WO2023221359A1

WO2023221359A1 - User security level identification method and apparatus based on multi-stage time sequence and multiple tasks

Info

Publication number: WO2023221359A1
Application number: PCT/CN2022/121543
Authority: WO
Inventors: 王磊; 宋孟楠; 苏绥绥; 郑彦
Original assignee: 北京淇瑀信息科技有限公司
Priority date: 2022-05-19
Filing date: 2022-09-27
Publication date: 2023-11-23
Also published as: CN114742645B; CN114742645A

Abstract

The present application relates to a user security level identification method and apparatus based on a multi-stage time sequence and multiple tasks. The method comprises: generating a plurality of stage sets according to all users and user stages corresponding thereto; sequentially arranging the plurality of stage sets according to a time sequence; performing multi-task training on an (n+1)th group of initial models according to an (n+1)th stage set and an nth group of model parameter vectors, so as to generate an (n+1)th group of model parameter vectors, wherein n is a positive integer; generating a plurality of groups of stage scoring models on the basis of the plurality of groups of model parameter vectors until the training of the plurality of stage sets is completed; and performing security level identification on the current user by means of the plurality of groups of stage scoring models. The present application can start from actual problems and application scenarios and integrally improve a multi-task machine learning method from the perspective of model samples and from model parameters, thereby ensuring the security of user data and transactions in an application system.

Description

User security level identification method and device based on multi-stage timing and multi-task

Technical field

The present application relates to the field of computer information processing. Specifically, it relates to a user security level identification method, device, electronic equipment and computer-readable medium based on multi-stage timing and multi-tasking.

Background technique

Machine learning, which uses useful information in historical data to help analyze future data, usually requires a large amount of labeled data to train a good learner. Deep learning models are a typical machine learning model. Because this type of model is a neural network with many hidden layers and many parameters, it usually requires millions of data samples to learn accurate parameters. However, some applications, including medical image analysis, cannot meet this data requirement because labeling the data requires a lot of manual labor. In these cases, multi-task learning (MTL) can help alleviate this data sparsity problem by using useful information from other related learning tasks.

The multi-task learning task is to predict the labels of unseen data based on a training data set (containing training data instances and their labels). The "quality" of the data in the training data set plays a crucial role in the effectiveness of multi-task learning. However, in actual application scenarios, the data in the training data set is difficult to accurately reflect the real situation.

The above information disclosed in the Background section is only for enhancement of understanding of the context of the application and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art.

Contents of the invention

In view of this, this application provides a user security level identification method, device, electronic equipment and computer-readable medium based on multi-stage timing and multi-task, which can be based on practical problems and application scenarios, and from the perspective of model samples and model parameters. Improve multi-task machine learning methods to ensure application system user data security and transaction security.

Additional features and advantages of the invention will be apparent from the detailed description which follows, or, in part, may be learned by practice of the invention.

According to one aspect of the present application, a user security level identification method based on multi-stage timing and multi-tasking is proposed. The method includes: generating multiple stage sets based on all users and their corresponding user stages; Arrange; perform multi-task training on the n+1th group of initial models based on the n+1th stage set and the nth group of model parameter vectors in turn, and generate the n+1th group of model parameter vectors, n is a positive integer; until the multiple After the stage set training is completed, multiple groups of stage scoring models are generated based on multiple groups of model parameter vectors; the security level of the current user is identified through the multiple groups of stage scoring models.

Optionally, generating multiple stage sets based on all users and their corresponding user stages includes: determining multiple user stages based on product characteristics; matching the user stage corresponding to each user among all users with the multiple user stages. ;Assign users to the stage set corresponding to their user stage based on the matching results.

Optionally, generating multiple stage sets based on the total number of users and their corresponding user stages also includes: determining a label strategy for each user stage; and assigning sample labels to users in each stage set according to the label strategy.

Optionally, perform multi-task training on the n+1 group of initial models based on the n+1 group of stage sets and the n group of model parameter vectors, and generate the n+1 group of model parameter vectors, including: extracting the multiple stages The first stage set in the set; input the first stage set into the first set of initial models to generate the first set of model parameter vectors; based on the n+1th stage set and the nth set of model parameter vectors, the n+1th set of initial The model performs multi-task training and generates the n+1th set of model parameter vectors, where n is a positive integer.

Optionally, perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, and generate the n+1 group of model parameter vectors, which also includes: Determine a set of machine learning models; assign sample labels to historical users according to the label strategy corresponding to each user stage; train the n+1th group of machine learning models through historical users with sample labels, and generate the n+1th group of initial Model, n is a positive integer.

Optionally, inputting the first-stage set into the first group of initial models to generate the first group of model parameter vectors includes: inputting the user information in the first-stage set into the first group of initial models respectively; Model training is performed based on the user information and its corresponding labels. After the training is completed, the first set of model parameter vectors is generated.

Optionally, perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, and generate the n+1 group of model parameter vectors, including: generating an update vector; The update vector is weighted and superimposed into the model parameter vector of the n+1th group of initial models; the user information in the n+1th stage set is input into the n+1th group of initial models after the model parameter vector is updated for multi-tasking. Training; after training, generate the n+1th set of model parameter vectors.

