Method, device, equipment and program product for identifying cyberbullying of teenagers

Abstract

The application relates to the technical field of Internet platforms, in particular to a youth cyberbullying identification method, device, equipment and program product. The application generates an object interaction graph by taking the interaction data and the content risk value associated with the to-be-identified youth object in the identification window as the center, and generates a historical normal relationship subgraph according to historical normal relationship data; determines the topological anomaly quantitative value of the newly added node or edge relative to the subgraph, and defines the corresponding node as a candidate pseudo-abnormal relationship node when the associated content risk value is less than a direct disposal threshold and the topological anomaly meets a preset condition; generates a virtual object neighborhood abnormal center node and a connection edge that aggregate each candidate node according to the topological anomaly quantitative value and the connection path of the youth object node, and obtains an enhanced object graph; and generates an identification result containing a network bullying risk state according to the graph. The application can more accurately and timely identify the network bullying behavior against the youth.

Description

Technical Field

[0001] This application relates to the field of Internet platform technology, and in particular to a method, device, equipment and program product for identifying cyberbullying among teenagers. Background Technology

[0002] With the development of short video, text-based communities, social media feeds, and instant messaging services, teenagers are increasingly posting, browsing, and interacting on online platforms. Platforms typically handle inappropriate content in comments, replies, reposts, or public posts through content moderation systems, user reporting systems, and manual review mechanisms. These systems generally collect text, images, audio, video, account behavior, reporting records, and handling records, and review content instances or account behavior according to platform rules.

[0003] In existing methods for identifying cyberbullying or handling content security, common practices of platforms include: classifying individual comments or replies by text, performing content recognition on images or videos, detecting abnormal account behavior, and determining review priorities based on the number of reports, interaction frequency, number of reposts, number of comments, blacklisting records, or historical violation records.

[0004] However, existing technologies still have some problems, making it impossible to more accurately and promptly identify cyberbullying targeting teenagers. Summary of the Invention

[0005] To address the aforementioned technical problems, or at least partially address them, this application provides a method for identifying cyberbullying against minors, which can more accurately and promptly identify cyberbullying behaviors targeting minors.

[0006] Firstly, this application provides a method for identifying cyberbullying among teenagers, the method being applied to a social media platform server, the method comprising: Obtain the interaction data associated with the adolescent to be identified within the current identification window, and the content risk value corresponding to the content instance in the interaction data; Based on the interaction data, an object interaction graph is generated with the adolescent to be identified as the center. The object interaction graph includes nodes of the adolescent to be identified, content instance nodes, interaction account nodes, and edges representing interaction behaviors. Based on the historical normal relationship data of the adolescents to be identified, a historical normal relationship subgraph is generated; Based on the object interaction graph and the historical normal relationship subgraph, determine the topological anomaly quantification value of the newly added node or newly added edge relative to the historical normal relationship subgraph within the current identification window; If the content risk value associated with the newly added node or edge is less than the direct handling threshold, and the topological anomaly quantification value of the newly added node or edge meets the preset anomaly conditions, the relation node corresponding to the newly added node or edge will be determined as a candidate pseudo-anomaly relation node. Based on the topological anomaly quantization value of the candidate pseudo-anomaly relationship node and the connection path between the candidate pseudo-anomaly relationship node and the adolescent object node to be identified, a virtual object neighborhood anomaly center node for aggregating the candidate pseudo-anomaly relationship node is generated in the object interaction graph, and a connection edge between the candidate pseudo-anomaly relationship node and the virtual object neighborhood anomaly center node is generated to obtain an enhanced object graph. Based on the enhanced object graph, a cyberbullying identification result for the adolescent to be identified is generated, and the cyberbullying identification result includes the cyberbullying risk status.

[0007] Optionally, generating the adolescent cyberbullying identification result corresponding to the adolescent to be identified based on the enhanced object graph includes: The cyberbullying risk score of the adolescent to be identified is determined based on the enhanced object map, and the cyberbullying risk score is compared with a judgment threshold to determine the cyberbullying risk status. The method further includes: The edge predictor is used to predict the subsequent edges to be added to the candidate pseudo-anomaly relation nodes, and the subsequent edges are added to the enhanced object graph to obtain the extended object graph; The extended object graph is reconstructed using a graph autoencoder to obtain the reconstruction confidence of the subsequent edges. Subsequent edges with reconstruction confidence less than the pruning threshold are pruned, while subsequent edges with reconstruction confidence greater than or equal to the pruning threshold are retained to obtain a robust subsequent object graph. The cyberbullying risk score corresponding to the robust subsequent object graph is determined according to the method of determining the cyberbullying risk score based on the enhanced object graph, thereby obtaining the cyberbullying risk tendency; An adjustment amount for the determination threshold is generated based on the cyberbullying risk tendency, and the determination threshold is adjusted using the adjustment amount. Based on the cyberbullying risk score corresponding to the enhanced object graph and the adjusted judgment threshold, the cyberbullying risk status of the adolescent to be identified is determined, and if it is determined that the adolescent to be identified has a cyberbullying risk, a manual review of the adolescent to be identified is triggered.

[0008] Optionally, the step of generating a threshold adjustment amount based on the cyberbullying risk propensity includes: If the cyberbullying risk tendency is higher than the first tendency threshold, a judgment threshold adjustment amount is generated to lower the judgment threshold, so that the adolescent subject to be identified is more likely to be judged as having a cyberbullying risk compared to before the judgment threshold was lowered. If the cyberbullying risk tendency is lower than the second tendency threshold, a judgment threshold adjustment amount is generated to raise the judgment threshold, so that the adolescent subject to be identified is less likely to be judged as having a cyberbullying risk compared to before the judgment threshold was raised. The second tendency threshold is less than the first tendency threshold.

[0009] Optionally, the edge prediction algorithm is trained through the following steps: The edge predictor is constructed, which includes an edge predictor input layer, a local neighborhood extraction layer, a graph convolutional coding layer, an edge breaking layer, and an edge predictor output layer. The edge prediction input layer is used to receive the enhanced object graph; The local neighborhood extraction layer is connected to the edge predictor input layer to obtain the augmented object graph. Taking the candidate pseudo-anomaly relationship node as the center, the local neighborhood subgraph of the candidate pseudo-anomaly relationship node within a preset number of hops in the augmented object graph is extracted. The graph convolutional coding layer is connected to the local neighborhood extraction layer to obtain the local neighborhood subgraph. The graph convolutional coding layer includes multiple graph convolution modules, each of which includes a graph convolution unit and a neighborhood aggregation unit. This allows for the aggregation of the structural and node features of the local neighborhood subgraph through multiple layers of graph convolution, resulting in a local node representation that characterizes the local connectivity tendency of the candidate pseudo-anomaly relation nodes. and the node representation of each node in the local neighborhood subgraph. ; Connecting edges to the graph convolutional coding layer to obtain local node representations and node representation of each node Based on local node representation Node representation of each node The similarity between them is used to determine the connection probability between the candidate pseudo-anomaly relationship node and each node in the local neighborhood subgraph, and the connection with a connection probability greater than or equal to the prediction threshold is determined as the newly added subsequent edge of the candidate pseudo-anomaly relationship node. The edge prediction output layer is connected to the edge breaking layer to receive the subsequent edges and outputs the newly added subsequent edges of the candidate pseudo-anomaly relation nodes. The augmented object graphs corresponding to multiple candidate pseudo-abnormal relation nodes and the actual newly added subsequent edges of the candidate pseudo-abnormal relation nodes are collected as training sets. The augmented object graphs are used as inputs to the edge predictor, and the actual newly added subsequent edges are used as training labels to train the edge predictor.

