Deep learning method and device, medium, and equipment for tabular transaction data

Abstract

The embodiments of this specification provide a deep learning method, device, medium, and equipment for tabular transaction data. The method includes: extracting N prior features from the tabular transaction data and converting each of the prior features into a corresponding feature vector; calculating a first attention score matrix corresponding to the tabular transaction data; fusing each of the feature vectors with the corresponding first attention score vector to obtain a corresponding second attention score vector; and performing deep learning on the tabular transaction data according to a second attention score matrix formed by N second attention score vectors to determine the relationship between transaction behavior information and label information in the tabular transaction data. The embodiments of this specification provide a deep learning solution suitable for tabular transaction data, reducing the loss of prior knowledge and improving the accuracy of deep learning results.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the field of deep learning technology, and in particular, to a method, apparatus, medium, and device for deep learning of tabular transaction data. Background Art

[0002] In previous risk control business models, ensemble tree models offered excellent performance and ease of use. While neural network models have demonstrated significant success in image and natural language processing, they struggle to outperform ensemble tree models for tabular transaction data. This is primarily due to the fact that most neural network models rely heavily on end-to-end high-order feature crossover to gain additional information and improve performance. However, the features of tabular transaction data are often heavily artificially designed, leading to significant loss of valuable information through multiple layers of deep crossover.

[0003] In order to better integrate tabular transaction data with other types of data based on deep learning capabilities and realize the fusion of multimodal data under a single framework, it is necessary to improve the existing prior knowledge fusion scheme and provide a deep learning scheme suitable for tabular transaction data. Summary of the Invention

[0004] One or more embodiments of this specification describe a method, apparatus, medium, and device for deep learning of tabular transaction data, providing a deep learning solution suitable for tabular transaction data.

[0005] According to the first aspect, the deep learning method for tabular transaction data provided by the embodiments of this specification includes:

[0006] Extracting N prior features from the tabular transaction data, and converting each of the prior features into a corresponding feature vector; wherein N is a positive integer greater than 1; the tabular transaction data is formed by at least one transaction data item; each transaction data item includes transaction behavior information and corresponding label information; the label information indicates whether the corresponding transaction behavior information has a transaction risk and, if so, the type of transaction risk; and the N prior features include a label feature and at least one transaction behavior feature.

[0007] Calculating a first attention score matrix corresponding to the tabular transaction data; wherein the first attention score matrix is formed by N first attention score vectors, each of which represents a fusion ratio of a priori feature in the N priori features;

[0008] Fusing each of the feature vectors with the corresponding first attention score vector to obtain a corresponding second attention score vector;

[0009] Based on the second attention score matrix formed by N of the second attention score vectors, perform deep learning on the tabular transaction data to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0010] According to a second aspect, the deep learning device for lattice transaction data provided by the embodiments of this specification includes:

[0011] A feature extraction module, configured to extract N prior features from the tabular transaction data, and convert each of the prior features into a corresponding feature vector; where N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information characterizes whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature;

[0012] A first calculation module, configured to calculate a first attention score matrix corresponding to the tabular transaction data; where the first attention score matrix is formed by N first attention score vectors, and each first attention score vector characterizes the fusion proportion of a prior feature among the N prior features;

[0013] A first fusion module, configured to fuse each feature vector and the corresponding first attention score vector to obtain a corresponding second attention score vector;

[0014] A first learning module, configured to perform deep learning on the tabular transaction data according to the second attention score matrix formed by N of the second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0015] According to a third aspect, the computer-readable storage medium provided by the embodiments of this specification stores a computer program, and when the computer program is executed on a computer, the computer is made to execute the method provided in the first aspect.

[0016] According to a fourth aspect, the computing device provided by the embodiments of this specification includes a memory and a processor, the memory stores executable code, and when the processor executes the executable code, the method provided in the first aspect is implemented.

