Risk information determination method, model training method, equipment, program product

Abstract

The risk information determination method, model training method, device, and program product provided by the present disclosure include: obtaining the business data to be evaluated of the business activity; the business data to be evaluated includes multiple data features, each data feature corresponds to a feature domain; inputting the data features corresponding to the same feature domain into the coding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain; inputting the feature vector corresponding to the feature domain into the attention layer to obtain the total vector including the relationship between each feature vector; inputting the total vector into the determination submodule, and using the determination submodule to determine the expectation corresponding to the business data to be evaluated; determining the risk information corresponding to the business data to be evaluated according to the expectation. The scheme provided by the present disclosure processes the data features belonging to the feature domain through the coding network corresponding to the feature domain to obtain the feature vector of the feature domain, and then accurately extracts the feature vector of the business activity, and can accurately obtain the risk assessment result.

Description

Technical Field

[0001] The present disclosure relates to artificial intelligence technology, and in particular to a risk information determination method, model training method, device, and program product. Background Art

[0002] At present, online shopping platforms have become more and more mature, and each online shopping platform will launch a variety of preferential activities to promote the transaction volume of goods. With the increase in preferential activities, the black and gray industries based on online shopping platforms have emerged. Some users reserve a large number of accounts on online shopping platforms and use these accounts to participate in preferential activities and conduct arbitrage.

[0003] In order to deal with the black and gray industries on online shopping platforms, there are risk assessment technologies in the existing technology. By conducting risk assessment on all links of the entire chain of users participating in preferential activities, it is an effective means to prevent the black and gray industries. Risk assessment technology is divided into rule-based risk assessment methods and algorithm-based risk assessment methods.

[0004] Among them, the rule-based risk assessment method is highly dependent on expert experience, and the rule release cycle is long, and the crackdown on the black and gray industries is relatively lagging. In the risk assessment method based on the algorithm model, because the data generated by participating in preferential activities not only has high feature dimensions but also has strong sparsity, traditional modeling or neural network structures cannot effectively learn the characteristics of this data. Summary of the invention

[0005] The present disclosure provides a risk information determination method, a model training method, a device, and a program product to solve the problem that the prior art cannot accurately assess the risks of business activities.

[0006] A first aspect of the present disclosure is to provide a method for determining risk information of a business activity, comprising:

[0007] Acquire the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes a plurality of data features, each of which corresponds to a feature domain;

[0008] Inputting the data features corresponding to the same feature domain into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0009] Inputting the feature vectors corresponding to the feature domain into the attention layer to obtain a total vector including the relationship between each of the feature vectors;

[0010] Inputting the total vector into a determination submodule, and using the determination submodule to determine an expectation corresponding to the business data to be evaluated;

[0011] Risk information corresponding to the business data to be evaluated is determined according to the expectation, where the risk information is used to indicate a risk level of the business data to be evaluated.

[0012] A second aspect of the present disclosure is to provide a method for training a model for assessing business activity risk, the model comprising: an encoding network corresponding to each feature domain, an attention layer, and a determination submodule;

[0013] The method comprises:

[0014] Acquire a plurality of sample data corresponding to the business activity; each of the sample data includes a plurality of data features, and each of the data features corresponds to a feature domain;

[0015] Input the data features corresponding to the same feature domain in each sample data into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0016] Inputting the feature vector corresponding to the feature domain into the attention layer to obtain a total vector corresponding to each sample data including the relationship between each feature vector;

[0017] The total vector of each sample data is input into the determination submodule, the expectation of each sample data is determined by using the determination submodule, and the model is trained according to the expectation of each sample data to obtain a model for evaluating the risk of business activities.

[0018] A third aspect of the present disclosure is to provide a risk information determination device for a business activity, comprising:

[0019] An acquisition module, used for acquiring the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes a plurality of data features, each of which corresponds to a feature domain;

[0020] An encoding module, used for inputting the data features corresponding to the same feature domain into an encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0021] An embedding module, used for inputting the feature vectors corresponding to the feature domain into the attention layer to obtain a total vector including the relationship between each of the feature vectors;

[0022] an expectation determination module, configured to input the total vector into a determination submodule, and use the determination submodule to determine an expectation corresponding to the service data to be evaluated;

[0023] A risk determination module is used to determine risk information corresponding to the business data to be evaluated according to the expectation, wherein the risk information is used to indicate the risk level of the business data to be evaluated.

[0024] A fourth aspect of the present disclosure is to provide a training device for a model for assessing business activity risk, the model comprising: an encoding network corresponding to each feature domain, an attention layer, and a determination submodule;

[0025] The device comprises:

[0026] An acquisition module, used for acquiring a plurality of sample data corresponding to the business activity; each of the sample data includes a plurality of data features, and each of the data features corresponds to a feature domain;

[0027] An encoding module, used for inputting the data features corresponding to the same feature domain in each sample data into an encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0028] An embedding module, used for inputting the feature vector corresponding to the feature domain into the attention layer to obtain a total vector corresponding to each sample data including the relationship between each feature vector;

[0029] an expectation determination module, used for inputting the total vector of each sample data into the determination submodule, and determining the expectation of each sample data by using the determination submodule;

[0030] The training module is used to train the model according to the expectations of each sample data to obtain a model for assessing the risk of business activities.

[0031] A fifth aspect of the present disclosure is to provide an electronic device, comprising:

[0032] Memory;

[0033] Processor; and

[0034] Computer programs;

[0035] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method for determining risk information of business activities as described in the first aspect above or the method for training a model for assessing business activity risks as described in the second aspect.

[0036] The sixth aspect of the present disclosure is to provide a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the method for determining risk information of business activities as described in the first aspect above or the method for training a model for assessing the risks of business activities as described in the second aspect.

[0037] The seventh aspect of the present disclosure is to provide a computer program product, including a computer program, which, when executed by a processor, implements the method for determining risk information of business activities according to the first aspect or the method for training a model for assessing the risks of business activities according to the second aspect.

[0038] The technical effects of the risk information determination method, model training method, device, and program product provided by the present disclosure are:

[0039] The risk information determination method, model training method, device, and program product provided by the present disclosure include: obtaining business data to be evaluated corresponding to business activities; the business data to be evaluated includes multiple data features, each data feature corresponds to a feature domain; the data features corresponding to the same feature domain are input into the coding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain; the feature vector corresponding to the feature domain is input into the attention layer to obtain the total vector including the relationship between each feature vector; the total vector is input into the determination submodule, and the expectation corresponding to the business data to be evaluated is determined by the determination submodule; the risk information corresponding to the business data to be evaluated is determined according to the expectation, and the risk information is used to indicate the risk degree of the business data to be evaluated. In the scheme provided by the present disclosure, the data features belonging to the feature domain are processed by the coding network corresponding to the feature domain, and the information in the domain can be fully combined and strengthened to obtain the feature vector of the feature domain, and then the feature vector of the business activity can be accurately extracted when the data features of the business activity are sparse, and the risk assessment result of the business activity can be accurately obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flowchart of a method for determining risk information of a business activity shown in an exemplary embodiment of the present application;

[0041] Figure 2 A schematic diagram of a risk assessment model shown as an exemplary embodiment of the present application;

[0042] Figure 3 This is a schematic diagram of the structure of each coding network shown in an exemplary embodiment of the present application;

[0043] Figure 4 A schematic diagram of the processing process of the attention layer shown as an exemplary embodiment of the present application;

[0044] Figure 5 A flowchart of a method for determining risk information of a business activity shown in another exemplary embodiment of the present application;

[0045] Figure 6 A flowchart of a method for training a model for assessing business activity risks according to an exemplary embodiment of the present application;

[0046] Figure 7 A schematic diagram of a risk assessment model shown as an exemplary embodiment of the present application;

[0047] Figure 8 A flowchart of a method for training a model for assessing business activity risks according to another exemplary embodiment of the present application;

[0048] Fig. 9 A structural diagram of a device for determining risk information of a business activity shown in an exemplary embodiment of the present application;

[0049] Fig.10 A structural diagram of a device for determining risk information of a business activity shown in another exemplary embodiment of the present application;

[0050] Fig.11 A structural diagram of a training device for a model for assessing business activity risks according to an exemplary embodiment of the present application;

[0051] Fig.12 A structural diagram of a training device for a model for assessing business activity risks according to another exemplary embodiment of the present application;

[0052] Fig.13 The figure is a structural diagram of an electronic device according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0053] At present, in order to deal with the black and gray industries on online shopping platforms, risk assessment technology is needed to effectively evaluate business activities to determine whether the business activities are easily exploited by the black and gray industries. Risk assessment technology conducts real-time risk assessment of marketing resources in all links of the entire chain of business activities, so that a series of risk control measures such as resource offline or account interception can be taken in time for medium and high-risk arbitrage behaviors. Risk assessment technology is divided into rule-based risk assessment methods and algorithm model-based risk assessment methods.

