Group risk identification method, device, computer equipment and storage medium
By acquiring and analyzing the data of users in interactive tasks, generating an associated network, and identifying risk groups, the problem of low accuracy in identifying risk groups in the prior art is solved, and the risk control capabilities and fairness of marketing activities are improved.
Patent Information
- Application Number
- CN202111061145.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-10
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-09-10
Smart Images

Figure CN113780413B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a group risk identification method, apparatus, computer equipment and storage medium. Background Art
[0002] At present, Internet platforms attract potential users by launching marketing activities. Some lawless elements participate in such marketing activities by registering members in batches. Such members who register in batches can be called risk groups. In order to avoid wasting resources required for launching marketing activities, it is usually necessary to identify risk groups.
[0003] In the related art, the scheme for identifying risk groups is usually to count the number of members registered under a certain dimension and identify risk groups based on the statistical results. For example, if the number of members registered under the same IP (Internet Protocol) address exceeds a first threshold, it is determined that the members registered under the same IP address belong to the same risk group; or, if the number of members registered in the same time period exceeds a second threshold, it is determined that the members registered in the same time period belong to the same risk group. However, the accuracy of the above scheme in identifying risk groups is low. Summary of the invention
[0004] Based on this, it is necessary to provide a group risk identification method, device, computer equipment and storage medium to address the above technical problems, which can improve the accuracy of risk group identification.
[0005] A method for identifying group risks, the method comprising:
[0006] Acquire user data generated by each user in the process of participating in the corresponding interactive task;
[0007] Extracting features from the user data to obtain a plurality of user features for each of the users;
[0008] According to the corresponding relationship between each of the users and the interactive tasks, feature screening is performed on the plurality of user features to obtain at least two feature set matrices;
[0009] According to each of the feature set matrices, respectively calculating the similarity between the users to obtain at least two similarity matrices;
[0010] generating a correlation network between the users based on a fusion matrix of at least two of the similarity matrices;
[0011] According to the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet, the risk level of the group corresponding to each subnet is identified.
[0012] A group risk identification device, the device comprising:
[0013] An acquisition module, used to acquire user data generated by each user in the process of participating in the corresponding interactive task;
[0014] An extraction module, used to extract features from the user data to obtain a plurality of user features for each of the users;
[0015] A screening module, configured to screen the plurality of user features respectively according to the corresponding relationship between each of the users and the interactive tasks, to obtain at least two feature set matrices;
[0016] A calculation module, used to calculate the similarity between the users according to each of the feature set matrices, to obtain at least two similarity matrices;
[0017] A generating module, configured to generate a correlation network between the users based on a fusion matrix of at least two similarity matrices;
[0018] The identification module is used to identify the risk level of the group corresponding to each subnet according to the number of groups in each subnet in the associated network or the weight of the connecting edge in each subnet.
[0019] In one embodiment, there are different corresponding relationships between each of the users and different of the interactive tasks;
[0020] The screening module is further used to determine at least two similarity conditions according to the corresponding relationship between each of the users and the interactive tasks; and perform feature screening on the plurality of user features based on the at least two similarity conditions.
[0021] In one embodiment, the calculation module is also used to determine the task vector of the interactive task in which each of the users participates based on each of the feature set matrices; determine the number of interactive tasks in which each of the users participates; input the task vector and the number of interactive tasks into a similarity algorithm for similarity calculation to obtain at least two similarity matrices; wherein different similarity matrices are used to reflect the similarity between the users under different similarity conditions.
[0022] In one embodiment, the users include a first user and a second user, and the second user is a user who participates in the interactive task at a target time after the first user; the calculation module is also used to determine the number of interactive tasks participated in by the first user and the second user between the target time and a preset deadline.
[0023] In one embodiment, the screening module is further used to screen each of the users based on at least two of the similarity matrices to filter out users whose similarity is less than a similarity threshold to obtain a target user; the generating module is further used to determine the importance of each of at least two of the similarity conditions; to fuse at least two of the similarity matrices based on the importance to obtain a fused matrix; and to generate an association network for the target user based on the fused matrix.
[0024] In one embodiment, the generating module is further used to construct an association network using the similarity in the fusion matrix as the weight of the connection edge and each of the target users as a network node.
[0025] In one embodiment, the identification module is also used to perform subnet division in the associated network to obtain multiple subnets; determine the number of network nodes in each of the subnets, and the number of network nodes is used to represent the number of groups; or determine the weight of the connecting edges in each of the subnets; and identify the overall risk level corresponding to the corresponding subnet based on the number of network nodes and the weight.
[0026] In one embodiment, the device further comprises:
[0027] A labeling module, configured to label the target users corresponding to some network nodes in the associated network with risk tags based on the risk tags;
[0028] A storage module, used for storing the risk label in a risk label library;
[0029] The identification module is further configured to, when the risk level reaches a risk threshold and there are network nodes that are not risk-labeled in the target subnet corresponding to the risk threshold, identify the risks of the network nodes that are not labeled;
[0030] The updating module is used to update the risk label library if the unlabeled network node is a network node corresponding to a secure user.