[0017] The deep learning method, device, medium, and equipment for tabular transaction data provided by the embodiments of this specification have the following beneficial effects: First, N prior features are extracted from the tabular transaction data, then the first attention score matrix corresponding to the tabular transaction data is calculated, the feature vector and the corresponding first attention score vector are fused to obtain the corresponding second attention score vector, and then deep learning is performed on the tabular transaction data according to the second attention score matrix formed by N second attention score vectors, so as to learn the relationship between transaction behavior information and label information. In this process, the prior knowledge fusion scheme is to fuse the feature vector corresponding to the prior feature and the first attention score vector to obtain the second attention score vector. It can be seen that the fusion scheme provided by the embodiments of this specification integrates prior knowledge into the attention mechanism and does not adopt an end-to-end high-order feature crossing method to obtain additional information. Therefore, prior knowledge will not be lost due to multi-layer deep crossing. Therefore, the embodiments of this specification provide a deep learning scheme suitable for tabular transaction data, reducing the loss of prior knowledge, making the relationship between transaction behavior information and label information in the finally learned tabular transaction data more accurate, and improving the performance of the neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of this specification. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 is a flowchart of the deep learning method for tabular transaction data in an embodiment of this specification;

[0020] Figure 2 is a structural block diagram of the deep learning device for tabular transaction data in an embodiment of this specification. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following describes the solutions provided by this specification with reference to the drawings.

[0022] As mentioned in the background art, the prior knowledge fusion scheme of the previous neural network models will cause some prior knowledge to be lost, which results in low learning ability of the neural network model for tabular transaction data and poor performance of the neural network model.

[0023] To this end, the embodiments of this specification provide a deep learning method for tabular transaction data. The method includes: extracting N prior features from the tabular transaction data, and converting each of the prior features into a corresponding feature vector; where N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information represents whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature; calculating a first attention score matrix corresponding to the tabular transaction data; where the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors represents the fusion proportion of a prior feature among the N prior features; fusing each of the feature vectors and the corresponding first attention score vectors to obtain a corresponding second attention score vector; and performing deep learning on the tabular transaction data according to the second attention score matrix formed by N second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0024] The following describes the specific implementation manners of the above concepts.

[0025] Figure 1 It is a flowchart of the deep learning method for tabular transaction data in an embodiment of the present invention. It can be understood that this method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. Refer to Figure 1 , the deep learning method for tabular transaction data includes the following steps S102 to S108:

[0026] S102. Extract N prior features from the tabular transaction data, and convert each of the prior features into a corresponding feature vector; where N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information represents whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature;

[0027] S104. Calculate a first attention score matrix corresponding to the tabular transaction data; where the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors represents the fusion proportion of a prior feature among the N prior features;

[0028] S106. Fuse each of the feature vectors and the corresponding first attention score vector to obtain a corresponding second attention score vector;

[0029] S108. Perform deep learning on the tabular transaction data according to the second attention score matrix formed by N second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0030] In Figure 1 In the deep learning method for tabular transaction data shown, first extract N prior features from the tabular transaction data, then calculate the first attention score matrix corresponding to the tabular transaction data, fuse the feature vector and the corresponding first attention score vector to obtain a corresponding second attention score vector, and then perform deep learning on the tabular transaction data according to the second attention score matrix formed by N second attention score vectors, so as to learn the relationship between the transaction behavior information and the label information. In this process, the prior knowledge fusion scheme is to fuse the feature vector corresponding to the prior feature and the first attention score vector to obtain the second attention score vector. It can be seen that the fusion scheme provided in the embodiments of this specification fuses prior knowledge into the attention mechanism and does not adopt an end-to-end high-order feature cross method to obtain additional information. Therefore, prior knowledge will not be lost due to multi-layer deep cross. Therefore, the embodiments of this specification provide a deep learning scheme suitable for tabular transaction data, reduce the loss of prior knowledge, make the relationship between the transaction behavior information and the label information in the finally learned tabular transaction data more accurate, and improve the performance of the neural network model.

[0031] The following describes Figure 1 the execution manners of each step.

[0032] S102. Extract N prior features from the tabular transaction data and convert each of the prior features into a corresponding feature vector; where N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information characterizes whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature;

[0033] Among them, the tabular transaction data includes at least one transaction data, and each transaction data includes transaction behavior information. For example, information such as the names of the two parties to the transaction, the transaction amount, and the transaction time. Each transaction data also includes whether the corresponding transaction behavior information has a transaction risk and the type of transaction risk when there is a transaction risk. For example, the label information in a set of transaction data is no transaction risk, and the label information in one transaction data is that there is a transaction risk and the transaction risk is gambling.

