Data processing method, training method, prediction method and related products

Abstract

The invention discloses a data processing method, a training method, a prediction method and a related product. The data processing method is used for determining a target parameter of a target model, the target model is used for predicting a target service feedback result based on a target feature of a target object in a target object set, and the target service feedback result is a feedback result of sending a target service to the target object in the target object set. The data processing method comprises the following steps: acquiring m first loss values; determining a target relationship based on the m first loss values, wherein the target relationship is a relationship between the first loss values and the second loss values; based on the target relationship, determining a first loss value corresponding to the second loss value meeting the preset condition as a target loss value; and taking a parameter of the to-be-confirmed model corresponding to the target loss value as a target parameter.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a data processing method, a training method, a prediction method and related products. Background Art

[0002] By sending a service to an object, the feedback result of the service can be obtained, and the feedback result can be predicted by using a machine learning model. When the parameters of the machine learning model are different, the accuracy of the prediction result of the machine learning model is also different. Therefore, how to determine the parameters of the machine learning model used to predict the feedback result of the service is of great significance. Summary of the invention

[0003] The present application provides a data processing method, a training method, a prediction method and related products, wherein the related products include a data processing device, a training device, a prediction device, an electronic device, a computer-readable storage medium and a computer program product.

[0004] In a first aspect, a data processing method is provided, the method being used to determine a target parameter of a target model, the target model being used to predict a target business feedback result based on a target feature of a target object in a target object set, the target business feedback result being a feedback result of a target business sent to the target object in the target object set, the method comprising:

[0005] Obtain m first loss values, where m is an integer greater than 1, and the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, where n is an integer greater than 1, and the model structures of the m models to be confirmed are the same as the model structure of the target model, and different parameters of the models to be confirmed are different; the first loss value is obtained based on n first differences, and the first differences correspond to the training features one by one, and the first differences are the difference between the individual predicted feedback results of sending the target business to the training objects and the individual actual feedback results of sending the target business to the training objects, and the individual predicted feedback results are the feedback results of sending the target business to the training objects in the training object set predicted by the model to be confirmed;

[0006] Determine a target relationship based on the m first loss values, the target relationship being a relationship between the first loss value and a second loss value, the second loss value being obtained based on a second difference, the second difference corresponding one-to-one to the model to be confirmed, the second difference being a difference between a group prediction feedback result of sending the target service to a target group and a group actual feedback result of sending the target service to the target group, the target group being a group including the n training objects; the group prediction feedback result being obtained based on the n individual prediction feedback results of the n training objects;

[0007] Based on the target relationship, determining that the first loss value corresponding to the second loss value that meets a preset condition is a target loss value;

[0008] The parameters of the model to be confirmed corresponding to the target loss value are used as the target parameters.

[0009] In combination with any implementation manner of the present application, determining the target relationship based on the m first loss values ​​includes:

[0010] Sampling from the m first loss values ​​to obtain p sampled loss values, where p is an integer less than m;

[0011] For each of the p sampled loss values, respectively determine n individual predicted feedback results corresponding to the sampled loss value and n individual actual feedback results corresponding to the sampled loss value;

[0012] Determine the group prediction feedback result of the target group based on the n individual prediction feedback results corresponding to the same sampling loss value, and obtain p group prediction feedback results;

[0013] Determine the group actual feedback result of the target group based on the n individual actual feedback results corresponding to the same sampling loss value, and obtain p group actual feedback results;

[0014] Determine the second loss value based on the difference between the group prediction feedback result and the group actual feedback result corresponding to the same sampled loss value, and obtain p second loss values;

[0015] The target relationship is obtained based on the p second loss values ​​and the p sampled loss values.

[0016] In combination with any implementation manner of the present application, obtaining the target relationship based on the p second loss values ​​and the p sampling loss values ​​includes:

[0017] Obtaining a priori distribution of the first loss value and the second loss value, where the prior distribution represents a distribution of the second loss value corresponding to the first loss value;

[0018] The prior distribution is updated based on the p second loss values ​​and the p sampled loss values ​​to obtain a posterior distribution as the target relationship.

