Data processing method, device and electronic device

Abstract

The embodiments of the present application disclose a data processing method, device and electronic device. The method includes: inputting corresponding user feature data into multiple expert networks in a multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks; obtaining the expert network features of the prediction task through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network; obtaining the prediction result of each prediction task through the expert network features of each prediction task and the tower network. Thus, in the process of obtaining the prediction result of each prediction task, in addition to using the features output by the task expert network of each prediction task itself, the features output by the shared expert network will also be used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more specifically, to a data processing method, device and electronic device. Background Art

[0002] In financial scenarios, user situations can be predicted so that users can be treated accordingly based on the prediction results. For example, when recommending credit flow distribution, user situations can be predicted and users can be recommended to lending institutions based on the prediction results. However, in related methods, there is still the problem of inaccurate prediction of user situations. Summary of the invention

[0003] In view of the above problems, the present application proposes a data processing method, device and electronic device to improve the above problems.

[0004] In a first aspect, the present application provides a data processing method, comprising: inputting corresponding user feature data into multiple expert networks in a multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, and the user feature data include data related to a target user; obtaining an expert network feature of the prediction task through a first output feature output by the task expert network of the prediction task and a second output feature output by the shared expert network; obtaining a prediction result of each prediction task through the expert network feature of each prediction task and a tower network, so as to determine a recommended object corresponding to the target user through the prediction result of each prediction task.

[0005] In a second aspect, the present application provides a data processing device, comprising: a data input unit, used to input corresponding user feature data into multiple expert networks in a multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, and the user feature data includes data related to the target user; a feature calculation unit, used to obtain the expert network feature of the prediction task through the first output feature output by the task expert network of the prediction task and the second output feature output by the shared expert network; a prediction unit, used to obtain the prediction result of each prediction task through the expert network feature of each prediction task and the tower network, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

[0006] In a third aspect, the present application provides an electronic device, comprising a processor and a memory; one or more programs are stored in the memory and configured to be executed by the processor to implement the above method.

[0007] In a fourth aspect, the present application provides a computer-readable storage medium having a program code stored therein, wherein the above method is executed when the program code is started and executed by a controller.

[0008] In a fifth aspect, the present application provides a computer program product, which includes computer instructions stored in a computer-readable storage medium; the computer program product is used to implement the above method.

[0009] The present application provides a data processing method, device and electronic device, in which, in the multi-target network involved, multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks. After inputting corresponding user feature data into the multiple expert networks, the expert network features of the prediction task can be obtained through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network, and then the prediction results of each prediction task can be obtained through the expert network features of each prediction task and the tower network. Thus, in the process of obtaining the prediction results of each prediction task, in addition to using the features (first output features) output by the task expert network of each prediction task itself, the features (second output features) output by the shared expert network will also be used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results, and thus, it is also possible to more accurately determine the recommended objects for the target users. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A schematic diagram showing an operating environment of the data processing method proposed in an embodiment of the present application;

[0012] Figure 2 A schematic diagram showing another operating environment of the data processing method proposed in an embodiment of the present application;

[0013] Figure 3 A flow chart of a data processing method proposed in an embodiment of the present application is shown;

[0014] Figure 4 A schematic diagram of the structure of a multi-objective model proposed in an embodiment of the present application is shown;

[0015] Figure 5 A schematic diagram of the structure of a multi-objective model proposed in an embodiment of the present application is shown;

[0016] Figure 6 A schematic diagram of the structure of a tower network in an embodiment of the present application is shown;

[0017] Figure 7 A flow chart of a data processing method proposed in another embodiment of the present application is shown;

[0018] Figure 8 A schematic diagram of the structure of a multi-objective model proposed in an embodiment of the present application is shown;

[0019] Fig. 9 A flow chart of a data processing method proposed in another embodiment of the present application is shown;

[0020] Fig.10 A flowchart of a data processing method proposed in another embodiment of the present application is shown;

[0021] Fig.11 A schematic diagram showing the effect of randomly ignoring neurons in an embodiment of the present application is shown;

[0022] Fig.12 A structural block diagram of a data processing device proposed in an embodiment of the present application is shown;

[0023] Fig.13 A structural block diagram of an electronic device for executing a data processing method according to an embodiment of the present application is shown;

[0024] Fig.14 It is a storage unit of an embodiment of the present application for storing or carrying program codes for implementing the data processing method according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0026] In financial scenarios, accurate prediction of user situations is of vital importance. Taking credit as an example, when making credit flow distribution recommendations, decisions can be made based on predictions of user situations. Specifically, users can be recommended to appropriate lending institutions based on the prediction results to improve the efficiency and accuracy of resource allocation. However, there are still some problems that cannot be ignored in related methods, the most prominent of which is the inaccurate prediction of user situations, which may lead to inappropriate or unreasonable recommendations, affecting user experience and the business results of financial institutions.

[0027] Therefore, the inventors have proposed a data processing method, device and electronic device in the present application that can improve the above-mentioned problems. In the multi-target network involved in the method, multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks. After inputting corresponding user feature data into the multiple expert networks, the expert network features of the prediction task can be obtained through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network, and then the prediction results of each prediction task can be obtained through the expert network features of each prediction task and the tower network. Thus, in the process of obtaining the prediction results of each prediction task, in addition to using the features (first output features) output by the task expert network of each prediction task itself, the features (second output features) output by the shared expert network will also be used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results, and thus, it is also possible to more accurately determine the recommended objects for the target users.

[0028] The following first introduces the application scenarios involved in the embodiments of the present application.

[0029] In the embodiment of the present application, the data processing method provided can be executed by an electronic device. In this way, all steps in the data processing method provided by the embodiment of the present application can be executed by the electronic device. For example, Figure 1 As shown, all steps in the data processing method provided in the embodiment of the present application can be executed by the processor of the electronic device 100.

