A training method, apparatus, and electronic equipment for a multi-task risk prediction model.

Abstract

This application provides a training method, apparatus, and electronic device for a multi-task risk prediction model. For predicting the probability of an event occurring at different time periods, the core of the multi-task risk prediction model provided in this application includes two layers: a general information extraction layer and a specific information extraction layer. The specific information extraction layer includes several expert networks corresponding to different prediction period nodes. Using the embodiments of this application, risk probability prediction for different prediction period nodes can be achieved without creating multiple independent models, effectively saving the hardware resources required for the prediction model. Furthermore, by penalizing the target loss function with a group constraint loss function, the multi-task risk prediction model trained based on this target loss function can accurately identify the group distribution of the input data, thereby outputting risk prediction results that conform to the group distribution, thus ensuring the accuracy and stability of the prediction results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method, apparatus and electronic device for a multi-task risk prediction model. Background Technology

[0002] With the development of artificial intelligence technology, especially the development of large-scale AI models, intelligent models are being widely applied in various industries and scenarios. In some practical applications, there is a need for an intelligent model to simultaneously predict the development of events across different time periods or stages, calculating the probability of a particular event occurring at each time period. For example, in the lending sector, intelligent models are needed to predict the probability of user delinquency risk for different loan durations; similarly, in the insurance sector, intelligent models are needed to predict the probability of users developing serious illnesses at different age levels.

[0003] Traditional methods for predicting the probability of an event occurring at different time periods often involve building separate intelligent models for each time period. These models then predict the data for their respective time periods and output the corresponding probability values. However, in practical applications, this approach is problematic. First, training and deploying multiple independent intelligent models consumes significant hardware resources. Second, because the constructed intelligent models are independent of each other, the predicted results fail to reflect the overall distribution of the data, resulting in poor accuracy. Summary of the Invention

[0004] In view of this, embodiments of this application provide a training method, apparatus, and electronic device for a multi-task risk prediction model, which saves the hardware resources required for training and deploying the multi-task risk prediction model and improves the prediction accuracy of the constructed multi-task risk prediction model.

[0005] In a first aspect, embodiments of this application provide a training method for a multi-task risk prediction model, wherein the multi-task risk prediction model includes at least: a general information extraction layer and a proprietary information extraction layer, wherein the proprietary information extraction layer includes: an expert network corresponding to several nodes with different prediction periods, and the method includes: The original information of each object to be analyzed is obtained and input into the general information extraction layer; Obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer to obtain the risk probability value output by each of the expert networks; Based on the risk probability values ​​of the N objects to be analyzed in the same batch at each of the prediction period nodes, calculate the average risk probability value of the group corresponding to each of the prediction period nodes. Based on a preset target loss function, the model parameters of the general information extraction layer and the proprietary information extraction layer are adjusted until the target loss function converges. Satisfy the following formula: ; in, For multi-task loss function, Let be the group constraint loss function, and λ be the adjustment weight of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

[0006] Secondly, embodiments of this application provide an apparatus for constructing a multi-task risk prediction model, wherein the multi-task risk prediction model includes at least: a general information extraction layer and a proprietary information extraction layer, wherein the proprietary information extraction layer includes: an expert network corresponding to several nodes with different prediction periods, and the apparatus includes: The data preprocessing module is used to obtain the raw information of each object to be analyzed and input it into the general information extraction layer; A general feature extraction module is used to obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer. A proprietary information extraction module is used to calculate the risk probability value output by each of the expert networks, and to calculate the group average risk probability value corresponding to each prediction period node based on the risk probability values ​​of the N objects to be analyzed in the same batch at each prediction period node. The model parameter adjustment module is used to adjust the model parameters of the general information extraction layer and the proprietary information extraction layer based on a preset target loss function until the target loss function converges. Satisfy the following formula: ; in, For multi-task loss function, Let be the group constraint loss function, and λ be the adjustment weight of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

[0007] Thirdly, embodiments of this application provide an electronic device, wherein the electronic device includes: a processor; and a memory storing a program; wherein the program includes instructions, which, when executed by the processor, cause the processor to perform the training method of the multi-task risk prediction model described in the first aspect.

[0008] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the training method of the multi-task risk prediction model described in the first aspect.

