Abnormal account identification method and device of comment account, storage medium and terminal

Abstract

The invention discloses an abnormal account identification method and device of a comment account, a storage medium and a terminal, relates to the technical field of network data analysis, and mainly aims to solve the problem of low identification efficiency of an abnormal account of an article comment account. The method mainly comprises the following steps: in response to a comment account exception identification instruction of a target theme, obtaining comment data of an article which is published in a preset time period and relates to the target theme; according to the comment data, screening out a comment account to be recognized from the global comment accounts; for the comment account to be recognized, basic dimension features and supplementary dimension features are extracted according to the account dimension information and the comment dimension information, feature relevancy division is carried out according to the basic dimension features and the supplementary dimension features, and multi-level recognition features of the comment account to be recognized are obtained; and utilizing the trained comment account exception identification model to identify the multi-level identification features to obtain an account identification result of the comment account to be identified. The method is mainly used for identifying abnormal comment accounts.

Description

Technical Field

[0001] The present invention relates to the technical field of network data analysis, and in particular to a method and device for identifying abnormal comment accounts, a storage medium, and a terminal. Background Art

[0002] Nowadays, social media platforms have become an important channel for the public to obtain information, express opinions, interact socially, and for enterprises to conduct commercial activities. However, with the rapid development of social media platforms, there are more and more abnormal accounts that forge online identities, publish false comments and opinions, and fabricate misleading and exaggerated information. The emergence of abnormal comment accounts has brought serious challenges to the entire network ecology. How to identify abnormal accounts in the comment area has also become an urgent problem to maintain the network ecological environment.

[0003] The existing identification of abnormal accounts in comment areas mainly relies on manual screening of comment accounts under articles. However, due to the large number of comments, a popular article may have tens of thousands or even hundreds of thousands of comments posted by accounts. The manual screening is very labor-intensive and time-consuming, resulting in low efficiency in identifying abnormal comment accounts. Summary of the invention

[0004] In view of this, the present invention provides a method and device for identifying abnormal comment accounts, a storage medium, and a terminal, the main purpose of which is to solve the problem of low efficiency in identifying abnormal comment accounts.

[0005] According to one aspect of the present invention, a method for identifying abnormal comment accounts is provided, comprising:

[0006] In response to an instruction for identifying abnormalities in comment accounts of a target topic, obtaining comment data of articles related to the target topic that were published within a preset time period;

[0007] Based on the comment data, filter out the comment account to be identified from the global comment accounts, and retrieve the account dimension information and comment dimension information of the comment account to be identified;

[0008] For the comment account to be identified, extract basic dimension features and supplementary dimension features according to the account dimension information and the comment dimension information, and perform feature relevance classification according to the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified;

[0009] The multi-level identification features are identified using the trained comment account anomaly identification model to obtain account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

[0010] Furthermore, the feature relevance division is performed based on the basic dimension features and the supplementary dimension features to obtain the multi-level identification features of the comment account to be identified, including:

[0011] Calculate the abnormal account relevance for different features under each dimension and obtain the correlation coefficient of different feature items;

[0012] Marking the feature item whose correlation coefficient is greater than the first correlation threshold as a first-level correlation identification feature;

[0013] Marking the feature item whose correlation coefficient is greater than a second correlation threshold and less than the first correlation coefficient threshold as a secondary correlation identification feature;

[0014] Marking the feature item whose correlation coefficient is smaller than the second correlation threshold as a third-level correlation identification feature;

[0015] The first correlation threshold is greater than the second correlation threshold.

[0016] Furthermore, the multi-level identification features are identified by using the trained comment account abnormality identification model to obtain the account identification result of the comment account to be identified, including:

[0017] In a case where the multi-level identification features include a first-level related identification feature, determining that the comment account to be identified is a suspected abnormal account;

[0018] In the case that the multi-level identification features do not include the primary related identification features, if the number of feature items of the secondary related identification features in the multi-level identification features is greater than the feature item number threshold, then determining that the comment account to be identified is an abnormal tendency account;

[0019] If the number of feature items of the secondary related identification features in the multi-level identification features is less than the feature item number threshold, it is determined that the comment account to be identified is a non-abnormal account.

[0020] Furthermore, the step of screening out the comment accounts to be identified from the global comment accounts based on the comment data includes:

[0021] Counting the number of comments from different comment accounts in the comment data to determine the number of comments published by each comment account;

[0022] If the number of comments published is greater than a preset comment publication threshold, or the proportion of the number of participating comment articles in the global number of articles is greater than a preset comment participation threshold, the comment account is determined as a candidate comment account;

[0023] Emotion recognition is performed on the comment content of the candidate comment account one by one. If the recognition result is that the number of positive emotions or negative emotions is greater than a preset emotion number threshold, the candidate comment account is determined as a comment account to be identified.

[0024] Furthermore, before using the trained comment account anomaly recognition model to recognize the multi-level recognition features, the method further includes:

[0025] Obtain sample comment data for multiple articles on the same topic and identify sample accounts;

[0026] Extracting multi-level identification sample features based on the account dimension data and comment dimension data of the sample account, and constructing a training sample set based on the target identification result label of the sample account and the multi-level identification sample features;

[0027] The extreme gradient boosting model is trained according to the training sample set until the model performance parameters of the extreme gradient boosting model meet the corresponding parameter thresholds, thereby obtaining a comment account anomaly recognition model that has completed training.

[0028] Further, the extreme gradient boosting model is trained according to the training sample set until the model performance parameter of the end gradient boosting model meets the corresponding parameter threshold, thereby obtaining a trained comment account anomaly recognition model, including:

[0029] Dividing the training sample set into a training set, a validation set and a test set according to a preset division ratio;

[0030] The extreme gradient boosting model is trained according to the training set and the initial parameters of the model to optimize the decision tree split point and the feature threshold, thereby obtaining a comment account anomaly recognition model to be verified;

[0031] The model performance test of the comment account anomaly recognition model to be verified is performed respectively according to the verification set and the test set to obtain model performance parameters, and when the model performance parameters meet the target performance parameters, a comment account anomaly recognition model that has completed training is obtained.

