Abnormal user identification method, device, apparatus and storage medium

Abstract

The application provides an abnormal user identification method and device, equipment and storage medium, relates to artificial intelligence technology, and can be used in the financial field. The method comprises the following steps: acquiring user asset related data to be identified; the user asset related data to be identified comprises at least one to-be-identified feature; the to-be-identified feature is a user asset related feature; inputting the user asset related data to be identified into a trained default user identification model, and adopting the trained default user identification model to output whether the to-be-identified user corresponding to the user asset related data to be identified is an abnormal user. The scheme of the application can improve the identification accuracy of the default user identification model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an abnormal user identification method, apparatus, device, and storage medium. Background Technology

[0002] As society develops, the market size of financial services is also growing. When handling financial transactions for users, banks and other financial institutions need to identify potential defaulters to ensure the recovery of funds.

[0003] Traditional human identification methods are inefficient and unable to adapt to the increasing number of users. Currently, artificial intelligence (AI) is often used to assist human identification. AI identifies abnormal users through recognition models, and the accuracy of these models is affected by many factors, including the amount and quality of data, and the model parameters. Currently, the model parameters of these recognition models are mostly randomly generated, which results in low accuracy. Summary of the Invention

[0004] This application provides an abnormal user identification method, apparatus, device, and storage medium to solve the problem of low identification accuracy in existing identification models.

[0005] According to a first aspect of this application, an abnormal user identification method is provided, comprising:

[0006] Acquire data related to the assets of the user to be identified; the data related to the assets of the user to be identified includes at least one feature to be identified; the feature to be identified is a user asset-related feature.

[0007] The user asset-related data to be identified is input into the trained default user identification model, and the trained default user identification model outputs whether the user corresponding to the user asset-related data is an abnormal user; the trained default user identification model includes at least one optimized parameter; the optimized parameter is obtained by training the user asset-related dataset and the parameter to be optimized using an improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0008] According to a second aspect of this application, an abnormal user identification device is provided, comprising:

[0009] The acquisition module is used to acquire data related to the assets of the user to be identified; the data related to the assets of the user to be identified includes at least one feature to be identified; the feature to be identified is a user asset-related feature;

[0010] The identification module is used to input the asset-related data of the user to be identified into the trained default user identification model, and to output whether the user to be identified corresponding to the asset-related data is an abnormal user. The trained default user identification model includes at least one optimized parameter. The optimized parameter is obtained by training the user asset-related dataset and the parameter to be optimized using an improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0011] According to a third aspect of this application, an electronic device is provided, comprising: a processor and a memory communicatively connected to the processor;

[0012] The memory stores computer-executed instructions;

[0013] The processor executes computer execution instructions stored in the memory to implement the method as described in the first aspect.

[0014] According to a fourth aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, which, when executed by a processor, are used to implement the method as described in the first aspect.

[0015] According to a fifth aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in the first aspect.

[0016] The abnormal user identification method, apparatus, device, and storage medium provided in this application acquire user asset-related data to be identified. This data includes at least one feature to be identified, which is a user asset-related feature. The user asset-related data is input into a trained default user identification model, and the model outputs whether the user corresponding to the data is an abnormal user. The trained default user identification model includes at least one optimized parameter. This optimized parameter is obtained by training the user asset-related dataset and an improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution. Because the optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm, the optimized parameter closely matches the user asset-related data and is the parameter found by the particle swarm optimization algorithm that makes the model most accurate. Therefore, the solution in this application can improve the identification accuracy of the default user identification model. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1A This is a network architecture diagram corresponding to an application scenario provided in the embodiments of this application;

[0019] Figure 1B This is a network architecture diagram corresponding to another application scenario provided in the embodiments of this application;

[0020] Figure 2 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 1 of this application;

[0021] Figure 3 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 2 of this application;

[0022] Figure 4 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 3 of this application;

[0023] Figure 5 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 4 of this application;

[0024] Figure 6 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 5 of this application;

[0025] Figure 7 This is a schematic diagram of the abnormal user identification device provided in Embodiment Six of this application;

[0026] Figure 8 This is a block diagram of an electronic device provided according to Embodiment 7 of this application.

[0027] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0028] First, the terms used in this application will be explained.

[0029] It should be noted that the abnormal user identification method, apparatus, device, and storage medium provided in this application can be used in the financial field. They can also be used in any field other than finance. The application areas of the abnormal user identification method, apparatus, device, and storage medium in this application are not limited.

[0030] The prior art involved in this application will be described in detail and analyzed below.

[0031] Banks and other financial institutions need to identify potentially defaulting users when processing financial transactions to improve transaction success rates, reduce failed fund recovery, and ensure fund recovery. With the development of internet finance technology and services, the market size of financial products is expanding, allowing users to conduct financial transactions, such as credit card transactions, directly online. This brings great convenience to users, but also presents greater challenges for banks and other financial institutions in verifying user information and identifying potentially defaulting users.

[0032] Currently, manual methods for identifying anomalous users are inefficient and cannot adapt to the increasing number of users. Therefore, artificial intelligence (AI) is often used to supplement manual identification. The high accuracy of AI identification can improve the efficiency of manual anomalous user identification. AI uses machine learning to build identification models, which then identify anomalous users. The accuracy of the identification model is affected by many factors, including the amount and quality of data, and the model parameters. Currently, the parameters of identification models are mostly randomly generated, resulting in low accuracy and failing to meet the accuracy requirements of practical applications.

[0033] Therefore, in the face of the problems in the existing technology, the inventors, through creative research, have found that to improve the recognition accuracy of the identification model, the parameters of the identification model need to be more accurate and more closely aligned with the application scenario. Thus, the inventors propose the technical solution of this application, which involves obtaining data related to the assets of the user to be identified; this data includes at least one feature to be identified; the feature to be identified is a user asset-related feature; the data is input into a trained default user identification model, and the trained default user identification model outputs whether the user corresponding to the data is an abnormal user; the trained default user identification model includes at least one optimized parameter; the optimized parameter is obtained by training the user asset-related dataset and an improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm trains the optimized parameter by finding the optimal solution to obtain the optimized parameter. Since the optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm, the optimized parameter is closely aligned with the user asset-related data and is the parameter found by the particle swarm optimization algorithm that makes the model most accurate. Therefore, the solution of this application can improve the recognition accuracy of the default user identification model.

[0034] The abnormal user identification method, apparatus, device, and storage medium provided in this application are intended to solve the above-mentioned technical problems of the prior art.

[0035] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solutions of this invention comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0036] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0037] The network architecture and application scenarios of the abnormal user identification method provided in the embodiments of this application will be described below. When the following description refers to the accompanying drawings, unless otherwise indicated, the same data in different drawings represent the same or similar elements.

