A population identification method, device, equipment and computer readable storage medium

Abstract

The application discloses a kind of crowd identification method, device, equipment and computer readable storage medium, applied to computer technology field, comprising: according to crowd case feature library, the basic data of historical customer is obtained, the basic data of historical customer is converted into historical customer feature, the historical customer feature adjacency matrix of historical customer and historical customer feature association is obtained, and the adjacency matrix index table of historical customer and historical customer feature corresponding relationship is constructed;According to the customer feature of the client to be investigated and historical customer feature adjacency matrix, seed guest group is obtained by screening;According to adjacency matrix index table, and the density block threshold value calculated by preset density block size, seed guest group is operated in two-way propagation, when meeting convergence condition, the customer set after convergence is used as target group;Two-way propagation operation is the selection operation of customer to feature and the selection operation of feature to customer.The application improves the accuracy, efficiency, adaptability and explainability of crowd identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a group identification method, device and equipment and computer readable storage medium. BACKGROUND

[0002] Traditional identification methods rely on expert rules, blacklists and basic statistics. Although machine learning models are mainstream, they all focus on single customer dimension and are difficult to mine complex correlations and hidden information of group abnormal behavior. At present, group identification methods revolve around correlation networks. For example, a multi-dimensional relationship network is constructed based on inventory customer data, and a multi-level traversal is performed to mark risky customers, but the correlation network is prone to low traversal efficiency due to data redundancy; or deep relationships in the multi-dimensional relationship network are dug by using a graph algorithm, but the graph algorithm is not adaptive to a small number of dynamically changing abnormal group relationships; or a classification model is constructed to predict risky customers by converting network structure through graph coding, but the probability value output by the model lacks explainability and has limited ability to identify a small number of dynamically changing abnormal groups.

[0003] Therefore, how to improve the accuracy, efficiency, adaptability and explainability of abnormal group identification has become a core technical problem that needs to be solved in the industry. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a group identification method, device and equipment and computer readable storage medium, which solves the problems of low efficiency, poor accuracy, insufficient adaptability and lack of explainability of the abnormal group identification method in the prior art.

[0005] To solve the above technical problems, the present application provides a group identification method, comprising:

[0006] Obtaining basic data of historical customers from a group case feature library, and converting the basic data of the historical customers into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers who have business transactions with automobile institutions;

[0007] Based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features is constructed;

[0008] Obtaining a to-be-investigated customer, and screening a seed customer group according to the customer features of the to-be-investigated customer and the historical customer feature adjacency matrix; the to-be-investigated customer is a customer with confirmed information doubts in automobile business;

[0009] According to the adjacency matrix index table and a density block threshold calculated based on a preset density block size, the seed customer group is subjected to an iterative bidirectional propagation operation, and when a convergence condition is met, a converged customer set is taken as the target group; the bidirectional propagation operation is a customer-to-feature selection operation and a feature-to-customer selection operation; the convergence condition is that the customer set and the feature set obtained by two adjacent iterative bidirectional propagation operations are completely consistent;

[0010] The density block threshold calculation formula is:

[0011] ;

[0012] The density block density threshold The calculation formula is as follows:

[0013] ;

[0014] S i is the current customer set of the i th round; F i is the current feature set of the i th round; and both represent the average density corresponding to the historical customer feature adjacency matrix A; represents the association parameter of customer i and feature j, which is called an edge; represents the density block size; M is the number of historical customers; N is the number of historical customer features.

[0015] Optionally, a to-be-investigated customer is obtained, and a seed customer group is screened according to a customer feature of the to-be-investigated customer and the historical customer feature adjacency matrix, including:

[0016] Basic data of the to-be-investigated customer are obtained, and the basic data of the to-be-investigated customer are converted into the customer feature;

[0017] The customer feature is compared with the historical customer feature in the historical customer feature adjacency matrix in terms of feature similarity, the first x customers with feature similarity greater than a preset threshold are screened, and the first x customers are taken as the seed customer group.

[0018] Optionally, according to the adjacency matrix index table and a density block threshold calculated based on a preset density block size, the seed customer group is subjected to an iterative bidirectional propagation operation, and when a convergence condition is met, a converged customer set is taken as the target group, including:

[0019] Step 1: A density block density threshold is calculated based on a preset density block size, and the density block threshold is calculated based on the density block density threshold, a current customer set or a current customer feature set;

[0020] Step 2: based on the adjacency matrix index table, count the first associated edge number of each feature in the customer set, filter the features with the first associated edge number greater than the density block threshold, and form the feature set; the initial set of the customer set is the seed customer group; the first associated edge number is the number of customers associated with each feature;

[0021] Step 3: if the feature set does not satisfy the first judgment condition, output an empty set and terminate; the first judgment condition is , F represents the feature set;

[0022] Step 4: if the feature set satisfies the first judgment condition, based on the adjacency matrix index table, count the second associated edge number of each customer in the feature set, filter the customers with the second associated edge number greater than the density block threshold, and form a new customer set; the second associated edge number is the number of features associated with each customer;

[0023] Step 5: if the new customer set does not satisfy the second judgment condition, output an empty set and terminate; the second judgment condition is , S represents the customer set;

[0024] Step 6: if the new customer set satisfies the second judgment condition, repeat the steps 1 to 6 until the convergence condition is satisfied, and terminate, and the converged customer set is taken as the target group.

[0025] Optionally, when the convergence condition is satisfied, the converged customer set is taken as the target group, comprising:

[0026] When the customer set and the feature set obtained by the bidirectional propagation operation of the adjacent two iterations are completely consistent, the customer set is taken as the target group.

[0027] Optionally, the basic data of the historical customer is obtained according to the mass case feature library, and the basic data of the historical customer is converted into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customer and the historical customer features, comprising:

[0028] The customer information of the abnormal case record is obtained, and based on the customer information, the typical features vulnerable to attack are obtained by analysis, and the mass case feature library is formed;

[0029] Based on the feature dimension of the mass case feature library, the basic data of the historical customer is extracted from the business system;

[0030] Based on the typical features of the mass case feature library, the basic data of the historical customer is converted into quantifiable historical customer features;

[0031] The historical customer features are classified according to data types, and a historical customer feature adjacency matrix is constructed based on the processed results.

[0032] Optionally, the historical customer features are classified according to data types, and a historical customer feature adjacency matrix is constructed based on the processed results, including:

[0033] A first initial Boolean matrix of historical customers and discrete features is constructed, positions in the first initial Boolean matrix that match discrete features in the historical customer features are valued as 1, and positions that do not match are valued as 0, to obtain a first Boolean matrix;

[0034] The numerical features in the historical customer features are binned according to a preset rule, and a second initial Boolean matrix of historical customers and feature bins is constructed according to the binning results; positions in the second initial Boolean matrix that correspond to the binning features of the numerical features are valued as 1, and positions that do not correspond are valued as 0, to obtain a second Boolean matrix;

[0035] The date features in the historical customer features are binned according to a year-month-week format, and a third initial Boolean matrix of historical customers and time bins is constructed according to the binning results; positions in the third initial Boolean matrix that correspond to the binning features of the date features are valued as 1, and positions that do not correspond are valued as 0, to obtain a third Boolean matrix;

[0036] The first Boolean matrix, the second Boolean matrix, and the third Boolean matrix are spliced along the feature dimension to form the historical customer feature adjacency matrix containing M historical customers and N historical customer features; the rows of the historical customer feature adjacency matrix correspond to the historical customers, and the columns correspond to the historical customer features.

