of the carrier texts is obtained based on the carrier text and the identifier of the propagating user, including:

Constructing a global propagation matrix [a ij ] n×m of the target message based on the carrier text and the identifier of the propagation user; wherein the a ij is the propagation of the jth carrier text by the ith propagation user a tag value; the n is the number of the propagating users; the m is the number of the carrier texts;

A submatrix composed of each of the global propagation matrices [a ij ] n × m is used as a text matrix of each of the carrier texts.
A device for identifying a false message, comprising:

a target message parameter obtaining unit, configured to acquire a plurality of carrier texts including the target message, and a propagation path of each of the carrier texts; the propagation path includes an identifier of a propagation user that propagates the carrier text;

a text matrix generating unit, configured to obtain a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

a text feature vector generating unit, configured to import each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

a user propagation matrix generating unit, configured to generate a user propagation matrix about the target message according to a propagation path of all the carrier texts; each element included in the user propagation matrix is specifically a carrier for each of the propagation users The number of texts;

a user propagation feature vector calculation unit, configured to import the user propagation matrix into a preset user feature calculation model, to obtain a user propagation feature vector corresponding to the target message;

An authenticity index calculation unit, configured to calculate an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

The false message identifying unit is configured to identify the target message as a fake message if the authenticity index is within a preset false index range.
The device for identifying a fake message according to claim 6, wherein the text feature vector generating unit comprises:

a text parameter obtaining unit, configured to respectively acquire a propagation number, a content feature parameter, and a propagation time parameter of each of the carrier texts;

And an import order determining unit, configured to sort each of the carrier texts based on the propagation time parameter, and determine an import order of each of the carrier texts;

a text timing vector calculation unit, configured to import the number of propagation times, the content feature parameter, the propagation time parameter, and the text matrix into a text time series vector conversion model to obtain a text timing vector of each of the carrier texts; The text time series vector conversion model is specifically:

among them,
To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x u is the text matrix; x t is the fusion of the carrier text with the import order t a matrix; x τ is the content feature parameter; W a and b a are preset adjustment coefficients of the text time series vector conversion model;

a text feature vector calculation unit, configured to import text timing vectors of each of the carrier texts into each layer of the multi-layer feedback loop neural network based on the import order, to obtain a text feature vector of the target message; The feedback cyclic neural network is specifically:

Where h 0 is a preset initial text vector;
a text timing vector for each of the carrier texts; h 1 , h 2 ... h t-1 are text feature iteration intermediate values output by each level of the multi-layer feedback loop neural network; h t is a text feature of the target message Vector; W, U, b are adjustment coefficients.
The device for identifying a fake message according to claim 6, wherein the user propagation feature vector calculation unit comprises:

a propagation coefficient determining unit, configured to perform singular value decomposition on the user propagation matrix to obtain a user propagation coefficient of each of the propagation users;

a user feature vector calculation unit, configured to import each of the user propagation coefficients into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

Where s i is the user feature vector of the i-th propagation user; y i is the user propagation coefficient of the i-th propagation user;
The user timing vector for the i-th propagation user; W u , b u ,
And b s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm;

a user feature matrix generating unit, configured to generate a user feature matrix based on user feature vectors of each of the propagation users;

a user propagation feature value calculation unit, configured to obtain a mask vector of each of the carrier texts according to a text matrix, and import the mask vector and the user feature matrix into a user propagation feature value calculation model to determine each of the carriers The user of the text propagates the feature value; the user propagation feature value calculation model is specifically:

Where [s i ] is the user feature matrix; m j is the mask vector of the carrier text described in the jth article; p j is the user propagation feature value of the carrier text described in the jth article; d([s i ] *m j ) is a non-empty element statistical function;

And a user propagation feature vector determining unit, configured to generate a user propagation feature vector of the target message according to each of the user propagation feature values.
The device for identifying a false message according to any one of claims 6-8, wherein the authenticity index calculation unit comprises:

An authenticity recognition matrix generating unit, configured to aggregate the user propagation feature vector and the text feature vector to obtain an authenticity recognition matrix of the target message;

The authenticity index calculation unit is configured to import the authenticity recognition matrix into the authenticity index calculation model to obtain an authenticity index of the target message; the authenticity index calculation model is specifically:

among them,
The authenticity index; [c j ] is the authenticity recognition matrix;
And b c is a preset coefficient of the authenticity index calculation model; e is a natural logarithm.
The device for identifying a fake message according to claim 6, wherein the text matrix generating unit comprises:

a global propagation matrix creating unit, configured to construct a global propagation matrix [a ij ] n×m of the target message based on the carrier text and the identifier of the propagation user; wherein the a ij is an ith propagation user a propagation tag value for the jth carrier text; the n is the number of the propagating users; the m is the number of the carrier texts;