[0009] The beneficial effects of this application are: This application provides a training method, apparatus, and electronic device for a multi-task risk prediction model. For the purpose of predicting the probability of an event occurring at different time periods, this application creates a multi-task risk prediction model. The core of this multi-task risk prediction model includes two layers: a general information extraction layer and a proprietary information extraction layer. The proprietary information extraction layer includes several expert networks corresponding to nodes with different prediction periods.

[0010] During the training of this multi-task risk prediction model, the original information of each object to be analyzed is acquired and input into the general information extraction layer. This layer extracts a general risk representation vector from the original information and inputs it into the specialized information extraction layer. Expert networks within the specialized information extraction layer calculate the risk probability value for each prediction period node based on the general risk representation vector. Then, based on the risk probability values ​​of N objects to be analyzed in the same batch, the group average risk probability value corresponding to each prediction period node is calculated. This average risk probability value is then used to determine whether it conforms to a preset trend of group average risk probability change, thus determining the penalty strength of the target loss function for the entire model. This penalty strength is represented by a group constraint loss function, which is used to adjust the target loss function. Finally, the model parameters of the general information extraction layer and the specialized information extraction layer are adjusted based on the target loss function until the target loss function converges, achieving the desired training effect for the multi-task risk prediction model.

[0011] By using the embodiments of this application, risk probabilities for different prediction period nodes can be predicted without creating multiple independent models, effectively saving the hardware resources required for the prediction model. Simultaneously, by penalizing the target loss function with a group constraint loss function, the risk differences and group distribution characteristics at different prediction period nodes can be accurately captured. The multi-task risk prediction model trained based on this target loss function can accurately identify the group distribution of the input data, thereby outputting risk prediction results that conform to the group distribution, thus ensuring the accuracy and stability of the prediction results. Attached Figure Description

[0012] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating a training method for the multi-task risk prediction model provided in this application is shown. Figure 2 This paper illustrates another flowchart of the training method for the multi-task risk prediction model provided in this application. Figure 3 A block diagram of a training device for the multi-task risk prediction model provided in this application is shown. Figure 4 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of this application is shown. Detailed Implementation

[0013] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0015] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0016] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] To reduce the hardware resource consumption of risk prediction models and improve their ability to handle population distribution to output prediction results that conform to the population distribution, thereby improving the prediction accuracy of risk prediction models, this application provides a training method, apparatus, and electronic device for a multi-task risk prediction model. In one aspect, this application provides a training method for a multi-task risk prediction model, applicable to any electronic device capable of constructing a multi-task risk prediction model, including but not limited to personal mobile terminals, computers, or servers. This application provides a novel multi-task risk prediction model comprising only two layers: a general information extraction layer and a proprietary information extraction layer. The proprietary information extraction layer has an expert network set up for each prediction cycle node. Compared to traditional multiple independent and unrelated risk prediction models, the multi-task risk prediction model provided in this application is smaller in scale and consumes fewer hardware resources.

[0018] In some possible embodiments, such as Figure 1 As shown, the method includes the following steps S11-S14: S11. Obtain the original information of each object to be analyzed and input it into the general information extraction layer; S12. Obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer to obtain the risk probability value output by each of the expert networks. S13. Based on the risk probability values ​​of the N objects to be analyzed in the same batch at each prediction period node, calculate the average risk probability value of the group corresponding to each prediction period node. S14. Based on the preset target loss function, adjust the model parameters of the general information extraction layer and the proprietary information extraction layer until the target loss function converges.

[0019] Wherein, the target loss function Satisfy the following formula: ; in, For multi-task loss function, For the group constraint loss function, Here, are the adjustment weights of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

[0020] By using the embodiments of this application, risk probabilities for different prediction period nodes can be predicted without creating multiple independent models, effectively saving the hardware resources required for the prediction model. Simultaneously, by penalizing the target loss function with a group constraint loss function, the risk differences and group distribution characteristics at different prediction period nodes can be accurately captured. The multi-task risk prediction model trained based on this target loss function can accurately identify the group distribution of the input data, thereby outputting risk prediction results that conform to the group distribution, thus ensuring the accuracy and stability of the prediction results.