[0054] Among them, the rule-based risk scoring method can be further divided according to the rule generation method, which can be divided into rules based on expert experience and rules generated by technologies such as feature importance mining; the rule use method can be divided into single rules and combined rules. Single rules can complete the judgment using one judgment rule, while combined rules require other conditions or rule combination judgments.

[0055] However, the rule-based risk scoring method relies heavily on expert experience, the rule formulation and release cycle is long, and there is a lag in the real-time confrontation with the black and gray industries. In addition, as the business changes, the rules need to be continuously iterated, otherwise the confrontation effect will be weakened. Faced with the massive dimensional characteristics of complex business scenarios, even experts find it difficult to formulate online rules that cover all risk issues. Risk assessment through rules does not have the overall portability of the solution and needs to be customized according to different scenarios.

[0056] Among them, the risk assessment method based on algorithm models uses artificial intelligence technologies such as machine learning and deep learning to process features and train models of marketing scenario data, and make risk predictions based on the model results.

[0057] Anomaly detection technology in the field of machine learning plays a vital role in the field of e-commerce risk control. Unlike traditional supervised modeling, unsupervised anomaly detection technology does not require specific label training, but can autonomously mine patterns and connections in the data and discover abnormal data faster and more extensively.

[0058] Generally speaking, anomaly detection methods can be divided into the following four categories: (1) Statistical test methods, which generally assume that normal data follows a normal distribution, while abnormal data has a larger deviation from normal data; (2) Time series detection methods, which discover abnormal data by detecting abnormal points in the sequence that are inconsistent with the pattern, such as sudden rises or falls, trend changes, level changes, etc.; (3) Supervised learning methods, which convert anomaly detection into a binary classification problem, with normal data in one category and abnormal data in another, and use mature binary classification machine learning or deep learning algorithms to perform anomaly detection; (4) Unsupervised learning methods, which generally use clustering methods for detection. If a certain data is relatively far from the class center, then the data is abnormal data.

[0059] However, the risk mining method based on clustering or spatial partitioning is still unable to handle massive high-dimensional features. The method based on first reducing the dimension and then mining outliers retains the information contained in the high-dimensional features as much as possible, and then finds outliers by clustering or spatial partitioning the low-dimensional vectors. The two training processes are decoupled and have two optimization goals in different directions. Some key information may be lost in the process of dimensionality reduction, which will cause the final effect to easily reach a suboptimal solution. The characteristics of business activities in online shopping platforms are not only high-dimensional and sparse, but also clearly divided by domain. Traditional modeling or neural network fitting cannot learn the relationship between sparse feature domains well. In addition, the abnormal score finally output by the model is usually expressed as the distance between the sample and the center of each cluster or the degree of discreteness of the subspace. The large floating range of the score cannot provide a good quantitative guidance for the business.

[0060] In order to solve the above technical problems, in the solution provided in the present application, the data features of business activities are divided according to feature domains, and the feature vectors of the data features of each feature domain are extracted through each coding network corresponding to the feature domain, and then the business activities are evaluated for risks based on the feature vectors corresponding to each feature domain. This method can accurately extract the feature vectors of business activities when the data features of business activities are highly sparse, and thus accurately obtain the risk assessment results of the business activities.

[0061] Figure 1 A flowchart of a method for determining risk information of a business activity shown as an exemplary embodiment of the present application.

[0062] like Figure 1 As shown, the method for determining risk information of business activities provided by this application includes:

[0063] Step 101 : obtaining to-be-evaluated business data corresponding to the business activity; the to-be-evaluated business data includes a plurality of data features, each of which corresponds to a feature domain.

[0064] Among them, the method provided in the present application can be executed by an electronic device with computing capabilities, and the electronic device can be, for example, a backend server of an e-commerce platform.

[0065] Specifically, a risk assessment model can be pre-trained and deployed in an electronic device. The electronic device can evaluate and process the business data to be evaluated of the business activity based on the deployed risk assessment model.

[0066] Figure 2 A schematic diagram of a risk assessment model shown as an exemplary embodiment of the present application.

[0067] like Figure 2 As shown, the risk assessment model provided by the present application may include multiple coding networks 21, attention layers 22, and determination submodules 23. The coding network 21, attention layer 22, and determination submodule 23 are connected in sequence, and the business data to be evaluated of the business activity can be input into the coding network 21. Based on the coding network 21, attention layer 22, and determination submodule 23, the expectation of the business data to be evaluated can be obtained, and the electronic device can then determine the risk information indicating the risk level of the business data to be evaluated based on the expectation.

[0068] Furthermore, the electronic device may obtain business data to be evaluated corresponding to the business activity, where the business data to be evaluated includes a plurality of data features, and each data feature corresponds to a feature domain.

[0069] In actual application, multiple feature domains can be pre-set. For example, a first feature domain related to registered account information can be set, and a second feature domain related to illegal reselling preferential information can also be set. The features related to the registered account in the data features can be divided into the first feature domain, so that these features have a corresponding relationship with the first feature domain, and the features of illegal reselling preferential information can also be divided into the second feature domain. In this way, the data features used to represent the same type of information can be divided into the same feature domain, and then the correlation between the data features used to represent the same type of information can be focused on.

[0070] For example, the business data to be evaluated includes N data features, and these N features can be divided into K feature domains according to the information represented by each data feature.

[0071] Step 102, inputting data features corresponding to the same feature domain into a coding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain.

[0072] Among them, in the solution provided by the present application, multiple encoding networks corresponding to each feature domain are set up, and the data features of the feature domain are processed through the encoding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain.

[0073] Figure 3 This is a schematic diagram of the structure of each coding network shown in an exemplary embodiment of the present application.

[0074] like Figure 3 As shown, for example, a first coding network 31 corresponding to the first feature domain is provided, a second coding network 32 corresponding to the second feature domain is provided, and a Kth coding network 3k corresponding to the Kth feature domain is provided.

[0075] The electronic device can input the data features corresponding to the same feature domain among multiple data features of the business data to be evaluated into the coding network corresponding to the feature domain, and then use the coding network to process the data features of the feature domain.

[0076] For example, a plurality of data features 311 divided into a first feature domain are input into a first encoding network 31 , and a plurality of data features 321 divided into a second feature domain are input into a second encoding network 32 .

[0077] The electronic device processes the input data feature of the coding network based on the coding network to obtain a feature vector. For example, the electronic device processes the input data feature 311 based on the first coding network 31 to obtain a feature vector 312. The electronic device processes the input data feature 321 based on the second coding network 32 to obtain a feature vector 322.

[0078] Specifically, the sparse features of each feature domain are encoded into a dense feature vector E of a specified dimension through the encoding network. K for:

[0079] E K =σ(X K θ f )

[0080] Where E k Represents the feature vector of the business data to be evaluated in K different feature domains.

[0081] Since the data features belonging to the same feature domain are used to represent the same type of information, for example, the data features belonging to the first feature domain are used to represent the features of the registered account, while the data features belonging to the second feature domain are used to represent the features of illegal reselling preferential information. Therefore, by processing the data features belonging to the feature domain through the encoding network corresponding to the feature domain, the information in the domain can be fully combined and strengthened to obtain the feature vector of the feature domain.

[0082] Therefore, through multiple coding networks, feature vectors corresponding to each feature domain can be extracted. Then, multiple feature vectors corresponding to multiple feature domains can be extracted from the business data to be evaluated, and then multiple feature vectors used to characterize the information categories corresponding to each feature domain can be obtained. For example, a feature vector used to characterize a registered account in the business data to be evaluated can be obtained, and another example is that a feature vector used to characterize reselling discount information in the business data to be evaluated can be obtained.

[0083] Specifically, the weight parameters in each coding network may be different, and may be adjusted and determined during the model training process. The structures of each coding network may be the same or different, and may be set according to the requirements.

[0084] Step 103, input the feature vector corresponding to the feature domain into the attention layer to obtain a total vector including the relationship between each feature vector.

[0085] Furthermore, in the solution provided by the present application, the risk assessment model also includes an attention layer, and the feature vectors corresponding to each feature domain can be input into the attention layer. The attention layer can process the feature vectors of each feature domain to obtain a total vector including the relationship between each feature vector.

[0086] Figure 4 A schematic diagram of the processing of the attention layer shown as an exemplary embodiment of the present application.

[0087] like Figure 4As shown, each feature vector 41 corresponding to each feature domain can be input into the attention layer 42, and the electronic device can process the feature vector input into the attention layer 42 according to the attention layer 42 to determine the relationship between these feature vectors, thereby obtaining a total vector 43 including the relationship between each of the feature vectors.

[0088] In actual application, the electronic device can determine the relationship between two feature vectors according to the weights in the attention layer, and can determine the new expression of the feature vector based on the relationship between the feature vector and other feature vectors, so that the information of other feature vectors can be embedded in the new expression. For example, the new expression of the first feature vector can be determined according to the relationship between the first feature vector and each other feature vector.