[0031] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0032] Acquire user data generated by each user in the process of participating in the corresponding interactive task;
[0033] Extracting features from the user data to obtain a plurality of user features for each of the users;
[0034] According to the corresponding relationship between each of the users and the interactive tasks, feature screening is performed on the plurality of user features to obtain at least two feature set matrices;
[0035] According to each of the feature set matrices, respectively calculating the similarity between the users to obtain at least two similarity matrices;
[0036] generating a correlation network between the users based on a fusion matrix of at least two of the similarity matrices;
[0037] According to the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet, the risk level of the group corresponding to each subnet is identified.
[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:
[0039] Acquire user data generated by each user in the process of participating in the corresponding interactive task;
[0040] Extracting features from the user data to obtain a plurality of user features for each of the users;
[0041] According to the corresponding relationship between each of the users and the interactive tasks, feature screening is performed on the plurality of user features to obtain at least two feature set matrices;
[0042] According to each of the feature set matrices, respectively calculating the similarity between the users to obtain at least two similarity matrices;
[0043] generating a correlation network between the users based on a fusion matrix of at least two of the similarity matrices;
[0044] According to the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet, the risk level of the group corresponding to each subnet is identified.
[0045] The above-mentioned group risk identification method, device, computer equipment and storage medium extract features from user data generated in the process of participating in corresponding interactive tasks, and perform feature screening on the extracted user features according to the corresponding relationship between each user and the interactive task to obtain at least two feature set matrices; calculate the similarity between each user according to each feature set matrix to obtain at least two similarity matrices, and generate an association network between each user based on the fusion matrix of at least two similarity matrices, so as to use the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet to identify the risk level of the group, thereby effectively improving the accuracy of identifying risky users. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 A diagram showing an application environment of a group risk identification method in an embodiment;
[0047] Figure 2 A schematic diagram of a process flow of a group risk identification method in one embodiment;
[0048] Figure 3 is a schematic diagram of an association network in one embodiment;
[0049] Figure 4 is a schematic diagram of a subnet in one embodiment;
[0050] Figure 5 A schematic diagram of a flow chart of a group risk identification method in another embodiment;
[0051] Figure 6 A schematic diagram of the structure of a group risk identification system in one embodiment;
[0052] Figure 7 is a structural block diagram of a group risk identification device in one embodiment;
[0053] Figure 8 is a structural block diagram of a group risk identification device in another embodiment;
[0054] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0056] The group risk identification method provided in this application can be applied to Figure 1 In the application environment shown, the application environment includes a terminal 102, a server 104 for performing task interaction with a user, and a server 106 for identifying group risks. The terminal 102, the server 104 and the server 106 can be connected by a communication connection such as a network.
[0057] The terminal 102 displays an interactive task page, which displays an interactive task (such as a lucky draw). The interactive task can be generated by the server 104 according to the settings of the merchant, and when the user participates in the interactive task, the server 104 will collect the user data of the user and then store it. The server 106 can obtain the user data generated by each user in the process of participating in the corresponding interactive task from the server 104; extract the features of the user data to obtain multiple user features for each user; according to the corresponding relationship between each user and the interactive task, the multiple user features are respectively screened to obtain at least two feature set matrices; according to each feature set matrix, the similarity between each user is calculated to obtain at least two similarity matrices; based on the fusion matrix of at least two similarity matrices, the association network between each user is generated; according to the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet, the risk level of the group corresponding to each subnet is identified.
[0058] Among them, the terminal 102 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to these.
[0059] Server 104 and server 106 can be independent physical servers or service nodes in a blockchain system. The service nodes in the blockchain system form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on the Transmission Control Protocol (TCP) protocol.
[0060] In addition, server 104 and server 106 can also be a server cluster composed of multiple physical servers, and can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), as well as big data and artificial intelligence platforms.
[0061] In one embodiment, Figure 2 As shown in the figure, a group risk identification method is provided, which can be applied to Figure 1 The server 106 in the example is used as an example to illustrate, and the following steps are included:
[0062] S202, obtaining user data generated by each user in the process of participating in the corresponding interactive task.
[0063] Among them, user data may include the account when the user participates in the interactive task (i.e., user account), account registration time, task identifier (ID), task time, device type, IP address, network type, mobile phone model and system version, etc.
[0064] Interaction tasks can be activities that users participate in before interacting with products, through which they can obtain corresponding coupons or other gifts, and then use the coupons or other gifts to interact with products. For example, when a user buys a certain product, he or she can participate in the preferential activities launched by the merchant, so that he or she can get discount coupons, discount coupons or cash coupons, etc., so that when purchasing the product, he or she can use the obtained discount coupons, discount coupons or cash coupons to purchase the product, thereby enjoying the merchant's discount.
[0065] In one embodiment, the server obtains from the database the user data generated by each user in the process of participating in the corresponding interactive task; or, the server samples the user data generated by users in the process of participating in the corresponding interactive task according to different regions in the database, thereby obtaining the user data of users in different regions; or, the server samples the user data generated by users in the process of participating in the corresponding interactive task according to different time periods in the database, thereby obtaining the user data generated by participating in activities at different time periods.
[0066] S204: extract features from the user data to obtain multiple user features for each user.