[0021] The following will provide a detailed explanation of steps S11 to S14 with specific examples: The multi-task risk prediction model provided in this application can be applied to any scenario where the risk probability of the same analytical object needs to be predicted at different prediction period nodes. For example, in the credit business scenario, the multi-task risk prediction model is used to predict the credit delinquency probability of a credit user in several future time periods; in the medical business scenario, the multi-task risk prediction model is used to predict the illness probability of a patient in several future time periods; and in the supply chain logistics business scenario, the multi-task risk prediction model is used to predict the probability of a certain item of goods experiencing timeout risk at various logistics turnover nodes.

[0022] In some possible embodiments, to facilitate an intuitive understanding of the multi-task risk prediction model provided in this application, it can be combined with, for example... Figure 2The flowchart shown illustrates that, in this application, the multi-task risk prediction model includes an input layer, a general information extraction layer, and a proprietary information extraction layer. Based on this, during the execution of steps S11 and S12, the method provided in this application further includes the following steps: Step 1: Obtain the original information through the input layer, perform feature classification on the structured features in the original information, convert the original information into a fixed-dimensional feature vector according to different types of features, and input the fixed-dimensional feature vector into the general information extraction layer; Step 2: Extract features from the fixed-dimensional feature vector through the general information extraction layer, extract risk information associated with each prediction period node from the fixed-dimensional feature vector, and generate the general risk representation vector based on the risk information.

[0023] It can be understood that this input layer is used to obtain the raw information of the object to be analyzed, such as... Figure 2 As shown, each input layer contains the original information of an object to be analyzed. The original information of N objects to be analyzed can be input in the same batch, thus enabling batch processing of risk prediction for the objects to be analyzed.

[0024] Due to the different sources of the original information, the original information may generate various data structures, specifically including numerical, categorical, sequential, and Boolean data. This input layer, after acquiring the original information, performs feature analysis based on the structured features within the original information, converting the original information into fixed-dimensional feature vectors according to different feature types.

[0025] For example, assuming this multi-task risk prediction model is used in the credit business field, the object to be analyzed can be a credit business user. The raw information of the object to be analyzed can be various credit-related data of the credit business user, including: occupation, age, gender, transaction behavior, credit report, etc. The input layer adapts and standardizes the raw information to obtain a feature vector. This feature vector is a fixed-dimensional feature vector.

[0026] Specifically, the input layer performs feature classification on structured features, mainly including: dividing numerical features, categorical features, sequence features, and Boolean features, among which: For numerical features, such as age, income, transaction amount, and number of overdue payments, normalization or standardization processing (such as Z-score standardization, Min-Max normalization, etc.) is performed to form feature vectors to prevent different scales from affecting the model's convergence speed and training stability.

[0027] For categorical features, such as gender, occupation type, marital status, and city level of the user, encoding processing is performed. Common methods include one-hot encoding, label encoding, or embedding encoding as feature vectors.

[0028] For sequence-based features, such as user behavior sequences, transaction logs, and loan repayment time series, the sequences are preserved in chronological order, and the length is truncated or padded according to model requirements to generate feature vectors.

[0029] For Boolean features, such as whether a user is a first-time borrower, whether an account is active, or whether a payment has been overdue, they can be used directly or encoded as 0 / 1 values ​​to generate a feature vector.

[0030] In this embodiment, the feature vectors generated from all types of features maintain the same dimension, and missing terms are padded with zeros. After the above processing, all features are uniformly converted into a fixed-dimensional real-valued vector. D represents the number of dimensions. This fixed-dimensional feature vector not only preserves the original information, but also improves the model's expressive and generalization capabilities through normalization, embedding, and other operations, providing high-quality input data for subsequent general information extraction layers and proprietary information extraction layers.

[0031] Since the overdue true labels of different prediction period nodes are different, but the data being processed has consistency and commonalities, the general information extraction layer provided in this application is used to extract the common content between the inputs corresponding to different prediction period nodes from the fixed-dimensional feature vector. It can be figuratively understood that this general information extraction layer is used to find commonalities in the input information.