[0032] Furthermore, in the case where the comment account abnormality identification instruction carries a negative abnormal account analysis requirement, after obtaining the account identification result of the comment account to be identified, the method further includes:

[0033] For a comment account whose account identification result is a suspected abnormal account or an abnormal account tendency, identifying the sentiment tendency category of the comment account;

[0034] If the account identification result is a suspected abnormal account, and the sentiment tendency category is a negative tendency, the comment account is determined as a suspected negative abnormal account;

[0035] If the account identification result is an abnormal account with a tendency to be negative, and the sentiment tendency category is negative, the comment account is determined to be an abnormal account with a tendency to be negative, and a negative abnormal account analysis result is generated based on the number of suspected negative abnormal accounts and the number of abnormal accounts with a tendency to be negative.

[0036] According to another aspect of the present invention, a device for identifying abnormal comment accounts is provided, comprising:

[0037] An acquisition module is used to respond to the comment account abnormality identification instruction of the target topic and acquire the comment data of the articles related to the target topic published within a preset time period;

[0038] A screening module, configured to screen out a to-be-identified comment account from the global comment accounts based on the comment data, and retrieve account dimension information and comment dimension information of the to-be-identified comment account;

[0039] A feature extraction module, for extracting basic dimension features and supplementary dimension features from the comment account to be identified based on the account dimension information and the comment dimension information, and performing feature relevance classification based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified;

[0040] The identification module is used to use the trained comment account abnormality identification model to identify the multi-level identification features and obtain the account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

[0041] Furthermore, the feature extraction module includes:

[0042] A calculation unit, used to calculate the abnormal account relevance for different features under each dimension, and obtain the correlation coefficients of different feature items; calculate the abnormal account relevance for different features under each dimension, and obtain the correlation coefficients of different feature items;

[0043] A first marking unit, used to mark the feature item whose correlation coefficient is greater than a first correlation threshold as a primary correlation identification feature;

[0044] A second marking unit, used to mark the feature item whose correlation coefficient is greater than a second correlation threshold and less than the first correlation coefficient threshold as a secondary correlation identification feature;

[0045] The third marking unit is used to mark the feature item whose correlation coefficient is less than the second correlation threshold as a third-level correlation identification feature.

[0046] Furthermore, the identification module includes:

[0047] A first determining unit is configured to determine that the comment account to be identified is a suspected abnormal account when the multi-level identification features include a first-level related identification feature;

[0048] A second determination unit is configured to determine that the comment account to be identified is an abnormal tendency account if the number of feature items of the second-level related identification features in the multi-level identification features is greater than a feature item number threshold when the first-level related identification features are not included in the multi-level identification features;

[0049] The third determination unit is used to determine that the comment account to be identified is a non-abnormal account if the number of feature items of the secondary related identification features in the multi-level identification features is less than the feature item number threshold.

[0050] Furthermore, the screening module includes:

[0051] A statistical unit, used to count the number of comments from different comment accounts in the comment data, and determine the number of comments published by each comment account;

[0052] a fourth determination unit, configured to determine the comment account as a candidate comment account if the number of comments published is greater than a preset comment publication threshold, or the proportion of the number of participating comment articles in the global number of articles is greater than a preset comment participation threshold;

[0053] The fifth determination unit is used to perform emotion recognition on the comment content of the candidate comment account one by one, and if the recognition result is that the number of positive emotions or negative emotions is greater than a preset emotion number threshold, the candidate comment account is determined as a comment account to be identified.

[0054] Furthermore, the device also includes:

[0055] The first determination module is also used to obtain sample comment data of multiple articles related to the same topic and determine sample accounts;

[0056] A construction module, used to extract multi-level identification sample features based on the account dimension data and comment dimension data of the sample account, and to construct a training sample set based on the target identification result label of the sample account and the multi-level identification sample features;

[0057] A training module is used to train the extreme gradient boosting model according to the training sample set until the model performance parameters of the extreme gradient boosting model meet the corresponding parameter thresholds, thereby obtaining a comment account anomaly recognition model that has completed training.

[0058] Furthermore, the training module includes:

[0059] A division unit, used to divide the training sample set into a training set, a validation set and a test set according to a preset division ratio;

[0060] An optimization unit, used for training the extreme gradient boosting model according to the training set and the initial parameters of the model, so as to optimize the splitting point and the feature threshold of the decision tree, and obtain an abnormal recognition model of the comment account to be verified;

[0061] The testing unit is used to perform model performance tests on the comment account anomaly recognition model to be verified based on the verification set and the test set, obtain model performance parameters, and obtain a trained comment account anomaly recognition model when the model performance parameters meet the target performance parameters.

[0062] Furthermore, the device also includes:

[0063] The identification module is further used to identify the sentiment tendency category of the comment account for the comment account whose account identification result is suspected to be an abnormal account or tends to be an abnormal account;

[0064] A second determination module is configured to determine the comment account as a suspected negative abnormal account if the account identification result is a suspected abnormal account and the sentiment tendency category is a negative tendency;

[0065] A generation module is used to determine the comment account as a negative abnormal account if the account identification result is an abnormal account with a tendency to be negative and the emotional tendency category is negative, and generate a negative abnormal account analysis result based on the number of suspected negative abnormal accounts and the number of negative abnormal accounts.

[0066] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored in the storage medium, and the executable instruction enables a processor to execute operations corresponding to the abnormal account identification method of the comment account as described above.

[0067] According to another aspect of the present invention, there is provided a terminal, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0068] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the above-mentioned abnormal account identification method of the comment account.

[0069] By means of the above technical solution, the technical solution provided by the embodiment of the present invention has at least the following advantages:

[0070] The present invention provides a method and device, storage medium, and terminal for identifying abnormal accounts of comment accounts. The embodiments of the present invention obtain comment data of articles related to a target topic and published within a preset time period in response to an abnormal identification instruction for comment accounts of a target topic; based on the comment data, a comment account to be identified is screened out from global comment accounts; for the comment account to be identified, basic dimension features and supplementary dimension features are extracted based on account dimension information and comment dimension information, and feature relevance is divided based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified; using a trained abnormal identification model for comment accounts, the multi-level identification features are identified to obtain an account identification result of the comment account to be identified, which greatly reduces the time for identifying abnormal accounts of comment accounts and reduces the dependence of comment account identification on manual work. At the same time, the accuracy of abnormal account identification is ensured, thereby greatly improving the efficiency of abnormal account identification of comment accounts.