[0067] Among them, user characteristics may be characteristics that reflect user behavior habits and preferences, and each user may correspond to multiple user characteristics.
[0068] In one embodiment, the server extracts five-tuple features such as task identifier, task time, IP address, device type and network type from the user data, and the five-tuple features are the user features obtained in S204. In addition, the server can also extract other features from the user data, such as extracting five-tuple features such as task identifier, task time, user account, account registration time and device type.
[0069] For example, suppose there are m users and n activities. It is known that some users will participate in one or more activities. At this time, a sparse matrix can be constructed. Suppose this matrix is N∈R m*n Since each activity has user behavior and status, the five-tuple features can be quickly extracted through User and Entity Behavior Analytics (UEBA) analysis. Let I ij is the attribute of user i to activity j, then we can get I ijIt is a five-tuple feature of activity ID, activity time, IP address, device type and network type. The matrix N is used as the training data of the similarity algorithm, and the purpose is to predict users with more user activities and more similar attributes.
[0070] In one embodiment, the server obtains user data generated by users participating in multiple interactive tasks, extracts the behavioral characteristics of users participating in multiple interactive tasks through UEBA analysis, and fully taps the value of user behavior data; overcomes the problems of single feature dimension, low correlation between features and risks, and difficulty in mining with traditional machine learning.
[0071] S206: Screen the multiple user features according to the corresponding relationship between each user and the interactive task to obtain at least two feature set matrices.
[0072] The corresponding relationship can be used to reflect the participation method adopted by the user when participating in the interactive task, and can reflect the number of times the user participates in the task. For example, the corresponding relationship can reflect that the user uses different devices such as mobile phones, desktop computers or tablet computers to participate in the corresponding interactive task, and can also reflect that the user uses different operating systems to participate in the corresponding interactive task, and can also reflect that the user uses virtual accounts or a large number of temporarily newly registered accounts to participate in the corresponding interactive task.
[0073] In one embodiment, different corresponding relationships exist between different users and different interactive tasks; S206 may specifically include: the server determines at least two similarity conditions according to the corresponding relationships between each user and the interactive tasks; and performs feature screening on multiple user features based on the at least two similarity conditions.
[0074] Among them, the similarity condition is used to filter out the user features with similarities between different users, such as filtering out the user features of using virtual accounts to participate in activities, or combining the user features of using newly registered user accounts to participate in activities. The similarity condition can be obtained in the following ways: the similarity condition can be determined according to the number of identical user features, such as a large number of users using newly registered user accounts to participate in activities, and the use of newly registered user accounts to participate in activities can be used as a similarity condition for feature screening; or, the user features can be combined and arranged, and the similarity condition can be determined according to the result obtained after the combination and arrangement; or, the similarity condition can be determined according to the mining demand, and the mining demand can refer to mining out large and small scalpers, equipment farm group crimes and automatons, etc. It should be pointed out that the corresponding relationship reflects the relationship between users and interactive tasks, such as users using mobile operating systems to participate in activities, or using virtual accounts to participate in activities, or participating in activities many times, etc. The similarity condition is a condition abstracted based on these corresponding relationships for filtering the user features corresponding to each user.
[0075] Feature screening refers to screening out user features that meet the corresponding similarity conditions from the user features of all users. For example, if the similarity condition is to mine the user features that participate in the activity using a newly registered account, then the user features of all users who participate in the activity using a newly registered account can be screened out through feature screening.
[0076] The feature set matrix refers to the set of all user features that meet the corresponding similarity conditions. For example, if the similarity condition is to mine the users who participate in the activities with newly registered accounts, then the user features of all users who participate in the activities with newly registered accounts are screened out through feature screening. The matrix composed of these screened user features is the feature set matrix.
[0077] Specifically, the five-tuple features shown by users in activities may be too strict when used as similarity matching. For example, the device farm group may not use the same model of mobile phone or system to commit crimes, and the automaton may also use a proxy IP when committing crimes. Therefore, the server can determine the similarity condition based on the correspondence between users and activities. For example, the most stringent similarity condition is that all the five-tuple features of an activity must match to count as one common activity. This is suitable for finding large and small risk groups; for device farm groups committing crimes, the similarity condition can ignore the mobile phone model and network type, but consider other user features, such as users participating in the activity, if they use virtual accounts or newly registered user accounts to participate in the activity, then the user features of these users are screened; for automatons, the similarity condition can ignore the device model and IP address, but consider the number of common activities (that is, the number of times different users participate in the same activity, such as user a participates in activities 1, 2, and 3, and user b participates in activities 2, 3, and 4, then the number of common activities is 2), that is, the user features corresponding to users participating in the same activity are screened. Therefore, by performing feature screening based on at least two similarity conditions, a feature set matrix corresponding to at least two different similarity conditions can be obtained.
[0078] S208, calculating the similarity between each user according to each feature set matrix, and obtaining at least two similarity matrices.
[0079] The similarity is used to indicate the similarity between users. The elements in the similarity matrix are the similarities between users.