[0032] For example, taking a multi-task risk prediction model applied to a credit business scenario, the general information extraction layer is used to extract underlying risk information that is common to both the credit loan cycle (also known as the credit aging Mob) from a large number of user features contained in the user's fixed-dimensional vector, thereby avoiding duplication of effort. Specifically, when various expert networks predict the probability of delinquency in the next few months, only the real labels differ; the input is the user's credit-related information. Therefore, the input information has commonalities. In this application, the general information extraction layer extracts the common content from the original information, extracts the common underlying feature representations for each prediction cycle node, and generates a general risk feature vector based on this common content. In this way, it avoids performing similar feature learning at each prediction cycle node, thereby improving information sharing efficiency.

[0033] In this application, the core of the general information extraction layer is to perform a nonlinear transformation on a fixed-dimensional feature vector using a multi-layer sensing mechanism (MLP). Through weight calculation and activation functions, the scattered features in the fixed-dimensional feature vector are integrated into a set of general feature vectors. For example, this feature vector can be a vector containing 30 numbers, where 10 numbers represent the user's basic credit score, 10 numbers represent the user's repayment ability level, and 10 numbers represent the user's credit rating. As one implementation method, step 2 above can be achieved through the following steps: The general risk representation vector is calculated using the multilayer perceptron (MLP) mechanism pre-built in the general information extraction layer, according to the following formula. : ; Where y is the fixed-dimensional feature vector processed and output by the input layer. Here, W1 is the weight of the first layer of the general information extraction layer, b1 is the bias coefficient of the first layer of the general information extraction layer, W2 is the weight of the second layer of the general information extraction layer, and b2 is the bias coefficient of the second layer of the general information extraction layer. The specific values ​​of W1, b1, W2, and b2 depend on the model training effect. That is, through step S14 above, the specific values ​​of W1, b1, W2, and b2 are continuously adjusted based on whether the target loss function converges, until the target loss function converges.

[0034] Furthermore, following step S13, the general risk representation vector is input into the proprietary information extraction layer. Each expert network in the proprietary information extraction layer calculates the risk probability value for its respective prediction period node, obtaining the risk probability value output by each expert network. The core function of each expert network is to accurately capture the risk for its assigned prediction period.

[0035] In this layer, the proprietary information extraction layer contains a corresponding number of expert networks based on the number of prediction periods required. Each expert network is only responsible for calculating the risk probability of the nodes in its assigned prediction period. For example, assuming the probability of tumor recurrence for a patient within the next 12 months needs to be predicted, the proprietary information extraction layer would have 12 expert networks, resulting in a total of 12 prediction periods. The first expert network would calculate the probability of tumor recurrence within the next month, the second expert network within the next two months, the third expert network within the next three months, and so on, with the xth expert network calculating the probability of tumor recurrence within the xth month.

[0036] For example, assuming this multi-task risk prediction model is applied to the credit business field, different expert networks are responsible for calculating the delinquency risk probability at different aging stages. Each expert network captures the unique risk characteristics of its respective stage and calculates the corresponding risk probability value. For instance, expert networks 1, 2, and 3 are responsible for calculating the short-term delinquency risk at the Mob1, Mob2, and Mob3 aging stages, respectively. In this case, expert networks 1, 2, and 3 capture information related to short-term repayment willingness in the general risk representation vector and calculate the delinquency probability as the risk probability value for the three prediction cycle nodes of Mob1, Mob2, and Mob3 based on this information.

[0037] In this application, each expert network is responsible for a single prediction period, and the input to each expert network is a general risk representation vector output by the general information extraction layer. Through separate weight calculations and activation functions, the risk probability value for each prediction period can be calculated. The risk probability value p output by each expert network is... x This can form a multi-stage risk value vector (or a list of risk probability values).

[0038] One implementation method involves the expert network calculating the risk probability value through the following steps: Using the expert network corresponding to the x-th prediction period node, the risk probability value corresponding to the x-th prediction period node is calculated based on the following formula. : ; in, The weights of the expert network corresponding to the x-th prediction period node are: Let be the bias coefficient of the expert network corresponding to the x-th prediction period node. Let this be the general risk representation vector. It is a sigmoid activation function. as well as The accuracy of the expert network's output is determined by the difference between the expert network's output and the true label. As one implementation method, this application uses a multi-task loss function. This multi-task loss function describes the difference between the expert network output and the true labels. The following formula constraints must be met:

[0039] In this formula, Let be the loss function for multiple prediction period nodes of the i-th object to be analyzed. Calculated based on the following formula: ; In this formula, Let be the true label value of the i-th object to be analyzed at the x-th prediction period node. Let M be the risk probability value of the i-th object to be analyzed at the x-th prediction period node, and M be the total number of prediction period nodes. This can be understood as the true label of the prediction stage corresponding to the i-th sample data at the x-th prediction period node. The value of is (0,1], corresponding to the risk probability value of the x-th prediction stage in the output. By summing the analysis results of all objects to be analyzed for the i-th sample, the multi-period crossover loss function under all prediction periods M can be obtained. Furthermore, through the formula The loss of N samples of the same batch of objects to be analyzed is summed and averaged. This multi-task loss function formula can optimize the error of the general information extraction layer and the proprietary information extraction layer when performing data classification. In this way, the general information extraction layer and the proprietary information extraction layer in the model can capture the risk differences corresponding to different prediction period stages, and better predict the risk of each prediction period.

[0040] Because risk is affected not only by individuals but also by changes in group distribution, meaning risk manifests not only at the individual level but also exhibits significant structural characteristics at the group level, the applicant, through statistical analysis of historical samples' risk prediction probability values ​​in practical application, discovered that users' risk probability values ​​at different prediction cycle nodes increase over time. Based on this, this application sets a group average risk probability change trend according to observed trends. Specifically, the preset group average risk probability change trend is that the corresponding risk probability value should increase as the prediction cycle node time increases. For the same batch of N objects to be analyzed, the preset group average risk probability change trend is that the risk probability value at each prediction cycle node increases over time; that is, the group average risk probability change trend is that the average risk probability value in that cycle stage increases as the preset cycle node time increases.

[0041] For example, if this multi-task risk prediction model is applied to the credit business, the probability of user delinquency shows an overall pattern of increasing over time at the account aging level, i.e. Therefore, this application requires guiding the model output to reflect a trend in the average risk probability of the group. Based on this, this application introduces a group-constrained loss function. The group constraint loss function guides the model's output to present prediction results that conform to the changing trend of the group's average risk probability. Specifically, in some possible embodiments, the group constraint loss function... Satisfy the following formula: ; in, Let x be the average risk probability value of the population at the x-th prediction period node. Let be the average risk probability value of the population at the (x+1)th prediction period node. The population constraint loss function can be viewed as a penalty term for the multi-task risk prediction model. This penalty forces adjustments to the model parameters, ensuring the model's output conforms to the monotonically increasing trend of the average population probability. Specifically, for adjacent prediction period nodes, if the increasing trend is not observed, a penalty is applied, with the penalty strength positively correlated with the difference between the two. ReLU is an activation function whose characteristic is that the output is 0 when the input ≤ 0, and equal to the input when the input > 0. If the value is ≤0, it indicates that the trend is not being followed. In this case, an output of 0 is used to penalize the entire objective loss function. >0, at this time the output result is In this way, the ReLU activation function can be used to filter out outputs that do not conform to the increasing trend, thus ensuring that the entire population constraint function is differentiable, and thereby guaranteeing that the model can be optimized through gradient descent.

[0042] Among them, the average risk probability value of the group It can be calculated using the following formula:

[0043] In this application, the group constraint loss function is used to correct biases in the model's output regarding the group distribution, thereby forcing the model to output predictions that conform to the preset trend of the group's average risk change. Through the constructed... The objective loss function serves as the guiding principle for the entire model training process, ensuring that the model can guarantee the prediction accuracy of individual objects under analysis while also ensuring that the model output conforms to the trend of the average risk probability of the group.

[0044] in, Used to adjust the proportion of the group constraint loss function in the entire objective loss function, for example A value of 0.3 indicates that the training objective for this multi-task risk prediction model is to prioritize the prediction accuracy of individual objects while also considering group dynamics. The larger the value of , the more necessary it is to consider the group dynamics. By iteratively training and continuously reducing the target loss function until it converges, the trained model can accurately calculate individual objects under analysis while also conforming to business development patterns.