[0038] like Figure 1A As shown, the network architecture corresponding to an application scenario provided in this application embodiment includes: a terminal 11, an electronic device 12, and a database 13. The terminal 11 carries a client 110 with an abnormal user identification method application software, or the terminal 11 accesses the URL of the abnormal user identification method. The terminal 11, electronic device 12, and database 13 are communicatively connected. The database 13 stores at least one of a training user asset-related dataset and a historical user asset-related dataset; the database 13 may also store user asset-related data to be identified. The electronic device 12 is configured with a trained defaulting user identification model.

[0039] Individuals or organizations that need to identify abnormal users, such as staff of banks and other financial institutions, can input data related to the assets of the user to be identified through the operation interface of client 110 or the corresponding webpage. Terminal 11 then sends this data to electronic device 12, enabling the electronic device to obtain the data. The data includes at least one feature to be identified, which is a user asset-related feature.

[0040] Electronic device 12 inputs the asset-related data of the user to be identified into the trained default user identification model, and uses the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user; the trained default user identification model includes at least one optimized parameter; the optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm to obtain the optimized parameter; the improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0041] Electronic device 12 sends the data related to the user's assets to be identified to terminal 11 to determine whether the user is an abnormal user. This allows terminal 11 to send the results of whether the data related to the user's assets to be identified is an abnormal user to staff of financial institutions such as banks, or other individuals or organizations that need to identify abnormal users.

[0042] Electronic device 12 can obtain a training user asset-related dataset from the database, and use an improved particle swarm optimization algorithm and the training user asset-related dataset to train at least one parameter to be optimized, thereby obtaining at least one optimized parameter and a trained defaulting user identification model. When the database only stores historical user asset-related datasets, electronic device 12 can obtain historical user asset-related datasets from the database and determine the training user asset-related dataset based on the historical user asset-related datasets.

[0043] like Figure 1B As shown, the network architecture corresponding to another application scenario provided in this application embodiment includes: an electronic device 12 and a database 13. The electronic device 12 is communicatively connected to the database 13. The electronic device 12 is configured with a trained defaulting user identification model. The database 13 stores data related to the user's assets to be identified, and also stores at least one of a training user asset-related dataset and a historical user asset-related dataset.

[0044] Electronic device 12 can send a data acquisition request to database 13 to obtain data related to the assets of a user to be identified. This data includes at least one feature to be identified; the feature to be identified is a user asset-related feature. After acquiring the data, electronic device 12 inputs it into a trained default user identification model and uses the model to output whether the user corresponding to the data is an abnormal user. The trained default user identification model includes at least one optimized parameter; the optimized parameter is obtained by training the user asset-related dataset and using an improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0045] The embodiments of this application will now be described with reference to the accompanying drawings. The embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0046] Example 1

[0047] Figure 2 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 1 of this application, as follows: Figure 2 As shown, the executing entity of this application is an abnormal user identification device, which is located in an electronic device. The abnormal user identification method provided in this embodiment includes steps 201 to 202.

[0048] Step 201: Obtain data related to the user's assets to be identified; the data related to the user's assets to be identified includes at least one feature to be identified; the feature to be identified is a user asset-related feature.

[0049] In this embodiment, user asset-related data refers to data that measures user assets and user behavior, and can be used to identify whether a user is an abnormal user. For example, user asset-related data may include at least one of the following: user identity information, user property information, user transaction information, user loan information, and user repayment information. User identity information may include at least one of the following: whether the user owns a house, a car, is marital, and has the highest level of education. User property information may include at least one of the following: user investment information, user deposit information, user loan information, and user income information. User transaction information may include at least one of the following: the user's bank transaction history. User loan information may include at least one of the following: the user's account loan limit and the user's account historical limit adjustment information. User repayment information may include at least one of the following: the user's repayment limit, the user's repayment delinquency probability, and the number of days the user's repayment is overdue.

[0050] The features to be identified are those used to identify whether the user corresponding to the user asset-related data is an abnormal user. The features to be identified are user asset-related features, which are features of user asset-related data. User asset-related features can be obtained by feature extraction from user asset-related data.

[0051] Step 202: Input the asset-related data of the user to be identified into the trained default user identification model, and use the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user; the trained default user identification model includes at least one optimized parameter; the optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm to obtain the optimized parameter; the improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0052] In this embodiment, the defaulting user identification model includes at least one adjustable parameter, which affects the identification accuracy of the defaulting user identification model. After optimizing the adjustable parameter as the parameter to be optimized, the optimized parameter is obtained. Replacing at least one adjustable parameter of the defaulting user identification model with the optimized parameter yields a trained defaulting user identification model.

[0053] The training user asset-related dataset can be pre-set, identified user asset-related data, meaning it is known whether the user corresponding to each piece of training user asset-related data in the dataset is an anomalous user. Furthermore, at least one piece of training user asset-related data in the dataset corresponds to a normal user, and at least one piece corresponds to an anomalous user.

[0054] The improved particle swarm optimization (PSO) algorithm can find the optimal solution for a variable in a variable space, which refers to the range of values ​​a variable can take. In this embodiment, the improved PSO algorithm is used to find the optimal solution for the parameter to be optimized to obtain the optimized parameter. Simultaneously, the recognition accuracy of the defaulting user identification model corresponding to the parameter to be optimized on the training user asset-related dataset can be used as the objective function of the improved PSO algorithm, so as to find the optimized parameter that maximizes the accuracy of the defaulting user identification model.

[0055] The abnormal user identification method provided in this embodiment obtains asset-related data of the user to be identified. The asset-related data includes at least one feature to be identified, which is a user asset-related feature. The asset-related data is input into a trained default user identification model, and the trained default user identification model outputs whether the user corresponding to the asset-related data is an abnormal user. The trained default user identification model includes at least one optimized parameter. The optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm trains the optimized parameter by finding the optimal solution to obtain the optimized parameter. Since the optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm, the optimized parameter can fit the user asset-related data closely and is the parameter that makes the model most accurate, found by the particle swarm optimization algorithm. Therefore, the solution of this application can improve the identification accuracy of the default user identification model.

[0056] As an optional implementation, the inertia weight ω of the improved particle swarm optimization algorithm is expressed as: Where, ω max ω represents the maximum value of the inertia weight. min f is the minimum value of the inertial weight. max For maximum group fitness, f min Let f be the minimum group fitness, and f be the particle fitness.

[0057] In this embodiment, the inertia weight ω affects the accuracy and efficiency of the improved particle swarm optimization algorithm in finding the optimal solution. Particle fitness is the value obtained by substituting the particle position into the fitness function; that is, when the parameter to be optimized in the default user identification model is the particle position, the identification accuracy of the default user identification model is achieved. By representing the inertia weight ω as a function of the maximum inertia weight, minimum inertia weight, maximum population fitness, minimum population fitness, and particle fitness, the particle's inertia weight changes with these parameters in each iteration. Consequently, the velocity of each particle also changes with these parameters in each iteration. This allows the local search capability of the improved particle swarm optimization algorithm to become stronger or weaker with the maximum population fitness, minimum population fitness, and particle fitness after each iteration, thus avoiding excessively fast or slow convergence speeds and preventing getting trapped in local optima while achieving efficient searching.