[0080] In one embodiment, S208 may specifically include: the server determines the task vector of the interactive task in which each user participates based on each feature set matrix; determines the number of interactive tasks in which each user participates; inputs the task vector and the number of interactive tasks into a similarity algorithm for similarity calculation to obtain at least two similarity matrices; wherein different similarity matrices are used to reflect the similarity between users under different similarity conditions.
[0081] The task vector may be a vector corresponding to the user's participation in an interactive task, such as a vector corresponding to the user's participation in various promotional activities. The task vector may be determined based on the feature set matrix, and specifically, the feature vectors of each user in the feature set matrix may be quantized to obtain the task vector of the interactive task in which each user participates.
[0082] For example, assuming that the feature set matrix is filtered according to whether the activity is participated in by a virtual account, the user features of each user in the feature set matrix include a virtual account. Therefore, the feature set matrix can be {user a: activity a, 3 times of participation in activities, mobile operating system, domestic IP address, virtual account; user b: activity b, 1 time of participation in activities, mobile operating system, overseas IP address, virtual account}. When vectorizing, different activities are assigned different values (activity a is 1, activity b is 2). When the number of times participating in the activity is less than 2, it is set to 0, and when it is greater than or equal to 2, it is set to 1. It is set to 1 when using a mobile operating system and to 0 when using a desktop system. It is set to 1 when using a domestic IP address and to 0 when using an overseas IP address. It is set to 1 when using a virtual account. Therefore, the task vector of user a can be (1,1,1,1,1), and the task vector of user b can be (2,0,1,0,1).
[0083] The number of interactive tasks can be the number of tasks in which each user participates in different interactive tasks. For example, if user a participates in promotional activities 1 to n, the number of interactive tasks of user a is n. Similarly, if user b participates in promotional activities 1 to i (where i is less than n), the number of interactive tasks of user b is i. Alternatively, the number of interactive tasks can be the number of tasks in which each user participates in different interactive tasks after time alignment. For example, if user a participates in promotional activities 1 to n at time t(0) to t(n), and participates in promotional activities i to n at time t(i) to t(n), the number of interactive tasks of user a is n-i+1. If user b starts to participate in promotional activities at t(i), and participates in promotional activities j to n at t(i) to t(n), the number of interactive tasks of user b is n-j+1.
[0084] In one embodiment, the user includes a first user and a second user, and the second user is a user who participates in the interactive task at a target time after the first user. The above step of determining the number of interactive tasks participated by each user may specifically include: the server determines the number of interactive tasks participated by the first user and the second user from the target time to a preset deadline.
[0085] Among them, since the second user starts to participate in the activity after the first user, the target time can be the time when the second user first participates in the activity, or it can be a certain time after the second user participates in the activity (such as the day after the second user participates in the activity).
[0086] The preset deadline may refer to a set deadline for calculating the number of tasks. When calculating the number of interactive tasks, the time when the first user and the second user participate in the activity are first aligned, that is, the number of interactive tasks participated by the first user and the second user is calculated from the target time until the preset deadline is reached, thereby obtaining the number of interactive tasks participated by the first user and the second user from the target time to the preset deadline.
[0087] For example, assuming that the activity platform holds different activities every day, if user a starts to participate in activities held by the activity platform on date t0 of this month, and user b starts to participate in activities held by the activity platform on date t1 of this month, and assuming that the preset deadline is date t2, when calculating the number of participants in the activities, the dates on which user a and user b participate in the activities are first aligned, that is, both start counting from date t1, that is, the calculated number of activities is the number of participants in the time period from t1 to t2.
[0088] For each activity, five tuple features can be extracted, and the set of these features (i.e., the feature set matrix) shows certain similarities in the same activity and at the same time, and the more activities each user participates in, the greater the difference in the feature set. Therefore, the number of similarities of feature sets of users in different activity scenarios reflects the degree of similarity between users. Assuming that N(u) represents the activity vector participated by user u, and N(v) represents the activity vector participated by user v, the Jaccard algorithm can be used to calculate the similarity.
[0089] For new users or users who join midway, the Jaccard algorithm uses the union of the participating activity sets as the denominator. The matching for new users or users who join midway is always poor. Therefore, some improvements are made here to define L N(u) represents the number of interactive tasks (number of activities) after the activity vector of user u is aligned with the time point when user v first participated in the activity, L N(v)The vector representing the user v’s participation in an activity is aligned with the time point at which user u first participated in the activity (number of activities), so that the similarity W(u,v) between users can be calculated.
[0090] Among them, the improved Jaccard algorithm is as follows:
[0091]
[0092] Among them, τ represents the regularization coefficient of the common activity frequency term, and the value of τ can be 0.01. It can be seen that the improved Jaccard algorithm consists of two parts. The first half is the proportion of common existence, which is the same as the Jaccard similarity coefficient, but the common time point factor is considered in the denominator, and the first activity is aligned, that is, the number of common activities between two users and the maximum number of activities after the first activity is aligned. The larger the proportion, the more common activities occur; the second half represents the value of the common quantity, which makes up for the defect that the first half cannot reflect the advantage of the common quantity, such as s = 4 / 5, y = 16 / 20, although s = y in the common ratio, but the common quantity s is less than y.
[0093] S210, generating a correlation network between users based on a fusion matrix of at least two similarity matrices.
