A behavior detection method and apparatus

Abstract

Embodiments of the present application provide a behavior detection method and device, the method comprising: obtaining a first credit score and N first sub-credit scores of a first user, a second credit score and N second sub-credit scores, the first credit score and the N first sub-credit scores being obtained by inputting M first feature domains into a credit score model, and the second credit score and the N second sub-credit scores being obtained by inputting M second feature domains into the credit score model. If the first credit score is different from the second credit score, S first sub-credit scores and S second sub-credit scores are determined according to the N first sub-credit scores and the N second sub-credit scores, and at least one target behavior information causing the first credit score to change relative to the second credit score is determined according to first feature data in a first feature domain corresponding to the S first sub-credit scores and second feature data in a second feature domain corresponding to the S second sub-credit scores. Through the above method, the user behavior causing the credit score to change can be accurately determined.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a behavior detection method and apparatus. Background Technology

[0002] A user's credit score is an indicator used to assess their creditworthiness. Financial institutions often use a user's credit score to provide corresponding financial services.

[0003] Currently, most financial companies only display credit scores to users. When a user's behavioral information changes, the displayed credit score also changes accordingly. However, the credit scores displayed by most financial companies lack interpretability; that is, the financial companies cannot determine which user behaviors led to the change in the credit score. Therefore, determining the user behaviors that caused the credit score change is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This application provides a behavior detection method and apparatus for determining user behaviors that lead to changes in credit scores.

[0005] In a first aspect, embodiments of this application provide a behavior detection method, which can be executed by a processor of a computer device. The method specifically includes the following steps: the processor acquires a first credit score and N first sub-credit scores of a first user, wherein the first credit score and the N first sub-credit scores are obtained by inputting M first feature domains of the first user acquired at a first moment into a credit score model; the first credit score is determined based on the N first sub-credit scores; each of the N first sub-credit scores corresponds to at least one of the M first feature domains; each first feature domain includes at least one first feature data; M is an integer greater than 1, and N is an integer greater than or equal to M; the credit score model is trained based on a first dataset, which includes at least one set of M preset feature domains labeled with credit scores. The processor can also acquire the first user's second credit score and N second sub-credit scores. The second credit score and N second sub-credit scores are obtained by inputting the first user's M second feature domains acquired at a second time step (before the first time step). The second credit score is determined based on the N second sub-credit scores, each of which corresponds to at least one of the M second feature domains. Each second feature domain includes at least one second feature data. If the first credit score differs from the second credit score, the processor compares the N first sub-credit scores and the N second sub-credit scores to determine S first sub-credit scores and S second sub-credit scores corresponding to them. The S first sub-credit scores and S second sub-credit scores are different, where S is a positive integer less than or equal to N. The processor then compares the first feature data in the first feature domains corresponding to the S first sub-credit scores with the second feature data in the second feature domains corresponding to the S second sub-credit scores to determine at least one target behavioral information that causes the change in the first credit score relative to the second credit score.

[0006] In the above method, when a user's credit score changes, the user behavior that caused the change in credit score can be accurately identified.

[0007] In one possible design, the processor acquires M first feature domains of the first user at the first moment, which may include: the processor acquires the first user data of the first user at the first moment, wherein the first user data includes multiple first feature data; the processor then divides the multiple first feature data according to a preset partitioning rule to obtain M first feature domains.

[0008] In one possible design, the M first feature domains of the first user are obtained by dividing the first user data of the first user acquired at the first moment based on a preset partitioning rule.

[0009] In one possible design, the credit score model may include a mapping module, a feature extraction module, and a credit score output module. The first credit score and N first sub-credit scores are obtained as follows: the processor first inputs at least one first feature data from M first feature domains into the mapping module to obtain the first feature domain vectors corresponding to the M first feature domains respectively. The processor then inputs the obtained M first feature domain vectors into the feature extraction module to obtain N first cross vectors. Finally, the processor inputs the N first cross vectors into the credit score output module to obtain the first credit score and N first sub-credit scores.

[0010] In one possible design, the mapping module may include an embedding layer and a first concatenation layer; the M first feature domain vectors are obtained as follows: For the j-th first feature domain among the M first feature domains, where j is an integer taking values ​​from [1, M], the following steps are performed: The processor inputs at least one first feature data from the j-th first feature domain into the embedding layer, determining that at least one first feature data in each first feature domain corresponds to a first feature vector. The processor then concatenates at least one first feature vector from the j-th first feature domain based on the first concatenation layer to obtain the first feature domain vector corresponding to the j-th first feature domain.

[0011] In one possible design, the feature extraction module may include a feature cross layer, L connection layers, and a second concatenation layer; where L is a non-negative integer; the N first cross vectors are obtained as follows: the processor inputs M first feature domain vectors into the feature cross layer to obtain N second cross vectors, the processor then inputs the N second cross vectors into the L connection layers to obtain N third cross vectors, and then the processor inputs the N third cross vectors and M first feature domain vectors into the second concatenation layer to obtain N first cross vectors.

[0012] In one possible design, the processor inputs M first feature domain vectors into a feature cross layer to obtain N second cross vectors. This can include: for the j-th first feature domain vector among the M first feature domain vectors, the processor performs a matrix dot product on the j-th first feature domain vector and the j-th first feature domain vector to obtain a first type vector corresponding to the j-th first feature domain vector, where j is an integer taking values ​​from [1, M]. The processor then determines at least one pair of related first feature domain vectors based on a preset cross rule, performs a matrix dot product on each pair of related first feature domain vectors, and obtains at least one second type vector. Finally, the first type vector corresponding to the j-th first feature domain vector and at least one second type vector are used as the N second cross vectors.

[0013] In one possible design, the processor inputs N third cross vectors and M first feature domain vectors into the second concatenation layer to obtain N first cross vectors. This may include: for the p-th third cross vector among the N third cross vectors, where p is an integer taking values ​​from [1, N], the processor performs the following steps: the processor determines at least one first feature domain vector corresponding to the p-th third cross vector as the target feature vector, and then concatenates the p-th third cross vector with at least one target feature domain vector to obtain the first cross vector corresponding to the p-th third cross vector.

[0014] In one possible design, the processor compares the first feature data in the first feature domain corresponding to S first sub-credit scores with the second feature data in the second feature domain corresponding to S second sub-credit scores to determine at least one target behavior information that causes a change in the first credit score relative to the second credit score. This can include: for the i-th first sub-credit score among the S first sub-credit scores and the i-th second sub-credit score among the S second sub-credit scores, where the i-th first sub-credit score and the i-th second sub-credit score correspond to each other, and i is an integer taking values ​​from [1, S], the processor performs the following steps: the processor determines at least one first comparison data, which is the first feature data in at least one first feature domain corresponding to the i-th first sub-credit score; the processor can also determine at least one second comparison data, which is the second feature data in at least one second feature domain corresponding to the i-th second sub-credit score; then, based on the at least one first comparison data and the at least one second comparison data, the processor determines at least one target behavior information that causes a change in the i-th first sub-credit score relative to the i-th second sub-credit score.

[0015] In one possible design, the above method may further include: outputting at least one target behavior information that causes a change in the first credit score relative to the second credit score.

[0016] In one possible design, the above method may also include at least one of the following: outputting the change of the first credit score relative to the second credit score, or outputting the first credit score.

[0017] In one possible design, the change of the first credit score relative to the second credit score includes at least one of the following: the amount of change of the first credit score relative to the second credit score, or the sign of the change of the first credit score relative to the second credit score.

[0018] Secondly, embodiments of this application also provide a behavior detection device, which includes units or means for performing the steps of the first aspect above.

[0019] Thirdly, embodiments of this application also provide a computer device, which includes a memory and a processor, wherein the memory is used to store programs and data, and the processor is used to run the programs stored in the memory and execute the method provided in the first aspect above based on the data stored in the memory.

[0020] Fourthly, embodiments of this application also provide a computer storage medium storing a computer program, which, when executed by a computer, causes the computer to perform the method provided in the first aspect above.

[0021] Fifthly, embodiments of this application also provide a chip for reading a computer program stored in a memory to implement the method described in the first aspect above.