[0058] The abnormal user identification method provided in this embodiment expresses the inertial weight ω of the improved particle swarm optimization algorithm as: By correlating the inertia weight ω with the maximum population fitness, minimum population fitness, and particle fitness, the search efficiency of the improved particle swarm optimization algorithm can be improved and local optima can be avoided, thus improving the parameter optimization effect of the defaulting user identification model.

[0059] Example 2

[0060] Figure 3 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 2 of this application, as follows: Figure 3 As shown, the abnormal user identification method provided in this embodiment, based on embodiment one, includes steps 301 to 302 before step 202, which involves inputting the asset-related data of the user to be identified into the trained default user identification model and using the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user.

[0061] Step 301: Train at least one parameter to be optimized using a training user asset related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter; the training user asset related dataset includes at least one training user asset related data for a normal user and at least one training user asset related data for an abnormal user; the training user asset related data includes at least one feature to be identified.

[0062] In this embodiment, the features to be identified in the training user asset-related data are the same as the features to be identified in the user asset-related data to be identified, so that the optimized parameters obtained by training the user asset-related data can match the user asset-related data to be identified, thereby improving the recognition accuracy of the trained default user identification model corresponding to the optimized parameters in the user asset-related data to be identified.

[0063] Step 302: Determine the trained defaulting user identification model based on at least one optimized parameter.

[0064] In this embodiment, since the defaulting user identification model can have multiple adjustable parameters, one or more adjustable parameters can be used as parameters to be optimized. Then, the improved particle swarm optimization algorithm and the training user asset related dataset are used to train one or more parameters to be optimized to obtain at least one optimized parameter. At least one parameter to be optimized in the defaulting user identification model is replaced with the optimized parameter, so that the defaulting user identification model corresponding to at least one optimized parameter is determined as the trained defaulting user identification model.

[0065] The abnormal user identification method provided in this embodiment trains at least one parameter to be optimized using a training user asset-related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter. The training user asset-related dataset includes at least one training user asset-related data for a normal user and at least one training user asset-related data for an abnormal user. The training user asset-related data includes at least one feature to be identified. A trained default user identification model is determined based on the at least one optimized parameter. Since the training user asset-related dataset and the improved particle swarm optimization algorithm are used to train at least one parameter to be optimized to obtain at least one optimized parameter, multiple optimized parameters can be quickly determined for default user identification models with multiple adjustable parameters, thereby improving the training efficiency of the default user identification model.

[0066] Example 3

[0067] Figure 4 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 3 of this application, as follows: Figure 4 As shown, the abnormal user identification method provided in this embodiment, based on embodiment two, refines step 301 by using a training user asset-related dataset and an improved particle swarm optimization algorithm to train at least one parameter to be optimized in order to obtain at least one optimized parameter. The refinement of step 301 includes the following steps.

[0068] For any parameter to be optimized in the defaulting user identification model, perform the following operations to obtain at least one optimized parameter.

[0069] Step 401: The parameters to be optimized are used as the position variables of the improved particle swarm algorithm, and the recognition accuracy of the defaulting user identification model corresponding to the parameters to be optimized on the training user asset-related dataset is used as the fitness of the improved particle swarm algorithm.

[0070] Since the independent variable of the improved particle swarm optimization algorithm is the particle position, and the algorithm finds the particle position with the highest or lowest fitness by changing the particle position, this embodiment can use the parameter to be optimized as the position variable of the improved particle swarm optimization algorithm, and use the recognition accuracy of the defaulting user identification model corresponding to the parameter to be optimized on the training user asset-related dataset as the fitness of the particle swarm optimization algorithm. Then, by using the improved particle swarm optimization algorithm to find the particle position with the highest fitness, the optimized parameter that maximizes the recognition accuracy of the defaulting user identification model on the training user asset-related dataset can be found.

[0071] Step 402: With the goal of maximizing fitness, solve for the optimal position of the improved particle swarm optimization algorithm.

[0072] Here, since the fitness of the improved particle swarm optimization algorithm is the recognition accuracy of the defaulting user identification model when particle position is used as the parameter to be optimized, the higher the fitness of the improved particle swarm optimization algorithm, the higher the recognition accuracy of the defaulting user identification model. Therefore, taking the maximization of the fitness of the particle swarm as the optimization objective of the improved particle swarm optimization algorithm, finding the particle position that maximizes the fitness will lead to finding the parameter to be optimized that maximizes the recognition accuracy of defaulting users.

[0073] Step 403: Determine the optimal location solution as the optimized parameters.

[0074] It is understandable that when the value of the parameter to be optimized is the optimal position solution found by the improved particle swarm algorithm within a preset number of iterations, the fitness of the particle swarm is maximized and the recognition accuracy of the defaulting user identification model is highest. Therefore, the value of the optimal position solution can be determined as the optimized parameter.

[0075] In this embodiment, for a defaulter identification model with multiple parameters to be optimized, the training order of each parameter is not limited. Each parameter can be trained in any order to determine the corresponding optimized parameters. Finally, the defaulter identification model is obtained by replacing each parameter to be optimized with its corresponding optimized parameter. Furthermore, when at least one optimized parameter exists, it can be substituted into the defaulter identification model during training before training the next parameter to be optimized.

[0076] The abnormal user identification method provided in this embodiment uses the parameters to be optimized as the position variables of the improved particle swarm optimization (PSO) algorithm, and the recognition accuracy of the default user identification model corresponding to the parameters to be optimized on the training user asset-related dataset as the fitness of the improved PSO algorithm. With the goal of maximizing fitness, the optimal position solution of the improved PSO algorithm is solved. The optimal position solution is then determined as the optimized parameters. Because the parameters to be optimized are used as the position variables of the improved PSO algorithm, and the recognition accuracy of the default user identification model corresponding to the parameters to be optimized on the training user asset-related dataset is used as the population fitness of the improved PSO algorithm, the improved PSO algorithm can find the optimal position solution that maximizes the population fitness within a preset number of iterations. Therefore, it can quickly find the optimized parameters that best match the training user asset-related dataset and maximize the recognition accuracy of the default user identification model.

[0077] As an optional implementation, based on Embodiment 4, step 402, which aims to maximize fitness and solve for the optimal position solution of the improved particle swarm optimization algorithm, is further refined. The refined step 402 includes steps 501 to 506.

[0078] Step 501: During the iteration period t, calculate the fitness of each particle in the iteration period t based on the position of each particle; t is a positive integer greater than or equal to 1.

[0079] In this embodiment, t is the iteration period number, therefore t is a natural number, and i represents the i-th particle in the particle swarm, therefore i is a natural number. The i-th particle in the improved particle swarm algorithm has a corresponding position in iteration period t. The fitness at iteration period t is

[0080] Step 502: Determine the maximum and minimum population fitness of the particle swarm in iteration period t based on the fitness of each particle in iteration period t.