[0071] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0073] Figure 1 A flow chart of a method for identifying abnormal comment accounts provided by an embodiment of the present invention is shown;

[0074] Figure 2 A flow chart of another abnormal account identification method for comment accounts provided by an embodiment of the present invention is shown;

[0075] Figure 3 A block diagram showing the composition of an abnormal account identification device for comment accounts provided by an embodiment of the present invention is shown;

[0076] Figure 4 A schematic diagram of the structure of a terminal provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0077] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0078] In view of the problem that the existing review account identification efficiency is low, the embodiment of the present invention provides a review account identification method for abnormal accounts, such as Figure 1 As shown, the method includes:

[0079] 101. In response to an instruction for identifying abnormalities in comment accounts of a target topic, obtain comment data of articles related to the target topic that were published within a preset time period.

[0080] In the embodiment of the present invention, the objects of abnormal account identification are mainly social media accounts that post comments under articles of specified topics. Comment data includes comment account information (account name, account ID, etc.) and comment content. The target topic can be about a certain corporate brand, a certain product under a certain brand, a certain hot event, a certain entertainment program, a certain film and television drama, etc., which is not specifically limited in the embodiment of the present invention. After receiving the abnormal identification instruction of the comment account of the target topic, the current execution subject obtains the comment data under all articles about the target topic within a preset time period in order to identify the comment account. Among them, the preset time period can be one week, one natural month, or it can be customized according to the actual application scenario; the publishing platform of the article can be the same social media platform or different social media platforms; the current execution subject can be a local server, a cloud platform service, etc., and the above content is not specifically limited in the embodiment of the present invention.

[0081] Taking the target topic as a certain corporate brand as an example, when it is necessary to understand the abnormal account status of comments under the introduction or promotion articles related to this brand, the abnormal account identification platform interface can be selected or input through the terminal device. The brand (target brand) that needs to be understood can be requested to identify the abnormal account of the comment account of the target brand to generate the abnormal account identification instruction of the comment account of the target topic. After receiving the abnormal account identification instruction of the comment account of the target topic, the abnormal account identification platform server captures the comment data published under all articles about the target brand published within one month.

[0082] 102. Filter out the comment account to be identified from the global comment accounts based on the comment data, and retrieve the account dimension information and comment dimension information of the comment account to be identified.

[0083] In an embodiment of the present invention, the comment data includes comment account information (account name, account ID, etc.) and comment content. Since the comment data volume is large and the number of accounts participating in the comment is also large, if all accounts that post comments are identified, it will cause a great consumption of the data processing resources of the current execution entity and will also reduce the efficiency of identification. Therefore, the comment accounts that need to be identified, that is, the comment accounts to be identified, are screened out from all the comment accounts involved in the comment data to reduce the amount of data processing and the number of account identifications. Among them, the screening of comment accounts can be based on one or more of the number of comments posted by the comment account, the emotional tendency of the comment content, the historical identification results of the comment account, and the level information of the comment account.

[0084] After determining the comment account to be identified, the account dimension information and comment dimension information of each comment account to be identified are obtained. In order to ensure the validity of the information, the obtained initial account dimension information and initial comment dimension information can also be filtered, including the elimination of null values, incomplete data, and abnormal values, so as to obtain account dimension information and comment dimension information containing only valid data.

[0085] 103. For the comment account to be identified, basic dimension features and supplementary dimension features are extracted based on the account dimension information and the comment dimension information, and feature relevance is divided based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified.

[0086] In the embodiment of the present invention, feature extraction is performed on the account dimension information and the comment dimension information respectively to obtain basic dimension features, that is, features that can be directly extracted from the account dimension signal and the comment dimension information, specifically including account dimension features and comment dimension features. For example, account dimension features such as account ID, account name, platform name, number of followers, number of followings, and number of posts can be directly extracted from the account dimension information. Comment dimension features such as comment content, comment time (the interval between two adjacent comments), and IP location can be directly extracted from the comment dimension information. Supplementary dimension information is a feature obtained by further identification based on the integrated dimension features, and may include comment content emotion category, media identity, number of negative comments, etc. Among them, the comment content emotion category can be the emotion classification prediction result of the comment content by the emotion recognition model, including positive emotion, negative emotion, and neutral emotion. The number of negative comments can be obtained by counting the number of comments whose comment emotion category is the negative emotion category. The media identity can be classified and identified by the identity tag or number of followers in the authentication information, and may include amateurs and self-media. After determining the basic dimension features and supplementary dimension features, in order to facilitate subsequent data processing, the features can also be normalized. For example, if the comment content sentiment category is positive, it is recorded as 1, if it is negative, it is recorded as -1, and if it is neutral, it is recorded as 0. For another example, if the media identity is amateur, it is recorded as 0, and if the media identity is self-media, it is recorded as 1.

[0087] In order to screen the feature items and evaluate the importance of different feature items, the correlation between each feature item and the abnormal account is evaluated, and the feature items in the basic dimension features and the supplementary dimension features are divided according to the evaluation results to obtain multi-level identification features. Among them, multiple levels of division can be made according to the correlation corresponding to different feature items. For example, posting multiple negative comments has a strong correlation with abnormal accounts, which can be classified as the highest level of identification features; the correlation between IP location anomalies and abnormal accounts is medium, which can be classified as a medium-level identification feature; the correlation between the number of posts and abnormal accounts is weak, so they can be deleted.

[0088] It should be noted that by dividing the extracted features according to their relevance to abnormal accounts, the importance of different features as identification basis in the subsequent abnormal account identification process can be determined. For example, the highest level identification feature can be used as the main identification basis, and the medium level identification feature can be used as the secondary identification basis. By dividing at different relevance levels, more effective features can be learned in the subsequent model identification process, thereby improving the efficiency and accuracy of model identification.

[0089] 104. Using the trained comment account anomaly recognition model, identify the multi-level recognition features to obtain an account recognition result of the comment account to be identified.