[0010] Optionally, the graph autoencoder is trained through the following steps: The graph autoencoder is constructed, which includes a graph autoencoder input layer, a graph convolutional encoder, a latent representation layer, an inner product decoder, and a graph autoencoder output layer. The graph autoencoder input layer is used to receive the extended object graph; The graph convolutional encoder is connected to the input layer of the graph autoencoder to obtain the extended object graph. The graph convolutional encoder includes multiple graph convolutional modules to aggregate the structural and node features of the object neighborhood of the adolescent object to be identified in the extended object graph through multi-layer graph convolution, thereby obtaining the hidden representation of each node that characterizes the global connectivity consistency of the object neighborhood. ; The hidden representation layer is connected to the graph convolutional encoder to obtain the hidden representation of each node. ; The inner product decoder connects to the hidden representation layer to obtain the hidden representation of each node. The subsequent edge is reconstructed based on the inner product between the implicit representations of the nodes at both ends of the subsequent edge to obtain the reconstruction confidence of the subsequent edge. The graph autoencoder output layer is connected to the inner product decoder to receive the reconstruction confidence and output the reconstruction confidence of the subsequent edges. Multiple extended object graphs corresponding to the neighborhoods of objects are collected as a training set. The extended object graphs are used as the input of the graph autoencoder. The reconstructing of existing edges in the extended object graphs is used as the training objective. The graph autoencoder is trained using a first loss function, where the first loss function is the reconstruction error between existing edges in the extended object graphs and the reconstructed edges.

[0011] Optionally, determining the topological anomaly quantification value of a newly added node or edge relative to the historical normal relationship subgraph within the current identification window, based on the object interaction graph and the historical normal relationship subgraph, includes: The newly added edge is quantized to obtain the relation node corresponding to the newly added edge, and the newly added node and the relation node are determined as nodes to be quantized; Based on the connection distance between the node to be quantized and the historical normal relationship subgraph, and the degree to which the node to be quantized lacks stable relationship support in the object interaction graph, the deviation score of the node to be quantized from the normal support structure is determined. The object pointing score of the node to be quantified to the adolescent object is determined based on the number or length of paths from the node to be quantified to the node to be identified via the content instance node or the interactive account node. The collaboration score of the node to be quantified is determined based on the degree to which the node to be quantified and other nodes to be quantified point to the same node of the adolescent to be identified within the current identification window, and the degree of time concentration of the node to be quantified entering the current identification window. The deviation score, the object pointing score, and the collaboration score are weighted and summed to obtain the topological anomaly quantization value of the node to be quantized, and the topological anomaly quantization value of the node to be quantized is determined as the topological anomaly quantization value of the corresponding newly added node or newly added edge.

[0012] Optionally, the step of generating a virtual object neighborhood anomaly center node for aggregating the candidate pseudo-anomaly relationship nodes in the object interaction graph based on the topological anomaly quantization value of the candidate pseudo-anomaly relationship nodes and the connection path between the candidate pseudo-anomaly relationship nodes and the adolescent object node to be identified, and generating connecting edges to obtain an enhanced object graph, includes: Based on the topological anomaly quantization value of each candidate pseudo-anomaly relation node, the first connection weight between each candidate pseudo-anomaly relation node and the anomaly center node of the virtual object's neighborhood is determined. According to the first connection weight, the node features of each candidate pseudo-anomaly relation node are weighted and aggregated to obtain the center node features of the anomaly center node in the neighborhood of the virtual object. Based on the number or length of paths between each candidate pseudo-anomaly relationship node and the adolescent object node to be identified, and the first connection weight corresponding to each candidate pseudo-anomaly relationship node, the second connection weight between the virtual object neighborhood anomaly center node and the adolescent object node to be identified is determined. A virtual object neighborhood anomaly center node with the central node characteristics is added to the object interaction graph. Connection edges between the candidate pseudo-anomaly relationship node and the virtual object neighborhood anomaly center node are generated according to the first connection weight. Connection edges between the virtual object neighborhood anomaly center node and the adolescent object node to be identified are generated according to the second connection weight, thus obtaining an enhanced object graph.

[0013] Secondly, this application provides a device for identifying cyberbullying among teenagers, the device comprising at least one module, the at least one module being used to perform any of the cyberbullying identification methods for teenagers described in the first aspect.

[0014] Thirdly, this application provides a computer device including a processor for executing a computer program stored in a memory to implement the method for identifying cyberbullying of minors as described in any of the first aspects.

[0015] Fourthly, this application provides a computer program product containing instructions that, when executed by a computer device, cause the computer device to perform the youth cyberbullying identification method as described in any of the first aspects.

[0016] The technical solution provided in this application has the following advantages compared with the prior art: One of its beneficial effects and its working principle is as follows: On social media platforms (such as short video, text and image communities, and instant messaging-related content platforms), cyberbullying among teenagers may not initially manifest as widespread insults and threats. Its early manifestations often involve multiple low-intensity interactions consistently targeting the same teenager. Before bullying develops, early participating accounts typically initiate targeted interactions to allow subsequent unrelated accounts to identify the teenager and learn about content that can be used for attacks. Early participating accounts often reply to pinned comments or mention nicknames and avatars. They may also reference old content or mention peer relationships. These interactions draw the attention of other accounts to the teenager. Because the risk of a single piece of content is low, it usually doesn't reach the threshold for direct intervention. Even when generating an object interaction graph centered on the teenager to be identified, these low-intensity new relationships are still calculated using corresponding nodes or edges. Between the new relationships and the teenager's object node, there are also content instance nodes or interacting account nodes. Each new relationship, as a dispersed weak connection, updates the representation of the teenager's object node separately. This representation is easily diluted by existing normal interactions and fails to reflect the abnormal structure of multiple new relationships pointing to the same object.

[0017] This application acquires interaction data within the current identification window and obtains the corresponding content risk value on the social platform server side. The server generates an object interaction graph centered on the adolescent subject. This graph includes adolescent subject nodes, content instance nodes, interactive account nodes, and interactive behavior edges. The server also generates a historical normal relationship subgraph based on historical normal relationship data. By comparing the two graph structures, the server identifies newly added nodes or edges within the current identification window and calculates their topological anomaly quantification value. When the content risk value corresponding to a newly added node or edge is lower than the direct handling threshold, the server identifies the corresponding relationship node as a candidate pseudo-anomaly relationship node according to the topological anomaly quantification value. This allows for the screening of weak relationships originating from historical normal structures from the interaction growth formed by low-risk content.

[0018] The server adds a virtual object neighborhood anomaly center node to the object interaction graph. The server aggregates all candidate pseudo-anomaly relationship nodes within the current identification window to this center node. All candidate nodes point to the same adolescent object, thus forming an enhanced object graph. Corresponding connecting edges are set between candidate nodes and the center node. The weight of the connecting edge is determined by the topological anomaly quantification value of the candidate node and the connection path between it and the adolescent object node. In this way, candidate nodes no longer act as scattered weak connections, but form a computable structure around the center node. This structure is used to represent two types of information: one is the degree of anomaly of the candidate node relative to the historical normal relationship structure; the other is whether the candidate node points to the same adolescent object via a connecting path.