[0037] Optionally, based on the historical customer feature adjacency matrix, an adjacency matrix index table of the historical customer and the corresponding relationship of the historical customer features is constructed, including:

[0038] The historical customer feature adjacency matrix is traversed by row, index information of features hit by each historical customer is extracted, and the adjacency matrix index table is generated in the form of historical customer index numbers and feature index numbers as edges.

[0039] The application further provides a population identification device, including:

[0040] An adjacency matrix construction module is configured to obtain basic data of historical customers from a mass case feature library, convert the basic data of the historical customers into historical customer features, and obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers having business transactions with automobile institutions.

[0041] an index table construction module, configured to construct an index table of an adjacency matrix of the historical customers and the historical customer features based on the historical customer feature adjacency matrix;

[0042] a seed customer group acquisition module, configured to acquire customer features of a customer to be investigated, and filter a seed customer group according to the historical customer feature adjacency matrix and the customer to be investigated; the customer to be investigated is a customer with confirmed information doubts in automobile business;

[0043] a target group determination module, configured to perform an iterative bidirectional propagation operation on the seed customer group according to the index table of the adjacency matrix and a density block threshold value calculated based on a preset density block size, and take a customer set after convergence as a target group when a convergence condition is met; the bidirectional propagation operation is a selection operation of customers to features and a selection operation of features to customers; the convergence condition is that a customer set and a feature set obtained by two adjacent iterative bidirectional propagation operations are completely consistent;

[0044] the density block threshold value calculation formula is as follows:

[0045] ;

[0046] a density block density threshold value the calculation formula is as follows:

[0047] ;

[0048] S i is a current customer set of the i th round; F i is a current feature set of the i th round; and both represent an average density corresponding to the historical customer feature adjacency matrix A; represents an association parameter of customer i and feature j, and is referred to as an edge; represents a density block size; M is the number of historical customers; and N is the number of historical customer features.

[0049] The application further provides a population identification device, which comprises:

[0050] a memory, configured to store a computer program;

[0051] a processor, configured to execute the computer program to realize the population identification method.

[0052] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores computer executable instructions; when the computer executable instructions are loaded and executed by a processor, the population identification method is realized.

[0053] It can be seen that the application obtains basic data of historical customers according to the group case feature library, converts the basic data of the historical customers into historical customer features, and obtains a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers having business dealings with the automobile mechanism; based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features is constructed; a to-be-investigated customer is obtained, and a seed customer group is obtained by screening according to customer features of the to-be-investigated customer and the historical customer feature adjacency matrix; the to-be-investigated customer is a customer having confirmed information doubts in automobile business; according to the adjacency matrix index table and a density block threshold value calculated based on a preset density block size, an iterative bidirectional propagation operation is performed on the seed customer group, and when a convergence condition is met, a converged customer set is taken as a target group; the bidirectional propagation operation is a customer-to-feature selection operation and a feature-to-customer selection operation; the convergence condition is that customer sets and feature sets obtained by two adjacent iterative bidirectional propagation operations are completely consistent; and a calculation formula of the density block threshold value is as follows: A calculation formula of the density block density threshold value is as follows: i is a current customer set of the i th round; F i is a current feature set of the i th round; and both represent an average density corresponding to the historical customer feature adjacency matrix A; represents an association parameter of the customer i and the feature j, and is referred to as an edge; represents a density block size; M is the number of the historical customers; and N is the number of the historical customer features.The application has the beneficial effects that: relying on the group case feature library to anchor core abnormal features, the associated group of the to-be-investigated customers is mined through the adjacency matrix and the bidirectional propagation, normal associated customer groups are avoided from being misjudged and missed, and the identification accuracy is improved, which is not only associated with the explainability of the association rules, but also has greater flexibility than the association rules; the adjacency matrix index table can simplify the association data of the customers and the features, reduce redundant calculation, the iterative bidirectional propagation does not need full traversal, the speed of screening the abnormal groups is accelerated, and the identification efficiency is improved; the abnormal group association is diffused and identified from the seed customer group to the associated customers, the group abnormal behavior can be effectively captured, and the actual business demand is met; based on the basic data conversion and the matrix operation, complex graph algorithms or model training are not needed, the algorithm dependence is reduced, the landing threshold is lowered, and the application in the business system is facilitated.

[0054] In addition, the application also provides a group identification device, equipment and computer readable storage medium, which also have the beneficial effects. BRIEF DESCRIPTION OF DRAWINGS

[0055] ​In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only aim at the embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of the provided drawings.

[0056] Figure 1 A flowchart of a population identification method provided for the embodiments of the present application;

[0057] Figure 2 An example diagram of a data preparation process provided for the embodiments of the present application;

[0058] Figure 3 An example diagram of a flow of a bidirectional propagation algorithm based on feature density provided for the embodiments of the present application;

[0059] Figure 4 An example diagram of a flow based on a deployment stage provided for the embodiments of the present application;

[0060] Figure 5 A structural schematic diagram of a population identification device provided for the embodiments of the present application;

[0061] Figure 6 A structural schematic diagram of a population identification device provided for the embodiments of the present application. DETAILED DESCRIPTION

[0062] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only aim at the embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of the provided drawings.

[0063] Traditional anomaly identification methods mainly rely on rules, blacklists and other expert experience and statistical methods based on expert experience and statistical methods. With the development of technology, machine learning models have gradually become mainstream. However, the above methods are mainly developed around a single customer dimension, and it is difficult to mine the complex correlations and hidden information between group behaviors. The current industry mainstream group identification method mainly includes:

[0064] (1) Graph traversal: based on the inventory of historical customers, an association network containing various association relationships is constructed, and multi-level traversal is performed on the association network for the to-be-detected customers, and the to-be-detected customers are associated to the marked risk nodes. However, the traversal complexity increases with the increase of the number of jumps in a polynomial complexity, so the traversal depth is limited.

[0065] (2) Graph algorithm: By pre-mining sub-graphs, the complexity of graph traversal is simplified, but the construction accuracy of sub-graphs is limited, and it also lacks interpretability.

[0066] (3) Graph encoding: In order to improve the computational efficiency of graph data and make the most of the information in the graph, the graph data structure is re-encoded to train a classification model, and the classification model is used to predict the risk of customers and determine whether the customer is a member of the abnormal group. However, the probability value output by the classification model lacks interpretability, and the ability to identify a small number of dynamically changing abnormal groups is limited.

[0067] In order to solve the above problems, the present application provides a group identification method, according to the characteristics of the group case, the relevant data is obtained from the business system, the data feature engineering is carried out, the adjacency matrix of the customer and the feature is constructed, the feature value of the current customer to be investigated is taken as the reference, a number of customers with high similarity of features are screened out in the adjacency matrix as seed customer groups, and the feature density propagation method is used to mine high-density customer feature blocks, and finally the customer groups in the block are output as the target group identified. Please refer to Figure 1 , Figure 1 The flowchart of a group identification method provided by the embodiment of the present application. The method can include:

[0068] S101: Obtain the basic data of historical customers according to the group case feature library, and convert the basic data of historical customers into historical customer features to obtain a historical customer feature adjacency matrix associated with historical customers and historical customer features.

