A method and system for detecting social robot groups based on online social platforms

Abstract

This invention proposes a method and system for detecting social robot groups based on a network social platform, comprising: constructing strong features for individual social robot identification, identifying accounts on the network social platform that conform to the strong features as seed accounts; performing extended line analysis based on the social network relationships of the seed accounts, and combining nickname detection to discover and summarize communities with similar nickname paradigms, thereby obtaining social robot communities based on seed accounts; employing a graph convolutional clustering community discovery method to integrate the graph structure and text signals of the network social platform into community detection, and using the social robot communities and seed accounts to identify key communities with robot coverage exceeding a threshold as the social robot group detection result.

Description

Technical Field

[0001] This invention relates to the field of social robot detection, and in particular to a method and apparatus for detecting social robot groups based on a network social platform. Background Technology

[0002] Social bots are robot accounts built on social media platforms that possess machine-controlled attributes. Unlike ordinary user accounts, they are characterized by large-scale, synchronous, and high-frequency operations. Social bots themselves have undergone multiple iterations. First-generation social bots had few social connections and exhibited significant automation. Second-generation bots became more trustworthy and popular, possessing numerous social connections and no longer repeatedly sending the same content. Third-generation bots show a deeper blend of human intervention and automation, even stealing information from other real accounts and using artificial intelligence to generate highly credible text or images, making them even more difficult to distinguish. Social bot detection technology has also evolved from traditional machine learning methods to deep learning methods. Traditional machine learning methods, such as SVM classifiers, identify social bots with obvious automation characteristics through feature construction, while deep learning methods can effectively detect second- and third-generation social bots. Social bots pose numerous potential risks, including influencing public opinion and spreading harmful information; therefore, the detection of social bots is receiving increasing attention. As bots continue to upgrade, corresponding detection technologies must also continuously iterate.

[0003] Social bot detection has long been a focus of research and industry. Traditional machine learning methods primarily employ ensemble learning, support vector machines, and other classifiers combined with feature construction for social bot detection. Some methods process the text of social accounts using natural language processing, while others emphasize the utilization of social network relationships. Deep learning-based research methods are also diverse, with many studies based on graph neural networks, such as graph convolutional neural networks and graph attention mechanisms. These methods combine graph network relationships with text features, user information features, and user behavior features to create vector representations for modeling and recognition. Algorithms based on Siamese networks, residual networks, transfer learning, and generative adversarial networks are also available. Besides methods for detecting individual social bot accounts, there are also methods for detecting social bot groups. Some researchers represent users using vectors and then use clustering methods to identify aggressor groups; others combine community detection methods and rules to identify social bot groups; still others first detect social bots by individual accounts and then detect social bot groups through vector similarity. Currently, there are relatively few methods for detecting social bot groups, and the interpretability and confidence of the detected groups are relatively poor, indicating room for improvement.

[0004] Current social bot detection methods largely focus on individual bot detection, with relatively little research on social bot groups. Existing social bot group detection methods primarily rely on converting various features into implicit vectors, combined with clustering or community detection techniques to arrive at the final identification result. These methods often use relatively simple community detection methods or simple vector approximations, failing to fully utilize community structure and user metadata. A social bot group detection method with high interpretability and confidence is lacking.

[0005] Furthermore, existing social bot group detection methods have room for optimization in each stage, including seed account selection, similar account community discovery, community discovery methods, and final bot determination, in order to improve the accuracy and efficiency of algorithm recognition. Summary of the Invention

[0006] There is relatively little research on existing methods for detecting social robot groups. Algorithms in each stage—seed account selection, similar account community discovery, community discovery methods, and final robot determination—have room for optimization. This invention proposes a method and apparatus for detecting social robot groups based on a network social platform. This method constructs a strong feature classifier to discover social robot seed accounts with high interpretability and confidence. It then discovers similar nickname communities based on social network relationships and nickname detection. Furthermore, it utilizes historical text information and network relationships to construct a graph convolutional clustering community discovery method to identify potential robot communities. Finally, the three methods are combined to achieve the detection and determination of social robot groups.