[0019] Increased exposure through general content dissemination, friend sharing, or short-term exposure can also generate new nodes or edges, but the relationships mostly revolve around content instances. The weights connecting these nodes to the central node are relatively dispersed, and the paths between them and the adolescent target nodes are not concentrated. Multiple candidate nodes in early bullying, while having low risk for individual content, repeatedly point to the same adolescent target. These candidate nodes exhibit topological anomalies relative to the historical normal relationship subgraph, and when aggregated, they form a weighted connection structure around the central node. When the server determines the cyberbullying risk score based on the enhanced object graph, it reads the above connection structure and can identify the bullying risk surrounding the adolescent target from the interaction growth.

[0020] Therefore, this application is able to more accurately and promptly identify cyberbullying against teenagers.

[0021] Its second beneficial effect and its working principle are as follows: In the early stages of the spread of content related to teenagers, both general content dissemination and the early stages of cyberbullying may see an increase in new accounts and a strengthening of target audiences, while the corresponding content risk remains relatively low. Based solely on interactions within the current identification window, the platform cannot determine whether subsequent new edges will shift towards content instances or continue to form around the teenager. The risk score determined by the enhanced target graph relies on the already appearing new nodes, edges, and related connection paths. Only by waiting for subsequent edges to accumulate and for the two types of relationships to show different trends can the platform reliably distinguish between them. Subsequent edges in general dissemination gradually shift towards content instances, while subsequent edges in the early stages of bullying continue to form around the teenager. Waiting for more data will cause risk verification to lag behind the formation of the bullying relationship.

[0022] To address the aforementioned lag, this application no longer waits for the actual accumulation of subsequent edges. The server utilizes an edge predictor to predict the most likely subsequent edges based on the local connectivity tendencies of candidate pseudo-anomaly relation nodes. The predicted subsequent edges are added to the enhanced object graph, forming an extended object graph. The server then uses a graph autoencoder to reconstruct the extended object graph based on the global connectivity consistency of the adolescent object's neighborhood. During reconstruction, subsequent edges lacking overall structural support are pruned, while those that can still be reconstructed after perturbation are retained, resulting in a robust subsequent object graph.

[0023] The social media platform server calculates a risk score on the robust subsequent object graph using the same method employed to determine cyberbullying risk scores based on the enhanced object graph, and uses this score as a cyberbullying risk propensity. Based on this risk propensity, the server generates a threshold adjustment amount and adjusts the risk assessment threshold for adolescent subjects. When the risk propensity leans towards being bullied, the server lowers the threshold, making it more likely to trigger manual review; when the risk propensity leans towards content dissemination, the server raises the threshold, reducing the likelihood of being identified as a cyberbully.

[0024] The edge predictor and graph autoencoder are trained on historical interaction data from social platforms, targeting two types of subsequent generation patterns: object-dependent bullying and ordinary content diffusion. The edge predictor provides the most likely subsequent edges for candidate pseudo-abnormal relationship nodes. The graph autoencoder reconstructs the extended object graph. Regardless of whether the subsequent edge points to a content instance or to the teenager, if it is not supported by the overall structure of the object's neighborhood and cannot be reconstructed after perturbation, it will be pruned, retaining only the subsequent edges supported by the structure.

[0025] After the above pruning process, the robust subsequent object graph will exhibit two different structures. For adolescents who already exhibit early bullying structural characteristics, subsequent edges around the individual can support each other and are preserved after reconstruction. The object's neighborhood thus maintains a dense coreference structure, resulting in a higher cyberbullying risk score. Even with limited data, the server will lower the judgment threshold accordingly, triggering manual review of relevant objects earlier. For object attention growth resulting from ordinary content diffusion, the object's neighborhood lacks mutually supporting dense coreferences. Scattered subsequent edges pointing to the individual lack structural support and will be pruned during reconstruction. The pruned robust subsequent object graph will not form a dense structure around the individual. The cyberbullying risk score determined accordingly will be lower, and the judgment threshold will not be arbitrarily lowered. This process does not require waiting for subsequent edges to actually appear, shortening the waiting time needed to stably distinguish between the two types of relationships. Therefore, early bullying continuing around the individual can be processed through manual review earlier. Ordinary content diffusion is also less likely to be misjudged.

[0026] Therefore, this application is able to more accurately and promptly identify cyberbullying against teenagers. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating an application scenario of the cyberbullying identification method for teenagers provided in the embodiments of this application; Figure 2 One of the flowcharts of the method for identifying cyberbullying of teenagers provided in this application embodiment; Figure 3 This is the second flowchart illustrating the method for identifying cyberbullying among teenagers provided in this application. Detailed Implementation

[0028] The technical solutions in this application will now be described with reference to the accompanying drawings.

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0030] Before providing a detailed explanation of the embodiments of this application, let's first introduce the application scenarios involved in the embodiments of this application.

[0031] Figure 1 This is a schematic diagram illustrating an application scenario of the cyberbullying identification method for teenagers provided in this application. For example... Figure 1 As shown, a social platform server is set up in the short video, text and image community, or instant messaging related content platform. The social platform server executes the adolescent cyberbullying identification method described in the following embodiment, processes the interaction data associated with the adolescent to be identified in the current identification window, and outputs the adolescent cyberbullying identification result corresponding to the adolescent to be identified. The adolescent cyberbullying identification result includes the cyberbullying risk status. When the cyberbullying risk status indicates that the adolescent to be identified is at risk of cyberbullying, the social platform server triggers manual review of the adolescent to be identified, so that the platform reviewers can review and handle the adolescent to be identified.

[0032] The cyberbullying identification method for teenagers provided in this application embodiment can be loaded and executed by a cyberbullying identification device for teenagers. The device includes at least one module, which is used to execute the cyberbullying identification method for teenagers described in the following embodiments.

[0033] Figure 2 This is one of the flowcharts illustrating the method for identifying cyberbullying of teenagers provided in this application. Figure 2 As shown: S201: Obtain the interaction data associated with the adolescent to be identified within the current identification window, and the content risk value corresponding to the content instance in the interaction data; Specifically, in this embodiment, the social platform server determines the duration of the recognition window. Define the current identification window and read the interaction data within this window that is associated with the adolescent to be identified from the platform's interaction log. The duration of the identification window is specified. These are configurable parameters, set by the social platform according to the recognition execution cycle and acceptable recognition latency, used to define the time range of recent interaction data evaluated in a single recognition.

[0034] The interactive data includes content instances and their interactive relationships. The content instances are records that carry text or multimedia, such as comments, replies, forwards, private messages, and posts. The interactive relationship records the account that issued the content instance and the object that the content instance points to. The object is read from the structured association field carried by the content instance in the platform's interactive log.

[0035] Specifically, the reference to a comment or reply instance is the content instance itself that was commented on or replied to, i.e., the post being commented on or the comment being replied to; when the comment or reply is sent directly to an account by mentioning that account, the reference is that account. The reference to a forwarded content instance is the original content instance that was forwarded; the reference to a private message content instance is the account that received the private message; and the reference to a content instance that includes a mention also includes the account that was mentioned. All of the above-mentioned fields are recorded by the platform when the content instance is generated. Accordingly, the reference to each interaction is an account or a content instance, where the account includes the account of the adolescent being identified.

[0036] The content risk value is given by the content moderation model deployed on the social platform after evaluating each content instance. This content moderation model is an existing content risk assessment technology based on text classification or multimodal classification. It takes content instances as input and outputs a numerical value representing the degree of violation or attack of the content itself. It is a mature technology in this field, and this step directly calls its output. The content risk value is used as input for subsequent steps along with the interaction data.

