Political-oriented network public opinion manipulator detection model training method, detection method and device

Abstract

The application provides a political orientation network public opinion manipulator detection model training method, a detection method and a device, comprising: constructing a training data set, each sample of which contains tweets collected from a social network platform; using a latent Dirichlet distribution to model the theme of the training data set, screening out political direction data to construct a political data set; constructing an initial detection model, which includes a feature selection module and a classification module; inputting the political data set, extracting features by the feature selection module, classifying by the classification module, and outputting a predicted classification result; training the initial detection model using the political data set, constructing a loss optimization model of the predicted classification result and the true classification; evaluating and selecting the best classification algorithm in terms of classification performance as the final classification algorithm of the classification model to obtain a detection model. The training method of the application guarantees the class balance ratio of the training set, has high training efficiency, and the detection model obtained by training has high precision and high performance.

Description

Technical Field

[0001] This invention relates to the field of online opinion manipulator detection technology, and in particular to a training method, detection method and device for a politically oriented online opinion manipulator detection model. Background Technology

[0002] In the information age, while social media has made content easily accessible to billions, it has not simultaneously enhanced the public's in-depth understanding of political issues. Politically oriented online manipulators can create false public opinion trends by massively publishing or forwarding content with specific political stances, influencing public perceptions of political events, policies, or candidates, leading to the polarization and distortion of public opinion. These manipulators frequently spread disinformation, rumors, and conspiracy theories, confusing the public and undermining the credibility of information. This information pollution not only misleads the public but can also damage the credibility of the media, leading to public distrust of authoritative sources.

[0003] Users' ability to distinguish between true and false information lags behind, and the development of social media platforms has accelerated the spread of misleading content. Simultaneously, the fragmented nature of social media exacerbates information chaos, allowing disinformation and political propaganda to spread widely. To address these challenges, social media platforms have strengthened usage guidelines and developed technologies to identify online manipulators. While the scientific community has used machine learning algorithms to detect automated accounts with significant success in mainstream languages, the identification of online manipulators in low-resource languages ​​still requires improvement. The information environment on social media is complex and ever-changing, necessitating continuous optimization of detection and filtering mechanisms to maintain information accuracy and social harmony.

[0004] Currently, there are two main methods commonly used to identify manipulators on social networks: First, behavioral analysis methods identify manipulators by analyzing user behavior patterns. For example, manipulator accounts may post a large amount of similar content within a short period or be active at specific times. Researchers analyze user posting frequency, interaction patterns, content similarity, and other behavioral characteristics to distinguish between genuine users and manipulators. Second, content analysis methods identify manipulators by analyzing the content posted by users. Manipulator accounts may post a large amount of repetitive or similar content or use specific keywords or phrases. Natural Language Processing (NLP) techniques, such as text similarity analysis, topic modeling, and sentiment analysis, can be used to identify these patterns.

[0005] However, current online opinion manipulator detection technologies, when applied to the political domain, still have some shortcomings: mainly in the selection of test datasets, the degree of domain segmentation, and accuracy. (1) The imbalance between the proportion of humans and bots in the directly collected datasets is not considered, as is the problem that training models on extremely imbalanced datasets may lead to excessively long training times, and how to balance these two issues, resulting in poor training effects and final performance of the models. (2) Most models suffer from overfitting and poor generalization ability. Currently, the models are more suitable for fields such as entertainment and daily life, but their generalization ability is poor in more segmented fields such as politics, experience, and law. (3) Methods based on behavioral features and content analysis usually require carefully designed features, and feature selection and engineering processes can be very time-consuming and require domain knowledge. The accuracy is low, and online opinion manipulators may adopt new strategies to evade detection, while the models may need to be updated and adjusted regularly to maintain their identification ability. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a training method, detection method and apparatus for a politically oriented online opinion manipulator detection model, so as to eliminate or improve one or more defects existing in the prior art.

[0007] On the one hand, the present invention provides a training method for a politically oriented online opinion manipulator detection model, the method comprising:

[0008] Construct a training dataset containing multiple samples, each containing tweets and metadata of an account collected from a social network platform, including normal user accounts and accounts of online opinion manipulators;

[0009] The training dataset is used to perform topic modeling using the pre-trained latent Dirichlet distribution, and data with political orientation is selected to construct a political dataset.

[0010] An initial detection model is constructed, which includes a feature selection module and a classification module. The classification model has multiple classification algorithms. The initial detection model takes the political dataset as input, extracts features by the feature selection module, and then inputs them into the classification module. The classification module performs classification and outputs the predicted classification result.