[0069] The execution subject of the embodiment is a terminal. The embodiment does not limit the type of terminal as long as it can complete the operation of the group identification method. It should be noted that the application scenario of the embodiment is the field of automobile consumption. The historical customers in the embodiment are customers who have business dealings with automobile institutions, which can also be understood as inventory customers, that is, customer groups that have business relationships with automobile enterprises / institutions.

[0070] The group case in the embodiment refers to a centralized fraud behavior launched by batch counterfeiting information and using a unified approach to a certain type of business (such as automobile loan submission) in the field of automobile consumption, resulting in a large number of case clusters with similar abnormal characteristics in a short period of time.

[0071] Further, the above-mentioned obtaining the basic data of historical customers according to the group case feature library, and converting the basic data of historical customers into historical customer features to obtain a historical customer feature adjacency matrix associated with historical customers and historical customer features can specifically include the following steps:

[0072] Step 11: Obtain customer information of abnormal application records, and analyze based on the customer information to obtain typical characteristics of vulnerable attacks, and form a mass case feature library;

[0073] Step 12: Based on the feature dimensions of the mass case feature library, extract the basic data of historical customers from the business system;

[0074] Step 13: Based on the typical characteristics of the mass case feature library, convert the basic data of historical customers into quantifiable historical customer characteristics;

[0075] Step 14: Classify and process the historical customer characteristics according to the data type, and construct a historical customer characteristic adjacency matrix based on the processed results.

[0076] The abnormal application in this embodiment is a business request in the automobile consumption field that does not conform to normal rules or has risk hidden dangers. For example, the automobile consumption loan application deviates from the conventional and compliant range in terms of process, content, characteristics, etc., and may imply false information, false behavior or operational errors, etc.

[0077] The credit investigation experts collect and organize the customer application information with abnormal points, evaluate possible concentrated false behaviors (fraudulent behaviors) through induction and summary, and summarize the key characteristics of the customers, such as consistent vehicle models, high vehicle prices, unified transaction schemes, out-of-town car purchases without reasonable explanations, concentration in certain channels, unreasonable application positioning, similar loan application times, etc. to form a mass case feature library. Based on the mass case feature library, the related basic data of historical customers is collected from the business system, such as vehicle model information, loan information, customer address information, application positioning information, car purchase channel and salesperson information, and credit derivative indicators. Further, the related basic data is converted into the required features, i.e. based on the corresponding features of the historical customers, i.e. historical customer characteristics, such as whether the vehicle price is high or not, whether the car is purchased out of town, etc. Finally, the historical customer characteristics are classified and processed according to the data type (such as discrete type, numerical type, etc.), i.e. different data types are processed by adopting the appropriate processing method. Based on the processed features, a feature adjacency matrix (historical customer feature adjacency matrix) is constructed to reflect the association between the customers (historical customers) and the corresponding features (historical customer characteristics).

[0078] Further, the above classification and processing of the historical customer characteristics according to the data type, and the construction of the historical customer feature adjacency matrix based on the processed results can include the following steps:

[0079] Step 411: Construct a first initial Boolean matrix of historical customers and discrete features, take the value of 1 at the position matching the discrete features in the historical customer characteristics in the first initial Boolean matrix, and take the value of 0 otherwise, to obtain a first Boolean matrix;

[0080] Step 412: The numerical features in the historical customer features are processed according to the preset rule, and a second initial Boolean matrix of the historical customer and feature binning is constructed according to the binning result; the position corresponding to the binning feature to which the numerical feature belongs in the second initial Boolean matrix is taken as 1, otherwise 0, to obtain a second Boolean matrix;

[0081] Step 413: The date type features in the historical customer features are processed according to the year-month-week format, and a third initial Boolean matrix of the historical customer and time binning is constructed according to the binning result; the position corresponding to the binning feature to which the date type feature belongs in the third initial Boolean matrix is taken as 1, otherwise 0, to obtain a third Boolean matrix;

[0082] Step 414: The first Boolean matrix, the second Boolean matrix and the third Boolean matrix are spliced along the feature dimension to form a historical customer feature adjacency matrix containing M historical customers and N historical customer features; the row of the historical customer feature adjacency matrix corresponds to the historical customer, and the column corresponds to the historical customer feature.

[0083] In order to better understand the above steps, the following examples can be referred to: (1) the processing method for discrete data is: taking the car purchase channel as an example, each channel is taken as a feature, and a Boolean matrix space B m*n1 of customers and channels is constructed, where m is the customer, n1 is the channel, and 1 is taken when the customer belongs to the channel, otherwise 0 is taken. (2) The processing method for numerical data is: taking the total loan amount as an example, the total loan amount is binned, each bin is taken as a feature, and a Boolean matrix space B m*n2 of customers and total loan amount bins is constructed, where m is the customer, n2 is the total loan amount bin, and 1 is taken when the customer's total loan amount belongs to the bin, otherwise 0 is taken. (3) The processing method for date type data is: taking the loan application time as an example, the loan application time is processed into year-month-week (such as 2025 June 1st week), each year-month-week is taken as a feature, and a Boolean matrix space B m*n3 of customers and loan application time bins is constructed, where m is the customer, n3 is the loan application time year-month-week, and 1 is taken when the customer belongs to the week, otherwise 0 is taken. Finally, the above Boolean matrices are horizontally spliced to obtain the final adjacency matrix A M*N is saved, where M is the number of historical customers, and N is the number of all features, i.e. the number of historical customer features. The historical customer feature adjacency matrix A can be obtained according to the above method.

[0084] S102: Based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customer and the historical customer feature is constructed.

[0085] The embodiment extracts and retains the effective association relationship between the customers and the features (such as which customers have which features) from the historical customer feature adjacency matrix A, arranges the association in the form of an index into a concise comparison table, that is, obtains the adjacency matrix index table E, and uses the adjacency matrix index table E to quickly query the corresponding relationship between the customers and the features.

[0086] Further, the above-mentioned adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features based on the historical customer feature adjacency matrix can specifically include the following steps: traversing the historical customer feature adjacency matrix by row, extracting the index information of the hit features of each historical customer, and generating the adjacency matrix index table in the form of the historical customer index number and the feature index number as the edge e ij Further, the above-mentioned adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features based on the historical customer feature adjacency matrix can specifically include the following steps: traversing the historical customer feature adjacency matrix by row, extracting the index information of the hit features of each historical customer, and generating the adjacency matrix index table in the form of the historical customer index number and the feature index number as the edge e

[0087] S103: Obtain a to-be-investigated customer, and screen a seed customer group according to the customer feature of the to-be-investigated customer and the historical customer feature adjacency matrix.

[0088] The to-be-investigated customer in the embodiment is a customer in the automobile business who has confirmed information doubts. The embodiment first obtains the customer to be investigated, and then screens the customer who meets the condition as the seed customer group by analyzing the association (such as the matching degree, the association strength, etc.) of the features in the historical customer feature adjacency matrix.

[0089] Further, the above-mentioned obtaining the to-be-investigated customer and screening the seed customer group according to the to-be-investigated customer and the historical customer feature adjacency matrix can specifically include the following steps:

[0090] Step 21: Obtain the basic data of the to-be-investigated customer, and convert the basic data of the to-be-investigated customer into customer features.