[0037] S202: Based on the interaction data, generate an object interaction diagram centered on the adolescent object to be identified; Specifically, in this embodiment of the application, the interactive data described in S201 is mapped to a heterogeneous graph centered on the adolescent object to be identified, namely, an object interaction graph.

[0038] In S201, each interaction relationship consists of an issuing account, a content instance, and a target object. These three elements form a relationship chain: an account generates a content instance, and the content instance points to a target object. This relationship chain is mapped to an object interaction graph: accounts are designated as account nodes, with the account of the adolescent to be identified serving as the central target adolescent node, and the remaining accounts as interaction account nodes; content instances are designated as content instance nodes; an edge is drawn between the issuing account node and the content instance node, indicating that the account generated the content instance; an edge is drawn between the content instance node and the node of its target object, indicating that the content instance points to that object, and the node pointing to the target object can be the target adolescent node, an interaction account node, or another content instance node. Multiple interaction relationships share the same account node, content instance node, and target adolescent node, thus forming an object interaction graph centered on the target adolescent node. Each edge has a direction: the edge where an account generates a content instance points from the account node to the content instance node, and the edge where a content instance points to an object points from the content instance node to the node of its target object. Constructing heterogeneous graphs from interactive data is a mature graph modeling technique in this field. This step completes the construction of the object interaction graph according to the above-mentioned mapping of nodes and edges, which serves as the data structure for subsequent topology analysis.

[0039] S203: Generate a historical normal relationship subgraph based on the historical normal relationship data of the adolescents to be identified; Specifically, in this embodiment, the interaction data of the adolescent to be identified within a historical period prior to the current identification window is taken, and records that have been judged by the platform as cyberbullying or have already been dealt with are removed to obtain historical normal relationship data; the length of the historical period is greater than the identification window duration. The social platform is configured to cover the long-term, stable interactive relationships of the adolescents to be identified. Following the same method as S202, the sending account, content instance, and target object in the historical normal relationship data are mapped to account nodes, content instance nodes, and edges representing the content instance generated by the account and the target object of the content instance, resulting in a historical normal relationship subgraph. The historical normal relationship subgraph and the object interaction graph use the same node and edge mapping and are aligned according to account identifiers, ensuring that the same account corresponds to the same node in both graphs. The historical normal relationship subgraph serves as a reference for the normal relationship structure of the adolescents to be identified, allowing subsequent steps to compare relationship changes relative to this reference within the current identification window. Constructing a subgraph from historical interaction data is a mature graph modeling technique in this field; this step completes the construction of the historical normal relationship subgraph in the manner described above.

[0040] S204: Based on the object interaction graph and the historical normal relationship subgraph, determine the topological anomaly quantification value of the newly added node or edge.

[0041] Specifically, in this embodiment, the object interaction graph and the historical normal relationship subgraph are aligned by account identifier and compared: nodes that exist in the object interaction graph but do not have corresponding nodes in the historical normal relationship subgraph are identified as newly added nodes; edges that exist in the object interaction graph but do not have corresponding edges in the historical normal relationship subgraph are identified as newly added edges. Newly added nodes and edges represent changes in relationships relative to the historical normal relationship structure within the current identification window. This step calculates a topological anomaly quantification value for each newly added node and each newly added edge to measure the degree to which the relationship change deviates from the normal relationship structure.

[0042] The topological anomaly quantification value is a weighted composite of three components: deviation score, object-directed score, and co-occurrence score. The deviation score describes the degree to which the relationship change deviates from the historical normal structure topologically. The object-directed score describes the degree to which it points to the adolescent object to be identified. The co-occurrence score describes the degree to which multiple relationship changes co-occur and point to the same object.

[0043] The three scores are determined as follows: Specifically, in the object interaction graph, each newly added edge is nodeized, resulting in a relation node corresponding to each newly added edge. All newly added nodes and all relation nodes are combined to form a set of nodes to be quantized, where each node is called a node to be quantized. After being nodeized, newly added edges participate in the measurement in the same form as newly added nodes. Both newly added nodes and newly added edges are calculated using the same set of three scores as nodes to be quantized.

[0044] Each node to be quantized in the set of nodes to be quantized The deviation score, object pointing score, and cooperative score are calculated in the following manner, and the topological anomaly quantization value of the node to be quantized is synthesized.

[0045] The node for identifying adolescents is denoted as The object interaction diagram is The set of nodes in the historical normal relation subgraph is ; exist The set of adjacent nodes in the array is , The moment of entering the current recognition window Take the generation time of the content instance or edge corresponding to the node to be quantized. The following connection distance, degree of lack of stable support, object pointing score, degree of common pointing and degree of time concentration are normalized to 0-1 according to the value range of all nodes to be quantized in the set of nodes to be quantized before participating in the weighting.

[0046] Deviation score measures the degree to which a node to be quantified deviates topologically from the normal relationship structure of the adolescent being identified, and is composed of two factors: connection distance and the degree of lack of stable support. Relationships in the normal state all fall within or near the historical normal relationship subgraph. The farther a node to be quantified is from the historical normal relationship subgraph, the less it belongs to the existing normal relationship; therefore, the node to be quantified is selected accordingly. arrive The minimum number of hops for the shortest path between all nodes is the connection distance. ,remember for and exist Shortest path hop count: Nodes with normal relationships are often supported by neighboring nodes that also belong to normal structures. The fewer neighboring nodes of a node to be quantified that belong to a historical normal relationship subgraph, the weaker its stability support. Therefore, [the following is taken as an example]. The adjacent nodes do not belong to The proportion of the lack of stable support : Deviation score Take the weighted sum of the two. Non-negative weights: The shortest path calculation and adjacent node selection are both mature graph algorithms in this field.

[0047] The object-pointing score measures the degree of concentration of the node to be quantified in the topology of the target youth object. A path from the node to be quantified through content instance nodes or interactive account nodes to the target youth object node is a pathway for this relationship change to point to the target youth object. The more paths and the shorter they are, the more concentrated the pointing. Based on this, [the score will be determined by...]. Departure, passing only content instance nodes or interactive account nodes along the way, arrival The number of jumps does not exceed The set of paths is denoted as , The maximum number of hops for the preset path; the object points to the score. Take the sum of the reciprocals of the hop counts for each path in the set. For path Number of hops: Bounded depth pathfinding is a mature graph traversal technique in this field.

[0048] The co-occurrence score measures whether the node to be quantified, along with other relationship changes occurring around the same time, points to the adolescent being identified. Multiple relationship changes pointing to the same adolescent being identified, occurring in a cluster within a short period, demonstrate a higher degree of co-occurrence than isolated single relationship changes. Based on this, Before and after entering the current recognition window Within range and exist to Other nodes to be quantized in the connection path constitute cooperative set , The predefined collaborative temporal neighborhood; the degree of common pointing is determined by the number of nodes in the collaborative set. The degree of time concentration is taken The span of the entry time of each node to be quantized The reciprocal, This is the difference between the maximum and minimum values ​​at these entry times. Co-score Take the weighted sum of the two. Non-negative weights To prevent division by zero of positive numbers: Nodes to be quantized Topological anomaly quantization value Take the weighted sum of the three fractions. The weights are non-negative, taking preset values ​​or determined by labeling known normal samples and known bullying samples on historical data: Will This serves as the topological anomaly quantification value for the corresponding newly added node or edge.

