Account obtaining method and device, computer device, and storage medium

Abstract

The present disclosure relates to an account obtaining method and device, computer equipment and a storage medium, and belongs to the technical field of computers. The method comprises: determining a first account that has performed a positive interaction behavior on a sensitive content item based on the sensitive content item published by a reference account; determining a candidate account that publishes a target content item based on the target content item on which the first account has performed the positive interaction behavior; generating an account relationship graph associated with the positive interaction behavior based on the reference account, the first account and the candidate account; and screening a target account that constitutes an account group with the reference account from the candidate account based on the account relationship graph. The present disclosure introduces the sensitive content item and the reference account as a supervision signal, and performs community mining under the action of the supervision signal, which can avoid mining account groups that are contrary to business significance, thereby greatly improving the identification accuracy of potential target accounts that have published sensitive content.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to an account acquisition method, apparatus, computer device, and storage medium. Background Technology

[0002] With the development of computer technology and the diversification of terminal functions, users can browse content published by creators (such as streamers and content creators) on platforms anytime and anywhere. However, some creators may also publish sensitive content on the platform. Therefore, how to accurately identify and crack down on accounts that publish sensitive content has gradually become an urgent problem for platforms to solve in the context of security risk control.

[0003] Currently, when identifying accounts that post sensitive content, unsupervised community mining algorithms are typically used, such as the Fast Unfolding algorithm (based on modularity) and the Label Propagation Algorithm (LPA). These algorithms attempt to identify communities (i.e., account groups) formed by accounts that frequently post sensitive content. However, in security risk control scenarios, these community mining algorithms may identify accounts that frequently post homogeneous content as a community, even though the accounts within these communities have never posted sensitive content. In other words, community mining algorithms are prone to outputting communities that contradict business logic, resulting in poor accuracy in identifying accounts that post sensitive content. Summary of the Invention

[0004] This disclosure provides a method, apparatus, computer device, and storage medium for obtaining accounts, to at least improve the accuracy of identifying accounts that publish sensitive content. The technical solution of this disclosure is as follows:

[0005] According to one aspect of the embodiments of this disclosure, an account acquisition method is provided, including:

[0006] Based on the sensitive content items published by the reference account, the first account that has performed positive interaction behavior on the sensitive content items is identified. The positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content items.

[0007] Based on the target content item for which the first account has performed the positive interactive behavior, candidate accounts for publishing the target content item are determined;

[0008] Based on the reference account, the first account, and the candidate account, an account relationship graph associated with the positive interaction behavior is generated. The account relationship graph is used to represent the topological structure of social relationships between publisher accounts of content items for which the first account has performed the positive interaction behavior.

[0009] Based on the account relationship diagram, target accounts that form an account group with the reference account are selected from the candidate accounts.

[0010] In some embodiments, generating an account relationship graph associated with the positive interaction behavior based on the reference account, the first account, and the candidate account includes:

[0011] Based on the reference account and the candidate account, construct the nodes in the account relationship graph;

[0012] Based on the positive interaction behavior performed by the first account, edges are constructed in the account relationship graph to connect nodes.

[0013] In some embodiments, constructing the edges in the account relationship graph for connecting nodes based on the positive interaction behavior performed by the first account includes:

[0014] If either the first account has performed the positive interaction behavior on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account.

[0015] In some embodiments, the step of filtering target accounts that form an account group with the reference accounts from the candidate accounts based on the account relationship diagram includes:

[0016] Obtain the maximum spanning tree of the account relationship graph, wherein the maximum spanning tree has the maximum weight among the multiple spanning trees of the account relationship graph;

[0017] Cluster the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups;

[0018] The multiple candidate account groups are merged to obtain the target account group;

[0019] The candidate accounts contained in the target account group are identified as the target accounts.

[0020] In some embodiments, obtaining the maximum spanning tree of the account relationship graph includes:

[0021] Based on the node similarity between the two nodes connected by each edge in the account relationship graph, the edges are assigned weights.

[0022] Based on the weights of each edge in the account relationship graph, the maximum spanning tree is generated, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the plurality of spanning trees.

[0023] In some embodiments, the node similarity refers to the ratio between the number of common neighbor nodes of the two nodes and the sum of the number of each node's individual neighbor nodes.

[0024] In some embodiments, generating the maximum spanning tree based on the weights of each edge in the account relationship graph includes:

[0025] Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is an empty set.

[0026] From the account relationship graph, obtain multiple candidate edges for connecting nodes within the node set and nodes outside the node set;

[0027] Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the nodes outside the node set connected by the target edge to the node set;

[0028] Repeat the operation of adding target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and the node set when adding stops is determined as the maximum spanning tree.

[0029] In some embodiments, clustering the accounts indicated by the nodes in the maximum spanning tree to obtain multiple candidate account groups includes:

[0030] Multiple core nodes are obtained by filtering from the nodes contained in the maximum spanning tree;

[0031] Multiple candidate account groups are determined by using the multiple core nodes as cluster centers;

[0032] Cluster all nodes in the maximum spanning tree except the core node into the candidate account group containing the core node with the highest path similarity to the core node.

[0033] In some embodiments, the step of selecting multiple core nodes from the nodes contained in the maximum spanning tree includes:

[0034] For any node contained in the maximum spanning tree, determine at least one neighbor node in the maximum spanning tree that is connected to the node by an edge;

[0035] Among the at least one neighboring node, target neighboring nodes whose weight of the connection edge between the node and the neighboring node is greater than or equal to a first weight threshold are selected;

[0036] If the sum of the weights of the connecting edges between the node and each of the target neighbor nodes is greater than the second weight threshold, the node is determined as the core node.

[0037] In some embodiments, the method further includes:

[0038] For any node in the maximum spanning tree other than the core node, determine the connected path between the node and any core node in the maximum spanning tree. The connected path refers to the path that can reach the core node from the node through multiple edges.

[0039] The path similarity between the node and the core node is determined based on the weight of each edge contained in the connected path.

[0040] In some embodiments, merging the plurality of candidate account groups to obtain the target account group includes:

[0041] For any candidate account group among the plurality of candidate account groups, if there is a connection edge between the core node of the candidate account group and the core node of other candidate account groups, and the weight of the connection edge is greater than a first weight threshold, the candidate account group and the other candidate account groups are merged.

[0042] Otherwise, the candidate account group will be merged with other candidate account groups that have the highest group similarity.

[0043] Repeat the operation of merging candidate account groups until no candidate account groups can be merged, and then select the target account group from the account groups obtained from each merge.

[0044] In some embodiments, the method further includes:

[0045] For any candidate account group among the plurality of candidate account groups, the sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group, wherein the degree parameter of the node represents the number of edges connected to the node in the maximum spanning tree;

[0046] Based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account groups, and the number of common edges between the candidate account group and the other candidate account groups, the group similarity between the candidate account group and the other candidate account groups is determined.

[0047] In some embodiments, the step of filtering the target account group from the account groups obtained from each merge includes:

[0048] For each merged account group, the modularity of the account group is obtained. The modularity is used to measure the quality of the division of the account group from the account relationship graph.

[0049] The group of accounts with the highest modularity is identified as the target account group.

[0050] In some embodiments, the positive interaction behavior includes at least one of the following: expressing a positive opinion about the content item, following the publisher's account of the content item, sharing, or downloading.

[0051] In some embodiments, after determining the candidate accounts for publishing the target content item, the method further includes:

[0052] Delete verified candidate accounts and candidate accounts registered by preset organizations.

[0053] According to another aspect of the embodiments of this disclosure, an account acquisition device is provided, comprising:

[0054] The determining unit is configured to execute sensitive content items published by reference accounts, and determine the first account that has performed positive interaction behavior on the sensitive content items, wherein the positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content items;

[0055] The determining unit is further configured to execute target content items based on the positive interactive behavior performed by the first account, and determine candidate accounts that publish the target content items;

[0056] The generation unit is configured to generate an account relationship graph associated with the positive interaction behavior based on the reference account, the first account, and the candidate account. The account relationship graph is used to characterize the topological structure of social relationships between publisher accounts of content items for which the first account has performed the positive interaction behavior.

[0057] The filtering unit is configured to perform a filtering operation based on the account relationship graph to select target accounts that form an account group with the reference account from the candidate accounts.

[0058] In some embodiments, the generation unit includes:

[0059] The node construction subunit is configured to perform the task of constructing nodes in the account relationship graph based on the reference account and the candidate account;

[0060] The edge constructs a sub-unit, which is configured to perform positive interactive behaviors based on the first account, and constructs edges in the account relationship graph to connect nodes.

[0061] In some embodiments, the edge-building subunit is configured to perform:

[0062] If either the first account has performed the positive interaction behavior on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account.

[0063] In some embodiments, the filtering unit includes:

[0064] The acquisition subunit is configured to acquire the maximum spanning tree of the account relationship graph, wherein the maximum spanning tree has the maximum weight among the multiple spanning trees of the account relationship graph;

[0065] The clustering subunit is configured to perform clustering on the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups;

[0066] The merging subunit is configured to merge the multiple candidate account groups to obtain the target account group;

[0067] The determination subunit is configured to perform the action of determining the candidate accounts contained in the target account group as the target account.

[0068] In some embodiments, the acquisition subunit includes:

[0069] The assignment sub-unit is configured to assign weights to the edges based on the node similarity between the two nodes connected by each edge in the account relationship graph;

[0070] The sub-unit is configured to generate the maximum spanning tree based on the weights of each edge in the account relationship graph, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the plurality of spanning trees.

[0071] In some embodiments, the node similarity refers to the ratio between the number of common neighbor nodes of the two nodes and the sum of the number of each node's individual neighbor nodes.

[0072] In some embodiments, the generating subunit is configured to perform:

[0073] Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is an empty set.

[0074] From the account relationship graph, obtain multiple candidate edges for connecting nodes within the node set and nodes outside the node set;

[0075] Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the nodes outside the node set connected by the target edge to the node set;

[0076] Repeat the operation of adding target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and the node set when adding stops is determined as the maximum spanning tree.

[0077] In some embodiments, the clustering subunit includes:

[0078] The filtering sub-unit is configured to perform filtering to obtain multiple core nodes from the nodes contained in the maximum spanning tree;

[0079] The sub-units are determined and configured to perform the task of identifying multiple candidate account groups, with the multiple core nodes serving as cluster centers.

[0080] The clustering sub-unit is configured to perform the task of clustering nodes in the maximum spanning tree, excluding the core node, into the candidate account group where the core node has the highest path similarity to the node.

[0081] In some embodiments, the filtering subunit is configured to perform:

[0082] For any node contained in the maximum spanning tree, determine at least one neighbor node in the maximum spanning tree that is connected to the node by an edge;

[0083] Among the at least one neighboring node, target neighboring nodes whose weight of the connection edge between the node and the neighboring node is greater than or equal to a first weight threshold are selected;

[0084] If the sum of the weights of the connecting edges between the node and each of the target neighbor nodes is greater than the second weight threshold, the node is determined as the core node.

[0085] In some embodiments, the determination of the sub-unit is further configured to perform:

[0086] For any node in the maximum spanning tree other than the core node, determine the connected path between the node and any core node in the maximum spanning tree. The connected path refers to the path that can reach the core node from the node through multiple edges.

[0087] The path similarity between the node and the core node is determined based on the weight of each edge contained in the connected path.

[0088] In some embodiments, the merging subunit includes:

[0089] The merging sub-unit is configured to perform the following operation: for any candidate account group among the plurality of candidate account groups, if there is a connection edge between the core node of the candidate account group and the core node of other candidate account groups, and the weight of the connection edge is greater than a first weight threshold, the candidate account group is merged with the other candidate account groups.

[0090] The merging sub-unit is also configured to otherwise merge the candidate account group with other candidate account groups with the highest group similarity;

[0091] The merging sub-unit is also configured to repeatedly perform the merging operation on candidate account groups until no candidate account groups can be merged.

[0092] The group filtering sub-unit is configured to perform filtering to obtain the target account group from the account groups obtained from each merge.

[0093] In some embodiments, the merging subunit is further configured to perform:

[0094] For any candidate account group among the plurality of candidate account groups, the sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group, wherein the degree parameter of the node represents the number of edges connected to the node in the maximum spanning tree;

[0095] Based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account groups, and the number of common edges between the candidate account group and the other candidate account groups, the group similarity between the candidate account group and the other candidate account groups is determined.

[0096] In some embodiments, the group filtering subunit is configured to perform:

[0097] For each merged account group, the modularity of the account group is obtained. The modularity is used to measure the quality of the division of the account group from the account relationship graph.