[0090] In an embodiment of the present invention, the comment account anomaly identification model is a pre-trained classification prediction model. The multi-level identification features corresponding to each comment account to be identified are input into the trained comment account anomaly identification model. The model predicts the type of comment account to be identified based on the classification strategy or judgment logic learned during the training process to obtain the account identification result of the comment account to be identified. The account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts. Suspected abnormal accounts are comment accounts that are more likely to be abnormal accounts; abnormal tendency accounts are comment accounts that have the possibility of being abnormal accounts, but further verification is required. Such accounts can be further paid attention to; non-abnormal accounts are normal comment accounts that exclude the possibility of abnormal accounts.

[0091] It should be noted that by extracting features from the multi-dimensional information of the account to be identified, dividing the features according to the correlation between different features and abnormal accounts, and then using the trained comment account anomaly identification model to identify the divided multi-level identification features, the degree of dependence of the identification process on manual work is greatly reduced, and the time and labor costs of comment account identification are reduced. At the same time, the accuracy of identifying whether a comment account is an abnormal account is ensured, thereby greatly improving the efficiency of comment account identification.

[0092] In an embodiment of the present invention, for further explanation and limitation, as Figure 2 As shown, the step of dividing the feature relevance according to the basic dimension features and the supplementary dimension features to obtain the multi-level identification features of the comment account to be identified includes:

[0093] 201. Calculate the abnormal account correlation for different features under each dimension respectively, and obtain the correlation coefficient of different feature items.

[0094] 202. Mark the feature item whose correlation coefficient is greater than the first correlation threshold as a primary correlation identification feature.

[0095] 203. Mark the feature item whose correlation coefficient is greater than the second correlation threshold and less than the first correlation coefficient threshold as a secondary correlation identification feature.

[0096] 204. Mark the feature item whose correlation coefficient is less than the second correlation threshold as a third-level correlation identification feature.

[0097] In the embodiment of the present invention, the basic dimension features and the supplementary dimension features include multiple features. In order to accurately evaluate the correlation between different feature items and abnormal account identification features, the Spearman correlation coefficient method is used to calculate the correlation coefficient corresponding to each feature item. For example, as shown in Table 1, the correlation coefficients corresponding to different features such as posting multiple negative comments, intensive posting of comments in a short period of time, IP location anomalies, and the number of followers can be calculated based on the Spearman correlation coefficient method, and the features can be divided into different correlation levels.

[0098] Table 1:

[0099]

[0100] Among them, the IP location anomaly in Table 1 is determined according to the frequency of IP location change of the account within the historical time period. For example, for an account that publishes multiple comments on a platform, the work data of this account within the historical 6 months is traced back. If the IP location change frequency of this account is greater than the preset change frequency, the IP location anomaly of this account is determined. The length of the historical time period can be customized according to actual needs; the preset change frequency can be changed 3 times within 1 month, or it can be customized according to actual needs, and the embodiment of the present invention does not make specific limitations. Intensive posting of comments in a short period of time is determined based on the comment time. For example, for an account on a platform, the comment time of this comment account is counted. If the time interval between multiple comments posted by this comment account is less than the preset duration, it can be determined that this account has the characteristics of intensive posting in a short period of time. Among them, the preset duration can be 1 minute, or it can be customized according to actual application needs, and the embodiment of the present invention does not make specific limitations.

[0101] After the correlation coefficient is calculated, the correlation level of each feature item is divided according to the comparison result of the correlation coefficient corresponding to the feature with the first correlation threshold and the second correlation threshold. Among them, the first correlation threshold is greater than the second correlation threshold, that is, the first-level correlation identification feature, which is used to characterize that the identification feature of this correlation level is strongly correlated with the abnormal account; the second-level correlation identification feature, which is used to characterize that the identification feature of this correlation level is moderately correlated with the abnormal account; the third-level correlation identification feature, which is used to characterize that the identification feature of this correlation level is weakly correlated with the abnormal account. After the features are divided into correlation levels, the third-level correlation identification feature can be discarded as an eliminated feature, and only the first-level correlation identification feature and the second-level correlation identification feature are retained as multi-level identification features and input into the abnormal recognition model. The third-level correlation identification feature can also be retained, and the first-level correlation identification feature, the second-level correlation identification feature, and the third-level correlation identification feature are collectively used as multi-level identification features and input into the abnormal recognition model to serve as secondary auxiliary identification features in the recognition process. The first correlation threshold may be set to a correlation coefficient of 0.6, and the second correlation threshold may be set to a correlation coefficient of 0.2. They may also be customized according to actual application requirements, and are not specifically limited in the embodiments of the present invention.

[0102] In one embodiment of the present invention, for further explanation and limitation, the use of the trained comment account abnormality recognition model to identify the multi-level recognition features to obtain the account recognition result of the comment account to be identified includes:

[0103] In a case where the multi-level identification features include a first-level related identification feature, determining that the comment account to be identified is a suspected abnormal account;

[0104] In the case that the multi-level identification features do not include the primary related identification features, if the number of feature items of the secondary related identification features in the multi-level identification features is greater than the feature item number threshold, then determining that the comment account to be identified is an abnormal tendency account;

[0105] If the number of feature items of the secondary related identification features in the multi-level identification features is less than the feature item number threshold, it is determined that the comment account to be identified is a non-abnormal account.

[0106] In the embodiment of the present invention, the first-level related identification feature is used as the main basis for identifying abnormal accounts. When the multi-level identification features of a comment account to be identified include a first-level related identification feature, the current comment account to be identified can be directly determined to be a suspected abnormal account. Of course, if the multi-level identification features include multiple first-level related identification features, the current comment account to be identified can be judged to be a suspected abnormal account. The second-level related identification feature is used as a secondary and auxiliary basis for judging abnormal accounts. When the multi-level identification features do not include the first-level related identification features, it is necessary to judge based on the number of second-level related identification features in the multi-level identification features. If the number of second-level related identification features is greater than the threshold value of the number of feature items, the current comment account to be identified can be judged to be a comment account with abnormal tendencies (abnormal tendency account); otherwise, the current comment account to be identified can be judged to be a non-abnormal account. Among them, the threshold value of the number of feature items can be customized according to specific application requirements, and the embodiment of the present invention does not make specific limitations.

