A method and system for deploying an e-commerce risk identification model

Abstract

The application discloses an e-commerce risk identification model deployment method and system, and belongs to the technical field of e-commerce data processing. By acquiring e-commerce risk identification requirements input by a staff member, a self-defined risk identification task can be realized. According to the e-commerce risk identification requirements, an improved deep learning algorithm is used to perform parameter optimization and deployment on the target e-commerce risk identification model, so that the deployment effect and deployment efficiency of the e-commerce risk identification model can be effectively improved, and the target e-commerce risk identification model can accurately identify e-commerce risks defined by a user.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of e-commerce data processing, and particularly relates to a deployment method and system of an e-commerce risk identification model. BACKGROUND

[0002] E-commerce risk identification is a process of using big data, machine learning, and other technologies to analyze transaction behavior, product information, and user data on e-commerce platforms to identify potential risks. Its main goal is to prevent false transactions, fraudulent behavior, and illegal goods, and to protect the healthy development of the platform and the rights and interests of users. By building a risk identification model, real-time monitoring of transaction activities can be achieved, and abnormal behavior can be detected and warned in a timely manner, thereby improving the transaction security and trustworthiness of e-commerce platforms. E-commerce risk identification is an important means of maintaining the e-commerce ecosystem and has important significance for promoting the sustainable development of the e-commerce industry. With the rapid development of the Internet, e-commerce has become an important part of modern commerce. However, various risks such as false transactions, fraudulent behavior, and illegal goods in e-commerce transactions have seriously affected the healthy development of e-commerce platforms. In order to effectively identify and prevent these risks, e-commerce platforms often need to deploy risk identification models. However, existing risk identification model deployment methods often have problems such as complex deployment, low efficiency, and poor deployment results, making it difficult to adapt to user-defined risk identification tasks. SUMMARY

[0003] The application provides a deployment method and system of an e-commerce risk identification model to solve the problem of low efficiency and poor deployment results in existing risk identification model deployment methods.

[0004] In one aspect, the application provides a deployment method of an e-commerce risk identification model, comprising:

[0005] Obtaining e-commerce risk identification requirements input by a staff member; wherein the e-commerce risk identification requirements at least include e-commerce risk identification sample data and e-commerce risk identification sample data corresponding e-commerce risk labels;

[0006] Obtaining selection instructions corresponding to the e-commerce risk identification model input by the staff member, and selecting a target e-commerce risk identification model in the deep learning model library based on the selection instructions;

[0007] Based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, using an improved deep learning algorithm to perform parameter optimization on the target e-commerce risk identification model to obtain target parameters corresponding to the target e-commerce risk identification model;

[0008] Deploying the target e-commerce risk identification model based on the target parameters corresponding to the target e-commerce risk identification model to complete the deployment process.

[0009] Further, the deep learning model library comprises one or more of an LSTM model, a CNN model, an RNN model, a CNN-LSTM model, a CNN-BP model, a BP model, a GRU model, and a SVM model.

[0010] Further, based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, an improved deep learning algorithm is used to perform parameter optimization on the target e-commerce risk identification model to obtain target parameters corresponding to the target e-commerce risk identification model, comprising:

[0011] The parameters of the target e-commerce risk identification model are initialized by using a chaotic mapping initialization method to obtain a plurality of different parameter individuals;

[0012] Based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, the fitness corresponding to each parameter individual is obtained;

[0013] According to the fitness corresponding to all parameter individuals, the optimal parameter individual is determined;

[0014] Based on the optimal parameter individual, a first variable curve search strategy is used to perform multi-information fusion search on the parameter individual to obtain the parameter individual after multi-information fusion search;

[0015] For the parameter individual after multi-information fusion search, a second variable curve search strategy is used to perform neighborhood information search on the parameter individual to obtain the parameter individual after neighborhood information search;

[0016] For the parameter individual after neighborhood information search, a good point attraction search strategy is used to guide the search of the parameter individual to obtain the parameter individual after guided search;

[0017] For the parameter individual after guided search, a Cauchy global mutation search strategy is used to perform global search on the parameter individual to obtain the parameter individual after global search;

[0018] It is judged whether the optimization end condition is met, if yes, the optimal parameter individual is re-determined according to the parameter individual after global search, and the re-determined optimal parameter individual is taken as the target parameter corresponding to the target e-commerce risk identification model, otherwise the step of obtaining fitness is returned.