[0098] The group of accounts with the highest modularity is identified as the target account group.

[0099] In some embodiments, the positive interaction behavior includes at least one of the following: expressing a positive opinion about the content item, following the publisher's account of the content item, sharing, or downloading.

[0100] In some embodiments, the apparatus further includes:

[0101] The deletion unit is configured to delete authenticated candidate accounts as well as candidate accounts registered with a preset organization.

[0102] According to another aspect of the embodiments of this disclosure, a computer device is provided, comprising:

[0103] One or more processors;

[0104] One or more memories for storing the one or more processor-executable instructions;

[0105] The one or more processors are configured to perform the account acquisition method in any of the possible implementations of the above-described aspects.

[0106] According to another aspect of the present disclosure, a computer-readable storage medium is provided such that, when at least one instruction in the computer-readable storage medium is executed by one or more processors of a computer device, the computer device is enabled to perform an account acquisition method in any possible implementation of the above aspect.

[0107] According to another aspect of the present disclosure, a computer program product is provided, including one or more instructions that can be executed by one or more processors of a computer device, enabling the computer device to perform the account acquisition method in any possible implementation of the above aspect.

[0108] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0109] By leveraging the characteristic that the first account has performed positive interaction behavior on different content items, reference accounts that publish sensitive content items and candidate accounts that publish target content items can be linked in the account relationship graph. This allows the construction of an account relationship graph based on the first account that has performed positive interaction behavior on content items published by different accounts. Community mining can then be performed based on this constructed account relationship graph. By introducing sensitive content items and reference accounts as monitoring signals, community mining can avoid uncovering account groups that contradict business objectives, thereby greatly improving the accuracy of identifying potential target accounts that have published sensitive content.

[0110] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0111] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0112] Figure 1 This is a schematic diagram illustrating the implementation environment of an account acquisition method according to an exemplary embodiment;

[0113] Figure 2 This is a flowchart illustrating an account acquisition method according to an exemplary embodiment;

[0114] Figure 3 This is a flowchart illustrating an account acquisition method according to an exemplary embodiment;

[0115] Figure 4 This is a comparative schematic diagram of isomorphic and heteromorphic diagrams provided in an embodiment of this disclosure;

[0116] Figure 5 This is a flowchart of a method for obtaining the maximum spanning tree provided in an embodiment of this disclosure;

[0117] Figure 6 This is a flowchart of obtaining a candidate account group provided in an embodiment of the present disclosure;

[0118] Figure 7 This is a flowchart illustrating a method for merging candidate account groups, as provided in this embodiment of the disclosure.

[0119] Figure 8 This is a schematic diagram illustrating the principle of negative account suppression after filtering out target accounts based on an account relationship diagram, as provided in this embodiment of the disclosure.

[0120] Figure 9 This is a schematic flowchart illustrating the principle of an account acquisition method provided in this embodiment.

[0121] Figure 10 This is a schematic diagram illustrating the effect of an account acquisition method provided in an embodiment of this disclosure;

[0122] Figure 11 This is a logical structure block diagram of an account acquisition device according to an exemplary embodiment;

[0123] Figure 12 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation

[0124] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0125] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0126] The information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals disclosed herein are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the accounts and interactive behaviors performed by the accounts mentioned in this disclosure were obtained with full authorization.

[0127] In some embodiments, the meaning of A and / or B includes three cases: A and B, and A and B.

[0128] The following explains the terminology used in the embodiments of this disclosure.

[0129] Degree parameter: In a topological graph, the degree parameter of each node refers to the number of edges connected to that node in the graph. In other words, the number of edges originating from that node in the graph is the value of the degree parameter of that node.

[0130] Modularity: Modularity, also known as the modularity metric, is a method for measuring the strength of a network's community structure. It is typically represented by the symbol "Q" in community mining algorithms. The value of modularity depends on the community allocation of nodes in the social network (i.e., the community partitioning of the network; in this embodiment, it refers to the partitioning of account groups). Modularity can quantitatively measure the quality of community partitioning. A higher modularity value indicates a stronger community structure and better partitioning quality; conversely, a lower modularity value indicates a weaker community structure and poorer partitioning quality. Therefore, selecting the account group with the highest modularity allows for the optimal partitioning of communities (i.e., account groups) in the social network; this process is also known as searching for the optimal segmentation of the social network.

[0131] With the development of computer technology and the diversification of terminal functions, users can browse content published by creators (such as streamers and content creators) on platforms anytime and anywhere. However, some creators may also publish sensitive content on the platform. Therefore, how to accurately identify and crack down on accounts that publish sensitive content has gradually become an urgent problem for platforms to solve in the context of security risk control.

[0132] Currently, when identifying accounts that post sensitive content, unsupervised community mining algorithms are often used, such as the Fast Unfolding algorithm and the Label Propagation Algorithm (LPA), which mine communities based on modularity, in an attempt to discover the communities (i.e., account groups) formed by accounts that frequently post sensitive content.

[0133] The Fast Unfolding algorithm suffers from several drawbacks. Firstly, during community discovery, it only groups accounts with homogeneous outputs into the same community. However, these communities may contain accounts that have never published sensitive content. This lack of a mechanism to remove invalid nodes hinders community building and can easily lead to the inclusion of many accounts that contradict business objectives, resulting in poor accuracy in identifying accounts that publish sensitive content. Secondly, as an unsupervised community discovery algorithm, Fast Unfolding presents challenges in security risk control scenarios, including evaluating communities, identifying business-meaningful communities, constructing business-meaning edges and nodes, and setting edge weights.

[0134] For the LPA algorithm, labels are propagated in the dataset and assigned to unlabeled nodes. In the initial stage, each node initializes a label. In each iteration, the label with the most tags among the connected nodes is updated to the node's own label. As community labels continue to propagate, closely connected nodes will eventually share a common label. However, the LPA algorithm itself lacks a mechanism to remove noisy or invalid nodes, and configuring a filtering mechanism for node labels is also a technical challenge. Furthermore, the node label update order is random, but obviously, updating more important nodes earlier will accelerate the convergence process. In addition, if there are more than one tag with the most tags among the other connected nodes, the node will randomly select the tag with the most tags as its own label. This randomness may lead to an avalanche effect, where the initial community division errors caused by this randomness are amplified as the labels propagate. Therefore, the LPA algorithm has relatively poor accuracy in identifying accounts that publish sensitive content.

[0135] In view of this, the present disclosure relates to an account acquisition method, specifically a method for mining common consumer communities. In the field of security risk control, the Mining (CCCM) algorithm, given a subset of reference accounts that have posted sensitive content, connects known reference accounts with unknown candidate accounts through common consumers to construct complex communities. From these complex communities, it then mines unknown target accounts that may have posted sensitive content. Common consumers are viewers who have performed the same positive interaction with both the reference accounts' sensitive content and the candidate accounts' target content. This allows the CCCM algorithm to assist platforms in discovering potential security risks. Complex communities can be constructed for various content scenarios (such as videos, live streams, and comments), enabling automated security risk discovery. Furthermore, by introducing known sensitive content and reference accounts as supervisory signals, this supervised CCCM algorithm, compared to unsupervised algorithms, avoids discovering communities (i.e., account groups) that only generate homogeneous content but are not actually sensitive. This prevents the output of communities that contradict the business implications of security risks, resulting in higher accuracy in identifying accounts that have posted sensitive content.

[0136] The system architecture involved in the embodiments of this disclosure will be explained below.

[0137] Figure 1 This is a schematic diagram illustrating an implementation environment for an account acquisition method according to an exemplary embodiment. See also... Figure 1 The implementation environment may include a terminal 101 and a server 102, both of which are computer devices, as detailed below.

[0138] Terminal 101 can be any computer device that supports content item browsing services. An application for browsing content items is installed on terminal 101. Content items refer to information items that carry certain content in the form of multimedia resources. For example, the form of content items includes, but is not limited to, audio, video, live broadcast, short video, news, comments, etc. For example, the application can include at least one of short video applications, live broadcast applications, audio and video applications, or social applications.

[0139] In some embodiments, after registering an account in the application, a user can log in to that account within the application. Then, the user can publish their own video works on the platform through that account, or they can start a live stream after being certified as a broadcaster.

[0140] In some embodiments, after a user logs into the account, the server 102 can push some content items to the account (e.g., in the form of video stream, feed stream, information stream, etc.). When the user is interested in the pushed content items, they can click to browse the content items and use the account to initiate interactive behavior on the content items. Taking a video work published by another person as an example, the interactive behavior can be divided into positive interactive behavior and negative interactive behavior. Positive interactive behavior refers to liking the video work, following the account of the publisher of the video work, sharing the video work, downloading the video work, etc., while negative interactive behavior refers to reporting the video work or other negative behavior.

[0141] Terminal 101 and server 102 can be connected via wired or wireless network.

[0142] Server 102 can be a computer device used to provide background services for the aforementioned applications. Server 102 may include at least one of a single server, multiple servers, a cloud computing platform, or a virtualization center. Optionally, server 102 may undertake the primary computing task, and terminal 101 may undertake the secondary computing task; or, server 102 may undertake the secondary computing task, and terminal 101 may undertake the primary computing task; or, terminal 101 and server 102 may collaborate on computing using a distributed computing architecture.

[0143] In some embodiments, when the server 102 performs security risk control on the platform, it typically obtains information about known and discovered reference accounts that frequently post sensitive content by filtering user reports and manually screening content items. For example, these reference accounts may refer to malicious accounts, negative accounts, or abnormal accounts. The account acquisition method provided in this disclosure embodiment can, based on known reference accounts and known sensitive content items, use the CCCM algorithm to mine potential target accounts that form account groups with the reference accounts on the platform, and can precisely suppress these target accounts to achieve the purpose of automatically suppressing abnormal accounts and negative videos. Compared to methods that simply assess the content of each account's posts as sensitive and then target those accounts, this method offers higher accuracy in account identification. Machine-based sensitive content identification is prone to errors and overlooks the connections between content creators. This method, however, utilizes a community mining algorithm based on shared consumers. By constructing nodes and edges in the account relationship graph based on shared consumers, and ensuring these nodes and edges have business significance, the overall community mining process achieves higher accuracy. It identifies the inherent connections between creators of sensitive content, facilitating the discovery and suppression of more abnormal accounts. Furthermore, it allows for batch processing of abnormal accounts, establishing complex relationship networks using identified reference accounts to uncover more potential and unknown target accounts. This results in precise community risk discovery and benefits risk management and improved business performance.

[0144] Optionally, terminal 101 may refer to one of a plurality of terminals, and the device type of terminal 101 includes, but is not limited to, at least one of the following: in-vehicle terminal, television, smartphone, smart speaker, tablet computer, e-book reader, MP3 (Moving Picture Experts Group Audio Layer III) player, MP4 (Moving Picture Experts Group Audio Layer IV) player, laptop computer, or desktop computer. The following embodiments use a smartphone as an example to illustrate terminal 101.

[0145] Those skilled in the art will understand that the number of terminals 101 described above can be more or less. For example, there may be only one terminal 101, or there may be dozens or hundreds of terminals 101, or even more. This disclosure does not limit the number or type of terminals 101.

[0146] Figure 2 This is a flowchart illustrating an account acquisition method according to an exemplary embodiment, see [link to flowchart]. Figure 2 This method of obtaining accounts applies to computer devices. The following explanation uses a computer device as a server as an example.

[0147] In step 201, the server determines the first account that has performed positive interaction behavior on the sensitive content item published by the reference account. The positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content item.

[0148] The reference account involved in this disclosure refers to an account that has been marked, is known, and has been discovered to have previously published sensitive content. For example, a reference account refers to a malicious account, a negative account, or an abnormal account. Usually, after a platform discovers and verifies that an account is a reference account, it will take certain measures to suppress the reference account and the sensitive content published by the reference account, such as banning the reference account or deleting the sensitive content.

[0149] The sensitive content items involved in the embodiments of this disclosure refer to content items that contain sensitive information or imply sensitive information. A content item refers to an information item that takes the form of a multimedia resource and carries specific content. For example, a content item can be a video, audio, live broadcast, news, comments, etc. The embodiments of this disclosure do not specifically limit this.