[0007] This invention also proposes a method for detecting social robot groups based on online social platforms, including:

[0008] Step 1: Construct strong features for individual identification of social robots, and identify accounts on online social platforms that match these strong features as seed accounts;

[0009] Step 2: Based on the social network relationships of the seed account, conduct extended line analysis, and combine nickname detection to discover and summarize similar nickname paradigm communities, thereby obtaining social robot communities based on the seed account;

[0010] Step 3: Employ the graph convolutional clustering community detection method to integrate the graph structure and text signals of the social networking platform into community detection. Then, identify key communities with robot coverage exceeding a threshold through the social robot community and seed accounts, and use these as the social robot group detection results.

[0011] The aforementioned method for detecting social robot groups based on online social platforms, wherein the strong features include:

[0012] Mutual follower rate, reposting frequency, cohesion of posting language, continuous login duration in the current month, frequency of posting IP changes, posting time, topic-driven activity and reposting goals.

[0013] The method for detecting social robot groups based on online social platforms includes step 2, which involves: performing extended analysis on accounts that follow each other with the seed account, identifying accounts with similar naming conventions that follow each other with the robot using a nickname revision method, and merging and clustering them according to similar nickname paradigms to form groups with different nickname paradigms as social robot communities.

[0014] The social robot group detection method based on the network social platform, wherein step 3 includes:

[0015] This paper presents a method for dividing historical texts posted by various accounts on a social networking platform into vectors. These vectors are then clustered into topic clusters, and cluster center vectors are extracted from each cluster to obtain the initial vector representation of the account. A graph network is constructed based on the follower and follower relationships of each account on the social networking platform. Initial unsupervised community partitioning is achieved through label propagation combined with the user's initial feature vectors, resulting in partitioned communities. The direction of label propagation is from followed users to followers. Graph convolution is performed on the node vectors within these partitioned communities, allowing nodes with graph network links to influence each other and updating the vector representation upwards. Clustering is then performed based on the new vector representations. The clustering results are measured by maximizing the average cohesion index, while a penalty term for community partitioning changes is added to the loss function. Finally, the loss function is adjusted in reverse, and the process from community partitioning to graph convolution to clustering is repeated until the loss function converges or the number of iterations is reached. The current community partitioning results and the final vector representations of the community nodes are then saved.

[0016] Based on the seed account and the accounts involved in the social robot community, the coverage rate of the current community segmentation results is calculated, and communities with a coverage rate exceeding the threshold are designated as key communities.

[0017] This invention also proposes a social robot group detection system based on a network social platform, including:

[0018] The initial module is used to construct strong features for individual identification of social robots, and to identify accounts in the online social platform that match the strong features as seed accounts;

[0019] The community discovery module is used to perform extended analysis based on the social network relationships of the seed account, and to discover and summarize communities with similar nickname paradigms by combining nickname detection, thereby obtaining social robot communities based on the seed account.

[0020] The group detection module is used to discover communities using graph convolutional clustering. It incorporates the graph structure and text signals of the social networking platform into the community detection process. Through the social robot community and seed accounts, it identifies key communities with robot coverage exceeding a threshold, which are then used as the social robot group detection results.

[0021] The aforementioned social robot group detection system based on a network social platform, wherein the strong feature includes:

[0022] Mutual follower rate, reposting frequency, cohesion of posting language, continuous login duration in the current month, frequency of posting IP changes, posting time, topic-driven activity and reposting goals.

[0023] The aforementioned social robot group detection system based on a network social platform includes a community discovery module that performs extended analysis on accounts that mutually follow the seed account, identifies accounts with similar naming conventions that mutually follow the robot using a nickname revision method, and merges and clusters them according to similar nickname paradigms to form groups with different nickname paradigms as social robot communities.