[0034] Among them, the N prior features include label features and at least one transaction behavior feature. It can be seen that the N prior features are important features or key features in a transaction data.

[0035] In one embodiment, the extracting N prior features from the tabular transaction data in S102 may specifically include: extracting N prior features from the tabular transaction data through an ensemble tree model.

[0036] Specifically, an appropriate ensemble tree model can be selected according to needs, and the tabular transaction data is input into the ensemble tree model. The ensemble tree model can then output N prior features, and these N prior features are important information in the tabular transaction data.

[0037] Here, compared with the LHUC (Learn Hidden Unit Contribution) model, the ensemble tree model can take into account both global and objective prior information.

[0038] First, by using the tree model, more global + objective prior information than LHUC can be obtained; improve the context of the native attention, and after secondary normalization, effectively fuse the prior information.

[0039] Among them, each of the prior features is transformed into a corresponding feature vector, that is, each prior feature is encoded in the form of a vector, and the corresponding prior feature is represented in the form of a feature vector. Since N prior features are extracted from the tabular transaction data, N encoded vectors can be obtained accordingly.

[0040] S104. Calculate the first attention score matrix corresponding to the tabular transaction data; among them, the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors represents the fusion proportion of a prior feature among the N prior features;

[0041] Among them, the first attention score matrix includes N first attention score vectors. One first attention score vector corresponds to one prior feature, and one first attention score vector reflects the fusion proportion of the corresponding prior feature among the N prior features, that is, it reflects the weight of one prior feature among the N prior features. Since one prior feature is represented by one encoding vector, the fusion proportion of one prior feature among the N prior features is also represented by one vector, and this vector is the above-mentioned first attention score vector.

[0042] In one embodiment, calculating the first attention score matrix corresponding to the tabular transaction data in S104 may include the following steps A1 to A3:

[0043] A1. Preprocess the tabular transaction data through two different data preprocessing methods to obtain a first vector and a second vector; wherein, both the first vector and the second vector are N-dimensional vectors;

[0044] That is to say, after preprocessing the tabular transaction data through one data preprocessing method, an N-dimensional vector is obtained, and this vector can be called the first vector. After preprocessing the tabular transaction data through another data preprocessing method, an N-dimensional vector is obtained, and this vector can be called the second vector.

[0045] Specifically, the tabular transaction data can be preprocessed according to the zero-mean processing method to obtain the first vector; and / or, the tabular transaction data can be preprocessed according to the equal-frequency binning processing method to obtain the second vector.

[0046] Among them, the zero-mean processing method means subtracting the average value of each data belonging to the same field from each data in the tabular transaction data to obtain a new value, and all the new values form a vector.

[0047] Among them, the equal-frequency binning processing method is a feature extraction method, which means sorting all the data in the same field of the tabular transaction data, dividing them into multiple intervals after sorting, and the number of values included in each interval is the same, and then extracting data from these intervals to form a vector.

[0048] It can be seen that the first vector and the second vector can reflect the data characteristics and distribution of the tabular transaction data.

[0049] A2. Expand the first vector into a first matrix and expand the second vector into a second matrix;

[0050] That is, expand the two vectors respectively to obtain two matrices.

[0051] Among them, there are many ways to expand, and one of them is introduced below:

[0052] Map each element in the first vector to M elements, so that the first vector is expanded into the first matrix with a dimension of N*M; where M is a positive integer greater than 1; and / or, map each element in the second vector to M elements, so that the second vector is expanded into the second matrix with a dimension of N*M.

[0053] That is to say, since there are N elements in the first vector, and each element is mapped to M elements, the first vector will be expanded into the first matrix with a dimension of N*M. Map each element in the second vector to M elements, so that the second vector will be expanded into the second matrix with a dimension of N*M.