[0019] Further, the parameters of the target e-commerce risk identification model are initialized by using a chaotic mapping initialization method to obtain a plurality of different parameter individuals, comprising:

[0020] A random position selection individual is generated by using a random initialization method, wherein a total dimension of parameters of the random position selection individual is same as a total dimension of parameters of the target e-commerce risk identification model.

[0021] On the basis of the random position selection individual, other position selection individuals are obtained as follows:

[0022]

[0023] wherein y i represents the i-th position selection individual, and when i = 1, y i is set as the random position selection individual; y i+1 represents the i+1-th position selection individual, mod represents a remainder function, π represents a circular constant, and sin represents a sine function.

[0024] The position selection individual is mapped to a solution space, and a parameter individual is determined as follows:

[0025] x i,d = y i+1,d *(ub d -lb d )+lb d

[0026] wherein y i+1,d represents a d-th dimension parameter of the i+1-th position selection individual, x i,d represents a d-th dimension parameter of the i-th parameter individual, d = 1, 2, …, D, D represents a total dimension of parameters of the target e-commerce risk identification model, lb d represents a lower limit of the d-th dimension parameter, and ub d represents an upper limit of the d-th dimension parameter.

[0027] Further, on the basis of the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, a fitness corresponding to each parameter individual is obtained, including:

[0028] For any one parameter individual, after applying the hyperparameters contained in the parameter individual to the e-commerce risk identification model, the actual output of the e-commerce risk identification model is obtained by taking the e-commerce risk identification sample data as input;

[0029] The e-commerce risk labels corresponding to the e-commerce risk identification sample data are taken as the expected output of the e-commerce risk identification model, and the cross-entropy loss function value is obtained according to the actual output and the expected output of the e-commerce risk identification model;

[0030] The cross-entropy loss function value is added to a preset constant to obtain a non-zero fitness calculation parameter, and the non-zero fitness calculation parameter is taken as reciprocal to obtain the fitness corresponding to the parameter individual;

[0031] Traverse all parameter individuals, get the fitness corresponding to each parameter individual.

[0032] Further, based on the optimal parameter individual, a first variable curve search strategy is adopted to perform multi-information fusion search on the parameter individuals, to obtain the parameter individuals after multi-information fusion search, including:

[0033]

[0034] λ = e hcos(π(1-t / T))

[0035] wherein, represents the kth parameter individual in the tth optimization process, k = 1, 2, …, N, N represents the total number of parameter individuals, represents the parameter individual after multi-information fusion search represents the optimal parameter individual, e represents a natural constant, cos represents a cosine function, λ represents a variable curve control parameter, r1 represents a first random number uniformly distributed between [-1, 1], h represents a constant term and is set to 5, T represents a preset maximum optimization number, π represents a circular constant, r2 represents a second random number uniformly distributed between [-2, 2], and R represents a random positive integer less than 2N / 3, represents the jth parameter individual in the tth optimization process.

[0036] Further, for the parameter individuals after multi-information fusion search, a second variable curve search strategy is adopted to perform neighborhood information search on the parameter individuals, to obtain the parameter individuals after neighborhood information search, including:

[0037]

[0038] wherein, represents the mth parameter individual after multi-information fusion search in the tth optimization process, m = 1, 2, …, N, N represents the total number of parameter individuals, represents the parameter individual after neighborhood information search ω represents an inertia weight, sin represents a sine function, r3 represents a third random number between (0, 1), and r4 represents a fourth random number between (0, 1), represents a random parameter individual other than the parameter individual c1 represents a first weighting factor between (0, 1), c2 represents a second weighting factor between (0, 1), α1 represents a first inertia weight adjustment factor, α2 represents a second inertia weight adjustment factor, ω max represents a preset maximum value of the inertia weight, ω minA preset minimum value representing an inertia weight, cos represents a cosine function, T represents a preset maximum optimization number, and pi represents a circular constant.

[0039] Further, for the parameter individual after the neighborhood information search, a good point attraction search strategy is adopted to guide the search of the parameter individual, to obtain the parameter individual after the guided search, including:

[0040]

[0041] wherein, represents the parameter individual after the neighborhood information search in the tth optimization process, n=1, 2, …, N, N represents the total number of parameter individuals, represents the parameter individual after the guided search, represents the optimal parameter individual, r5 represents a fifth random number between (0, 1), and r6 represents a sixth random number uniformly distributed in [-1, 1], represents a random parameter individual other than the parameter individual .