[0081] In this embodiment, the particle swarm exhibits a maximum and a minimum population fitness in each iteration period. The maximum population fitness refers to the highest fitness among all particles in the swarm, and the minimum population fitness refers to the lowest fitness among all particles. Therefore, the maximum population fitness of the particle swarm in iteration period t... The maximum fitness of each particle within the iteration period t is, i.e. Minimum population fitness of particle swarm optimization in iteration period t Let be the minimum fitness of each particle in the iteration period t, i.e.

[0082] Step 503: Determine the historical maximum fitness of each particle based on the t fitness values ​​of each particle over t iteration cycles.

[0083] In this embodiment, the historical maximum fitness is the maximum fitness of a single particle across all iteration cycles. It is understood that each particle has a fitness value in each iteration cycle. In iteration period t, each particle has a total of t fitness values. Then the historical maximum fitness of the i-th particle over a total of t iterations. The maximum value of the maximum population fitness for the i-th particle in each iteration cycle is denoted as .

[0084] Step 504: Determine the historical maximum population fitness of the particle swarm based on the t maximum population fitnesss in t iteration cycles.

[0085] In this embodiment, the historical maximum population fitness is the maximum population fitness of the particle swarm in each iteration cycle. The maximum value in the range, i.e., the maximum fitness of the historical population.

[0086] Step 505: Iterate the velocity and position of each particle according to the inertia weight ω until the historical maximum group fitness is greater than or equal to the preset fitness, or the iteration period t is greater than or equal to the preset number of iteration periods.

[0087] In this embodiment, the preset fitness can be the recognition accuracy of the defaulting user identification model, and the preset iteration cycle number is the maximum number of iterations of the improved particle swarm algorithm.

[0088] In each iteration cycle, the maximum fitness F of the historical population is determined. max Then, the velocities V of each particle are iterated according to the inertial weights. i Based on the velocity V of each particle i Iterate over the positions of each particle To obtain the position of each particle in the iteration period t+1 For example,

[0089] The inertia weight ω can change with the iteration period, and the inertia weight ω of each particle... i These can vary; for example, the inertia weight of the i-th particle during iteration period t. It can be represented as f t ω max and ω min Functions, such as,

[0090]

[0091] Where, ω max and ω min It can be preset, ω max The inertia weight, ω, can be used for the first iteration cycle. min It can be the inertia weight for the last iteration cycle.

[0092] Step 506: Determine the position corresponding to the maximum fitness of the historical population as the optimal solution position of the improved particle swarm optimization algorithm.

[0093] The maximum fitness of the historical population is the highest fitness that exists in all iteration cycles. Therefore, the position corresponding to the maximum fitness of the historical population is the optimal solution position of the improved particle swarm optimization algorithm.

[0094] The abnormal user identification method provided in this embodiment calculates the fitness of each particle in iteration period t based on its position; t is a positive integer greater than or equal to 1. It determines the maximum and minimum group fitness of the particle swarm in iteration period t based on the fitness of each particle in iteration period t; it determines the historical maximum fitness of each particle based on the t fitness values ​​of each particle in t iteration periods; it determines the historical maximum group fitness of the particle swarm based on the t maximum group fitness values ​​of the particle swarm in t iteration periods; iterates the velocity and position of each particle according to the inertia weight ω until the historical maximum group fitness is greater than or equal to a preset fitness, or the iteration period t is greater than or equal to a preset number of iteration periods; and determines the position corresponding to the historical maximum group fitness as the optimal solution position of the particle swarm algorithm.

[0095] As an optional implementation, step 505 is refined by iterating the velocity and position of each particle according to the inertia weight ω. The refinement of step 505 includes the following steps.

[0096] Iterate through the velocity and position of each particle using the following formula:

[0097]

[0098]

[0099]

[0100]

[0101] Where i represents the i-th particle, t is the iteration period number, and ω t Let c1 and c2 be the inertia weights for the iteration period t, and c1 and c2 be learning factors that are non-negative. Let be the velocity of particle i during iteration period t. Let be the velocity of particle i after the iteration, and rand(0,1) be a random number between (0,1). Let i be the particle position corresponding to the historical maximum fitness of particle i over t iteration cycles. For particle position, gbest t The position of the particle corresponding to the historical maximum population fitness of the particle swarm over t iterations.

[0102] In this embodiment, during the first iteration cycle, the velocity of each particle in the particle swarm can be 0, and the position of each particle can be randomly generated within a preset variable space. The preset variable space represents the range of values ​​for the parameter to be optimized, and different parameters to be optimized can have different preset variable spaces. Furthermore, in each iteration cycle, the position of each particle also falls within the preset variable space.

[0103] In the first iteration, the i-th particle in the improved particle swarm optimization algorithm has a particle velocity. Particle position and corresponding fitness Then, the maximum population fitness of the particle swarm in the first iteration period. and minimum group fitness The fitness of each particle in the first iteration period. The maximum and minimum values ​​in, i.e. The historical maximum fitness of each particle The highest population fitness in the history of particle swarm optimization The historical maximum fitness The corresponding particle position is determined as the historical best position of the i-th particle within one iteration cycle. The historical maximum population fitness F max The corresponding particle position determines the historical best population position of the particle swarm within a common iteration cycle (gbest). 1 .

[0104] In this embodiment, ω max It can be 0.8, ω min The c1 and c2 can both be 1, so for the i-th particle, The particle velocity of the i-th particle in the improved particle swarm optimization algorithm was calculated in the second iteration cycle. Particle position

[0105] In the second iteration, the i-th particle in the improved particle swarm optimization algorithm has a particle velocity. Particle position and the corresponding fitness Then, the maximum population fitness of the particle swarm in the second iteration cycle. and minimum group fitness Fitness of each particle The maximum and minimum values ​​in, i.e. and The historical maximum fitness of each particle for and The maximum value in, i.e. The highest population fitness in the history of particle swarm optimization The historical maximum fitness The corresponding particle position is determined as the historical best position of the i-th particle over two iteration cycles. The historical maximum population fitness F max The corresponding particle position determines the historical best population position of the particle swarm within a common iteration cycle (gbest). 2 .

[0106] Example 4

[0107] Figure 5 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 4 of this application, as follows: Figure 5 As shown, the abnormal user identification method provided in this embodiment, based on any of the above embodiments, includes steps 701 to 704 before step 301, which involves training a user asset-related dataset and an improved particle swarm optimization algorithm to train at least one parameter to be optimized in order to obtain at least one optimized parameter.

[0108] Step 701: Obtain the historical user asset related dataset; the historical user asset related dataset includes at least one historical user asset related data for a normal user and at least one historical user asset related data for an abnormal user; the historical user asset related data includes at least one user asset related attribute information.