[0150] The interactive behaviors involved in this disclosure refer to human-computer interaction behaviors used to view sensitive content items. Optionally, the interactive behaviors include positive interactive behaviors and negative interactive behaviors. Positive interactive behaviors refer to interactive behaviors used to express a positive opinion on sensitive content items, and negative interactive behaviors refer to interactive behaviors used to express a negative opinion on sensitive content items. For example, positive interactive behaviors include at least one of the following: expressing a positive opinion on a content item, following the publisher's account of a content item, sharing, or downloading, etc., wherein the positive opinion behavior includes, but is not limited to, liking, rewarding, or collecting a content item. For example, negative interactive behaviors include at least one of the following: reporting a sensitive content item, complaining about a sensitive content item, reporting the publisher's account of a content item, or complaining about the publisher's account of a content item, etc. This disclosure does not specifically limit these behaviors.

[0151] In some embodiments, the server can add sensitive tags to accounts that have been discovered, known, or detected as having published sensitive content, and identify accounts with sensitive tags in the database as reference accounts. For example, the server adds sensitive tags to accounts that have been reported and verified to have published sensitive content, or the server adds sensitive tags to accounts that have been banned within a target time period and have published sensitive content. This disclosure does not specifically limit the method of obtaining reference accounts. The target time period is a period set by technical personnel, such as the most recent week or the most recent month; this disclosure does not specifically limit this period.

[0152] In some embodiments, after obtaining a reference account, if the sensitive content items published by the reference account have already been marked, the marked sensitive content items can be obtained directly. Alternatively, if the content items published by the reference account have not been marked, the content items published by the reference account can be detected to determine whether the current content item is a sensitive content item, thereby obtaining the detected sensitive content item.

[0153] In some embodiments, after the server obtains the reference accounts marked in the platform and the marked sensitive content items published by the reference accounts, it then determines the accounts that have performed positive interaction behavior on the sensitive content items as the first accounts. Optionally, since there may be more than one type of positive interaction behavior, the server can perform the method provided in this disclosure embodiment on each type of positive interaction behavior, so as to identify different first accounts for different positive interaction behaviors, so as to discover more potential target accounts that have published sensitive content items from multiple dimensions of positive interaction behavior.

[0154] In step 202, the server determines candidate accounts to publish the target content item based on the target content item for which the first account has performed the positive interaction behavior.

[0155] The target content items involved in this embodiment refer to content items other than sensitive content items for which the first account has performed positive interaction behavior. For example, when the positive interaction behavior is a positive expression of opinion on the content item, such positive expression behavior includes, but is not limited to, liking, rewarding, or collecting the content item. Taking liking as an example, after obtaining the reference account and the sensitive content item published by the reference account, the first account that liked the sensitive content item is identified. Then, other content items liked by the first account are obtained as target content items. Since these target content items may or may not be sensitive content items, the candidate accounts that publish the target content items may or may not be target accounts, and need to be screened and mined through subsequent steps.

[0156] In some embodiments, after the server obtains the first account through step 201 above, it can determine the target content item for which the first account has performed a positive interactive behavior. For example, it can query the target content item for which the positive interactive behavior has been performed from the first account's historical browsing history of the content item, or it can maintain a historical interaction list for each content item for which the first account has performed the positive interactive behavior, and determine all content items on the historical interaction list as target content items. This embodiment of the disclosure does not specifically limit the method of obtaining target content items.

[0157] Optionally, when the technician sets a statistical time period, the server can only obtain the target content items that the first account has performed the positive interaction behavior within the statistical time period. This is because the interests and preferences of the first account may change dynamically at different times. Therefore, obtaining the target content items only for the pre-set statistical time period can realize a pre-screening mechanism for the target content items. For example, the statistical time period is the most recent week or the most recent month, or the statistical time period is the week before and the week after the first account browsed the sensitive content items. This embodiment of the disclosure does not specifically limit the setting method of the statistical time period.

[0158] In some embodiments, after the server determines the target content item, it determines the account that published the target content item, i.e., the publisher account of the target content item, as a candidate account. It should be noted that since the reference account involved in the above step 201 may be one or more, the first account that has performed positive interaction behavior on the sensitive content item published by each reference account may also be one or more, and the target content item that each first account has also performed positive interaction behavior on may also be one or more. Therefore, the candidate account that finally publishes the target content item may also be one or more. This embodiment of the disclosure does not specifically limit this.

[0159] In step 203, the server generates an account relationship graph associated with the positive interaction behavior based on the reference account, the first account, and the candidate account. The account relationship graph is used to represent the topological structure of social relationships between the publisher accounts of the content items for which the first account has performed the positive interaction behavior.

[0160] In some embodiments, the server constructs nodes in the account relationship graph using reference accounts and candidate accounts, and constructs edges in the account relationship graph to connect nodes based on whether the same first account has performed positive interaction behavior on sensitive content items published by the reference account and target content items published by the candidate account. Finally, an account relationship graph formed by each node and each edge can be obtained. For example, the account relationship graph can be represented by G(V,E), where G represents the account relationship graph, V represents the set of nodes in the account relationship graph, and E represents the set of edges in the account relationship graph.

[0161] In step 204, the server, based on the account relationship graph, selects target accounts from the candidate accounts that form an account group with the reference account.

[0162] In some embodiments, after constructing the original account relationship graph, the server can perform community mining based on the CCCM algorithm to find account groups formed by clustering known reference accounts and unknown candidate accounts. The candidate accounts contained in these account groups are the target accounts that can form account groups with the reference accounts.

[0163] In the above process, by leveraging the characteristic that the first account has performed positive interaction behavior on different content items, it is possible to connect the reference accounts that publish sensitive content items and the candidate accounts that publish target content items in the account relationship graph, thereby constructing an account relationship graph based on common consumers (referring to the first account that has performed positive interaction behavior on content items published by different accounts). Then, community mining can be carried out based on this constructed account relationship graph, which helps to improve the accuracy of target account identification and also helps to discover more potential target accounts.

[0164] The method provided in this disclosure leverages the characteristic that a first account has performed positive interactive behaviors on different content items. This allows the reference account that publishes sensitive content items and the candidate account that publishes target content items to be linked in an account relationship graph. This constructs an account relationship graph based on the first account that has performed positive interactive behaviors on content items published by different accounts. Community mining is then performed based on this constructed account relationship graph. By introducing sensitive content items and reference accounts as monitoring signals, community mining can avoid discovering account groups that contradict business significance, thereby greatly improving the accuracy of identifying potential target accounts that have published sensitive content.

[0165] In one possible implementation, generating an account relationship diagram associated with the positive interaction behavior based on the reference account, the first account, and the candidate account includes:

[0166] Based on the reference account and the candidate account, construct the nodes in the account relationship graph;

[0167] Based on the positive interaction behavior performed by the first account, construct the edges in the account relationship graph to connect the nodes.

[0168] In one possible implementation, based on the positive interaction behavior performed by the first account, the edges used to connect nodes in the account relationship graph include:

[0169] If either the first account has performed the positive interaction behavior on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account.

[0170] In one possible implementation, based on the account relationship diagram, target accounts that form an account group with the reference account are selected from the candidate accounts, including:

[0171] Find the maximum spanning tree of the account relationship graph, which has the maximum weight among the multiple spanning trees of the account relationship graph;

[0172] Cluster the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups;

[0173] The multiple candidate account groups are merged to obtain the target account group;

[0174] The candidate accounts included in the target account group are identified as the target account.

[0175] In one possible implementation, obtaining the maximum spanning tree of the account relationship graph includes:

[0176] Based on the node similarity between the two nodes connected by each edge in the account relationship graph, a weight is assigned to the edge;

[0177] Based on the weights of each edge in the account relationship graph, the maximum spanning tree is generated, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the multiple spanning trees.

[0178] In one possible implementation, the node similarity refers to the ratio between the number of common neighbor nodes of the two nodes and the sum of the number of each node's individual neighbor nodes.

[0179] In one possible implementation, generating the maximum spanning tree based on the weights of each edge in the account relationship graph includes:

[0180] Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is empty.

[0181] From the account relationship graph, obtain multiple candidate edges to connect nodes within the node set and nodes outside the node set;

[0182] Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the node outside the node set connected by the target edge to the node set;

[0183] Repeat the operation of adding target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and node set at the point where addition stops is determined as the maximum spanning tree.

[0184] In one possible implementation, the accounts indicated by the nodes contained in the maximum spanning tree are clustered to obtain multiple candidate account groups, including:

[0185] Multiple core nodes are obtained by filtering from the nodes contained in the maximum spanning tree;

[0186] Using these multiple core nodes as cluster centers, multiple candidate account groups are determined;

[0187] Cluster all nodes in the maximum spanning tree except for the core node into the candidate account group containing the core node with the highest path similarity to the core node.

[0188] In one possible implementation, selecting multiple core nodes from the nodes contained in the maximum spanning tree includes:

[0189] For any node contained in the maximum spanning tree, determine at least one neighboring node in the maximum spanning tree that is connected to that node by an edge;

[0190] Among the at least one neighboring node, target neighboring nodes whose weight of the connection edge between the node and the neighboring node is greater than or equal to a first weight threshold are selected;

[0191] If the sum of the weights of the connecting edges between this node and each of its target neighbor nodes is greater than the second weight threshold, then this node is determined as the core node.

[0192] In one possible implementation, the method further includes:

[0193] For any node in the maximum spanning tree other than the core node, determine the connected path between the node and any core node in the maximum spanning tree. The connected path refers to the path that can reach the core node from the node through multiple edges.

[0194] Based on the weight of each edge contained in the connected path, the path similarity between the node and the core node is determined.

[0195] In one possible implementation, the multiple candidate account groups are merged to obtain the target account group, which includes:

[0196] If any candidate account group among the multiple candidate account groups has a connection edge between its core node and the core nodes of other candidate account groups, and the weight of the connection edge is greater than the first weight threshold, then the candidate account group and the other candidate account groups are merged.

[0197] Otherwise, merge the candidate account group with the other candidate account groups with the highest group similarity;

[0198] Repeat the process of merging candidate account groups until no candidate account groups can be merged. Select the target account group from the account groups obtained from each merge.

[0199] In one possible implementation, the method further includes:

[0200] For any candidate account group among the multiple candidate account groups, the sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group. The degree parameter of the node represents the number of edges connected to the node in the maximum spanning tree.

[0201] Based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account groups, and the number of common edges in the candidate account group and the other candidate account groups, the group similarity between the candidate account group and the other candidate account groups is determined.

[0202] In one possible implementation, the process of selecting the target account group from the account groups obtained from each merge includes:

[0203] For each merged account group, obtain the modularity of the account group. The modularity is used to measure the quality of the division of the account group from the account relationship graph.

[0204] The group of accounts with the highest modularity is identified as the target group of accounts.

[0205] In one possible implementation, the positive interaction behavior includes at least one of the following: expressing a positive opinion about the content item, following the publisher's account of the content item, sharing, or downloading.

[0206] In one possible implementation, after identifying candidate accounts to publish the target content item, the method further includes:

[0207] Delete verified candidate accounts and candidate accounts registered by preset organizations.

[0208] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0209] Figure 3 This is a flowchart illustrating an account acquisition method according to an exemplary embodiment, such as... Figure 3 As shown, the account acquisition method is used for computer devices. Taking a computer device as a server as an example, this embodiment includes the following steps.

[0210] In step 301, the server determines the first account that has performed positive interaction behavior on the sensitive content item published by the reference account. The positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content item.

[0211] In some embodiments, the interactive behaviors that an account can initiate with the content items it views include positive interactive behaviors and negative interactive behaviors. Positive interactive behaviors refer to interactive behaviors used to express a positive opinion about the content items, while negative interactive behaviors refer to interactive behaviors used to express a negative opinion about the content items.

[0212] Optionally, positive interaction behaviors include at least one of the following: expressing positive opinions about content items, following the publisher's account of content items, sharing, or downloading, etc., wherein the positive expression of opinions includes, but is not limited to, liking, rewarding, or collecting content items; Optionally, negative interaction behaviors include at least one of the following: reporting content items, complaining about content items, reporting the publisher's account of content items, or complaining about the publisher's account of content items, etc., and this disclosure embodiment does not specifically limit these.

[0213] In some embodiments, the server can add sensitive tags to accounts that have been discovered, known, or detected as having published sensitive content, and identify accounts with sensitive tags in the database as reference accounts. For example, the server adds sensitive tags to accounts that have been reported and verified to have published sensitive content, or the server adds sensitive tags to accounts that have been banned within a target time period and have published sensitive content. This disclosure does not specifically limit the method of obtaining reference accounts. The target time period is a period set by technical personnel, such as the most recent week or the most recent month; this disclosure does not specifically limit this period.