[0022] Sixthly, embodiments of this application provide a chip system including a processor for supporting a computer device in implementing the functions involved in the first aspect above. In one possible design, the chip system further includes a memory for storing necessary programs and data of the computer device. This chip system may be composed of chips or may include chips and other discrete devices. Attached Figure Description

[0023] Figure 1 A schematic diagram of the structure of a computer device provided in an embodiment of this application;

[0024] Figure 2 A flowchart illustrating a behavior detection method provided in an embodiment of this application;

[0025] Figure 3 A flowchart illustrating a method for determining M first feature domains provided in an embodiment of this application;

[0026] Figure 4 This is a schematic diagram of the structure of a credit score model provided in an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of the structure of a credit score model provided in an embodiment of this application;

[0028] Figure 6 A flowchart illustrating a method for determining a first credit score and N first sub-credit scores, provided in an embodiment of this application;

[0029] Figure 7 A flowchart illustrating a method for determining N first cross vectors provided in an embodiment of this application;

[0030] Figure 8 A schematic diagram of a structure for displaying a first credit score is provided in an embodiment of this application;

[0031] Figure 9 A schematic diagram of a structure for displaying a first credit score is provided in an embodiment of this application;

[0032] Figure 10 A schematic diagram of a structure for displaying a first credit score is provided in an embodiment of this application;

[0033] Figure 11 This is a schematic diagram of the structure of a first model provided in an embodiment of this application;

[0034] Figure 12 This is a schematic diagram of the structure of a second model provided in an embodiment of this application;

[0035] Figure 13 This is a schematic diagram of the structure of a third model provided in an embodiment of this application;

[0036] Figure 14 This is a schematic diagram of the structure of a behavior detection device provided in an embodiment of this application;

[0037] Figure 15 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in further detail below with reference to the accompanying drawings. The specific operating methods and functional descriptions in the method embodiments can also be applied to the device embodiments or system embodiments.

[0039] To better explain the embodiments of this application, the relevant terms or technologies used in this application will be explained first:

[0040] I. Row Vectors

[0041] A row vector is a 1×p matrix, where p is any positive integer, for example: x = [x1 x2 … x p ].

[0042] II. Column Vectors

[0043] The column vector is a q×1 matrix, where q is any positive integer, for example:

[0044] III. Matrix Size / Scale

[0045] A p×q matrix is ​​a rectangular array consisting of p rows and q columns of elements. For example:

[0046]

[0047] In a matrix, each number that makes up the matrix is ​​called an element of the matrix. For example, A 11 A 12 A pq All of these are elements of matrix A. The index (or coordinate) of an element indicates its position within the matrix; it can be the row number (or row coordinate) or column number (or column coordinate). For example, A 11 The subscript "11" indicates that the element is located in the first row and first column of matrix A. 21 This indicates that the element is located in the second row and first column of matrix A. Additionally, the element's subscript can have different representations, such as A... 11 It can also be written as A 1，1 A 21 It can also be written as A 2，1 Similar details will not be repeated below.

[0048] It should be noted that the row number of the first row of a matrix is ​​not limited to 1, but can also be other values, such as 0. Similarly, the column number of the first column of a matrix is ​​not limited to 1, but can also be 0 or other data. For example, if the row number of the first row and the column number of the first column of the matrix above are 0 and 0 respectively, then the index of the first element in the matrix is ​​"00", which means that the element is in the 0th row and 0th column of the matrix.

[0049] IV. Matrix dot product (element-wise product, entry-wise product)

[0050] Matrix dot product refers to the element-wise multiplication of two matrices of the same size (i.e., the two matrices have the same number of rows and columns). For example, A and B are both p×q matrices. The matrix dot product of matrices A and B results in matrix C.

[0051]

[0052]

[0053] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0054] Figure 1 This diagram illustrates a possible computer device to which the behavior detection method provided in this application is applicable. (See also...) Figure 1 As shown, the computer device includes components such as a processor 110, a memory 120, a communication module 130, an input unit 140, a display unit 150, and a power supply 160. Those skilled in the art will understand that... Figure 1The structure of the computer device shown in the figure does not constitute a limitation on the computer device. The computer device provided in the embodiments of this application may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0055] The following is combined with Figure 1 A detailed introduction to each component of a computer device:

[0056] The communication module 130 can connect to other devices via wireless or physical connection to enable data transmission and reception by the computer device. Optionally, the communication module 130 may include any one or a combination of radio frequency (RF) circuit, wireless fidelity (WiFi) module, communication interface, Bluetooth module, etc., and this application embodiment does not limit this.

[0057] The memory 120 can be used to store program instructions and data. The processor 110 executes various functional applications and data processing of the computer device by running the program instructions stored in the memory 120. Among these program instructions are those that enable the processor 110 to execute the behavior detection method provided in the following embodiments of this application.

[0058] Optionally, the memory 120 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, various application programs, and program instructions; the data storage area may store user data and other data. Furthermore, the memory 110 may include high-speed random access memory, and may also include non-volatile memory, such as disk storage devices, flash memory devices, or other volatile solid-state storage devices.

[0059] The input unit 140 can be used to receive user input data or operation commands, etc. Optionally, the input unit 140 may include input devices such as a touch panel, function keys, physical keyboard, mouse, camera, or monitor.

[0060] The display unit 150 enables human-computer interaction, displaying user-inputted information and other content via a user interface. The display unit 150 may include a display panel 151. Optionally, the display panel 151 may be configured as a liquid crystal display (LCD), an organic light-emitting diode (OLED), or similar device.

[0061] Furthermore, when the input unit includes a touch panel, the touch panel may cover the display panel 151. When the touch panel detects a touch event on or near it, it transmits the information to the processor 110 to determine the type of touch event and perform the corresponding operation.

[0062] The processor 110 is the control center of the computer device, connecting the various components via various interfaces and lines. The processor 110 can execute program instructions stored in the memory 120 and access data stored in the memory 120 to perform various functions of the computer device and implement the behavior detection method provided in this application embodiment.

[0063] Optionally, the processor 110 may include one or more processing units. Specifically, the processing unit may include hardware devices such as a CPU and / or a GPU capable of floating-point operations. This processing unit is able to determine and output the user's credit score. When the processor 110 of the computer device implements the behavior detection method, the processing unit reads the user's user data from the storage data area of ​​the memory 120 and outputs the user's credit score based on the user's user data.

[0064] The computer device also includes a power supply 160 (such as a battery) for supplying power to the various components. Optionally, the power supply 160 can be logically connected to the processor 110 through a power management system, thereby enabling the power management system to perform functions such as charging and discharging of the computer device.

[0065] Although not shown, the computer device may also include components such as a camera, sensors, and audio acquisition devices, which will not be described in detail here.

[0066] This application provides a behavior detection method, which can be derived from... Figure 1 The processor in the computer device shown performs the operation. See also... Figure 2 As shown, the process of this method includes:

[0067] S201, the processor obtains the first user's first credit score and N first sub-credit scores.

[0068] In this embodiment, the first credit score and N first sub-credit scores are obtained by inputting the M first feature domains of the first user acquired at the first time moment into the credit score model, where M is an integer greater than 1 and N is an integer greater than or equal to M. The first time moment can be the current time moment or any time moment before the current time moment, and is not limited here.

[0069] This application provides a method for determining M first feature domains of a first user, which may include, for example: Figure 3 The following steps are shown:

[0070] S301, the processor acquires the first user data of the first user at the first moment, wherein the first user data includes multiple first feature data.

[0071] In this embodiment, each user corresponds to multiple identical user features, and each user feature corresponds to a feature data. User features can be used to describe a user's identity information, behavioral information, etc. For example, user features can be a user's name, age, gender, salary, company, savings amount, total loan amount, amount repaid, etc., and are not limited here.

[0072] For any given user characteristic, at different times, the feature data corresponding to that user characteristic for the same user can be the same or different. At the same time, the feature data corresponding to that user characteristic for different users can be the same or different.

[0073] To facilitate user behavior detection, user feature data can be stored in a database or other storage media. When changes in user behavior cause corresponding changes in user feature data, the user feature data stored in the database is also updated accordingly.

[0074] With user feature data stored in a database, the processor can access the database at the first moment and retrieve multiple first feature data of the first user from the database.

[0075] For example, Table 1 shows the first user "Zhang San" and his first feature data obtained by the processor when accessing the database at the first moment.

[0076] Table 1.

[0077] Name age gender Salary Company of employment Deposit amount Total loan amount Amount already repaid Zhang San 25 male 20 Company xx 0 300 100

[0078] In Table 1, the first user "Zhang San" has a deposit amount of 0. After "Zhang San" deposits 200 into the bank, the deposit amount stored in the database is 200, as shown in Table 2.