[0030] Alternatively, the data processing method provided in the embodiment of the present application may also be executed by a server. Correspondingly, in this mode of execution by the server, the server may start executing the steps in the data processing method provided in the embodiment of the present application in response to a trigger instruction. The trigger instruction may be sent by an electronic device used by a user, or may be triggered locally by the server in response to some automation events.

[0031] In addition, the data processing method provided in the embodiment of the present application can also be executed by the electronic device and the server in collaboration. In this way, some steps in the data processing method provided in the embodiment of the present application are executed by the electronic device, while other steps are executed by the server. For example, Figure 2 As shown, the electronic device 100 can execute the data processing method including: obtaining user feature data (including user feature data corresponding to each expert network). Afterwards, the electronic device 100 transmits the user feature data to the server 200. After receiving the user feature data, the server 200 can execute subsequent steps to obtain the prediction results and recommended objects of each prediction task.

[0032] It should be noted that in this method of collaborative execution by the electronic device and the server, the steps respectively executed by the electronic device and the server are not limited to the methods introduced in the above examples. In actual applications, the steps respectively executed by the electronic device and the server can be dynamically adjusted according to actual conditions.

[0033] It should be noted that the electronic device 100 is Figure 1 and Figure 2 In addition to the smartphone shown in , it can also be a tablet computer, an intelligent voice assistant and other devices. Server 200 can be an independent physical server, or it can be a server cluster or distributed system composed of multiple physical servers, as well as a cloud server for basic cloud computing services such as an artificial intelligence platform. Among them, in the case where the data processing method provided in the embodiment of the present application is executed by a server cluster or distributed system composed of multiple physical servers, different steps in the data processing method can be executed by different physical servers respectively, or can be executed in a distributed manner by a server built based on a distributed system.

[0034] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0035] See also Figure 3 , a data processing method provided in an embodiment of the present application, the method comprising:

[0036] S110: inputting corresponding user feature data into a plurality of expert networks in the multi-target network, wherein the plurality of expert networks include a shared expert network and task expert networks of respective plurality of prediction tasks, and the user feature data include data related to the target user.

[0037] In the embodiment of the present application, the multi-objective network may also be referred to as a multi-objective learning (Multi-Task Learning) network. Optionally, the multi-objective network may be a CGC (Customized Gate Control model) network.

[0038] In a multi-objective network, there can be task expert networks and shared expert networks.

[0039] In a multi-objective network, a task expert network is an expert network used to extract and represent features related to a specific task. Therefore, in this application, each prediction task can have its own task expert network, and the design of the task expert network enables it to capture feature representations that are closely related to the task, thereby helping to improve the performance of the task. In this case, the task expert network can specifically extract features related to a specific task from the input user feature data. Wherein, in the case where there are multiple prediction tasks, there will also be multiple task expert networks, and the user feature data input into each task expert network can be user feature data related to the corresponding prediction task. Exemplarily, if multiple prediction tasks can include a credit prediction task, a loan prediction task, and an overdue prediction task, wherein the credit prediction task can be understood as a task for predicting whether the target user can obtain a loan, and the approved credit limit in the case of obtaining a loan, and the loan prediction task can be understood as a task for predicting the loan rate of the target user after obtaining the loan, and the loan rate refers to the proportion of the actual credit limit used by the target user after obtaining the credit limit. The overdue prediction task can be understood as a task for predicting the probability that the target user will repay the loan overdue after obtaining the loan.

[0040] The multiple expert networks may include a task expert network corresponding to the credit limit prediction task, a task expert network corresponding to the expenditure prediction task, and a task expert network corresponding to the overdue prediction task. Correspondingly, the user feature data related to the credit limit prediction task may be input into the task expert network corresponding to the credit limit prediction task, the user feature data related to the expenditure prediction task may be input into the task expert network corresponding to the expenditure prediction task, and the user feature data related to the overdue prediction task may be input into the task expert network corresponding to the overdue prediction task.

[0041] In a multi-target network, a shared expert network is an expert network used by multiple prediction tasks. The design of shared expert networks enables them to capture feature representations shared between different tasks, thereby reusing these features. Therefore, shared expert networks are responsible for learning common feature representations and knowledge across tasks. Among them, shared expert networks help improve the generalization ability of multi-target networks by capturing common features across tasks. In addition, the common features obtained by shared expert networks are used in multiple tasks, which helps to enhance the performance of the model on unseen data. By combining with task expert networks, shared expert networks can realize customized sharing of expert knowledge of different tasks, thereby further improving the performance of the model.

[0042] Based on the foregoing content, the structure of a multi-target network involved in the embodiment of the present application can be as follows: Figure 4 As shown. Figure 4 As shown, the multiple prediction tasks may include a prediction task C, a prediction task W, and a prediction task R. In this case, the multi-objective network may include a task expert network C corresponding to the prediction task C, a task expert network W corresponding to the prediction task W, and a task expert network R corresponding to the prediction task R. In addition, Figure 4 The structure shown also includes a shared expert network S.

[0043] In an embodiment of the present application, user feature data is data that characterizes the characteristics of a user (e.g., a target user). Optionally, any data that can be used to characterize the characteristics of a user can be used as user feature data. Exemplarily, user feature data may include application usage behavior data, browser search behavior data, lending behavior data, demographic feature data, and the like. In an embodiment of the present application, the user feature data input into each expert network (including a task expert network and a shared expert network) may be the same or different. For example, in one approach, the user feature data required for different prediction tasks may be different, and thus, the user feature data input into the task expert networks of different prediction tasks may be different.

[0044] In the embodiment of the present application, the target user can be understood as the user currently being predicted. Optionally, when there are multiple users to be predicted, the multiple users to be predicted can be predicted as target users in sequence. Optionally, when there are multiple users to be predicted, multiple users to be predicted can also be predicted synchronously. It should be noted that in the case of synchronous prediction of multiple target users, a multi-target network can be run in multiple devices, so that multiple target users can be predicted synchronously through the multi-target network running in the multiple devices.