[0107] It should be noted that, when the multi-level identification features include the third-level related identification features, the number of the third-level related identification features can also be used as a secondary auxiliary basis in the process of distinguishing the second-level related identification features, and judged together with the number of the second-level related identification features. For example, when the number of feature items of the second-level related identification features is equal to the feature item number threshold, the number of the third-level related identification features is counted. If the number of the third-level related identification features is greater than the feature item number threshold, it can be judged that the current comment account to be identified is an abnormal tendency account; otherwise, it can be judged that the current comment account to be identified is a non-abnormal account.

[0108] It should be noted that comment data can involve multiple social media platforms. As shown in Table 2, some recognition results of abnormal account comments on a certain topic are shown, including recognition features and recognition results.

[0109] Table 2:

[0110]

[0111] In one embodiment of the present invention, for further explanation and limitation, the step of selecting the comment account to be identified from the global comment accounts based on the comment data includes:

[0112] Counting the number of comments from different comment accounts in the comment data to determine the number of comments published by each comment account;

[0113] If the number of comments published is greater than a preset comment publication threshold, or the proportion of the number of participating comment articles in the global number of articles is greater than a preset comment participation threshold, the comment account is determined as a candidate comment account;

[0114] Emotion recognition is performed on the comment content of the candidate comment account one by one. If the recognition result is that the number of positive emotions or negative emotions is greater than a preset emotion number threshold, the candidate comment account is determined as a comment account to be identified.

[0115] In an embodiment of the present invention, in order to screen the comment accounts, the comment accounts are screened based on two dimensions: the comment account's participation in the comments on the target topic and the emotional tendency of the comment content. Specifically, the comment publishing volume of each comment account can be extracted from the comment data, that is, the number of comments published by the comment account on the articles related to the target topic. The comment account whose comment publishing volume is greater than the preset comment publishing volume threshold is determined as a candidate comment account. The number of articles related to the target topic commented on by the comment account can also be extracted from the comment data, and the proportion of the commented articles in the global articles is calculated, and the comment account whose proportion is greater than the preset comment participation threshold is determined as the account to be identified. Among them, the global number of articles is the number of all articles related to the target topic. For example, if the target topic is a brand, the global number of articles is the number of all articles published by a brand or related to a brand.

[0116] After screening the comment accounts from the dimension of the target topic comment participation of the comment account, the candidate comment accounts are further screened from the dimension of the emotional tendency of the comment content. Specifically, the emotional recognition model is used to perform emotional classification on the comment content published by the candidate comment account. If the emotional classification is non-neutral, that is, positive or negative, the candidate comment account is determined as an account that needs to be identified, that is, a comment account to be identified. Among them, the emotional recognition model can be an emotional classification prediction model obtained by training based on the NLP (Natural Language Processing) model. Since abnormal accounts can be divided into two types of negative abnormal accounts and positive abnormal accounts, the purpose of negative abnormal accounts is to publish negative and malicious remarks on the target topic, and the purpose of positive abnormal accounts is to publish exaggerated propaganda remarks that exaggerate the advantages of the target topic. Once the comment content published by the comment account to be identified has an emotional tendency, whether it is positive or negative, it is possible to be an abnormal account. Therefore, the candidate comment accounts whose emotional recognition results are that the number of comments with positive emotions is greater than the preset emotional number threshold and the number of comments with negative emotions is greater than the preset emotional number threshold are determined as comment accounts to be identified. Among them, the preset comment publication volume threshold, the preset comment participation threshold and the preset emotion quantity threshold can all be customized according to actual needs. For example, the preset comment publication volume threshold is 2, the preset comment participation threshold is 20%, and the preset emotion quantity threshold is 2. The embodiment of the present invention does not make specific limitations.

[0117] In one embodiment of the present invention, for further explanation and limitation, before using the trained comment account abnormality recognition model to recognize the multi-level recognition features, the method further includes:

[0118] Obtain sample comment data for multiple articles on the same topic and identify sample accounts;

[0119] Extracting multi-level identification sample features based on the account dimension data and comment dimension data of the sample account, and constructing a training sample set based on the target identification result label of the sample account and the multi-level identification sample features;

[0120] The extreme gradient boosting model is trained according to the training sample set until the model performance parameters of the extreme gradient boosting model meet the corresponding parameter thresholds, thereby obtaining a comment account anomaly recognition model that has completed training.

[0121] In an embodiment of the present invention, in order to construct a training sample set for the model, sample comment data of multiple articles on the same topic are obtained, sample accounts are determined based on the sample comment data, and feature extraction is performed on the account dimension data and comment dimension data of the sample account to obtain multi-level identification sample features. Whether the sample account is an abnormal account is identified by manual means, the true identity of the sample account is determined, and the identity of the sample account is used as the target identification result to mark the sample account. Among them, the target identification result includes suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts. Among them, the identification of the sample is exactly the same as the identification of the comment account to be identified by the same technical means; the feature extraction of the account dimension data and comment dimension data of the sample account is the same as the feature extraction process of the comment account to be identified described in steps 103 and steps 201 to 204, which will not be repeated here.

[0122] End-to-end gradient boosting is used as the basic algorithm model for the comment account anomaly recognition model. The extreme gradient boosting algorithm (XGBoost) is an integrated machine learning algorithm based on decision trees. This model can efficiently process large-scale complex data and improve the processing speed through parallel computing and memory optimization. In addition, when modeling and processing data, it can effectively prevent data overfitting and improve the generalization ability of the model. In addition, extreme gradient boosting can process features in both numerical and categorical forms at the same time, and can better adapt to the diverse user behavior indicators in abnormal account recognition. In addition, the good processing ability and incremental learning characteristics of end-to-end gradient boosting for unbalanced data enable this algorithm to solve the imbalance problem in which non-abnormal accounts dominate the comment data, so as to adapt to the dynamic changes in abnormal account behavior and continuously improve the recognition accuracy. Therefore, end-to-end gradient boosting, as the basic algorithm model for the comment account anomaly recognition model, can meet the application scenarios of comment abnormal account recognition and better identify abnormal accounts.