[0011] The initial detection model is trained using the political dataset, and a loss between the predicted classification result and the true classification is constructed. The initial detection model is then optimized using the loss. The classification algorithm with the best classification performance is evaluated and selected as the final classification algorithm for the classification model, thus obtaining the network opinion manipulator detection model.

[0012] In some embodiments of the present invention, constructing a training dataset includes:

[0013] Use the application programming interface of the social network platform to obtain tweets and metadata of each account;

[0014] The number of normal user accounts and network opinion manipulator accounts is controlled according to a preset class balance ratio.

[0015] In some embodiments of the present invention, a latent Dirichlet distribution is used to perform topic modeling on the training dataset, and the pre-training step includes:

[0016] Establish political topics or events to build a training set;

[0017] The number of topics is estimated through exploratory data analysis, and the performance of the initial implicit Dirichlet distribution model under different numbers of topics is evaluated using the topic coherence index to determine the optimal number of topics.

[0018] Using the training set and the determined optimal number of subjects, the initial latent Dirichlet distribution model is trained to obtain the final latent Dirichlet distribution model.

[0019] In some embodiments of the present invention, features extracted by the feature selection module are input into the classification module, including:

[0020] Information gain analysis is used to calculate the information gain of each feature relative to the target variable. Features are then sorted according to their information gains to select those with discriminative power. The target variable includes whether the tweet was sent by the account that manipulates online opinion. The higher the information gain, the greater the contribution of the corresponding feature to the classification task.

[0021] In some embodiments of the present invention, after the features are extracted by the feature selection module and input into the classification module, the method further includes: using a violin plot, fusing a box plot and kernel density estimation to show the distribution differences of each feature between different categories.

[0022] In some embodiments of the present invention, before the political dataset is input into the initial detection model, the following is also included:

[0023] The political dataset is preprocessed to remove unnecessary elements from the tweets and then segmented using a word segmentation tool; the unnecessary elements include stop words, punctuation marks, Uniform Resource Locators (URLs), and user mention symbols.

[0024] The preprocessed data is text-encoded using a preset language model and input into the initial detection model with a preset fixed length.

[0025] In some embodiments of the present invention, the method in the initial detection model training includes:

[0026] The initial detection model was trained using hierarchical K-fold cross-validation;

[0027] A binary cross-entropy loss is constructed between the predicted classification result and the true classification, and the initial detection model is optimized by minimizing the binary cross-entropy loss.

[0028] In some embodiments of the present invention, evaluating and selecting the classification algorithm with the best classification performance as the final classification algorithm for the classification model includes:

[0029] The performance of models using various classification algorithms was evaluated using model accuracy and F1 score as the main evaluation criteria, in order to select the classification algorithm with the best classification performance.

[0030] On the other hand, the present invention also provides a method for detecting politically oriented online opinion manipulators, the method comprising:

[0031] Retrieve the tweets to be detected;

[0032] The tweet is input into a network opinion manipulator detection model trained based on any of the politically oriented network opinion manipulator detection model training methods mentioned above, and the classification result of the account corresponding to the tweet is obtained. The classification result includes normal user accounts and network opinion manipulator accounts.

[0033] On the other hand, the present invention also provides a device for detecting politically oriented online opinion manipulators, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device / equipment / system implements the steps of any of the methods mentioned above.

[0034] This invention provides a training method, detection method, and apparatus for a politically-oriented online opinion manipulator detection model. When constructing the training dataset, it considers the class balance ratio to ensure a reasonable proportion of normal users and online opinion manipulators, thereby improving model training efficiency and performance. It uses a latent Dirichlet distribution model to perform topic modeling on the training dataset, eliminating the need for manual annotation and automatically discovering thematic structures hidden in large amounts of text data, filtering out political data to ensure the model's accuracy and interpretability. During training, the loss between the predicted classification results and the actual classification of each classification algorithm is optimized, selecting the classification algorithm with the best performance to finally obtain the online opinion manipulator detection model. The training method of this invention guarantees the class balance ratio of the training set, has high training efficiency, and the trained online opinion manipulator detection model possesses high accuracy and high performance.

[0035] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0036] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0038] Figure 1 This is a schematic diagram illustrating the steps of a training method for a politically oriented online opinion manipulator detection model in one embodiment of the present invention.