[0019] In combination with any implementation manner of the present application, the loss values ​​other than the p sampled loss values ​​among the m first loss values ​​are loss values ​​to be sampled, and determining, based on the target relationship, the first loss value corresponding to the second loss value that satisfies a preset condition as a target loss value includes:

[0020] Based on the target relationship, determining an acquisition function, the acquisition function being used to estimate a value of the first loss value being sampled;

[0021] Based on the acquisition function, determine the sampled value of the loss value to be sampled, and obtain (mp) sampling values;

[0022] Based on the (mp) sampling values, q sampling loss values ​​are obtained by sampling from the (mp) loss values ​​to be sampled, where q is an integer less than or equal to (mp);

[0023] Determine the second loss value that meets the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as a reference loss value;

[0024] The first loss value corresponding to the reference loss value is determined as the target loss value.

[0025] In combination with any embodiment of the present application, the acquisition function satisfies at least one of the following conditions: the mean of the second loss value corresponding to the first loss value in the target relationship is negatively correlated with the value of the first loss value sampled, the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value sampled, and the value of the sampled loss value in the first loss value is lower than the value of the loss value in the first loss value other than the sampled loss value.

[0026] In combination with any embodiment of the present application, the determining the second loss value that satisfies the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as the reference loss value includes:

[0027] When the sum of p and q is greater than or equal to the sampling threshold, the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​is determined as the reference loss value.

[0028] In combination with any implementation manner of the present application, the first loss value is obtained based on the sum of the squares of n first differences.

[0029] In a second aspect, a training method is provided, the method comprising:

[0030] A model to be trained is obtained, wherein the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, which are obtained according to the first aspect and any one of the embodiments thereof; n training features and n labels of n training objects are obtained, and the labels represent individual actual feedback results of sending the target business to the training objects; the n training features are input into the model to be trained so that the model to be trained predicts the feedback results of sending the target business to the training objects, and obtains n training feedback results; based on the difference between the n training feedback results and the n labels, the training loss of the model to be trained is obtained; based on the training loss, the parameters of the model to be trained are updated to obtain a prediction model.

[0031] A third aspect provides a prediction method, the method comprising:

[0032] Obtaining a set of features to be predicted, wherein the set of features to be predicted is a set of features to be predicted, and the features to be predicted are features of objects to be predicted in a group to be predicted;

[0033] Inputting the feature set to be predicted into a prediction model so that the prediction model predicts the individual prediction feedback result of sending the target business to the object to be predicted in the group to be predicted, and obtaining an individual prediction feedback result set, wherein the prediction model is trained according to the method of the second aspect;

[0034] Based on the individual prediction feedback results in the individual prediction feedback result set, a group actual feedback result of sending the target service to the group to be predicted is obtained.

[0035] In a fourth aspect, a data processing device is provided, the data processing device being used to determine a target parameter of a target model, the target model being used to predict a target business feedback result based on a target feature of a target object in a target object set, the target business feedback result being a feedback result of a target business sent to the target object in the target object set, the data processing device comprising:

[0036] an acquisition unit, for acquiring m first loss values, where m is an integer greater than 1, and the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, where n is an integer greater than 1, and the model structures of the m models to be confirmed are the same as the model structure of the target model, and the parameters of different models to be confirmed are different; the first loss value is obtained based on n first differences, and the first differences correspond to the training features one by one, and the first differences are the difference between the individual predicted feedback results of sending the target business to the training objects and the individual actual feedback results of sending the target business to the training objects, and the individual predicted feedback results are the feedback results of sending the target business to the training objects in the training object set predicted by the model to be confirmed;

[0037] a determination unit, configured to determine a target relationship based on the m first loss values, the target relationship being a relationship between the first loss value and a second loss value, the second loss value being obtained based on a second difference, the second difference corresponding one-to-one to the model to be confirmed, the second difference being a difference between a group prediction feedback result of sending the target service to a target group and a group actual feedback result of sending the target service to the target group, the target group being a group including the n training objects; the group prediction feedback result being obtained based on the n individual prediction feedback results of the n training objects;

[0038] The determining unit is configured to determine, based on the target relationship, that the first loss value corresponding to the second loss value that satisfies a preset condition is a target loss value;

[0039] A processing unit is used to use the parameters of the model to be confirmed corresponding to the target loss value as the target parameters.