[0123] In one embodiment of the present invention, for further explanation and limitation, the step of training the extreme gradient boosting model according to the training sample set until the model performance parameter of the end gradient boosting model meets the corresponding parameter threshold, and obtaining a trained comment account anomaly recognition model includes:

[0124] Dividing the training sample set into a training set, a validation set and a test set according to a preset division ratio;

[0125] The extreme gradient boosting model is trained according to the training set and the initial parameters of the model to optimize the decision tree split point and the feature threshold, thereby obtaining a comment account anomaly recognition model to be verified;

[0126] The model performance test of the comment account anomaly recognition model to be verified is performed respectively according to the verification set and the test set to obtain model performance parameters, and when the model performance parameters meet the target performance parameters, a comment account anomaly recognition model that has completed training is obtained.

[0127] In an embodiment of the present invention, after the training sample set is extracted, in order to balance the sample sizes of the training set, the validation set, and the test set and avoid the risk of overfitting or underfitting during the model training process, the training sample set is divided according to a preset division ratio. Preferably, the preset division ratio may be 70% for the training set, 15% for the validation set, and 15% for the test set. Division according to this preset division ratio can ensure that the training set has sufficient data to train the model, and that the validation set and the test set have sufficient independence to evaluate the generalization ability of the model. Before training, the classifier or regressor of the extreme gradient boosting model needs to be initialized, and the model initial parameters of the extreme gradient boosting model need to be configured. The model initial parameters include, but are not limited to, the learning rate, the depth of the tree, and the subsample ratio.

[0128] During the training process, the extreme gradient boosting model repeatedly traverses all input features and tries different split points and thresholds to find the split point that maximizes the information gain until the preset number of trees is reached or other stopping conditions are met. Through repeated training and optimization, the most reasonable threshold of the key features is determined to achieve the highest recognition accuracy and obtain the abnormal recognition model of the comment account to be verified. Further based on the performance parameters of the abnormal recognition model of the comment account to be verified on the validation set, the model parameters and feature preprocessing steps are tuned, and a new round of training and evaluation is carried out until the model performance parameters reach the preset target performance parameters. Finally, the performance of the model tuned on the validation set is evaluated using the test set to ensure that the model can also show good generalization ability on unseen data.

[0129] It should be noted that the model performance parameters include accuracy, recall, and F1 score. Table 3 shows a set of experimental result data.

[0130] Table 3:

[0131] Dimensions Accuracy Recall F1 score Normal user 0.93 0.91 0.92 Abnormal account 0.87 0.85 0.86

[0132] From the data in the table above, we can see that the model has an accuracy of 0.93, a recall rate of 0.91, and an F1 score of 0.92 for normal users. The model has an accuracy of 0.87, a recall rate of 0.85, and an F1 score of 0.85 for suspected abnormal accounts. This shows that the model has high accuracy and recall rate in identifying suspected abnormal account users and non-abnormal account users, and has good recognition effect. It can be used to identify abnormal accounts of comment accounts.

[0133] In one embodiment of the present invention, for further explanation and limitation, when the comment account abnormality identification instruction carries a negative abnormal account analysis requirement, after obtaining the account identification result of the comment account to be identified, the method further includes:

[0134] For a comment account whose account identification result is a suspected abnormal account or an abnormal account tendency, identifying the sentiment tendency category of the comment account;

[0135] If the account identification result is a suspected abnormal account, and the sentiment tendency category is a negative tendency, the comment account is determined as a suspected negative abnormal account;

[0136] If the account identification result is an abnormal account with a tendency to be negative, and the sentiment tendency category is negative, the comment account is determined to be an abnormal account with a tendency to be negative, and a negative abnormal account analysis result is generated based on the number of suspected negative abnormal accounts and the number of abnormal accounts with a tendency to be negative.

[0137] In the case where the target topic is a certain brand, the demand for abnormal account identification is often proposed by the brand side. The brand side needs to understand whether there are any negative abnormal accounts on the Internet such as suspected manipulation, misleading public opinion, and hype topics in the comments on articles related to its brand, and understand the clues of negative abnormal accounts to grasp the real public opinion situation and customer needs about the brand. In view of this situation, in an embodiment of the present invention, the emotional tendency categories of suspected abnormal accounts and accounts with a tendency to abnormal are further identified, and can be specifically identified based on the emotional characteristics of the comment content. After determining the suspected abnormal accounts and the accounts with a tendency to negative abnormality, the number of suspected negative abnormal accounts and the number of accounts with a tendency to negative abnormality are extracted, and then the negative abnormal account analysis results are generated. The negative abnormal account analysis results may include statistical data such as the number of suspected negative abnormal accounts, the number of accounts with a tendency to negative abnormality, the proportion of suspected negative abnormal accounts relative to all comment accounts, and the proportion of the number of accounts with a tendency to negative abnormality relative to all comment accounts, and can be presented in the negative abnormal account analysis results in the form of charts, thereby providing comprehensive and accurate data support for the brand side to grasp the real public opinion situation and customer needs of the brand and maintain brand reputation and competitiveness.

[0138] The present invention provides a method for identifying abnormal accounts of comment accounts. An embodiment of the present invention obtains comment data of articles related to a target topic and published within a preset time period in response to an abnormal identification instruction for comment accounts of a target topic; based on the comment data, a comment account to be identified is screened out from global comment accounts; for the comment account to be identified, basic dimension features and supplementary dimension features are extracted based on account dimension information and comment dimension information, and feature relevance is divided based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified; using a trained abnormal identification model for comment accounts, the multi-level identification features are identified to obtain an account identification result of the comment account to be identified, which greatly reduces the time for identifying abnormal accounts of comment accounts and reduces the dependence of comment account identification on manual work. At the same time, the accuracy of abnormal account identification is ensured, thereby greatly improving the efficiency of abnormal account identification of comment accounts.