[0091] Step 22: Compare the customer features with the historical customer features in the historical customer feature adjacency matrix in terms of feature similarity, screen the top x customers with the feature similarity greater than a preset threshold, and take the top x customers as the seed customer group.

[0092] In the embodiment, the customer to be investigated is collected from the business system to obtain the related basic information and processed into the data structure of the above-mentioned historical customer feature adjacency matrix, that is, the customer to be investigated is converted into the customer feature in the form of , y is the number of the to-be-investigated customers. In the historical customer feature adjacency matrix A, the to-be-investigated customer feature similarity is sorted from high to low, and the historical customer index number of the top x customers is taken as the seed customer group If the similarity is too low, there is no seed customer group, and an empty set is returned. The similarity calculation formula is as follows:

[0093] .

[0094] where I is an indicator function; , indicates that the ith customer has the jth feature; M is the number of rows, i.e., the number of historical customers; N is the number of columns, i.e., the number of historical customer features; Sim is a similarity value; the index set of the customer to be investigated ; indicates a continuous multiplication.

[0095] S104: According to the adjacency matrix index table and the density block threshold value calculated based on the preset density block size, a bidirectional propagation operation is performed on the seed customer group iteratively, and when a convergence condition is met, a converged customer set is taken as a target group.

[0096] In this embodiment, when the seed customer group is obtained in step S103, a high-density feature block mining algorithm is used to calculate a feature-dense target group S i , which is fed back to the business system as output. The high-density feature block mining algorithm is a bidirectional propagation mining algorithm based on feature density. The bidirectional propagation operation refers to a customer-to-feature selection operation and a feature-to-customer selection operation. In this embodiment, the density block threshold value can be obtained by defining a density block size , a density block density threshold calculation formula, and a density block threshold value calculation formula, to determine whether to stop the propagation operation.

[0097] The calculation formula of the density block density threshold value is as follows:

[0098] .

[0099] The density block threshold value calculation formula is:

[0100] .

[0101] where S i is the current customer set of the ith round; F i is the current feature set of the ith round; and both represent the average density corresponding to the historical customer feature adjacency matrix A; indicates the association parameter of customer i and feature j, which can be referred to as an edge.

[0102] Based on the adjacency matrix index table E obtained in step S102 and the density block threshold value d obtained by the above calculation formula, the seed customer group The iterative bidirectional propagation operation is performed until a convergence condition or a stop condition is met, and then the operation is ended. In service deployment, the historical customer feature adjacency matrix A obtained in step S101 and the adjacency matrix index table E obtained in step S102 can be loaded into the memory, and the input and output formats of the service can be defined by using an API (Application Programming Interface) mode.

[0103] Further, the seed customer group is subjected to the iterative bidirectional propagation operation according to the adjacency matrix index table and the density block threshold value calculated based on the preset density block size, and when the convergence condition is met, the converged customer set is taken as the target group, which can specifically include the following steps:

[0104] Step 1: The density block density threshold value is calculated based on the preset density block size, and the density block threshold value is calculated based on the density block density threshold value, the current customer set or the current customer feature set;

[0105] Step 2: Based on the adjacency matrix index table, the first associated edge number of each feature in the customer set is counted, the features with the first associated edge number greater than the density block threshold value are screened, and a feature set is formed; the initial set of the customer set is the seed customer group;

[0106] Step 3: If the feature set does not satisfy the first judgment condition, an empty set is output and the operation is terminated;

[0107] Step 4: If the feature set satisfies the first judgment condition, the second associated edge number of each customer in the feature set is counted based on the adjacency matrix index table, the customers with the second associated edge number greater than the density block threshold value are screened, and a new customer set is formed;

[0108] Step 5: If the new customer set does not satisfy the second judgment condition, an empty set is output and the operation is terminated;

[0109] Step 6: If the new customer set satisfies the second judgment condition, steps 1 to 5 are repeatedly executed until the convergence condition is met, and the converged customer set is taken as the target group.

[0110] The embodiment first starts from an initial seed customer group, counts the number of customers associated with each feature, and filters a feature set based on a density block threshold requirement; if the feature set does not meet the requirement, the process is terminated and an empty set is returned; if the feature set meets the requirement, the process continues to propagate, counts the number of features associated with each customer based on the feature set, and filters a new customer set based on a density block threshold requirement; if the new customer set does not meet the requirement, the process is terminated and an empty set is returned; if the new customer set meets the requirement, the process repeats the steps of selecting features from the customer group and selecting customers from the feature group (i.e., the steps of bidirectional propagation) until the result is stable (converged), and the final customer group is the target group. It should be noted that the first number of associated edges in the embodiment refers to the number of customers associated with each feature; the first judgment condition refers to , F represents the feature set; the second number of associated edges refers to the number of features associated with each customer; the second judgment condition refers to , S represents the customer set; m and n represent the preset number of customers and the preset number of features in the density block , which can be set according to actual conditions. The convergence condition refers to the customer set and the feature set obtained by the bidirectional propagation operation of the two adjacent iterations being completely consistent.

[0111] Specifically, the bidirectional propagation in the embodiment includes:

[0112] (1) The density block threshold required for adding the current customer or the current customer feature to the current customer set or the current feature set in each propagation is defined as .

[0113] (2) Propagation of the customer set S to the feature set F S2F(S, d): initialize F as an empty set, calculate the number of edges c of each feature f in the customer set S, select f with the number of edges greater than the density block threshold d, and return . If , the density block requirement is not met, an empty set is output, and the algorithm is ended.

[0114] (3) Propagation of the feature set F to the customer set S F2S(F, d): initialize S as an empty set, calculate the number of edges c of each customer s in F, select s with the number of edges greater than the density block threshold d, and return . If , the density block requirement is not met, an empty set is output, and the algorithm is ended.

[0115] (4) Repeat until convergence, i.e., S i =S i -1 and F i =F i -1, S i ​​As the customer set output in the high-density feature block. That is, when the customer set and the feature set obtained by the two-iterative bidirectional propagation operation are completely consistent, the customer set is taken as the target group.

[0116] By applying the group identification method provided by the embodiment of the application, S101, the basic data of the historical customer is obtained according to the group case feature library, and the basic data of the historical customer is converted into the historical customer feature to obtain the historical customer feature adjacency matrix associated with the historical customer; S102, based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customer and the historical customer feature is constructed; S103, the customer to be investigated is obtained, and the seed customer group is obtained by screening according to the customer feature of the customer to be investigated and the historical customer feature adjacency matrix; S104, according to the adjacency matrix index table and the density block threshold value calculated based on the preset density block size, the seed customer group is subjected to an iterative bidirectional propagation operation, and when the convergence condition is met, the converged customer set is taken as the target group. This method relies on the group case feature library to anchor the core abnormal features, and mines the associated group of the customer to be investigated through the adjacency matrix and bidirectional propagation, avoids misjudgment and omission of normal associated customers, and improves the identification accuracy. This method has the explainability of the association rule and greater flexibility than the association rule. Moreover, the adjacency matrix index table can simplify the associated data of the customer and the feature, reduce redundant calculation, and accelerate the abnormal group screening speed and improve the identification efficiency through the iterative bidirectional propagation without full traversal. In addition, the abnormal group association is identified from the seed customer group to the associated customer, which can effectively capture the group abnormal behavior and meet the actual business needs. Moreover, based on the basic data conversion and matrix operation, the algorithm power dependence is reduced, the landing threshold is lowered, and the application in the business system is facilitated.