[0214] In some embodiments, after obtaining a reference account, if the sensitive content items published by the reference account have already been marked, the marked sensitive content items can be obtained directly. Alternatively, if the content items published by the reference account have not been marked, the content items published by the reference account can be detected to determine whether the current content item is a sensitive content item, thereby obtaining the detected sensitive content item.

[0215] In some embodiments, after the server obtains the reference accounts marked in the platform and the marked sensitive content items published by the reference accounts, it then identifies the accounts that have performed positive interaction behavior on the sensitive content items as the first accounts. Optionally, since there may be more than one type of positive interaction behavior, the server can execute the method provided in this disclosure embodiment for each type of positive interaction behavior, so as to identify different first accounts for different positive interaction behaviors and construct different account relationship graphs, so as to mine potential target accounts that have published sensitive content items from more dimensions of positive interaction behavior from the account relationship graphs associated with various positive interaction behaviors, thereby improving the coverage of target account identification.

[0216] In step 302, the server determines candidate accounts to publish the target content item based on the target content item for which the first account has performed the positive interaction behavior.

[0217] In some embodiments, after the server obtains the first account through step 301 above, it can determine the target content item for which the first account has performed a positive interactive behavior. For example, it can query the target content item for which the positive interactive behavior has been performed from the first account's historical browsing history of the content item, or it can maintain a historical interaction list for each content item for which the first account has performed the positive interactive behavior, and determine all content items on the historical interaction list as target content items. This embodiment of the disclosure does not specifically limit the method of obtaining target content items.

[0218] Optionally, when the technician sets a statistical time period, the server can only obtain the target content items that the first account has performed the positive interaction behavior within the statistical time period. This is because the interests and preferences of the first account may change dynamically at different times. Therefore, obtaining the target content items only for the pre-set statistical time period can realize a pre-screening mechanism for the target content items. For example, the statistical time period is the most recent week or the most recent month, or the statistical time period is the week before and the week after the first account browsed the sensitive content items. This embodiment of the disclosure does not specifically limit the setting method of the statistical time period.

[0219] In some embodiments, after the server determines the target content item, it determines the account that published the target content item, i.e., the publisher account of the target content item, as a candidate account. It should be noted that since the reference account involved in the above step 301 may be one or more, the first account that has performed positive interaction behavior on the sensitive content item published by each reference account may also be one or more, and the target content item that each first account has also performed positive interaction behavior on may also be one or more. Therefore, the candidate account that finally publishes the target content item may also be one or more. This embodiment of the disclosure does not specifically limit this.

[0220] In some embodiments, after identifying multiple candidate accounts, the server can perform preliminary screening of these candidate accounts. For example, the server can delete certified candidate accounts and candidate accounts registered by preset organizations. Optionally, the preset organization can refer to an official institution or an organization certified by the platform. The server can create a whitelist and add official institutions and organizations certified by the platform to the whitelist. This allows for the direct deletion of candidate accounts registered by preset organizations on the whitelist, thereby improving the efficiency of account deletion. Since certified candidate accounts and candidate accounts registered by preset organizations usually receive positive interactions from many accounts, the positive interactions performed by the first account on the content published by these candidate accounts can be considered invalid interference information. Therefore, it is necessary to remove certified candidate accounts and candidate accounts registered by preset organizations from all candidate accounts, thereby enabling the pre-screening of some invalid nodes in the account relationship graph.

[0221] In step 303, the server constructs nodes in an account relationship graph based on the reference account and the candidate account. This account relationship graph is used to represent the topological structure of social relationships between publisher accounts of content items for which the first account has performed the positive interaction behavior.

[0222] In some embodiments, the server treats each reference account obtained in step 301 and each candidate account obtained in step 302 as nodes in the account relationship graph, with each node having a one-to-one correspondence with an account (reference account or candidate account). In other words, a corresponding node is constructed in the account relationship graph for each reference account, and similarly, a corresponding node is constructed in the account relationship graph for each candidate account.

[0223] In some embodiments, the symbol G is used to represent the account relationship graph. When storing the account relationship graph, a set of nodes V and a set of edges E can be used for storage. That is, G(V, E) represents the account relationship graph G determined by the set of nodes V and the set of edges E, where V represents the set of nodes in the account relationship graph and E represents the set of edges in the account relationship graph.

[0224] In the above scenario, the server can add all reference accounts and all candidate accounts to an account set. Then, it assigns a unique associated node identifier to each account in the account set and adds the node identifiers associated with each account to the node set V, thereby constructing nodes in the account relationship graph based on reference accounts and candidate accounts.

[0225] In step 304, the server constructs edges in the account relationship graph to connect nodes based on the positive interaction behavior performed by the first account.

[0226] In some embodiments, the server constructs edges in the account relationship graph to connect nodes based on whether the same first account has performed positive interaction behavior on sensitive content items published by the reference account and target content items published by the candidate account. Ultimately, an account relationship graph formed by each node and each edge can be obtained.

[0227] In some embodiments, for any reference account and any candidate account, if any first account has performed the positive interaction behavior on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account, that is, the edge connecting the node of the reference account and the node of the candidate account is added to the edge set E.

[0228] In an exemplary scenario, for reference account i and candidate account j, if there exists a first account k that has performed positive interaction behavior (such as liking) on ​​sensitive content items published by reference account i and also performed the same positive interaction behavior (such as liking) on ​​target content items published by candidate account j, that is, first account k liked both the sensitive content items published by reference account i and the target content items published by candidate account j, then an edge will be constructed in the account relationship graph to connect node u of reference account i and node v of candidate account j.<u,v> and the edge<u,v> Add it to the edge set E.

[0229] From a business perspective, accounts that publish content (including reference and candidate accounts) can be called content producers (such as video producers, live streamers, or broadcasters), and accounts that browse content (such as the primary account) can be called content consumers (such as video consumers or live stream viewers). In constructing the account relationship graph, it is equivalent to using content producers as nodes and constructing edges to connect nodes based on content consumers. Optionally, by removing accounts that have been certified or registered by a pre-defined organization, it is equivalent to excluding accounts that have been certified or registered by a pre-defined organization, thus achieving a filtering mechanism for invalid nodes. Optionally, by removing negative interaction behaviors performed by content consumers and constructing edges only for positive interaction behaviors, it is equivalent to removing negative consumers (since negative consumers have performed negative interaction behaviors on sensitive content items, it means that they are not interested in sensitive content items, thus introducing noise signals). This achieves a filtering mechanism for invalid edges, thereby improving the accuracy of the constructed account relationship graph.

[0230] Taking videos as an example, the process of constructing an account relationship graph can be illustrated as shown in Table 1:

[0231] Table 1

[0232]

[0233] Based on Table 1, video producers are used as nodes in the account relationship graph. Edges in the graph are constructed based on positive consumers. The resulting account relationship graph is a heterogeneous graph of the community. Heterogeneous graphs are those that do not distinguish between node and edge types (node ​​type + edge type = 2), while heterogeneous graphs distinguish between node and edge types (node ​​type + edge type > 2). Figure 4 As shown, Figure 4This is a comparative schematic diagram of isomorphic and heteromorphic graphs provided in an embodiment of this disclosure, showing isomorphic graph 401 and heteromorphic graph 402. It can be seen that in isomorphic graph 401, the social topology relationship of four accounts (A to D) is constructed based on social friend relationships. Accounts A and D are mutual social friends, accounts B and C are mutual social friends, and accounts B and D are also mutual friends. In isomorphic graph 401, there is no need to distinguish between different types of nodes and different types of edges. In heteromorphic graph 402, account A performs a positive interaction behavior of consuming videos (i.e., browsing videos), account B performs two positive interaction behaviors: commenting and sending private messages, and account C performs a positive interaction behavior such as watching live streams. Based on the different types of positive interactive behaviors, four different account relationship graphs can be constructed. That is, according to the different types of positive interactive behaviors, different edge types can be distinguished in the account relationship graph, thereby constructing account relationship graphs associated with different positive interactive behaviors. In this embodiment of the disclosure, from the perspective of heterogeneous graphs, by connecting the reference accounts that published sensitive content items and the candidate accounts that published target content items with positive consumers who consumed different videos, an account relationship graph is constructed. Based on the constructed account relationship graph, a community mining algorithm is used to obtain a spanning tree, cluster nodes, merge communities, and finally mine the target account.

[0234] On the one hand, by removing accounts registered by verified or pre-defined organizations, it is equivalent to deleting some invalid nodes from the account relationship graph. Since both the primary accounts interested in sensitive content and accounts uninterested in sensitive content will generally pay attention to the content published by accounts registered by verified or pre-defined organizations (such as news, trending topics, and information), these accounts may actually introduce noise interference into community mining. By removing accounts registered by verified or pre-defined organizations in the preprocessing stage, invalid nodes in the account relationship graph can be effectively suppressed, improving the accuracy of subsequent community mining.

[0235] On the other hand, since only positive consumers (i.e., the first account that performed positive interaction behavior) are used to construct the edges in the account relationship graph, negative consumers are essentially eliminated. This is because since negative consumers performed negative interaction behavior on sensitive content items, it means that they are not interested in sensitive content items in the first place, and therefore they will introduce noise interference. On the other hand, since positive consumers performed positive interaction behavior on sensitive content items, it means that positive consumers are likely to be interested in sensitive content items and are likely to browse some similar sensitive content items. Therefore, by linking the publisher accounts of target content items that positive consumers have also performed positive interaction behavior on with the corresponding reference accounts, the account relationship graph can be constructed through common positive consumers, thereby effectively suppressing invalid edges in the account relationship graph and improving the accuracy of subsequent community mining.

[0236] In steps 303-304 above, a possible implementation is shown for generating an account relationship graph associated with the positive interaction behavior based on the reference account, the first account, and the candidate account. Since the positive interaction behavior may include more than one type, the server can determine the corresponding first account for each positive interaction behavior and generate an account relationship graph associated with this positive interaction behavior. For example, an account relationship graph G1 is generated for the liking behavior, an account relationship graph G2 is generated for the following behavior of the publisher account of the content item (i.e. following the reference account or following the candidate account), an account relationship graph G3 is generated for the sharing behavior, and an account relationship graph G4 is generated for the download behavior. For each account relationship graph, the potential target accounts that have published sensitive content, which are filtered by the corresponding positive interaction behavior, can be mined through steps 305-308 below. This can improve the coverage of target account identification and comprehensively mine and discover the potential security risks in the platform from multiple dimensions.

[0237] In step 305, the server obtains the maximum spanning tree of the account relationship graph, which has the highest weight among the multiple spanning trees of the account relationship graph.

[0238] It should be noted that a connected graph without cycles is called a tree. If a spanning subgraph T of a graph G is a tree, then tree T is called a spanning tree of graph G.

[0239] In this embodiment of the disclosure, the connected and non-cyclic spanning subgraph in the account relationship graph G is the spanning tree T of the account relationship graph G. Furthermore, when the account relationship graph G is an undirected weighted graph, each edge in the account relationship graph is an undirected edge carrying a weight. Therefore, for each spanning tree T, the sum of the weights of all edges contained in the spanning tree T can be obtained. Among them, the spanning tree T* with the largest weight (i.e., the sum of the weights of all edges in the spanning tree is the largest) is called the maximum spanning tree.

[0240] In some embodiments, since steps 303-304 above only describe how to construct the nodes and edges in the account relationship graph, but in order to obtain the maximum spanning tree, it is also necessary to assign weights to each edge. Therefore, the server can assign weights to each edge through step 3051 below, and further, obtain the maximum spanning tree based on the assigned weighted undirected weighted account relationship graph through step 3052 below.

[0241] Figure 5 This is a flowchart of a method for obtaining the maximum spanning tree provided in an embodiment of this disclosure, such as... Figure 5 As shown, one possible implementation of obtaining the maximum spanning tree is as follows, involving steps 3051 and 3052:

[0242] In step 3051, the server assigns a weight to each edge based on the node similarity between the two nodes connected by each edge in the account relationship graph.

[0243] In some embodiments, for each edge contained in the edge set of the account relationship graph, the server can determine the two nodes connected by the edge, then obtain the node similarity between the two nodes, and then assign a weight to the edge according to the node similarity.

[0244] In an exemplary scenario, to connect the k-th node v k and the (k+1)th node v k+1 edge <v k v k+1 Taking the k-th node v as an example, let's illustrate this further. k and the (k+1)th node v k+1 The similarity between nodes is used as the edge <v k v k+1 The weight of > can be denoted as ω(υ k υ k+1 Each edge in the edge set E of the account relationship graph G can be assigned a weight in the above way. After traversing all edges in the edge set E, the weights of all edges in the account relationship graph G can be assigned.