And a text matrix dividing unit configured to use a sub-matrix formed by each of the global propagation matrices [a ij ] n×m as a text matrix of each of the carrier texts.
An identification device for a fake message, characterized in that the identification device of the fake message comprises a memory, a processor, and computer readable instructions stored in the memory and operable on the processor, the processor The following steps are implemented when the computer readable instructions are executed:

Acquiring a plurality of carrier texts containing the target message, and a propagation path of each of the carrier texts; the propagation path comprising an identifier of a propagation user propagating the carrier text;

Obtaining a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

Importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

Generating, according to a propagation path of all the carrier texts, a user propagation matrix about the target message; each element included in the user propagation matrix is specifically a number of carrier texts propagated by each of the propagation users;

Importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message;

Calculating an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

If the authenticity index is within a preset false index range, the target message is identified as a false message.
The device for identifying a fake message according to claim 11, wherein the importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message comprises:

Obtaining, respectively, the number of times of propagation of each of the carrier texts, content feature parameters, and propagation time parameters;

And sorting each of the carrier texts according to the propagation time parameter, and determining an import order of each of the carrier texts;

Importing the number of times of propagation, the content feature parameter, the propagation time parameter, and the text matrix into a text time series vector conversion model to obtain a text time series vector of each of the carrier texts; the text time series vector conversion model is specific for:

among them,
To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x u is the text matrix; x t is the fusion of the carrier text with the import order t a matrix; x τ is the content feature parameter; W a and b a are preset adjustment coefficients of the text time series vector conversion model;

And displaying, according to the importing order, a text time series vector of each of the carrier texts to each level of the multi-layer feedback loop neural network to obtain a text feature vector of the target message; the multi-layer feedback loop neural network is specifically:

Where h 0 is a preset initial text vector;
a text timing vector for each of the carrier texts; h 1 , h 2 ... h t-1 are text feature iteration intermediate values output by each level of the multi-layer feedback loop neural network; h t is a text feature of the target message Vector; W, U, b are adjustment coefficients.
The device for identifying a false message according to claim 11, wherein the importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message comprises:

Performing singular value decomposition on the user propagation matrix to obtain user propagation coefficients of each of the propagation users;

Introducing each of the user propagation coefficients into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

Where s i is the user feature vector of the i-th propagation user; y i is the user propagation coefficient of the i-th propagation user;
The user timing vector for the i-th propagation user; W u , b u ,
And b s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm;

Generating a user feature matrix based on user feature vectors of each of the propagation users;

Obtaining, according to the text matrix, a mask vector of each of the carrier texts, and importing the mask vector and the user feature matrix into a user propagation feature value calculation model, and determining user propagation feature values of each of the carrier texts; The user propagation feature value calculation model is specifically as follows:

Where [s i ] is the user feature matrix; m j is the mask vector of the carrier text described in the jth article; p j is the user propagation feature value of the carrier text described in the jth article; d([s i ] *m j ) is a non-empty element statistical function;

Generating a user propagation feature vector of the target message according to each of the user propagation feature values.
The device for identifying a false message according to any one of claims 11 to 13, wherein the calculating the authenticity index of the target message according to the user propagation feature vector and the text feature vector comprises:

And synthesizing the user propagation feature vector and the text feature vector to obtain an authenticity recognition matrix of the target message;

The authenticity recognition matrix is imported into the authenticity index calculation model to obtain the authenticity index of the target message; the authenticity index calculation model is specifically:

among them,
The authenticity index; [c j ] is the authenticity recognition matrix;
And b c is a preset coefficient of the authenticity index calculation model; e is a natural logarithm.
The device for identifying a false message according to claim 11, wherein the text matrix of each of the carrier texts is obtained based on the carrier text and the identifier of the user, including:

Constructing a global propagation matrix [a ij ] n×m of the target message based on the carrier text and the identifier of the propagation user; wherein the a ij is the propagation of the jth carrier text by the ith propagation user a tag value; the n is the number of the propagating users; the m is the number of the carrier texts;

A submatrix composed of each of the global propagation matrices [a ij ] n × m is used as a text matrix of each of the carrier texts.
A computer readable storage medium storing computer readable instructions, wherein the computer readable instructions, when executed by a processor, implement the following steps:

Acquiring a plurality of carrier texts containing the target message, and a propagation path of each of the carrier texts; the propagation path comprising an identifier of a propagation user propagating the carrier text;

Obtaining a text matrix of each of the carrier texts based on the carrier text and the identifier of the propagating user;

Importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message;

Generating, according to a propagation path of all the carrier texts, a user propagation matrix about the target message; each element included in the user propagation matrix is specifically a number of carrier texts propagated by each of the propagation users;

Importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message;

Calculating an authenticity index of the target message according to the user propagation feature vector and the text feature vector;

If the authenticity index is within a preset false index range, the target message is identified as a false message.
The computer readable storage medium according to claim 16, wherein the importing each of the text matrices into a preset feature vector calculation model to obtain a text feature vector of the target message comprises:

Obtaining, respectively, the number of times of propagation of each of the carrier texts, content feature parameters, and propagation time parameters;

And sorting each of the carrier texts according to the propagation time parameter, and determining an import order of each of the carrier texts;

Importing the number of times of propagation, the content feature parameter, the propagation time parameter, and the text matrix into a text time series vector conversion model to obtain a text time series vector of each of the carrier texts; the text time series vector conversion model is specific for:

among them,
To introduce the text time series vector of the carrier text of order t; η is the number of propagations; ΔT is the propagation time parameter; x u is the text matrix; x t is the fusion of the carrier text with the import order t a matrix; x τ is the content feature parameter; W a and b a are preset adjustment coefficients of the text time series vector conversion model;

And displaying, according to the importing order, a text time series vector of each of the carrier texts to each level of the multi-layer feedback loop neural network to obtain a text feature vector of the target message; the multi-layer feedback loop neural network is specifically:

Where h 0 is a preset initial text vector;
a text timing vector for each of the carrier texts; h 1 , h 2 ... h t-1 are text feature iteration intermediate values output by each level of the multi-layer feedback loop neural network; h t is a text feature of the target message Vector; W, U, b are adjustment coefficients.
The computer readable storage medium according to claim 16, wherein the importing the user propagation matrix into a preset user feature calculation model to obtain a user propagation feature vector corresponding to the target message comprises:

Performing singular value decomposition on the user propagation matrix to obtain user propagation coefficients of each of the propagation users;

Introducing each of the user propagation coefficients into a propagation feature vector transformation model to determine a user feature vector of each of the propagation users; the user feature vector transformation model is specifically:

Where s i is the user feature vector of the i-th propagation user; y i is the user propagation coefficient of the i-th propagation user;
The user timing vector for the i-th propagation user; W u , b u ,
And b s is a preset coefficient of the user feature vector conversion model; e is a natural logarithm;

Generating a user feature matrix based on user feature vectors of each of the propagation users;

Obtaining, according to the text matrix, a mask vector of each of the carrier texts, and importing the mask vector and the user feature matrix into a user propagation feature value calculation model, and determining user propagation feature values of each of the carrier texts; The user propagation feature value calculation model is specifically as follows:

Where [s i ] is the user feature matrix; m j is the mask vector of the carrier text described in the jth article; p j is the user propagation feature value of the carrier text described in the jth article; d([s i ] *m j ) is a non-empty element statistical function;

Generating a user propagation feature vector of the target message according to each of the user propagation feature values.
The computer readable storage medium according to any one of claims 16 to 18, wherein the calculating the authenticity index of the target message according to the user propagation feature vector and the text feature vector comprises:

And synthesizing the user propagation feature vector and the text feature vector to obtain an authenticity recognition matrix of the target message;

The authenticity recognition matrix is imported into the authenticity index calculation model to obtain the authenticity index of the target message; the authenticity index calculation model is specifically:

among them,
The authenticity index; [c j ] is the authenticity recognition matrix;
And b c is a preset coefficient of the authenticity index calculation model; e is a natural logarithm.
The computer readable storage medium according to claim 16, wherein the text matrix of each of the carrier texts is obtained based on the carrier text and the identifier of the propagating user, including:

Constructing a global propagation matrix [a ij ] n×m of the target message based on the carrier text and the identifier of the propagation user; wherein the a ij is the propagation of the jth carrier text by the ith propagation user a tag value; the n is the number of the propagating users; the m is the number of the carrier texts;

A submatrix composed of each of the global propagation matrices [a ij ] n × m is used as a text matrix of each of the carrier texts.