[0117] In order to make the application more convenient to understand, it can be divided into several parts for example description:

[0118] The first part is the data preparation stage, which can be referred to in detail in Figure 2 , Figure 2 The example diagram of the data preparation process provided by the embodiment of the application.

[0119] 1. Group case customer feature analysis: domain experts collect and sort customer information with abnormal points, determine whether they have been subjected to concentrated attacks through induction and summary, and summarize key features such as consistent vehicle models, high virtual vehicle prices, unified transaction schemes, no reasonable explanation for out-of-town vehicle purchases, concentrated in a few channels, unreasonable order positioning, similar loan application times, etc.

[0120] 2. Collection of historical customer data: Extract data of a certain vehicle model from the business system for the past month, including customer information table (residence address, work address, residence address), loan application table (loan application time, address, down payment ratio, down payment, period, loan amount), dealer information table (dealer name, dealer address), vehicle information table (vehicle model, manufacturer's guide price), credit information table (whether there is credit, self-check credit frequency in the past 30 days, institutional query frequency in the past 30 days, institutional loan refusal frequency in the past 90 days).

[0121] 3. Feature engineering: Based on the characteristics of mass case customers, the basic data is converted into quantifiable features.

[0122] (1) Off-site personnel: Determine whether the residence address and work address, residence address are in the same province and city, and if not, it is an off-site personnel.

[0123] (2) Abnormal life trajectory: Determine whether the work address and residence address are in the same province and city, and if not, it is abnormal.

[0124] (3) Off-site car purchase: Determine whether the residence address, work address and dealer address are in the same province and city, and if not, it is off-site.

[0125] (4) Abnormal order: Determine whether the order address and dealer address are in the same province and city, and if not, it is abnormal.

[0126] (5) Vehicle price deviation: The formula is (down payment / down payment ratio) / manufacturer's guide price.

[0127] 4. Construction of historical customer feature adjacency matrix A, classified processing according to data type.

[0128] (1) Processing of discrete data: Taking the car purchase channel as an example, each channel is regarded as a feature, and a Boolean matrix space of customers and channels is constructed, where m is the customer, n1 is the channel, and the value is 1 when the customer belongs to the channel, otherwise 0. All features include dealer name, off-site personnel, abnormal life trajectory, off-site car purchase, off-site order, loan period, and whether there is credit.

[0129] (2) Processing of numerical data: Taking the total loan amount as an example, it is binned, and each bin is regarded as a feature, and a Boolean matrix space of customers and total loan amount bins is constructed, where m is the customer, n2 is the total loan amount bin, and the value is 1 when the customer's total loan amount belongs to the bin, otherwise 0. All features include total loan amount, down payment ratio, vehicle price deviation, self-check credit frequency in the past 30 days, institutional query frequency in the past 30 days, and institutional loan refusal frequency in the past 90 days.

[0130] (3) Date type data processing: taking the loan application time as an example, it is processed into year-month-week (such as 2025 June 1st week), each year-month-week is taken as a feature, and a Boolean matrix space of the customer and the loan application time binning is constructed, wherein m is the customer, n3 is the loan application time year-month-week, and when the customer belongs to the week, the value is 1, otherwise 0. All features include the loan application time.

[0131] (4) Construction of historical customer feature adjacency matrix A: the above Boolean matrix space is spliced to obtain the final adjacency matrix A, wherein M is the customer (the number of historical customers), N is all features (the number of historical customer features, that is, the feature dimension of the mass case feature library), M=8584, N=704.

[0132] 5, Construction of adjacency matrix index table E: the historical customer feature adjacency matrix A is traversed by row, for each customer hit feature f, the corresponding historical customer index number (that is, each historical customer has its corresponding index number) and the feature index number are saved as edges, finally a total of 256476 edges, the format is as follows:

[0133] Table 1 Adjacency matrix index example table

[0134]

[0135] For example, historical customer 1 hits features 4, 8 and 9, wherein feature 4 is loan amount (135000, 164900], feature 8 is down payment ratio (-0.1, 15.3], and feature 9 is online direct channel.

[0136] 6, Screening of seed customer group: for the customer to be investigated, the relevant basic information is collected from the business system and processed into the data structure of the historical customer feature adjacency matrix. In the historical customer feature adjacency matrix A, the similarity of the customer to be investigated is sorted from high to low, and the top 5 historical customer index numbers with a similarity greater than 70% are taken as the seed customer group .

[0137] The second part is the mining stage. For details, please refer to Figure 3 , Figure 3 A flowchart of a bidirectional propagation algorithm based on feature density provided by an embodiment of the application.

[0138] 1, Prepare the seed customer group and the adjacency matrix index table E.

[0139] 2, Define the size of the density block , and calculate the block density threshold .

[0140] 3、 define the density block threshold value required for adding the current customer or the current customer feature to the current customer set or the current feature set in each propagation as .

[0141] 4、 Propagation S2F(S, d) of the customer set S to the feature set F: initialize F as an empty set, for each feature f, calculate the number of edges c1 that exist in S, select f with the number of edges greater than the density block threshold value d, and finally F = {641, 514, 3, 4, 98, 614, 7, 8, 9, 10, 266, 2}.

[0142] 5、 Propagation F2S(F, d) of the feature set F to the customer set S: initialize S as an empty set, for each customer s, calculate the number of edges c2 that exist in F, select s with the number of edges greater than the density block threshold value d, and finally S = {8192, 1, 7171, 5636, 3077, 5, 7, 520, 1032, 13, 4110, 4623, 1041, 530,...}.

[0143] 6、 Continuously perform the above propagation process to update the customer set and the feature set until convergence, that is, . As a customer set in a high-density feature block, that is, the target group output, = {8192, 1, 7171, 5636, 3077, 5, 7, 520, 1032, 125, 130, 8324, 132,...}.

[0144] The third part is the deployment stage. For details, refer to Figure 4 , Figure 4 A flowchart example provided by an embodiment of the application based on the deployment stage.

[0145] 1、 Service deployment: load the calculated historical customer feature adjacency matrix A and adjacency matrix index table E into the memory, and define the input and output formats of the service using the API method.

[0146] 2、 Feature engineering: collect relevant basic information from the business system for the customers to be investigated and process it into the data structure of the adjacency matrix A.

[0147] 3、 Screening of seed customer group: in the historical customer feature adjacency matrix A, sort the similarity of the customer features to be investigated from high to low, and take the top 5 customers with a similarity greater than 70% as the seed customer group If the similarity is too low, there is no seed customer group, and an empty set is returned.

[0148] 4、 Customer group mining: if there is Figure 3 ​​algorithm steps, calculate the feature dense target group , as output feedback to the business system.