[0039] Figure 2 This is a flowchart illustrating the training method for a politically oriented online opinion manipulator detection model in one embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of feature analysis based on violin diagrams in one embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0042] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0043] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0044] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0045] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0046] To address the problems of long training times and poor performance caused by the imbalanced ratio of legitimate user accounts to manipulator accounts in the test datasets of existing online opinion manipulator detection technologies, as well as the issues of overfitting, poor generalization ability, and low detection accuracy in existing online opinion manipulator detection models, this invention provides a training method for a politically oriented online opinion manipulator detection model, such as... Figure 1 As shown, the method includes the following steps S101 to S104:

[0047] Step S101: Construct a training dataset containing multiple samples, each containing tweets and metadata from an account collected from a social network platform. The accounts include legitimate user accounts and accounts manipulated by online opinion manipulators.

[0048] Step S102: Use the pre-trained hidden Dirichlet distribution to perform topic modeling on the training dataset, filter out data with political orientation, and construct a political dataset.

[0049] Step S103: Construct an initial detection model, which includes a feature selection module and a classification module. The classification model incorporates various classification algorithms. The initial detection model takes a political dataset as input. Features are extracted by the feature selection module and then input into the classification module. The classification module performs classification and outputs the predicted classification result.

[0050] Step S104: Train the initial detection model using a political dataset, construct the loss between the predicted classification result and the true classification, and optimize the initial detection model using the loss; evaluate and select the classification algorithm with the best classification performance, and use it as the final classification algorithm for the classification model to obtain the network opinion manipulator detection model.

[0051] like Figure 2 The diagram shown is a flowchart illustrating the training method for a politically-oriented online opinion manipulator detection model.

[0052] In step S101, the training dataset is constructed.

[0053] In some embodiments, application programming interfaces (APIs) of social networking platforms, such as the Twitter API, are used to obtain tweets and metadata from each account, and the number of normal user accounts and accounts manipulated by online opinion manipulators is controlled according to a preset class balance ratio. Metadata refers to additional information related to tweets, which helps in understanding the context and characteristics of the tweets. Tweet metadata includes not only the text content of the tweet but also various other attributes, providing richer background information to aid in analysis and classification. For example, metadata includes tweet ID, publisher information, tweet content information, interaction information, tweet type, content tags, and tweet geolocation information.

[0054] In step S102, the training dataset is modeled using a pre-trained Latent Dirichlet Allocation (LDA) distribution to filter out data with political themes and construct a political dataset. LDA is a commonly used topic modeling method that can automatically extract latent topics from large amounts of text data.

[0055] In some embodiments, training on a latent Dirichlet distribution includes:

[0056] Based on professional materials in the political field, establish a relatively comprehensive set of political topics or events, such as elections, policy discussions, and international relations, to build a training set.

[0057] Through exploratory data analysis, such as word clouds and word frequency distributions, the potential number of topics is initially estimated. Then, topic coherence metrics, such as c_v and c_uci, are used to evaluate the performance of the initial latent Dirichlet distribution model under different numbers of topics, thus determining the optimal number of topics. c_v is a topic coherence scoring method based on word embeddings and context vectors, measuring coherence by calculating the similarity between words in a set of topics. c_uci is based on the UCI (University of California, Irvine) metric, which uses the co-occurrence frequency of words throughout the corpus to calculate coherence, considering the co-occurrence patterns of words within a certain window in a given corpus.

[0058] Train an initial latent Dirichlet distribution model using a preprocessed training set and a determined optimal number of topics. For example, the gensim library in Python can be used for model training. Adjust the parameters of the initial latent Dirichlet distribution model, such as the number of iterations and the random seed, based on model performance. Analyze the word distribution of each topic to understand its meaning; typically, a topic can be described by its top few most probable words. Continuously adjust the model parameters and preprocessing steps based on model performance and application feedback to optimize topic modeling, ultimately obtaining the latent Dirichlet distribution model.

[0059] The trained latent Dirichlet distribution model can perform LDA topic modeling on the training dataset constructed in step S101, such as Twitter data, to filter out politically oriented data and construct a political dataset.

[0060] In step S103, an initial detection model is constructed, which includes a feature selection module and a classification module. In the initial state, the classification model has multiple classification algorithms. In subsequent training, the performance of each classification algorithm is compared and the optimal one is selected.

[0061] The initial detection model takes the political dataset constructed in step S102 as input, extracts features in the feature selection module, classifies according to the extracted features in the classification module, and finally generates a predicted classification result as output.

[0062] In some embodiments, information gain analysis is used to select key features that distinguish political online opinion manipulators, normal users, and other categories of online opinion manipulators. Information gain analysis is a commonly used method for feature selection, especially in classification tasks, as it helps identify the data features that best differentiate between different categories. By calculating the information gain of each feature relative to the target variable (e.g., whether or not a user is a political online opinion manipulator), and ranking the features according to their information gain, the most discriminative features are selected. Higher information gain means a greater contribution of the feature to the classification task.