[0054] In an actual scenario, when the first vector and the second vector are input into the underlying model structure, the width part of the underlying model structure will expand the first vector into the first matrix, and the depth part of the underlying model structure will expand the second vector into the second matrix. The way the width part and the depth part expand the vector is the above-mentioned way of mapping one element to M elements.

[0055] A3. Calculate the first attention score matrix according to the first matrix and the second matrix.

[0056] That is, after obtaining the first matrix and the second matrix, the first attention score matrix can be calculated using these two matrices.

[0057] Specifically, the first matrix and the second matrix can be multiplied in pairs to obtain the first attention score matrix with a dimension of N*M; where each first attention score vector includes M elements.

[0058] That is to say, multiply the elements in the same position of the first matrix and the second matrix to obtain the elements in the same position of the first attention score matrix. In this way, a first attention score matrix with a dimension of N*M can be obtained. Each row in the first attention score matrix is a first attention score vector, and each first attention score vector includes M elements.

[0059] It can be understood that after calculating the first attention score matrix, the prior features can be fused into the attention mechanism.

[0060] S106. Fuse each of the feature vectors and the corresponding first attention score vectors to obtain the corresponding second attention score vectors;

[0061] That is, the feature vector corresponding to a prior feature and the corresponding first attention score vector are fused to obtain an attention score vector incorporating the prior feature, which is referred to as the second attention score vector.

[0062] Specifically, each of the feature vectors and the corresponding first attention score vectors can be fused by element-wise multiplication; wherein, the number of elements in the feature vector and the first attention score vector is the same.

[0063] That is, the elements at the same positions in the feature vector corresponding to a prior feature and the corresponding first attention score vector are multiplied to obtain the elements at the same positions in the corresponding second attention score vector.

[0064] It can be seen that the embodiments of this specification do not fuse the prior features into the loss function, nor do they perform fusion through high-order feature crossing, so the loss of prior features can be reduced.

[0065] S108. According to the second attention score matrix formed by N second attention score vectors, perform deep learning on the tabular transaction data to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0066] That is, N second attention score vectors are formed into a matrix, which is referred to as the second attention score matrix. Then, the second attention score matrix is input into a neural network for deep learning, so as to learn the relationship between the transaction behavior information and the label information.

[0067] Among them, tabular transaction data refers to a numerical data table obtained based on statistical methods for describing transaction problems.

[0068] Among them, the attention mechanism is a technique in neural networks that mimics cognitive attention. This mechanism can enhance the weights of certain parts in the input data of the neural network while weakening the weights of other parts, thus making the focus of the neural network concentrate on the most important small part of the data. Which parts are more important than others depends on the context, and the attention mechanism can be trained by the gradient descent method.

[0069] Among them, knowledge fusion refers to knowledge fusion technology, which is mainly applied in the field of knowledge graphs. It refers to fusing the descriptive information about the same entity or concept from multiple data sources, integrating heterogeneous data and disambiguating the knowledge from different data sources under a unified specification, and is an important method for providing comprehensive knowledge sharing. Generally speaking, knowledge fusion includes steps such as concept fusion, entity alignment, attribute alignment, and attribute value fusion.

[0070] It can be seen that the embodiments of this specification embed the prior features obtained by the integrated tree model into the attention mechanism, enabling the prior features to play the role of a referee, guiding the neural network model to focus on other similar information in the data, and being able to fit the relationship between the transaction behavior information and the label information in the tabular features based on a relatively shallow neural network, thereby obtaining a model with stronger performance.

[0071] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain implementations, multitasking and parallel processing are also possible or may be advantageous.

[0072] In a second aspect, an embodiment of this specification provides a deep learning device for tabular transaction data. Refer to Figure 2 , the device 200 includes:

[0073] A feature extraction module 202, configured to extract N prior features from the tabular transaction data and convert each of the prior features into a corresponding feature vector; where N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information characterizes whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature;

[0074] A first calculation module 204, configured to calculate a first attention score matrix corresponding to the tabular transaction data; where the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors characterizes the fusion proportion of a prior feature among the N prior features;

[0075] A first fusion module 206, configured to fuse each of the feature vectors and the corresponding first attention score vector to obtain a corresponding second attention score vector;

[0076] A first learning module 208, configured to perform deep learning on the tabular transaction data according to a second attention score matrix formed by N second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data.