[0094] Among them, the association network depicts the social relationships and economic transactions between corporate customers and between enterprises and natural persons, helping financial institutions discover hidden common characteristics between individuals, thereby locating and identifying fraud members with gang characteristics.
[0095] In one embodiment, before S210, the method further includes: the server screens each user based on at least two similarity matrices to filter out users whose similarity is less than a similarity threshold, and obtains a target user; S210 may specifically include: the server determines the importance of each of the at least two similarity conditions; based on the importance, fuses the at least two similarity matrices to obtain a fusion matrix; generates an association network for the target user based on the fusion matrix, such as Figure 3 shown.
[0096] The improved Jaccard algorithm can be used to calculate several similarity matrices, and then filter out users whose similarity is less than a similarity threshold (such as 0.5), because these users do not show similarity, and it can also save storage space and speed up the calculation. For example, assuming that similarity matrix a is composed of similarities between users 1-10, and similarity matrix b is composed of similarities between users 6-15, then users with similarity greater than 0.5 can be filtered out in similarity matrix a, and users with similarity greater than 0.5 can be filtered out in similarity matrix b, thereby obtaining target users with similarity greater than 0.5.
[0097] It should be noted that, for the screening of target users, in at least two similarity matrices, if the similarity of a user in one of the similarity matrices is greater than a preset threshold, and the similarity in the other similarity matrices is less than the preset threshold, then the user will be screened out as a target user. For example, if the similarity of user 6 in similarity matrix a is greater than 0.5, and the similarity of user 6 in similarity matrix b is less than 0.5, then the screened target users include user 6.
[0098] Specifically, when screening target users, the screening can be performed in the following manner: the server sorts the corresponding users based on at least two similarity matrices to obtain at least two user sets; in each user set, the users whose similarity is less than the similarity threshold are filtered, and the repeated users are filtered out, so as to obtain the target users. For example, assuming that the similarity matrix a is composed of the similarities between users 1-10, and the similarity matrix b is composed of the similarities between users 6-15, users 1-10 and users 6-15 can be sorted in order from large to small in similarity to obtain two sorted user sets. Then, in each user set, the users whose similarity is less than 0.5 are filtered respectively to obtain the users whose similarity is greater than 0.5 in each user set. If the users whose similarity is greater than 0.5 in user set a include users 3-7, and the users whose similarity is greater than 0.5 in user set b include users 6-10, then the repeated users are removed to obtain users 3-10, and the users 3-10 are the target users.
[0099] Since each feature set matrix is calculated under different similarity conditions, these results will be constrained by similarity conditions, resulting in differences in similarity results. For example, the fewer similarity condition constraints, the higher the similarity between users, the more users there are, and the greater the mismatch. At the same time, similarity conditions also have a certain order of importance, that is, the degree of similarity conditions. For example, the number of users with a similarity value greater than 0.5 obtained by satisfying the five-tuple feature set will definitely be less than the number of users with a similarity value greater than 0.5 obtained by the feature set without considering IP attributes, and the number of users with a similarity value greater than 0.5 obtained by the feature set without considering IP attributes will definitely be less than the number of users with a similarity value greater than 0.5 obtained by the feature set that only considers the network type. Therefore, when merging similarity matrices, it is necessary to consider the impact of conditional constraints. The fewer similarity constraints, the greater the penalty.
[0100] In one embodiment, the server sorts the feature set matrices according to the importance of the similarity condition. N(u)∩N(v) is the number of activities shared by users u and v, then is the number of common activities between users u and v in the similarity matrix 1, is the similarity between u and v in the similarity matrix 1, as shown in Table 1:
[0101] Table 1
[0102]
[0103]
[0104] As shown in Table 1, the similarity of users u and v after fusion is:
[0105]
[0106] The improved Jaccard algorithm can only be established when both conditions 1 and 2 are met, and at the same time, the purpose of filtering low similarities is achieved. At this time, the final user association similarity matrix can be obtained.
[0107] In one embodiment, the server fuses at least two similarity matrices to obtain a fused matrix; uses the similarities in the fused matrix as the weights of the connecting edges, and constructs an association network with each target user as a network node.
[0108] Among them, the above-mentioned fusion can refer to summing up the similarity matrices in at least two similarity matrices, then calculating the average value, deleting the similarities whose average values are less than the similarity threshold, and obtaining a fusion matrix composed of the similarities between target users; or, it can also refer to extracting the similarities greater than the similarity threshold in at least two similarity matrices, and forming a new similarity matrix based on the extracted similarities, and the new similarity matrix is the fusion matrix.
[0109] A connection edge may refer to an edge used to connect two different network nodes, such as an edge used to connect a network node corresponding to a target user u and a network node corresponding to a target user v.
[0110] The weight of a connection edge may refer to: the similarity between the target users corresponding to the two network nodes connected by the connection edge (the similarity is the similarity in the fusion matrix). For example, the weight of a connection edge between network node u and network node v may be the similarity between target user u and target user v corresponding to the network node u and network node v.
[0111] S212, identifying the risk level of the group corresponding to each subnet according to the number of groups in each subnet in the associated network or the weight of the connecting edge in each subnet.