[0045] S120: Obtaining expert network features of the prediction task through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network.

[0046] As mentioned above, the features output by the task expert network are mainly used to extract and represent information about specific tasks. These features are mainly used to help the multi-objective network better understand and process specific tasks. For example, for prediction task C, the first output feature of the corresponding task expert network output is the feature related to prediction task C, for prediction task W, the first output feature of the corresponding task expert network output is the feature related to prediction task W, and for prediction task R, the first output feature of the corresponding task expert network output is the feature related to prediction task R. For example, in the video recommendation scenario, the task expert network for the task of "video completion" may output features related to video content quality (such as clarity, picture stability, etc.) or user viewing behavior (such as viewing time, whether to quit midway, etc.).

[0047] Among them, the shared expert network is mainly responsible for learning common feature representations across multiple tasks. These features are usually shared by all tasks and reflect the commonality or relevance between tasks. For example, in the video recommendation scenario, the shared expert network may output common features about the user's basic information (such as age, gender, interest preferences, etc.) or the basic attributes of the video (such as type, duration, director, etc.).

[0048] For example, in the case where multiple prediction tasks include a credit limit prediction task, a spending prediction task, and an overdue prediction task, the first output feature of the task expert network corresponding to the credit limit prediction task may be a feature that characterizes the user's income and liabilities and the user's asset status, and the first output feature of the task expert network corresponding to the spending prediction task may be a feature that characterizes the user's consumption habits and the user's capital flow. The first output feature of the task expert network of the overdue prediction task may characterize the user's overdue history and the user's credit score.

[0049] The first output feature of each task expert network can be fused with the second output feature of the shared expert network to obtain the input of the tower network for each prediction task, so that the input of the tower network can be fused with the output content of the corresponding task expert network and the output content of the shared expert network.

[0050] As a feature fusion method, obtaining the expert network feature of the prediction task through the first output feature output by the task expert network of the prediction task and the second output feature output by the shared expert network may include: obtaining the expert network feature of each prediction task through the first output feature output by the task expert network of the prediction task, the weight corresponding to the first output feature, the second output feature output by the shared expert network, and the weight of the second output feature. Optionally, the weight corresponding to the first output feature and the weight of the second output feature can be obtained through the gate control network (Gating Network) included in the multi-objective network.

[0051] In the embodiment of the present application, the gate control network is responsible for dynamically obtaining the weights of different expert networks (including task expert networks and shared expert networks) according to the input data (for example, the first output feature and the second output feature input to the gate control network), thereby realizing personalized processing of different tasks. This mechanism enables the multi-objective network to process multiple tasks at the same time and flexibly adjust the network structure according to the requirements of the task.

[0052] For example, Figure 5 As shown, the gate control network includes a selection layer (Selector), a fully connected layer (Fully Connected Layer) and an activation function layer, wherein the activation function included in the activation function layer can be Softmax.

[0053] Among them, the main function of the selection layer is to act as a selector to generate selection probabilities based on the input data, which will be used to fuse the outputs from different expert modules. In a multi-target network, the tower network of each prediction task can acquire knowledge from all expert networks (including shared expert networks and task expert networks). The selection layer generates the selection probability for each expert module by calculating the correlation or matching degree between the input data and each expert network.

[0054] Among them, the fully connected layer is responsible for receiving the output of the selection layer and further processing it to generate the final fusion result. In a multi-target network, the fully connected layer can be a single-layer feedforward network, which receives the input from the selection layer (i.e., the selection probability) and performs a linear transformation through the weight matrix and bias term, and then applies the activation function to generate the output. Therefore, the output of the fully connected layer may be directly used as the input of the activation function layer.

[0055] The activation function layer is used to convert the output of the fully connected layer into a probability distribution, and ensures that the sum of all output values ​​is 1 and each value is between 0 and 1, which helps the gating network to perform smoother weighted sum calculations when selecting the output of the expert module. The activation function layer converts the output of the fully connected layer into a probability distribution by calculating the ratio of the exponent of each output value to the sum of the exponents of all output values.

[0056] The input of the gate control network can include two parts: one part is the original input data (for example, the user feature data that will be input into the expert network), and the other part is the output from different expert networks. These input data are pre-processed and sent to the gate control network. The gate control network generates a weight value for each expert network through a series of calculations. These weight values ​​reflect the contribution of each expert network to the final output under the current input.

[0057] According to the calculated weight values, the gate control network can perform weighted summation on the outputs of different expert networks to obtain the final output result. For example, the expert network feature for each prediction task can be obtained based on the first output feature, the weight corresponding to the first output feature, the second output feature of the shared expert network output, and the weighted summation of the second output feature.

[0058] Exemplarily, for prediction task C, after obtaining the first output feature output by task expert network C and the second output feature output by shared expert network S, weighted summation can be performed in the aforementioned manner to obtain the expert network feature of prediction task C. For prediction task W, after obtaining the first output feature output by task expert network W and the second output feature output by shared expert network S, weighted summation can be performed in the aforementioned manner to obtain the expert network feature of prediction task W. For prediction task R, after obtaining the first output feature output by task expert network R and the second output feature output by shared expert network S, weighted summation can be performed in the aforementioned manner to obtain the expert network feature of prediction task R.

[0059] S130: Obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

[0060] After obtaining the expert network features of each prediction task, the expert network features of each prediction task can be input into the corresponding tower network to obtain the prediction result of each prediction task through the processing of the tower network. Among them, each task expert network will have its own corresponding tower network, that is, there will be a tower network corresponding to each prediction task.

[0061] In the embodiment of the present application, the tower network may also be referred to as a task tower. The tower network may be composed of a multi-layer neural network, and these network layers may be fully connected layers, convolutional layers, recurrent layers, etc., wherein the specific structure of the tower network may depend on the nature of the task to be performed and the characteristics of the input data. The width (i.e., the number of neurons) and depth (i.e., the number of network layers) of the tower network are both adjustable hyperparameters that can be optimized according to the requirements of the specific task.