[0089] The electronic device can construct a total vector including the relationship between each feature vector according to the new expression of each feature vector. The total vector includes the relationship between each feature vector, so the feature vector can integrate the feature vectors of multiple information categories in the data to be evaluated, and then accurately extract the information included in the data to be evaluated.

[0090] Specifically, the weight value in the attention layer can be obtained by training the model.

[0091] Step 104: input the total vector into a determination submodule, and use the determination submodule to determine the expectation corresponding to the business data to be evaluated.

[0092] Furthermore, the risk assessment model also includes a determination submodule, and the electronic device can input the total vector determined in step 103 into the determination submodule, and process the total vector according to the weight value in the determination submodule to obtain the expectation of the business data to be evaluated.

[0093] In one implementation, the electronic device may first compress the total vector to reduce the dimension of the total vector, and then process the vector after the dimension reduction to obtain the expectation of the total vector, that is, the expectation corresponding to the business data to be evaluated.

[0094] In actual application, Gaussian distribution parameters can be set in the determination submodule, and the Gaussian distribution parameters can be obtained by training the model. For example, if there are k feature domains in total, k Gaussian distribution parameters can be set.

[0095] The total vector or the compressed total vector can be processed according to the parameters of each Gaussian distribution to determine the expectation corresponding to the business data to be evaluated, and the expectation is used to characterize the distance between the business data to be evaluated and each Gaussian distribution. Further, whether the business data to be evaluated is abnormal can be determined according to the distance between the business data to be evaluated and each Gaussian distribution.

[0096] Step 105: Determine risk information corresponding to the business data to be evaluated as desired, where the risk information is used to indicate the risk level of the business data to be evaluated.

[0097] Specifically, the expectation determined by the risk assessment model can be converted into risk information, and then the risk level of the business data to be assessed can be indicated by the risk information. For example, a mapping function can be pre-set, and the expectation of the business data to be assessed can be mapped to a score in the range of 0-1 through the mapping function. For example, the expectation can be mapped to the range of 0-1 through a shift logarithmic transformation, a sigmoid mapping, etc.

[0098] In this implementation mode, if the floating range between the expectations of different data to be evaluated is large, these values ​​can be mapped to the same score scale, thereby better representing the risk level of the data to be evaluated.

[0099] In one implementation, a risk threshold may also be set, and if the determined risk information exceeds the risk threshold, a warning may be issued.

[0100] The method provided in this embodiment is used for risk information of business activities. The method is executed by a device equipped with the method provided in this embodiment, and the device is usually implemented in hardware and / or software.

[0101] The risk information determination method of the business activity provided by the present application includes: obtaining the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes multiple data features, each data feature corresponds to a feature domain; the data features corresponding to the same feature domain are input into the coding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain; the feature vector corresponding to the feature domain is input into the attention layer to obtain the total vector including the relationship between each feature vector; the total vector is input into the determination submodule, and the expectation corresponding to the business data to be evaluated is determined by the determination submodule; the risk information corresponding to the business data to be evaluated is determined according to the expectation, and the risk information is used to indicate the risk degree of the business data to be evaluated. In the method provided by the present application, the data features belonging to the feature domain are processed by the coding network corresponding to the feature domain, and the information in the domain can be fully combined and strengthened to obtain the feature vector of the feature domain, and then the feature vector of the business activity can be accurately extracted when the data features of the business activity are sparse, and the risk assessment result of the business activity can be accurately obtained.

[0102] Figure 5 A flowchart of a method for determining risk information of a business activity shown in another exemplary embodiment of the present application.

[0103] like Figure 5 As shown, the method for determining risk information of business activities provided by this application includes:

[0104] Step 501, obtaining offline features corresponding to business activities; obtaining real-time data corresponding to business activities.

[0105] Step 502, determining the data to be evaluated of the business activity according to the offline features and the real-time data; the business data to be evaluated includes a plurality of data features, each of which corresponds to a feature domain.

[0106] The features in the data to be evaluated may include both offline features and real-time features. For example, offline features corresponding to business activities may be pre-set. When real-time data corresponding to the business activity exists, the offline features of the business activity may be obtained and processed together with the real-time data to obtain the business data to be evaluated corresponding to the business activity.

[0107] Specifically, offline features can accurately reflect the characteristics of business activities, while online real-time data can be generated in real time. By processing the real-time features and offline features of business activities, it is possible to extract accurate feature information related to the business activities and ensure the real-time nature of the business data to be evaluated, thereby improving the accuracy of the evaluation of business activities.

[0108] Furthermore, a feature processing engine may be provided. For example, after generating real-time data related to the first business activity, offline features of the first business activity may be obtained, and the feature processing engine may be used to process the offline features and the real-time data to obtain business data X to be evaluated including multiple data features.

[0109] Step 503: input the data features corresponding to the same feature domain into the encoding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain.

[0110] The execution principle and method of step 503 are similar to those of step 102 and will not be described in detail.

[0111] Step 504: Determine the feature similarity between every two feature vectors based on the feature vectors corresponding to each feature domain.

[0112] Step 505: Determine a total vector including the relationship between each feature vector according to the similarity of each feature vector.

[0113] The electronic device may input each feature vector corresponding to each feature domain obtained in step 503 into the attention layer of the model, and determine the total vector including the relationship between each feature vector through the attention layer. Specifically, steps 504 and 505 may be performed.

[0114] The electronic device may determine the feature similarity between every two feature vectors based on the feature vectors corresponding to each feature domain. For example, a function for calculating the similarity may be set to determine the similarity between every two feature vectors in each feature vector.

[0115] For example, a similarity function φ may be set, and for two feature vectors E1 and E2, a feature similarity φ(E1, E2) between E1 and E2 may be determined by the similarity function φ. Feature similarity can characterize the relationship between feature vectors belonging to different feature domains.

[0116] For any combination of two eigenvectors Em and Ek, for:

[0117]

[0118] Among them, W query , W key It is a transformation matrix, and the values ​​in the matrix can be updated during the model training process.

[0119] Specifically, the total vector including the relationship between each feature vector can be determined according to the feature similarity between each feature vector. Specifically, the new expression of each feature vector can be determined according to each feature similarity, and then the total vector P can be determined according to the new expression of each feature vector, and the total vector includes the information of each feature vector itself and the information between the feature vectors.

[0120] Furthermore, the relationship coefficient between each feature vector and other feature vectors can be determined based on the feature similarity between each feature vector and other feature vectors. For example, for feature vector Em, the similarity φ(Em, Ek) between it and any other feature vector Ek can be determined. The relationship coefficient between Em and each other feature vector can be determined based on each feature similarity φ(Em, Ek) of Em.

[0121] In practical applications, a method for determining the relationship coefficient can be pre-set, and then the relationship coefficient between each eigenvector and other eigenvectors can be determined according to the method, and the relationship coefficient is used to characterize the relationship between the two eigenvectors. Specifically, the relationship coefficient between the eigenvector Em and the eigenvector Ek can be determined based on the following formula:

[0122]

[0123] The feature vector Em and the feature vector Ek are any two of the feature vectors determined by the encoding networks. K is the number of feature domains.

[0124] The feature domain embedding vector of each feature vector may also be determined based on the relationship coefficient between each feature vector and other feature vectors and each feature vector; wherein the total vector includes the feature domain embedding vector of each feature vector.

[0125] Specifically, any two eigenvectors have a relationship coefficient. Therefore, the feature domain embedding vector of a eigenvector can be determined based on the relationship coefficient between the eigenvector and other eigenvectors. The feature domain embedding vector includes information about the eigenvector and other eigenvectors.

[0126] Furthermore, for any feature vector Em, its feature domain embedding vector can be:

[0127]

[0128] Among them, W value It is a transformation matrix, and the values ​​in the matrix can be updated during the model training process.

[0129] In practical applications, the total vector can be generated based on the feature domain embedding vector of each feature vector. The total vector represents the embedding vector of each different feature domain. The result of global pooling.

[0130] The total vector P can be:

[0131]

[0132] Step 506: input the total vector into the compression network to obtain a compressed vector.

[0133] Among them, in the solution provided by the present application, the determination submodule includes a compression network, and the electronic device can use the compression network to process the total vector to obtain a compressed vector corresponding to the total vector.

[0134] Specifically, the compression network has weight parameters, which are obtained by training the model. The compression network is responsible for reducing the dimension of the total vector P.

[0135] Furthermore, when the total vector is processed based on the compression network, the total vector can be encoded according to the encoding function and the first neural network parameter to obtain the encoded vector. The first neural network parameter is obtained by training the model.

[0136] The encoding vector Zc of the total vector P is:

[0137] Z c =h(P;θ e )

[0138] Where h() is the encoding function and θe is the first neural network parameter.

[0139] In actual application, a decoding function may also be set in the compression network, and the electronic device may decode the encoding vector according to the decoding function and the second neural network parameter to obtain a decoding vector. The second neural network parameter is obtained by training the model.

[0140] The decoded vector X' of the encoded vector Zc is:

[0141] X′=g(Z c θ d )

[0142] Where g() is the decoding function and θd is the second neural network parameter.