[0149] Therefore, the beneficial effects of the present application can be summarized as follows:

[0150] (1) Identification of mass correlation: by constructing a historical customer feature adjacency matrix, the customer is propagated based on feature density, thereby identifying the customer group with the same feature, which has the interpretability of association rules and greater flexibility than rules;

[0151] (2) Incremental update of features: common features summarized by experts can be directly expanded to the historical customer feature adjacency matrix, ensuring the accuracy and timeliness of the algorithm to identify dense feature groups;

[0152] (3) Incremental update of artificial marking: risk customers marked by experts in real time can be used as customer samples to be investigated, as a feature basis for screening seed customer groups, ensuring the accuracy and timeliness of the algorithm to identify dense feature groups;

[0153] (4) The time complexity of the propagation algorithm is O(N): the propagation algorithm is linearly related to the number of edges, ensuring the timeliness of the result calculation.

[0154] The group identification method provided by the present application fully utilizes business data and artificial expert experience to solve the accuracy and timeliness problems of group identification under the current difficulty of constructing an effective strong causal relationship network. Based on the method, different density blocks are set for testing on business data, and the algorithm can be selected between accuracy and recall rate. The accuracy is 20%-47%, and the recall rate is 40%-92%, which greatly improves the efficiency of artificial accurate identification of abnormal groups (abnormal groups can be understood as groups with fraudulent behavior / fraudulent behavior), and reduces the subsequent disposal cost of abnormal groups.

[0155] The group identification device provided by the embodiment of the present application will be described below. The group identification device described below can be correspondingly referred to the group identification method described above.

[0156] For details, Figure 5 , Figure 5 The structure diagram of a group identification device provided by the embodiment of the present application can include:

[0157] The adjacency matrix construction module 100 is used to obtain the basic data of the historical customers from the mass case feature library, and convert the basic data of the historical customers into historical customer features to obtain the historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers who have business transactions with automobile organizations;

[0158] The index table construction module 200 is used to construct an adjacency matrix index table that corresponds the historical customers to the historical customer features based on the historical customer feature adjacency matrix.

[0159] The seed customer acquisition module 300 is used to acquire customers to be investigated and to filter and obtain seed customers based on the customer characteristics of the customers to be investigated and the adjacency matrix of the historical customer characteristics; the customers to be investigated are customers in the automotive business with confirmed information doubts.

[0160] The target group determination module 400 is used to perform an iterative bidirectional propagation operation on the seed customer group based on the adjacency matrix index table and a density block threshold calculated based on a preset density block size. When the convergence condition is met, the converged customer set is taken as the target group. The bidirectional propagation operation consists of a customer-to-feature selection operation and a feature-to-customer selection operation. The convergence condition is that the customer set and feature set obtained from two adjacent iterative bidirectional propagation operations are completely identical. The density block threshold is calculated using the following formula: The density threshold of the density block The calculation formula is as follows: S i Let F be the current customer set in round i; i Let be the current feature set in the i-th round; and All represent the average density corresponding to the historical customer feature adjacency matrix A; The parameters representing the association between customer i and feature j are called edges; This represents the density block size; M is the number of historical customers; and N is the number of historical customer features.

[0161] Based on the above embodiments, the seed customer acquisition module 300 may include:

[0162] An adjacency matrix construction unit is used to obtain the basic data of the customer to be investigated and convert the basic data of the customer to be investigated into the customer features;

[0163] The comparison unit is used to compare the customer features with the historical customer features in the historical customer feature adjacency matrix, filter out the top x customers whose feature similarity is greater than a preset threshold, and use the top x customers as the seed customer group.

[0164] Based on any of the above embodiments, the target group determination module 400 may include:

[0165] The first unit is configured to perform step 1: calculating a density block density threshold based on a preset density block size, and calculating the density block threshold based on the density block density threshold, a current customer set or a current customer feature set;

[0166] The second unit is configured to perform step 2: based on the adjacency matrix index table, counting a first associated edge number of each feature in the customer set, screening features with the first associated edge number greater than the density block threshold, and forming the feature set; the initial set of the customer set is the seed customer group; the first associated edge number is the number of customers associated with each feature;

[0167] The third unit is configured to perform step 3: if the feature set does not satisfy a first judgment condition, outputting an empty set and terminating; the first judgment condition is , F represents the feature set;

[0168] The fourth unit is configured to perform step 4: if the feature set satisfies the first judgment condition, based on the adjacency matrix index table, counting a second associated edge number of each customer in the feature set, screening customers with the second associated edge number greater than the density block threshold, and forming a new customer set; the second associated edge number is the number of features associated with each customer;

[0169] The fifth unit is configured to perform step 5: if the new customer set does not satisfy a second judgment condition, outputting an empty set and terminating; the second judgment condition is , S represents the customer set;

[0170] The sixth unit is configured to perform step 6: if the new customer set satisfies the second judgment condition, repeating the steps 1 to 6 until a convergence condition is satisfied, and terminating, and taking the converged customer set as the target group.

[0171] Based on the above embodiment, the target group determination module 400 can include:

[0172] The judgment unit is configured to, when the customer set and the feature set obtained by the bidirectional propagation operation of two adjacent iterations are completely consistent, take the customer set as the target group.

[0173] Based on the above embodiment, the adjacency matrix construction module 100 can include:

[0174] The analysis unit is configured to obtain customer information of abnormal input records, and analyze based on the customer information to obtain typical features vulnerable to attacks, and form the mass case feature library;

[0175] An extraction unit is configured to extract basic data of the historical customer from a business system based on a feature dimension of the group case feature library;

[0176] A conversion unit is configured to convert the basic data of the historical customer into quantifiable historical customer features based on the typical features of the group case feature library.

[0177] A classification processing unit is configured to classify the historical customer features according to data types and construct a historical customer feature adjacency matrix based on a processed result.

[0178] Based on the above embodiment, the classification processing unit can include:

[0179] A discrete processing unit is configured to construct a first initial Boolean matrix of the historical customer and discrete features, take a value of 1 for a position in the first initial Boolean matrix that matches a discrete feature in the historical customer features, and take a value of 0 otherwise, to obtain a first Boolean matrix.

[0180] A numerical processing unit is configured to perform a binning processing on a numerical feature in the historical customer features according to a preset rule, construct a second initial Boolean matrix of the historical customer and the feature binning according to a binning result, take a value of 1 for a position in the second initial Boolean matrix that corresponds to a binning feature to which the numerical feature belongs, and take a value of 0 otherwise, to obtain a second Boolean matrix.

[0181] A time processing unit is configured to perform a binning processing on a date feature in the historical customer features according to a year-month-week format, construct a third initial Boolean matrix of the historical customer and the time binning according to a binning result, take a value of 1 for a position in the third initial Boolean matrix that corresponds to a binning feature to which the date feature belongs, and take a value of 0 otherwise, to obtain a third Boolean matrix.

[0182] A splicing unit is configured to splice the first Boolean matrix, the second Boolean matrix and the third Boolean matrix along a feature dimension, to form the historical customer feature adjacency matrix containing M historical customers and N historical customer features. Rows of the historical customer feature adjacency matrix correspond to the historical customers, and columns correspond to the historical customer features.

[0183] Based on the above embodiment, the index table construction module 200 can include:

[0184] An index table forming unit is configured to traverse the historical customer feature adjacency matrix by rows, extract index information of a feature hit by each historical customer, and generate the adjacency matrix index table in a form of historical customer index numbers and feature index numbers as edges.

[0185] It should be noted that the order of the modules and units in the above-mentioned group identification device can be changed without affecting the logic.