[0109] In this embodiment, the historical user asset-related dataset can be the user's asset-related data within a historical time period, for example, the user's asset-related data over the past year, three years, or five years. User asset-related attribute information is a specific representation of the user asset-related data, and can be a field within the user asset-related data, or a numerical, textual, or logical value with a date. For example, the user asset-related attribute information for "whether a user owns a car" can be a logical value with a date, either yes or no, indicating whether the user owns or does not own a car on a certain date; the user asset-related attribute information for "user transaction information" can be different numerical values ​​on different dates, representing different transaction amounts for the user on different dates; the user asset-related attribute information for "user's highest education level" can be different text on different dates, representing the user's highest education level on different dates.

[0110] Step 702: Determine at least one candidate feature of historical user asset-related data based on the user asset-related attribute information of each user.

[0111] In this embodiment, candidate features are features extracted from historical user asset-related data. These can be represented by features of various user information items, such as user transaction information features and user repayment information features. For user asset-related attribute information expressed numerically, its average value or weighted average value over a historical time period can be determined as a candidate feature of the historical user asset-related data. For user asset-related attribute information expressed in text or logic, its latest value can be determined as a candidate feature of the historical user asset-related data. Alternatively, it can be converted into numerical form and then its average value or weighted average value over a historical time period can be used as a candidate feature. When using the weighted average value of user asset-related attribute information as a candidate feature, the weight of more recent information can be greater, and the weight of older information can be smaller. This embodiment does not impose any limitations on this.

[0112] Step 703: Use the Relief algorithm to calculate the importance values ​​of each candidate feature in the historical user asset related dataset.

[0113] In this embodiment, the Relief algorithm is a feature weighting algorithm that assigns different importance values ​​to features based on their relevance, in order to identify more representative features from candidate features to determine whether a user is an anomalous user. Since the format of each historical user asset-related data is the same and they have the same fields—specifically, the number of items in user identity information, user property information, user transaction information, user loan information, and user repayment information is the same, and the items including whether the user owns property and user deposit information are also identical—the Relief algorithm can be used to calculate the importance value of each candidate feature. This allows for the selection of candidate features with higher importance values ​​as the features to be identified, while removing candidate features with lower importance values, thus reducing the amount of data and achieving higher accuracy.

[0114] For example, the Relief algorithm can randomly select a sample and a feature from a historical user asset-related dataset. The sample is a historical user asset-related data point. The algorithm records the value of the sample under the first feature, identifies the values ​​of similar samples under the same feature, and calculates the similarity difference for that sample under that feature. Then, it identifies the values ​​of dissimilar samples under the same feature and calculates the dissimilarity difference for that sample under that feature. Finally, it squares the similarity difference, takes the negative, and sums it with the square of the dissimilarity difference to obtain the weight of that sample under that feature. This process is repeated to calculate the weights of all samples under that feature. Finally, the sum of all sample weights under that feature is divided by the number of samples to obtain the importance value of that feature. In this embodiment, the method for calculating the importance values ​​of each candidate feature in the historical user asset-related dataset using the Relief algorithm is not limited.

[0115] Step 704: Candidate features with importance values ​​greater than a preset importance threshold are identified as features to be identified, and users and their features to be identified in the historical user asset related dataset are identified as the training user asset related dataset.

[0116] In this embodiment, a preset importance threshold is used to determine the features that are more representative of whether a user is an aberrant user from the candidate features. The preset importance threshold can be a pre-set value, or it can be related to the importance value of each candidate feature. For example, the preset importance threshold can be the median, quartile, or other importance values ​​of each candidate feature.

[0117] The abnormal user identification method provided in this embodiment obtains a historical user asset-related dataset. This dataset includes at least one record of historical user asset-related data for a normal user and at least one record of historical user asset-related data for an abnormal user. The historical user asset-related data includes at least one user asset-related attribute. Based on the user asset-related attribute information of each user, at least one candidate feature of the historical user asset-related data is determined. The Relief algorithm is used to calculate the importance value of each candidate feature in the historical user asset-related dataset. Candidate features with importance values ​​greater than a preset importance threshold are identified as features to be identified, and the users and their features to be identified in the historical user asset-related dataset are used as the training dataset. Because the Relief algorithm is used to calculate the importance value of each candidate feature, and candidate features with importance values ​​greater than a preset importance threshold are identified as features to be identified, the features to be identified are those in the user asset-related data that have a higher correlation with whether a user is an abnormal user. Therefore, more accurate features to be identified can be determined, which in turn helps to determine more accurate optimized parameters and obtain a default user identification model with higher identification accuracy.

[0118] As an optional implementation, based on Embodiment 4, step 703, which uses the Relief algorithm to calculate the importance values ​​of each candidate feature in the historical user asset related dataset, is refined. The refinement of step 703 includes the following steps.

[0119] For each candidate feature in the historical user asset-related dataset, perform the following operations:

[0120] For each piece of historical user asset-related data, perform the following operations:

[0121] Obtain the candidate feature values ​​of the historical user asset-related data itself, the candidate feature values ​​of the adjacent similar samples of the historical user asset-related data, and the candidate feature values ​​of the adjacent dissimilar samples of the historical user asset-related data.

[0122] Based on the candidate feature values ​​of the historical user's assets, the candidate feature values ​​of the same category, and the candidate feature values ​​of different categories, calculate the difference between the same category and the difference between the different categories of the historical user's assets.

[0123] The average value of the candidate feature is determined by subtracting the squared differences between the squared differences of different categories from the squared differences of the same category in the historical user asset-related data.

[0124] In this embodiment, for each piece of historical user asset-related data, its own candidate feature value is the specific numerical value of each feature it possesses. Similar samples are historical user asset-related data belonging to either normal or abnormal users, similar to the historical user asset-related data. Adjacent similar samples are those among similar samples of historical asset-related data whose feature value for the candidate feature is closest to the historical user asset-related data's own candidate feature value. Similar candidate feature values ​​are the feature values ​​of adjacent similar samples of historical asset-related data under that feature. Dissimilar samples are historical user asset-related data belonging to both normal and abnormal users, respectively. Adjacent dissimilar samples are those among dissimilar samples of historical asset-related data whose feature value for the candidate feature is closest to the historical user asset-related data's own candidate feature value. Dissimilar candidate feature values ​​are the feature values ​​of adjacent dissimilar samples of historical asset-related data under that feature.

[0125] The difference between similar categories is the difference between the candidate feature value of the historical asset-related data under this feature and the candidate feature value of the same category. The difference between dissimilar categories is the difference between the candidate feature value of the historical asset-related data under this feature and the candidate feature value of dissimilar categories.

[0126] The average of the squared differences between different categories and the squared differences between the same category in the historical user asset-related data under this feature is determined as the importance value of the candidate feature. The larger the importance value, the stronger the classification ability of the feature; the smaller the importance value, the weaker the classification ability of the feature.