[0063] For example, by using information gain analysis, a ranking of the information gains of each feature is obtained, and features with higher information gains are selected as key factors for analysis, including:

[0064] Topic Concentration: Whether a large amount of content is published under a specific topic (such as political topics).

[0065] Release timing pattern: Whether activity surges during specific time periods (such as before or after political events).

[0066] Interaction pattern: Whether a highly concentrated social network group is formed, with frequent interactions among members.

[0067] Emotional polarization: Whether the content tends to extreme emotions (such as strong positive or negative emotions).

[0068] Content duplication: Whether the same templates or keywords are used repeatedly.

[0069] Furthermore, such as Figure 3As shown, a violin plot, combined with box plots and kernel density estimation, is used to accurately depict the distribution of each variable and reveal differences in characteristics. Kernel density estimation displays the probability density distribution of variables, helping to understand whether the data exhibits certain regularities (for example, political online opinion manipulators may have more concentrated patterns of activity in certain specific contexts). Box plots further illustrate the quantiles, interquartile ranges, and outliers of the data, identifying central tendency and the degree of variability. By combining these two methods, the distribution of each feature can be analyzed in depth, and significant differences between different categories can be identified.

[0070] For example, taking tweet length as a feature, political online opinion manipulators may have relatively consistent tweet lengths, while ordinary users' tweet lengths will show a wider distribution.

[0071] Based on the feature selection and data visualization results above, the main characteristics of political online opinion manipulators are analyzed as follows:

[0072] Highly focused on specific topics: Tweets from political online opinion manipulators are usually highly focused on specific political topics, especially during events such as elections and policy debates. This focus is often manifested in the high repetition of certain keywords and hashtags.

[0073] Regularity in Behavioral Patterns: The behavioral patterns of political online opinion manipulators exhibit significant regularity, such as a large volume of content being published at specific times (e.g., before and after political events), showing a certain "peak" pattern. In contrast, the behavioral patterns of normal users are more random and unpredictable.

[0074] Content repetition and template-based nature: Political online opinion manipulators often use preset templates or keywords, resulting in highly repetitive content that lacks personalization. In contrast, normal users' tweets are more diverse and personalized, using different expressions.

[0075] Social network cluster effect: Political network opinion manipulators often form specific social network clusters with a high frequency of interaction, such as the concentrated phenomenon of mutual forwarding, commenting, and liking.

[0076] Emotional polarization: The tweets of political online opinion manipulators often exhibit strong emotional polarization, tending to amplify positive or negative emotions, while the emotional expressions of normal users are more centralized and diverse.

[0077] Differences in interaction models: Political online opinion manipulators have different interaction models than normal users. For example, when using hashtags (#hashtag), forwarding, and URL links, they may exhibit more concentrated behavior, while normal users are more casual and diverse.

[0078] Tweet length: The length of tweets from political online opinion manipulators is usually relatively consistent and shows a relatively concentrated distribution, while the length of tweets from normal users shows a large difference, with short, long and other forms.

[0079] Based on the above explanation, the feature selection module inputs the extracted features into the classification module, which then performs classification to obtain the predicted classification result.

[0080] In step S104, the initial detection model is trained using a political dataset.

[0081] In some embodiments, to ensure the balance and diversity of the sample proportions, hierarchical folded cross-validation, such as hierarchical K-fold cross-validation, is used to train the initial detection model. Each cross-validation divides the data into K subsets, where the label proportions in each subset remain consistent, ensuring that each model is trained and evaluated on representative samples.

[0082] In some embodiments, the initial classification module includes eight classification algorithms: Support Vector Machine (SVM), Decision Tree, Random Forest, K-Nearest Neighbors (KNN), Naive Bayes Classifier, Neural Networks, Gradient Boosting Machines (GBM), and Deep Learning Model (LSTM). During training, each model is trained, and the model that performs best on the classification task is selected.

[0083] In some embodiments, for each classification algorithm model, a binary cross-entropy loss is constructed between the predicted classification result and the true classification to optimize the model by minimizing the binary cross-entropy loss.

[0084] In some embodiments, before inputting the political dataset into the initial detection model for training, the political dataset undergoes data preprocessing, including:

[0085] Unnecessary elements in the tweets, such as stop words, punctuation marks, Unicode, and user mentions, are removed, and the tweets are segmented using a word segmentation tool. Pre-defined language models, such as FastText BETO, Twitter 32, and XLM-T53, are used to text-encode the preprocessed data, converting the tweets into a labeled sequence, which is then input into the initial detection model with a pre-defined fixed length.