[0245] In some embodiments, for each edge connecting two nodes in the account relationship graph, the node similarity between the two nodes can be determined based on the number of common neighbor nodes and the number of individual neighbor nodes of the two nodes. For example, the ratio between the number of common neighbor nodes and the sum of the number of individual neighbor nodes of the two nodes can be used as the node similarity between the two nodes. In other words, the node similarity refers to the ratio between the number of common neighbor nodes and the sum of the number of individual neighbor nodes of the two nodes.

[0246] Schematic, for the k-th node v k and the (k+1)th node V k+1 Let's assume node v k Having N1 neighbor nodes means that in the account relationship graph, a node v can be connected to it via an edge. k The number of connected nodes is N1. In other words, in the edge set of the account relationship graph, node v... k The number of edges serving as endpoints is N1, assuming node v k+1 Having N2 neighbor nodes means that in the account relationship graph, a node v can be connected to it via an edge. k+1 The number of connected nodes is N². In other words, in the edge set of the account relationship graph, node v...k+1 The number of edges serving as endpoints is N², assuming node v k and node v k+1 The number of common neighbor nodes is N3, meaning there are N3 neighbor nodes that are simultaneously connected to node v. k and node v k+1 Connected, therefore, node v k and node v k+1 The node similarity between them is ω(υ k υ k+1 = N3 / (N1+N2).

[0247] In the above process, by using the ratio of the number of common neighbor nodes between two nodes to the sum of the number of their respective neighbor nodes as node similarity, we can characterize how many shared accounts have built a social topology relationship based on common positive consumers between the two accounts indicated by the two nodes. The ratio of the number of shared accounts to the sum of the number of their respective related accounts symbolizes the overlap of common consumers of the content items published by the two accounts. This provides a business-meaning node similarity between the two accounts and improves the expressive power of node similarity.

[0248] In step 3052, the server generates the maximum spanning tree based on the weights of each edge in the account relationship graph, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the multiple spanning trees.

[0249] In some embodiments, after assigning a sum of weights to each edge in the account relationship graph G through the above step 3051, the account relationship graph G will be transformed from an undirected graph into an undirected weighted graph. At this time, all spanning trees of the account relationship graph G can be obtained first, and then the sum of weights of all edges in each spanning tree can be calculated. The spanning tree with the largest sum of weights is obtained as the maximum spanning tree.

[0250] In other embodiments, the server can also generate a maximum spanning tree containing n nodes and n-1 edges by first constructing an empty tree T and then selecting and adding the n-1 edges with the largest weights to the set of empty trees T, as described below:

[0251] A1. Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is empty.

[0252] In some embodiments, when initializing an empty tree, a starting node x is first randomly sampled from the account relationship graph, and then the node set of the empty tree is initialized to V. new={x}, initializes the edge set of the empty tree to an empty set, i.e., E. new ={}.

[0253] A2. From the account relationship graph, obtain multiple candidate edges to connect nodes within the node set and nodes outside the node set.

[0254] In some embodiments, for the node set V new The element in the set V is node u, which determines the node set used to connect the nodes. new The node u and the node set V within new Candidate edges of the outer node v<u,v> For the node set V new The above operation is performed on each element, i.e., each node, in the set. Each node within the set may form one or more candidate edges with nodes outside the set. After traversing the node set V, new After considering all elements (i.e., all nodes in the set), the multiple candidate edges obtained for each node are summarized. Here, node v is an element in the node set V of the account relationship graph, but not in the node set V of the current tree. new The elements in the set V, in other words, v∈V but v is not in the node set V. new middle.

[0255] A3. Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the node outside the node set connected by the target edge to the node set.

[0256] In some embodiments, for all candidate edges obtained in step A2 above, the weight assigned to each candidate edge in step 3051 above is obtained, the candidate edge with the largest weight is determined as the target edge, and then the target edge is added to the edge set E of the maximum spanning tree. new In this context, since the candidate edges are themselves elements of the edge set in the account relationship graph, it is assumed that the edge with the largest weight is...<u,v> This is equivalent to selecting multiple candidate edges from the edge set E of the account relationship graph through step A2, and then choosing the candidate edge with the largest weight in edge set E as the target edge.<u,v> Add to the edge set E of the current tree new In the middle, at this time, due to the target edge<u,v> It was added to the edge set E new In the middle, it is also necessary to correspondingly target the edge.<u,v> The node v outside the included set is also added to the node set V. new middle.

[0257] A4. Repeat the operations of A2-A3 to add target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and node set when adding stops is determined as the maximum spanning tree.

[0258] In some embodiments, steps A2-A3 above are repeated, and each time a set of edges E is selected from multiple candidate edges. new Add the target edge with the largest weight to the node set V, and simultaneously add the target edge with the largest weight to the node set V. new Add nodes outside the set connected to the target edge to the edge set E until no target edge can be found, then stop adding nodes to the edge set E. new Adding a new edge to the node set V naturally stops adding edges to the node set V. new Adding a new node to the edge set E is equivalent to adding all target edges to the edge set E. new In the process, all interconnected nodes are added to the node set V. new In this process, the node set V formed at this time new The set of edges E new This allows us to determine a unique maximum spanning tree.

[0259] By using steps A1-A4 above to obtain the maximum spanning tree, it is not necessary to obtain all spanning trees of the account relationship graph and calculate the sum of the weights of all edges contained in each spanning tree. This can also determine the maximum spanning tree, reduce the amount of computation when obtaining the maximum spanning tree, and improve the calculation efficiency of the maximum spanning tree.

[0260] In steps 3051-3052 above, weights are assigned to edges based on node similarity, and a maximum spanning tree is generated based on the assigned account relationship graph. This allows nodes outside the maximum spanning tree to be excluded during subsequent community mining, which not only reduces the computational load of the community mining algorithm but also further eliminates invalid nodes, preventing invalid nodes from causing noise interference to the subsequent community mining algorithm.

[0261] In other embodiments, instead of obtaining the maximum spanning tree, the community mining algorithm can be performed directly on the original account relationship graph, thereby saving the computational cost of obtaining the maximum spanning tree.

[0262] In step 306, the server clusters the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups.

[0263] In some embodiments, the server clusters the accounts indicated by each node on the maximum spanning tree, thereby obtaining candidate account groups formed by account clustering when clustering stops. These candidate account groups are also called local communities to be merged. Before clustering begins, all core nodes can be found from the maximum spanning tree. Then, the core nodes are used as the cluster center of each candidate account group, and other non-core nodes are clustered into the candidate account group where the most similar core node is located. In this way, local communities can be gradually clustered with the core nodes as the cluster center.

[0264] Figure 6 This is a flowchart of obtaining a candidate account group provided in an embodiment of this disclosure, such as... Figure 6 As shown, before starting node clustering, all core nodes in the maximum spanning tree are found, and then clustering begins with the core nodes as cluster centers. This will be explained below.

[0265] In step 3061, the server selects multiple core nodes from the nodes contained in the maximum spanning tree.

[0266] In some embodiments, for any node contained in the maximum spanning tree, the server determines at least one neighbor node in the maximum spanning tree that is connected to the node by an edge. In other words, it finds all neighbor nodes that can be connected to the node by an edge in the maximum spanning tree. That is, it finds the other endpoint of all edges that start from the node as an endpoint in the maximum spanning tree and uses them as the neighbor nodes of the node. Since the maximum spanning tree is a connected spanning subgraph of the account relationship graph, the number of neighbor nodes of each node is at least one. In other words, the number of neighbor nodes of each node is greater than or equal to 1.

[0267] In some embodiments, the server filters out target neighbor nodes from the at least one neighbor node, where the weight of the connecting edge between the node and the neighbor node is greater than or equal to a first weight threshold. In other words, after obtaining at least one neighbor node for each node, the server can determine whether the weight of the connecting edge between the node and the neighbor node is greater than or equal to the first weight threshold for each neighbor node. If the weight of the connecting edge is greater than or equal to the first weight threshold, the current neighbor node is determined as a target neighbor node. If the weight of the connecting edge is less than the first weight threshold, the above judgment operation is performed on the next neighbor node. This operation is repeated until all neighbor nodes of the node have been traversed. At this point, all the filtered target neighbor nodes can be obtained. The first weight threshold is a value preset by an engineer, or it is a value sampled from the weight set composed of the weights of all edges in the maximum spanning tree. This embodiment does not specifically limit the method of obtaining the first weight threshold.

[0268] In some embodiments, if the sum of the weights of the connection edges between the node and each of the target neighbor nodes is greater than a second weight threshold, the server determines the node as the core node. In other words, after the server has filtered out all the target neighbor nodes, the connection edges between the node and each target neighbor node can form a set of connection edges: Γ ε(u, υ) = {υ∈Nghb(u)|ω(u, υ)≥ε}, where ε represents the first weight threshold, ω(u, υ) represents the weight of the connecting edge between node u and its target neighbor node v, and Nghb(u) represents the neighbor nodes of node u. That is, the target neighbor node v is a node that satisfies the condition that the weight ω(u, υ) of the connecting edge among the neighbor nodes of Nghb(u) is greater than or equal to the first weight threshold ε. The connecting edges formed by all target neighbor nodes v and node u satisfying the above condition are...<u,v> A set of connecting edges Γ can be formed ε (u, υ), if the set of edges Γ is connected ε All connecting edges contained in (u, υ)<u,v> The sum of the weights is greater than the second weight threshold, assuming |Γ ε (u, υ)| represents Γ ε All connecting edges contained in (u, υ)<u,v> Let μ represent the second weight threshold, and then if |Γ ε If (u, υ)|>μ, then the current node u is determined as a kernel node.

[0269] In the above process, for each node, the connecting edges with a weight greater than or equal to the first weight threshold are added to the connecting edge set. When the sum of the weights of the connecting edges contained in the connecting edge set is greater than the second weight threshold, the current node is regarded as a core node. The above operation is repeated until all nodes in the maximum spanning tree have been traversed. This can complete the screening of all core nodes in the maximum spanning tree. That is, if the total weight of the connecting edges with large weights where a node is located exceeds the second weight threshold, it means that this node is a relatively important and core node in the maximum spanning tree. By screening core nodes in this way, it can be ensured that the total weight of the connecting edges with large weights where the core node is located is also large. At this time, either the number of connecting edges with large weights where the core node is located is large, or the individual weight of the connecting edges with large weights where the core node is located is large, thereby optimizing the selection method of core nodes.

[0270] In other embodiments, the server may select nodes in the set of connected edges whose number of connected edges is greater than a threshold as core nodes. This ensures that the number of connected edges with larger weights at the core node is always greater than the threshold. The threshold can be any value greater than or equal to 1 preset by an engineer; this disclosure does not impose specific limitations on it.

[0271] In step 3062, the server uses the multiple core nodes as cluster centers to determine multiple candidate account groups.

[0272] In some embodiments, for each core node selected in step 3061 above, a candidate account group containing only the account indicated by the core node is initialized using the core node as the cluster center. The above operation is performed on each core node to obtain all the initialized candidate account groups. Before the clustering starts, each candidate account group contains only the account indicated by the corresponding core node. Each candidate account group has a one-to-one correspondence with a core node.

[0273] In step 3063, the server clusters all nodes in the maximum spanning tree except for the core node into the candidate account group where the core node with the highest path similarity to the core node is located.

[0274] In some embodiments, the server needs to perform node clustering on all nodes in the maximum spanning tree except for the core nodes (i.e., ordinary nodes that are not core nodes) to assign each node to the candidate account group containing the core node with the highest path similarity. Optionally, for each node other than the core nodes, the path similarity between the node and all core nodes selected in step 3061 above is first calculated, and then the node is clustered into the candidate account group containing the core node with the highest path similarity. Alternatively, the core nodes are sorted in descending order of path similarity, and the node is assigned to the candidate account group containing the first-ranked core node.