[0079] Table 2.

[0080] Name age gender Salary Company of employment Deposit amount Total loan amount Amount already repaid Zhang San 25 male 20 Company xx 200 300 100

[0081] S302, the processor divides multiple first feature data according to a preset partitioning rule to obtain M first feature domains.

[0082] In the embodiments of this application, the preset partitioning rule can be determined based on human experience, or it can be determined by other methods, which are not limited here. The preset partitioning rule can classify feature data with similar attribute information into the same feature domain.

[0083] For example, primary characteristic data falling within the scope of personal information can be grouped into the same primary characteristic domain. For instance, name (“Zhang San”), age (“25”), and gender (“Male”) can be grouped into the same primary characteristic domain. Primary characteristic data falling within the scope of employment can be grouped into the same primary characteristic domain. For instance, salary (“20”) and company (“xx Company”) can be grouped into the same primary characteristic domain. Primary characteristic data falling within the scope of funds can be grouped into the same primary characteristic domain. For instance, deposit amount (“200”), total loan amount (“300”), and repaid amount (“100”) can be grouped into the same primary characteristic domain.

[0084] It should be understood that a single first feature data point can be divided into one first feature domain or multiple first feature domains. Each first feature domain includes at least one first feature data point.

[0085] The above method provides a method for determining M first feature domains, such that the obtained first feature domains may include feature data with similar attribute information.

[0086] In this embodiment, the credit score model can be trained based on a first dataset, wherein the first dataset includes at least one set of M pre-defined feature domains labeled with credit scores. How to determine the structure of the credit score model and how to train the credit score model using the first dataset will be explained later and will not be described in detail here.

[0087] The method for obtaining M preset feature domains is similar to the method for obtaining M first feature domains, as follows:

[0088] For any set of M preset feature domains with credit score labels in the first dataset, the M preset feature domains can be obtained through the following steps: The processor acquires preset user data for any preset user at a third time point. This preset user data includes multiple preset feature data. The processor divides the multiple preset feature data based on preset partitioning rules to obtain M preset feature domains. The third time point can be any time point before the first time point.

[0089] In this embodiment of the application, the structure of the credit score model is as follows: Figure 4As shown, the system includes a mapping module, a feature extraction module, and a credit score output module. The mapping module includes an embedding layer and a first concatenation layer. The feature extraction module includes a feature cross-layer, L connection layers, and a second concatenation layer, where L is a non-negative integer. The credit score output module includes P connection layers and a fully connected layer, where P is a non-negative integer.

[0090] To facilitate a detailed description of the structure of the credit scoring model, Figure 5 A detailed example is given in conjunction with the embodiment. The first user corresponds to three first feature domains, namely first feature domain 1, first feature domain 2 and first feature domain 3. Among them, first feature domain 1 includes name ("Zhang San"), age ("25"), and gender ("male"); first feature domain 2 includes salary ("20") and company ("xx company"); first feature domain 3 includes deposit amount ("200"), total loan amount ("300"), and repaid amount ("100"). The three first feature domains are input into the credit score model, and the credit score model can output the first user's first credit score and four first sub-credit scores.

[0091] refer to Figure 4 or Figure 5 The credit score model structure in this application provides a method for obtaining a first user's first credit score and N first sub-credit scores, which may include... Figure 6 The following steps are shown.

[0092] S601, the processor inputs at least one first feature data from the M first feature domains to the mapping module to obtain the first feature domain vectors corresponding to the M first feature domains respectively.

[0093] For the j-th first feature field among M first feature fields, where j is an integer taking values ​​from [1, M], the following steps are performed: The processor first inputs at least one first feature data from the j-th first feature field into the embedding layer to determine that each of the at least one first feature data in the j-th first feature field corresponds to a first feature vector. Then, based on the first concatenation layer, the processor concatenates at least one first feature vector from the j-th first feature field to obtain the first feature field vector corresponding to the j-th first feature field. The concatenation order of the at least one first feature vector from the j-th first feature field is not limited.

[0094] It should be understood that since the embedding layer and the first concatenation layer do not have trainable network parameters, each of the M first feature domains can use the same embedding layer and the same first concatenation layer. The embedding layer can be a trained word2vec, etc., and is not limited here.

[0095] In this embodiment, the dimensions of the first feature vectors corresponding to different first feature data are the same. When different first feature domains include the same number of first feature data, the dimensions of the first feature domain vectors corresponding to different first feature domains are the same. When different first feature domains include different numbers of first feature data, the dimensions of the first feature domain vectors corresponding to different first feature domains are different.

[0096] For example, in Figure 5 In the first feature domain 1, three first feature data are included: name ("Zhang San"), age ("25"), and gender ("male"). The first feature domain 2 includes two first feature data: salary ("20") and company ("xx company"). The first feature domain 3 includes three first feature data: deposit amount ("200"), total loan amount ("300"), and repaid amount ("100"). Inputting the first feature domains 1, 2, and 3 into the mapping module yields three first feature domain vectors: first feature domain vector 1 corresponding to first feature domain 1, first feature domain vector 2 corresponding to first feature domain 2, and first feature domain vector 3 corresponding to first feature domain 3.

[0097] If the dimension of the first feature vector corresponding to each first feature data is 16, then the dimension of the first feature domain vector 1 is 48 because the first feature domain 1 includes 3 first feature data; the dimension of the first feature domain vector 2 is 32 because the first feature domain 2 includes 2 first feature data; and the dimension of the first feature domain vector 3 is 48 because the first feature domain 3 includes 3 first feature data.

[0098] S602, the processor inputs the obtained M first feature domain vectors into the feature extraction module to obtain N first cross vectors.

[0099] This application provides a method for obtaining N first cross vectors, which may include, for example: Figure 7 The following steps are shown:

[0100] S701, the processor inputs M first feature domain vectors into the feature cross layer to obtain N second cross vectors.

[0101] This application provides a method for obtaining N second cross vectors, which may include the following steps: For the j-th first feature domain vector among M first feature domain vectors, the processor performs a matrix dot product on the j-th first feature domain vector and the j-th first feature domain vector to obtain a first type vector corresponding to the j-th first feature domain vector. The processor then determines at least one pair of related first feature domain vectors based on a preset cross rule, and performs a matrix dot product on each pair of related first feature domain vectors to obtain at least one second type vector. Finally, the processor uses the first type vector corresponding to the j-th first feature domain vector and at least one second type vector as N second cross vectors.

[0102] In this embodiment, the preset cross-rule is determined based on a first dataset. The preset cross-rule can be determined based on multiple sets of M preset feature domains with credit score labels in the first dataset, or it can be determined based on a single set of M preset feature domains with credit score labels in the first dataset. The preset cross-rule includes at least one set of identifiers for two related preset feature domains. The processor can obtain at least one set of identifiers for two related preset feature domains from the preset cross-rule, and determine at least one set of two related first feature domain vectors corresponding to the identifiers of the at least one set of two related preset feature domains. The processor then performs matrix multiplication on each set of two related first feature domain vectors to obtain at least one second-type vector.

[0103] For example, the preset cross rules include <preset feature domain 2, preset feature domain 3>, in Figure 5 In the process, after the processor inputs three first feature domain vectors (first feature domain vector 1, first feature domain vector 2, and first feature domain vector 3) into the feature cross layer, the processor performs a matrix dot product on first feature domain vector 1 to obtain second cross vector 1. The processor then performs a matrix dot product on first feature domain vector 2 to obtain second cross vector 2. The processor then performs a matrix dot product on first feature domain vector 3 to obtain second cross vector 3. Next, the processor obtains the identifiers of two associated preset feature domains (preset feature domain 2 and preset feature domain 3) from a preset cross rule, and determines the two associated first feature domain vectors (first feature domain vector 2 and first feature domain vector 3) based on these identifiers. The processor then performs a matrix dot product on first feature domain vector 2 and first feature domain vector 3 to obtain second cross vector 4. Therefore, the processor inputs the three first feature domain vectors... Figure 5 The feature cross layer can obtain four second cross vectors, namely second cross vector 1, second cross vector 2, second cross vector 3 and second cross vector 4.

[0104] In this embodiment, when performing matrix dot product on two associated first feature domain vectors, if the two associated first feature domain vectors have the same dimension, they can be directly multiplied. If the two associated first feature domain vectors have different dimensions, a first preset length is extracted from the two associated first feature domain vectors before performing the matrix dot product. Generally, the first preset length can be determined based on the smaller dimension first feature domain vector among the two associated first feature domain vectors.