[0040] In combination with any implementation manner of the present application, the determining unit is specifically configured to:

[0041] Sampling from the m first loss values ​​to obtain p sampled loss values, where p is an integer less than m;

[0042] For each of the p sampled loss values, respectively determine n individual predicted feedback results corresponding to the sampled loss value and n individual actual feedback results corresponding to the sampled loss value;

[0043] Determine the group prediction feedback result of the target group based on the n individual prediction feedback results corresponding to the same sampling loss value, and obtain p group prediction feedback results;

[0044] Determine the group actual feedback result of the target group based on the n individual actual feedback results corresponding to the same sampling loss value, and obtain p group actual feedback results;

[0045] Determine the second loss value based on the difference between the group prediction feedback result and the group actual feedback result corresponding to the same sampled loss value, and obtain p second loss values;

[0046] The target relationship is obtained based on the p second loss values ​​and the p sampled loss values.

[0047] In combination with any implementation manner of the present application, the determining unit is specifically configured to:

[0048] Obtaining a priori distribution of the first loss value and the second loss value, where the prior distribution represents a distribution of the second loss value corresponding to the first loss value;

[0049] The prior distribution is updated based on the p second loss values ​​and the p sampled loss values ​​to obtain a posterior distribution as the target relationship.

[0050] In combination with any implementation manner of the present application, the loss values ​​other than the p sampled loss values ​​among the m first loss values ​​are loss values ​​to be sampled, and the determining unit is specifically configured to:

[0051] Based on the target relationship, determining an acquisition function, the acquisition function being used to estimate a value of the first loss value being sampled;

[0052] Based on the acquisition function, determine the sampled value of the loss value to be sampled, and obtain (mp) sampling values;

[0053] Based on the (mp) sampling values, q sampling loss values ​​are obtained by sampling from the (mp) loss values ​​to be sampled, where q is an integer less than or equal to (mp);

[0054] Determine the second loss value that meets the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as a reference loss value;

[0055] The first loss value corresponding to the reference loss value is determined as the target loss value.

[0056] In combination with any embodiment of the present application, the acquisition function satisfies at least one of the following conditions: the mean of the second loss value corresponding to the first loss value in the target relationship is negatively correlated with the value of the first loss value sampled, the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value sampled, and the value of the sampled loss value in the first loss value is lower than the value of the loss value in the first loss value other than the sampled loss value.

[0057] In combination with any implementation manner of the present application, the determining unit is specifically configured to:

[0058] When the sum of p and q is greater than or equal to the sampling threshold, the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​is determined as the reference loss value.

[0059] In combination with any implementation manner of the present application, the first loss value is obtained based on the sum of the squares of n first differences.

[0060] In a fifth aspect, a training device is provided, the training device comprising:

[0061] An acquisition unit, used for acquiring a model to be trained, wherein the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, and the target parameters are obtained according to the first aspect and any embodiment thereof;

[0062] The acquisition unit is used to acquire n training features and n labels of n training objects, wherein the labels represent individual actual feedback results of sending the target service to the training objects;

[0063] A prediction unit, used for inputting the n training features into the model to be trained, so that the model to be trained predicts the feedback result of sending the target service to the training object, and obtains n training feedback results;

[0064] A processing unit, configured to obtain a training loss of the model to be trained based on a difference between the n training feedback results and the n labels;

[0065] An updating unit is used to update the parameters of the model to be trained based on the training loss to obtain a prediction model.

[0066] In a sixth aspect, a prediction device is provided, the prediction device comprising:

[0067] An acquisition unit, used for acquiring a set of features to be predicted, wherein the set of features to be predicted is a set of features to be predicted, and the features to be predicted are features of objects to be predicted in a group to be predicted;

[0068] A prediction unit, used for inputting the feature set to be predicted into a prediction model, so that the prediction model predicts the individual prediction feedback result of sending the target business to the object to be predicted in the group to be predicted, and obtains an individual prediction feedback result set, wherein the prediction model is trained according to the method of the second aspect;

[0069] The processing unit is used to obtain a group actual feedback result of sending the target service to the group to be predicted based on the individual prediction feedback results in the individual prediction feedback result set.

[0070] In the seventh aspect, an electronic device is provided, comprising: a processor and a memory, the memory being used to store computer program code, the computer program code comprising computer instructions, and when the processor executes the computer instructions, the electronic device executes the first aspect and any embodiment thereof, the electronic device either executes the method of the second aspect, or the electronic device either executes the method of the third aspect.

[0071] In an eighth aspect, another electronic device is provided, comprising: a processor, a sending device, an input device, an output device and a memory, wherein the memory is used to store computer program code, and the computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes the first aspect and any embodiment thereof, or the electronic device executes the method of the second aspect, or the electronic device executes the method of the third aspect.

[0072] In the ninth aspect, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor is caused to execute the first aspect and any embodiment thereof, or the processor is caused to execute the method of the second aspect, or the processor is caused to execute the method of the third aspect.