[0049] S205: If the content risk value associated with the newly added node or the newly added edge is less than the direct handling threshold, and the topological anomaly quantification value of the newly added node or the newly added edge meets the preset anomaly conditions, the relation node corresponding to the newly added node or the newly added edge is determined as a candidate pseudo-anomaly relation node. Specifically, in this embodiment, for each node to be quantified, its associated content risk value is taken: when the node to be quantified is a relation node corresponding to a newly added edge, the content risk value of the content instance corresponding to that newly added edge is taken; when the node to be quantified is a newly added node and itself is a content instance node, its own content risk value is taken; when the node to be quantified is a newly added node and itself is an account node, the content risk value of the content instance node connected to that account node is taken, and when multiple content instances are associated, the maximum value is taken. The node to be quantified is recorded as follows: The associated content risk value is The topological anomaly quantization value is .

[0050] Content with a risk value reaching the direct handling threshold will be handled directly by the social platform's content review and will not proceed to this step. Nodes with a risk value below the direct handling threshold but a high topological anomaly quantification value are those whose content does not meet the direct handling threshold but whose relationship structure is abnormal. The direct handling threshold is set accordingly. With abnormal threshold When the node to be quantized Simultaneously satisfy When the node to be quantified is determined, it is identified as a candidate pseudo-anomaly relation node.

[0051] Among them, the direct disposal threshold This refers to the current content risk threshold used by social media platforms to trigger automatic content moderation. A content risk value reaching [a certain threshold] is considered a trigger. The content has already been automatically processed by content moderation; this step only handles content with a risk value lower than [a certain threshold]. The nodes to be quantized The value depends on the platform's content review and handling strategy.

[0052] Abnormal threshold To define at what level a topology anomaly quantization value meets a preset anomaly condition, this embodiment uses historical data for identification: Normal samples from historical data that are known not to have experienced cyberbullying are taken; the topology anomaly quantization value for each normal sample corresponding to the node to be quantized is calculated; and the high quantization percentile of these topology anomaly quantization values ​​is taken as... For example, take Quantiles ensure that the proportion of normal samples selected as candidates does not exceed a predetermined proportion; the higher the quantile, the fewer nodes are selected as candidates for quantization. In other implementations, an anomaly threshold... Alternatively, a preset fixed value can be used, or the quantile of the topological anomaly quantization values ​​of all nodes to be quantized within the current identification window can be used, and the nodes to be quantized with the top preset proportion of topological anomaly quantization values ​​can be included as candidates.

[0053] S206: Based on the topological anomaly quantization value of the candidate pseudo-anomaly relationship node and the connection path between the candidate pseudo-anomaly relationship node and the adolescent object node to be identified, generate a virtual object neighborhood anomaly center node for aggregating the candidate pseudo-anomaly relationship node in the object interaction graph, and generate the connection edge between the candidate pseudo-anomaly relationship node and the virtual object neighborhood anomaly center node to obtain an enhanced object graph. Specifically, this step adds a virtual object neighborhood anomaly center node to the object interaction graph, aggregates the scattered candidate pseudo-anomaly relationship nodes to this center node, and then connects this center node to the adolescent object node to be identified. This ensures that anomalies pointing to the adolescent object are converged on a single node and then connected to the adolescent object node to be identified. Let the nodes in the candidate pseudo-anomaly relationship node set be denoted as _____. Its topological anomaly quantization value is Node characteristics are The abnormal center node of the virtual object's neighborhood is The node to be identified is the adolescent. .

[0054] First connection weight. The higher the topological anomaly quantization value of a candidate pseudo-anomaly relation node, the greater its proportion in the object's neighborhood anomaly. Based on this, each candidate pseudo-anomaly relation node is determined according to its topological anomaly quantization value. With the central node First connection weight between And normalize within the candidate set: Central node features. The node features of each candidate pseudo-anomaly relation node are weighted and aggregated according to the first connection weight to obtain the central node. Central node features : The second connection weight. The more concentrated the candidate pseudo-anomaly relationship nodes point to the adolescent to be identified, and the larger the first connection weight, the stronger the association between the central node and the adolescent to be identified after aggregation. Based on this, the central node is determined according to the number or length of paths between each candidate pseudo-anomaly relationship node and the adolescent to be identified node, as well as the first connection weight. Nodes of adolescents to be identified The second connection weight between Select candidate pseudo-anomaly relation nodes. The object pointing score determined in the manner described in S204 is its path contribution. ,but Construct an enhanced object graph. Add a central node feature to the object interaction graph. The abnormal center node of the virtual object neighborhood For each candidate pseudo-anomaly relation node According to the first connection weight generate and The connecting edges between them; according to the second connection weight generate Nodes of adolescents to be identified The connecting edges between these nodes and edges. The resulting object interaction graph is called the enhanced object graph.

[0055] S207: Based on the enhanced object map, generate a cyberbullying identification result for the adolescent to be identified, wherein the cyberbullying identification result includes a cyberbullying risk status; Specifically, the enhanced object graph, compared to the object interaction graph, adds a virtual object neighborhood anomaly center node that aggregates candidate pseudo-anomaly relationship nodes, as well as connection edges generated between the center node and the adolescent object node to be identified according to a second connection weight. Dispersed anomalies are aggregated through the center node and connected to the adolescent object node to be identified. This step reads the representation of the adolescent object node to be identified from the enhanced object graph, and generates an adolescent cyberbullying identification result that includes the cyberbullying risk status.

[0056] Specifically, in this embodiment of the application, the cyberbullying risk score of the adolescent to be identified is determined based on the enhanced object map, and the cyberbullying risk score is compared with the judgment threshold to determine the cyberbullying risk status.

[0057] Each node's node feature is its attribute representation in the object interaction graph. For example, an account node takes its interaction statistics within the current recognition window, and a content instance node takes the intermediate representation between its content risk value and the content review model output. The construction of node features is a conventional method in this field; the node features of the aforementioned candidate pseudo-anomaly relationship nodes... Features of the central node All of these are derived from this.

[0058] Multi-layer message passing is performed on the node representations of the augmented object graph using graph convolution to read out the representations of the nodes of the adolescent objects to be identified. The initial representation of each node is taken from its node features, and the initial representation of the abnormal center node in the neighborhood of the virtual object is taken from the features of its center node. Record nodes In the The layer is represented as , No. Layer to the first The message passing mechanism of the layer is as follows: in, For nodes In adjacent nodes of the augmented object graph, For the edge The weights are calculated as follows: the edge between the candidate pseudo-anomaly node and the center node takes the first connection weight; the edge between the center node and the node of the adolescent to be identified takes the second connection weight; and the remaining edges take the preset weights. For the first Layer transformation parameters. Number of message passing layers. Choosing 3 allows the nodes of the adolescent objects to be identified to be aggregated to each candidate pseudo-anomaly relation node and its connecting edges. After layer message passing, retrieve the node of the adolescent object to be identified. The representation of Cyberbullying risk score: For the value to fall within arrive The sigmoid function between , These are the scoring parameters. Transformation parameters for each layer. With scoring parameters , The training process yields the following: Samples of adolescents identified as cyberbullies or otherwise normal are taken from historical records. An augmented object map is constructed according to steps S201 to S206, and the results are read from the augmented object map. The system calculates a cyberbullying risk score and uses the binary cross-entropy between the cyberbullying risk score and the sample label as the loss. Gradient descent is then used to optimize the transformation parameters and scoring parameters of each layer.