[0127] The abnormal user identification method provided in this embodiment performs the following operations for each candidate feature in the historical user asset-related dataset: For each piece of historical user asset-related data, the following operations are performed: The method acquires the candidate feature value of the historical user asset-related data itself, the candidate feature values ​​of adjacent similar samples of the historical user asset-related data, and the candidate feature values ​​of adjacent dissimilar samples of the historical user asset-related data. Based on the candidate feature values ​​of the historical user asset-related data itself, the similar candidate feature values, and the dissimilar candidate feature values, the similar difference and the abnormal difference of the historical user asset-related data are calculated. The average of the squared dissimilar difference and the squared similar difference of each historical user asset-related data is determined as the importance value of the candidate feature. Since the similar difference and the dissimilar difference can reflect the contribution of the feature to the sample classification, determining the average of the squared dissimilar difference and the squared similar difference of each historical user asset-related data as the importance value of the candidate feature can accurately reflect the contribution of each candidate feature to determining whether a user is an abnormal user.

[0128] Example 5

[0129] Figure 6 This is a flowchart illustrating the abnormal user identification method provided in Embodiment 5 of this application, as follows: Figure 6 As shown, the abnormal user identification method provided in this embodiment, based on any of the above embodiments, includes steps 901 to 903 before step 701, which involves obtaining a dataset related to historical user assets.

[0130] Step 901: Convert the user asset-related attribute information, which is not represented by a numerical value, into user asset-related attribute information that is represented by a numerical value.

[0131] In this embodiment, non-numerical user asset-related attribute information refers to user asset-related attribute information that is textual or logical values. For user asset-related attribute information represented by logical values, it can be converted into 1 and 0 according to the logical yes or no. For example, having a car is represented as 1, and not having a car is represented as 0. For user asset-related attribute information represented by text, it can be converted into different numbers according to the category of text content. For example, a user's highest education level of high school, bachelor's degree, master's degree, and doctorate is converted into 1, 2, 3, and 4.

[0132] Step 902: Calculate the average value of user asset-related attribute information, expressed in numerical form, over at least one preset number of days.

[0133] In this embodiment, for user asset-related attribute information represented by numerical values, its average value over a preset number of days can be obtained. For example, for user income information, its average value over the most recent 7 days, 30 days, 60 days, 90 days, 180 days, 360 days, 720 days, 1080 days, and 1800 days can be obtained respectively. For user asset-related attribute information that is not represented by numerical values, since it has been converted into numerical user asset-related attribute information in step 901, its average value over at least one preset number of days can also be calculated.

[0134] Step 903: Normalize the average value of the user asset-related attribute information represented by numerical values ​​over each preset number of days to obtain at least one candidate feature.

[0135] In this embodiment, for any user asset-related attribute information, its mean over each preset number of days is normalized to obtain multiple candidate features. For example, for user income information, its mean over the past 7 days and 30 days is calculated, and the mean over the 7 days and 30 days is normalized respectively to obtain the 7-day mean feature and the 30-day mean feature of user income information.

[0136] The abnormal user identification method provided in this embodiment converts user asset-related attribute information, which is not represented by numerical values, into user asset-related attribute information that is represented by numerical values; calculates the mean of the user asset-related attribute information represented by numerical values ​​over at least one preset number of days; and normalizes the mean of the user asset-related attribute information represented by numerical values ​​over each preset number of days to obtain at least one candidate feature. Since converting the user asset-related attribute information, which is not represented by numerical values, into numerical values ​​facilitates the subsequent extraction of candidate features of the user asset-related attribute information, and normalizing the mean of the user asset-related attribute information represented by numerical values ​​over each preset number of days, multiple candidate features of the user asset-related attribute information in different time periods can be obtained, so as to determine the features to be identified that truly affect whether a user is an abnormal user from multiple candidate features.

[0137] As an optional implementation, based on any of the above embodiments, the defaulting user identification model is a deep neural network-isolated forest model, which consists of a deep neural network and an isolated forest algorithm, and the output of the deep neural network is used as the input of the isolated forest algorithm; and for step 202, inputting the asset-related data of the user to be identified into the trained defaulting user identification model, and using the output of the trained defaulting user identification model to determine whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user, the refinement of step 202 includes steps 1001 to 1002.

[0138] Step 1001: Input the asset-related data of the user to be identified into the deep neural network, and output the outliers of the user to be identified by the isolated forest model.

[0139] In this embodiment, the defaulting user identification model uses a deep neural network (DNN) – the Isolation Forest model – as input to the Isolation Forest algorithm (iForestd). The DNN can be a four-layer network: an input layer, three hidden layers, and an output layer. The DNN's output serves as the input to iForestd. After inputting the asset-related data of the user to be identified into the deep neural network, the Isolation Forest model can output outliers corresponding to the user's asset-related data. These outliers can indicate the user's potential default.

[0140] Step 1002: If it is determined that the abnormal value of the user to be identified is greater than the preset abnormal threshold, then the user to be identified is identified as an abnormal user.

[0141] In this embodiment, the preset anomaly threshold represents the threshold at which a user is likely to default, and is used to determine whether a user is an abnormal user with the potential to default. When the anomaly value of the user to be identified is greater than the preset anomaly threshold, the user is likely to default and will default; in this case, the user to be identified is determined to be an abnormal user. When the anomaly value of the user to be identified is less than or equal to the preset anomaly threshold, the user is not likely to default and will not default; in this case, the user to be identified is determined to be a normal user.

[0142] For example, when using the trained defaulting user identification model provided in this application to predict whether a user is likely to default, the preset anomaly threshold can be 0.5. If the anomaly value of the user to be identified is greater than 0.5, it means that the user will default; if it is less than or equal to 0.5, it means that the user will not default.

[0143] For abnormal users identified by the trained defaulter identification model who are likely to default, users can be notified in advance to repay their loans, thus providing greater assurance for the recovery of funds by banks and other financial institutions.

[0144] The abnormal user identification method provided in this embodiment inputs the asset-related data of the user to be identified into a deep neural network and outputs the abnormal value of the user to be identified by the isolated forest model. If the abnormal value of the user to be identified is determined to be greater than a preset abnormal threshold, the user to be identified is identified as an abnormal user. Since the method of inputting the asset-related data of the user to be identified into a deep neural network and outputting the abnormal value of the user to be identified by the isolated forest model can quickly determine whether the user is an abnormal user based on the abnormal value of the user to be identified, thus improving the identification efficiency of the default user identification model.

[0145] Example 6

[0146] Figure 7 This is a structural schematic diagram of the abnormal user identification device provided in Embodiment Six of this application, as shown below. Figure 7 As shown, the abnormal user identification device 70 provided in this embodiment includes an acquisition module 71 and an identification module 72.

[0147] The acquisition module 71 is used to acquire data related to the assets of the user to be identified; the data related to the assets of the user to be identified includes at least one feature to be identified; the feature to be identified is a feature related to user assets.