Optionally, generating an update vector includes: nonlinearly transforming the nth group of model parameter vectors to generate an update vector; or generating the first to nth groups through nonlinear transformation of the first to nth group of model parameter vectors. update vector.

Optionally, input the user information in the n+1th stage set into the n+1th group of initial models after the model parameter vector is updated for multi-task training, including: inputting the user information in the n+1th stage set Input the updated model parameter vectors into the n+1 initial model respectively; the n+1 initial model performs multi-task training based on user information and its corresponding labels; when the loss function during the training process does not meet the convergence conditions, Redetermine the initial model parameters of the n+1 initial model to perform multi-task training again; when the loss function meets the convergence condition, complete the multi-task training of the n+1 initial model.

Optionally, re-determine the initial model parameters of the (n+1)th group of initial models to perform model training again, including: re-training the model on the (n+1)th group of initial models to generate new initial model parameters; or re-determining the convergence conditions Model training is performed again on the n+1 group of initial models to generate new initial model parameters.

According to one aspect of the present application, a user security level identification device based on multi-stage timing and multi-tasking is proposed. The device includes: a stage module for generating multiple stage sets based on all users and their corresponding user stages; a sorting module, It is used to arrange multiple stage sets in sequence; the training module is used to perform multi-task training on the n+1th group of initial models based on the n+1th stage set and the nth group of model parameter vectors, and generate the n+1th group of initial models. A set of model parameter vectors, n is a positive integer; a model module, used to generate multiple sets of scoring models based on multiple sets of model parameter vectors until the multiple stages of collective training are completed; a grading module, used to use the multiple sets of scoring models to The current user performs security level identification, and the security level of the current user is determined based on the identification result.

According to one aspect of the present application, an electronic device is proposed. The electronic device includes: one or more processors; a storage device for storing one or more programs; when one or more programs are processed by one or more processors, Execution causes one or more processors to implement the method as above.

According to one aspect of the present application, a computer-readable medium is proposed, on which a computer program is stored. When the program is executed by a processor, the above method is implemented.

According to the user security level identification method, device, electronic equipment and computer-readable medium based on multi-stage timing and multi-task of the present application, multiple stage sets are generated according to the total number of users and their corresponding user stages; the multiple stage sets are generated in time sequence Arrange in order; perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, generate the n+1 group of model parameter vectors, and generate the n group of model parameter vectors, n is a positive integer; until the multi-stage set training is completed, multiple groups of stage scoring models are generated based on multiple groups of model parameter vectors; through the multi-group stage scoring model, the security level identification of the current user can be based on actual problems and Starting from the application scenario, the multi-task machine learning method is improved as a whole from the perspective of model samples and model parameters, thereby ensuring the security of user data and transaction security in the application system.

It should be understood that the above general description and the following detailed description are only exemplary and do not limit the present application.

Description of the drawings

The above and other objects, features and advantages of the present application will become more apparent by describing in detail example embodiments thereof with reference to the accompanying drawings. The drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic diagram of a sample space according to an exemplary embodiment.

Figure 2 is a schematic diagram of a sample space according to another exemplary embodiment.

Figure 3 is a system block diagram of a user security level identification method and device based on multi-stage sequential multi-tasking according to an exemplary embodiment.

FIG. 4 is a flow chart of a user security level identification method based on multi-stage sequential multi-tasking according to an exemplary embodiment.

FIG. 5 is a schematic diagram of a user security level identification method based on multi-stage sequential multi-tasking according to another exemplary embodiment.

FIG. 6 is a flowchart of a user security level identification method based on multi-stage sequential multi-tasking according to another exemplary embodiment.

FIG. 7 is a schematic diagram of a user security level identification method based on multi-stage sequential multi-tasking according to another exemplary embodiment.

FIG. 8 is a block diagram of a user security level identification device based on multi-stage sequential multi-tasking according to an exemplary embodiment.

FIG. 9 is a block diagram of an electronic device according to an exemplary embodiment.

Figure 10 is a block diagram of a computer-readable medium according to an exemplary embodiment.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in various forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concepts of the example embodiments. To those skilled in the art. The same reference numerals in the drawings represent the same or similar parts, and thus their repeated description will be omitted.

Furthermore, the described features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to provide a thorough understanding of embodiments of the present application. However, those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other instances, well-known methods, apparatus, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the present application.

The block diagrams shown in the figures are functional entities only and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and/or processor devices and/or microcontroller devices. entity.

The flowcharts shown in the drawings are only illustrative, and do not necessarily include all contents and operations/steps, nor must they be performed in the order described. For example, some operations/steps can be decomposed, and some operations/steps can be merged or partially merged, so the actual order of execution may change according to the actual situation.

It will be understood that, although the terms first, second, third, etc. may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another component. Accordingly, a first component discussed below may be referred to as a second component without departing from the teachings of the present concepts. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of exemplary embodiments, and the modules or processes in the accompanying drawings are not necessarily necessary to implement the present application, and therefore cannot be used to limit the protection scope of the present application.