[0062] The tower network can receive information from the expert network (for example, the expert network features obtained above), which can be obtained by weighted combination of the output of the expert network through the gate control network, reflecting the importance of different expert networks for the current task. The neural network layer within the tower network can further process and extract features from the received information to form a higher-level feature representation. Based on the extracted features, the tower network can output the final prediction results or decisions. These results can be classification labels, regression values, probability distributions, etc., depending on the type of task.

[0063] Optionally, in a multi-objective network, the tower network and the expert network interact through a gate control network. The gate control network can dynamically adjust the output weights of different expert networks according to the needs of the current task, and pass the weighted output (expert network features) to the task tower. This interaction mechanism enables the task tower to make full use of the knowledge of different expert networks, thereby improving the accuracy of prediction.

[0064] For example, the structure of a tower network involved in the embodiments of the present application can be as follows: Figure 6 As shown. Figure 6 As shown, the tower network structure uses a feedforward neural network, which consists of an input layer, a hidden layer, and an output layer, in which each neuron is fully connected to the neurons in the next layer. The neurons in each layer receive input from the previous layer and perform weighted summation, and realize nonlinear transformation through activation functions.

[0065] As a method, the recommended object includes a loan providing object, the multiple prediction tasks include a loan amount prediction task, and the activation function of the tower network of the loan amount prediction task includes a linear part and a zero value part, the linear part is used to output the approved loan amount of the target user when the loan review is passed, and the zero value part is used to output that the target user has not passed the loan review. Optionally, the activation function can be a ReLu function.

[0066] Among them, the expression of the ReLu function can be:

[0067]

[0068] Among them, x in the above expression can be understood as the aforementioned linear part, and 0 can be understood as the aforementioned zero value part.

[0069] A data processing method provided in this embodiment enables, through the above-mentioned method, in the process of obtaining the prediction result of each prediction task, in addition to using the features output by the task expert network of each prediction task itself (first output features), the features output by the shared expert network (second output features) are also used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results, and further, it is also possible to more accurately determine the recommended objects for the target users.

[0070] See also Figure 7 , a data processing method provided in an embodiment of the present application, the method comprising:

[0071] S210: Inputting corresponding user feature data into multiple expert networks in the multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, and the user feature data includes data related to the target user, wherein the prediction target of each prediction task is different.

[0072] S220: Determine weights of features output by multiple first sub-networks in the task expert network, wherein the features output by the multiple first sub-networks are different.

[0073] In the embodiment of the present application, each expert network may include multiple sub-networks. For example, for each task expert network, multiple first sub-networks may be included. For a shared expert network, multiple second sub-networks may be included. For example, Figure 8 As shown, the task expert network C, the task expert network S, and the task expert network R each include multiple first subnetworks. The shared expert network S may include multiple second subnetworks. The number of subnetworks included in each expert network is not specifically limited in the embodiment of the present application. In one embodiment, the number of subnetworks included in each expert network may be the same. Figure 8 The "vector 1 .....vector M" involved in the gate control network can be understood as data input into the gate control network, for example, it can be the features output by multiple sub-networks in a certain expert network.

[0074] It should be noted that in the embodiment of the present application, the output of each subnetwork (the first subnetwork or the second subnetwork) is a vector. Since each subnetwork has different weights and bias parameters, even if the data input to the subnetwork is the same, each subnetwork will learn different feature representations and eventually output different vectors. Exemplarily, the task expert network C may include a first subnetwork C1, a first subnetwork C2, and a first subnetwork C3. When the user feature data input to the first subnetwork C1, the first subnetwork C2, and the first subnetwork C3 are the same, the features output by the first subnetwork C1, the first subnetwork C2, and the first subnetwork C3 may be different.

[0075] As a method, the weights of the features output by the multiple first sub-networks are determined by the gate control network based on the features output by the multiple first sub-networks.

[0076] Optionally, in the same task expert network, the weights of the features of each first subnetwork are determined by the gate control network based on the features output by each first subnetwork itself. Exemplarily, in the case where the task expert network C includes the first subnetwork C1, the first subnetwork C2, and the first subnetwork C3, the gate control network can determine the weight of the features output by the first subnetwork C1 based on the features output by the first subnetwork C1. Correspondingly, the gate control network can determine the weight of the features output by the first subnetwork C2 based on the features output by the first subnetwork C2, and the gate control network can determine the weight of the features output by the first subnetwork C3 based on the features output by the first subnetwork C3.

[0077] Optionally, in the same task expert network, the weights of the features of each first subnetwork are determined by the gate control network based on the features of all the first subnetworks in the same task expert network. Exemplarily, in the case where the task expert network C includes the first subnetwork C1, the first subnetwork C2, and the first subnetwork C3, the gate control network can determine the weight of the feature output by the first subnetwork C1 based on the features output by the first subnetwork C1, the features output by the first subnetwork C2, and the features output by the first subnetwork C3. Correspondingly, the gate control network can determine the weight of the feature output by the first subnetwork C2 based on the features output by the first subnetwork C1, the features output by the first subnetwork C2, and the features output by the first subnetwork C3, and the gate control network can determine the weight of the feature output by the first subnetwork C3 based on the features output by the first subnetwork C1, the features output by the first subnetwork C2, and the features output by the first subnetwork C3.

[0078] As a method, the weights of the features output by the multiple first sub-networks are determined by the gate control network based on the features output by the multiple first sub-networks and the features output by the multiple second sub-networks.