[0275] In some embodiments, for any node in the maximum spanning tree other than the core node, the server can calculate the path similarity between the node and each core node selected in step 3061 as follows: In the maximum spanning tree, determine the connected paths between the node and any core node. A connected path refers to a path from the node to the core node via multiple edges. For example, for non-core node s and core node t, determine one or more connected paths p(s,t) that connect node s and core node t. Then, based on the weight of each edge contained in the connected path, determine the path similarity between the node and the core node. For example, for each connected path p(s,t), the k-th node v... k and its neighboring node v on the connected path k+1 The weight ω(υ) of the connecting edge between k υ k+1 Taking the reciprocal, the sum of the reciprocals of the weights of all connected edges is taken as the contribution of the current connected path p(s,t). The reciprocal of the sum of the contributions of all connected paths between node s and core node t is taken as the path similarity S between node s and core node t. In other words, the path similarity S(s,t) between node s and core node t can be expressed as the following formula:

[0276]

[0277] Where S(s,t) represents the path similarity between node s and core node t, p(s,t) represents the connected path between node s and core node t, and k represents the k-th node v on the connected path p(s,t). k The sequence number, ω(υ) k υ k+1 ) represents the k-th node v k and its neighboring node v on the connected path k+1 The weight of the connecting edges between them.

[0278] In the above process, by taking into account the connected paths between each node and the core node when determining the path similarity, we can take into account both the local information between the nodes and the core node and the global topology of the entire maximum spanning tree. This allows for a more accurate measurement of the similarity between the nodes and the core node after introducing the connected path factor. Because the path similarity is more accurate, the nodes can be accurately assigned to the candidate account group where the most similar core node is located, thus greatly improving the accuracy of the division of local communities in the maximum spanning tree.

[0279] In other embodiments, the server may not use path similarity for node clustering, but instead perform node clustering based on the node similarity described in step 3051 above. Since the node similarity has already been obtained in step 3051, the node similarity already calculated in step 3051 can be directly reused, thereby saving the server's computing resources.

[0280] In steps 3061-3063 above, by first selecting core nodes from the maximum spanning tree and then using the core nodes as cluster centers to divide local communities, i.e., candidate account groups, this ensures that when initializing each candidate account group, the core nodes are used to ensure that the cluster centers of the candidate account groups are core nodes with a large weight contribution in the entire maximum spanning tree. Compared with randomly sampling multiple nodes in the maximum spanning tree as initial cluster centers, this improves the clustering efficiency of converging to candidate account groups. Furthermore, by reducing the randomness of initializing cluster centers, the clustering accuracy is also improved under the influence of core nodes.

[0281] In other embodiments, the server may also randomly sample multiple nodes from the maximum spanning tree as cluster centers, and determine multiple candidate account groups based on the randomly sampled nodes. Then, node clustering is performed based on some KNN (K-Nearest Neighbor) algorithm or K-means algorithm. This disclosure does not specifically limit the node clustering method.

[0282] In step 307, the server merges the multiple candidate account groups to obtain the target account group.

[0283] In some embodiments, after clustering nodes into candidate account groups through step 306 above, it is also necessary to merge similar candidate account groups to obtain one or more final target account groups. If the candidate account groups are regarded as local communities in the maximum spanning tree, this merging process is equivalent to merging local communities to obtain a large community that can no longer be merged.

[0284] In some embodiments, before merging candidate account groups, it can be determined whether there are any candidate account groups whose core nodes are directly connected. If there are two candidate account groups whose core nodes are directly connected and the weight of the connecting edge is greater than the first weight threshold, then the two candidate account groups are directly merged. After all candidate account groups with connected core nodes and connecting edge weights greater than the first weight threshold have been merged, for the remaining candidate account groups, each candidate account group is iteratively merged into another candidate account group with the highest group similarity. Finally, when all candidate account groups can no longer be merged, the entire network will present a single large community. At this time, the account group with the largest modularity can be selected as the target account group from the account groups generated by each iteration of merging.

[0285] Figure 7 This is a flowchart illustrating a method for merging candidate account groups, as provided in this embodiment of the disclosure. Figure 7 The diagram illustrates the server's merging process for candidate account groups under the aforementioned method, which will be explained below.

[0286] In step 3071, the server merges any candidate account group among the multiple candidate account groups with the other candidate account groups if there is a connection edge between the core node of the candidate account group and the core node of other candidate account groups, and the weight of the connection edge is greater than the first weight threshold.

[0287] In some embodiments, before merging candidate account groups, it is first determined whether there are any candidate account groups whose core nodes are directly connected. If there are two candidate account groups whose core nodes are directly connected, and the weight of the connection edge between the two core nodes is greater than the first weight threshold, then the two candidate account groups are directly merged.

[0288] In some embodiments, the server traverses all candidate account groups in all candidate account groups. If there is a connection edge between the core node of the candidate account group and the core node of any other candidate account group, it means that the core node of the candidate account group is directly connected to the core node of the other candidate account group. In this case, it continues to determine whether the weight of the connection edge between the core node of the candidate account group and the core node of the other candidate account group is greater than a first weight threshold. If the weight of the connection edge between the two core nodes is greater than the first weight threshold, the candidate account group and the other candidate account group are merged to obtain a new candidate account group. Otherwise, if the weight of the connection edge between the two core nodes is less than or equal to the first weight threshold, the account groups cannot be merged directly even if the two core nodes are directly connected.

[0289] It should be noted that each core node of this candidate account group may have connections to more than one core node of other candidate account groups. Therefore, if the weight of each connection between a core node and more than one other core node is greater than the first weight threshold, the candidate account group and other candidate account groups directly connected to multiple core nodes with connection weights greater than the first weight threshold can be collectively merged. For example, core node u of candidate account group A... A Core nodes u of other candidate account groups B, C, and D respectively B u C u D There are connecting edges between them, and the core node u A With kernel node u B Connection edge A u B > weight ω(u A u B If the value is greater than the first weight threshold ε, the kernel node u A With kernel node u C Connection edge A u C > weight ω(u A u C If the value is greater than the first weight threshold ε, the kernel node u A With kernel node u D Connection edge A u D > weight ω(u A u​​​D If the value is less than the first weight threshold ε, then candidate account group A can be merged with other candidate account groups B and C, but cannot be merged with other candidate account group D. That is, the three candidate account groups A, B, and C can eventually be merged into a new candidate account group.

[0290] It should be noted that if there are no connecting edges between the core node of the current candidate account group and the core nodes of all other candidate account groups, it means that the core node of the current candidate account group does not have any candidate account groups directly connected to it. Alternatively, although the core node of the current candidate account group is connected to the core nodes of other candidate account groups, the weight of any connecting edge is greater than the first weight threshold. In this case, it is necessary to merge them through step 3072 below. After traversing all candidate account groups, it is possible to find all candidate account groups that have core nodes directly connected and whose connecting edge weights are greater than the first weight threshold and merge them. After one round of traversal, it is possible to merge all candidate account groups that simultaneously satisfy the following two conditions: 1) There are connecting edges between the core nodes of two candidate account groups; 2) All candidate account groups whose connecting edge weights are greater than the first weight threshold have been merged in one round. That is, when it is guaranteed that no two original candidate account groups or merged candidate account groups simultaneously satisfy the above two conditions 1) and 2), proceed to step 3072 below.

[0291] In other embodiments, during the first round of merging, the condition 2) that the weight of the connecting edge is greater than the first weight threshold can be modified to: the path similarity between the two core nodes is greater than the first weight threshold, that is, the path similarity between the two core nodes is judged, rather than the weight of the connecting edge between the two core nodes (as can be seen from step 3051, the weight of the connecting edge actually represents the node similarity between the two core nodes). In this way, since the path similarity can take into account the connected path between the two core nodes, the stringency of directly merging candidate account groups can be increased, and the accuracy of the order of community merging can also be improved to a certain extent.

[0292] In step 3072, otherwise, the server merges the candidate account group with other candidate account groups with the highest group similarity.

[0293] In some embodiments, when no two account groups (which may be the original candidate account groups or the candidate account groups obtained by merging through step 3071) simultaneously satisfy the above two conditions 1) and 2), the server calculates the group similarity between each candidate account group remaining after one round of merging in step 3071 and each other candidate account group, and then merges the candidate account group with the other candidate account group with the highest group similarity.

[0294] In some embodiments, for any candidate account group among the plurality of candidate account groups, the server can calculate the group similarity between the candidate account group and any other candidate account group in the following manner: The sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group. The degree parameter of a node represents the number of edges connected to that node in the maximum spanning tree. In other words, first, the degree parameters of the nodes associated with each account in the candidate account group are obtained, and then the sum of the degree parameters of each node is obtained as the group degree parameter of the candidate account group. Then, the similarity is obtained for the other candidate account groups... The degree parameters of the nodes associated with each account are included, and the sum of the degree parameters of each node is obtained as the group degree parameter of the other candidate account group. Then, based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account group, and the number of common edges between the candidate account group and the other candidate account group, the group similarity between the candidate account group and the other candidate account group is determined. That is, a mapping relationship is established so that the group similarity can be mapped to the group degree parameter of the candidate account group, the group degree parameter of the other candidate account group, and the number of common edges between the candidate account group and the other candidate account group under the influence of three variables.

[0295] It should be noted that since a candidate account group is essentially a subgraph on a maximum spanning tree, a common edge refers to the number of identical edges in the subgraphs corresponding to the two candidate account groups. It also represents the number of identical elements in the edge set of the subgraphs corresponding to the two candidate account groups, which is equivalent to reflecting the number of common edges shared by the two local communities.

[0296] In some embodiments, the server assumes that the candidate account group is the i-th candidate account group, and the other candidate account groups are the j-th candidate account group c. j First, identify the candidate account group c. i and the other candidate account group c j The number of common edges in the middle, num(c i c j Next, obtain the candidate account group c. i Group degree parameter and the other candidate account group c j Group degree parameter The product between Next, the product Obtain the arithmetic square root Then calculate the number of common edges num(c) i c jDivide by the arithmetic square root The obtained value is used as the candidate account group c. i Group c of other candidate accounts j The group similarity between groups, in other words, the group similarity can be represented by the following formula:

[0297]

[0298]

[0299] Wherein, S(c i c j ) represents the i-th candidate account group c i Group c with the j-th candidate account j Group similarity between them, num(c i c j ) represents candidate account group c i and candidate account group c j The number of common edges in the middle, Characterizing candidate account group c i The group degree parameter, Characterizing candidate account group c j The group degree parameter, |c i |Characterizes candidate account group c i The total number of accounts included (i.e., the total number of nodes), υ j Characterized in candidate account group c i The j-th node in the array, degree(υ j ) Representation node υ j The degree parameter.

[0300] In the above process, by using the group degree parameters of each of the two candidate account groups and the number of common edges in the two candidate account groups, we can obtain a precise group similarity that represents the degree of similarity between the two groups. The larger the group similarity value, the higher the similarity between the two candidate account groups, and the smaller the group similarity value, the lower the similarity between the two candidate account groups. Thus, we can further merge the candidate account groups under the guidance of group similarity. The merging order of the candidate account groups determines the modularity of the account groups obtained after each merge. Therefore, by prioritizing the merging of the candidate account groups with the highest group similarity under the guidance of group similarity, it is beneficial to select the target account group with the optimal partitioning method.

[0301] It should be noted that since each candidate account group can find another candidate account group with the highest group similarity, when deciding which pair of candidate account groups to merge first, the group similarity of all paired candidate account groups can be compared horizontally. Thus, the pair of candidate account groups with the highest group similarity is merged first. For example, if the group similarity of candidate account group A and candidate account group B is the highest and equal to 0.8, the group similarity of candidate account group C and candidate account group D is the highest and equal to 0.5, and the group similarity of candidate account group E and candidate account group F is the highest and equal to 0.6, then in the entire merging process, the candidate account group A and candidate account group B with the highest group similarity after horizontal comparison are merged first, resulting in a new candidate account group A+B.

[0302] In some embodiments, after merging each candidate account group with the other candidate account group with the highest similarity, it is possible to record which two candidate account groups were merged and the modularity of the merged account group. The above merging operation on account groups is performed iteratively until no candidate account groups can be merged. At this point, it is equivalent to all candidate account groups being aggregated into a large account group, i.e., a large community. At this point, based on the modularity recorded in each merged account group, it can be selected which merged account group is the optimal target account group.

[0303] In step 3073, the server repeatedly performs the operation of merging candidate account groups until no candidate account groups can be merged, and selects the target account group from the account groups obtained from each merge.

[0304] In some embodiments, after merging any two candidate account groups through step 3072 above, it is recorded which candidate account groups were merged into the account group obtained in this merge, and the modularity of the account group obtained in this merge is also recorded. Thus, under the guidance of the modularity, the account group with the largest modularity value can be selected as the target account group.

[0305] In other words, after each merge operation in step 3072 above, the server obtains the modularity of the account group obtained from the merge. The modularity is used to measure the quality of the division of the account group from the account relationship graph. The server iteratively executes the merge operation of candidate account groups in step 3072 above until no candidate account group can continue to be merged. At this time, the server can query the modularity of the account groups obtained from each merge and determine the account group with the highest modularity as the target account group.