In order to facilitate the understanding of the content of this application, the Internet financial service platform will be taken as an example to illustrate the actual application. As shown in Figure 1, in the Internet financial platform, after the user registers as a website member, he will apply for financial resources before actual services. The Internet service platform will score the user's financial risk based on the user's basic information. The score is higher than the threshold. High-quality users are allowed to borrow financial resources, and users whose scores are below the threshold will no longer provide financial services. Not everyone of high-quality users will actually borrow financial resources, and only some users will occupy financial resources when they are actually needed. High-quality users may use financial resources on the first day when their qualifications for borrowing financial resources are approved, or they may use financial resources within 30 days of being approved to use financial resources, or they may use financial resources at a further time. Of course, there are Quality users who do not use financial resources at all. Among high-quality users who use financial resources, some users may default after the use period of the financial resources expires. After the default exceeds a certain period, the user will enter the collection process. Some users may in a shorter period of time The defaulted resources will be returned within a certain period of time, and some users may take longer to return the resources.

The applicant in this case found that in each of the above steps, there will be user losses (users are denied service or users actively choose not to provide subsequent services), and in the modeling samples of each stage, there will be losses in this process. Users who performed well in each stage were used as positive and negative samples, and users screened in the previous step were used as unlabeled samples. Positive and negative samples and unlabeled samples were unified to form a sample set, and semi-supervised machine learning methods were used for modeling.

In each of the above stages, the training data set is established by the users in the current stage to train the machine learning model. That is to say, in each stage, the training set data is established based on the users after multiple screenings. thereby establishing an evaluation model. As shown in Figure 2, the real sample space is the full sample space, while the actual training set data is a biased sample space. In this case, the trained samples cannot truly reflect the actual situation.

From the above description, it can be seen that in the sample set in the actual application stage, due to limitations of the actual situation, the actual modeling sample space is far smaller than the real full sample space, which results in the accuracy and precision of the built model. There is a bias. In order to solve this problem, the applicant of this case proposed a user security level identification method based on multi-stage timing and multi-tasking. The content of this application will be described in detail below with the help of specific embodiments.

Figure 3 is a system block diagram of a user security level identification method and device based on multi-stage interaction sequence (MSIS) according to an exemplary embodiment.

As shown in Figure 3, the system architecture 30 may include

terminal devices

301, 302, 303, a network 304 and a server 305. The network 304 is used as a medium for providing communication links between the

terminal devices

301, 302, 303 and the server 305. Network 304 may include various connection types, such as wired, wireless communication links, fiber optic cables, etc.

Users can use

terminal devices

301, 302, 303 to interact with the server 305 through the network 304 to receive or send messages, etc. Various communication client applications can be installed on the

terminal devices

301, 302, and 303, such as Internet service applications, shopping applications, web browser applications, instant messaging tools, email clients, social platform software, etc.

The

terminal devices

301, 302, and 303 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

The server 305 may be a server that provides various services, such as a backend management server that provides support for Internet service websites browsed by users using the

terminal devices

301, 302, and 303. The background management server can analyze and process the received user data, and feed back the processing results (such as security level, resource quota) to the administrator of the Internet service website and/or the

terminal device

301, 302, 303.

The server 305 can, for example, obtain user data from the

terminal devices

301, 302, and 303 as full user data; the server 305 can, for example, generate multiple stage sets based on the full number of users and their corresponding user stages; the server 305 can, for example, combine the multiple stage sets in time sequence. Arranged in sequence; the server 305 can, for example, perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors in sequence, and generate the n+1 group of model parameter vectors, where n is a positive integer, n is a positive integer; until the multiple stage set training is completed, multiple sets of stage scoring models are generated based on multiple sets of model parameter vectors; the server 305 can, for example, use the multiple sets of stage scoring models to evaluate the performance of the

terminal devices

301, 302 and 303. Users perform security level identification.

It should be noted that the user security level identification method based on multi-stage timing and multi-tasking provided by the embodiment of the present application can be executed by the server 305 and/or the

terminal devices

301, 302, 303. Correspondingly, the method based on multi-stage timing and multi-tasking The user security level identification device may be provided in the server 305 and/or the

terminal devices

301, 302, and 303. The web pages provided for users to browse the Internet service platform are generally located in

terminal devices

301, 302, and 303.

FIG. 4 is a flow chart of a user security level identification method based on multi-stage sequential multi-tasking according to an exemplary embodiment. The user security level identification method 40 based on multi-stage sequential multi-tasking at least includes steps S402 to S410.

As shown in Figure 4, in S402, multiple stage sets are generated based on all users and their corresponding user stages. Multiple user stages can be determined based on product characteristics; the user stage corresponding to each user among all users can be matched with the multiple user stages; and users can be assigned to stage sets corresponding to their user stages based on the matching results.