[0073] In the tenth aspect, a computer program product is provided, which includes a computer program or instructions, and when the computer program or instructions are run on a computer, the computer is caused to execute the above-mentioned first aspect and any embodiment thereof, or the computer is caused to execute the method of the above-mentioned second aspect, or the computer is caused to execute the method of the above-mentioned third aspect.

[0074] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application.

[0075] In the embodiment of the present application, m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, n is an integer greater than 1, the model structures of the m models to be confirmed are the same as the model structure of the target model, and the parameters of different models to be confirmed are different. The first loss value is obtained based on n first differences, the first difference corresponds to the training feature one by one, the first difference is the difference between the individual prediction feedback result of sending the target service to the training object and the individual actual feedback result of sending the target service to the training object, and the individual prediction feedback result is the feedback result of sending the target service to the training object in the training object set predicted by the model to be confirmed. After obtaining m first loss values, the data processing device determines the target relationship based on the m first loss values, the target relationship is the relationship between the first loss value and the second loss value, the second loss value is obtained based on the second difference, wherein the second difference corresponds to the model to be confirmed one by one, the second difference is the difference between the group prediction feedback result of sending the target service to the target group and the group actual feedback result of sending the target service to the target group, the target group is a group including n training objects, and the group prediction feedback result is obtained based on the n individual prediction feedback results of the n training objects. Based on the target relationship, the first loss value corresponding to the second loss value that meets the preset condition is determined as the target loss value. The parameters of the to-be-confirmed model corresponding to the target loss value are used as the target parameters. In this way, when the first loss value represents the error of the individual prediction feedback result and the second loss value represents the error of the group prediction feedback result, the target parameter for predicting the group feedback result can be determined based on the first loss value. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background technology, the drawings required for use in the embodiments of the present application or the background technology will be described below.

[0077] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0078] Figure 1 A flowchart of a data processing method provided in an embodiment of the present application;

[0079] Figure 2 A flowchart of another data processing method provided in an embodiment of the present application;

[0080] Figure 3 A flowchart of a training method provided in an embodiment of the present application;

[0081] Figure 4 A flowchart of a prediction method provided in an embodiment of the present application;

[0082] Figure 5 A schematic diagram of the structure of a data processing device provided in an embodiment of the present application;

[0083] Figure 6 A training device provided in an embodiment of the present application;

[0084] Figure 7 A prediction device provided in an embodiment of the present application;

[0085] Figure 8 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0086] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings 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.

[0087] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0088] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0089] An embodiment of the present application provides a data processing method, a training method, and a prediction method, wherein the data processing method is used to determine the target parameters of a target model, the target model is used to predict a target business feedback result based on the target characteristics of a target object in a target object set, and the target business feedback result is a feedback result of a target business sent to a target object in the target object set.

[0090] In an embodiment of the present application, the target business may be any business. In one possible implementation, the target business is advertising. In another possible implementation, the target business is multimedia content, wherein the multimedia content includes one or more of the following: images, text, and videos. The target object set is a set of target objects, wherein the number of target objects in the target object set is greater than 1. The target object is a receiving object of the target business. In one possible implementation, the target business is advertising, and the target object is an object for delivering the advertisement. In another possible implementation, the target business is multimedia content, and the target object is an object for receiving the multimedia content.

[0091] In an embodiment of the present application, the target feedback result is a feedback result obtained when a target service is sent to a target object in a target object set. In one possible implementation, the target service is advertising, and the target feedback result is the revenue obtained by the advertisement publisher from the target object by placing advertisements to the target object. For example, the target feedback result is the life cycle value (LTV) of the target object. In another possible implementation, the target service is multimedia content, and the target feedback result is an indicator that characterizes the target object's interest in the multimedia content. For example, the target feedback result is the number of times the multimedia content is viewed by the target object, the target feedback result may be the number of times the multimedia content is forwarded by the target object, or the target feedback result may be the duration of time the multimedia content is browsed by the target object.

[0092] In the embodiment of the present application, the data processing method is executed by a data processing device, wherein the data processing device can be any electronic device that can execute the technical solution disclosed in the embodiment of the method of the present application. Optionally, the data processing device can be one of the following: a computer, a server.

[0093] It should be understood that the method embodiment of the present application can also be implemented by a processor executing a computer program code. The following describes the embodiment of the present application in conjunction with the drawings in the embodiment of the present application. Figure 1 , Figure 1 A flowchart of a data processing method provided in an embodiment of the present application.