[0042] Further, for the parameter individual after the guided search, a Cauchy global mutation search strategy is adopted to perform global search on the parameter individual, to obtain the parameter individual after the global search, including:

[0043] On the basis of the current optimization number, the adaptive mutation factor is obtained as:

[0044]

[0045] wherein, p is the adaptive mutation factor, sin represents a sine function, pi represents a circular constant, T represents a preset maximum optimization number, and t represents a current optimization number;

[0046] On the basis of the adaptive mutation factor, the global search is performed on the parameter individual, to obtain the parameter individual after the global search as:

[0047]

[0048] wherein, represents the parameter individual after the guided search in the tth optimization process, q=1, 2, …, N, N represents the total number of parameter individuals, represents the parameter individual after the global search cauchy(0, 1) represents a random number generated by a standard Cauchy distribution.

[0049] On the other hand, the present application provides a deployment system of an e-commerce risk identification model, comprising: an identification requirement acquisition module, a model selection module, a parameter optimization module, and a model deployment module.

[0050] The identification requirement acquisition module is configured to acquire an e-commerce risk identification requirement input by a staff member; wherein the e-commerce risk identification requirement comprises at least e-commerce risk identification sample data and e-commerce risk labels corresponding to the e-commerce risk identification sample data;

[0051] The model selection module is configured to acquire a selection instruction corresponding to an e-commerce risk identification model input by a staff member, and select a target e-commerce risk identification model from a deep learning model library based on the selection instruction;

[0052] The parameter optimization module is configured to perform parameter optimization on the target e-commerce risk identification model by using an improved deep learning algorithm based on the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, and acquire target parameters corresponding to the target e-commerce risk identification model;

[0053] The model deployment module is configured to deploy the target e-commerce risk identification model by using the target parameters corresponding to the target e-commerce risk identification model, and complete a deployment process.

[0054] The deployment method and system of the e-commerce risk identification model provided by the present application can realize a self-defined risk identification task by acquiring an e-commerce risk identification requirement input by a staff member, and can effectively improve the deployment effect and efficiency of the e-commerce risk identification model by performing parameter optimization and deployment on the target e-commerce risk identification model based on the e-commerce risk identification requirement by using an improved deep learning algorithm, so that the target e-commerce risk identification model can accurately identify a user-defined e-commerce risk. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0056] Figure 1 A flowchart of a deployment method of an e-commerce risk identification model provided by an embodiment of the present application.

[0057] Figure 2 A structural diagram of a deployment system of an e-commerce risk identification model provided by an embodiment of the present application.

[0058] The above-described drawings have shown specific embodiments of the present application, and more detailed descriptions will be given hereinafter. These drawings and written descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0059] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The following description is presented in connection with the drawings, wherein the same numbers on different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments are not meant to be all-inclusive of all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0060] Embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0061] As Figure 1 illustrated, the present application provides a deployment method of an e-commerce risk identification model, comprising:

[0062] S11, obtaining e-commerce risk identification requirements input by a staff; wherein the e-commerce risk identification requirements at least include e-commerce risk identification sample data and e-commerce risk labels corresponding to the e-commerce risk identification sample data;

[0063] For example, when the staff needs to identify prohibited or sensitive goods to avoid e-commerce illegal risks, the e-commerce risk identification sample data input by the staff can be e-commerce commodity images, and the e-commerce risk labels corresponding to the e-commerce risk identification sample data input by the staff can be image types, i.e. including normal categories and various abnormal categories. Thus, a self-defined identification task can be realized.

[0064] It is worth noting that the above e-commerce risk identification sample data and e-commerce risk labels corresponding to the e-commerce risk identification sample data are only examples of the embodiments of the present application. The staff can also input other types of data to deploy the e-commerce risk identification model to realize different e-commerce risk identification.

[0065] S12, obtaining selection instructions corresponding to the e-commerce risk identification model input by the staff, and selecting a target e-commerce risk identification model in the deep learning model library based on the selection instructions;

[0066] The deep learning model library can include a variety of deep learning models, which at least include models for data identification and models for image identification, so that the staff can select different models to realize self-defined e-commerce risk identification tasks, thereby improving the flexibility of e-commerce risk identification.