[0105] For example, the dimension of the first feature vector 2 is 32, and the dimension of the first feature vector 3 is 48. When the processor performs matrix multiplication on the first feature vector 2 and the first feature vector 3, since the dimensions of the first feature vector 2 and the first feature vector 3 are different, the dimension of the first feature vector 2 (32) is smaller than the dimension of the first feature vector 3 (48). Therefore, the processor can set the first preset length to 32. The processor can extract 32 dimensions from the first feature vector 3 as a temporary feature vector, and then perform matrix multiplication on the first feature vector 2 and the temporary feature vector to obtain the second cross vector 4.

[0106] It should be understood that when truncating the first feature domain vector to a first preset length, the first preset length can be truncated from the first half of the first feature domain vector as a temporary feature domain vector, or the first preset length can be truncated from the second half of the first feature domain vector as a temporary feature domain vector, or the first preset length can be randomly truncated from the first feature domain vector as a temporary feature domain vector, without any limitation here.

[0107] In the above method, each first feature domain vector has a strong correlation with itself. In addition, according to the preset cross rule, two related first feature domain vectors can be determined. The feature cross layer performs matrix dot product between the first feature domain vector and itself, and performs matrix dot product between the two related first feature domain vectors, which can obtain more and more refined features between each first feature domain vector.

[0108] The following describes how to determine the preset crossover rules based on a set of M preset feature domains labeled with credit scores in the first dataset.

[0109] The processor can obtain any set of M preset feature domains with credit score labels from the first dataset, and obtain the r-th preset feature domain and the s-th preset feature domain from the M preset feature domains, where r is an integer that takes all integers in the range [1, M] and s is an integer that takes all integers in the range [1, M]. r and s are not the same. The r-th preset feature domain is taken as the first preset feature domain and the s-th preset feature domain is taken as the second preset feature domain. The processor calculates the first similarity between the first preset feature domain and the second preset feature domain. If the first similarity is greater than the first similarity value, the association between the first preset feature domain and the second preset feature domain is added to the preset cross rule. If the first similarity is less than or equal to the first similarity value, no processing is performed.

[0110] For example, the first dataset includes a set of three preset feature domains labeled with credit scores: preset feature domain 1, preset feature domain 2, and preset feature domain 3. The processor can calculate the first similarity between preset feature domain 1 and preset feature domain 2, the first similarity between preset feature domain 1 and preset feature domain 3, and the first similarity between preset feature domain 2 and preset feature domain 3. If the first similarity between preset feature domain 1 and preset feature domain 2 is less than the first similarity value, no processing is performed; if the similarity between preset feature domain 1 and preset feature domain 3 is less than the first similarity value, no processing is performed; if the similarity between preset feature domain 2 and preset feature domain 3 is greater than the first similarity value, the association between preset feature domain 2 and preset feature domain 3 is added to the preset cross-rule.

[0111] Based on the above calculations, the preset cross rules can include the following:

[0112] <Preset feature domain 2, preset feature domain 3>

[0113] The method for determining the preset cross rules based on multiple sets of M preset feature domains with credit score labels in the first dataset is similar to the method for determining the preset cross rules based on a set of M preset feature domains with credit score labels in the first dataset, and will not be described in detail here.

[0114] In this embodiment, the first similarity between the first preset feature domain and the second preset feature domain can be determined based on the calculation formula of cosine similarity. The processor can first determine the first preset feature domain vector A corresponding to the first preset feature domain and the second preset feature domain vector B corresponding to the second preset feature domain based on the embedding layer and the first splicing layer, wherein the dimension of the first preset feature domain vector A and the second preset feature domain vector B are both n. The first similarity between the first preset feature domain and the second preset feature domain is determined by calculating the similarity between the first preset feature domain vector A and the second preset feature domain vector B. The similarity between the first preset feature domain vector A and the second preset feature domain vector B can be determined using formula (1):

[0115] Where cos(A,B) represents the similarity between the first preset feature domain vector A and the second preset feature domain vector B, a k b represents the k-th dimension feature in the first preset feature domain vector A. k This represents the k-th dimension feature in the second preset feature domain vector B.

[0116] It should be understood that other methods can also be used to determine the first similarity between the first preset feature domain and the second preset feature domain, which are not limited here.

[0117] In this embodiment, since the feature cross layer does not have trainable network parameters, the M first feature domain vectors can use the same feature cross layer.

[0118] In S702, the processor inputs N second cross vectors to L connection layers to obtain N third cross vectors.

[0119] In this embodiment, since the connection layers in the feature extraction module have trainable network parameters, the N second cross vectors do not share the same connection layer. Therefore, each second cross vector corresponds to L connection layers, and the network parameters in the L connection layers corresponding to different second cross vectors are not the same.

[0120] S703, the processor inputs N third cross vectors and M first feature domain vectors into the second concatenation layer to obtain N first cross vectors.

[0121] In this embodiment of the application, for the p-th third cross vector among N third cross vectors, where p is an integer taking values ​​from [1, N], the following steps are performed: the processor determines at least one first feature domain vector corresponding to the p-th third cross vector as the target feature vector, and concatenates the p-th third cross vector with at least one target feature domain vector to obtain the first cross vector corresponding to the p-th third cross vector.

[0122] For example, in Figure 5 In the process, the third cross vector 1 is only related to the first feature domain vector 1. Therefore, the target feature vector corresponding to the third cross vector 1 is the first feature domain vector 1. The processor concatenates the third cross vector 1 and the first feature domain vector 1 to obtain the first cross vector 1 corresponding to the third cross vector 1.

[0123] The third cross vector 2 is only related to the first feature domain vector 2. Therefore, the target feature vector corresponding to the third cross vector 2 is the first feature domain vector 2. The processor concatenates the third cross vector 2 and the first feature domain vector 2 to obtain the first cross vector 2 corresponding to the third cross vector 2.

[0124] The third cross vector 3 is only related to the first feature domain vector 3. Therefore, the target feature vector corresponding to the third cross vector 3 is the first feature domain vector 3. The processor concatenates the third cross vector 3 and the first feature domain vector 3 to obtain the first cross vector 3 corresponding to the third cross vector 3.

[0125] The third cross vector 4 is related to the first feature domain vector 2 and the first feature domain vector 3. Therefore, the target feature vectors corresponding to the third cross vector 4 are the first feature domain vector 2 and the first feature domain vector 3. The processor concatenates the third cross vector 4, the first feature domain vector 2 and the first feature domain vector 3 to obtain the first cross vector 4 corresponding to the third cross vector 4.

[0126] In this embodiment, since there are no trainable network parameters in the second splicing layer, the N third cross vectors can use the same second splicing layer.

[0127] S603, the processor inputs N first cross vectors to the credit score output module to obtain the first credit score and N first sub-credit scores.

[0128] In this embodiment, the processor inputs N first cross vectors to P connection layers to obtain N reference scores, and then inputs the N reference scores to a fully connected layer to obtain a first credit score and N first sub-credit scores. The first credit score is determined based on the N first sub-credit scores; that is, the sum of the N first sub-credit scores can be used as the first credit score.

[0129] In one possible implementation, the fully connected layer includes N fully connected layer parameters, namely fully connected layer parameter 1, ..., fully connected layer parameter N. After inputting N reference scores into the fully connected layer, each reference score is multiplied by its corresponding fully connected layer parameter to obtain N first sub-credit scores. The N first sub-credit scores are then added together to obtain the first credit score. The N fully connected layer parameters are all determined after training.

[0130] For example, in Figure 5 In the process, reference score 1 is multiplied by fully connected layer parameter 1 to obtain the first sub-credit score 1, reference score 2 is multiplied by fully connected layer parameter 2 to obtain the first sub-credit score 2, reference score 3 is multiplied by fully connected layer parameter 3 to obtain the first sub-credit score 3, and reference score 4 is multiplied by fully connected layer parameter 4 to obtain the first sub-credit score 4. The first sub-credit score 1, the first sub-credit score 2, the first sub-credit score 3 and the first sub-credit score 4 are added together to obtain the first credit score.

[0131] In this embodiment, since the connection layers in the output module have trainable network parameters, the N first cross vectors do not share the same connection layer. Therefore, each first cross vector corresponds to P connection layers, and the network parameters in the P connection layers corresponding to different first cross vectors are not the same. Furthermore, the N reference scores share the same connection layer.