[0094] 101. Obtain m first loss values.

[0095] In an embodiment of the present application, the first loss value is obtained by inputting n training features of n training objects into the model to be confirmed, wherein the training features correspond one-to-one to the training objects, the training features are the features of the training objects, and the training features include one or more of the following: behavioral characteristics of the training objects, attributes of the training objects, and characteristics of the sending methods, wherein the sending method is a method of sending a target service to the target object, for example, the target service is sent to the training object by delivering it on the target channel, then the characteristics of the sending method include the characteristics of delivering the target service on the target channel.

[0096] Optionally, the training object is an object that accepts the target business. In one possible implementation, the target business is an advertisement, and the training object is a conversion object of the advertisement. For example, the advertisement is a promotional advertisement for an application software. When the training object becomes a user of the application software through the advertisement, the training object is a conversion object of the advertisement. In another possible implementation, the target business is multimedia content, and the training object is an object that views the multimedia content. The target business is a business delivered by the target platform. In this case, the training object becomes a user of the target platform after accepting the target business. For example, the target business is a promotional advertisement for an application software, and the platform that delivers the advertisement is the platform that runs the application software. In this case, the platform is the target platform. After becoming a conversion object of the promotional advertisement for the application software by accepting the target business, the training object becomes a user of the target platform.

[0097] The behavior characteristics of the training object include the behavior characteristics of the training object within a preset time period after becoming a user of the target platform. Optionally, the behavior characteristics include one or more of the following: the active duration of the training object, the duration of the training object using the target platform, and the training object's operation record on the content in the target platform. For example, if the target platform is a platform for displaying multimedia content, the preset time period can be 7 days. The active duration of the training object is the duration of the training object's operation on the content of the target platform, wherein the operation generated on the content of the target platform includes: clicking on the content, browsing the content, commenting on the content, forwarding the content, and collecting the content. The duration of the training object using the target platform is the duration of the training object logging into the target platform. The attributes of the training object include one or more of the following: the hobbies of the training object, the age of the training object, the gender of the training object, and the information of the device used by the training object. The characteristics of the sending method include one or more of the following: the characteristics of the target channel, the placement position of the target business in the target channel, the category of the material used to display the target business, and the characteristics of the account of the target business placed in the target channel. For example, the target business is advertising, the target channel is the channel for delivering advertising, the characteristics of the target channel include the information of the target channel's audience group, and the delivery position includes the display position of the target business in the target channel. The materials used to display the target business include: images, videos, and texts, and the categories of the materials used to display the target business are the categories to which the materials belong. The characteristics of the account that delivers the target business in the target channel include the level of the account in the target channel, where the higher the level, the higher the probability that the account will compete to get the business delivered in the target channel.

[0098] The model to be confirmed can be any deep learning model. Optionally, the model to be confirmed is a gradient boosting decision tree model (Light Gradient Boosting Machine, LGBM). Inputting the training features of the training object into the model to be confirmed can enable the model to be confirmed to predict the target feedback result of the target business sent to the training object based on the training features, which is the individual prediction feedback result. Inputting the training features of different training objects into the model to be confirmed can enable the model to be confirmed to send the target feedback results of the target business to different training objects based on the prediction, and obtain individual prediction feedback results of different training objects. It should be understood that by inputting the training features of one training object into the model to be confirmed, the individual prediction feedback results of one training object can be obtained, and by inputting the training features of n training objects into the model to be confirmed, the individual prediction feedback results of n training objects can be obtained.

[0099] Optionally, when the model to be confirmed is LGBM, the parameters of the target model include one or more of the following hyperparameters: the maximum number of leaf nodes of the tree (num_leaves), the maximum depth of the tree (max_depth), the learning rate (learning_rate), the number of trees (n_estimators), the proportion of samples used in training (subsample), and the frequency of sampling (subsample_freq).