[0139] Furthermore, as a response to the above Figure 1 The implementation of the method shown in the figure, the embodiment of the present invention provides a device for identifying abnormal account of comment account, such as Figure 3 As shown, the device comprises:

[0140] An acquisition module 31 is used to respond to an abnormality identification instruction of a comment account of a target topic and acquire comment data of articles related to the target topic published within a preset time period;

[0141] A screening module 32 is used to screen out the comment accounts to be identified from the global comment accounts according to the comment data, and retrieve the account dimension information and comment dimension information of the comment accounts to be identified;

[0142] The feature extraction module 33 is used to extract basic dimension features and supplementary dimension features for the comment account to be identified according to the account dimension information and the comment dimension information, and to perform feature relevance classification according to the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified;

[0143] The identification module 34 is used to use the trained comment account abnormality identification model to identify the multi-level identification features and obtain the account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

[0144] Furthermore, the feature extraction module 33 includes:

[0145] A calculation unit, used to calculate the abnormal account relevance for different features under each dimension, and obtain the correlation coefficients of different feature items; calculate the abnormal account relevance for different features under each dimension, and obtain the correlation coefficients of different feature items;

[0146] A first marking unit, used to mark the feature item whose correlation coefficient is greater than a first correlation threshold as a primary correlation identification feature;

[0147] A second marking unit, used to mark the feature item whose correlation coefficient is greater than a second correlation threshold and less than the first correlation coefficient threshold as a secondary correlation identification feature;

[0148] A third marking unit, used for marking the feature item whose correlation coefficient is less than the second correlation threshold as a third-level correlation identification feature;

[0149] The first correlation threshold is greater than the second correlation threshold.

[0150] Furthermore, the identification module 34 includes:

[0151] A first determining unit is configured to determine that the comment account to be identified is a suspected abnormal account when the multi-level identification features include a first-level related identification feature;

[0152] A second determination unit is configured to determine that the comment account to be identified is an abnormal tendency account if the number of feature items of the second-level related identification features in the multi-level identification features is greater than a feature item number threshold when the first-level related identification features are not included in the multi-level identification features;

[0153] The third determination unit is used to determine that the comment account to be identified is a non-abnormal account if the number of feature items of the secondary related identification features in the multi-level identification features is less than the feature item number threshold.

[0154] Furthermore, the screening module 32 includes:

[0155] A statistical unit, used to count the number of comments from different comment accounts in the comment data, and determine the number of comments published by each comment account;

[0156] a fourth determination unit, configured to determine the comment account as a candidate comment account if the number of comments published is greater than a preset comment publication threshold, or the proportion of the number of participating comment articles in the global number of articles is greater than a preset comment participation threshold;

[0157] The fifth determination unit is used to perform emotion recognition on the comment content of the candidate comment account one by one, and if the recognition result is that the number of positive emotions or negative emotions is greater than a preset emotion number threshold, the candidate comment account is determined as a comment account to be identified.

[0158] Furthermore, the device also includes:

[0159] The first determination module is also used to obtain sample comment data of multiple articles related to the same topic and determine sample accounts;

[0160] A construction module, used to extract multi-level identification sample features based on the account dimension data and comment dimension data of the sample account, and to construct a training sample set based on the target identification result label of the sample account and the multi-level identification sample features;

[0161] A training module is used to train the extreme gradient boosting model according to the training sample set until the model performance parameters of the extreme gradient boosting model meet the corresponding parameter thresholds, thereby obtaining a comment account anomaly recognition model that has completed training.

[0162] Furthermore, the training module includes:

[0163] A division unit, used to divide the training sample set into a training set, a validation set and a test set according to a preset division ratio;

[0164] An optimization unit, used for training the extreme gradient boosting model according to the training set and the initial parameters of the model, so as to optimize the splitting point and the feature threshold of the decision tree, and obtain an abnormal recognition model of the comment account to be verified;

[0165] The testing unit is used to perform model performance tests on the comment account anomaly recognition model to be verified based on the verification set and the test set, obtain model performance parameters, and obtain a trained comment account anomaly recognition model when the model performance parameters meet the target performance parameters.

[0166] Furthermore, the device also includes:

[0167] The identification module is further used to identify the sentiment tendency category of the comment account for the comment account whose account identification result is suspected to be an abnormal account or tends to be an abnormal account;

[0168] A second determination module is configured to determine the comment account as a suspected negative abnormal account if the account identification result is a suspected abnormal account and the sentiment tendency category is a negative tendency;

[0169] A generation module is used to determine the comment account as a negative abnormal account if the account identification result is an abnormal account with a tendency to be negative and the emotional tendency category is negative, and generate a negative abnormal account analysis result based on the number of suspected negative abnormal accounts and the number of negative abnormal accounts.

[0170] The present invention provides an abnormal account identification device for comment accounts. An embodiment of the present invention obtains comment data of articles related to a target topic and published within a preset time period in response to an abnormal comment account identification instruction for a target topic; based on the comment data, a comment account to be identified is screened out from global comment accounts; for the comment account to be identified, basic dimension features and supplementary dimension features are extracted based on account dimension information and comment dimension information, and feature relevance is divided based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified; using a trained abnormal comment account identification model, the multi-level identification features are identified to obtain an account identification result of the comment account to be identified, which greatly reduces the time for abnormal comment account identification and reduces the reliance on manual labor for comment account identification. At the same time, the accuracy of abnormal account identification is ensured, thereby greatly improving the efficiency of abnormal comment account identification.

[0171] According to one embodiment of the present invention, a storage medium is provided, wherein the storage medium stores at least one executable instruction, and the computer executable instruction can execute the abnormal account identification method of the comment account in any of the above method embodiments.

[0172] Figure 4 A schematic diagram of the structure of a terminal provided according to an embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the terminal.

[0173] like Figure 4As shown, the terminal may include: a processor (processor) 402 , a communication interface (Communications Interface) 404 , a memory (memory) 406 , and a communication bus 408 .

[0174] The processor 402 , the communication interface 404 , and the memory 406 communicate with each other via a communication bus 408 .

[0175] The communication interface 404 is used for network communication with other devices such as a client or other servers.

[0176] The processor 402 is used to execute the program 410, and specifically can execute the relevant steps in the embodiment of the abnormal account identification method of the comment account.

[0177] Specifically, the program 410 may include program codes, which include computer operation instructions.