[0132] In this embodiment, each of the N first sub-credit scores corresponds to at least one of the M first feature domains, meaning each first sub-credit score is determined by at least one first feature domain. To facilitate determining the correspondence between each first sub-credit score and at least one first feature domain, this correspondence can be stored in a storage area.

[0133] by Figure 5 For example, the correspondence between the first sub-credit score and the first feature domain stored in the storage area is as follows:

[0134] <First Sub-Credit Score 1: First Feature Domain 1>

[0135] <First Sub-Credit Score 2: First Feature Domain 2>

[0136] <First Sub-Credit Score 3: First Feature Domain 3>

[0137] <First Sub-Credit Score 4: First Feature Domain 2, First Feature Domain 3>

[0138] The storage area can also store the correspondence between the first feature domain and the first feature data, as follows:

[0139] <First Feature Domain 1: Name (“Zhang San”), Age (“25”), Gender (“Male”)>

[0140] <First Feature Domain 2: Salary (“20”), Company (“xx Company”)>

[0141] <First Feature Domain 3: Deposit Amount (“200”), Total Loan Amount (“300”), Amount Repaid (“100”)>

[0142] In the above method, since the storage area stores the correspondence between the first sub-credit score and the first feature domain, as well as the correspondence between the first feature domain and the first feature data, when the first sub-credit score changes, the first feature domain that causes the change in the first sub-credit score can be conveniently determined according to the correspondence between the first sub-credit score and the first feature domain, and then the first feature data that causes the change in the first sub-credit score can be determined according to the correspondence between the first feature domain and the first feature data.

[0143] S202, the processor obtains the first user's second credit score and N second sub-credit scores.

[0144] In this embodiment, the second credit score and N second sub-credit scores are obtained by inputting the M second feature domains of the first user acquired at the second time point into the credit score model. The second credit score is determined based on the N second sub-credit scores, and each of the N second sub-credit scores corresponds to at least one of the M second feature domains. Each second feature domain includes at least one second feature data. The second time point can be any time point prior to the first time point, and is not limited thereto.

[0145] The method for determining the second credit score and N second sub-credit scores of the first user through the credit scoring model is the same as the method for determining the first credit score and N first sub-credit scores of the first user through the credit scoring model, and will not be described in detail here.

[0146] If the first user data of the first user obtained by the processor at the first moment is the same as the first user data of the first user obtained by the processor at the second moment, then the first user's first credit score and second credit score are the same, and the N first sub-credit scores and N second sub-credit scores are also the same.

[0147] S203, if the first credit score and the second credit score are different, the processor compares the N first sub-credit scores and the N second sub-credit scores to determine the S first sub-credit scores and the S second sub-credit scores corresponding to the S first sub-credit scores, where the S first sub-credit scores and the S second sub-credit scores are different, and S is a positive integer less than or equal to N.

[0148] In one possible implementation, if the first credit score and the second credit score are the same, the processor can directly output the first credit score, and can also output the change of the first credit score relative to the second credit score. If the processor is located in the terminal device, the first credit score and the change of the first credit score relative to the second credit score can be directly displayed on the terminal device's display interface.

[0149] In this application embodiment, the change of the first credit score relative to the second credit score includes at least one of the following: the amount of change of the first credit score relative to the second credit score, or the sign of the change of the first credit score relative to the second credit score, wherein the sign of the change includes an upward sign or a downward sign.

[0150] For example, if the first moment is the current moment and the second moment is one day ago, the processor can obtain the first user's second credit score of 720 one day ago. The processor can also obtain the first user's first credit score of 720 at the current moment. After obtaining the first credit score, the processor determines that the first credit score and the second credit score are the same, that is, the change in the first credit score relative to the second credit score is 0. Figure 8 As shown, the processor can directly output the first credit score of 720, along with an upward sign, and display the change in the first credit score relative to the second credit score as 0 points near the upward sign. The processor can also output an explanatory description, namely, "The credit score has increased by 0 points compared to 1 day ago."

[0151] S204, the processor compares the first feature data in the first feature domain corresponding to the S first sub-credit scores with the second feature data in the second feature domain corresponding to the S second sub-credit scores, and determines at least one target behavior information that causes the first credit score to change relative to the second credit score.

[0152] In this embodiment of the application, after determining at least one target behavior information that causes a change in the first credit score relative to the second credit score, the processor can output at least one target behavior information that causes a change in the first credit score relative to the second credit score, and can also output the first credit score and the change in the first credit score relative to the second credit score.

[0153] In this embodiment, for the i-th first sub-credit score among S first sub-credit scores and the i-th second sub-credit score among S second sub-credit scores, where the i-th first sub-credit score and the i-th second sub-credit score correspond, and i is an integer taking values ​​from [1, S], the following steps are performed: The processor determines at least one first comparison data, wherein the at least one first comparison data is first feature data in at least one first feature domain corresponding to the i-th first sub-credit score. The processor can also determine at least one second comparison data, wherein the at least one second comparison data is second feature data in at least one second feature domain corresponding to the i-th second sub-credit score. Finally, based on the at least one first comparison data and the at least one second comparison data, the processor determines at least one target behavior information that causes the i-th first sub-credit score to change relative to the i-th second sub-credit score.

[0154] Furthermore, after determining at least one target behavior piece of information that causes a change in the i-th first sub-credit score relative to the i-th second sub-credit score, the processor can also determine the credit score corresponding to each target behavior piece of information, which can be determined in the following way:

[0155] If the change in the i-th first sub-credit score relative to the i-th second sub-credit score is caused by a single target behavior information, then the difference between the i-th first sub-credit score and the i-th second sub-credit score can be directly used as the credit score corresponding to that target behavior information. If the change in the i-th first sub-credit score relative to the i-th second sub-credit score is caused by W target behavior information, where W is an integer greater than 1, then the difference between the i-th first sub-credit score and the i-th second sub-credit score is used as a first reference value, and the result of dividing the first reference value by W is used as the credit score corresponding to each of the W target behavior information.

[0156] In the embodiments of this application, i can be an integer that takes the values ​​of [1, S]. As the value of i changes, at least one target behavior information can be determined, and the credit score corresponding to each target behavior information can also be determined.

[0157] When the value of i is different, the determined target behavior information can be the same or different. If there is one target behavior information corresponding to one credit score, then that credit score is directly used as the total credit score of that target behavior information. If there are multiple credit scores corresponding to one target behavior information, then the sum of the multiple credit scores is used as the total credit score of that target behavior information.

[0158] For example, if the first moment is the current moment and the second moment is one day ago, the processor can obtain the first user's first credit score and four first sub-credit scores at the current moment, as shown in Table 3.

[0159] Table 3.

[0160] First Credit Score First Sub-Credit Score 1 First Sub-Credit Score 2 First child credit score 3 First child credit score 4 740 360 200 60 120

[0161] The processor can obtain the first user's second credit score and four second sub-credit scores from the previous day, as shown in Table 4.

[0162] Table 4.

[0163] Second Credit Score Second Sub-Credit Score 1 Second child credit score 2 Second child credit score 3 Second child credit score 4 720 360 200 50 110

[0164] Since the first credit score 740 is different from the second credit score 720, the four first sub-credit scores in Table 3 are compared with the four second sub-credit scores in Table 4 to determine two different first sub-credit scores (i.e., first sub-credit score 3 and first sub-credit score 4) and two different second sub-credit scores (i.e., second sub-credit score 3 and second sub-credit score 4).