[0079] Optionally, in the same task expert network, the weights of the features of each first subnetwork can be determined based on the features output by each first subnetwork itself and the features output by multiple second subnetworks. Exemplarily, the task expert network C includes the first subnetwork C1, the first subnetwork C2 and the first subnetwork C3, and the shared expert network S includes the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3, wherein the gate control network can determine the weight of the feature output by the first subnetwork C1 through the features output by the first subnetwork C1 and the features output by the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3. Correspondingly, the gate control network can determine the weight of the feature output by the first subnetwork C2 through the features output by the first subnetwork C2 and the features output by the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3. The gate control network can determine the weight of the feature output by the first subnetwork C3 through the features output by the first subnetwork C3 and the features output by the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3.

[0080] Optionally, in the same task expert network, the weights of the features of each first subnetwork are determined by the gate control network based on the features of all first subnetworks in the same task expert network and the features output by all second subnetworks. Exemplarily, the task expert network C includes the first subnetwork C1, the first subnetwork C2 and the first subnetwork C3, and the shared expert network S includes the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3. In this case, the weights of the features output by the first subnetwork C1, the first subnetwork C2 and the first subnetwork C3 can be obtained based on the features output by the first subnetwork C1, the first subnetwork C2 and the first subnetwork C3, and the features output by the second subnetwork S1, the second subnetwork S2 and the second subnetwork S3.

[0081] S230: Based on the features output by the multiple first sub-networks in the task expert network and the weight of each feature, a first output feature of the task expert network output is obtained.

[0082] As a method, in a task expert network, the first output feature of the task expert network can be obtained by weighted summing the features output by multiple first sub-networks.

[0083] Exemplarily, the task expert network C includes a first sub-network C1, a first sub-network C2 and a first sub-network C3. In this case, the first output feature output by the task expert network C can be obtained by weighted summing the features output by the first sub-network C1, the first sub-network C2 and the first sub-network C3.

[0084] S240: Obtaining expert network features of the prediction task through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network.

[0085] S250: Obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

[0086] The data processing method provided in this embodiment is conducive to improving the accuracy of the prediction results through the above-mentioned method, and further, it can also more accurately determine the recommended objects for the target users. In addition, in this embodiment, the task expert network and the shared expert network each include a sub-network, and further, when determining the output features of each task expert network, in addition to referring to the output features of the sub-network included in the task expert network, the output features of each sub-network in the shared expert network are also referred to, so that the output features of each task expert network are more refined and comprehensive.

[0087] See also Fig. 9 , a data processing method provided in an embodiment of the present application, the method comprising:

[0088] S310: Inputting corresponding user feature data into multiple expert networks in the multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, and the user feature data includes data related to the target user, wherein the multiple prediction tasks include at least two of a quota prediction task, a spending prediction task, and an overdue prediction task.

[0089] S320: Obtaining expert network features of the prediction task through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network.

[0090] S330: Obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task, and the recommended object includes the loan provider recommended to the target user, or includes the loan provider recommended to the target user.

[0091] This embodiment provides a data processing method, in which multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, wherein the multiple prediction tasks include at least two of a credit limit prediction task, a spending prediction task and an overdue prediction task. After inputting corresponding user feature data into the multiple expert networks, the expert network features of the prediction task can be obtained through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network, and then the prediction results of each prediction task can be obtained through the expert network features of each prediction task and the tower network, for example, the approved credit limit prediction result can be obtained through the credit limit prediction task, the spending prediction result of the user can be obtained through the spending prediction task, and the overdue rate prediction result of the target user can be obtained through the overdue prediction task.

[0092] Moreover, because in the process of obtaining the prediction results of each prediction task, in addition to using the features output by the task expert network of each prediction task itself (the first output features), the features output by the shared expert network (the second output features) are also used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results (approved amount prediction results, utilization rate prediction results and overdue rate prediction results), and thus, it can also more accurately determine the loan provider for the target user.

[0093] See also Fig.10 , a data processing method provided in an embodiment of the present application, the method comprising:

[0094] S410: Inputting corresponding user feature data and object feature data into multiple expert networks in the multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of respective multiple prediction tasks, and the user feature data includes data related to the target user.

[0095] S420: Obtaining expert network features of the prediction task through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network.

[0096] S430: Obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

[0097] It should be noted that in addition to inputting corresponding user feature data into the expert network in this embodiment, corresponding object feature data may also be input. The object feature data may be understood as object feature data of the recommended object to be predicted. In this case, the prediction result obtained for each prediction task may be understood as the prediction result for the recommended object to be predicted.

[0098] Optionally, a plurality of candidate recommendation objects may be determined in advance. In this case, a corresponding multi-task prediction result may be determined for each candidate recommendation object in turn, and the multi-task prediction result includes a prediction result of each prediction task.

[0099] For example, if multiple prediction tasks include a quota prediction task, a withdrawal prediction task, and an overdue prediction task, and the recommended object is a loan provider, and multiple candidate recommended objects include a loan provider D1, a loan provider D2, and a loan provider D3. Then, multi-task prediction results can be obtained for the loan provider D1, the loan provider D2, and the loan provider D3, respectively. In the process of obtaining the multi-task prediction result for the recommended object D1, the object feature data corresponding to the recommended object D1 can be input into the expert network, and in the obtained multi-task results, the prediction result of the quota prediction task represents the target user's quota approval status for the loan provider D1, the prediction result of the withdrawal prediction task represents the target user's withdrawal status for the loan provided by the loan provider D1, and the prediction result of the overdue prediction task represents the target user's overdue status for the loan provided by the loan provider D1.

[0100] Similarly, in the process of obtaining the multi-task prediction results for the recommended object D2, the object feature data corresponding to the recommended object D2 can be input into the expert network, and in the obtained multi-task results, the prediction result of the credit limit prediction task represents the credit limit approval status of the target user for the loan provider D2, the prediction result of the expenditure prediction task represents the expenditure status of the target user for the loan provided by the loan provider D2, and the prediction result of the overdue prediction task represents the overdue status of the target user for the loan provided by the loan provider D2.