[0024] The aforementioned social robot group detection system based on a network social platform, wherein the group detection module includes:

[0025] This paper presents a method for dividing historical texts posted by various accounts on a social networking platform into vectors. These vectors are then clustered into topic clusters, and cluster center vectors are extracted from each cluster to obtain the initial vector representation of the account. A graph network is constructed based on the follower and follower relationships of each account on the social networking platform. Initial unsupervised community partitioning is achieved through label propagation combined with the user's initial feature vectors, resulting in partitioned communities. The direction of label propagation is from followed users to followers. Graph convolution is performed on the node vectors within these partitioned communities, allowing nodes with graph network links to influence each other and updating the vector representation upwards. Clustering is then performed based on the new vector representations. The clustering results are measured by maximizing the average cohesion index, while a penalty term for community partitioning changes is added to the loss function. Finally, the loss function is adjusted in reverse, and the process from community partitioning to graph convolution to clustering is repeated until the loss function converges or the number of iterations is reached. The current community partitioning results and the final vector representations of the community nodes are then saved.

[0026] Based on the seed account and the accounts involved in the social robot community, the coverage rate of the current community segmentation results is calculated, and communities with a coverage rate exceeding the threshold are designated as key communities.

[0027] The present invention also proposes a storage medium for storing a program that executes any of the social robot group detection methods based on a network social platform.

[0028] This invention also proposes a client for any social robot group detection system based on a network social platform.

[0029] As can be seen from the above solutions, the advantages of the present invention are:

[0030] Compared with existing methods, this method can achieve efficient, reliable, and interpretable identification of social robot groups.

[0031] This invention first constructs strong features for the identification of individual social robots, and obtains extremely confident individuals (confirmed robot accounts) to form seed accounts. Based on the seed accounts and the groups of accounts that follow each other, nickname detection is performed to quickly form nickname paradigm groups and nickname paradigm knowledge bases for subsequent detection. Finally, an unsupervised graph convolutional clustering community discovery method is used to form possible social robot groups. The group is interpreted by combining the results of the first two tests, thereby further expanding the group detection results.

[0032] This invention optimizes every aspect of social bot group detection, especially the seed account-based extended identification method. For accounts that are already identified as social bots, they can be directly extended, and for suspected bots that are not yet identified, nickname paradigm detection can be performed to form new seed accounts, enabling continuous social bot group detection and achieving a virtuous cycle. Attached Figure Description

[0033] Figure 1 This is a diagram illustrating the seed account classification process of the present invention.

[0034] Figure 2 This is a diagram illustrating the nickname detection process of the present invention.

[0035] Figure 3 This is a diagram illustrating the community discovery process of the present invention. Detailed Implementation

[0036] There are relatively few existing methods for detecting social bot groups. Existing research shows that there is room for optimization in the algorithms of each stage, including seed account selection, similar account community discovery, community discovery methods, and final bot determination.

[0037] Through in-depth research on social robot group detection, the inventors discovered that the above problems can be solved in the following way: First, a batch of strong features with high discriminative power for social robot detection is constructed, and a robot seed account classifier is built by combining explicit feature engineering with ensemble learning. Second, based on the social network relationships of the seed accounts, line analysis is performed, and nickname detection is combined to discover and summarize communities with similar nickname paradigms, forming social robot communities based on seed accounts. Finally, an optimized graph convolutional clustering community detection algorithm is proposed, which integrates graph structure and text signals into community detection, and identifies key communities with high coverage through the account. Accounts in these communities are used as the third type of detected social robots, enhancing the interpretability of robot detection within communities and optimizing the social robot group detection method.

[0038] To achieve the above-mentioned technical effects, the present invention includes the following key technical points:

[0039] Key Point 1: A set of strong features with high discriminative power for detecting social bots was constructed, including abnormal mutual following rate (mutual following count +10 / account follower count +100), abnormal reposting frequency (similar content reposted more than 3 times in a short period of time), cohesion of posting language (averaging the similarity of 30 random posting content vectors), abnormal continuous login duration in the current month (single continuous posting or interaction exceeding 48 hours and more than 3 times per month), extreme abnormal changes in posting IP (more than 5 changes in IP location in a single day), abnormal posting time (posting time is fixed within a small range of a certain hour each day and is posted frequently), abnormal topic promotion activity (participation in the first 1000 comments or interactions of a certain topic more than 3 times), and abnormal reposting targets (more than 80% of the reposted content comes from other posts by the same account), etc. These features were combined with the random forest method to classify social bots. Technical effect: A social bot individual account classifier based on strong features was formed, and high-confidence accounts with high classification thresholds were selected as seed accounts for further use. The accounts combined with strong features have strong interpretability.