[0148] The identification module 72 is used to input the asset-related data of the user to be identified into the trained default user identification model, and to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user. The trained default user identification model includes at least one optimized parameter. The optimized parameter is obtained by training the user asset-related dataset and the improved particle swarm optimization algorithm to obtain the optimized parameter. The improved particle swarm optimization algorithm trains the parameter to be optimized by finding the optimal solution to obtain the optimized parameter.

[0149] As an optional implementation, the inertia weight ω of the improved particle swarm optimization algorithm is expressed as: Where, ω max ω represents the maximum value of the inertia weight. min f is the minimum value of the inertial weight. max For maximum group fitness, f min Let f be the minimum group fitness, and f be the particle fitness.

[0150] As an optional implementation, the abnormal user identification device 70 further includes a training module, which trains the model to train at least one parameter to be optimized using a training user asset-related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter; the training user asset-related dataset includes at least one training user asset-related data for a normal user and at least one training user asset-related data for an abnormal user; the training user asset-related data includes at least one feature to be identified; and the trained default user identification model is determined based on the at least one optimized parameter.

[0151] As an optional implementation, the training module is specifically used to perform the following operations for any parameter to be optimized in the default user identification model, in order to obtain at least one optimized parameter: taking the parameter to be optimized as the position variable of the improved particle swarm optimization algorithm, and taking the identification accuracy of the default user identification model corresponding to the parameter to be optimized on the training user asset-related dataset as the fitness of the improved particle swarm optimization algorithm; solving for the optimal position solution of the improved particle swarm optimization algorithm with the goal of maximizing the fitness; and determining the optimal position solution as the optimized parameter.

[0152] As an optional implementation, the training module is further configured to: calculate the fitness of each particle in iteration period t based on its position; t is a positive integer greater than or equal to 1; determine the maximum and minimum population fitness of the particle swarm in iteration period t based on the fitness of each particle in iteration period t; determine the historical maximum fitness of each particle based on the t fitnesss of each particle in t iteration periods; determine the historical maximum population fitness of the particle swarm based on the t maximum population fitnesss of the particle swarm in t iteration periods; iterate the velocity and position of each particle according to the inertia weight ω until the historical maximum population fitness is greater than or equal to a preset fitness, or the iteration period t is greater than or equal to a preset number of iteration periods; and determine the position corresponding to the historical maximum population fitness as the optimal solution position of the particle swarm optimization algorithm.

[0153] As an optional implementation, the training module is further used to iterate the velocity and position of each particle according to the following formula:

[0154] Where i represents the i-th particle, t is the iteration period number, and ω t Let c1 and c2 be the inertia weights for the iteration period t, and c1 and c2 be learning factors that are non-negative. Let be the velocity of particle i during iteration period t. Let be the velocity of particle i after the iteration, and rand(0,1) be a random number between (0,1). Let be the position corresponding to the historical maximum fitness of particle i over t iteration cycles. The position of the particle. This represents the position corresponding to the historical maximum population fitness of the particle swarm over t iteration cycles.

[0155] As an optional implementation, the training module is further configured to: acquire a historical user asset-related dataset; the historical user asset-related dataset includes at least one historical user asset-related data for a normal user and at least one historical user asset-related data for an abnormal user; the historical user asset-related data includes at least one user asset-related attribute information; determine at least one candidate feature of the historical user asset-related data based on the user asset-related attribute information of each user; calculate the importance value of each candidate feature in the historical user asset-related dataset using the Relief algorithm; determine the candidate features with importance values ​​greater than a preset importance threshold as features to be identified, and determine the users and their features to be identified in the historical user asset-related dataset as the training user asset-related dataset.

[0156] As an optional implementation, the training module is further configured to perform the following operations for each candidate feature in the historical user asset related dataset: For each piece of historical user asset related data, the following operations are performed: Obtain the candidate feature value of the historical user asset related data itself, the candidate feature values ​​of the adjacent similar samples of the historical user asset related data, and the candidate feature values ​​of the adjacent dissimilar samples of the historical user asset related data; Calculate the similarity difference and the outlier difference of the historical user asset related data based on the candidate feature value of the historical user asset related data itself, the similarity candidate feature value, and the outlier candidate feature value; Determine the average value of the square of the outlier difference minus the square of the similarity difference as the importance value of the candidate feature.

[0157] As an optional implementation, the training module is further configured to: convert user asset-related attribute information represented in non-numerical form into user asset-related attribute information represented in numerical form; calculate the mean of the user asset-related attribute information represented in numerical form over at least one preset number of days; and normalize the mean of the user asset-related attribute information represented in numerical form over each preset number of days to obtain at least one candidate feature.

[0158] As an optional implementation, the defaulting user identification model is a deep neural network-isolation forest model, which consists of a deep neural network and an isolation forest algorithm, with the output of the deep neural network serving as the input of the isolation forest algorithm. The identification module is specifically used to input the asset-related data of the user to be identified into the deep neural network, and the isolation forest algorithm outputs the outlier value of the user to be identified. If the outlier value of the user to be identified is determined to be greater than a preset outlier threshold, then the user to be identified is identified as an outlier user.

[0159] The abnormal user identification device provided in this embodiment can execute the abnormal user identification method provided in any one of the embodiments one to five above. The specific implementation method and principle are similar, and will not be described again here.

[0160] Example 7

[0161] Figure 8 This is a block diagram of an electronic device provided according to Embodiment Seven of this application, such as Figure 8 As shown, the electronic device 80 provided in this embodiment includes a processor 81 and a memory 82 that is communicatively connected to the processor 81.

[0162] Memory 82 stores computer-executed instructions.

[0163] The processor 81 executes computer execution instructions stored in the memory 82 to implement the abnormal user identification method provided in any of the above embodiments.

[0164] The memory 82 and the processor 81 can be connected via a bus.

[0165] The memory 82 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk, etc.

[0166] As an optional implementation, the electronic device 80 also includes a transceiver.

[0167] The transceiver is connected to the memory and processor circuitry. It is used to send and receive data.

[0168] Optionally, the electronic device 80 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the abnormal user identification method provided in any of the above embodiments.

[0169] This invention also provides a computer-readable storage medium storing computer-executable instructions. When executed by a processor, these instructions are used to implement the abnormal user identification method provided in any of the above embodiments. Exemplarily, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), magnetic tape, floppy disk, or optical data storage device, etc.

[0170] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the abnormal user identification method provided in any of the above embodiments.

[0171] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can be implemented in other ways. For example, the module division in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, unless otherwise specified, the functional units in the various embodiments of this application can be integrated into one module, or each module can exist physically separately, or two or more modules can be integrated together. The integrated modules described above can be implemented in hardware or as software program modules.