[0306] In one example, let Q represent the modularity of an account group. Assume that in the first merge, candidate account groups A and B are merged to obtain a new candidate account group A+B, with a modularity of Q1. In the second merge, candidate account group C is merged again with the previously merged candidate account group A+B, resulting in a new candidate account group A+B+C, with a modularity of Q2. In the third merge, candidate account group D is merged again with the previously merged candidate account group A+B+C, resulting in a new candidate account group A+B+C+D, with a modularity of Q3. After this, no other candidate account groups can be merged. At this point, the query returns the maximum value Q in the queue {Q1, Q2, Q3}. max =Q2, which means that the account group A+B+C obtained from the second merge is the optimal partition of the account group in the current maximum spanning tree (i.e., it has the highest community partitioning quality). Therefore, the account group A+B+C obtained from the second merge will be selected as the target account group for the final output.

[0307] In the above process, by recording the modularity of the account groups obtained by each merging during the iterative execution of the merging operation, the target account groups with the best community partitioning quality can be selected by using the modularity as a guide. This makes the output target account groups represent the optimal partitioning of account groups in the maximum spanning tree, thereby improving the partitioning accuracy of the selected target account groups and ensuring that the target account groups have the best community partitioning quality.

[0308] In steps 3071-3073 above, a second round of merging of candidate account groups is performed based on the connection status of the core nodes and the weights of the connecting edges. Then, based on the group similarity between different candidate account groups, the second round of merging is iteratively executed. Finally, based on the modularity of the account groups obtained by each merging, the target account group of the final output is selected, so that account groups with closer connections and higher similarity are merged first, and the target account group with the best community division quality is selected, thereby ensuring the accuracy of target account group identification and the quality of community division.

[0309] In some embodiments, steps 305-307 above are equivalent to providing a community mining algorithm based on common consumers. The input to this algorithm is an undirected, unweighted network, i.e., the original unweighted undirected account relationship graph G(V, E). The output of this algorithm is a target account group (i.e., the mined community structure), the modularity Q-value of the target account group, and a first weight threshold ε. Illustratively, this algorithm includes the following steps:

[0310] (I) Generating the maximum spanning tree: First, calculate the weight of each edge in the edge set E of the account relationship graph G to transform the undirected unweighted network G(V, E) into an undirected weighted network G(V, E, ω). Then, generate a maximum spanning tree T(V, E) for the undirected weighted network G(V, E, ω). T );

[0311] (II) Determine the kernel node: For the edge set E of the maximum spanning tree T(V, ET) T Sort the weights of the edges and record each edge in the candidate queue W in the order of sorting. T In the middle, then, in queue W T Choose a new edge weight as the first weight threshold ε, and then in the maximum spanning tree T(V, E) T In step 3061, all core nodes are filtered out.

[0312] (III) Node clustering: Calculate the path similarity S(s,t) from all nodes to the core node, and select the core node with the largest path similarity for each node to merge, thereby generating various local communities (i.e., candidate account groups); then, for each candidate account group generated after node clustering, determine whether there is a direct connection between the core nodes and the weight of the connecting edge is greater than the first weight threshold ε. If the above conditions are met, the local communities where the two core nodes are located are directly merged; or, determine whether there is a direct connection between the core nodes and the path similarity between the two core nodes is greater than the first weight threshold ε. If the above conditions are met, the local communities where the two core nodes are located are directly merged.

[0313] (IV) Local community merging: Calculate the group similarity between local communities, select the two local communities with the highest group similarity to merge, that is, calculate the group similarity between all local communities pairwise, then merge the two local communities with the highest group similarity, and then calculate and record the modularity Q value of the merged community.

[0314] (V) Repeat step (IV) until the entire network merges into one large community, find the largest Q value and record it in queue Q. S middle;

[0315] (VI) Determine the candidate queue W T If all edges in the queue have been traversed, and there are still untraversed edges, then repeat steps (II) to (V). If there are no untraversed edges, it means that the traversal has been completed, then the algorithm ends, and queue Q is selected. S Take the largest Q value (e.g., Q) * Output Q. * The corresponding community structure, i.e., the modularity, is Q.* The target account group, and the first weight threshold ε in this case.

[0316] For the above algorithm, assuming the total number of nodes in the account relationship graph is n and the total number of edges is m, the worst-case complexity of calculating the maximum spanning tree is O(mlogn). Assuming the number of core nodes in the maximum spanning tree is k, the complexity of calculating the path similarity between each node and each core node is nk. Clustering k local communities based on k core nodes requires k-1 steps to gradually cluster these k local communities into one community; therefore, the time complexity of merging local communities is k-1. Since the algorithm uses the maximum spanning tree T(V, E)... T The weights of the edges on the first weight threshold ε are used as the candidate set. Therefore, in the worst case, the above process needs to be repeated n-1 times. In summary, the time complexity of the community partitioning algorithm based on the maximum spanning tree is O(mn) throughout the entire process.

[0317] In step 308, the server identifies the candidate accounts contained in the target account group as the target accounts.

[0318] In some embodiments, after obtaining the target account group through step 307 above, since the target account group may include reference accounts and candidate accounts, and the reference accounts are known accounts that have published sensitive content items, when performing security risk control, it is only necessary to output the candidate accounts included in the target account group as target accounts, so as to accurately suppress and inhibit the target accounts, and also accurately suppress the content items published by the target accounts, thus achieving accurate identification and accurate suppression of target accounts.

[0319] In steps 305-308 above, a possible implementation method is shown for selecting target accounts that form an account group with the reference account from the candidate accounts based on the account relationship graph. That is, by performing node clustering and account group merging on the maximum spanning tree of the account relationship graph, the optimal segmentation of the target account group is mined out, and then the target account is selected from the target account group. In this way, some invalid nodes located outside the maximum spanning tree can be removed by using the maximum spanning tree. Then, node clustering is used to merge scattered nodes into the candidate account group where the most similar core node is located. Then, similar candidate account groups are merged. With the modularity of the merged account group as the screening index, the target account group can be realized as the optimal segmentation of the maximum spanning tree, thereby greatly improving the accuracy of target account identification.

[0320] Figure 8This is a schematic diagram illustrating the principle of negative account suppression after filtering target accounts based on an account relationship graph, as provided in this embodiment. Figure 8 As shown, assuming two reference accounts are known: video producer 801 and video producer 802, then, through the n (n≥1) positive consumers C1 to Cn shared between video producer 801 and video producer 802, the target account (i.e. the potential producer of negative videos) 803 with potential sensitive content items is mined in the account relationship graph. Next, 30-day bans or silences can be imposed on target account 803 to promptly discover and control potential security risks within the platform.

[0321] Figure 9 This is a schematic flowchart illustrating the principle of an account acquisition method provided in this disclosure embodiment, such as... Figure 9 As shown, in the account relationship graph construction phase, nodes in the graph are constructed using content producer A (e.g., reference account) and content producer B (e.g., candidate account), and edges connecting these nodes are constructed based on common consumers (e.g., the first account). Next, in the pruning phase, invalid nodes are removed by eliminating verified or officially registered accounts. Furthermore, invalid edges are removed by eliminating negative consumers (i.e., common consumers who have performed negative interactive behaviors). Then, in the community mining phase, node clustering and account group merging are performed using the core node as the cluster center to output the target account group with the highest Q-value. Finally, in the stage of discovering unknown negative accounts, the known reference accounts A and B are excluded from the target account group, leaving the unknown target account C discovered in this mining phase. Target account C is the machine-identified potential negative account that has published sensitive content, thus allowing for precise identification and negative suppression of target account C to reduce platform security risks.

[0322] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0323] The method provided in this disclosure leverages the characteristic that a first account has performed positive interactive behaviors on different content items. This allows the reference account that publishes sensitive content items and the candidate account that publishes target content items to be linked in an account relationship graph. This constructs an account relationship graph based on the first account that has performed positive interactive behaviors on content items published by different accounts. Community mining is then performed based on this constructed account relationship graph. By introducing sensitive content items and reference accounts as monitoring signals, community mining can avoid discovering account groups that contradict business significance, thereby greatly improving the accuracy of identifying potential target accounts that have published sensitive content.

[0324] Figure 10 This is a schematic diagram illustrating the effect of an account acquisition method provided in an embodiment of this disclosure, such as... Figure 10 As shown, black nodes represent known reference accounts, white nodes represent mined target accounts, and gray nodes represent normal accounts. The account acquisition method provided in this embodiment can automatically mine potential target accounts when reference accounts serve as monitoring signals. These target accounts are very likely to be negative accounts that have also published sensitive content. Therefore, after manual verification of the target accounts, they can be precisely suppressed. This can reduce the human cost in the process of suppressing negative accounts. That is, users only need to verify the target accounts mined by the algorithm, instead of verifying a large number of platform accounts one by one. This greatly reduces the human cost of security risk control and improves the efficiency of security risk control.

[0325] During testing of the account acquisition method based on the embodiments of this disclosure, after being deployed to the platform, approximately 400 new target accounts were identified daily. The accuracy rate for identifying negative accounts reached 72%, while the business recall rate was only 24%. The business recall rate refers to the gain rate of newly recalled negative accounts after using the CCCM algorithm compared to before using the CCCM algorithm. Furthermore, during testing using sensitive videos as an example, the test results are shown in Table 2.

[0326] Table 2

[0327] Typical Clubs Number of accounts negative account ratio Algorithm gain Club Tags Reference account 19 100％ / Sensitive Anime Target account 40 72％ 49％ Sensitive Anime

[0328] Among them, the negative account ratio refers to the ratio between the actual negative accounts and the target accounts identified by the algorithm, and the algorithm gain refers to the gain value brought about by the identification of target accounts before and after the application of the CCCM algorithm. That is, how much the number of additional negative accounts identified before and after the application of the CCCM algorithm has increased compared to the number of known reference accounts before the application of the CCCM algorithm.

[0329] In this disclosed embodiment, a new approach to constructing account relationship graphs is proposed by connecting different content producers based on common consumers. It also introduces supervised reference accounts to mine malicious communities (i.e., target account groups) composed of negative accounts under the influence of supervisory signals. Furthermore, different account relationship graphs can be constructed and malicious communities discovered from different dimensions can be mined based on different positive interactive behaviors (such as liking, following, sharing, and downloading) performed by common consumers. This enables automated security risk control and suppression of negative accounts on the platform.

[0330] Figure 11 This is a logical structure block diagram of an account acquisition device according to an exemplary embodiment. (Refer to...) Figure 11 The device includes a determining unit 1101, a generating unit 1102, and a filtering unit 1103.

[0331] The determination unit 1101 is configured to perform actions based on sensitive content items published by reference accounts, and to determine the first account that has performed positive interaction behavior on the sensitive content item, wherein the positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content item.

[0332] The determining unit 1101 is also configured to execute the target content item based on the positive interaction behavior performed by the first account, and determine the candidate account for publishing the target content item;

[0333] The generation unit 1102 is configured to generate an account relationship graph associated with the positive interaction behavior based on the reference account, the first account, and the candidate account. The account relationship graph is used to represent the topological structure of social relationships between publisher accounts of content items for which the first account has performed the positive interaction behavior.

[0334] The filtering unit 1103 is configured to perform a filtering operation based on the account relationship graph to obtain target accounts that form an account group with the reference account from the candidate accounts.

[0335] The apparatus provided in this embodiment leverages the characteristic that a first account has performed positive interactive behaviors on different content items. This allows it to connect reference accounts that publish sensitive content items and candidate accounts that publish target content items in an account relationship graph. This constructs an account relationship graph based on the first account that has performed positive interactive behaviors on content items published by different accounts. Community mining is then performed based on this constructed account relationship graph. Because sensitive content items and reference accounts are introduced as monitoring signals, community mining under the influence of these signals avoids uncovering account groups that contradict business objectives, thereby significantly improving the accuracy of identifying potential target accounts that have published sensitive content.

[0336] In some embodiments, based on Figure 11 The apparatus comprises, wherein the generating unit 1102 includes:

[0337] The node building subunit is configured to perform the task of building nodes in the account relationship graph based on the reference account and the candidate account;

[0338] The edge constructs a sub-unit, which is configured to perform positive interactive behaviors based on the first account, and constructs the edges in the account relationship graph to connect nodes.

[0339] In some embodiments, the edge building subunit is configured to perform:

[0340] If the first account has performed the positive interaction on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account.