[0112] Among them, a subnet can refer to a small network in an associated network, such as Figure 4 As shown. (T means medium black, F means no medium black label)
[0113] In one embodiment, S212 may specifically include: the server performs subnet division in the associated network to obtain multiple subnets; determines the number of network nodes in each subnet, where the number of network nodes is used to represent the number of groups; or determines the weight of the connecting edge in each subnet; and identifies the overall risk level corresponding to the corresponding subnet based on the number of network nodes and the weight.
[0114] Define the association network G = {V, E}, where the network node V is used to represent the user, the connection edge E is used to represent the similarity between users, and the weight of the connection edge E is at least greater than 0.5, and the similarity is the weight. Since the weight of the edge reflects the similarity of the synchronization behavior between users, the number of network nodes (i.e., the number of network nodes) reflects the common similarity between the groups corresponding to each subnet. The more network nodes in the subnet or the larger the weight of the edge, the higher the risk of the group corresponding to the subnet.
[0115] In one embodiment, after S212, the method further includes: when risk labels exist for target users corresponding to some network nodes in the associated network, the server performs risk labeling on some network nodes based on the risk labels; the risk labels are stored in a risk label library; when the risk level reaches a risk threshold, and there are no risk-labeled network nodes in the target subnet corresponding to the risk threshold, risk identification is performed on the unlabeled network nodes; if the unlabeled network nodes are network nodes corresponding to safe users, the risk label library is updated.
[0116] Furthermore, if risk labels are added, the risk level of the group can be obtained more intuitively, and then the risk label library can be updated through the risk level of the subnet and the users marked as normal, thereby improving the accuracy and interpretability of risk judgment.
[0117] The above group risk identification method can be applied to coupon activities, red envelope activities, sign-in activities, points activities, risk prediction, risk visualization and group discovery, etc., which can significantly improve the accuracy of user group discovery, enhance risk control capabilities and the fairness and authenticity of marketing activities. Figure 5 and Figure 6 For further description, Figure 5 As shown, the group identification method includes:
[0118] S1, obtain the data of user participation in activities and obtain the required user data.
[0119] S2, similarity calculation, obtain the similarity matrix.
[0120] S3, user screening and similarity matrix fusion.
[0121] S4, build the association network.
[0122] S5, identify risk groups based on the association network.
[0123] Next, the above group identification method is further explained in combination with the group risk identification system structure, such as Figure 6 As shown, the details are as follows:
[0124] 1) First, obtain the user data generated by users in various activities.
[0125] The user data includes the user's behavior data and basic data. In addition, in the process of acquiring the user, the user's risk label can also be acquired.
[0126] 2) Extract valid five-tuple features such as activity ID, activity time, IP address, device type and network type from user data according to the UEBA analysis method;
[0127] 3) According to the different activities that users participate in every day, extract the five-tuple features of each activity to form a five-tuple feature set;
[0128] 4) Constructing a feature set matrix based on the corresponding relationship between users and activities, such as forming different feature set matrices through different similarity conditions;
[0129] 5) Using a similarity algorithm, the similarity between users is calculated based on each feature set matrix to obtain a similarity matrix;
[0130] 6) The similarity matrices are merged and the similarities less than the similarity threshold are filtered out to obtain the final similarity matrix;
[0131] 7) When users who reach the similarity threshold meet the similarity, a correlation graph can be established. At this time, the group risk can be determined based on the number of network nodes in the subnet;
[0132] 8) Add the user's risk tag to the risk tag library to more accurately extract the group risk situation;
[0133] 9) Feedback users who are individually normal but have high group risk to the risk tag library.
[0134] In the above embodiment, by extracting features from user data generated in the process of participating in corresponding interactive tasks, the extracted user features are screened according to the corresponding relationship between each user and the interactive task to obtain at least two feature set matrices; the similarity between each user is calculated according to each feature set matrix to obtain at least two similarity matrices, and an association network between each user is generated based on a fusion matrix of at least two similarity matrices, thereby using the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet to identify the risk level of the group, thereby effectively improving the accuracy of identifying risky users.
[0135] In common marketing activity scenarios, especially when user behavior data is not abundant, most of them can only obtain the risk status of users through external data, and fail to fully tap the behavioral value generated by users on their own platforms. At the same time, the results generated by calling external interfaces may be general results, lacking consideration of factors such as industry background, activity status, and environment, such as the conversion of black users in marketing activities. Therefore, a robust method is needed to fully tap the data value and improve the risk identification effect by using existing user risk labels and combining them with the behavioral data of marketing activities.