[0045] Based on the method provided in the first aspect, in the second aspect, this application provides a construction apparatus for a multi-task risk prediction model. The multi-task risk prediction model includes at least: a general information extraction layer and a proprietary information extraction layer, wherein the proprietary information extraction layer includes: several expert networks corresponding to nodes with different prediction periods, such as... Figure 3 As shown, the device 30 includes: Data preprocessing module 301 is used to acquire the original information of each object to be analyzed and input it into the general information extraction layer; The general feature extraction module 302 is used to obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer. The proprietary information extraction module 303 is used to calculate the risk probability value output by each of the expert networks, and to calculate the group average risk probability value corresponding to each prediction period node based on the risk probability values ​​of the N objects to be analyzed in the same batch at each prediction period node. The model parameter adjustment module 304 is used to adjust the model parameters of the general information extraction layer and the proprietary information extraction layer based on a preset target loss function until the target loss function converges. Satisfy the following formula: ; in, For multi-task loss function, Let be the group constraint loss function, and λ be the adjustment weight of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

[0046] In some possible embodiments, the multi-task risk prediction model further includes: an input layer, wherein the data preprocessing module is specifically used for: The original information is obtained through the input layer, the structured features in the original information are classified, the original information is converted into a fixed-dimensional feature vector according to different types of features, and the fixed-dimensional feature vector is input into the general information extraction layer. The general information extraction layer extracts features from the fixed-dimensional feature vector, extracts risk information associated with each prediction period node from the fixed-dimensional feature vector, and generates the general risk representation vector based on the risk information.

[0047] In some possible embodiments, the group constraint loss function Satisfy the following formula:

[0048] in, Let x be the average risk probability value of the population at the x-th prediction period node. This represents the average risk probability value of the population at the (x+1)th prediction period node.

[0049] In some possible embodiments, the general feature extraction module is specifically used for: The general risk representation vector is calculated using the multi-layer perception mechanism pre-set in the general information extraction layer, according to the following formula. : ; Where y is the fixed-dimensional feature vector. The activation function is defined as follows: W1 is the weight of the first layer network in the general information extraction layer, b1 is the bias coefficient of the first layer network in the general information extraction layer, W2 is the weight of the second layer network in the general information extraction layer, and b2 is the bias coefficient of the second layer network in the general information extraction layer.

[0050] In some possible embodiments, the proprietary information extraction module is specifically used for: Using the expert network corresponding to the x-th prediction period node, the risk probability value corresponding to the x-th prediction period node is calculated based on the following formula. : ; in, The weights of the expert network corresponding to the x-th prediction period node are: Let be the bias coefficient of the expert network corresponding to the x-th prediction period node. Let be the general risk representation vector.

[0051] In some possible embodiments, the preset trend of the average risk probability of the group is that the risk probability value of each prediction period node increases over time.

[0052] In some possible embodiments, the multi-task loss function The following formula constraints must be met: ; in, Let be the loss function for multiple prediction period nodes of the i-th object to be analyzed. Calculated based on the following formula: ; in, Let be the true label value of the i-th object to be analyzed at the x-th prediction period node. Let be the risk probability value of the i-th object to be analyzed at the x-th prediction period node.

[0053] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application comply with relevant laws and regulations and do not violate public order and good morals.

[0054] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0055] Thirdly, exemplary embodiments of this application also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the electronic device to perform a method according to an embodiment of this application.

[0056] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of this application.

[0057] An exemplary embodiment of this application also provides a computer program product, including a computer program, wherein, when executed by a computer's processor, the computer program is used to cause the computer to perform a method according to an embodiment of this application.

[0058] refer to Figure 4The present invention describes a structural block diagram of an electronic device 400 that can serve as a server or client of this application, which is an example of a hardware device that can be applied to various aspects of this application. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0059] like Figure 4 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM 402) or a computer program loaded from a storage unit 408 into a random access memory (RAM 403). The RAM 403 may also store various programs and data required for the operation of the electronic device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output interface (I / O interface 405) is also connected to the bus 404.

[0060] Multiple components in electronic device 400 are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information to electronic device 400. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows electronic device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0061] The computing unit 401 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the aforementioned training method for the multi-task risk prediction model can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 400 via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the aforementioned training method for the multi-task risk prediction model by any other suitable means (e.g., by means of firmware).