[0040] Existing technologies do not utilize features such as textual cohesion, abnormal topic-driven activity, and mutual follower rate. These are effective features discovered through engineering experiments. Other feature definitions and thresholds differ from existing technologies. For example, regarding abnormal posting times, there are many definitions of "abnormal," but this invention uses the threshold determined through repeated experiments and verification to achieve the best results. This invention uses newly discovered feature values ​​and optimized values ​​from existing feature extraction methods, which can improve the model's classification accuracy.

[0041] Key Point 2: Through the seed account of the social robot, expand the analysis of accounts that follow each other with the robot account, and combine the nickname revision method to identify accounts with similar naming conventions that follow each other with the robot. Then, merge and cluster them according to similar nickname paradigms to form groups of different nickname paradigms. Technical effect: The seed account can detect robot communities with similar nicknames that follow each other, and accumulate a robot nickname paradigm knowledge base for future robot nickname detection and recognition.

[0042] Key Point 3 proposes an optimized graph convolutional clustering community detection algorithm. This algorithm combines label propagation, k-means clustering, text content, network structure, and graph convolutional operations to discover graph communities. Then, it combines the robot communities discovered in the first two key points to identify key robot communities. Technical effect: It can realize graph convolutional clustering using text information and network structure features, and interpret and judge the clustering results by combining the detection results of other social robots, thus realizing another dimension of social robot group detection.

[0043] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0044] This invention proposes a method and apparatus for detecting social robot groups based on a network social platform, comprising the following steps:

[0045] Step 1: Construct strong features with high discriminative power through feature engineering, and combine them with the XGBoost algorithm to form a social robot classifier. Select high-threshold confidence accounts to form seed accounts. For example... Figure 1 As shown, features such as whether the mutual follow rate is abnormal, whether the reposting is abnormal, the cohesion of the posting language, whether the continuous login time in the current month is abnormal, whether the posting IP changes are extremely abnormal, whether the posting time is abnormal, whether the topic promotion is active and abnormal, and whether the reposting target is abnormal, along with the manual account labels, are input into the xgboost model to construct a social robot individual detection classifier. Confidence accounts with high thresholds, such as prediction probabilities greater than 0.9, are set as seed accounts.

[0046] Step Two: Expand from the seed account to accounts that mutually follow it, forming nickname detection groups and paradigms through nickname detection. For example... Figure 2 As shown, accounts that follow each other and seed robot accounts are selected to form a candidate set for nickname detection. Similar nickname pairs are generated using the minHash method. The accounts are aggregated upwards to be classified into a unique similar nickname group and nickname paradigms are extracted. The longest nickname string with the highest frequency of occurrence in the community is selected as the nickname paradigm and accumulated into the nickname knowledge base for subsequent detection. At the same time, the social robot group corresponding to the nickname paradigm is formed.

[0047] Step 3: Based on text content and network structure, an optimized graph convolutional clustering method is proposed for unsupervised community discovery. For example... Figure 3 As shown, firstly, the historical texts posted by the account are vectorized using a fastText pre-trained model. This is then clustered into topic clusters, and cluster center vectors are extracted from large clusters. Next, mean-pooling is used to obtain the initial vector representation of the account. Secondly, a graph network is constructed based on the follower / follower relationship. Label propagation, combined with the user's initial feature vectors, forms the initial unsupervised community partitioning, where the label propagation direction is from followed users to followers. Thirdly, graph convolution is performed on the node vectors within the partitioned community, allowing nodes with graph network links to influence each other and aggregate upwards to update the vector representation. Based on the new vector representation, k-means clustering is performed. The clustering result is measured by maximizing the average cohesion index (cohesion is the average of pairwise similarities between vectors within a cluster). Simultaneously, a penalty term for community partitioning changes is added to the loss function, meaning the magnitude of the change in the k-means clustering result compared to the previous clustering result affects each iteration, while simultaneously stabilizing the clustering results. Finally, the loss function is adjusted in reverse, and the process of dividing the community into graph convolution operations and k-means clustering is repeated to produce the community division and the final vector representation of the community node; the final vector representation of the community node is the object to be adjusted during iteration.