[0077] In one embodiment, the first calculation module includes:

[0078] A preprocessing unit for preprocessing the tabular transaction data through two different data preprocessing methods to obtain a first vector and a second vector; wherein, both the first vector and the second vector are N-dimensional vectors;

[0079] A first expansion unit for expanding the first vector into a first matrix and expanding the second vector into a second matrix;

[0080] A first calculation unit for calculating the first attention score matrix according to the first matrix and the second matrix.

[0081] Further, the preprocessing unit is specifically configured to: preprocess the tabular transaction data according to the zero-mean processing method to obtain the first vector; and / or, preprocess the tabular transaction data according to the equal-frequency binning processing method to obtain the second vector.

[0082] Further, the first expansion unit is specifically configured to: map each element in the first vector to M elements, so that the first vector is expanded into the first matrix with a dimension of N*M; wherein, M is a positive integer greater than 1; and / or, map each element in the second vector to M elements, so that the second vector is expanded into the second matrix with a dimension of N*M.

[0083] Further, the first calculation unit is specifically configured to: perform element-wise multiplication on the first matrix and the second matrix to obtain the first attention score matrix with a dimension of N*M; wherein, each first attention score vector includes M elements.

[0084] In one embodiment, the first fusion module is specifically configured to: fuse each feature vector and the corresponding first attention score vector by element-wise multiplication; wherein, the number of elements in the feature vector and the first attention score vector is the same.

[0085] In one embodiment, the feature extraction module is specifically configured to: extract N prior features from the tabular transaction data through an ensemble tree model.

[0086] It can be understood that the explanations, specific implementation manners, beneficial effects, examples, etc. of the relevant content in the device provided by the embodiments of the present invention can refer to the corresponding parts in the method provided in the first aspect, and will not be repeated here.

[0087] In a third aspect, an embodiment of this specification provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed on a computer, the computer is made to execute the method in any one of the embodiments in the specification.

[0088] Specifically, a system or device equipped with a storage medium can be provided. On this storage medium, software program code for implementing the functions of any one of the above-described embodiments is stored, and the computer (or CPU or MPU) of the system or device is caused to read and execute the program code stored in the storage medium.

[0089] In this case, the program code read from the storage medium itself can implement the functions of any one of the above-described embodiments. Therefore, the program code and the storage medium storing the program code constitute a part of the present invention.

[0090] Furthermore, it should be clear that not only can the functions of any one of the above-described embodiments be implemented by executing the program code read by the computer, but also by causing an operating system or the like operating on the computer based on the instructions of the program code to complete part or all of the actual operations.

[0091] In addition, it can be understood that the program code read from the storage medium is written into a memory provided in an expansion board inserted into the computer or into a memory provided in an expansion module connected to the computer, and then based on the instructions of the program code, a CPU or the like installed on the expansion board or the expansion module is caused to execute part and all of the actual operations, thereby implementing the functions of any one of the above-described embodiments.

[0092] It can be understood that for the explanations, specific implementation manners, beneficial effects, examples, etc. of the content in the computer-readable medium provided in the embodiments of the present invention, reference can be made to the corresponding parts in the method provided in the first aspect, and details are not described herein again.

[0093] Fourthly, an embodiment of this specification provides a computing device, including a memory and a processor. An executable code is stored in the memory. When the processor executes the executable code, the method in any one of the embodiments in the specification is implemented.

[0094] It can be understood that for the explanations, specific implementation manners, beneficial effects, examples, etc. of the content in the computing device provided in the embodiments of the present invention, reference can be made to the corresponding parts in the method provided in the first aspect, and details are not described herein again.

[0095] It can be understood that the structure schematically shown in the embodiments of this specification does not constitute a specific limitation on the devices in the embodiments of this specification. In other embodiments of the specification, the above-described devices may include more or fewer components than those shown in the figures, or combine certain components, or split certain components, or have different component arrangements. The components shown in the figures can be implemented in hardware, software, or a combination of software and hardware.