The following explanation will continue to take the financial service platform as an example. The users can be individual users or corporate users, and the resources can be financial resources, power resources, water resources, data resources, etc. User information may include basic information authorized by the user, which may be, for example, business account information, user's terminal device identification information, user location information, etc. User information may also include behavioral information, which may be, for example, the user's page operation data, user information, etc. The length of business access, the frequency of user’s business access, etc. The specific content of user information can be determined according to the actual application scenario and is not limited here. In the financial service platform, users can be divided into "service application stage", "resource payment stage" and "overdue stage". The above three stages are related in time sequence according to the business content.

The user information can include the current stage of the user. When the user is in the "overdue stage", it can be seen that the user has passed the "service application stage" and "resource activation stage". At this time, the user needs to be placed separately. into the set of the corresponding stage. In the same way, when a user is in the "resource mobilization stage", he must have passed the "service application stage", and the user also needs to be put into the collection of the corresponding stage.

In one embodiment, the method further includes: determining a label strategy for each user stage; and assigning sample labels to users in each stage set according to the label strategy. More specifically, the "pass" and "reject" labels can be determined for the "service application stage"; the "first day of expenditure", "expenditure within 30 days", and "expenditure within 60 days" labels can be determined for the "resource expenditure stage" ; "Overdue stage" determines "repayment upon reminder", "repayment upon reminder within 30 days", "repayment upon reminder within 60 days"; according to the user performance in the user information, users at different stages are allocated corresponding to that stage Tag of.

In S404, multiple stage sets are arranged in sequence in time sequence. Arrange the above-mentioned stage sets according to their corresponding order of business occurrence, that is, the three stage sets are arranged in the order of "service application stage", "resource mobilization stage" and "overdue stage".

In S406, multi-task training is performed on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, and the n+1 group of model parameter vectors are generated.

In one embodiment, the method further includes: determining a set of machine learning models for each user stage; assigning sample labels to historical users according to label strategies corresponding to each user stage; and classifying the nth group of machines through historical users with sample labels. The learning model performs multi-task training and generates the nth group of initial models, where n is a positive integer. The initial model can be trained separately for each stage and each label through historical data in advance. For example, the initial application model can be generated by training corresponding to the "service application stage", and this group of models can be generated by training corresponding to the "resource mobilization stage", which can include The initial model of energy expenditure on the first day, the initial model of energy expenditure within 30 days, the initial model of energy expenditure within 60 days, etc.

Specifically, for each training set, an initial model is constructed respectively, the user information of each user in the training set is input into the initial model to obtain a predicted label, and the predicted label is compared with the corresponding real label. Yes, determine whether the predicted labels are consistent with the real labels, count the number of predicted labels that are consistent with the real labels, and calculate the proportion of the number of predicted labels that are consistent with the real labels in the number of all predicted labels, if If the proportion is greater than or equal to the preset proportion value, the initial model converges and the initial model after training is obtained. If the proportion is less than the preset proportion value, the parameters in the initial model are adjusted. After adjustment, The initial model re-predicts the predicted label of each object until the proportion is greater than or equal to the preset proportion value. Wherein, the method of adjusting the parameters in the initial model can be carried out by using a stochastic gradient descent algorithm, a gradient descent algorithm or a normal equation.

The first stage set among the multiple stage sets can be extracted; the first stage set is input into the first group of initial models to generate the first group of model parameter vectors; and then the n+1th stage set and the nth group of model parameters are generated. The vector performs multi-task training on the n+1 group of initial models and generates the n+1 group of model parameter vectors, where n is a positive integer.

Figure 5 shows the multi-stage sequential multi-task machine learning model framework introduced in this application. Multi-stage refers to the multiple product stages mentioned above. The above multiple product stages are arranged in time sequence and combined with multi-task learning. Model training is performed in this way, that is, the multi-stage sequential multi-task machine learning model framework in this application is generated.

In a specific application, the users in the first stage set and their corresponding user tags can be input from the input layer. The sharing layer often organizes the user data and then inputs it into the corresponding application initial model in the first stage to obtain the corresponding The model parameter vector of this input data.

Then the users in the second stage set and their corresponding user labels are input from the input layer. The sharing layer frequently organizes the user data, and then inputs it together with the model parameter vector obtained in the first stage into the corresponding first-day dynamic expenditure in the second stage. From the initial model, the initial model of movement and expenditure within 30 days, and the initial model of movement and expenditure within 60 days, the model parameter vectors of the three models corresponding to this input data are obtained.

In subsequent stages, data and models are processed sequentially. Among them, the details of "Perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors to generate the n+1 group of model parameter vectors, n is a positive integer" will be in Detailed descriptions are provided in the corresponding embodiments of FIG. 6 and FIG. 7 .

In S408, until the multiple stage set training is completed, multiple sets of stage scoring models are generated based on multiple sets of model parameter vectors. After all stage sets are trained, multiple sets of trained stage scoring models can be generated based on the model parameters in each current set of initial models.

In S410, the security level of the current user is identified through the multiple sets of stage scoring models. For example, in an actual application scenario, the user information of the current user can be obtained; multiple groups of stage scoring models can be extracted according to the user stage in the user information; for example, if the current user is in the "service application stage", then the "resource dynamic support stage" can be extracted Multiple initial models corresponding to the "overdue stage".