[0062] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0063] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0064] As used in this application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0065] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0066] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0067] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

Claims

1. A training method for a multi-task risk prediction model, characterized in that, The multi-task risk prediction model includes at least: a general information extraction layer and a proprietary information extraction layer, wherein the proprietary information extraction layer includes: several expert networks corresponding to nodes with different prediction periods, and the method includes: The original information of each object to be analyzed is obtained and input into the general information extraction layer; Obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer to obtain the risk probability value output by each of the expert networks; Based on the risk probability values ​​of the N objects to be analyzed in the same batch at each of the prediction period nodes, calculate the average risk probability value of the group corresponding to each of the prediction period nodes. Based on a preset target loss function, the model parameters of the general information extraction layer and the proprietary information extraction layer are adjusted until the target loss function converges. Satisfy the following formula: ； in, For multi-task loss function, For the group constraint loss function, Here, are the adjustment weights of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

2. The method according to claim 1, characterized in that, The multi-task risk prediction model further includes an input layer, and the method further includes: The original information is obtained through the input layer, the structured features in the original information are classified, the original information is converted into a fixed-dimensional feature vector according to different types of features, and the fixed-dimensional feature vector is input into the general information extraction layer. The general information extraction layer extracts features from the fixed-dimensional feature vector, extracts risk information associated with each prediction period node from the fixed-dimensional feature vector, and generates the general risk representation vector based on the risk information.

3. The method according to claim 1, characterized in that, The group constraint loss function Satisfy the following formula: ； in, Let x be the average risk probability value of the population at the x-th prediction period node. Let be the average risk probability value of the population at the (x+1)th prediction period node, and ReLU be the activation function.

4. The method according to claim 2, characterized in that, The step of obtaining the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer includes: The general risk representation vector is calculated using the multi-layer perception mechanism pre-set in the general information extraction layer, according to the following formula. : ； Where y is the fixed-dimensional feature vector. The activation function is defined as follows: W1 is the weight of the first layer network in the general information extraction layer, b1 is the bias coefficient of the first layer network in the general information extraction layer, W2 is the weight of the second layer network in the general information extraction layer, and b2 is the bias coefficient of the second layer network in the general information extraction layer.

5. The method according to claim 1, characterized in that, The process of obtaining the risk probability values ​​output by each of the expert networks includes: Using the expert network corresponding to the x-th prediction period node, the risk probability value corresponding to the x-th prediction period node is calculated based on the following formula. : ； in, The weights of the expert network corresponding to the x-th prediction period node are: Let be the bias coefficient of the expert network corresponding to the x-th prediction period node. This is the general risk representation vector. It is a sigmoid activation function.

6. The method according to claim 1, characterized in that, The preset trend of the average risk probability of the group is that the risk probability value of each prediction period node increases over time.

7. The method according to claim 1, characterized in that, The multi-task loss function The following formula constraints must be met: ； in, Let be the loss function for multiple prediction period nodes of the i-th object to be analyzed. Calculated based on the following formula: ； in, Let be the true label value of the i-th object to be analyzed at the x-th prediction period node. Let M be the risk probability value of the i-th object to be analyzed at the x-th prediction period node, and M be the total number of prediction period nodes.

8. A training device for a multi-task risk prediction model, characterized in that, The multi-task risk prediction model includes at least: a general information extraction layer and a proprietary information extraction layer, wherein the proprietary information extraction layer includes: an expert network corresponding to several nodes with different prediction periods, and the device includes: The data preprocessing module is used to obtain the raw information of each object to be analyzed and input it into the general information extraction layer; A general feature extraction module is used to obtain the general risk representation vector corresponding to each of the objects to be analyzed generated by the general information extraction layer, and input the general risk representation vector into the proprietary information extraction layer. A proprietary information extraction module is used to calculate the risk probability value output by each of the expert networks, and to calculate the group average risk probability value corresponding to each prediction period node based on the risk probability values ​​of the N objects to be analyzed in the same batch at each prediction period node. The model parameter adjustment module is used to adjust the model parameters of the general information extraction layer and the proprietary information extraction layer based on a preset target loss function until the target loss function converges. Satisfy the following formula: ； in, For multi-task loss function, Let be the group constraint loss function, and λ be the adjustment weight of the group constraint loss function, where The function value is used to characterize the loss function for the target. The severity of the punishment The function value is positively correlated with the severity of the punishment, which is determined based on whether the average risk probability value of the group conforms to a preset trend of change in the average risk probability of the group.

9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing a program; wherein the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images