[0067] S13, based on the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, using an improved deep learning algorithm to perform parameter optimization on the target e-commerce risk identification model to obtain target parameters corresponding to the target e-commerce risk identification model;

[0068] In the prior art, optimization algorithms such as particle swarm optimization algorithms, genetic algorithms and the like are generally used to deploy deep learning models, and there are often technical problems of poor deployment effect and poor deployment efficiency, which result in that the e-commerce risk identification task cannot be accurately completed in the end, and therefore the embodiment of the present application uses an improved deep learning algorithm to perform parameter optimization on the target e-commerce risk identification model to solve the technical problems in the prior art and improve the e-commerce risk identification accuracy.

[0069] S14, deploying the target e-commerce risk identification model with the target parameters corresponding to the target e-commerce risk identification model to complete the deployment process.

[0070] The e-commerce risk identification model deployment method and system provided by the present application can realize a self-defined risk identification task by obtaining the e-commerce risk identification requirements input by the staff, and can effectively improve the deployment effect and deployment efficiency of the e-commerce risk identification model by using an improved deep learning algorithm to perform parameter optimization and deployment on the target e-commerce risk identification model according to the e-commerce risk identification requirements, so that the target e-commerce risk identification model can accurately identify the e-commerce risks defined by the user.

[0071] In the embodiment of the present application, the deep learning model library includes one or more of an LSTM (Long Short-Term Memory) model, a CNN (Convolutional Neural Networks) model, an RNN (Recurrent Neural Network) model, a CNN-LSTM model, a CNN-BP (Back Propagation) model, a BP model, a GRU (Gate Recurrent Unit) model and a SVM (Support Vector Machine) model.

[0072] It should be noted that the above deep learning model is only a preferred case of the embodiment of the present application, and the deep learning model library can also include other deep learning models. When it is detected that the e-commerce risk identification sample data input by the staff is image data, only the deep learning model capable of identifying images is allowed to be selected by the staff.

[0073] In the embodiment of the present application, the improved deep learning algorithm is used to perform parameter optimization on the target e-commerce risk identification model based on the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, to obtain the target parameters corresponding to the target e-commerce risk identification model, including:

[0074] The parameters of the target e-commerce risk identification model are initialized by using a chaotic mapping initialization method, and a plurality of different parameter individuals are obtained;

[0075] Based on the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, the fitness corresponding to each parameter individual is obtained;

[0076] The optimal parameter individual is determined according to the fitness corresponding to all parameter individuals;

[0077] Based on the optimal parameter individual, a first variable curve search strategy is used to perform multi-information fusion search on the parameter individual, and the parameter individual after the multi-information fusion search is obtained;

[0078] For the parameter individual after the multi-information fusion search, a second variable curve search strategy is used to perform neighborhood information search on the parameter individual, and the parameter individual after the neighborhood information search is obtained;

[0079] For the parameter individual after the neighborhood information search, a good point attraction search strategy is used to perform guided search on the parameter individual, and the parameter individual after the guided search is obtained;

[0080] For the parameter individual after the guided search, a Cauchy global mutation search strategy is used to perform global search on the parameter individual, and the parameter individual after the global search is obtained;

[0081] It is judged whether the optimization end condition is met, if yes, the optimal parameter individual is re-determined according to the parameter individual after the global search, and the re-determined optimal parameter individual is taken as the target parameter corresponding to the target e-commerce risk identification model, otherwise the step of obtaining the fitness is returned.

[0082] In the prior art, optimization algorithms such as particle swarm algorithm and genetic algorithm are generally used to deploy deep learning models, which often have the technical problems of poor deployment effect and poor deployment efficiency, resulting in the final inability to accurately complete the e-commerce risk identification task, therefore, the improved deep learning algorithm is used to optimize the parameters of the target e-commerce risk identification model in the embodiment of the present application to solve the technical problems existing in the prior art and improve the e-commerce risk identification accuracy.

[0083] In the embodiment of the present application, the parameters of the target e-commerce risk identification model are initialized by using a chaotic mapping initialization method, and a plurality of different parameter individuals are obtained, including:

[0084] A random position selection individual is generated by using a random initialization method; wherein the total dimension of the parameters of the random position selection individual is the same as the total dimension of the parameters of the target e-commerce risk identification model; for example, an individual can be randomly generated in the solution space, and the individual is directly taken as the random position selection individual.