[0100] The individual actual feedback result is the true value (groundtruth, GT) of the feedback result of the target business sent to the training object, that is, the individual actual feedback result can be used to measure the accuracy of the individual prediction feedback result. Specifically, the larger the first difference between the individual prediction feedback result and the individual actual feedback result, the lower the accuracy of the individual prediction feedback result, that is, the first difference can represent the accuracy of the individual prediction feedback result output by the model to be confirmed. Since the number of individual prediction feedback results is n, the number of first differences is also n, and the first loss value is obtained based on the n first differences, that is, the first loss value can represent the accuracy of the n individual prediction feedback results output by the model to be confirmed. In a possible implementation method, the first loss value is obtained based on the sum of the squares of the n first differences. Optionally, the first loss value is obtained based on the mean square error (MSE) of the n individual prediction feedback results. Specifically, the MSE is calculated based on the n individual prediction feedback results and the n individual actual feedback results as the first loss value. Since the calculation of MSE does not involve a regularization term, the first difference between the individual predicted feedback result and the individual actual feedback result can be represented by MSE, which can reduce the error of the first difference, and then the first loss value obtained based on the first difference can improve the accuracy of the first loss value, thereby improving the accuracy of the target parameter subsequently determined based on the first loss value. In another possible implementation, the first loss value is the average of n first differences.

[0101] Different parameters of the target model will lead to different structures of the target model output. Specifically, when n training features are input into target models with different parameters, the first loss values ​​obtained are different. Therefore, when the model structures of m models to be confirmed are the same as the model structure of the target model, and different models to be confirmed among the m models to be confirmed have different parameters, the n training features are respectively input into the m models to be confirmed, and m first loss values ​​can be obtained, that is, the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed.

[0102] 102. Determine a target relationship based on the m first loss values.

[0103] In an embodiment of the present application, the target relationship is the relationship between the first loss value and the second loss value, the second loss value is obtained based on the second difference, the second difference corresponds one-to-one to the model to be confirmed, and the second difference is the difference between the group prediction feedback result of sending the target business to the target group and the group actual feedback result of sending the target business to the target group, and the target group is a group including n training objects. In one possible implementation, the target business is advertising, and the objects in the target group are all objects that send advertisements to the target group through the target channel. In another possible implementation, the target business is advertising, and the objects in the target group are objects that meet the preset conditions. For example, the preset condition is that the age is less than 35 years old, then the objects in the target group are objects that are less than 35 years old.

[0104] As described in step 101, the training features of n training objects are input into the model to be confirmed, and the individual prediction feedback results of the n training objects can be obtained. Then, the group prediction feedback result of the target group including the n training objects can be determined based on the individual prediction results of the n training objects. In one possible implementation method, the mean value of the individual prediction results of the n training objects is determined to obtain the group prediction feedback result of the target group including the n training objects. In another possible implementation method, the median value of the individual prediction results of the n training objects is determined as the group prediction feedback result of the target group including the n training objects. Therefore, the n training features are input into the model to be confirmed, and the n individual prediction feedback results of the n training objects can be obtained, and then a group prediction feedback result of the target group including the n training objects can be determined based on the n individual prediction feedback results.

[0105] Different parameters of the target model will lead to different output results of the target model. Specifically, when n training features are input into target models with different parameters, the n individual prediction feedback results of n training objects are different. Different models to be confirmed have different parameters, so the group prediction feedback results of the target group obtained based on different models to be confirmed are different. In other words, m group prediction feedback results can be obtained based on m models to be confirmed.

[0106] The actual feedback result of the group is the true value (ground truth, GT) of the feedback result of the target group obtained by sending the target service to the training objects in the target group. That is to say, the actual feedback result of the group can be used to measure the accuracy of the group prediction feedback result. Specifically, the larger the second difference between the group prediction feedback result and the group actual feedback result, the lower the accuracy of the group prediction feedback result, that is, the second difference can represent the accuracy of the group prediction feedback result output by the model to be confirmed. Since the number of group prediction feedback results is m, the number of second differences is also m, and the second loss value is obtained based on the m second differences, that is, the second loss value can represent the accuracy of the m group prediction feedback results output by the m models to be confirmed, wherein the second loss value corresponds one-to-one to the model to be confirmed.

[0107] Optionally, the actual feedback result of the group is the average of the actual feedback results of n individuals, and the second difference is the difference between the average of the predicted feedback results of n individuals and the average of the actual feedback results of n individuals. In this case, the second loss value obtained based on the second difference is the Mean Absolute Percentage Error (MAPE) based on the predicted feedback results of n individuals and the actual feedback results of n individuals.