[0143] Specifically, the reconstruction error Zr may be determined according to the features included in the data to be evaluated X and the decoding vector X′. The reconstruction error Zr is used to characterize the difference between the features of the decoding vector X′ and the data to be evaluated X.

[0144] Furthermore, r =f(X, X′). The function f() is used to determine the difference between the encoding vector and the characteristics of the data to be evaluated. For example, the decoding vector can be subtracted from the characteristics of the data to be evaluated to obtain the reconstruction error Zr.

[0145] In practical applications, the encoding vector Z c , reconstruction error z r , determine the compression vector Z. The compression vector Z is:

[0146] Z=[Z c , Z r ]

[0147] Among them, Z c , Z r Concatenate and obtain the compressed vector Z.

[0148] Specifically, by processing the total vector in this way to obtain a compressed vector, it is possible to avoid the problem that feature information is lost when extracting features of each feature domain in the data to be evaluated through the encoding network, resulting in the total vector P obtained not being able to accurately reflect the information included in the data to be evaluated.

[0149] Step 507, determining the expectation corresponding to the data to be evaluated according to the compression vector and the preset parameters corresponding to each Gaussian distribution; the number of Gaussian distributions is the number of feature domains; the expectation is used to characterize the distance between the data to be evaluated and each Gaussian distribution.

[0150] Furthermore, preset parameters corresponding to each Gaussian distribution may be set in the determination submodule, and the preset parameters may be determined by model training.

[0151] In actual application, a total of K preset parameters of Gaussian distribution are set, that is, Gaussian distributions consistent with the number of feature domains are set. By training the model, the preset parameters of each Gaussian distribution can characterize the characteristics of a feature domain.

[0152] Among them, the expectation corresponding to the data to be evaluated can be determined according to the compression vector and the preset parameters corresponding to each Gaussian distribution; the expectation is used to characterize the distance between the data to be evaluated and each Gaussian distribution. Through the expectation, it can be determined whether the data to be evaluated is abnormal.

[0153] Specifically, the preset parameters of each Gaussian distribution may include the probability of occurrence of the Gaussian distribution. Mean Covariance matrix The compression coding can be processed according to the preset parameters of each Gaussian distribution to determine the expectation, that is, the expectation of the data to be evaluated.

[0154] Furthermore, the expectation E(z) of the i-th data to be evaluated is:

[0155]

[0156] Where zi represents the compressed vector of the i-th data to be evaluated.

[0157] Specifically, the expectation is used to characterize the distance between the data to be evaluated and each Gaussian distribution. Therefore, whether the data to be evaluated has an abnormality can be determined through the expectation.

[0158] Step 508: Map the expectation of the data to be evaluated into a risk score according to a preset mapping function.

[0159] In practical application, the solution provided by the present application is also provided with a preset mapping function, which is used to map the expectations of the data to be evaluated to the risk score. Since there is a large floating range between the expectations of different data to be evaluated, the expectations of the data to be evaluated are mapped to the risk score, and each expectation can be mapped to the same scale range, so as to provide a reference risk score.

[0160] Among them, the i-th data to be evaluated z i The risk score score(z i ) can be:

[0161] mscore(z i )=sigmoid(log(E(z)+abs(γ(E(z))+ε))

[0162] Specifically, γ(E(z)) is the threshold partition function of all sample spaces, and min, average, k-σ, etc. can be selected. ε can be taken as 0.01 as a hyperparameter.

[0163] Step 508: If the risk score is greater than the threshold, a warning is issued.

[0164] Furthermore, a threshold for evaluating the risk score can be set, and if the risk score is greater than the threshold, an early warning can be issued. For example, an early warning message can be sent to users related to the business activity.

[0165] Figure 6 A flowchart of a method for training a model for assessing business activity risks, shown as an exemplary embodiment of the present application.

[0166] like Figure 6 As shown, the training method of the model for assessing business activity risks provided by the present application includes:

[0167] Step 601, obtaining a plurality of sample data corresponding to business activities; each sample data includes a plurality of data features, and each data feature corresponds to a feature domain.

[0168] The method provided in the present application may be executed by an electronic device with computing capabilities, such as a computer.

[0169] Specifically, a risk assessment model can be pre-built, which includes: an encoding network, an attention layer, and a determination submodule corresponding to each feature domain. Then the built model is trained. The trained risk assessment model can be deployed in the server of the e-commerce platform, and the server can evaluate and process the business data to be evaluated of the business activities based on the deployed risk assessment model.

[0170] Figure 7 A schematic diagram of a risk assessment model shown as an exemplary embodiment of the present application.

[0171] like Figure 7 As shown, the risk assessment model provided by the present application may include multiple encoding networks 71, attention layers 72, and determination submodules 73. The encoding network 71, attention layer 72, and determination submodule 73 are connected in sequence, and the model can be trained using sample data of business activities.

[0172] Furthermore, the electronic device may obtain multiple sample data corresponding to the business activity, each sample data may include multiple data features, and each data feature corresponds to a feature domain. For example, M pieces of sample data may be obtained, and each sample data may include N features.

[0173] In actual application, multiple feature domains can be pre-set. For example, a first feature domain related to registered account information can be set, and a second feature domain related to illegal reselling discount information can also be set. Among the data features of the sample data, the features related to the registered account can be divided into the first feature domain, so that these features have a corresponding relationship with the first feature domain, and the features of illegal reselling discount information can also be divided into the second feature domain. In this way, the data features used to represent the same type of information can be divided into the same feature domain, so that the model can learn the association between the data features used to represent the same type of information.

[0174] For example, the business data to be evaluated includes N data features, and these N features can be divided into K feature domains according to the information represented by each data feature.

[0175] Step 602: Input the data features corresponding to the same feature domain in each sample data into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain.

[0176] Among them, the constructed model is provided with multiple encoding networks corresponding to each feature domain. Through the encoding network corresponding to the feature domain, the data features of the feature domain are processed to obtain the feature vector corresponding to the feature domain.

[0177] The electronic device can input the data features corresponding to the same feature domain among the multiple data features of the sample data into the coding network corresponding to the feature domain, and then use the coding network to process the data features of the feature domain. For example, multiple data features divided into the first feature domain are input into the first coding network, and multiple data features divided into the second feature domain are input into the second coding network.

[0178] The electronic device processes the input data features of the coding network based on the coding network to obtain a feature vector. For example, the electronic device processes the input data features based on the first coding network to obtain a feature vector. The electronic device processes the input data features based on the second coding network to obtain a feature vector.

[0179] Therefore, by processing the data features belonging to the feature domain through the encoding network corresponding to the feature domain, the information in the domain can be fully combined and enhanced to obtain the feature vector of the feature domain.

[0180] Specifically, the sparse features of each feature domain are encoded into a dense feature vector E of a specified dimension through the encoding network. K for:

[0181] E K =σ(X K θ f )

[0182] X K is the feature vector of the kth sample data, θ f are the parameters in the encoding network.

[0183] Specifically, the weight parameters in each encoding network can be continuously updated during the training iteration process. The structures of each encoding network can be the same or different, and can be specifically set according to needs.

[0184] Step 603: Input the feature vector corresponding to the feature domain into the attention layer to obtain a total vector corresponding to each sample data including the relationship between each feature vector.

[0185] Furthermore, in the solution provided by the present application, the constructed model also includes an attention layer, and the feature vectors corresponding to each feature domain can be input into the attention layer. The attention layer can process the feature vectors of each feature domain to obtain a total vector including the relationship between each feature vector.

[0186] In actual application, the electronic device can determine the relationship between two feature vectors according to the weights in the attention layer, and the weights in the attention layer can be updated during the training process. The attention layer can determine the new expression of the feature vector based on the relationship between the feature vector and other feature vectors, so that the information of other feature vectors can be embedded in the new expression. For example, the new expression of the first feature vector can be determined based on the relationship between the first feature vector and each other feature vector.

[0187] The electronic device can construct a total vector including the relationship between each feature vector according to the new expression of each feature vector. The total vector includes the relationship between each feature vector, so the feature vector can integrate the feature vectors of multiple information categories in each sample data, and then accurately extract the information included in the sample data.

[0188] Step 604, input the total vector of each sample data into the determination submodule, use the determination submodule to determine the expectation of each sample data, and train the model according to the expectation of each sample data to obtain a model for evaluating the risk of business activities.

[0189] Furthermore, the constructed model also includes a determination submodule, and the electronic device can input the determined total vector into the determination submodule, and process the total vector according to the weight value in the determination submodule to obtain the expectation of each sample data.

[0190] In one implementation, the electronic device may first compress the total vector of each sample data to reduce the dimension of the total vector, and then process the vector after the dimension reduction to obtain the expectation of each total vector, that is, the expectation corresponding to each sample data.

[0191] In practical applications, the probability vector of each sample belonging to each Gaussian distribution can be determined based on the total vector after compression of each sample data. For example, if there are k feature domains in total, k Gaussian distributions can be set. For each batch of sample data, the Gaussian distribution to which each sample data in this batch belongs can be determined, and then the probability vector of the sample belonging to each Gaussian distribution can be obtained.