[0136] This application proposes a mining method that combines UEBA and association graphs. In marketing activities, UEBA is used to extract user features, and then the extracted features are used to construct association graphs, and association groups are constructed through association graphs. The advantages of this method are: 1) It avoids the problem of insufficient user behavior data mentioned above, because UEBA is a method for analyzing user behavior and can effectively extract behavioral features; 2) It has strong interpretability. The association graph mines group users through the commonalities between user entities, which is different from the traditional machine learning that mines features based on a single user dimension, which is difficult to interpret; 3) User behavior similarity can be converted into a strong correlation between users. Birds of a feather flock together. When weak features such as IP, mobile phone model, and system version appear to a certain frequency in the same activity and the same time slice, weak associations can gradually be transformed into strong associations; 4) Risk labels can be corrected to improve accuracy and recall. Through user risk labels, risk groups can be extracted, and then the user risk labels can be corrected to improve the accuracy and recall of awakening. In typical application scenarios such as the Spring Festival auto insurance coupon redemption event, this method tracks the behavioral characteristics of users' continuous participation in activities, extracts the five tuple features of UEBA for the time when users participate in each activity to build an association network, and mines group users through weak relationships such as time point, IP, mobile phone model, and activity ID, thereby fully leveraging the value of existing data and improving accuracy.
[0137] It should be understood that although Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.
[0138] In one embodiment, Figure 7 As shown, a group risk identification device is provided, which can adopt a software module or a hardware module, or a combination of the two to become a part of a computer device. The device specifically includes: an acquisition module 702, an extraction module 704, a screening module 706, a calculation module 708, a generation module 710 and an identification module 712, wherein:
[0139] An acquisition module 702 is used to acquire user data generated by each user in the process of participating in a corresponding interactive task;
[0140] An extraction module 704 is used to extract features from user data to obtain multiple user features for each user;
[0141] A screening module 706 is used to screen the multiple user features according to the corresponding relationship between each user and the interactive task, and obtain at least two feature set matrices;
[0142] A calculation module 708, configured to calculate the similarity between each user according to each feature set matrix, and obtain at least two similarity matrices;
[0143] A generating module 710, configured to generate a correlation network between users based on a fusion matrix of at least two similarity matrices;
[0144] The identification module 712 is used to identify the risk level of the group corresponding to each subnet according to the number of groups in each subnet in the associated network or the weight of the connecting edge in each subnet.
[0145] In one embodiment, there are different corresponding relationships between different users and different interactive tasks;
[0146] The screening module 706 is further configured to determine at least two similarity conditions according to the correspondence between each user and the interactive task; and perform feature screening on multiple user features based on the at least two similarity conditions.
[0147] In one embodiment, the calculation module 708 is also used to determine the task vector of the interactive task in which each user participates according to each feature set matrix; determine the number of interactive tasks in which each user participates; input the task vector and the number of interactive tasks into the similarity algorithm for similarity calculation to obtain at least two similarity matrices; wherein different similarity matrices are used to reflect the similarity between users under different similarity conditions.
[0148] In one embodiment, the users include a first user and a second user, and the second user is a user who participates in the interactive task at a target time after the first user; the calculation module 708 is also used to determine the number of interactive tasks participated in by the first user and the second user from the target time to a preset deadline.
[0149] In one embodiment, the screening module 706 is further used to screen each user based on at least two similarity matrices to filter out users whose similarity is less than a similarity threshold, and obtain target users;
[0150] The generation module 710 is further used to determine the importance of each of at least two similarity conditions; fuse at least two similarity matrices based on the importance to obtain a fusion matrix; and generate an association network for the target user based on the fusion matrix.
[0151] In one embodiment, the generating module 710 is further configured to use the similarity in the fusion matrix as the weight of the connection edge and to construct an association network with each target user as a network node.
[0152] In one embodiment, the identification module 712 is further used to perform subnet division in the associated network to obtain multiple subnets; determine the number of network nodes in each subnet, the number of network nodes is used to represent the number of groups; or, determine the weight of the connecting edges in each subnet; and identify the overall risk level corresponding to the corresponding subnet based on the number of network nodes and the weight.
[0153] In one embodiment, Figure 8 As shown, the device also includes:
[0154] The marking module 714 is used to mark the risk of some network nodes based on the risk label when the target users corresponding to some network nodes in the associated network have risk labels;
[0155] A storage module 716, used to store the risk label in a risk label library;
[0156] The identification module 712 is further configured to identify the risks of the unlabeled network nodes when the risk level reaches the risk threshold and there are network nodes that are not labeled for risk in the target subnet corresponding to the risk threshold;
[0157] The updating module 718 is configured to update the risk label library if the unlabeled network node is a network node corresponding to a secure user.
[0158] In the above embodiment, by extracting features from user data generated in the process of participating in corresponding interactive tasks, the extracted user features are screened according to the corresponding relationship between each user and the interactive task to obtain at least two feature set matrices; the similarity between each user is calculated according to each feature set matrix to obtain at least two similarity matrices, and an association network between each user is generated based on a fusion matrix of at least two similarity matrices, thereby using the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet to identify the risk level of the group, thereby effectively improving the accuracy of identifying risky users.
[0159] For the specific definition of the group risk identification device, please refer to the definition of the group risk identification method above, which will not be repeated here. Each module in the above group risk identification device can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0160] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig. 9 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store user data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a group risk identification method is implemented.
[0161] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0162] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.
[0163] In one embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above method embodiments when executed by a processor.
[0164] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned method embodiments.