[0096] For the information interaction, execution process, etc. between the modules in the above-mentioned device and system, since they are based on the same concept as the method embodiments of this specification, the specific content can be referred to the description in the method embodiments of this specification, and will not be elaborated here.

[0097] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0098] Those skilled in the art should be able to realize that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, add-ons, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0099] The specific implementation manners described above further elaborate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only the specific implementation manners of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning method for tabular transaction data, comprising: extracting N prior features from the tabular transaction data and converting each of the prior features into a corresponding feature vector; wherein, N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, and each transaction data includes transaction behavior information and corresponding label information, the label information characterizing whether there is a transaction risk for the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature; calculating a first attention score matrix corresponding to the tabular transaction data; wherein, the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors characterizes the fusion proportion of a prior feature among the N prior features; fusing each of the feature vectors and the corresponding first attention score vector to obtain a corresponding second attention score vector; performing deep learning on the tabular transaction data according to a second attention score matrix formed by N second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data; wherein, the extracting N prior features from the tabular transaction data includes: extracting N prior features from the tabular transaction data through an ensemble tree model.

2. The method according to claim 1, wherein, The calculating the first attention score matrix corresponding to the tabular transaction data includes: preprocessing the tabular transaction data through two different data preprocessing methods respectively to obtain a first vector and a second vector; wherein, both the first vector and the second vector are N-dimensional vectors; expanding the first vector into a first matrix and expanding the second vector into a second matrix; calculating the first attention score matrix according to the first matrix and the second matrix.

3. The method according to claim 2, wherein The preprocessing the tabular transaction data through two different data preprocessing methods respectively to obtain a first vector and a second vector includes: preprocessing the tabular transaction data according to a zero-mean processing method to obtain the first vector; and / or, preprocessing the tabular transaction data according to an equal-frequency binning processing method to obtain the second vector.

4. The method according to claim 2, wherein The expanding the first vector into a first matrix and expanding the second vector into a second matrix includes: mapping each element in the first vector to M elements so that the first vector is expanded into the first matrix with a dimension of N*M; wherein, M is a positive integer greater than 1; and / or, mapping each element in the second vector to M elements so that the second vector is expanded into the second matrix with a dimension of N*M.

5. The method according to claim 4, wherein The calculating the first attention score matrix according to the first matrix and the second matrix includes: performing element-wise multiplication on the first matrix and the second matrix to obtain the first attention score matrix with a dimension of N*M; wherein, each of the first attention score vectors includes M elements.

6. The method according to claim 1, wherein Fusing each of the feature vectors with the corresponding first attention score vector includes: Fusing each of the feature vectors with the corresponding first attention score vector by element-wise multiplication; wherein, the number of elements in the feature vector and the first attention score vector is the same.

7. A deep learning device for tabular transaction data, comprising: A feature extraction module, configured to extract N prior features from the tabular transaction data and convert each of the prior features into a corresponding feature vector; wherein, N is a positive integer greater than 1, the tabular transaction data is tabular transaction data formed by at least one transaction data, each transaction data includes transaction behavior information and corresponding label information, the label information represents whether there is a transaction risk in the corresponding transaction behavior information and the type of transaction risk when there is a transaction risk, and the N prior features include label features and at least one transaction behavior feature; A first calculation module, configured to calculate a first attention score matrix corresponding to the tabular transaction data; wherein, the first attention score matrix is formed by N first attention score vectors, and each of the first attention score vectors represents the fusion proportion of a prior feature among the N prior features; A first fusion module, configured to fuse each of the feature vectors with the corresponding first attention score vector to obtain a corresponding second attention score vector; A first learning module, configured to perform deep learning on the tabular transaction data according to a second attention score matrix formed by N second attention score vectors to determine the relationship between the transaction behavior information and the label information in the tabular transaction data; Wherein, the feature extraction module is configured to: extract N prior features from the tabular transaction data through an ensemble tree model.

8. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed on a computer, the computer is made to execute the method according to any one of claims 1 to 6.

9. A computing device, comprising a memory and a processor, wherein an executable code is stored in the memory, and when the processor executes the executable code, the method according to any one of claims 1 to 6 is implemented.

Images