[0059] Set the decision threshold ,when The current cyberbullying risk status is "there is a risk of cyberbullying". There is no risk of cyberbullying at this time. Judgment threshold. Labeling based on historical data: Bullying samples already labeled as cyberbullying and normal samples already labeled as normal are taken from historical data. Cyberbullying risk scores are calculated for each sample according to steps S201 to S207. At candidate thresholds, the proportion of normal samples judged to have a cyberbullying risk and the proportion of bullying samples judged to have a cyberbullying risk are calculated separately. The candidate threshold that results in the highest value for the latter, provided the former does not exceed a preset upper limit, is taken as the final judgment threshold. The higher the threshold, the fewer teenagers will be identified as having a risk of cyberbullying. The cyberbullying identification result for teenagers includes the aforementioned cyberbullying risk status.

[0060] Figure 3 This is a second flowchart illustrating the method for identifying cyberbullying among teenagers provided in this application embodiment. Figure 3 As shown: S301: Use the edge predictor to predict the new subsequent edges for the candidate pseudo-anomaly relation nodes, and add the subsequent edges to the enhanced object graph to obtain the extended object graph.

[0061] Specifically, in this step, the enhanced object graph obtained in S206 is input into the edge predictor. The edge predictor predicts the most likely new subsequent edge for each candidate pseudo-anomaly relation node in the enhanced object graph, using each candidate pseudo-anomaly relation node as the center, and outputs the result. The subsequent edges output by the edge predictor are then added to the enhanced object graph, resulting in an extended object graph that simultaneously contains both currently occurring relationships and predicted subsequent relationships. The structure, processing of each layer, and training method of the edge predictor are as follows.

[0062] Specifically, in the embodiments of this application, an edge predictor is constructed, which includes an edge predictor input layer, a local neighborhood extraction layer, a graph convolutional coding layer, an edge breaking layer, and an edge predictor output layer.

[0063] The edge predictor input layer obtains the augmented object graph.

[0064] The local neighborhood extraction layer uses candidate pseudo-anomaly relationship nodes as the center and extracts the number of hops of the candidate pseudo-anomaly relationship node in the augmented object graph. The local neighborhood subgraph within.

[0065] The graph convolutional coding layer comprises multiple graph convolutional modules. Each module includes a graph convolutional unit and a neighborhood aggregation unit. The graph convolutional unit performs a linear transformation on the node features, and the neighborhood aggregation unit performs a weighted average aggregation of the transformed representations of adjacent nodes according to their connection weights. After multiple layers of graph convolutional aggregation, the structure and node features of the local neighborhood subgraph are combined to obtain a local node representation that characterizes the local connectivity tendency of candidate pseudo-anomaly relation nodes. And the node representation of each node in the local neighborhood subgraph. .

[0066] Layering of edges based on local nodes Node representation of each node The similarity between nodes determines the connection probability between candidate pseudo-anomaly nodes and nodes in the local neighborhood subgraph; the nodes in the local neighborhood subgraph are denoted by their similarity. The node is represented as Candidate pseudo-anomaly relation nodes and nodes Connection probability between: For the value to fall within arrive The sigmoid function between; connects with a probability greater than or equal to the prediction threshold. The connection is determined as the newly added subsequent edge of the candidate pseudo-anomaly relation node.

[0067] The edge prediction layer outputs the newly added subsequent edges of the candidate pseudo-anomaly relation nodes.

[0068] The parameters of the edge prediction algorithm are obtained by training it using the following steps.

[0069] Constructing training samples: Take multiple observation times from historical data, and for each observation time, construct an augmented object graph of the adolescent object to be identified according to S201 to S206 as the input of the edge prediction machine; for each candidate pseudo-anomaly relation node in the augmented object graph... Nodes in its local neighborhood subgraph Within a preset time period after the observation time and Whether an actual new edge is added as a label between them The actual increase is recorded as No new additions are recorded as ;Retrieve all tags The node pairs are used as positive samples, and the labels are... The nodes are based on the relative positive samples Compare The proportional sampling is used as negative samples, and the positive samples and the sampled negative samples constitute the training pair set. Calculate the connection probability for each training pair. , Candidate pseudo-anomaly relation nodes The local node representation, with binary cross-entropy as the loss: Using the Adam optimizer and learning rate The parameters of each layer of the edge predictor are iteratively optimized using mini-batch gradient descent until the loss converges.

[0070] In this embodiment, the graph convolutional coding layer of the edge predictor has two graph convolutional modules, the inter-layer activation function is ReLU, and the output dimensions of the two graph convolutional modules are 128 and 64 respectively; the local neighborhood extraction layer has a preset number of hops. Take 2; Local node representation Node representation of each node The 64-dimensional representation of the corresponding node is obtained from the graph convolutional coding layer; the similarity of the edges is calculated using the inner product, and the prediction threshold is... Take 0.5.

[0071] S302: Reconstruct the extended object graph using a graph autoencoder, and prune or retain subsequent edges according to the comparison between the reconstruction confidence and the pruning threshold to obtain a robust subsequent object graph; Specifically, in this step, the extended object graph obtained in S301 is input into the graph autoencoder, which performs global reconstruction on the extended object graph and outputs the reconstruction confidence of each subsequent edge in the extended object graph. The reconstruction confidence represents the degree to which the subsequent edge is supported by the overall structure of the object neighborhood of the adolescent object to be identified; a reconstruction confidence greater than or equal to the pruning threshold is selected. Subsequent edges are retained, and the value is less than the pruning threshold. The subsequent edges are pruned, and the remaining subsequent edges, together with the enhanced object graph, form a robust subsequent object graph.

[0072] The structure, layer processing, and training methods of the graph autoencoder are as follows: The graph autoencoder includes a graph autoencoder input layer, a graph convolutional encoder, a hidden representation layer, an inner product decoder, and a graph autoencoder output layer.

[0073] The diagram shows the extended object obtained from the encoder input layer.

[0074] The graph convolutional encoder includes multiple graph convolutional modules. Through multiple layers of graph convolution, it aggregates the structural and node features of the object neighborhood of the adolescent object to be identified in the extended object graph, and obtains the hidden representation of each node that represents the global connectivity consistency of the object neighborhood. .

[0075] The hidden representation layer obtains the hidden representation of each node. .

[0076] The inner product decoder reconstructs the node pair based on the inner product between the implicit representations of the node and its two endpoints, obtaining the reconstruction confidence of the node pair; let the implicit representation of the node pair be denoted as . , The reconstruction confidence of this node pair For the value to fall within arrive The sigmoid function between.

[0077] The reconstruction confidence of each subsequent edge output by the encoder output layer.

[0078] The parameters of the graph autoencoder are obtained by training using the following steps: Multiple observation times from historical data are taken, and for each observation time, an extended object graph of the adolescent to be identified is constructed according to S201 to S301 as the input of the graph autoencoder. The existing edges in the extended object graph are used as positive samples, and unconnected node pairs are sampled according to a preset ratio relative to the positive samples as negative samples. The positive samples and the sampled negative samples constitute the training pair set. The forward graph convolutional encoder obtains the hidden representation of each node. The reconstruction confidence of each training pair is then obtained through the inner product decoder. The first loss function is the reconstruction error between the existing edges in the expanded object graph and the reconstructed edges, using the binary cross-entropy of the training pairs: in, For node pairs The label indicating whether an edge already exists; if it does, denoted as . Not connected is recorded as The Adam optimizer and learning rate are used. The parameters of each layer of the graph autoencoder are iteratively optimized using mini-batch gradient descent until the loss converges. After training, the graph autoencoder provides high reconstruction confidence for edges that support the overall structure of the object's neighborhood.