[0048] Step 4: Identify key robot communities in graph convolution clustering based on seed accounts and robot community accounts with similar nicknames. For example... Figure 3 As shown, based on the detection of seed accounts and bot community accounts with similar nicknames, the coverage rate of bot accounts for the communities discovered by graph convolutional clustering has been determined. For communities with a coverage rate of over 80%, new social bot communities are formed, and finally, other accounts in the group are identified as social bot accounts, thus completing the interpretation of unsupervised social bot community detection.

[0049] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0050] This invention also proposes a social robot group detection system based on a network social platform, including:

[0051] The initial module is used to construct strong features for individual identification of social robots, and to identify accounts in the online social platform that match the strong features as seed accounts;

[0052] The community discovery module is used to perform extended analysis based on the social network relationships of the seed account, and to discover and summarize communities with similar nickname paradigms by combining nickname detection, thereby obtaining social robot communities based on the seed account.

[0053] The group detection module is used to discover communities using graph convolutional clustering. It incorporates the graph structure and text signals of the social networking platform into the community detection process. Through the social robot community and seed accounts, it identifies key communities with robot coverage exceeding a threshold, which are then used as the social robot group detection results.

[0054] The aforementioned social robot group detection system based on a network social platform, wherein the strong feature includes:

[0055] Mutual follower rate, reposting frequency, cohesion of posting language, continuous login duration in the current month, frequency of posting IP changes, posting time, topic-driven activity and reposting goals.

[0056] The aforementioned social robot group detection system based on a network social platform includes a community discovery module that performs extended analysis on accounts that mutually follow the seed account, identifies accounts with similar naming conventions that mutually follow the robot using a nickname revision method, and merges and clusters them according to similar nickname paradigms to form groups with different nickname paradigms as social robot communities.

[0057] The aforementioned social robot group detection system based on a network social platform, wherein the group detection module includes:

[0058] This paper presents a method for dividing historical texts posted by various accounts on a social networking platform into vectors. These vectors are then clustered into topic clusters, and cluster center vectors are extracted from each cluster to obtain the initial vector representation of the account. A graph network is constructed based on the follower and follower relationships of each account on the social networking platform. Initial unsupervised community partitioning is achieved through label propagation combined with the user's initial feature vectors, resulting in partitioned communities. The direction of label propagation is from followed users to followers. Graph convolution is performed on the node vectors within these partitioned communities, allowing nodes with graph network links to influence each other and updating the vector representation upwards. Clustering is then performed based on the new vector representations. The clustering results are measured by maximizing the average cohesion index, while a penalty term for community partitioning changes is added to the loss function. Finally, the loss function is adjusted in reverse, and the process from community partitioning to graph convolution to clustering is repeated until the loss function converges or the number of iterations is reached. The current community partitioning results and the final vector representations of the community nodes are then saved.

[0059] Based on the seed account and the accounts involved in the social robot community, the coverage rate of the current community segmentation results is calculated, and communities with a coverage rate exceeding the threshold are designated as key communities.

[0060] The present invention also proposes a storage medium for storing a program that executes any of the social robot group detection methods based on a network social platform.

[0061] This invention also proposes a client for any social robot group detection system based on a network social platform.

[0062] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many modifications under the guidance of the present invention without departing from the spirit of the present invention, and these modifications all fall within the protection scope of the present invention.