[0085] Based on the random position selection individual, other position selection individuals are obtained:

[0086]

[0087] wherein y i represents the i-th position selection individual, and when i = 1, y i is set as a random position selection individual; y i+1 represents the i+1-th position selection individual, mod represents a remainder function, represents a circular constant, and sin represents a sine function.

[0088] The position selection individual is mapped to a solution space, and a parameter individual is determined as:

[0089] x i,d = y i+1,d *(ub d -lb d )+lb d

[0090] wherein y i+1,d represents the d-th dimension parameter of the i+1-th position selection individual, x i,d represents the d-th dimension parameter of the i-th parameter individual, d = 1, 2, …, D, D represents the total dimension of parameters of the target e-commerce risk identification model, lb d represents the lower limit of the d-th dimension parameter, and ub d represents the upper limit of the d-th dimension parameter.

[0091] Compared with the conventional chaotic mapping, the chaotic mapping initialization method provided in the embodiment of the application is more uniformly distributed, can effectively make the initial solution more uniformly distributed in the solution space, can effectively improve the deployment efficiency of the target e-commerce risk identification model, and can effectively avoid the algorithm from being stuck.

[0092] In the embodiment of the application, based on the e-commerce risk identification sample data and the e-commerce risk label corresponding to the e-commerce risk identification sample data, the fitness corresponding to each parameter individual is obtained, including:

[0093] For any one parameter individual, after the hyperparameters contained in the parameter individual are applied to the e-commerce risk identification model, the actual output of the e-commerce risk identification model is obtained by taking the e-commerce risk identification sample data as input;

[0094] The e-commerce risk label corresponding to the e-commerce risk identification sample data is taken as the expected output of the e-commerce risk identification model, and the cross-entropy loss function value is obtained according to the actual output and the expected output of the e-commerce risk identification model;

[0095] Add the cross-entropy loss function value to the preset constant to obtain a non-zero fitness calculation parameter, and take the reciprocal of the non-zero fitness calculation parameter to obtain the fitness corresponding to the parameter individual;

[0096] Traverse all parameter individuals to obtain the fitness corresponding to each parameter individual.

[0097] In the embodiment of the application, based on the optimal parameter individual, a first variable curve search strategy is used for multi-information fusion search on the parameter individual to obtain the parameter individual after the multi-information fusion search, including:

[0098]

[0099] Lambda = e hcos(π(1-t / T))

[0100] Wherein, The k-th parameter individual in the t-th optimization process is represented by k = 1, 2, …, N, N represents the total number of parameter individuals, The parameter individual after the multi-information fusion search is represented by The optimal parameter individual is represented by e, which represents a natural constant, cos represents a cosine function, lambda represents a variable curve control parameter, r1 represents a first random number uniformly distributed between [-1, 1], h represents a constant term, and is set to 5, T represents a preset maximum optimization number, pi represents a circular constant, r2 represents a second random number uniformly distributed between [-2, 2], and R represents a random positive integer less than 2N / 3, The j-th parameter individual in the t-th optimization process is represented by

[0101] The first variable curve search strategy provided by the embodiment of the application can make the parameter individual fuse the information of other parameter individuals, make the outlying parameter individual obtain higher search efficiency, effectively improve the search speed of the algorithm, and at the same time, combine the optimal parameter individual to perform variable curve search, which can provide the ability to jump out of the local optimum in the process of searching for the optimal position, and in the later stage of the algorithm, the algorithm can be converged to improve the deployment effect of the target e-commerce risk identification model.

[0102] In the embodiment of the application, for the parameter individual after the multi-information fusion search, a second variable curve search strategy is used for neighborhood information search on the parameter individual to obtain the parameter individual after the neighborhood information search, including:

[0103]

[0104] Wherein, The m-th parameter individual after the multi-information fusion search in the t-th optimization process is represented by m = 1, 2, …, N, N represents the total number of parameter individuals, Parameter individual after neighborhood information search ω represents an inertia weight, sin represents a sine function, r3 represents a third random number between (0, 1), r4 represents a fourth random number between (0, 1), Random parameter individual other than parameter individual c1 represents a first weighting factor between (0, 1), c2 represents a second weighting factor between (0, 1), α1 represents a first inertia weight adjustment factor, α2 represents a second inertia weight adjustment factor, ω max ω represents a preset maximum value of an inertia weight, min ω represents a preset minimum value of an inertia weight, cos represents a cosine function, T represents a preset maximum optimization number, and π represents a circular constant.