[0079] In this embodiment, the graph autoencoder's graph convolutional encoder has two graph convolutional modules, both employing convolution operations of a graph convolutional network and using connection weights as adjacency weights. The first graph convolutional module has an output dimension of 64 and uses ReLU as the activation function. The implicit representation... The dimension is set to 32; the inner product decoder obtains the reconstruction confidence by taking the inner product of the implicit representations of the two endpoints and then applying it to the sigmoid function; pruning threshold. Take 0.5.

[0080] S303: Determine the cyberbullying risk score corresponding to the robust subsequent object graph according to the method of determining the cyberbullying risk score based on the enhanced object graph, and obtain the cyberbullying risk tendency; Specifically, in this step, based on the robust follow-up object graph obtained in S302, the representation of the adolescent object node to be identified and the cyberbullying risk score are read from the enhanced object graph in the same way as in S207, i.e., using the same multi-layer graph convolutional message passing, the same transformation parameters and scoring parameters, to read the representation of the adolescent object node to be identified in the robust follow-up object graph. Calculate the cyberbullying risk score: The cyberbullying risk score is used as a cyberbullying risk propensity.

[0081] The robust follow-up object graph, based on the enhanced object graph, retains follow-up edges supported by the overall structure of the object's neighborhood. These follow-up edges, through message passing, change the representation of the adolescent object node to be identified. This allows the cyberbullying risk propensity to reflect the level of cyberbullying risk among the adolescents to be identified after these subsequent incidents.

[0082] S304: Generate a threshold adjustment amount based on the cyberbullying risk tendency, and adjust the threshold based on the threshold adjustment amount; Specifically, this step generates an adjustment amount for the judgment threshold based on the risk propensity for cyberbullying. This adjustment amount is then added to the original judgment threshold to obtain the adjusted threshold. The adjustment amount is a scalar; its sign determines whether the judgment threshold is lowered or raised, and its magnitude determines the adjustment range.

[0083] If the cyberbullying risk tendency is higher than the first tendency threshold, a threshold adjustment amount is generated to lower the threshold, making it easier for the adolescent to be identified to be judged as having a cyberbullying risk compared to before the threshold was lowered; if the cyberbullying risk tendency is lower than the second tendency threshold, a threshold adjustment amount is generated to raise the threshold, making it less likely for the adolescent to be identified to be judged as having a cyberbullying risk compared to before the threshold was raised; the second tendency threshold is less than the first tendency threshold.

[0084] Record the risk propensity for cyberbullying as The first tendency threshold is The second tendency threshold is and The threshold adjustment amount is The adjustment coefficient is a preset positive number. .exist hour, A negative value lowers the judgment threshold; hour, If the value is positive, the judgment threshold is increased; hour, The threshold for determining cyberbullying risk remains unchanged; the greater the deviation of the risk propensity from the threshold, the larger the adjustment. Adjusted threshold: S305: Based on the cyberbullying risk score corresponding to the enhanced object graph and the adjusted judgment threshold, determine the cyberbullying risk status of the adolescent to be identified, and if it is determined that there is a cyberbullying risk, trigger manual review of the adolescent to be identified; Specifically, this step will assign a cyberbullying risk score to the enhanced object graph obtained in S207. With the adjusted judgment threshold Comparison: In At that time, the cyberbullying risk status was "there is a risk of cyberbullying"; in At that time, the cyberbullying risk status was "no cyberbullying risk".

[0085] The final cyberbullying risk status is determined by the cyberbullying risk score corresponding to the augmented object graph and the adjusted judgment threshold. The cyberbullying risk tendency is changed by adjusting the judgment threshold, thereby changing the difficulty of making a cyberbullying judgment.

[0086] If it is determined that a minor is at risk of cyberbullying, a manual review of the minor will be triggered, and the platform's reviewers will verify the minor's relevant interactions.

[0087] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0088] In the various embodiments of this application, unless otherwise specified or logically conflicting, the terminology and / or descriptions between different embodiments are consistent and can be referenced mutually. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. In the textual description of the embodiments of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. In this application, "first," "second," and various numerical designations are only for ease of description and are not used to limit the scope of the embodiments of this application. For example, they are used to distinguish different messages, rather than to describe a specific order or sequence.

[0089] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application. The order of the process numbers does not imply the order of execution; the execution order of each process should be determined by its function and internal logic.

[0090] Finally, it should be noted that the above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying cyberbullying among teenagers, characterized in that, The method is applied to a social platform server, and the method includes: Obtain the interaction data associated with the adolescent to be identified within the current identification window, and the content risk value corresponding to the content instance in the interaction data; Based on the interaction data, an object interaction graph is generated with the adolescent to be identified as the center. The object interaction graph includes nodes of the adolescent to be identified, content instance nodes, interaction account nodes, and edges representing interaction behaviors. Based on the historical normal relationship data of the adolescents to be identified, a historical normal relationship subgraph is generated; Based on the object interaction graph and the historical normal relationship subgraph, determine the topological anomaly quantification value of the newly added node or newly added edge relative to the historical normal relationship subgraph within the current identification window; If the content risk value associated with the newly added node or edge is less than the direct handling threshold, and the topological anomaly quantification value of the newly added node or edge meets the preset anomaly conditions, the relation node corresponding to the newly added node or edge will be determined as a candidate pseudo-anomaly relation node. Based on the topological anomaly quantization value of the candidate pseudo-anomaly relationship node and the connection path between the candidate pseudo-anomaly relationship node and the adolescent object node to be identified, a virtual object neighborhood anomaly center node for aggregating the candidate pseudo-anomaly relationship node is generated in the object interaction graph, and a connection edge between the candidate pseudo-anomaly relationship node and the virtual object neighborhood anomaly center node is generated to obtain an enhanced object graph. Based on the enhanced object graph, a cyberbullying identification result for the adolescent to be identified is generated, and the cyberbullying identification result includes the cyberbullying risk status.

2. The method for identifying cyberbullying among teenagers according to claim 1, characterized in that, The step of generating a cyberbullying identification result for the adolescent to be identified based on the enhanced object map includes: The cyberbullying risk score of the adolescent to be identified is determined based on the enhanced object map, and the cyberbullying risk score is compared with a judgment threshold to determine the cyberbullying risk status. The method further includes: The edge predictor is used to predict the subsequent edges to be added to the candidate pseudo-anomaly relation nodes, and the subsequent edges are added to the enhanced object graph to obtain the extended object graph; The extended object graph is reconstructed using a graph autoencoder to obtain the reconstruction confidence of the subsequent edges. Subsequent edges with reconstruction confidence less than the pruning threshold are pruned, while subsequent edges with reconstruction confidence greater than or equal to the pruning threshold are retained to obtain a robust subsequent object graph. The cyberbullying risk score corresponding to the robust subsequent object graph is determined according to the method of determining the cyberbullying risk score based on the enhanced object graph, thereby obtaining the cyberbullying risk tendency; An adjustment amount for the determination threshold is generated based on the cyberbullying risk tendency, and the determination threshold is adjusted using the adjustment amount. Based on the cyberbullying risk score corresponding to the enhanced object graph and the adjusted judgment threshold, the cyberbullying risk status of the adolescent to be identified is determined, and if it is determined that the adolescent to be identified has a cyberbullying risk, a manual review of the adolescent to be identified is triggered.