[0341] In some embodiments, based on Figure 11 The device comprises a screening unit 1103 including:

[0342] The sub-unit is configured to retrieve the maximum spanning tree of the account relationship graph, which has the maximum weight among the multiple spanning trees of the account relationship graph.

[0343] The clustering subunit is configured to perform clustering of the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups;

[0344] The merge subunit is configured to merge the multiple candidate account groups to obtain the target account group;

[0345] The determination subunit is configured to perform the task of identifying the candidate accounts contained in the target account group as the target account.

[0346] In some embodiments, based on Figure 11 The device comprises, wherein the acquisition subunit includes:

[0347] The sub-unit is configured to assign a weight to each edge based on the node similarity between the two nodes connected by each edge in the account relationship graph.

[0348] The sub-unit is configured to generate the maximum spanning tree based on the weights of each edge in the account relationship graph, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the multiple spanning trees.

[0349] In some embodiments, the node similarity refers to the ratio between the number of common neighbor nodes of the two nodes and the sum of the number of each node's individual neighbor nodes.

[0350] In some embodiments, the generating subunit is configured to perform:

[0351] Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is empty.

[0352] From the account relationship graph, obtain multiple candidate edges to connect nodes within the node set and nodes outside the node set;

[0353] Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the node outside the node set connected by the target edge to the node set;

[0354] Repeat the operation of adding target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and node set at the point where addition stops is determined as the maximum spanning tree.

[0355] In some embodiments, based on Figure 11 The device comprises the following components:

[0356] The filtering sub-unit is configured to perform filtering to obtain multiple core nodes from the nodes contained in the maximum spanning tree;

[0357] The sub-units are determined and configured to perform the task of identifying multiple candidate account groups with the multiple core nodes serving as cluster centers.

[0358] The clustering sub-unit is configured to perform clustering of all nodes in the maximum spanning tree except the core node into the candidate account group where the core node has the highest path similarity to the current node.

[0359] In some embodiments, the filtering subunit is configured to perform:

[0360] For any node contained in the maximum spanning tree, determine at least one neighboring node in the maximum spanning tree that is connected to that node by an edge;

[0361] Among the at least one neighboring node, target neighboring nodes whose weight of the connection edge between the node and the neighboring node is greater than or equal to a first weight threshold are selected;

[0362] If the sum of the weights of the connecting edges between this node and each of its target neighbor nodes is greater than the second weight threshold, then this node is determined as the core node.

[0363] In some embodiments, the determining subunit is further configured to perform:

[0364] For any node in the maximum spanning tree other than the core node, determine the connected path between the node and any core node in the maximum spanning tree. The connected path refers to the path that can reach the core node from the node through multiple edges.

[0365] Based on the weight of each edge contained in the connected path, the path similarity between the node and the core node is determined.

[0366] In some embodiments, based on Figure 11 The device comprises, wherein the merging subunit includes:

[0367] The merge sub-unit is configured to merge any candidate account group among the multiple candidate account groups if there is a connection edge between the core node of the candidate account group and the core node of other candidate account groups, and the weight of the connection edge is greater than a first weight threshold.

[0368] This merge sub-unit is also configured to perform an alternative, merge the candidate account group with the other candidate account group with the highest group similarity;

[0369] This merge subunit is also configured to repeatedly perform the merge operation on candidate account groups until no candidate account groups can be merged.

[0370] The group filtering sub-unit is configured to perform filtering to obtain the target account group from the account groups obtained from each merge.

[0371] In some embodiments, the merging subunit is further configured to perform:

[0372] For any candidate account group among the multiple candidate account groups, the sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group. The degree parameter of the node represents the number of edges connected to the node in the maximum spanning tree.

[0373] Based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account groups, and the number of common edges in the candidate account group and the other candidate account groups, the group similarity between the candidate account group and the other candidate account groups is determined.

[0374] In some embodiments, the group filtering sub-unit is configured to perform:

[0375] For each merged account group, obtain the modularity of the account group. The modularity is used to measure the quality of the division of the account group from the account relationship graph.

[0376] The group of accounts with the highest modularity is identified as the target group of accounts.

[0377] In some embodiments, the positive interaction behavior includes at least one of the following: expressing a positive opinion about the content item, following the publisher's account of the content item, sharing, or downloading.

[0378] In some embodiments, based on Figure 11 The device comprises:

[0379] The deletion unit is configured to delete authenticated candidate accounts as well as candidate accounts registered with a preset organization.

[0380] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0381] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments concerning the account acquisition method, and will not be elaborated upon here.

[0382] Figure 12 This is a schematic diagram of the structure of a computer device 1200 provided in an embodiment of this disclosure. The computer device 1200 can vary significantly due to differences in configuration or performance. It may include one or more Central Processing Units (CPUs) 1201 and one or more memories 1202. The memory 1202 stores at least one line of program code, which is loaded and executed by the processor 1201 to implement the account acquisition method provided in the various embodiments described above. Of course, the computer device 1200 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The computer device 1200 may also include other components for implementing device functions, which will not be elaborated upon here.

[0383] In an exemplary embodiment, a computer-readable storage medium including at least one instruction is also provided, such as a memory including at least one instruction, which can be executed by a processor in a computer device to complete the account acquisition method in the above embodiments. Optionally, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as ROM (Read-Only Memory), RAM (Random-Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk, and optical data storage device, etc.

[0384] In an exemplary embodiment, a computer program product is also provided, including one or more instructions that can be executed by a processor of a computer device to perform the account acquisition method provided in the above embodiments.

[0385] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0386] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for obtaining an account, characterized in that, include: Based on the sensitive content items published by the reference account, the first account that has performed positive interaction behavior on the sensitive content items is identified. The positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content items. Based on the target content item for which the first account has performed the positive interactive behavior, candidate accounts for publishing the target content item are determined; Based on the reference account and the candidate account, construct the nodes in the account relationship graph; Based on the positive interaction behavior performed by the first account, the edges used to connect nodes are constructed in the account relationship graph. The account relationship graph is used to represent the topological structure of the social relationships between the publisher accounts of the content items for which the first account has performed the positive interaction behavior. Obtain the maximum spanning tree of the account relationship graph, wherein the maximum spanning tree has the maximum weight among the multiple spanning trees of the account relationship graph; Cluster the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups; The multiple candidate account groups are merged to obtain the target account group; The candidate accounts contained in the target account group are identified as the target accounts.

2. The account acquisition method according to claim 1, characterized in that, The construction of edges in the account relationship graph for connecting nodes based on the positive interaction behavior performed by the first account includes: If either the first account has performed the positive interaction behavior on both the sensitive content item published by the reference account and the target content item published by the candidate account, an edge is generated in the account relationship graph to connect the node of the reference account and the node of the candidate account.

3. The account acquisition method according to claim 1, characterized in that, The step of obtaining the maximum spanning tree of the account relationship graph includes: Based on the node similarity between the two nodes connected by each edge in the account relationship graph, the edges are assigned weights. Based on the weights of each edge in the account relationship graph, the maximum spanning tree is generated, wherein the sum of the weights of the edges contained in the maximum spanning tree is the largest among the plurality of spanning trees.

4. The account acquisition method according to claim 3, characterized in that, The node similarity refers to the ratio between the number of common neighbor nodes of the two nodes and the sum of the number of each node's individual neighbor nodes.

5. The account acquisition method according to claim 3, characterized in that, The step of generating the maximum spanning tree based on the weights of each edge in the account relationship graph includes: Initialize an empty tree, which includes a set of nodes and a set of edges. The set of nodes includes a randomly selected starting node, and the set of edges is an empty set. From the account relationship graph, obtain multiple candidate edges for connecting nodes within the node set and nodes outside the node set; Add the target edge with the largest weight among the multiple candidate edges to the edge set, and add the nodes outside the node set connected by the target edge to the node set; Repeat the operation of adding target edges to the edge set until all target edges have been added to the edge set. The spanning tree formed by the edge set and the node set when adding stops is determined as the maximum spanning tree.

6. The account acquisition method according to claim 1, characterized in that, The clustering of accounts indicated by nodes in the maximum spanning tree to obtain multiple candidate account groups includes: Multiple core nodes are obtained by filtering from the nodes contained in the maximum spanning tree; Multiple candidate account groups are determined by using the multiple core nodes as cluster centers; Cluster all nodes in the maximum spanning tree except the core node into the candidate account group containing the core node with the highest path similarity to the core node.

7. The account acquisition method according to claim 6, characterized in that, The step of selecting multiple core nodes from the nodes contained in the maximum spanning tree includes: For any node contained in the maximum spanning tree, determine at least one neighbor node in the maximum spanning tree that is connected to the node by an edge; Among the at least one neighboring node, target neighboring nodes whose weight of the connection edge between the node and the neighboring node is greater than or equal to a first weight threshold are selected; If the sum of the weights of the connecting edges between the node and each of the target neighbor nodes is greater than the second weight threshold, the node is determined as the core node.

8. The account acquisition method according to claim 6, characterized in that, The method further includes: For any node in the maximum spanning tree other than the core node, determine the connected path between the node and any core node in the maximum spanning tree. The connected path refers to the path that can reach the core node from the node through multiple edges. The path similarity between the node and the core node is determined based on the weight of each edge contained in the connected path.

9. The account acquisition method according to claim 1, characterized in that, The process of merging the multiple candidate account groups to obtain the target account group includes: For any candidate account group among the plurality of candidate account groups, if there is a connection edge between the core node of the candidate account group and the core node of other candidate account groups, and the weight of the connection edge is greater than a first weight threshold, the candidate account group and the other candidate account groups are merged. Otherwise, the candidate account group will be merged with other candidate account groups that have the highest group similarity. Repeat the operation of merging candidate account groups until no candidate account groups can be merged, and then select the target account group from the account groups obtained from each merge.

10. The account acquisition method according to claim 9, characterized in that, The method further includes: For any candidate account group among the plurality of candidate account groups, the sum of the degree parameters of the nodes associated with each account in the candidate account group is determined as the group degree parameter of the candidate account group, wherein the degree parameter of the node represents the number of edges connected to the node in the maximum spanning tree; Based on the group degree parameter of the candidate account group, the group degree parameter of the other candidate account groups, and the number of common edges between the candidate account group and the other candidate account groups, the group similarity between the candidate account group and the other candidate account groups is determined.

11. The account acquisition method according to claim 9, characterized in that, The process of filtering the target account group from the account groups obtained from each merge includes: For each merged account group, the modularity of the account group is obtained. The modularity is used to measure the quality of the division of the account group from the account relationship graph. The group of accounts with the highest modularity is identified as the target account group.

12. The account acquisition method according to claim 1, characterized in that, The positive interactive behaviors include at least one of the following: expressing a positive opinion about the content item, following the publisher's account of the content item, sharing, or downloading.

13. The account acquisition method according to claim 1, characterized in that, After determining the candidate accounts for publishing the target content item, the method further includes: Delete verified candidate accounts and candidate accounts registered by preset organizations.

14. An account acquisition device, characterized in that, include: The determining unit is configured to execute sensitive content items published by reference accounts, and determine the first account that has performed positive interaction behavior on the sensitive content items, wherein the positive interaction behavior refers to the interaction behavior of expressing a positive opinion on the sensitive content items; The determining unit is further configured to execute target content items based on the positive interactive behavior performed by the first account, and determine candidate accounts that publish the target content items; The generation unit is configured to construct nodes in the account relationship graph based on the reference account and the candidate account; Based on the positive interaction behavior performed by the first account, the edges used to connect nodes are constructed in the account relationship graph. The account relationship graph is used to represent the topological structure of the social relationships between the publisher accounts of the content items for which the first account has performed the positive interaction behavior. The filtering unit is configured to perform the task of obtaining the maximum spanning tree of the account relationship graph, wherein the maximum spanning tree has the maximum weight among the multiple spanning trees of the account relationship graph; Cluster the accounts indicated by the nodes contained in the maximum spanning tree to obtain multiple candidate account groups; The multiple candidate account groups are merged to obtain the target account group; The candidate accounts contained in the target account group are identified as the target accounts.

15. A computer device, characterized in that, include: One or more processors; One or more memories for storing the one or more processor-executable instructions; The one or more processors are configured to execute the instructions to implement the account acquisition method as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, When at least one instruction in the computer-readable storage medium is executed by one or more processors of a computer device, the computer device is enabled to perform the account acquisition method as described in any one of claims 1 to 13.

17. A computer program product, characterized in that, It includes one or more instructions, which are executed by one or more processors of a computer device, enabling the computer device to perform the account acquisition method as described in any one of claims 1 to 13.

Images