[0108] Based on step 101 and step 102, it can be known that the first loss value and the second loss value both correspond to the model to be confirmed one by one. For the convenience of expression, the first loss value and the second loss value corresponding to the same model to be confirmed are referred to as the first loss value and the second loss value with a corresponding relationship. The target relationship is the relationship between the first loss value and the second loss value with a corresponding relationship, that is, the second loss value corresponding to the first loss value can be determined based on the target relationship. For example, the first loss value obtained based on the model to be confirmed is s1, then based on the target relationship and s1, the second loss value obtained based on the model to be confirmed can be determined.

[0109] In one possible implementation, the data processing device determines the target relationship by executing the following steps: sampling from m first loss values ​​to obtain p sampled loss values, where p is an integer less than m; for each of the p sampled loss values, respectively determining n individual predicted feedback results corresponding to the sampled loss value and n individual actual feedback results corresponding to the sampled loss value; determining the group predicted feedback result of the target group based on the n individual predicted feedback results corresponding to the same sampled loss value, and obtaining p group predicted feedback results; determining the group actual feedback result of the target group based on the n individual actual feedback results corresponding to the same sampled loss value, and obtaining p group actual feedback results; determining the second loss value based on the difference between the group predicted feedback result and the group actual feedback result corresponding to the same sampled loss value, and obtaining p second loss values; obtaining the target relationship based on the p second loss values ​​and the p sampled loss values.

[0110] For example, the m first loss values ​​include a first loss value a and a first loss value b. The first loss value a is sampled from the m first loss values, and p is 1 at this time. The first loss value a is the average value of n individual prediction feedback results, wherein the n individual prediction feedback results include individual prediction feedback result 1 and individual prediction feedback result 2, individual prediction feedback result 1 is the individual prediction feedback result of training object 1, and individual prediction feedback result 2 is the individual prediction feedback result of training object 2. The individual actual feedback result of training object 1 is individual actual feedback result 1, and the individual actual feedback result of training object 2 is individual actual feedback result 2. The average of individual prediction feedback result 1 and individual prediction feedback result 2 is determined to obtain the group prediction feedback result of the target group including training object 1 and training object 2, and the average of individual actual feedback result 1 and individual actual feedback result 2 is determined to obtain the group actual feedback result of the target group including training object 1 and training object 2. Based on the difference between the group prediction feedback result and the group actual feedback result, a second loss value is determined. Thus, a second loss value can be obtained based on a sampled loss value.

[0111] As an optional implementation, the data processing device obtains the target relationship by fitting p second loss values ​​and p sampling loss values.

[0112] As another optional implementation, the data processing device implements "obtaining a target relationship based on p second loss values ​​and p sampling loss values" by executing the following steps: obtaining a prior distribution of the first loss value and the second loss value, the prior distribution representing the distribution of the second loss value corresponding to the first loss value; updating the prior distribution based on the p second loss values ​​and the p sampling loss values ​​to obtain a posterior distribution as the target relationship.

[0113] The above prior distribution, in which a first loss value may correspond to multiple second loss values, that is, the interval of the second loss value corresponding to the first loss value can be determined based on the prior distribution. For example, the interval of the second loss value corresponding to the first loss value determined based on the prior distribution is 5 to 10, that is, the second loss value corresponding to the first loss value is distributed in the interval of 5 to 10. Optionally, the prior distribution is a Gaussian Process (GP).

[0114] Since the distribution interval of the second loss value corresponding to the first loss value is determined based on the prior distribution, it is difficult to determine the second loss value corresponding to the first loss value, and the p second loss values ​​and the p sampled loss values ​​are all determined based on the results of the output of the model to be confirmed, that is, the p second loss values ​​and the p sampled loss values ​​are all observed values, so the data processing device updates the prior distribution based on the p second loss values ​​and the p sampled loss values ​​to obtain the posterior distribution, so that the second loss value corresponding to the first loss value can be more accurately determined based on the posterior distribution compared to the prior distribution, and the posterior distribution is used as the target relationship. Optionally, the data processing device updates the prior distribution to obtain the posterior distribution based on the maximum a posteriori estimate (MAP), the p second loss values ​​and the p sampled loss values.

[0115] 103. Based on the target relationship, determine that the first loss value corresponding to the second loss value that meets a preset condition is a target loss value.

[0116] In the embodiment of the present application, the second loss value meets the preset condition, which means that the second loss value meets expectations, that is, the group prediction feedback result obtained based on the result output by the model to be confirmed meets expectations, which means that the group prediction feedback result corresponding to the second loss value has high accuracy. The preset condition can be one of the following: the minimum second loss value, the second loss value less than or equal to the convergence threshold. Optionally, the preset condition is the minimum second loss value, and the data processing device determines the first loss value corresponding to the minimum second loss value as the target loss value based on the target relationship.