Claims

1. A method for detecting a social robot swarm based on a network social platform, characterized in that, The method comprises the following steps: Step 1, constructing a strong feature for social robot individual identification, and identifying an account meeting the strong feature in a network social platform as a seed account; Step 2, performing expansion analysis based on the social network relationship of the seed account, and combining nickname detection to discover and summarize a similar nickname paradigm community, to obtain a social robot community based on the seed account; Step 3, adopting a graph convolution clustering community discovery method, integrating the graph structure and text signal of the network social platform into community detection, and determining a key community with a robot coverage rate greater than a threshold value as a social robot group detection result through the social robot community and the seed account; The step 2 comprises: performing expansion analysis on accounts mutually followed with the seed account, combining a nickname revision method to identify accounts with similar naming specifications and mutually followed with robots, and merging and clustering according to similar nickname paradigms to form different nickname paradigm groups as social robot communities; The step 3 comprises: Performing vector conversion on historical text published by each account in the network social platform, and forming a topic cluster by clustering the vectors, extracting a cluster center vector of the topic cluster, and obtaining an initial vector expression of the account; constructing a graph network based on the follow and followed relationship of each account in the network social platform, forming an initial unsupervised community division by label propagation combined with the initial feature vector of the user, and obtaining a divided community, wherein the direction of the label propagation is from the followed to the follower; performing graph convolution operation on the node vectors in the divided community, so that the nodes in the community that exist in the graph network link influence each other, and the vector representation is updated upward; based on the new vector representation, clustering is performed, and the clustering result is measured by maximizing the average cohesion degree index, and a community division change penalty term is added to the loss function; finally, the loss function is adjusted in reverse, and the process of dividing the community to the graph convolution operation to the clustering is repeated until the loss function converges or the iteration number is reached, and the current community division result and the final vector expression of the community node are saved; According to the seed account and the social robot community involved accounts, the coverage rate of the current community division result is counted, and the community with a coverage rate greater than a threshold value is taken as the key community.

2. The network social platform based social bot swarm detection method of claim 1, wherein, The strong feature comprises: The mutual liking rate, the forwarding frequency, the text language cohesion degree, the continuous login time length in the month, the text ip change frequency, the text time, the topic boosting activity and the forwarding target.

3. A social bot swarm detection system based on a network social platform, characterized in that, The method comprises the following steps: An initial module is configured to construct a strong feature for social robot individual identification, and identify an account meeting the strong feature in a network social platform as a seed account; A community discovery module is configured to perform expansion analysis based on the social network relationship of the seed account, and combine nickname detection to discover and summarize a similar nickname paradigm community, to obtain a social robot community based on the seed account; A group detection module is configured to adopt a graph convolution clustering community discovery method, integrate the graph structure and text signal of the network social platform into community detection, and determine a key community with a robot coverage rate greater than a threshold value as a social robot group detection result through the social robot community and the seed account. The community discovery module includes: performing an expansion analysis on the accounts that follow the seed account, identifying accounts with similar naming conventions and following the robot in combination with the nickname revision method, and merging and clustering according to similar nickname norms to form different nickname norm groups as social robot communities; The group detection module includes: The historical text published by each account in the network social platform is converted into a vector, and a topic cluster is formed by clustering the vector, and the cluster center vector of the cluster is extracted to obtain the initial vector representation of the account; a graph network is constructed based on the follow and followed relationship of each account in the network social platform, and an initial unsupervised community division is formed by label propagation combined with the initial feature vector of the user to obtain the divided community, wherein the direction of label propagation is from the follower to the follower; the node vector in the divided community is subjected to graph convolution operation, so that the nodes in the community that exist in the graph network link influence each other, and the vector representation is updated upwardly; clustering is performed based on the new vector representation, and the clustering result is measured by maximizing the average cohesion index, and the loss function increases the community division change penalty term; finally, the loss function is adjusted in the reverse direction, and the process of dividing the community to the graph convolution operation to the clustering is repeated until the loss function converges or the iteration number is reached, and the current community division result and the final vector representation of the community node are saved; According to the seed account and the social robot community involved accounts, the coverage rate of the current community division result is calculated, and the community with a coverage rate exceeding a threshold is taken as the key community.

4. The network social platform based social bot swarm detection system of claim 3, wherein, The strong features include: The mutual favor rate, the forwarding frequency, the text message cohesion degree, the continuous login time in the month, the text message IP change frequency, the text message time, the topic boost activity and the forwarding target.

5. A storage medium for storing a program for executing the network social platform based social robot group detection method according to claim 1 or 2.

6. A client for the network social platform based social robot group detection system according to claim 3 or 4.

Images