[0186] The group identification device provided in this embodiment of the invention uses an adjacency matrix construction module 100 to obtain basic data of historical customers from a group case feature database, and convert the basic data of historical customers into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers who have business dealings with the automotive organization; an index table construction module 200 is used to construct an adjacency matrix index table of the correspondence between the historical customers and the historical customer features based on the historical customer feature adjacency matrix; and a seed customer group acquisition module 300 is used to obtain the customer features of the customers to be investigated, and to determine the seed customer group based on the seed customer group. The adjacency matrix of the customers and the historical customer features is used to filter out seed customer groups; the customers to be investigated are those with confirmed suspicious information in the automotive business; the target group determination module 400 is used to perform iterative bidirectional propagation operations on the seed customer groups according to the adjacency matrix index table and the density block threshold calculated based on the preset density block size. When the convergence condition is met, the converged customer set is taken as the target group; the bidirectional propagation operation is a customer-to-feature selection operation and a feature-to-customer selection operation; the convergence condition is that the customer set and feature set obtained by two adjacent iterative bidirectional propagation operations are completely consistent; the density block threshold calculation formula is: The density threshold of the density block The calculation formula is as follows: S i Let F be the current customer set in round i; i Let be the current feature set in the i-th round; and All represent the average density corresponding to the historical customer feature adjacency matrix A; The parameters representing the association between customer i and feature j are called edges; D represents the density block size; M represents the number of historical customers; and N represents the number of historical customer features. The device anchors core abnormal features by relying on the group case feature library, mines the associated group of customers to be investigated through an adjacency matrix and bidirectional propagation, avoids misjudgment and missed judgment of normal associated customer groups, and improves the identification accuracy. This has the interpretability of the association rule and greater flexibility than the association rule. In addition, the adjacency matrix index table can simplify the associated data of customers and features, reduce redundant calculation, and accelerate the abnormal group screening speed and improve the identification efficiency through iterative bidirectional propagation without full traversal. In addition, the abnormal group association is identified from the seed customer group to the associated customers, which can effectively capture group abnormal behaviors and meet actual business needs. In addition, based on the basic data conversion and matrix operation, the device does not need complex graph algorithms or model training, reduces the dependence on computing power, lowers the landing threshold, and is convenient for landing application in the business system.

[0187] The group identification device provided by the embodiment of the application is described below. The group identification device described below can be correspondingly referred to the group identification method described above.

[0188] Please refer to Figure 6 , Figure 6 The structure diagram of the group identification device provided by the embodiment of the application can include:

[0189] The memory 10 is configured to store a computer program.

[0190] The processor 20 is configured to execute the computer program to implement the group identification method described above.

[0191] The memory 10, the processor 20, and the communication interface 31 can communicate with each other through the communication bus 32.

[0192] In the embodiment of the application, the memory 10 stores one or more programs. The program can include program code, and the program code includes computer operation instructions. In the embodiment of the application, the memory 10 can store programs for implementing the following functions:

[0193] The basic data of the historical customers is obtained from the group case feature library, and the basic data of the historical customers is converted into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features. The historical customers are customers having business with the automobile organization.

[0194] Based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features is constructed.

[0195] Obtain a to-be-investigated customer, and filter a seed customer group according to a customer feature and a historical customer feature adjacency matrix of the to-be-investigated customer; the to-be-investigated customer is a customer with confirmed information doubts in automobile business;

[0196] According to an adjacency matrix index table and a density block threshold value calculated based on a preset density block size, performing an iterative bidirectional propagation operation on the seed customer group, and when a convergence condition is met, taking a converged customer set as a target group; the bidirectional propagation operation is a customer-to-feature selection operation and a feature-to-customer selection operation; the convergence condition is that a customer set and a feature set obtained by two adjacent iterative bidirectional propagation operations are completely consistent;

[0197] The density block threshold value calculation formula is:

[0198] ;

[0199] The density block density threshold value The calculation formula is as follows:

[0200] ;

[0201] S i is a current customer set of the i th round; F i is a current feature set of the i th round; and both represent the average density corresponding to the historical customer feature adjacency matrix A; represents an association parameter between customer i and feature j, and is referred to as an edge; represents a density block size; M is the number of historical customers; and N is the number of historical customer features.

[0202] In a possible implementation manner, the memory 10 can include a storage program area and a storage data area, where the storage program area can store an operating system, and application programs required by at least one function, and the like; and the storage data area can store data created in a use process.

[0203] In addition, the memory 10 can include a read-only memory and a random access memory, and provide instructions and data for the processor. A part of the memory can also include an NVRAM. The memory stores an operating system and operation instructions, executable modules or data structures, or subsets of them, or an extended set of them, where the operation instructions can include various operation instructions for implementing various operations. The operating system can include various system programs for implementing various basic tasks and processing hardware-based tasks.

[0204] The processor 20 can be a central processing unit (CPU), an application specific integrated circuit, a digital signal processor, a field programmable gate array or other programmable logic device. The processor 20 can be a microprocessor or any conventional processor, etc. The processor 20 can invoke a program stored in the memory 10.

[0205] The communication interface 31 can be an interface of a communication module, used for connecting with other devices or systems.

[0206] Of course, it needs to be explained that, Figure 6 The structure shown does not constitute a limitation on the group identification device in the embodiments of the present application, and in actual applications, the group identification device can include more or less components than Figure 6 those shown, or combine certain components.

[0207] The computer readable storage medium provided by the embodiments of the present application is introduced as follows, and the computer readable storage medium described below can be referred to in correspondence with the group identification method described above.

[0208] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the group identification method described above.

[0209] The computer readable storage medium can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0210] In the present specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0211] The skilled person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly show the interchangeability of hardware and software, the components and steps of each example have been described in general in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0212] Finally, it is to be understood that, wherever used, the terms "comprising", "including", or other any variations thereof are intended to be non-exclusive in nature, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0213] The above describes in detail the population identification method, device, equipment and computer readable storage medium provided by the present application. The principles and implementation manners of the present application are described by applying specific examples. The above example description is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method of population identification, characterized by, The method comprises the following steps: According to the mass case feature library, the basic data of the historical customers are obtained, and the basic data of the historical customers are converted into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers who have business transactions with automobile institutions; the basic data include data in the customer information table, the loan application table, the dealer information table, the vehicle information table and the credit derivative information table extracted from the business system; the historical customer features include the dealer name, the out-of-town personnel, the abnormal life trajectory, the out-of-town car purchase, the out-of-town order, the loan period, whether there is credit, the total loan amount, the down payment ratio, the vehicle type and price deviation, the number of self-check credit in the last 30 days, the number of institutional inquiries in the last 30 days, the number of institutional loan refusals in the last 90 days, and the loan application time; Based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features is constructed; A to-be-investigated customer is obtained, and a seed customer group is screened according to the customer features of the to-be-investigated customer and the historical customer feature adjacency matrix; The to-be-investigated customer is a customer in automobile business who has confirmed information doubts; According to the adjacency matrix index table and a density block threshold value calculated based on a preset density block size, an iterative bidirectional propagation operation is performed on the seed customer group, and when a convergence condition is met, a converged customer set is taken as a target group; the bidirectional propagation operation is a customer-to-feature selection operation and a feature-to-customer selection operation; the convergence condition is that the customer set and the feature set obtained by two adjacent iterative bidirectional propagation operations are completely consistent; The density block threshold value calculation formula is: ； The density block density threshold The calculation formula is as follows: ； S i is the current customer set for the i-th round; F i is the current feature set for the i-th round; and both represent the average density corresponding to the historical customer feature adjacency matrix A; represents the association parameter between customer i and feature j, called edge; represents the density block size; M is the number of historical customers; N is the number of historical customer features; m and n represent the preset number of customers and the preset number of features in the density block , respectively.