[0105] The second variable curve search strategy provided by the embodiment of the present application can perform collaborative search in combination with different parameter individuals, and simultaneously perform adaptive inertia weight adjustment in the collaborative search process, so as to gradually improve the convergence ability of the algorithm with the increase of the optimization number, and can effectively search the area between two parameter vectors and improve the ability of searching a global optimal solution.

[0106] In the embodiment of the present application, for the parameter individual after neighborhood information search, a good point attraction search strategy is adopted to guide search of the parameter individual, to obtain a parameter individual after guided search, including:

[0107]

[0108] wherein, The parameter individual after neighborhood information search in the t-th optimization process is represented as x n t, n = 1, 2, …, N, N represents a total number of parameter individuals, The parameter individual after guided search is represented as x n t+1, The optimal parameter individual is represented as x * t, r5 represents a fifth random number between (0, 1), and r6 represents a sixth random number uniformly distributed in [-1, 1], Random parameter individual other than parameter individual x n t+2.

[0109] The good point attraction search strategy provided by the embodiment of the present application can make parameter individuals interact with other parameter individuals, and simultaneously search around an optimal position, so as to improve the optimization precision of the algorithm, and in the later stage of the algorithm, all parameter individuals search together, which can greatly improve the probability of finding a global optimal solution.

[0110] In the embodiment of the present application, for the parameter individual after guided search, a Cauchy global mutation search strategy is adopted to perform global search on the parameter individual, to obtain a parameter individual after global search, including:

[0111] The adaptive mutation factor is obtained based on the current optimization number as follows:

[0112]

[0113] Wherein, p is the adaptive mutation factor, sin represents a sine function, represents a circular constant, T represents a preset maximum optimization number, and t represents a current optimization number;

[0114] The parameter individual is globally searched based on the adaptive mutation factor, and the parameter individual after global searching is obtained as follows:

[0115]

[0116] Wherein, The parameter individual after the qth guided search in the tth optimization process is represented as q = 1, 2, …, N, N represents a total number of parameter individuals, The parameter individual after global searching is represented as cauchy(0, 1) represents a random number generated by a standard cauchy distribution.

[0117] The cauchy global mutation search strategy provided by the embodiment of the application can effectively improve the global search capability of the algorithm, so as to avoid the algorithm from falling into a local optimum, and the global search capability gradually decreases in the later stage of the algorithm, which can ensure the convergence of the algorithm.

[0118] To sum up, the embodiment of the application improves the deployment efficiency and deployment effect of the e-commerce risk identification model, can effectively increase the e-commerce risk identification capability, assist the staff to find the e-commerce risk, and improve the work efficiency.

[0119] As shown in Figure 2 The application provides a deployment system of an e-commerce risk identification model, which comprises an identification requirement acquisition module 21, a model selection module 22, a parameter optimization module 23 and a model deployment module 24.

[0120] The identification requirement acquisition module 21 is used for acquiring the e-commerce risk identification requirements input by the staff, wherein the e-commerce risk identification requirements at least comprise e-commerce risk identification sample data and e-commerce risk identification sample data corresponding e-commerce risk labels.

[0121] The model selection module 22 is used for acquiring the selection instruction corresponding to the e-commerce risk identification model input by the staff, and selecting the target e-commerce risk identification model in the deep learning model library based on the selection instruction.

[0122] The parameter optimization module 23 is configured to perform parameter optimization on the target e-commerce risk identification model by using an improved deep learning algorithm based on e-commerce risk identification sample data and e-commerce risk identification sample data corresponding e-commerce risk labels, and obtain target parameters corresponding to the target e-commerce risk identification model.

[0123] The model deployment module 24 is configured to deploy the target e-commerce risk identification model by using the target parameters corresponding to the target e-commerce risk identification model, and complete the deployment process.

[0124] The deployment system of the e-commerce risk identification model provided in the embodiments of the present application can execute the method technical solution described above, and has similar principles and beneficial effects, which will not be described here again.