3. The method for identifying cyberbullying among teenagers according to claim 2, characterized in that, The step of generating a threshold adjustment amount based on the cyberbullying risk propensity includes: If the cyberbullying risk tendency is higher than the first tendency threshold, a judgment threshold adjustment amount is generated to lower the judgment threshold, so that the adolescent subject to be identified is more likely to be judged as having a cyberbullying risk compared to before the judgment threshold was lowered. If the cyberbullying risk tendency is lower than the second tendency threshold, a judgment threshold adjustment amount is generated to raise the judgment threshold, so that the adolescent subject to be identified is less likely to be judged as having a cyberbullying risk compared to before the judgment threshold was raised. The second tendency threshold is less than the first tendency threshold.

4. The method for identifying cyberbullying among teenagers according to claim 2, characterized in that, The edge prediction algorithm is trained through the following steps: The edge predictor is constructed, which includes an edge predictor input layer, a local neighborhood extraction layer, a graph convolutional coding layer, an edge breaking layer, and an edge predictor output layer. The edge prediction input layer is used to receive the enhanced object graph; The local neighborhood extraction layer is connected to the edge predictor input layer to obtain the augmented object graph. Taking the candidate pseudo-anomaly relationship node as the center, the local neighborhood subgraph of the candidate pseudo-anomaly relationship node within a preset number of hops in the augmented object graph is extracted. The graph convolutional coding layer is connected to the local neighborhood extraction layer to obtain the local neighborhood subgraph. The graph convolutional coding layer includes multiple graph convolution modules, each of which includes a graph convolution unit and a neighborhood aggregation unit. This allows for the aggregation of the structural and node features of the local neighborhood subgraph through multiple layers of graph convolution, resulting in a local node representation that characterizes the local connectivity tendency of the candidate pseudo-anomaly relation nodes. and the node representation of each node in the local neighborhood subgraph. ; Connecting edges to the graph convolutional coding layer to obtain local node representations and node representation of each node Based on local node representation Node representation of each node The similarity between them is used to determine the connection probability between the candidate pseudo-anomaly relationship node and each node in the local neighborhood subgraph, and the connection with a connection probability greater than or equal to the prediction threshold is determined as the newly added subsequent edge of the candidate pseudo-anomaly relationship node. The edge prediction output layer is connected to the edge breaking layer to receive the subsequent edges and outputs the newly added subsequent edges of the candidate pseudo-anomaly relation nodes. The augmented object graphs corresponding to multiple candidate pseudo-abnormal relation nodes and the actual newly added subsequent edges of the candidate pseudo-abnormal relation nodes are collected as training sets. The augmented object graphs are used as inputs to the edge predictor, and the actual newly added subsequent edges are used as training labels to train the edge predictor.

5. The method for identifying cyberbullying among teenagers according to claim 2, characterized in that, The graph autoencoder is trained through the following steps: The graph autoencoder is constructed, which includes a graph autoencoder input layer, a graph convolutional encoder, a latent representation layer, an inner product decoder, and a graph autoencoder output layer. The graph autoencoder input layer is used to receive the extended object graph; The graph convolutional encoder is connected to the input layer of the graph autoencoder to obtain the extended object graph. The graph convolutional encoder includes multiple graph convolutional modules to aggregate the structural and node features of the object neighborhood of the adolescent object to be identified in the extended object graph through multi-layer graph convolution, thereby obtaining the hidden representation of each node that characterizes the global connectivity consistency of the object neighborhood. ; The hidden representation layer is connected to the graph convolutional encoder to obtain the hidden representation of each node. ; The inner product decoder connects to the hidden representation layer to obtain the hidden representation of each node. The subsequent edge is reconstructed based on the inner product between the implicit representations of the nodes at both ends of the subsequent edge to obtain the reconstruction confidence of the subsequent edge. The graph autoencoder output layer is connected to the inner product decoder to receive the reconstruction confidence and output the reconstruction confidence of the subsequent edges. Multiple extended object graphs corresponding to the neighborhoods of objects are collected as a training set. The extended object graphs are used as the input of the graph autoencoder. The reconstructing of existing edges in the extended object graphs is used as the training objective. The graph autoencoder is trained using a first loss function, where the first loss function is the reconstruction error between existing edges in the extended object graphs and the reconstructed edges.

6. The method for identifying cyberbullying among teenagers according to claim 1, characterized in that, The step of determining the topological anomaly quantification value of newly added nodes or edges relative to the historical normal relationship subgraph within the current identification window, based on the object interaction graph and the historical normal relationship subgraph, includes: The newly added edge is quantized to obtain the relation node corresponding to the newly added edge, and the newly added node and the relation node are determined as nodes to be quantized; Based on the connection distance between the node to be quantized and the historical normal relationship subgraph, and the degree to which the node to be quantized lacks stable relationship support in the object interaction graph, the deviation score of the node to be quantized from the normal support structure is determined. The object pointing score of the node to be quantified to the adolescent object is determined based on the number or length of paths from the node to be quantified to the node to be identified via the content instance node or the interactive account node. The collaboration score of the node to be quantified is determined based on the degree to which the node to be quantified and other nodes to be quantified point to the same node of the adolescent to be identified within the current identification window, and the degree of time concentration of the node to be quantified entering the current identification window. The deviation score, the object pointing score, and the collaboration score are weighted and summed to obtain the topological anomaly quantization value of the node to be quantized, and the topological anomaly quantization value of the node to be quantized is determined as the topological anomaly quantization value of the corresponding newly added node or newly added edge.

7. The method for identifying cyberbullying among teenagers according to claim 1, characterized in that, The step of generating a virtual object neighborhood anomaly center node for aggregating the candidate pseudo-anomaly relationship nodes in the object interaction graph based on the topological anomaly quantization value of the candidate pseudo-anomaly relationship nodes and the connection path between the candidate pseudo-anomaly relationship nodes and the adolescent object node to be identified, and generating connection edges to obtain an enhanced object graph, includes: Based on the topological anomaly quantization value of each candidate pseudo-anomaly relation node, the first connection weight between each candidate pseudo-anomaly relation node and the anomaly center node of the virtual object's neighborhood is determined. According to the first connection weight, the node features of each candidate pseudo-anomaly relation node are weighted and aggregated to obtain the center node features of the anomaly center node in the neighborhood of the virtual object. Based on the number or length of paths between each candidate pseudo-anomaly relationship node and the adolescent object node to be identified, and the first connection weight corresponding to each candidate pseudo-anomaly relationship node, the second connection weight between the virtual object neighborhood anomaly center node and the adolescent object node to be identified is determined. A virtual object neighborhood anomaly center node with the central node characteristics is added to the object interaction graph. Connection edges between the candidate pseudo-anomaly relationship node and the virtual object neighborhood anomaly center node are generated according to the first connection weight. Connection edges between the virtual object neighborhood anomaly center node and the adolescent object node to be identified are generated according to the second connection weight, thus obtaining an enhanced object graph.

8. A device for identifying cyberbullying among teenagers, characterized in that, The device includes at least one module for performing the adolescent cyberbullying identification method according to any one of claims 1-7.

9. A computer device, characterized in that, The computer device includes a processor for executing a computer program stored in a memory to implement the method for identifying cyberbullying of minors as described in any one of claims 1-7.

10. A computer program product containing instructions, characterized in that, When the instruction is executed by a computer device, the computer device performs the method for identifying cyberbullying of minors as described in any one of claims 1-7.

Images