[0101] Among them, in the process of obtaining the multi-task prediction results for the recommended object D3, the object feature data corresponding to the recommended object D3 can be input into the expert network, and in the obtained multi-task results, the prediction result of the credit limit prediction task represents the credit limit approval status of the target user for the loan provider D3, the prediction result of the expenditure prediction task represents the expenditure status of the target user for the loan provided by the loan provider D3, and the prediction result of the overdue prediction task represents the overdue status of the target user for the loan provided by the loan provider D3.

[0102] It should be noted that, in the embodiment of the present application, the process of recommending the target user to the corresponding recommended object can also be understood as a process of traffic distribution, that is, distributing the user as a kind of traffic to each recommended object. In this embodiment, because the input data of the multi-target network can include object feature data, the recommended object of each target user can be determined directly based on the multi-task prediction results of each target user.

[0103] Exemplarily, the candidate recommendation object is a loan provider, and the multiple candidate recommendation objects include loan provider D1, loan provider D2 and loan provider D3. The multiple prediction tasks may include a credit limit prediction task, a expenditure prediction task and an overdue prediction task. Then for the target user, the corresponding multi-task prediction results may include the prediction results of the credit limit prediction task, the prediction results of the expenditure prediction task and the prediction results of the overdue prediction task. Furthermore, the corresponding loan provider can be determined for the target user from these prediction results from loan provider D1, loan provider D2 and loan provider D3.

[0104] Among them, after obtaining the prediction results of the credit limit prediction task, the expenditure prediction task and the overdue prediction task, the benefits of recommending the target user to different loan providers can also be calculated, and then the loan provider with the largest benefit can be used as the loan provider corresponding to the target user. Among them, the formula for calculating the benefit can be: credit limit × expenditure rate × (1-overdue rate) × interest rate × profit sharing ratio. Among them, the credit limit is the approved credit limit predicted by the credit limit prediction task. Exemplarily, if the benefit of recommending the target user to the loan provider D1 is calculated based on this formula as benefit S1, the benefit of recommending the target user to the loan provider D2 is benefit S2, and the benefit of recommending the target user to the loan provider D3 is benefit S3, if S3 is the largest, the loan provider D3 will be used as the loan provider corresponding to the target user, and the target user will be recommended to the loan provider D3.

[0105] This embodiment provides a data processing method, so that the expert network features input into the tower network can include richer information through the above method, so as to improve the accuracy of the prediction results, and further, the recommended objects can be determined more accurately for the target users. Moreover, in this embodiment, the input data input into each expert network can include object feature data in addition to the corresponding user feature data, so that in this embodiment, each task prediction result obtained can be understood as a task prediction result for each recommended object, and further, the recommended object corresponding to the target user can be determined directly based on the prediction results output by the multi-objective network.

[0106] In addition, as a method, if the input data input into the multi-target network does not include object feature data, after obtaining the prediction result of each prediction task, the prediction result of each prediction task can be matched with the recommendation conditions provided by the candidate recommendation object, so as to use the candidate recommendation object that can successfully match the prediction result of each task as the recommendation object corresponding to the target user.

[0107] The recommendation condition for each recommended object may include requirements for the prediction results of each prediction task.

[0108] Exemplarily, the candidate recommended objects include recommended object T1, recommended object T2, and recommended object T3, and the prediction results of each prediction task may include the prediction results of prediction task C, prediction task W, and prediction task R. In this case, the recommendation condition of each recommended object may include the requirements of the prediction results of the corresponding prediction task C, prediction task W, and prediction task R, so that the recommended object corresponding to the target user can be determined by the requirements. For example, if the prediction results of each task of the target user match the requirements in the recommendation conditions of the recommended object T1, the target user can be recommended to the recommended object T1, or the recommended object T1 can be recommended to the target user.

[0109] It should be noted that in the embodiment of the present application, the multi-objective network used to execute the data processing method is obtained in advance through training. During the training process, the loss function of the network (expert network or tower network) corresponding to some prediction tasks (for example, the aforementioned amount prediction task and expenditure prediction task) can be set to the Huber loss function. Among them, the Huber loss function combines the characteristics of mean square error loss and mean absolute error loss. It is a quadratic function when the error is small and a linear function when the error is large. The Huber loss function also has the following advantages: when the error is small, the smoothness characteristic of the mean square error loss is retained; when the error is large, the mean absolute error loss is insensitive to outliers; it has the characteristic of being differentiable near 0, which is conducive to optimizing the stability of the algorithm. By using the Huber loss function, the error processing can be better balanced, while reducing the impact of outliers on model training, thereby improving the robustness and generalization ability of the model. The expression of the Huber loss function is as follows:

[0110]

[0111] In the above expression, a represents the difference between the true value and the predicted value, and δ is a hyperparameter that controls the turning point of the loss function.

[0112] In an embodiment of the present application, in order to solve the problem that the neural network (multi-target network) is not well-fitted to the regression problem, a pre-training mechanism is added. In this case, the XGBoost model can be used for pre-training, and then its leaf nodes are introduced as features in the training process of the multi-target network, which helps to alleviate the problem of poor fit of the neural network. Among them, XGBoost is a gradient boosting algorithm based on a decision tree, which performs well in regression and classification problems. The leaf nodes of the XGBoost model are input into the neural network as features, which can provide richer information and feature representation, which helps the neural network to better learn the complex relationship of data.

[0113] In an embodiment of the present application, in order to solve the problem that neural networks (multi-target networks) are prone to overfitting, a Dropout mechanism can be introduced in the hidden layer (including the hidden layer in the expert network and the tower network). The Dropout mechanism is a commonly used neural network regularization method. During the model training phase, it randomly ignores some neurons with a certain probability to reduce the joint adaptive relationship between neurons. The Dropout mechanism forces the neurons in the network not to depend on specific neurons, thereby forcing the network to learn more robust and generalized feature representations. This randomness helps prevent the network from over-relying on certain features, improves the generalization ability of the model, and reduces the risk of overfitting. Among them, the effect of the Dropout mechanism can be as follows: Fig.11 As shown. Among them, Fig.11 The circles in the figure represent neurons.