[0125] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the application cover any and all variations of the application that come within the scope of the general concept of the application and that the application is not limited by the specific embodiments described herein. It is to be understood that the application is not limited to particular details described herein and / or illustrated in the drawings. Rather, the application is intended to cover any and all variations that come within the scope of the application including what is claimed as the application. The scope of the application is limited only by the claims.

Claims

1. A method for deploying an e-commerce risk identification model, characterized in that, The method comprises the following steps: obtaining the e-commerce risk identification requirements input by the staff; wherein the e-commerce risk identification requirements at least include e-commerce risk identification sample data and e-commerce risk identification sample data corresponding e-commerce risk labels; obtaining the selection instruction corresponding to the e-commerce risk identification model input by the staff, and selecting the target e-commerce risk identification model in the deep learning model library based on the selection instruction; based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, the improved deep learning algorithm is used to optimize the parameters of the target e-commerce risk identification model, and the target parameters corresponding to the target e-commerce risk identification model are obtained; deploying the target e-commerce risk identification model with the target parameters corresponding to the target e-commerce risk identification model, and completing the deployment process; based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, the improved deep learning algorithm is used to optimize the parameters of the target e-commerce risk identification model, and the target parameters corresponding to the target e-commerce risk identification model are obtained, which comprises: using the chaotic mapping initialization method to initialize the parameters of the target e-commerce risk identification model, and obtaining a plurality of different parameter individuals; based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, the fitness of each parameter individual is obtained; determining the optimal parameter individual according to the fitness of all parameter individuals; based on the optimal parameter individual, the first variable curve search strategy is used to perform multi-information fusion search on the parameter individual, and the parameter individual after multi-information fusion search is obtained; for the parameter individual after multi-information fusion search, the second variable curve search strategy is used to perform neighborhood information search on the parameter individual, and the parameter individual after neighborhood information search is obtained; for the parameter individual after neighborhood information search, the good point attraction search strategy is used to guide the search of the parameter individual, and the parameter individual after guided search is obtained; for the parameter individual after guided search, the Cauchy global mutation search strategy is used to perform global search on the parameter individual, and the parameter individual after global search is obtained; determine whether the optimization end condition is met, if yes, the optimal parameter individual is determined again according to the parameter individual after global search, and the optimal parameter individual determined again is taken as the target parameter corresponding to the target e-commerce risk identification model, otherwise, return to the step of obtaining the fitness; using the chaotic mapping initialization method to initialize the parameters of the target e-commerce risk identification model, and obtaining a plurality of different parameter individuals, which comprises: generating a random position selection individual by using the random initialization method; wherein the parameter total dimension of the random position selection individual is the same as the parameter total dimension of the target e-commerce risk identification model; based on the random position selection individual, the other position selection individuals are obtained as: where y i represents the i-th position selection individual, and when i = 1, y i is set as a random position selection individual; y i+1 represents the i+1-th position selection individual, mod represents a remainder function, π represents a circular constant, and sin represents a sine function; mapping the position selection individual to the solution space to determine the parameter individual as: x i,d = y i+1,d *(ub d - lb d )+ lb d wherein y i+1,d represents the d-th dimensional parameter of the i+1-th position selection individual, x i,d represents the d-th dimensional parameter of the i-th parameter individual, d = 1, 2, …, D, D represents the total dimension of parameters of the target e-commerce risk identification model, lb d represents the lower limit of the d-th dimensional parameter, ub d represents the upper limit of the d-th dimensional parameter; based on the e-commerce risk identification sample data and the e-commerce risk identification sample data corresponding e-commerce risk labels, the fitness of each parameter individual is obtained, which comprises: For any one parameter individual, the hyperparameters contained in the parameter individual are applied to the e-commerce risk identification model, and then the actual output of the e-commerce risk identification model is obtained by taking the e-commerce risk identification sample data as input; The e-commerce risk label corresponding to the e-commerce risk identification sample data is taken as the expected output of the e-commerce risk identification model, and the cross-entropy loss function value is obtained according to the actual output and the expected output of the e-commerce risk identification model; The cross-entropy loss function value is added to the preset constant to obtain a non-zero fitness calculation parameter, and the non-zero fitness calculation parameter is taken to obtain the fitness corresponding to the parameter individual by taking the reciprocal of the non-zero fitness calculation parameter; The fitness corresponding to each parameter individual is obtained by traversing all parameter individuals; Based on the optimal parameter individual, the first variable curve search strategy is used to perform multi-information fusion search on the parameter individual to obtain the parameter individual after multi-information fusion search, including: λ = e hcos(π(1-t / T)) wherein, denotes the kth parameter individual in the tth optimization process, k = 1, 2, …, N, N denotes the total number of parameter individuals, denotes the parameter individual after multi-information fusion search denotes the optimal parameter individual, e denotes a natural constant, cos denotes a cosine function, λ denotes a variable curve control parameter, r1 denotes a first random number uniformly distributed between [-1, 1], h denotes a constant term and is set to 5, T denotes a preset maximum optimization number, π denotes a circular constant, r2 denotes a second random number uniformly distributed between [-2, 2], and R denotes a random positive integer less than 2N / 3, denotes the jth parameter individual in the tth optimization process; Based on the parameter individual after multi-information fusion search, the second variable curve search strategy is used to perform neighborhood information search on the parameter individual to obtain the parameter individual after neighborhood information search, including: wherein, denotes the parameter individual after the mth multi-item information fusion search in the tth optimization process, m = 1, 2, …, N, N denotes the total number of parameter individuals, denotes the parameter individual after the neighborhood information search ω denotes the inertia weight, sin denotes the sine function, r3 denotes the third random number between (0, 1), r4 denotes the fourth random number between (0, 1), denotes the random parameter individual other than the parameter individual c1 denotes the first weighting factor between (0, 1), c2 denotes the second weighting factor between (0, 1), α1 denotes the first inertia weight adjustment factor, α2 denotes the second inertia weight adjustment factor, ω max denotes the preset maximum value of the inertia weight, ω min denotes the preset minimum value of the inertia weight, cos denotes the cosine function, T denotes the preset maximum optimization number, π denotes the circular constant; Based on the parameter individual after neighborhood information search, the good point attraction search strategy is used to perform guided search on the parameter individual to obtain the parameter individual after guided search, including: wherein, denotes the parameter individual after the n-th neighborhood information search in the t-th optimization process, n = 1, 2, …, N, N denotes the total number of parameter individuals, denotes the parameter individual after the guided search, denotes the optimal parameter individual, r5 denotes the fifth random number between (0, 1), and r6 denotes the sixth random number uniformly distributed between [-1, 1], denotes the random parameter individual other than the parameter individual . Based on the parameter individual after guided search, the Cauchy global mutation search strategy is used to perform global search on the parameter individual to obtain the parameter individual after global search, including: Based on the current optimization number, the adaptive mutation factor is obtained as: Wherein, p is less than 1, sin represents a sine function, π represents a circular constant, T represents a preset maximum optimization number, and t represents a current optimization number; Based on the adaptive mutation factor, the parameter individual is globally searched to obtain the parameter individual after global search as: wherein, denotes the parameter individual after the qth guided search in the tth optimization process, q = 1, 2, …, N, N denotes the total number of parameter individuals, denotes the parameter individual after the global search cauchy(0, 1) denotes a random number generated by a standard cauchy distribution.