[0086] In some embodiments, the model performance using each classification algorithm is evaluated using the accuracy and F1 score of each model as the main evaluation criteria, so as to select the classification algorithm with the best classification performance.

[0087] Based on the above explanation, the training and selection of the model were completed, and the final model for detecting online public opinion manipulators was obtained.

[0088] Therefore, the present invention also provides a method for detecting politically oriented online opinion manipulators, the method comprising the following steps S201 to S202:

[0089] Step S201: Obtain the tweets to be detected.

[0090] Step S202: Input the tweet into the network opinion manipulator detection model trained based on steps S101 to S104 to obtain the classification result of the account corresponding to the tweet. The classification result includes normal user accounts and network opinion manipulator accounts.

[0091] Corresponding to the above method, the present invention also provides a device for detecting politically oriented online opinion manipulators. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device performs the steps of the method described above.

[0092] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0093] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0094] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0095] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A training method for a politically-oriented online public opinion manipulator detection model, characterized in that, The method includes: Construct a training dataset containing multiple samples, each containing tweets and metadata of an account collected from a social network platform, including normal user accounts and accounts of online opinion manipulators; The training dataset is used to perform topic modeling using the pre-trained latent Dirichlet distribution, and data with political orientation is selected to construct a political dataset. An initial detection model is constructed, which includes a feature selection module and a classification module. The classification model has multiple classification algorithms. The initial detection model takes the political dataset as input, extracts features by the feature selection module, and then inputs them into the classification module. The classification module performs classification and outputs the predicted classification result. The initial detection model is trained using the political dataset, and a loss between the predicted classification result and the true classification is constructed. The initial detection model is then optimized using the loss. The classification algorithm with the best classification performance is evaluated and selected as the final classification algorithm for the classification model, thus obtaining the network opinion manipulator detection model.

2. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, Construct the training dataset, including: Use the application programming interface of the social network platform to obtain tweets and metadata of each account; The number of normal user accounts and network opinion manipulator accounts is controlled according to a preset class balance ratio.

3. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, Topic modeling is performed on the training dataset using a latent Dirichlet distribution. The pre-training steps include: Establish political topics or events to build a training set; The number of topics is estimated through exploratory data analysis, and the performance of the initial implicit Dirichlet distribution model under different numbers of topics is evaluated using the topic coherence index to determine the optimal number of topics. Using the training set and the determined optimal number of subjects, the initial latent Dirichlet distribution model is trained to obtain the final latent Dirichlet distribution model.

4. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, Features extracted by the feature selection module are then input into the classification module, including: Information gain analysis is used to calculate the information gain of each feature relative to the target variable. Features are then sorted according to their information gains to select those with discriminative power. The target variable includes whether the tweet was sent by the account that manipulates online opinion. The higher the information gain, the greater the contribution of the corresponding feature to the classification task.

5. The training method for the politically oriented online opinion manipulator detection model according to claim 4, characterized in that, The feature selection module extracts features, which are then input into the classification module, and the system further includes: We use violin plots, combined with box plots and kernel density estimation, to show the distribution differences of features among different categories.

6. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, Before the political dataset is input into the initial detection model, it also includes: The political dataset is preprocessed to remove unnecessary elements from the tweets and then segmented using a word segmentation tool; the unnecessary elements include stop words, punctuation marks, Uniform Resource Locators (URLs), and user mention symbols. The preprocessed data is text-encoded using a preset language model and input into the initial detection model with a preset fixed length.

7. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, In the initial detection model training, the method includes: The initial detection model was trained using hierarchical K-fold cross-validation; A binary cross-entropy loss is constructed between the predicted classification result and the true classification, and the initial detection model is optimized by minimizing the binary cross-entropy loss.

8. The training method for the politically oriented online opinion manipulator detection model according to claim 1, characterized in that, Evaluate and select the classification algorithm with the best classification performance as the final classification algorithm for the classification model, including: The performance of models using various classification algorithms was evaluated using model accuracy and F1 score as the main evaluation criteria, in order to select the classification algorithm with the best classification performance.

9. A method for detecting politically oriented online opinion manipulators, characterized in that, The method includes: Retrieve the tweets to be detected; The tweet is input into a detection model trained based on the politically oriented online opinion manipulator detection model training method according to any one of claims 1 to 8, and the classification result of the account corresponding to the tweet is obtained. The classification result includes normal user accounts and online opinion manipulator accounts.

10. A device for detecting politically oriented online opinion manipulators, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device / equipment / system implements the steps of the method as described in any one of claims 1 to 9.

Images