In one embodiment, the multiple sets of stage models can also be arranged in sequence according to their corresponding stage timings; the user information can be input into the multiple sets of stage scoring models in turn to generate multiple sets of stage scores; according to the multiple sets of stage Rating to determine the service provided to said user.

In existing technologies, a user may have fewer ratings during the application phase and may be initially rejected. In this application, a multi-stage sequential multi-task model is used to train multiple models in different stages. In practical applications, these models can be used to calculate the user's score respectively. Choosing the largest score is likely to be the final corresponding situation of the user. , for example, if a user has the highest probability of spending money on the 30th day during the overdue phase, then the user will be assigned preferential information and strategies based on this situation to promote the user to spend money. If the user has the highest risk score of overdue payment for 30 days during the overdue phase, then the user will be determined. Deferred repayment strategies and more.

According to the user security level identification method based on multi-stage timing and multi-task of this application, multiple stage sets are generated according to the total number of users and their corresponding user stages; the multiple stage sets are arranged in sequence according to the n+1th stage set; , the nth group of model parameter vectors perform multi-task training on the n+1th group of initial models, and generate the n+1th group of model parameter vectors, where n is a positive integer; until the multiple stages of collective training are completed, based on the multiple groups of model parameters The vector generates multiple groups of stage scoring models; the method of identifying the security level of the current user through the multi-group stage scoring model can comprehensively improve the multi-task machine learning method from the perspective of model samples and model parameters based on actual problems and application scenarios. , thereby ensuring the security of application system user data and transaction security.

It should be clearly understood that this application describes how to make and use specific examples, but that the principles of this application are not limited to any details of these examples. Rather, these principles can be applied to many other embodiments based on the teachings of this disclosure.

FIG. 6 is a flowchart of a user security level identification method based on multi-stage sequential multi-tasking according to another exemplary embodiment. The process 60 shown in Figure 6 is to perform multi-task training on the n+1th group of initial models based on the n+1th stage set and the nth group of model parameter vectors in sequence in S406 of the process shown in Figure 4, and generate the n+th group of initial models. 1 set of model parameter vectors, n is a positive integer" detailed description.

As shown in Figure 6, in S602, the user information in the first stage set is input into the first set of initial models respectively. As mentioned above, the users in the first stage set and their corresponding user tags can be input from the input layer. The sharing layer often organizes the user data and then inputs it into the corresponding application initial model in the first stage to obtain the corresponding A vector of model parameters for the input data.

In S604, the first set of initial models performs model training based on the user information and its corresponding labels. After the training is completed, the first set of model parameter vectors are generated.

In S606, the model parameter vector of the n+1 group of initial models is updated through the n group of model parameter vectors.

In one embodiment, the model parameter vector of the n+1 group of initial models can be updated through the nonlinear transformation function sigmoid function and the nth group of model parameter vectors; in some embodiments, other nonlinear transformation methods can also be used to The nth set of model parameter vectors is used to update the model parameter vectors of the n+1th set of initial models. This application is not limited to this.

The parameter transfer formula can be expressed as follows:

e _in =g(e _out );

Among them, e _in represents the data input to the current stage after parameter transformation, e _out represents the model parameters of the previous stage, and g() is a nonlinear function. Among them, the purpose of performing nonlinear transformation instead of linear transformation is to make the function corresponding to the model parameters more consistent with the actual situation, rather than a simple straight line division.

For example, if the calculated model parameters are a linear function, y=ax+b, then users above the straight line can be considered good, and users below the straight line are considered bad. In this case, there are some users close to the straight line. Feature points may be divided incorrectly; therefore, when passing parameters, using nonlinear functions to fit the model parameters can make the division of feature points more accurate and consistent with the actual situation.

In one embodiment, the nth set of model parameters can be nonlinearly transformed to generate an update vector; the update vector is weighted and superimposed into the model parameter vector of the n+1th set of initial models. In this case, the formula for parameter output can be written as:

Among them, α _m is the weight of the model parameters of the m-th model in the D-1 stage, that is, the weight of the model parameters of the m-th model in the previous stage, D is the current stage,

is the model parameter vector of the m-th model in the previous stage, and m is the number of models in the D-1 stage.

In another embodiment, the first to nth groups of update vectors may be generated through nonlinear transformation of the first to nth groups of model parameters; the update vectors of the first to nth groups are weighted and superimposed on the In the model parameter vector of n+1 groups of initial models. In this case, the formula can be written as:

in,

Represents the weight of the model parameters of the m1-th model in stage D-1,

Represents the model parameter vector of the m1-th model in stage D-1, and m1 represents the number of models in stage D-1;

Represents the weight of the model parameters of the m2-th model in the D-2 stage,

Represents the model parameter vector of the m2-th model in stage D-2, and m2 represents the number of models in stage D-2;

Represents the weight of the model parameters of the mnth model in stage 1,

Represents the model parameter vector of the mn-th model in the first stage, and mn indicates the number of models in the first stage;

That is, the model parameters of this stage are jointly generated based on the model parameters of multiple models in the D-1 stage (previous stage) to the model parameters of multiple models in the first stage.