[0192] The parameters of each Gaussian distribution can also be estimated based on the probability vector, the number of sample data, and the total vector after compression of each sample data. During the training iteration process, the parameters of each Gaussian distribution will also be updated.

[0193] The expectation of each sample data can be determined based on the estimated Gaussian distribution parameters and the total vector after compression of each sample data. The expectation can be used to represent the probability that the sample belongs to K Gaussian distributions. Among them, by training the model with a large number of samples, the probability that each sample data belongs to K Gaussian distributions can be constructed in the model, that is, multiple Gaussian distribution spaces are constructed. When the model is used to process the business data to be evaluated, the probability that the business data to be evaluated belongs to each Gaussian distribution can also be determined, and then the distance between the business data to be evaluated and each Gaussian distribution space can be determined, so that abnormal sample data can be identified.

[0194] You can also set the objective function, which can be constructed based on the expectations of each sample, and use the objective function to perform gradient backpropagation to update the encoding network, attention layer, and weight values ​​of the sub-modules corresponding to each feature domain in the model.

[0195] Figure 8 A flowchart of a method for training a model for assessing business activity risks is shown as another exemplary embodiment of the present application.

[0196] like Figure 8 As shown, the training method of the model for assessing business activity risks provided by the present application includes:

[0197] Step 801, obtaining offline features corresponding to business activities; obtaining each real-time data corresponding to the business activities.

[0198] Step 802, determining each sample data of the business activity according to the offline feature and each real-time data; each sample data includes a plurality of data features, and each data feature corresponds to a feature domain.

[0199] The features in the sample data may include both offline features and real-time features. For example, offline features corresponding to business activities may be pre-set, and when real-time data corresponding to the business activity exists, the offline features of the business activity may be obtained and processed together with the real-time data to obtain sample data corresponding to the business activity.

[0200] Specifically, offline features can accurately reflect the characteristics of business activities, while online real-time data can be generated in real time. By processing the real-time features of business activities and offline features, we can not only extract feature information that is accurate with the business activities, but also ensure the real-time nature of sample data, thereby improving the accuracy of the trained model.

[0201] Furthermore, a feature processing engine may be provided. For example, after generating real-time data related to the first business activity, offline features of the first business activity may be obtained, and the feature processing engine may be used to process the offline features and the real-time data to obtain sample data X including multiple data features.

[0202] Step 803: input the data features of each sample data corresponding to the same feature domain into the encoding network corresponding to the feature domain to obtain the feature vector corresponding to the feature domain.

[0203] The execution principle and method of step 803 are similar to those of step 602 and will not be described in detail.

[0204] Step 804: Determine the feature similarity between every two feature vectors based on the feature vectors corresponding to each feature domain.

[0205] Step 805: Determine a total vector including the relationship between each feature vector according to the similarity of each feature vector.

[0206] The electronic device may input each feature vector corresponding to each feature domain obtained in step 803 into the attention layer of the model, and determine the total vector including the relationship between each feature vector through the attention layer. Specifically, steps 804 and 805 may be performed.

[0207] The electronic device may determine the feature similarity between every two feature vectors in each sample data according to the feature vectors corresponding to each feature domain in each sample data. For example, a function for calculating similarity may be set to determine the similarity between every two feature vectors in each feature vector.

[0208] For example, a similarity function φ may be set, and for two feature vectors E1 and E2, a feature similarity φ(E1, E2) between E1 and E2 may be determined by the similarity function φ. Feature similarity can characterize the relationship between feature vectors belonging to different feature domains.

[0209] For any combination of two eigenvectors Em and Ek in a sample data, for:

[0210]

[0211] Among them, W query , W key To transform the matrix, the values ​​in the matrix can be updated during model training.

[0212] Specifically, the total vector of the relationship between each feature vector of each sample data can be determined according to the similarity of each feature of each sample data. Specifically, the new expression of each feature vector can be determined according to the similarity of each feature of a sample data, and then the total vector P can be determined according to the new expression of each feature vector, and the total vector includes the information of each feature vector itself and the information between the feature vectors.

[0213] Furthermore, the relationship coefficient between each feature vector and other feature vectors can be determined based on the feature similarity between each feature vector and other feature vectors, and the relationship coefficient is used to characterize the relationship between two feature vectors. For example, for a feature vector Em in a sample data, the similarity φ(Em, Ek) between it and any other feature vector Ek in the sample data can be determined. The relationship coefficient between Em and other feature vectors in the sample data can be determined based on the feature similarities φ(Em, Ek) of Em.

[0214] In practical applications, a method for determining the relationship coefficient can be pre-set, and the relationship between each eigenvector and other eigenvectors can be determined based on the method. Specifically, the relationship coefficient between the eigenvector Em and the eigenvector Ek can be determined based on the following formula:

[0215]

[0216] The feature vector Em and the feature vector Ek are any two of the feature vectors of the sample data determined by each encoding network. K is the number of feature domains.

[0217] The feature domain embedding vector of each feature vector in the sample data can also be determined based on the relationship coefficient between each feature vector and other feature vectors in the sample data and each feature vector; wherein the total vector includes the feature domain embedding vector of each feature vector.

[0218] Specifically, any two feature vectors in the sample data have a relationship coefficient. Therefore, the feature domain embedding vector of a feature vector in the sample data can be determined based on the relationship coefficient between the feature vector and other feature vectors in the sample data. The feature domain embedding vector includes information about the feature vector and information about other feature vectors in the sample data.

[0219] Furthermore, for any feature vector Em, its feature domain embedding vector can be:

[0220]

[0221] Among them, W value It is a transformation matrix, and the values ​​in the matrix can be updated during the model training process.

[0222] In practical applications, the total vector of each sample data can be generated based on the feature domain embedding vector of each feature vector in each sample data. The total vector represents the embedding vector of each different feature domain. The result of global pooling.

[0223] The total vector P can be:

[0224]

[0225] Step 806: input the total vector of each sample data into the compression network to obtain the compressed vector of each sample data.

[0226] Among them, in the solution provided by the present application, the determination submodule includes a compression network, and the electronic device can use the compression network to process the total vector of each sample data to obtain a compressed vector of each sample data.

[0227] Specifically, the compression network has weight parameters, which can be updated during the model training process. The compression network is responsible for reducing the dimension of the total vector P of each sample data.

[0228] Furthermore, when the total vector of the sample data is processed based on the compression network, the total vector of each sample data can be encoded according to the encoding function and the first neural network parameter to obtain the encoding vector of each sample data. The first neural network parameter can be updated during the model training process.

[0229] The encoding vector Zc of the total vector P of a sample data is:

[0230] Z c =h(P;θ e )

[0231] Where h() is the encoding function and θe is the first neural network parameter.

[0232] In actual application, a decoding function can also be set in the compression network, and the electronic device can decode the encoding vector of a sample data according to the decoding function and the second neural network parameters to obtain the decoding vector of the sample data. The second neural network parameters are updated during the model training process.

[0233] The decoding vector X' of the encoding vector Zc of a sample data is:

[0234] X′=g(Z c θ d )

[0235] Where g() is the decoding function and θd is the second neural network parameter.

[0236] Specifically, the reconstruction error Zr of each sample data may be determined according to the data features included in each sample data X and the decoding vector X′ of each sample data. The reconstruction error Zr is used to characterize the difference between the decoding vector X′ of each sample data and the features of each sample data X.

[0237] Furthermore, r =f(X, X′). The function f() is used to determine the difference between the decoded vector of the sample data and the data feature of the sample data. For example, the decoded vector and the feature of the sample data can be subtracted to obtain the reconstruction error Zr.

[0238] In practical applications, the encoding vector Z of each sample data can be used c , the reconstruction error z of each sample data r , determine the compression vector Z of each sample data. The compression vector Z is:

[0239] Z=[Z c , Z r ]

[0240] Among them, Z c , Z r Concatenate and obtain the compressed vector Z.

[0241] Specifically, by processing the total vector of each sample data in this way to obtain a compressed vector, it is possible to avoid the problem of losing feature information when extracting features of each feature domain in each sample data through the encoding network, resulting in the total vector P being unable to accurately reflect the information included in the sample data.

[0242] In an optional implementation, the reconstruction error Z of each sample data may also be calculated based on r Modify the parameters in the encoding network corresponding to each feature domain.

[0243] Furthermore, the data features X corresponding to each feature domain in the sample data can be processed by the encoding network to obtain the feature vectors E corresponding to each feature domain. By processing the feature vectors E corresponding to each feature domain, the decoding vector X′ can be obtained. If the encoding network can accurately extract the features in the sample data and obtain accurate feature vectors, then the final decoding vector X′ can represent the data features X in the sample data. Then the reconstruction error Zr determined based on the decoding vector X′ and the data features X of each sample data should be close to 0. The preset condition of the reconstruction error Zr can be set to 0, and the parameters θ in each encoding network can be updated based on this constraint. f .