[0114] It should be noted that in the embodiments of the present application, the specific examples of multiple prediction tasks introduced (for example, amount prediction tasks, expenditure prediction tasks, and overdue prediction tasks) are only for illustrative purposes. The application does not limit the fields or scenarios in which the multi-objective network is applied. For example, the multiple prediction tasks applied by the multi-objective network in the embodiments of the present application can also be used in video recommendation scenarios, product recommendation scenarios, etc. Among them, when the data processing method provided in the embodiments of the present application is applied to different scenarios, the required user feature data may be different, and the prediction tasks configured for the task expert network may be different.

[0115] See also Fig.12 , an embodiment of the present application provides a data processing device 500, the device 500 comprising:

[0116] The data input unit 510 is used to input corresponding user feature data into multiple expert networks in the multi-target network, wherein the multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks, and the user feature data includes data related to the target user, wherein the prediction target of each prediction task is different.

[0117] The feature calculation unit 520 is used to obtain the expert network feature of the prediction task through the first output feature output by the task expert network of the prediction task and the second output feature output by the shared expert network.

[0118] The prediction unit 530 is used to obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

[0119] As a method, the recommended objects include loan provision objects, and the multiple prediction tasks include a loan amount prediction task. The activation function of the tower network of the loan amount prediction task includes a linear part and a zero value part. The linear part is used to output the approved loan amount of the target user when the loan passes the review, and the zero value part is used to output that the target user fails the loan review.

[0120] As a method, the task expert network includes multiple first subnetworks. In this method, the feature calculation unit 520 is also used to determine the weights of the features output by the multiple first subnetworks in the task expert network, wherein the features output by the multiple first subnetworks are different; based on the features output by the multiple first subnetworks in the task expert network, and the weight of each feature, the first output feature output by the task expert network is obtained. Optionally, the multi-target network also includes a gate control network, and the weights of the features output by the multiple first subnetworks are determined by the gate control network based on the features output by the multiple first subnetworks. Optionally, the shared expert network includes multiple second subnetworks, and the weights of the features output by the multiple first subnetworks are determined by the gate control network based on the features output by the multiple first subnetworks and the features output by the multiple second subnetworks.

[0121] As a method, the feature calculation unit 520 is specifically used to obtain the expert network feature of each prediction task through the first output feature output by the task expert network of the prediction task, the weight corresponding to the first output feature, the second output feature output by the shared expert network, and the weight of the second output feature.

[0122] The present embodiment provides a data processing device, so that in the process of obtaining the prediction result of each prediction task, in addition to using the features output by the task expert network of each prediction task itself (first output features), the features output by the shared expert network (second output features) are also used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results, and further, it is also possible to more accurately determine the recommended objects for the target users.

[0123] It should be noted that the device embodiment in the present application corresponds to the aforementioned method embodiment. The specific principles in the device embodiment can be found in the contents of the aforementioned method embodiment and will not be repeated here.

[0124] The following will be combined Fig.13 An electronic device provided by the present application is described.

[0125] See also Fig.13 Based on the above data processing method and device, the embodiment of the present application also provides another electronic device 1000 that can execute the above data processing method. The electronic device 1000 includes one or more (only one is shown in the figure) processors 102, a memory 104, and a network module 106 that are coupled to each other. Among them, the memory 104 stores a program that can execute the content of the above embodiment, and the processor 102 can execute the program stored in the memory 104.

[0126] Among them, the processor 102 may include one or more processing cores. The processor 102 uses various interfaces and lines to connect various parts of the entire electronic device 1000, and executes various functions and processes data of the electronic device 1000 by running or executing instructions, programs, code sets or instruction sets stored in the memory 104, and calling data stored in the memory 104. Optionally, the processor 102 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA), or the processor 102 can be an application processor (Application Processor). The processor 102 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 102, but may be implemented separately through a communication chip.

[0127] The memory 104 may include a random access memory (RAM) or a read-only memory (ROM). The memory 104 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 104 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the following various method embodiments, etc.

[0128] The network module 106 is used to implement information exchange between the electronic device 1000 and other devices, for example, to transmit device control instructions, operation request instructions, and status information acquisition instructions, etc. When the electronic device 1000 is specifically a different device, the corresponding network module 106 may be different.

[0129] In addition, in one case, the electronic device 1000 may also include a screen, a sensor module, and an audio acquisition device.

[0130] The screen can be used to display the screen. For example, when a game program is running, it can be used to display the game screen corresponding to the game program. In one embodiment, the electronic device 1000 can capture multiple frames of screens continuously displayed on the screen to obtain multiple frames of screens required for detecting whether the electronic device 100 is displaying the event animation.

[0131] The sensor module may include at least one sensor. Specifically, the sensor module may include but is not limited to: a light sensor, a motion sensor, a pressure sensor, an infrared heat sensor, a distance sensor, an acceleration sensor, and other sensors. Among them, the pressure sensor can be a sensor that detects the pressure generated by pressing on the electronic device 1000. That is, the pressure sensor detects the pressure generated by the contact or pressing between the user and the audio playback device, such as the pressure generated by the contact or pressing between the user's ear and the mobile terminal. Therefore, the pressure sensor can be used to determine whether contact or pressing occurs between the user and the electronic device 1000, and the magnitude of the pressure. Among them, the acceleration sensor can detect the magnitude of acceleration in each direction (generally three axes), and can detect the magnitude and direction of gravity when stationary, which can be used to identify the application of the posture of the electronic device 1000 (such as horizontal and vertical screen conversion, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, knocking), etc. In addition, the electronic device 1000 can also be configured with other sensors such as gyroscopes, barometers, hygrometers, thermometers, etc., which will not be repeated here.