More specifically, the weight of each stage above

and

The calculation method can refer to α _m , and α _m can be processed according to the following formula:

in,

represents the weight of the model parameters of the m-th model in the D-1 stage before normalization, m is the number of models in the D-1 stage, <> represents the dot product function, α _m is the m-th model in the D-1 stage the weights of the model parameters of the model,

are the model parameters of the m-th model in the D-1 stage (previous stage), and g() is a nonlinear function.

The model parameters of the previous stage and the model parameters of this stage are weighted and added, and the formula for generating the model parameters of this stage can be as follows:

Among them, the g′() function is a nonlinear transformation function, which is the same as the above g() function. They are all nonlinear transformation functions, such as the sigmoid function. Of course, other nonlinear transformation methods can also be used. This plan does not do this. Specially limited.

in,

is the updated model parameter vector of the first initial model of the current stage,

The model parameter vector of the original first initial model at the current stage. If you need to determine the model parameter vector of the updated nth initial model, you can refer to the calculation method of this formula and replace the model parameter vector, that is, in the formula The model parameter vector of the first initial model is replaced with the model parameter vector of the n-th initial model, and the updated model parameter vector of the n-th initial model is obtained. β ₁ and β ₂ are the weights, and the specific values of β ₁ and β ₂ are The calculation method can refer to α _m , or the user can set it according to the importance of the model parameter vector, which will not be described in detail here in this application.

In S608, the user information in the n+1th stage set is respectively input into the n+1th group of initial models after the model parameter vector is updated for multi-task training.

As mentioned above, in actual calculations, the parameter vector of the model in the previous stage can be transformed and passed into the initial model of the current stage. The parameter vector of the previous stage and the parameter vector of the current stage are weighted and summed to obtain the current stage. Parameter vector of the new initial model.

It is worth noting that the model parameters of the current initial model are models trained in advance based on existing features and labels, directly added to the parameters of the model from the previous stage uploaded, and the new model parameters are used for the current calculation.

On the one hand, the convergence of the model does not accurately judge all features. On the other hand, the initial model parameters after training are not fixed. When there are existing features and labels, as long as the convergence conditions are met, the output will be result. Moreover, for the input user features, the output result of the model is a decimal number between 0 and 1. Therefore, adding the parameters of the imported model to the original model parameters will not necessarily affect the judgment results of the original features. Therefore, whether the parameters of the new model will misjudge existing features and labels.

In S610, after training is completed, the n+1th set of model parameter vectors are generated.

In one embodiment, the user information in the n+1th stage set can be input into the n+1th group of initial models after the model parameter vector is updated; the n+1th group of initial models are based on the user information and their corresponding tags. Perform multi-task training; when the loss function during the training process does not meet the convergence conditions, re-determine the initial model parameters of the n+1 initial model to perform multi-task training again; when the loss function meets the convergence conditions, complete the n+ Multi-task training of an initial set of models.

In actual calculations, if the new model does not meet the convergence conditions for the judgment of the original features, for example, the prediction accuracy does not meet the requirements, the original model can be retrained to obtain a new original model. model parameters, and then pass in the model parameters of the previous stage for judgment, and iterate until the new model obtained by the parameters of the original model and the passed model satisfies the conditions for the judgment of the original features.

In one embodiment, redetermining the initial model parameters of the n+1th group of initial models to perform model training again includes: performing model training again on the n+1th group of initial models to generate new initial model parameters; or redetermining Convergence conditions are used to perform model training again on the n+1th group of initial models to generate new initial model parameters.

In one embodiment, as mentioned above, the trained model will output results as long as it meets the convergence conditions when it already has features and labels. Since the model parameters corresponding to each trained model are not necessarily the same, when the training loss function of multi-task training does not meet the convergence conditions, the initial model can be trained again and a set of initial model models can be regenerated according to the training stage. parameters, and then use the model parameters of the regenerated initial model to train the model again until the convergence conditions of multi-task training are met.

In another embodiment, when the training loss function of multi-task training does not meet the convergence conditions, the convergence conditions during initial model training can be adjusted so that the initial model can be trained again to obtain the model parameters of the regenerated initial model, Then the model parameters of the regenerated initial model are used to train the model again until the convergence conditions of multi-task training are met.

Those skilled in the art can understand that all or part of the steps for implementing the above-described embodiments are implemented as computer programs executed by a CPU. When the computer program is executed by the CPU, the above-mentioned functions defined by the above-mentioned method provided by this application are executed. The program can be stored in a computer-readable storage medium, which can be a read-only memory, a magnetic disk or an optical disk.

In addition, it should be noted that the above-mentioned drawings are only schematic illustrations of processes included in the methods according to the exemplary embodiments of the present application, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal sequence of these processes. In addition, it is also easy to understand that these processes may be executed synchronously or asynchronously in multiple modules, for example.