[0172] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0173] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0174] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0175] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for identifying abnormal users, characterized in that, include: Obtain asset-related data of the user to be identified; The user asset-related data to be identified includes at least one feature to be identified; The features to be identified are user asset-related features; The asset-related data of the user to be identified is input into the trained default user identification model, and the trained default user identification model is used to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user; the trained default user identification model includes at least one optimized parameter; The optimized parameters were obtained by training a user asset-related dataset and an improved particle swarm optimization algorithm on the parameters to be optimized. The improved particle swarm optimization algorithm trains the parameters to be optimized by finding the optimal solution to obtain the optimized parameters; the defaulting user identification model is a deep neural network-isolated forest model, which consists of a deep neural network and an isolated forest algorithm, and the output of the deep neural network is used as the input of the isolated forest algorithm; The inertia weight of the improved particle swarm algorithm Represented as: ;in, This represents the maximum value of the inertia weight. This represents the minimum inertial weight. To maximize group fitness, For minimum group fitness, For particle fitness.

2. The method according to claim 1, characterized in that, Before inputting the asset-related data of the user to be identified into the trained default user identification model, and before using the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user, the process further includes: At least one parameter to be optimized is trained using a training user asset-related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter; the training user asset-related dataset includes at least one training user asset-related data for a normal user and at least one training user asset-related data for an abnormal user; the training user asset-related data includes at least one feature to be identified. The trained defaulting user identification model is determined based on at least one optimized parameter.

3. The method according to claim 2, characterized in that, The process of training at least one parameter to be optimized using a training user asset-related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter includes: For any parameter to be optimized in the defaulting user identification model, perform the following operations to obtain at least one optimized parameter: The parameters to be optimized are used as the position variables of the improved particle swarm optimization algorithm, and the recognition accuracy of the defaulting user identification model corresponding to the parameters to be optimized on the training user asset-related dataset is used as the fitness of the improved particle swarm optimization algorithm. The optimal position solution of the improved particle swarm optimization algorithm is found with the goal of maximizing fitness. The optimal location solution is determined as the optimized parameters.

4. The method according to claim 3, characterized in that, The process of finding the optimal position solution in the improved particle swarm optimization algorithm with the goal of maximizing fitness includes: In iteration period t, the fitness of each particle in iteration period t is calculated based on the position of each particle; where t is a positive integer greater than or equal to 1. The maximum and minimum population fitness of the particle swarm at iteration period t are determined based on the fitness of each particle at iteration period t. The historical maximum fitness of each particle is determined based on the t fitness values ​​of each particle over t iterations. The historical maximum population fitness of the particle swarm is determined based on the t maximum population fitnesss in t iteration cycles. Based on inertia weight Iterate over the velocity and position of each particle until the historical maximum population fitness is greater than or equal to a preset fitness, or the iteration period t is greater than or equal to a preset number of iteration periods; The position corresponding to the maximum fitness of the historical population is determined as the optimal solution position of the particle swarm optimization algorithm.

5. The method according to claim 4, characterized in that, The inertial weight Iterate over the velocity and position of each particle, including: Iterate through the velocity and position of each particle using the following formula: ； ； ； ； Where i represents the i-th particle, and t is the number of iteration periods. The inertia weight is the inertia weight for the iteration period t. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. , For learning factors and , It is a non-negative number. Let be the velocity of particle i during iteration period t. Let be the velocity of particle i after the iteration, and rand(0,1) be a random number between (0,1). Let i be the particle position corresponding to the historical maximum fitness of particle i over t iteration cycles. For particle position, The position of the particle corresponding to the historical maximum population fitness of the particle swarm over t iterations.

6. The method according to any one of claims 2-5, characterized in that, Before training on at least one parameter to be optimized using a training user asset-related dataset and an improved particle swarm optimization algorithm to obtain at least one optimized parameter, the method further includes: Obtain a historical user asset-related dataset; the historical user asset-related dataset includes at least one set of historical user asset-related data for a normal user and at least one set of historical user asset-related data for an abnormal user; the historical user asset-related data includes at least one user asset-related attribute information. Based on the user asset-related attribute information of each user, at least one candidate feature of historical user asset-related data is determined; The Relief algorithm is used to calculate the importance values ​​of each candidate feature in the historical user asset-related dataset; Candidate features with importance values ​​greater than a preset importance threshold are identified as features to be identified, and users and their features to be identified in the historical user asset related dataset are identified as the training user asset related dataset.

7. The method according to claim 6, characterized in that, The calculation of importance values ​​for each candidate feature in the historical user asset-related dataset using the Relief algorithm includes: For each candidate feature in the historical user asset-related dataset, perform the following operations: For each piece of historical user asset-related data, perform the following operations: Obtain the candidate feature values ​​of the historical user asset-related data itself, the candidate feature values ​​of the same type of adjacent similar samples of the historical user asset-related data, and the candidate feature values ​​of the opposite type of adjacent dissimilar samples of the historical user asset-related data; Calculate the similarity difference and abnormality difference of the historical user asset-related data based on its own candidate feature values, similar candidate feature values, and different candidate feature values; The average value of the candidate feature is determined by subtracting the squared differences between the squared differences of different categories from the squared differences of the same category in the historical user asset-related data.

8. The method according to claim 6, characterized in that, The step of determining at least one candidate feature of historical user asset-related data based on the user asset-related attribute information of each user also includes: Convert user asset-related attribute information that is not represented by numbers into user asset-related attribute information that is represented by numbers; Calculate the average of user asset-related attribute information, expressed in numerical form, over at least one preset number of days; The mean values ​​of user asset-related attribute information, expressed in numerical form, are normalized over each preset number of days to obtain at least one of the candidate features.

9. The method according to claim 1, characterized in that, The step of inputting the asset-related data of the user to be identified into the trained default user identification model, and using the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user, includes: The asset-related data of the user to be identified is input into a deep neural network, and the outlier value of the user to be identified is output by the isolated forest algorithm. If the outlier value of the user to be identified is determined to be greater than the preset outlier threshold, then the user to be identified will be identified as an outlier user.

10. An abnormal user identification device, characterized in that, include: The acquisition module is used to acquire data related to the assets of the user to be identified; the data related to the assets of the user to be identified includes at least one feature to be identified; The features to be identified are user asset-related features; The identification module is used to input the asset-related data of the user to be identified into the trained default user identification model, and use the trained default user identification model to output whether the user to be identified corresponding to the asset-related data of the user to be identified is an abnormal user; the trained default user identification model includes at least one optimized parameter; The optimized parameters were obtained by training a user asset-related dataset and an improved particle swarm optimization algorithm on the parameters to be optimized. The improved particle swarm optimization algorithm trains the parameters to be optimized by finding the optimal solution to obtain the optimized parameters; the defaulting user identification model is a deep neural network-isolated forest model, which consists of a deep neural network and an isolated forest algorithm, and the output of the deep neural network is used as the input of the isolated forest algorithm; The inertia weight of the improved particle swarm algorithm Represented as: ;in, This represents the maximum value of the inertia weight. This represents the minimum inertial weight. To maximize group fitness, For minimum group fitness, For particle fitness.

11. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-9.

13. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-9.

Images