[0244] Step 807, input the compressed vector of each sample data into the estimation network to obtain the probability vector that each sample data belongs to each Gaussian distribution; wherein the number of Gaussian distributions is the number of feature domains.

[0245] In actual application, an estimation network is further provided in the determination submodule, and the estimation network is used to estimate the probability vector of each Gaussian distribution according to the compression vector of each sample data.

[0246] Among them, the third neural network parameters and preset functions are set in the estimation network, and the electronic device can process the compression vector of each sample data according to the parameters set in the estimation network.

[0247] Specifically, the output feature p of the sample data is:

[0248] p=MLN(Z;θ m )

[0249] Furthermore, Z is the compression vector of sample data, θ m is the third neural network parameter. MLN() is a neural network, and the electronic device takes Z as input, processes Z using MLN() and the third neural network parameter to obtain p. MLN() may be a subnetwork in the estimation network.

[0250] For example, if m sample data are used in a training process, then m compressed vectors can be obtained for these m sample data, and then m output features p can be obtained.

[0251] In actual application, a preset function may also be set in the estimation network, and the preset function is used to determine the probability vector that each sample data belongs to each Gaussian distribution.

[0252] Among them, the probability vector of each Gaussian distribution for:

[0253]

[0254] Specifically, softmax() is a preset function, and p is the output feature of the sample data. For example, each time an output feature p is obtained, the obtained output feature is processed by softmax() to obtain the probability vector of each Gaussian distribution.

[0255] If there are m sample data and K Gaussian distributions, then the probability vector Can be m*K dimensional. Probability vector Each value in may correspond to a combination of a sample data and a Gaussian distribution, and the value is used to indicate whether the sample data belongs to the Gaussian distribution. For example, if the value is 1, it indicates that the sample data belongs to the Gaussian distribution space.

[0256] for example, Corresponding to the first sample data and the second Gaussian distribution space, It can characterize whether the first sample data belongs to the second Gaussian distribution space.

[0257] Furthermore, K Gaussian distributions can be set, where K is the number of feature domains. Through a large amount of sample data, the probability vectors of K Gaussian distributions can be obtained.

[0258] Step 808, determining the expectation of each sample data according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data; wherein the expectation is used to characterize the distance between the sample data and each Gaussian distribution.

[0259] In practical applications, we can first construct the space of each Gaussian distribution based on the probability vector of each sample data belonging to each Gaussian distribution and the compression vector of each sample data, and then determine the expectation of each sample data based on the space of each Gaussian distribution. The expectation is used to characterize the distance between the sample data and each Gaussian distribution.

[0260] Among them, the parameters of each Gaussian distribution can be determined according to the probability vector of each sample data belonging to each Gaussian distribution and the compression vector of each sample number, and then each Gaussian distribution space can be constructed.

[0261] Specifically, the parameters of each Gaussian distribution may include occurrence probability, mean, and covariance matrix. The probability, mean, and covariance matrix of each Gaussian distribution may be determined based on the probability vector of each sample data belonging to each Gaussian distribution and the compression vector of each sample data.

[0262] Furthermore, the occurrence probability of each Gaussian distribution is determined according to the probability vector that each sample data belongs to each Gaussian distribution and the number of sample data.

[0263] The probability of occurrence of the kth Gaussian distribution for:

[0264]

[0265] Among them, if the i-th sample data belongs to the k-th Gaussian distribution, then otherwise, There are M sample data in total. Multiple Gaussian distributions corresponding to the kth Gaussian distribution can be determined based on the probability vector of each sample data belonging to each Gaussian distribution. Then determine the occurrence probability of the kth Gaussian distribution. It is used to characterize the occurrence probability of the kth Gaussian distribution in multiple samples, and then the characteristics of each Gaussian distribution space can be reflected through the occurrence probability of each Gaussian distribution.

[0266] Specifically, the mean of each Gaussian distribution may also be determined according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data.

[0267] Furthermore, the mean of the kth Gaussian distribution is for:

[0268]

[0269] In practical applications, zi is the compressed vector of the i-th sample data. It can represent the mean of the compressed vector of the sample data belonging to the kth Gaussian distribution. Through the mean of each sample data, the characteristic information of each Gaussian distribution can be represented.

[0270] The covariance matrix of each Gaussian distribution may also be determined according to the mean of each Gaussian distribution, the probability vector that each sample data belongs to each Gaussian distribution, and the compression vector of each sample data.

[0271] Specifically, the covariance matrix of the kth Gaussian distribution is for:

[0272]

[0273] Furthermore, the probability, mean, and covariance matrix of each Gaussian distribution can be determined by the above method, and then the Gaussian distribution parameters can be obtained. The parameters of the Gaussian distribution can characterize the characteristics of a Gaussian distribution. Each time the model is trained with samples, the Gaussian distribution parameters can be updated according to the compression vector of the sample data. For example, the parameters of the Gaussian distribution can be obtained by training the first batch of sample data. When processing the second batch of sample data, the parameters of the Gaussian distribution can be updated by combining the compression variables of the first batch of sample data and the compression variables of the second batch of sample data.

[0274] In practical applications, when training the model using sample data, each time after determining the parameters of the Gaussian distribution, the expectation of each sample data can be determined based on the compression vector of each sample data in the current batch and the parameters of each Gaussian distribution.

[0275] The expectation is used to characterize the distance between the sample data and each Gaussian distribution. For example, the first expectation for the first sample is used to characterize the distance between the first sample and each Gaussian distribution.

[0276] Specifically, the compression coding of each sample data may be processed according to the determined parameters of each Gaussian distribution to determine the expectation of each sample data.

[0277] Furthermore, the expectation E(z) of the i-th sample data is:

[0278]

[0279] Where zi represents the compressed vector of the i-th sample data.

[0280] Step 809, according to the expectation of each sample data, modify the attention layer, determine the network parameters in the submodule, and obtain a model for assessing the risk of business activities. The determination submodule includes a compression network and an estimation network.

[0281] Specifically, a function for training the model can be pre-set, which can be used to train the encoding network, the attention layer, and determine the network parameters in the submodule. Specifically, the model can be trained using the following formula:

[0282]

[0283]

[0284]

[0285] Among them, θe is the first neural network parameter in the compressed network, θd is the second neural network parameter in the compressed network, and θ m To estimate the parameters of the third neural network in the network.

[0286] The model trained by this method can be applied in the method shown in the figure.

[0287] Fig. 9 This is a structural diagram of a device for determining risk information of a business activity shown as an exemplary embodiment of the present application.

[0288] like Fig. 9 As shown, the risk information determination device 900 for business activities provided in this embodiment includes:

[0289] The acquisition module 910 is used to acquire the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes a plurality of data features, each of which corresponds to a feature domain;

[0290] An encoding module 920, configured to input the data features corresponding to the same feature domain into an encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0291] An embedding module 930, configured to input the feature vectors corresponding to the feature domain into an attention layer to obtain a total vector including the relationship between each of the feature vectors;

[0292] An expectation determination module 940, configured to input the total vector into a determination submodule, and use the determination submodule to determine an expectation corresponding to the business data to be evaluated;

[0293] The risk determination module 950 is used to determine risk information corresponding to the business data to be evaluated according to the expectation, and the risk information is used to indicate the risk level of the business data to be evaluated.

[0294] The specific principles and implementation methods of the risk information determination device for business activities provided in this embodiment are similar to Figure 1 The embodiments shown are similar and will not be described again here.

[0295] Fig.10 This is a structural diagram of a device for determining risk information of a business activity shown in another exemplary embodiment of the present application.

[0296] like Fig.10 As shown, based on the above embodiment, the risk information determination device 1000 for business activities provided in this embodiment, the embedded module 930 includes:

[0297] A similarity determination unit 931 is used to determine the feature similarity between every two feature vectors according to the feature vectors corresponding to each feature domain;

[0298] The embedding unit 932 is used to determine a total vector including the relationship between each feature vector according to the similarity of each feature.

[0299] Optionally, the embedding unit 932 is specifically used for:

[0300] Determine, according to the feature similarity between each feature vector and other feature vectors, a relationship coefficient between each feature vector and other feature vectors, wherein the relationship coefficient is used to characterize the relationship between the two feature vectors;

[0301] Determine a feature domain embedding vector for each feature vector according to the relationship coefficient between each feature vector and other feature vectors and each feature vector; wherein the feature domain embedding vector is used to represent the feature vector embedded with other feature vector information;

[0302] According to each feature domain embedding vector, a total vector including the relationship between each of the feature vectors is determined.

[0303] Optionally, the expectation determination module 940 includes:

[0304] A compression unit 941, configured to input the total vector into a compression network to obtain a compressed vector;

[0305] The expectation determination unit 942 is used to determine the expectation corresponding to the data to be evaluated based on the compression vector and the preset parameters corresponding to each of the Gaussian distributions; the number of the Gaussian distributions is the number of the feature domains; the expectation is used to characterize the distance between the data to be evaluated and each of the Gaussian distributions.