The following are device embodiments of the present application, which can be used to execute method embodiments of the present application. For details not disclosed in the device embodiments of this application, please refer to the method embodiments of this application.

FIG. 8 is a block diagram of a user security level identification device based on multi-stage sequential multi-tasking according to an exemplary embodiment. As shown in Figure 8, the user security level identification device 80 based on multi-stage sequential multi-tasking includes: stage module 802, sorting module 804, training module 806, model module 808, and grading module 810.

The stage module 802 is used to generate multiple stage sets based on the total number of users and their corresponding user stages;

The sorting module 804 is used to arrange multiple stage sets in sequence;

The training module 806 is used to perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, and generate the n+1 group of model parameter vectors, where n is a positive integer;

The model module 808 is used to generate multiple sets of scoring models based on multiple sets of model parameter vectors until the multiple stages of collective training are completed;

The classification module 810 is configured to identify the security level of the current user through the multiple sets of scoring models, and determine the security level of the current user based on the identification results.

According to the user security level identification device based on multi-stage timing and multi-task of the present application, multiple stage sets are generated according to the total number of users and their corresponding user stages; the multiple stage sets are arranged in sequence according to the n+1th stage set; , the nth group of model parameter vectors perform multi-task training on the n+1th group of initial models, and generate the n+1th group of model parameter vectors, where n is a positive integer; until the multiple stages of collective training are completed, based on the multiple groups of model parameters The vector generates multiple groups of stage scoring models; the method of identifying the security level of the current user through the multi-group stage scoring model can comprehensively improve the multi-task machine learning method from the perspective of model samples and model parameters based on actual problems and application scenarios. , thereby ensuring the security of application system user data and transaction security.

An electronic device 900 according to this embodiment of the present application is described below with reference to FIG. 9 . The electronic device 900 shown in FIG. 9 is only an example and should not impose any limitations on the functions and usage scope of the embodiments of the present application.

As shown in Figure 9, electronic device 900 is embodied in the form of a general computing device. The components of the electronic device 900 may include, but are not limited to: at least one processing unit 910, at least one storage unit 920, a bus 930 connecting different system components (including the storage unit 920 and the processing unit 910), a display unit 940, and the like.

Wherein, the storage unit stores program code, and the program code can be executed by the processing unit 910, so that the processing unit 910 performs the steps in this specification according to various exemplary embodiments of the present application. For example, the processing unit 910 may perform the steps shown in FIG. 4 and FIG. 6 .

The storage unit 920 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 9201 and/or a cache storage unit 9202, and may further include a read-only storage unit (ROM) 9203.

The storage unit 920 may also include a program/utility 9204 having a set of (at least one) program modules 9205 including, but not limited to: an operating system, one or more applications, other program modules, and programs. Data, each of these examples or some combination may include an implementation of a network environment.

Bus 930 may be a local area representing one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or using any of a variety of bus structures. bus.

The electronic device 900 may also communicate with one or more external devices 900' (e.g., a keyboard, a pointing device, a Bluetooth device, etc.) so that the user can communicate with the device that the electronic device 900 interacts with, and/or the electronic device 900 can communicate with a Any device (such as a router, modem, etc.) with which multiple other computing devices communicate. This communication may occur through an input/output (I/O) interface 950. Furthermore, the electronic device 900 may also communicate with one or more networks (eg, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet) through the network adapter 960. Network adapter 960 may communicate with other modules of electronic device 900 via bus 930. It should be understood that, although not shown in the figures, other hardware and/or software modules may be used in conjunction with electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives And data backup storage system, etc.

Through the above description of the embodiments, those skilled in the art can easily understand that the example embodiments described here can be implemented by software, or can be implemented by software combined with necessary hardware. Therefore, as shown in Figure 10, the technical solution according to the embodiment of the present application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, U disk, mobile hard disk etc.) or on the network, including several instructions to cause a computing device (which can be a personal computer, a server, or a network device, etc.) to execute the above method according to the embodiment of the present application.

The software product may take the form of any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: electrical connection with one or more conductors, portable disk, hard disk, random access memory (RAM), read only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

The computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave carrying readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transport the program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

Program code for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., as well as conventional procedural formulas. Programming language—such as "C" or a similar programming language. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server execute on. In situations involving remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device, such as provided by an Internet service. (business comes via Internet connection).

The above-mentioned computer-readable medium carries one or more programs. When the above-mentioned one or more programs are executed by a device, the computer-readable medium realizes the following functions: generate multiple stages according to the total number of users and their corresponding user stages. Set; arrange multiple stage sets in sequence; perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors, and generate the n+1 group of model parameter vectors, n is a positive integer; until the multiple stage set training is completed, multiple sets of stage scoring models are generated based on multiple sets of model parameter vectors; the current user is identified as a security level through the multiple sets of stage scoring models.

Those skilled in the art can understand that the above-mentioned modules can be distributed in devices according to the description of the embodiments, or can be modified accordingly in one or more devices that are only different from this embodiment. The modules of the above embodiments can be combined into one module, or further divided into multiple sub-modules.