[0165] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0166] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0167] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A method for identifying group risks, characterized in that: The method comprises: Acquire user data generated by each user in the process of participating in the corresponding interactive task, where different corresponding relationships exist between each user and different interactive tasks; Extracting features from the user data to obtain a plurality of user features for each user, wherein the user features are used to reflect the user's behavior habits and hobbies in participating in corresponding interactive tasks; Determining at least two similarity conditions according to the corresponding relationship between each of the users and the interactive tasks; Based on at least two of the similarity conditions, feature screening is performed on the plurality of user features to obtain at least two feature set matrices; According to each of the feature set matrices, respectively calculating the similarity between the users to obtain at least two similarity matrices; Screening each of the users based on at least two of the similarity matrices to filter out users whose similarity is less than a similarity threshold, and obtaining a target user; determining the importance of each of at least two of the similarity conditions; Based on the importance, at least two similarity matrices are merged to obtain a fusion matrix; generating an association network for the target user based on the fusion matrix; According to the number of groups in each subnet in the association network or the weight of the connecting edge in each subnet, the risk level of the group corresponding to each subnet is identified.
2. The method according to claim 1, characterized in that The similarity condition is used to filter out user features with similarities between different users, and the determining of at least two similarity conditions according to the corresponding relationship between each of the users and the interactive tasks includes: Determine similarity criteria based on the number of identical user features; or Perform combination arrangement based on user characteristics, and determine similarity conditions based on the result of the combination arrangement; or Determine similarity conditions based on mining requirements.
3. The method according to claim 1, characterized in that The calculating the similarity between the users according to each of the feature set matrices to obtain at least two similarity matrices comprises: Determine a task vector of an interactive task in which each user participates according to each of the feature set matrices; Determining the number of interactive tasks in which each of the users participates; Inputting the task vector and the number of interactive tasks into a similarity algorithm to perform similarity calculation to obtain at least two similarity matrices; The different similarity matrices are used to reflect the similarities between the users under different similarity conditions.
4. The method according to claim 3, characterized in that The users include a first user and a second user, wherein the second user is a user who participates in the interactive task at a target time after the first user; Determining the number of interactive tasks in which each user participates includes: The number of interactive tasks in which the first user and the second user participate between the target time and a preset deadline is determined.
5. The method according to claim 2, characterized in that: The user is screened based on at least two similarity matrices to filter out users whose similarity is less than a similarity threshold, and the target users include: In at least two similarity matrices, users whose similarity in one similarity matrix is greater than a preset threshold and whose similarity in the other similarity matrix is less than the preset threshold are screened out as target users.
6. The method according to claim 5, characterized in that Generating the association network for the target user based on the fusion matrix includes: The similarity in the fusion matrix is used as the weight of the connection edge, and the target users are used as network nodes to build an association network.
7. The method according to any one of claims 1 to 6, characterized in that The step of identifying the risk level of the group corresponding to each subnet according to the number of groups in each subnet in the associated network or the weight of the connecting edge in each subnet includes: Performing subnet division in the associated network to obtain a plurality of subnets; Determine the number of network nodes in each of the subnetworks, where the number of network nodes is used to represent the number of groups; or determine the weight of the connecting edge in each of the subnetworks; The risk level of the entire subnet is identified based on the number of network nodes and the weight.
8. The method according to any one of claims 1 to 6, characterized in that After identifying the risk level of the group corresponding to each of the subnets, the method further includes: When target users corresponding to some network nodes in the associated network have risk labels, risk marking the some network nodes based on the risk labels; Storing the risk tag in a risk tag library; When the risk level reaches a risk threshold, and there are network nodes without risk marking in the target subnet corresponding to the risk threshold, risk identification is performed on the network nodes without marking; If the unlabeled network node is a network node corresponding to a secure user, the risk label library is updated.
9. A group risk identification device, characterized in that: The device comprises: An acquisition module, used to acquire user data generated by each user in the process of participating in a corresponding interactive task, wherein different corresponding relationships exist between each user and different interactive tasks; An extraction module, used to extract features from the user data to obtain a plurality of user features for each user, wherein the user features are used to reflect the behavior habits and hobbies of the user in participating in the corresponding interactive task; A screening module, configured to determine at least two similarity conditions according to the correspondence between each of the users and the interactive tasks, and perform feature screening on the plurality of user features based on the at least two similarity conditions to obtain at least two feature set matrices; A calculation module, used for calculating the similarity between each of the users according to each of the feature set matrices, to obtain at least two similarity matrices; the screening module is also used for screening each of the users based on at least two of the similarity matrices, to filter out users whose similarity is less than a similarity threshold, to obtain target users; A generating module, used to determine the importance of each of at least two of the similarity conditions, fuse at least two of the similarity matrices based on the importance to obtain a fusion matrix, and generate an association network for the target user based on the fusion matrix; The identification module is used to identify the risk level of the group corresponding to each subnet according to the number of groups in each subnet in the associated network or the weight of the connecting edge in each subnet.
10. The device according to claim 9, characterized in that The similarity condition is used to filter out user features that are similar between different users. When at least two similarity conditions are determined according to the corresponding relationship between each of the users and the interactive tasks, the screening module is further used to: Determine similarity criteria based on the number of identical user features; or Perform combination arrangement based on user characteristics, and determine similarity conditions based on the result of the combination arrangement; or Determine similarity conditions based on mining requirements.
11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.
12. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.
13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.