[0306] Optionally, the compression unit 941 is specifically used for:

[0307] Encoding the total vector according to the encoding function and the first neural network parameter to obtain an encoding vector;

[0308] Decoding the encoding vector according to the decoding function and the second neural network parameter to obtain a decoding vector;

[0309] Determine a reconstruction error according to the characteristics of the data to be evaluated and the decoding vector; the reconstruction error is used to characterize the difference between the characteristics of the decoding vector and the data to be evaluated;

[0310] The compression vector is determined according to the encoding vector and the reconstruction error.

[0311] The risk determination module 950 is specifically used for:

[0312] According to a preset mapping function, the expectation of the data to be evaluated is mapped into the risk score.

[0313] Optionally, the acquisition module 910 includes:

[0314] An acquisition unit 911 is used to acquire offline features corresponding to the business activity and acquire real-time data corresponding to the business activity;

[0315] The determination unit 912 is used to determine the data to be evaluated of the business activity according to the offline characteristics and the real-time data.

[0316] Optionally, the device further includes an early warning module 960, which is used to:

[0317] If the risk score is greater than a threshold, an early warning is issued.

[0318] The specific principles and implementation methods of the device provided in this embodiment are Figure 5 The embodiments shown are similar and will not be described again here.

[0319] Fig.11 A structural diagram of a training device for a model for assessing business activity risks, shown as an exemplary embodiment of the present application.

[0320] The model includes: an encoding network corresponding to each feature domain, an attention layer, and a determination submodule;

[0321] like Fig.11 As shown, the training device 1100 of the model for assessing business activity risks provided by the present application includes:

[0322] The acquisition module 1110 is used to acquire a plurality of sample data corresponding to the business activity; each of the sample data includes a plurality of data features, and each of the data features corresponds to a feature domain;

[0323] The encoding module 1120 is used to input the data features corresponding to the same feature domain in each sample data into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain;

[0324] An embedding module 1130, configured to input a feature vector corresponding to the feature domain into the attention layer to obtain a total vector corresponding to each sample data including a relationship between each feature vector;

[0325] An expectation determination module 1140, configured to input the total vector of each sample data into the determination submodule, and determine the expectation of each sample data using the determination submodule;

[0326] The training module 1150 is used to train the model according to the expectations of each sample data to obtain a model for assessing the risk of business activities.

[0327] The specific principles and implementation methods of the device provided in this embodiment are Figure 6 The embodiments shown are similar and will not be described again here.

[0328] Fig.12 A structural diagram of a training device for a model for assessing business activity risks, shown as another exemplary embodiment of the present application.

[0329] like Fig.12 As shown, based on the above embodiment, on the basis of the training device 1200 for the model for assessing business activity risks provided by the present application, the embedding module 1130 includes:

[0330] A similarity determination unit 1131 is used to determine the feature similarity between every two feature vectors in each sample data according to the feature vectors corresponding to each feature domain in each sample data;

[0331] The embedding unit 1132 is used to determine, according to the similarities of the features in each sample data, a total vector of the relationship between each feature vector included in each sample data.

[0332] Optionally, the embedding unit 1132 is specifically used for:

[0333] Determine the relationship coefficient between each feature vector and other feature vectors in each sample data according to the similarity of each feature of each sample data; the relationship coefficient is used to characterize the relationship between the two feature vectors;

[0334] Determine a feature domain embedding vector for each feature vector in each sample data according to the relationship coefficient between each feature vector and other feature vectors in each sample data and each feature vector;

[0335] According to the feature domain embedding vector of each feature vector in each sample data, a total vector of each sample data is generated.

[0336] Optionally, the expectation determination module 1140 includes:

[0337] A compression unit 1141, configured to input the total vector of each sample data into a compression network to obtain a compressed vector of each sample data;

[0338] An estimation unit 1142, configured to input the compression vector of each sample data into an estimation network to obtain a probability vector that each sample data belongs to each Gaussian distribution; wherein the number of the Gaussian distributions is the number of the feature domains;

[0339] The expectation determination unit 1143 is used to determine the expectation of each sample data according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data; wherein the expectation is used to characterize the distance between the sample data and each Gaussian distribution.

[0340] Optionally, the compression unit 1141 is specifically used for:

[0341] Encoding the total vector of each sample data according to the encoding function and the first neural network parameter to obtain an encoding vector of each sample data;

[0342] Decoding the encoding vector of each sample data according to the decoding function and the second neural network parameters to obtain a decoding vector of each sample data;

[0343] Determine a reconstruction error of each sample data according to the data feature of each sample data and the decoding vector of each sample data; the reconstruction error is used to characterize the difference between the data feature of each sample data and the decoding vector of each sample data;

[0344] The compression vector of each sample data is determined according to the encoding vector of each sample data and the reconstruction error of each sample data.

[0345] Optionally, the training module 1150 includes a first training unit 1151, which is used to:

[0346] The parameters in the encoding network corresponding to each of the feature domains are corrected according to the reconstruction error of each sample data.

[0347] Optionally, the estimating unit 1142 is specifically configured to:

[0348] According to the third neural network parameters, converting the compressed vector of each sample data into an output feature;

[0349] According to the preset function and the output feature of each sample data, the probability vector of each sample data belonging to each Gaussian distribution is determined.

[0350] Optionally, the expectation determination unit 1143 is specifically configured to:

[0351] Determine the parameters of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample number;

[0352] The expectation of each of the sample data is determined according to the compression vector of each of the sample data and the parameters of each of the Gaussian distributions.

[0353] Optionally, the parameters of the Gaussian distribution include occurrence probability, mean, and covariance matrix;

[0354] The expectation determination unit 1143 is specifically used for:

[0355] Determine the occurrence probability of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the number of the sample data;

[0356] Determine the mean of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data;

[0357] The covariance matrix of each Gaussian distribution is determined according to the mean of each Gaussian distribution, the probability vector that each sample data belongs to each Gaussian distribution, and the compression vector of each sample data.

[0358] Optionally, the training module 1150 includes a second training unit 1152, which is used to:

[0359] According to the expectation of each sample data, the network parameters in the attention layer and the determination submodule are modified to obtain a model for evaluating the risk of business activities.

[0360] Optionally, the acquisition module 1110 includes:

[0361] The acquisition unit 1111 is used to acquire the offline features corresponding to the business activity and acquire each real-time data corresponding to the business activity;

[0362] The determination unit 1112 is used to determine each sample data of the business activity according to the offline feature and each real-time data.

[0363] The specific principles and implementation methods of the device provided in this embodiment are Figure 8 The embodiments shown are similar and will not be described again here.

[0364] Fig.13 The figure is a structural diagram of an electronic device according to an exemplary embodiment of the present invention.

[0365] like Fig.13 As shown, the electronic device provided in this embodiment includes:

[0366] Memory 131;

[0367] Processor 132; and

[0368] Computer programs;

[0369] The computer program is stored in the memory 131 and is configured to be executed by the processor 132 to implement any of the risk information determination methods for business activities or the training methods for models for assessing business activity risks as described above.

[0370] This embodiment also provides a computer-readable storage medium having a computer program stored thereon.

[0371] The computer program is executed by a processor to implement any of the above-described methods for determining risk information of a business activity or methods for training a model for assessing the risk of a business activity.

[0372] This embodiment also provides a computer program, including program code. When a computer runs the computer program, the program code executes any of the risk information determination methods for business activities or the training methods for models for assessing business activity risks as described above.

[0373] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0374] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining risk information of business activities on an online shopping platform, characterized in that: include: Acquire the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes a plurality of data features, and the features related to the registered account among the plurality of data features are divided into a first feature domain, and the features related to the reselling preferential information among the plurality of data features are divided into a second feature domain, and each feature domain corresponds to at least one data feature; Inputting the data features corresponding to the same feature domain into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain; Inputting the feature vectors corresponding to the feature domains into the attention layer, and determining the feature similarity between every two feature vectors according to the feature vectors corresponding to each feature domain; Determine, according to the feature similarity between each feature vector and other feature vectors, a relationship coefficient between each feature vector and other feature vectors, wherein the relationship coefficient is used to characterize the relationship between the two feature vectors; Determine a feature domain embedding vector for each feature vector according to the relationship coefficient between each feature vector and other feature vectors and each feature vector; wherein the feature domain embedding vector is used to represent the feature vector embedded with other feature vector information; Determining a total vector including the relationship between each of the feature vectors according to the embedding vectors of each feature domain; Inputting the total vector into a determination submodule, and using the determination submodule to determine an expectation corresponding to the business data to be evaluated; Risk information corresponding to the business data to be evaluated is determined according to the expectation, where the risk information is used to indicate a risk level of the business data to be evaluated.