2. The method of claim 1, wherein, The deep learning model library includes one or more of an LSTM model, a CNN model, an RNN model, a CNN-LSTM model, a CNN-BP model, a BP model, a GRU model, and a SVM model. 3.A system for deploying an e-commerce risk identification model, capable of performing the method for deploying an e-commerce risk identification model according to claim 1 or 2, characterized in that, Including: An identification requirement acquisition module, a model selection module, a parameter optimization module, and a model deployment module; The identification requirement acquisition module is configured to acquire e-commerce risk identification requirements input by a worker, wherein the e-commerce risk identification requirements at least include e-commerce risk identification sample data and e-commerce risk labels corresponding to the e-commerce risk identification sample data; The model selection module is configured to acquire a selection instruction corresponding to a target e-commerce risk identification model input by a worker, and select the target e-commerce risk identification model from a deep learning model library based on the selection instruction; The parameter optimization module is configured to perform parameter optimization on the target e-commerce risk identification model by using an improved deep learning algorithm based on the e-commerce risk identification sample data and the e-commerce risk labels corresponding to the e-commerce risk identification sample data, and obtain target parameters corresponding to the target e-commerce risk identification model; The model deployment module is configured to deploy the target e-commerce risk identification model by using the target parameters corresponding to the target e-commerce risk identification model, and complete a deployment process.

Images