[0165] The correspondence between the first sub-credit score stored in the storage area and the first feature domain is set as follows:

[0166] <First Sub-Credit Score 1, First Feature Domain 1>

[0167] <First Sub-Credit Score 2, First Feature Domain 2>

[0168] <First Sub-Credit Score 3, First Feature Domain 3>

[0169] <First Sub-Credit Score 4, First Feature Domain 2, First Feature Domain 3>

[0170] The correspondence between the first feature domain and the first feature data stored in the storage area is as follows:

[0171] <First Feature Domain 1: Name (“Zhang San”), Age (“25”), Gender (“Male”)>

[0172] <First Feature Domain 2: Salary (“20”), Company (“xx Company”)>

[0173] <First Feature Domain 3: Deposit Amount (“200”), Total Loan Amount (“300”), Amount Repaid (“100”)>

[0174] The correspondence between the second sub-credit score and the second feature domain stored in the storage area is set as follows:

[0175] <Second Sub-Credit Score 1, Second Feature Domain 1>

[0176] <Second Sub-Credit Score 2, Second Feature Domain 2>

[0177] <Second Sub-Credit Score 3, Second Feature Domain 3>

[0178] <Second Sub-Credit Score 4, Second Feature Domain 2, Second Feature Domain 3>

[0179] The correspondence between the second feature domain and the second feature data stored in the storage area is as follows:

[0180] <Second Feature Domain 1: Name (“Zhang San”), Age (“25”), Gender (“Male”)>

[0181] <Second Feature Domain 2: Salary (“20”), Company (“xx Company”)>

[0182] <Second Feature Field 3: Deposit Amount (“400”), Total Loan Amount (“300”), Amount Repaid (“100”)>

[0183] The processor can determine the correspondence between the first sub-credit score 3 and the first feature domain 3 based on the correspondence between the first sub-credit score and the first feature domain. It can also determine the first feature data corresponding to the first feature domain 3, including deposit amount (“200”), total loan amount (“300”), and repaid amount (“100”), based on the correspondence between the first feature domain and the first feature data. Therefore, the three first comparison data points determined by the processor include deposit amount (“200”), total loan amount (“300”), and repaid amount (“100”).

[0184] The processor can also determine the correspondence between the second sub-credit score 3 and the second feature domain 3 based on the correspondence between the second sub-credit score and the second feature domain. Based on the correspondence between the second feature domain and the second feature data, the processor can determine that the corresponding second feature data for the second feature domain 3 includes deposit amount (“400”), total loan amount (“300”), and repaid amount (“100”). Therefore, the three second comparison data points determined by the processor include deposit amount (“400”), total loan amount (“300”), and repaid amount (“100”).

[0185] Based on the three first comparison data and three second comparison data, it can be determined that the deposit amount has changed. Therefore, the target behavioral information that caused the change in the first sub-credit score 3 relative to the second sub-credit score 3 can be identified, namely, "the deposit amount increased from 200 to 400, and the deposit amount increased by 200".

[0186] Since the change in the first sub-credit score 3 relative to the second sub-credit score 3 is caused by a target behavior information, the difference between the first sub-credit score 3 and the second sub-credit score 3 can be directly used as the credit score corresponding to the target behavior information "the deposit amount increased from 200 to 400, and the deposit amount increased by 200". That is, the credit score corresponding to the target behavior information "the deposit amount increased from 200 to 400, and the deposit amount increased by 200" is 10 (i.e., 60-50=10).

[0187] The processor can determine the correspondence between the first sub-credit score 4 and the first feature domains 2 and 3 based on the correspondence between the first feature domain and the first feature data. It can also determine the first feature data corresponding to first feature domain 2, including salary (“20”) and company (“xx company”), and the first feature data corresponding to first feature domain 3, including deposit amount (“200”), total loan amount (“300”), and repaid amount (“100”). Therefore, the processor can determine five first comparison data items, including salary (“20”), company (“xx company”), deposit amount (“200”), total loan amount (“300”), and repaid amount (“100”).

[0188] The processor can determine the correspondence between the second sub-credit score 4 and the second feature domains 2 and 3 based on the correspondence between the second feature domain and the second feature data. It can also determine the second feature data corresponding to second feature domain 2, including salary (“20”) and company (“xx company”), and the second feature data corresponding to second feature domain 3, including deposit amount (“400”), total loan amount (“300”), and repaid amount (“100”). Therefore, the processor can determine five second sets of data to be compared, including salary (“20”), company (“xx company”), deposit amount (“400”), total loan amount (“300”), and repaid amount (“100”).

[0189] Based on the five first comparison data points and five second comparison data points identified, it can be determined that the deposit amount has changed. Therefore, the target behavioral information that caused the change in the first sub-credit score 4 relative to the second sub-credit score 4 can be identified, namely, "the deposit amount increased from 200 to 400, and the deposit amount increased by 200".

[0190] Since the change in the first sub-credit score 4 relative to the second sub-credit score 4 is caused by a target behavior information, the difference between the first sub-credit score 4 and the second sub-credit score 4 can be directly used as the credit score corresponding to the target behavior information "the deposit amount increased from 200 to 400, and the deposit amount increased by 200". That is, the credit score corresponding to the target behavior information "the deposit amount increased from 200 to 400, and the deposit amount increased by 200" is 10 (i.e., 120-110=10).

[0191] In summary, since the target behavior information "the deposit amount increased from 200 to 400, and the deposit amount increased by 200" corresponds to two credit scores, both of which are 10, the total credit score for this target behavior information is 20 (i.e., 10+10=20).

[0192] The processor can determine that the first credit score is 740, the change in the first credit score relative to the second credit score is 20, the sign of the change in the first credit score relative to the second credit score is an upward sign, the target behavior information that caused the change in the first credit score relative to the second credit score is "the deposit amount increased from 200 to 400, the deposit amount increased by 200", and the corresponding total credit score for this target behavior information is 20. The processor can output the above information, such as... Figure 9 As shown.

[0193] In this embodiment, the processor can also provide suggestions to the first user based on the target behavior information and the total credit score corresponding to the target behavior information. For example, if the total credit score corresponding to the target behavior information "deposit amount increased from 200 to 400, deposit amount increased by 200" is 20, the processor can provide suggestions, such as if the deposit amount increases by another 200, the credit score can increase by another 20 points accordingly.

[0194] Furthermore, the processor can set different loan amounts and interest rates for different credit scores. The higher the credit score, the higher the corresponding loan amount and the lower the interest rate.

[0195] In addition, the processor can output the first user's credit score according to a preset period, which can be one month, one week, one day, etc., without limitation. For example, if the preset period is one day, the processor can determine and output the first user's credit score every day.

[0196] To facilitate viewing by the first user, the processor determines the first user's credit score according to a preset cycle and displays the determined credit score in the form of a chart. For example... Figure 10 As shown, the preset period is set to one day, and the processor can determine the credit score of the first user every day. The credit score of the first user is 724 from February 1, 2023 to February 3, 2023, and the credit score of the first user is 728 from February 4, 2023 to February 6, 2023.

[0197] The above method can accurately identify user behaviors that lead to changes in credit scores, making the credit scores determined in this application interpretable.

[0198] In this embodiment of the application, the structure of the credit score model is as follows: Figure 4 As shown, the system includes a mapping module, a feature extraction module, and a credit score output module. The mapping module includes an embedding layer and a first concatenation layer. The feature extraction module includes a feature cross-layer, L connection layers, and a second concatenation layer, where L is a non-negative integer. The credit score output module includes P connection layers and one fully connected layer, where P is a non-negative integer.

[0199] Generally, different values ​​for L and P result in different structures and performance of the credit scoring model after training. In this embodiment, P is set to 1 to improve training speed. This embodiment provides a method for determining the value of L.

[0200] In this embodiment of the application, the following is first constructed: Figure 11The first model shown has a structure including a mapping module, a feature extraction module, a credit score output module, and a loss function module. The mapping module includes an embedding layer and a first concatenation layer. The feature extraction module includes a feature cross-interaction layer and one connection layer. The credit score output module includes one connection layer and a fully connected layer. The loss function module includes a loss function layer, which can be a mean squared error (MSE) loss function, etc., and is not limited here.

[0201] The first model is trained using a first dataset. After training, the first model can be validated using a second dataset. The second dataset includes at least one set of M pre-defined feature domains labeled with credit scores. No two sets of data are the same in the first and second datasets.

[0202] When using the second dataset for validation, the processor can input a set of M preset feature domains with credit score labels into the trained first model. If the difference between the preset credit score output by the first model and the credit score label is within a first preset range, then the constructed first model structure is usable. For example, if the preset credit score output by the first model is 728, the credit score label is 725, the first preset range is [-10, 10], and the difference between the preset credit score 728 and the credit score label 725 is 3, the difference 3 is within the first preset range [-10, 10], therefore the constructed first model structure is usable.

[0203] If the difference between the preset credit score output by the first model and the credit score label exceeds the first preset range, the constructed first model structure is unusable. A connection layer can be added to the feature extraction module of the first model to construct a structure like... Figure 12 The second model shown has a structure including a mapping module, a feature extraction module, a credit score output module, and a loss function module. The mapping module, credit score output module, and loss function module in the second model are the same as those in the first model. The feature extraction module in the second model includes two connection layers, while the feature extraction module in the first model includes one connection layer.