[0132] The audio acquisition device is used to acquire audio signals. Optionally, the audio acquisition device includes a plurality of audio acquisition devices. The audio acquisition device may be a microphone.

[0133] As a method, the network module of the electronic device 1000 is a radio frequency module, which is used to receive and send electromagnetic waves, realize the mutual conversion between electromagnetic waves and electrical signals, and thus communicate with a communication network or other devices. The radio frequency module may include various existing circuit elements for performing these functions, such as antennas, radio frequency transceivers, digital signal processors, encryption / decryption chips, user identity modules (SIM) cards, memories, etc. For example, the radio frequency module can exchange information with an external device through the electromagnetic waves sent or received, and then receive audio signals sent by the external device.

[0134] Please refer to Fig.14 , which shows a structural block diagram of a computer-readable storage medium provided in an embodiment of the present application. The computer-readable medium 1100 stores program codes, which can be called by a processor to execute the method described in the above method embodiment.

[0135] The computer readable storage medium 1100 may be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read-only memory), an EPROM, a hard disk, or a ROM. Optionally, the computer readable storage medium 1100 includes a non-transitory computer-readable storage medium. The computer readable storage medium 1100 has storage space for program code 1110 that executes any method step in the above method. These program codes can be read from or written to one or more computer program products. The program code 1110 can be compressed, for example, in an appropriate form.

[0136] In summary, the present application provides a data processing method, device and electronic device, in which, in the multi-target network involved, multiple expert networks include a shared expert network and task expert networks of multiple prediction tasks. After inputting corresponding user feature data into the multiple expert networks, the expert network features of the prediction task can be obtained through the first output features output by the task expert network of the prediction task and the second output features output by the shared expert network, and then the prediction results of each prediction task can be obtained through the expert network features of each prediction task and the tower network. Thus, in the process of obtaining the prediction results of each prediction task, in addition to using the features (first output features) output by the task expert network of each prediction task itself, the features (second output features) output by the shared expert network will also be used, so that the expert network features input into the tower network can include richer information, which is conducive to improving the accuracy of the prediction results, and thus, it is also possible to more accurately determine the recommended objects for the target users.

[0137] As a method, this application proposes a multi-objective credit field traffic distribution method. Compared with traditional practices, by introducing a mechanism for sharing some parameters, it is possible to better capture common features and related information in the data, thereby improving the accuracy and generalization of predictions. Compared with traditional models, this method is more accurate in prediction. At the algorithmic level, the loss function and activation function are optimized to make them more consistent with biological neural activation mechanisms and credit traffic distribution scenarios, while avoiding complex exponential operations, which can improve model training efficiency. Adding the XGBoost pre-training step can effectively alleviate the fitting problem during regression prediction and improve model performance. In addition, the introduction of the Dropout mechanism can alleviate the overfitting problem during binary classification prediction, enhance generalization ability, and enable the model to maintain performance on cross-period samples.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical feature diagrams therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data processing method, characterized in that: The method comprises: Inputting corresponding user feature data into a plurality of expert networks in the multi-target network, wherein the plurality of expert networks include a shared expert network and task expert networks of respective plurality of prediction tasks, and the user feature data include data related to the target user; Obtaining an expert network feature of the prediction task by using a first output feature of the task expert network output of the prediction task and a second output feature of the shared expert network output; The prediction result of each prediction task is obtained through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

2. The method according to claim 1, characterized in that The recommended objects include loan provision objects, the multiple prediction tasks include a loan amount prediction task, the activation function of the tower network of the loan amount prediction task includes a linear part and a zero value part, the linear part is used to output the approved loan amount of the target user if the target user passes the loan review, and the zero value part is used to output that the target user has not passed the loan review.

3. The method according to claim 1, characterized in that The task expert network includes a plurality of first sub-networks, and the method further includes: Determining weights of features output by a plurality of first subnetworks in the task expert network, wherein the features output by the plurality of first subnetworks are different; Based on the features output by the multiple first sub-networks in the task expert network and the weight of each feature, a first output feature of the task expert network output is obtained.

4. The method according to claim 3, characterized in that The multi-objective network also includes a gate control network, and the weights of the features output by the multiple first sub-networks are determined by the gate control network based on the features output by the multiple first sub-networks.

5. The method according to claim 4, characterized in that The shared expert network includes multiple second sub-networks, and the weights of the features output by the multiple first sub-networks are determined by the gate control network based on the features output by the multiple first sub-networks and the features output by the multiple second sub-networks.

6. The method according to claim 1, characterized in that The expert network feature of the prediction task is obtained by predicting the first output feature of the task expert network output of the task and the second output feature of the shared expert network output, including: The expert network feature of each prediction task is obtained by predicting the first output feature output by the task expert network of the task, the weight corresponding to the first output feature, the second output feature output by the shared expert network, and the weight of the second output feature.

7. The method according to claim 1, characterized in that The multiple prediction tasks include at least two of a credit limit prediction task, a spending prediction task, and an overdue prediction task, and the recommended object includes a loan provider recommended to the target user, or includes a loan provider recommended to the target user.

8. A data processing device, characterized in that: The device comprises: A data input unit, used to input corresponding user feature data into a plurality of expert networks in the multi-target network, wherein the plurality of expert networks include a shared expert network and task expert networks of respective plurality of prediction tasks, and the user feature data includes data related to the target user; A feature calculation unit, used to obtain an expert network feature of the prediction task through a first output feature output by the task expert network of the prediction task and a second output feature output by the shared expert network; The prediction unit is used to obtain the prediction result of each prediction task through the expert network features and the tower network of each prediction task, so as to determine the recommended object corresponding to the target user through the prediction result of each prediction task.

9. An electronic device, characterized in that: including a processor and a memory; One or more programs are stored in the memory and configured to be executed by the processor to implement the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The computer program product comprises computer instructions, and the computer instructions are stored in a computer-readable storage medium; the computer program product is used to implement the method according to any one of claims 1 to 7.

Images