2. The method according to claim 1, characterized in that The step of inputting the total vector into a determination submodule and using the determination submodule to determine an expectation corresponding to the data to be evaluated includes: Inputting the total vector into a compression network to obtain a compressed vector; According to the compression vector and preset parameters corresponding to each Gaussian distribution, an expectation corresponding to the data to be evaluated is determined; the number of the Gaussian distributions is the number of the feature domains; and the expectation is used to characterize the distance between the data to be evaluated and each Gaussian distribution.

3. The method according to claim 2, characterized in that The step of inputting the total vector into a compression network to obtain a compressed vector comprises: Encoding the total vector according to the encoding function and the first neural network parameter to obtain an encoding vector; Decoding the encoding vector according to the decoding function and the second neural network parameter to obtain a decoding vector; Determine a reconstruction error according to the characteristics of the data to be evaluated and the decoding vector; the reconstruction error is used to characterize the difference between the characteristics of the decoding vector and the data to be evaluated; The compression vector is determined according to the encoding vector and the reconstruction error.

4. The method according to claim 1, characterized in that: The step of determining a risk score corresponding to the data to be evaluated according to the expectation includes: According to a preset mapping function, the expectation of the data to be evaluated is mapped into the risk score.

5. The method according to any one of claims 1 to 4, characterized in that: Obtaining the data to be evaluated corresponding to the business activity, including: Acquire offline features corresponding to the business activities, and acquire real-time data corresponding to the business activities; The data to be evaluated of the business activity is determined according to the offline characteristics and the real-time data.

6. The method according to any one of claims 1 to 4, characterized in that: Also includes: If the risk score is greater than a threshold, an early warning is issued.

7. A training method for a model for assessing business activity risks on an online shopping platform, characterized in that: The model includes: an encoding network corresponding to each feature domain, an attention layer, and a determination submodule; The method comprises: Acquire multiple sample data corresponding to the business activity; each of the sample data includes multiple data features, and the features related to the registered account among the multiple data features are divided into a first feature domain, and the features related to the reselling preferential information among the multiple data features are divided into a second feature domain; each feature domain corresponds to at least one data feature; Input the data features corresponding to the same feature domain in each sample data into the encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain; Inputting the feature vector corresponding to the feature domain into the attention layer, and determining the feature similarity between every two feature vectors in each sample data according to the feature vector corresponding to each feature domain in each sample data; Determine the relationship coefficient between each feature vector and other feature vectors in each sample data according to the similarity of each feature of each sample data; the relationship coefficient is used to characterize the relationship between the two feature vectors; Determine a feature domain embedding vector for each feature vector in each sample data according to the relationship coefficient between each feature vector and other feature vectors in each sample data and each feature vector; Generate a total vector for each sample data according to the feature domain embedding vector of each feature vector in each sample data; The total vector of each sample data is input into the determination submodule, the expectation of each sample data is determined by using the determination submodule, and the model is trained according to the expectation of each sample data to obtain a model for evaluating the risk of business activities.

8. The method according to claim 7, characterized in that The step of inputting the total vector of each sample data into a determination submodule, and using the determination submodule to determine the expectation of each sample data comprises: Inputting the total vector of each sample data into a compression network to obtain a compressed vector of each sample data; Input the compressed vector of each sample data into the estimation network to obtain a probability vector that each sample data belongs to each Gaussian distribution; wherein the number of the Gaussian distributions is the number of the feature domains; The expectation of each sample data is determined according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data; wherein the expectation is used to characterize the distance between the sample data and each Gaussian distribution.

9. The method according to claim 8, characterized in that The step of inputting the total vector of each sample data into a compression network to obtain a compressed vector of each sample data comprises: Encoding the total vector of each sample data according to the encoding function and the first neural network parameter to obtain an encoding vector of each sample data; Decoding the encoding vector of each sample data according to the decoding function and the second neural network parameters to obtain a decoding vector of each sample data; Determine a reconstruction error of each sample data according to the data feature of each sample data and the decoding vector of each sample data; the reconstruction error is used to characterize the difference between the data feature of each sample data and the decoding vector of each sample data; The compression vector of each sample data is determined according to the encoding vector of each sample data and the reconstruction error of each sample data.

10. The method according to claim 9, characterized in that Also includes: The parameters in the encoding network corresponding to each of the feature domains are corrected according to the reconstruction error of each sample data.

11. The method according to claim 8, characterized in that The step of inputting the compressed vector of each sample data into an estimation network to obtain a probability vector of each sample data belonging to each Gaussian distribution includes: According to the third neural network parameters, converting the compressed vector of each sample data into an output feature; According to the preset function and the output feature of each sample data, the probability vector of each sample data belonging to each Gaussian distribution is determined.

12. The method according to claim 8, characterized in that The determining the expectation of each sample data according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data comprises: Determine the parameters of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample number; The expectation of each of the sample data is determined according to the compression vector of each of the sample data and the parameters of each of the Gaussian distributions.

13. The method according to claim 12, characterized in that The parameters of the Gaussian distribution include probability of occurrence, mean, and covariance matrix; Determining the parameters of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample number, including: Determine the occurrence probability of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the number of the sample data; Determine the mean of each Gaussian distribution according to the probability vector that each sample data belongs to each Gaussian distribution and the compression vector of each sample data; The covariance matrix of each Gaussian distribution is determined according to the mean of each Gaussian distribution, the probability vector that each sample data belongs to each Gaussian distribution, and the compression vector of each sample data.

14. The method according to claim 7, characterized in that The model is trained according to the expectation of each sample data to obtain a model for assessing the risk of business activities, including: According to the expectations of each sample data, the network parameters in the attention layer and the determination submodule are modified.

15. The method according to any one of claims 7 to 14, characterized in that: Acquire multiple sample data corresponding to the business activity, including: Acquire offline features corresponding to the business activity, and acquire each real-time data corresponding to the business activity; Each sample data of the business activity is determined according to the offline feature and each real-time data.

16. A device for determining risk information of business activities on an online shopping platform, characterized in that: include: an acquisition module, configured to acquire the business data to be evaluated corresponding to the business activity; the business data to be evaluated includes a plurality of data features, wherein the features related to the registered account among the plurality of data features are divided into a first feature domain, and the features related to the reselling preferential information among the plurality of data features are divided into a second feature domain, and each feature domain corresponds to at least one data feature; An encoding module, used for inputting the data features corresponding to the same feature domain into an encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain; An embedding module, configured to input the feature vectors corresponding to the feature domains into an attention layer, and determine the feature similarity between every two feature vectors according to the feature vectors corresponding to each feature domain; Determine, according to the feature similarity between each feature vector and other feature vectors, a relationship coefficient between each feature vector and other feature vectors, wherein the relationship coefficient is used to characterize the relationship between the two feature vectors; Determine a feature domain embedding vector for each feature vector according to the relationship coefficient between each feature vector and other feature vectors and each feature vector; wherein the feature domain embedding vector is used to represent the feature vector embedded with other feature vector information; Determining a total vector including the relationship between each of the feature vectors according to the embedding vectors of each feature domain; an expectation determination module, configured to input the total vector into a determination submodule, and use the determination submodule to determine an expectation corresponding to the service data to be evaluated; A risk determination module is used to determine risk information corresponding to the business data to be evaluated according to the expectation, wherein the risk information is used to indicate the risk level of the business data to be evaluated.

17. A training device for a model for assessing business activity risks on an online shopping platform, characterized in that: The model includes: an encoding network corresponding to each feature domain, an attention layer, and a determination submodule; The device comprises: An acquisition module, used to acquire a plurality of sample data corresponding to the business activity; each of the sample data includes a plurality of data features, wherein the features related to the registered account among the plurality of data features are divided into a first feature domain, and the features related to the reselling preferential information among the plurality of data features are divided into a second feature domain, and each feature domain corresponds to at least one data feature; An encoding module, used for inputting the data features corresponding to the same feature domain in each sample data into an encoding network corresponding to the feature domain to obtain a feature vector corresponding to the feature domain; An embedding module, used for inputting the feature vector corresponding to the feature domain into the attention layer, and determining the feature similarity between every two feature vectors in each sample data according to the feature vector corresponding to each feature domain in each sample data; Determine the relationship coefficient between each feature vector and other feature vectors in each sample data according to the similarity of each feature of each sample data; the relationship coefficient is used to characterize the relationship between the two feature vectors; Determine a feature domain embedding vector for each feature vector in each sample data according to the relationship coefficient between each feature vector and other feature vectors in each sample data and each feature vector; Generate a total vector for each sample data according to the feature domain embedding vector of each feature vector in each sample data; an expectation determination module, used for inputting the total vector of each sample data into the determination submodule, and determining the expectation of each sample data by using the determination submodule; The training module is used to train the model according to the expectations of each sample data to obtain a model for assessing the risk of business activities.

18. An electronic device, characterized in that: include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement the method according to any one of claims 1-6 or 7-15.

19. A computer-readable storage medium, characterized in that: A computer program is stored thereon, The computer program is executed by a processor to implement the method according to any one of claims 1-6 or 7-15.

20. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1-6 or 7-15.

Images