[0204] The second model is trained using the same method as the first model. After training, the second model can be validated using a second dataset. If the difference between the preset credit score output by the second model and the credit score label exceeds a first preset range, the constructed second model structure is unusable. In this case, an additional connection layer can be added to the feature extraction module of the second model, and so on. If the difference between the preset credit score output by the second model and the credit score label is within the first preset range, the constructed second model structure is usable.

[0205] If the constructed second model structure is confirmed to be usable, the second model is retrained using the first dataset. During the training process, the gradient value of the parameter corresponding to the last connection layer in the feature extraction module of the second model is monitored. If the gradient value of the parameter corresponding to the last connection layer in the feature extraction module of the second model is less than a first preset gradient value, a second concatenation layer is added after the connection layer in the feature extraction module of the second model, thus constructing a second model. Figure 13 The third model is shown. After determining the structure of the third model, the third model is fine-tuned and trained using the first dataset to obtain the trained third model, which is then used as the credit score model in this embodiment.

[0206] The above methods can be used to determine a more accurate and efficient credit scoring model.

[0207] Based on the above embodiments, this application also provides a behavior detection device, which is applied to, for example... Figure 1 The processor of the computer device shown is used to implement Figure 2 The illustrated embodiment provides a behavior detection method. See also... Figure 14 As shown, the device includes a transceiver unit 1401 and a processing unit 1402. Wherein:

[0208] The transceiver unit 1401 is used to acquire the first credit score and N first sub-credit scores of the first user. The first credit score and N first sub-credit scores are obtained by inputting the M first feature domains of the first user acquired at the first moment into the credit score model. The first credit score is determined based on the N first sub-credit scores. Each of the N first sub-credit scores corresponds to at least one of the M first feature domains. Each first feature domain includes at least one first feature data. M is an integer greater than 1, and N is an integer greater than or equal to M. The credit score model is trained based on a first dataset, which includes at least one set of M preset feature domains with credit score labels.

[0209] The transceiver unit 1401 is also used to obtain the second credit score and N second sub-credit scores of the first user. The second credit score and N second sub-credit scores are obtained by inputting the M second feature domains of the first user obtained at the second time point into the credit score model. The second time point is before the first time point. The second credit score is determined based on the N second sub-credit scores. Each of the N second sub-credit scores corresponds to at least one of the M second feature domains. Each second feature domain includes at least one second feature data.

[0210] The processing unit 1402 is further configured to, when the first credit score and the second credit score are different, compare N first sub-credit scores and N second sub-credit scores to determine S first sub-credit scores and S second sub-credit scores corresponding to the S first sub-credit scores, wherein the S first sub-credit scores and the S second sub-credit scores are different; wherein S is a positive integer less than or equal to N.

[0211] The processing unit 1402 is further configured to compare the first feature data in the first feature domain corresponding to the S first sub-credit scores with the second feature data in the second feature domain corresponding to the S second sub-credit scores, and determine at least one target behavior information that causes the first credit score to change relative to the second credit score.

[0212] In one possible implementation, the M first feature domains of the first user are obtained by dividing the first user data of the first user acquired at the first moment based on a preset partitioning rule.

[0213] In one possible implementation, the credit score model includes a mapping module, a feature extraction module, and a credit score output module. The processing unit 1402 can obtain a first credit score and N first sub-credit scores in the following way: inputting at least one first feature data from M first feature domains into the mapping module to obtain first feature domain vectors corresponding to the M first feature domains respectively; inputting the obtained M first feature domain vectors into the feature extraction module to obtain N first cross vectors; and inputting the N first cross vectors into the credit score output module to obtain the first credit score and N first sub-credit scores.

[0214] In one possible implementation, the mapping module includes an embedding layer and a first concatenation layer; the processing unit 1402 can obtain M first feature domain vectors in the following way: for the j-th first feature domain among the M first feature domains, where j is an integer taking values ​​from [1, M], the following steps are performed: input at least one first feature data in the j-th first feature domain into the embedding layer, and determine that at least one first feature data in the j-th first feature domain corresponds to a first feature vector respectively; based on the first concatenation layer, concatenate at least one first feature vector corresponding to the j-th first feature domain to obtain the first feature domain vector corresponding to the j-th first feature domain.

[0215] In one possible implementation, the feature extraction module includes a feature cross layer, L connection layers, and a second concatenation layer, where L is a non-negative integer. The processing unit 1402 can obtain N first cross vectors by: inputting M first feature domain vectors into the feature cross layer to obtain N second cross vectors; inputting the N second cross vectors into the L connection layers to obtain N third cross vectors; and inputting the N third cross vectors and M first feature domain vectors into the second concatenation layer to obtain N first cross vectors.

[0216] In one possible implementation, when the processing unit 1402 inputs M first feature domain vectors into the feature cross layer to obtain N second cross vectors, it specifically performs the following: for the j-th first feature domain vector among the M first feature domain vectors, it performs a matrix dot product between the j-th first feature domain vector and the j-th first feature domain vector to obtain a first type vector corresponding to the j-th first feature domain vector, where j is an integer that takes values ​​from [1, M]; based on a preset cross rule, it determines at least one set of two related first feature domain vectors, performs a matrix dot product between each set of two related first feature domain vectors to obtain at least one second type vector; and uses the first type vector corresponding to the j-th first feature domain vector and at least one second type vector as N second cross vectors.

[0217] In one possible implementation, the processing unit 1402 inputs N third cross vectors and M first feature domain vectors into the second concatenation layer. When obtaining N first cross vectors, it specifically performs the following steps for the p-th third cross vector among the N third cross vectors, where p is an integer taking values ​​from [1, N]: determining at least one first feature domain vector corresponding to the p-th third cross vector as the target feature vector; concatenating the p-th third cross vector with at least one target feature domain vector to obtain the first cross vector corresponding to the p-th third cross vector.

[0218] In one possible implementation, when processing unit 1402 compares the first feature data in the first feature domain corresponding to S first sub-credit scores with the second feature data in the second feature domain corresponding to S second sub-credit scores to determine at least one target behavior information that causes a change in the first credit score relative to the second credit score, it specifically performs the following steps for the i-th first sub-credit score among S first sub-credit scores and the i-th second sub-credit score among S second sub-credit scores, where the i-th first sub-credit score and the i-th second sub-credit score correspond to each other, and i is an integer taking values ​​from [1, S]: determining at least one first comparison data, wherein the at least one first comparison data is the first feature data in at least one first feature domain corresponding to the i-th first sub-credit score; determining at least one second comparison data, wherein the at least one second comparison data is the second feature data in at least one second feature domain corresponding to the i-th second sub-credit score; and determining at least one target behavior information that causes a change in the i-th first sub-credit score relative to the i-th second sub-credit score based on the at least one first comparison data and the at least one second comparison data.

[0219] In one possible implementation, the processing unit 1402 may also output at least one target behavior information that causes the first credit score to change relative to the second credit score.

[0220] In one possible implementation, the processing unit 1402 may also output at least one of the following: output the change of the first credit score relative to the second credit score, or output the first credit score.

[0221] In one possible implementation, the change in the first credit score relative to the second credit score includes at least one of the following: the amount of change in the first credit score relative to the second credit score, or the sign of the change in the first credit score relative to the second credit score.

[0222] This application provides a behavior detection device. When a first credit score and a second credit score are different, the device can determine at least one target behavior that causes a change in the first credit score. It can accurately determine which user behavior changes will lead to a change in the credit score, making the credit score determined by this application interpretable.

[0223] It should be noted that the module division in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, exist as separate physical entities, or have two or more units integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0224] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0225] Based on the above embodiments, this application also provides a computer device for implementing the following: Figure 2 The behavior detection method shown has the following characteristics: Figure 14 The function of the behavior detection device shown. (See also...) Figure 15 As shown, the computer device includes: a processor 1501 and a memory 1502.

[0226] The processor 1501 and the memory 1502 are interconnected. Optionally, the processor 1501 and the memory 1502 can be interconnected via a bus 1503; the bus 1503 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 15 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0227] Optionally, the computer device 1500 further includes a communication module 1504 for communicating and interacting with other devices. Optionally, the communication module 1504 can communicate with other devices wirelessly; for example, the communication module 1504 can be an RF circuit, a WiFi module, etc. The communication module 1504 can also communicate with other devices via a physical connection; for example, the communication module 1504 can be a communication interface.