2. The method of claim 1, wherein, A to-be-investigated customer is obtained, and a seed customer group is screened according to the customer features of the to-be-investigated customer and the historical customer feature adjacency matrix, comprising: Basic data of the to-be-investigated customer are obtained, and the basic data of the to-be-investigated customer are converted into the customer features; The customer features are compared with the historical customer features in the historical customer feature adjacency matrix in terms of feature similarity, the first x customers with feature similarity greater than a preset threshold value are screened, and the first x customers are taken as the seed customer group.

3. The method of claim 1 or 2, wherein the method further comprises: According to the adjacency matrix index table and a density block threshold value calculated based on a preset density block size, an iterative bidirectional propagation operation is performed on the seed customer group, and when a convergence condition is met, a converged customer set is taken as a target group, comprising: Step 1: a density block density threshold value is calculated based on a preset density block size, and the density block threshold value is calculated based on the density block density threshold value, the current customer set or the current customer feature set; Step 2: based on the adjacency matrix index table, the number of first associated edges of each feature in the customer set is counted, features with the number of first associated edges greater than the density block threshold value are screened, and the feature set is formed; the initial set of the customer set is the seed customer group; the number of first associated edges is the number of customers associated with each feature; Step 3: if the feature set does not satisfy a first judging condition, output an empty set and terminate; the first judging condition is , F represents a feature set; Step 4: If the feature set meets the first judgment condition, the number of second associated edges of each customer in the feature set is counted based on the adjacency matrix index table, customers with the number of second associated edges greater than the density block threshold are screened, and a new customer set is formed; Step 5: If the new customer set does not satisfy a second judging condition, output an empty set and terminate; the second judging condition is , S represents a customer set; Step 6: If the new customer set meets the second judgment condition, the steps 1 to 6 are repeatedly executed until the convergence condition is met, and the converged customer set is taken as the target group.

4. The method of claim 1, wherein, When the convergence condition is met, the converged customer set is taken as the target group, comprising: When the customer set and the feature set obtained by the bidirectional propagation operation of the adjacent two iterations are completely consistent, the customer set is taken as the target group.

5. The method of claim 1, wherein, According to the group case feature library, the basic data of the historical customer is obtained, and the basic data of the historical customer is converted into historical customer features to obtain a historical customer feature adjacency matrix associated with the historical customer and the historical customer features, comprising: Obtaining customer information of abnormal application records, and analyzing based on the customer information to obtain typical features vulnerable to attack, forming the group case feature library; the typical features vulnerable to attack include consistent vehicle models, high virtual vehicle prices, unified transaction schemes, no reasonable explanation for out-of-town vehicle purchases, concentration in certain channels, unreasonable positioning of application records, and similar loan application times; Based on the feature dimensions of the group case feature library, the basic data of the historical customer is extracted from the business system; Based on the typical features of the group case feature library, the basic data of the historical customer is converted into quantifiable historical customer features; The historical customer features are classified and processed according to the data type, and the historical customer feature adjacency matrix is constructed based on the processed results.

6. The method of claim 5, wherein, The historical customer features are classified and processed according to the data type, and the historical customer feature adjacency matrix is constructed based on the processed results, comprising: A first initial Boolean matrix of historical customers and discrete features is constructed, the positions in the first initial Boolean matrix matching the discrete features in the historical customer features are valued as 1, and the positions are valued as 0 otherwise, to obtain a first Boolean matrix; The numerical features in the historical customer features are binned according to a preset rule, and a second initial Boolean matrix of historical customers and feature bins is constructed according to the binning results; the positions in the second initial Boolean matrix corresponding to the binning features to which the numerical features belong are valued as 1, and the positions are valued as 0 otherwise, to obtain a second Boolean matrix; The date type features in the historical customer features are binned according to the year-month-week format, and a third initial Boolean matrix of historical customers and time bins is constructed according to the binning results; the positions in the third initial Boolean matrix corresponding to the binning features to which the date type features belong are valued as 1, and the positions are valued as 0 otherwise, to obtain a third Boolean matrix; Splice the first Boolean matrix, the second Boolean matrix and the third Boolean matrix along the feature dimension to form the historical customer feature adjacency matrix containing M historical customers and N historical customer features; the rows of the historical customer feature adjacency matrix correspond to the historical customers, and the columns correspond to the historical customer features.

7. The method of claim 1, wherein, Based on the historical customer feature adjacency matrix, an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features is constructed, including: Traverse the historical customer feature adjacency matrix by row, extract the index information of the hit features of each historical customer, and generate the adjacency matrix index table in the form of historical customer index numbers and feature index numbers as edges.

8. A group identification apparatus characterized by comprising: Including: An adjacency matrix construction module is configured to obtain basic data of historical customers from a mass case feature library, convert the basic data of the historical customers into historical customer features, and obtain a historical customer feature adjacency matrix associated with the historical customers and the historical customer features; the historical customers are customers having business transactions with automobile institutions; the basic data includes data in a customer information table, a loan application table, a dealer information table, a vehicle information table and a credit derivative information table extracted from a business system; the historical customer features include dealer name, off-site personnel, life trajectory anomaly, off-site car purchase, off-site order, loan period, whether there is credit, total loan amount, down payment ratio, vehicle type and price deviation, number of credit inquiries in the last 30 days, number of institutional inquiries in the last 30 days, number of institutional loan rejections in the last 90 days, and loan application time; An index table construction module is configured to construct an adjacency matrix index table of the corresponding relationship between the historical customers and the historical customer features based on the historical customer feature adjacency matrix; A seed customer group acquisition module is configured to acquire customer features of a customer to be investigated, and filter a seed customer group according to the customer to be investigated and the historical customer feature adjacency matrix; The customer to be investigated is a customer having confirmed information doubts in automobile business; A target group determination module is configured to perform an iterative bidirectional propagation operation on the seed customer group according to the adjacency matrix index table and a density block threshold value calculated based on a preset density block size, and when a convergence condition is met, a customer set after convergence is taken as a target group; the bidirectional propagation operation is a selection operation of customers to features and a selection operation of features to customers; the convergence condition is that the customer set and the feature set obtained by the bidirectional propagation operation of the two adjacent iterations are completely consistent; The density block threshold value calculation formula is: ； Density block density threshold The formula for calculating the density threshold is as follows: ； S i is the current customer set for the i-th round; F i is the current feature set for the i-th round; and both represent the average density corresponding to the historical customer feature adjacency matrix A; represents the association parameter between customer i and feature j, called edge; represents the density block size; M is the number of historical customers; N is the number of historical customer features; m and n represent the preset number of customers and the preset number of features in the density block , respectively.

9. A group identification device, characterized by Including: A memory is configured to store a computer program; A processor is configured to execute the computer program to implement the group identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and when the computer executable instructions are loaded and executed by the processor, the group identification method according to any one of claims 1 to 7 is implemented.

Images