Through the description of the above embodiments, those skilled in the art can easily understand that the example embodiments described here can be implemented by software, or can be implemented by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, U disk, mobile hard disk, etc.) or on the network. , including several instructions to cause a computing device (which may be a personal computer, server, mobile terminal, or network device, etc.) to execute the method according to the embodiment of the present application.

Exemplary embodiments of the present application have been specifically shown and described above. It is to be understood that the present application is not limited to the detailed structures, arrangements, or implementation methods described herein; on the contrary, the present application is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

A user security level identification method based on multi-stage timing and multi-tasking, which is characterized by including:

Generate multiple stage sets based on the total number of users and their corresponding user stages;

Arrange multiple stage sets in sequence;

Perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors in sequence, and generate the n+1 group of model parameter vectors, where n is a positive integer;

Until the multiple stage set training is completed, multiple sets of stage scoring models are generated based on the multiple sets of model parameter vectors;

The security level of the current user is identified through the multi-group stage scoring model.
The method according to claim 1, characterized in that multiple stage sets are generated based on all users and their corresponding user stages, including:

Identify multiple user stages based on product characteristics;

Match the user stage corresponding to each user among all users with the multiple user stages;

According to the matching results, users are assigned to the stage set corresponding to their user stage.
The method of claim 2, wherein generating multiple stage sets based on all users and their corresponding user stages further includes:

Determine labeling strategies for each user stage;

Sample labels are assigned to users in each stage set according to the label policy.
The method according to claim 1, characterized in that multi-task training is performed on the n+1 group of initial models according to the n+1 stage set and the n group of model parameter vectors in order to generate the n+1 group of model parameter vectors. ,include:

extracting a first stage set among the plurality of stage sets;

Input the first stage set into the first set of initial models to generate the first set of model parameter vectors;

Perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors to generate the n+1 group of model parameter vectors, where n is a positive integer.
The method according to claim 1, characterized in that multi-task training is performed on the n+1 group of initial models according to the n+1 stage set and the n group of model parameter vectors in order to generate the n+1 group of model parameter vectors. ,Also includes:

Identify a set of machine learning models for each user stage;

Assign sample tags to historical users according to the tag strategy corresponding to each user stage;

The n+1th group of machine learning models are trained through historical users with sample labels to generate the n+1th group of initial models, where n is a positive integer.
The method of claim 4, characterized in that the first stage set is input into the first group of initial models to generate the first group of model parameter vectors, including:

Enter the user information in the first stage set into the first set of initial models respectively;

The first set of initial models perform model training based on user information and their corresponding labels. After the training is completed, the first set of model parameter vectors are generated.
The method according to claim 4, characterized in that multi-task training is performed on the n+1 group of initial models according to the n+1 stage set and the n group of model parameter vectors to generate the n+1 group of model parameter vectors, include:

Generate update vector;

Weight the update vector and add it to the model parameter vector of the n+1 initial model;

Enter the user information in the n+1th stage set into the n+1th group of initial models after the model parameter vector is updated for multi-task training;

After training is completed, the n+1th set of model parameter vectors are generated.
The method of claim 7, wherein generating an update vector includes:

Perform nonlinear transformation on the nth set of model parameter vectors to generate an update vector; or

The first group to the nth group of update vectors are generated through nonlinear transformation of the first group to the nth group of model parameter vectors.
The method according to claim 7, characterized in that the user information in the n+1th stage set is respectively input into the n+1th group of initial models after the model parameter vector is updated for multi-task training, including:

Enter the user information in the n+1th stage set into the n+1th group of initial models after the model parameter vector is updated;

The n+1 initial model performs multi-task training based on user information and its corresponding labels;

When the loss function during the training process does not meet the convergence conditions, re-determine the initial model parameters of the n+1 initial model to perform multi-task training again;

When the loss function meets the convergence condition, the multi-task training of the n+1 initial model is completed.
The method of claim 9, wherein re-determining the initial model parameters of the (n+1)th group of initial models to perform model training again includes:

Perform model training again on the n+1 initial model to generate new initial model parameters; or

Redetermine the convergence conditions to perform model training again on the n+1th group of initial models to generate new initial model parameters.
A user security level identification device based on multi-stage timing and multi-tasking, which is characterized by including:

The stage module is used to generate multiple stage sets based on all users and their corresponding user stages;

The sorting module is used to arrange multiple stage sets in sequence;

The training module is used to perform multi-task training on the n+1 group of initial models based on the n+1 stage set and the n group of model parameter vectors in sequence, and generate the n+1 group of model parameter vectors, where n is a positive integer;

A model module, configured to generate multiple sets of scoring models based on multiple sets of model parameter vectors until the multiple stages of collective training are completed;

A classification module, configured to identify the security level of the current user through the multiple sets of scoring models, and determine the security level of the current user based on the identification results.
An electronic device, characterized by including:

one or more processors;

A storage device for storing one or more programs;

When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method as described in any one of claims 1-10.
A computer-readable medium with a computer program stored thereon, characterized in that when the program is executed by a processor, the method according to any one of claims 1-10 is implemented.