[0228] The processor 1501 is used to implement, for example Figure 2The behavior detection method shown can be referred to the specific description in the above embodiments, and will not be repeated here.

[0229] The memory 1502 is used to store programs and data, etc. Specifically, the program may include program code, which includes instructions for computer operation. The memory 1502 may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. The processor 1501 executes the program stored in the memory 1502 to implement the above functions, thereby achieving the following: Figure 2 The behavior detection method shown.

[0230] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, causes the computer to perform actions such as... Figure 2 The behavior detection method shown.

[0231] In summary, the embodiments of this application provide a behavior detection method and apparatus. In this method, when the first credit score and the second credit score are different, at least one target behavior that causes the change in the first credit score can be determined. It can accurately determine which changes in the user's behavior will lead to changes in the credit score, so that the credit score determined by this application has interpretability.

[0232] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0233] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0234] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0235] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0236] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A behavior detection method, characterized in that, The method includes: The system obtains a first credit score and N first sub-credit scores for a first user. The first credit score and the N first sub-credit scores are obtained by inputting M first feature domains of the first user, acquired at a first moment, into a credit score model. The first credit score is determined based on the N first sub-credit scores. Each of the N first sub-credit scores corresponds to at least one of the M first feature domains, and each first feature domain includes at least one first feature data. M is an integer greater than 1, and N is an integer greater than or equal to M. The credit score model is trained based on a first dataset, which includes at least one set of M preset feature domains labeled with credit scores. The second credit score and N second sub-credit scores of the first user are obtained. The second credit score and the N second sub-credit scores are obtained by inputting M second feature domains of the first user obtained at a second time point into the credit score model. The second time point is before the first time point. The second credit score is determined based on the N second sub-credit scores. Each of the N second sub-credit scores corresponds to at least one of the M second feature domains. Each second feature domain includes at least one second feature data. If the first credit score is different from the second credit score, the N first sub-credit scores and the N second sub-credit scores are compared to determine S first sub-credit scores and S second sub-credit scores corresponding to the S first sub-credit scores. The S first sub-credit scores and the S second sub-credit scores are different, and S is a positive integer less than or equal to N. The first feature data in the first feature domain corresponding to the S first sub-credit scores is compared with the second feature data in the second feature domain corresponding to the S second sub-credit scores to determine at least one target behavior information that causes the first credit score to change relative to the second credit score.

2. The method as described in claim 1, characterized in that, The M first feature domains of the first user are obtained by dividing the first user data of the first user acquired at the first moment based on a preset division rule.

3. The method as described in claim 1, characterized in that, The credit score model includes a mapping module, a feature extraction module, and a credit score output module; The first credit score and the N first sub-credit scores are obtained in the following manner: Input at least one first feature data from the M first feature domains into the mapping module to obtain the first feature domain vectors corresponding to the M first feature domains respectively; The obtained M first feature domain vectors are input into the feature extraction module to obtain N first cross vectors; The N first cross vectors are input into the credit score output module to obtain the first credit score and the N first sub-credit scores.

4. The method as described in claim 3, characterized in that, The mapping module includes an embedding layer and a first splicing layer; The M first feature domain vectors are obtained in the following way: For the j-th first feature field among the M first feature fields, where j is an integer that takes all elements in the range [1, M], the following steps are performed: At least one first feature data in the j-th first feature domain is input into the embedding layer to determine that at least one first feature data in the j-th first feature domain corresponds to a first feature vector; Based on the first splicing layer, at least one first feature vector corresponding to the j-th first feature domain is spliced ​​to obtain the first feature domain vector corresponding to the j-th first feature domain.

5. The method as described in claim 3, characterized in that, The feature extraction module includes a feature cross layer, L connection layers, and a second concatenation layer, where L is a non-negative integer. The N first cross vectors are obtained in the following way: The M first feature domain vectors are input into the feature cross layer to obtain N second cross vectors; The N second cross vectors are input into the L connection layers to obtain N third cross vectors; The N third cross vectors and the M first feature domain vectors are input into the second concatenation layer to obtain the N first cross vectors.

6. The method as described in claim 5, characterized in that, The step of inputting the M first feature domain vectors into the feature cross layer to obtain N second cross vectors includes: For the j-th first feature field vector among the M first feature field vectors, perform a matrix dot product between the j-th first feature field vector and the j-th first feature field vector to obtain the first type vector corresponding to the j-th first feature field vector, where j is an integer that takes all elements in [1, M]. Based on the preset cross rules, at least one set of two related first feature domain vectors is determined, and matrix dot product is performed on each set of two related first feature domain vectors to obtain at least one second type vector. The first type vector corresponding to the j-th first feature domain vector and the at least one second type vector are used as the N second cross vectors.

7. The method as described in claim 5 or 6, characterized in that, The step of inputting the N third cross vectors and the M first feature domain vectors into the second concatenation layer to obtain the N first cross vectors includes: For the p-th third cross vector among the N third cross vectors, where p is an integer that takes all elements in the range [1, N], perform the following steps: Determine at least one first feature domain vector corresponding to the p-th third cross vector as the target feature vector; The p-th third cross vector is concatenated with at least one target feature domain vector to obtain the first cross vector corresponding to the p-th third cross vector.

8. The method according to any one of claims 1-6, characterized in that, The first feature data in the first feature domain corresponding to the S first sub-credit scores is compared with the second feature data in the second feature domain corresponding to the S second sub-credit scores to determine at least one target behavior information that causes a change in the first credit score relative to the second credit score, including: For the i-th first sub-credit score among the S first sub-credit scores, and the i-th second sub-credit score among the S second sub-credit scores, where the i-th first sub-credit score and the i-th second sub-credit score correspond, and i is an integer taking values ​​from [1, S], the following steps are performed: Determine at least one first data to be compared, wherein the at least one first data to be compared is a first feature data in at least one first feature domain corresponding to the i-th first sub-credit score; Determine at least one second data to be compared, wherein the at least one second data to be compared is a second feature data in at least one second feature domain corresponding to the i-th second sub-credit score; Based on the at least one first comparison data and the at least one second comparison data, determine at least one target behavior information that causes the i-th first sub-credit score to change relative to the i-th second sub-credit score.

9. The method according to any one of claims 1-6, characterized in that, Also includes: Output at least one target behavior information that causes the first credit score to change relative to the second credit score.

10. The method as described in claim 9, characterized in that, It also includes at least one of the following: Output the change of the first credit score relative to the second credit score, or... Output the first credit score.

11. The method as described in claim 10, characterized in that, The change in the first credit score relative to the second credit score includes at least one of the following: The change in the first credit score relative to the second credit score, or, The sign of the change in the first credit score relative to the second credit score.

12. A behavior detection device, characterized in that, The device includes: The transceiver unit is used to acquire a first user's first credit score and N first sub-credit scores. The first credit score and the N first sub-credit scores are obtained by inputting M first feature domains of the first user acquired at a first moment into a credit score model. The first credit score is determined based on the N first sub-credit scores. Each of the N first sub-credit scores corresponds to at least one of the M first feature domains, and each first feature domain includes at least one first feature data. M is an integer greater than 1, and N is an integer greater than or equal to M. The credit score model is trained based on a first dataset, which includes at least one set of M preset feature domains labeled with credit scores. The transceiver unit is further configured to acquire the second credit score and N second sub-credit scores of the first user. The second credit score and the N second sub-credit scores are obtained by inputting M second feature domains of the first user acquired at a second time point into the credit score model. The second time point is prior to the first time point. The second credit score is determined based on the N second sub-credit scores. Each of the N second sub-credit scores corresponds to at least one of the M second feature domains, and each second feature domain includes at least one second feature data. The processing unit is further configured to, when the first credit score is different from the second credit score, compare the N first sub-credit scores and the N second sub-credit scores to determine S first sub-credit scores and S second sub-credit scores corresponding to the S first sub-credit scores, wherein the S first sub-credit scores and the S second sub-credit scores are different, and S is a positive integer less than or equal to N; The processing unit is further configured to compare the first feature data in the first feature domain corresponding to the S first sub-credit scores with the second feature data in the second feature domain corresponding to the S second sub-credit scores, and determine at least one target behavior information that causes the first credit score to change relative to the second credit score.

13. A computer device, characterized in that, include: Memory, used to store programs and data; A processor for running a program stored in the memory, and performing the method as described in any one of claims 1-11 based on data stored in the memory.

14. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a computer, causes the computer to perform the method provided by any one of claims 1-11.

Images