[0178] The processor 402 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiment of the present invention. The one or more processors included in the terminal may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0179] The memory 406 is used to store the program 410. The memory 406 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0180] The program 410 may be specifically configured to enable the processor 402 to perform the following operations:

[0181] In response to an instruction for identifying abnormalities in comment accounts of a target topic, obtaining comment data of articles related to the target topic that were published within a preset time period;

[0182] Based on the comment data, filter out the comment account to be identified from the global comment accounts, and retrieve the account dimension information and comment dimension information of the comment account to be identified;

[0183] For the comment account to be identified, extract basic dimension features and supplementary dimension features according to the account dimension information and the comment dimension information, and perform feature relevance classification according to the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified;

[0184] The multi-level identification features are identified using the trained comment account anomaly identification model to obtain account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

[0185] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying abnormal comment accounts, characterized in that: include: In response to an instruction for identifying abnormalities in comment accounts of a target topic, obtaining comment data of articles related to the target topic that were published within a preset time period; Based on the comment data, filter out the comment account to be identified from the global comment accounts, and retrieve the account dimension information and comment dimension information of the comment account to be identified; For the comment account to be identified, extract basic dimension features and supplementary dimension features according to the account dimension information and the comment dimension information, and perform feature relevance classification according to the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified; The multi-level identification features are identified using the trained comment account anomaly identification model to obtain account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

2. The method according to claim 1, characterized in that The basic dimension features include account dimension features and comment dimension features, the supplementary dimension features include account supplementary features, comment supplementary features and comment content sentiment features, and the feature relevance division is performed based on the basic dimension features and the supplementary dimension features to obtain the multi-level identification features of the comment account to be identified, including: Calculate the abnormal account relevance for different features under each dimension and obtain the correlation coefficient of different feature items; Marking the feature item whose correlation coefficient is greater than the first correlation threshold as a first-level correlation identification feature; Marking the feature item whose correlation coefficient is greater than a second correlation threshold and less than the first correlation coefficient threshold as a secondary correlation identification feature; Marking the feature item whose correlation coefficient is smaller than the second correlation threshold as a third-level correlation identification feature; The first correlation threshold is greater than the second correlation threshold.

3. The method according to claim 2, characterized in that The method of using the trained comment account abnormality recognition model to recognize the multi-level recognition features and obtaining the account recognition result of the comment account to be recognized includes: In a case where the multi-level identification features include a first-level related identification feature, determining that the comment account to be identified is a suspected abnormal account; In the case that the multi-level identification features do not include the primary related identification features, if the number of feature items of the secondary related identification features in the multi-level identification features is greater than the feature item number threshold, then determining that the comment account to be identified is an abnormal tendency account; If the number of feature items of the secondary related identification features in the multi-level identification features is less than the feature item number threshold, it is determined that the comment account to be identified is a non-abnormal account.

4. The method according to claim 1, characterized in that The step of screening out the comment accounts to be identified from the global comment accounts based on the comment data includes: Counting the number of comments from different comment accounts in the comment data to determine the number of comments published by each comment account; If the number of comments published is greater than a preset comment publication threshold, or the proportion of the number of participating comment articles in the global number of articles is greater than a preset comment participation threshold, the comment account is determined as a candidate comment account; Emotion recognition is performed on the comment content of the candidate comment account one by one. If the recognition result is that the number of positive emotions or negative emotions is greater than a preset emotion number threshold, the candidate comment account is determined as a comment account to be identified.

5. The method according to claim 1, characterized in that Before using the trained comment account anomaly recognition model to recognize the multi-level recognition features, the method further includes: Obtain sample comment data for multiple articles on the same topic and identify sample accounts; Extracting multi-level identification sample features based on the account dimension data and comment dimension data of the sample account, and constructing a training sample set based on the target identification result label of the sample account and the multi-level identification sample features; The extreme gradient boosting model is trained according to the training sample set until the model performance parameters of the extreme gradient boosting model meet the corresponding parameter thresholds, thereby obtaining a comment account anomaly recognition model that has completed training.

6. The method according to claim 5, characterized in that The extreme gradient boosting model is trained according to the training sample set until the model performance parameter of the end gradient boosting model meets the corresponding parameter threshold, thereby obtaining a trained comment account anomaly recognition model, including: Dividing the training sample set into a training set, a validation set and a test set according to a preset division ratio; The extreme gradient boosting model is trained according to the training set and the initial parameters of the model to optimize the decision tree split point and the feature threshold, thereby obtaining a comment account anomaly recognition model to be verified; The model performance test of the comment account anomaly recognition model to be verified is performed respectively according to the verification set and the test set to obtain model performance parameters, and when the model performance parameters meet the target performance parameters, a comment account anomaly recognition model that has completed training is obtained.

7. The method according to any one of claims 1 to 6, characterized in that In the case where the comment account abnormality identification instruction carries an abnormal account analysis requirement, after obtaining the account identification result of the comment account to be identified, the method further includes: For a comment account whose account identification result is a suspected abnormal account or an abnormal account tendency, identifying the sentiment tendency category of the comment account; If the account identification result is a suspected abnormal account, and the sentiment tendency category is a negative tendency, the comment account is determined as a suspected negative abnormal account; If the account identification result is an abnormal account with a tendency to be negative, and the sentiment tendency category is negative, the comment account is determined to be an abnormal account with a tendency to be negative, and a negative abnormal account analysis result is generated based on the number of suspected negative abnormal accounts and the number of abnormal accounts with a tendency to be negative.

8. A device for identifying abnormal account of comment account, characterized in that: include: An acquisition module is used to respond to the comment account abnormality identification instruction of the target topic and acquire the comment data of the articles related to the target topic published within a preset time period; A screening module, configured to screen out a to-be-identified comment account from the global comment accounts based on the comment data, and retrieve account dimension information and comment dimension information of the to-be-identified comment account; A feature extraction module, for extracting basic dimension features and supplementary dimension features from the comment account to be identified based on the account dimension information and the comment dimension information, and performing feature relevance classification based on the basic dimension features and the supplementary dimension features to obtain multi-level identification features of the comment account to be identified; The identification module is used to use the trained comment account abnormality identification model to identify the multi-level identification features and obtain the account identification results of the comment account to be identified, and the account identification results include suspected abnormal accounts, abnormal tendency accounts and non-abnormal accounts.

9. A storage medium storing at least one executable instruction, wherein the executable instruction enables a processor to execute an operation corresponding to the abnormal account identification method for comment accounts as described in any one of claims 1-7.

10. A terminal, comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the abnormal account identification method of the comment account according to any one of claims 1-7.

Images