[0117] In a possible implementation, the loss values ​​other than the p sampled loss values ​​among the m first loss values ​​are loss values ​​to be sampled. The data processing device performs the following steps in the process of executing step 103: based on the target relationship, determine the acquisition function, wherein the acquisition function is used to estimate the value of the first loss value being sampled; based on the acquisition function, determine the value of the loss value to be sampled to obtain (mp) sampling values; based on the (mp) sampling values, sample from the (mp) loss values ​​to be sampled to obtain q sampling loss values, q being an integer less than or equal to (mp); determine the second loss value that meets the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as the reference loss value; determine the first loss value corresponding to the reference loss value as the target loss value.

[0118] The greater the sampling value of the loss value to be sampled, the higher the probability that the loss value to be sampled is sampled. Optionally, the data processing device determines the largest q of the (mp) sampling values ​​as q sampling loss values.

[0119] Optionally, the acquisition function satisfies at least one of the following conditions: the mean of the second loss value corresponding to the first loss value in the target relationship is negatively correlated with the value of the first loss value sampled, the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value sampled, and the value of the sampled loss value in the first loss value is lower than the value of the loss value other than the sampled loss value in the first loss value.

[0120] The larger the mean of the second loss value, the larger the second loss value, and the expectation of the second loss value is to make the second loss value smaller. Therefore, the smaller the second loss value, the higher the sampled value of the first loss value corresponding to the second loss value. In this way, by sampling the first loss value and making it an observed value, the probability that the observed value contains the first loss value corresponding to the second loss value that meets the preset conditions can be increased.

[0121] A large variance of the second loss value corresponding to the first loss value indicates that the uncertainty of the second loss value corresponding to the first loss value is high, and high uncertainty will lead to a larger error in the target relationship. Therefore, if the first loss value is sampled to make the first loss value an observed value, the second loss value corresponding to the first loss value can be made a constant value, thereby reducing the error of the target relationship and improving the accuracy of the target relationship. Based on the fact that the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value being sampled, the accuracy of the target relationship can be improved.

[0122] Since the first loss value can be sampled to become an observed value, the uncertainty of the second loss value corresponding to the first loss value can be reduced, thereby improving the accuracy of the target relationship. Therefore, sampling the unsampled first loss value is more conducive to improving the accuracy of the target relationship than sampling the sampled first loss value. Therefore, by making the sampled value of the sampled loss value in the first loss value lower than the sampled value of the loss value in the first loss value other than the sampled loss value, the accuracy of the target relationship can be improved.

[0123] Optionally, the acquisition function is one of the following: a probability of improvement (PI) function, an expected improvement (EI) function.

[0124] As an optional implementation, the data processing device implements "determining the second loss value that meets the preset conditions from the second loss values ​​corresponding to the p sampling loss values ​​and the second loss values ​​corresponding to the q sampling loss values ​​as the reference loss value" by executing the following steps: when the sum of p and q is greater than or equal to the sampling threshold, determining the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​as the reference loss value. Among them, the sum of p and q is greater than or equal to the sampling threshold, which means that the number of sampling loss values ​​is greater than or equal to the sampling threshold, which also means that the number of sampling loss values ​​meets expectations. At this time, determining the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​as the reference loss value can increase the probability that the reference loss value is the second loss value that meets the preset conditions.

[0125] As another optional implementation, the data processing device implements "determining the second loss value that meets the preset conditions from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as the reference loss value" by executing the following steps: when the number of samplings is greater than or equal to the number threshold, determining the minimum value of the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as the reference loss value, wherein one sampling process includes: sampling from m first loss values ​​to obtain p sampled loss values, obtaining p second loss values ​​based on the p sampled loss values, updating the prior distribution based on the p sampled loss values ​​and the p second loss values ​​to obtain the posterior distribution, determining the acquisition function based on the posterior distribution, and determining the sampling loss value of the next sampling based on the acquisition function. The sampling number being greater than or equal to the number threshold indicates that the sampling number has reached the expectation, and at this time determining the minimum value of the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as the reference loss value can increase the probability that the reference loss value is the second loss value that meets the preset conditions.

[0126] 104. Using the parameters of the model to be confirmed corresponding to the target loss value as the target parameters.

[0127] As described in step 101, the first loss value corresponds to the model to be confirmed one by one, and the first loss value is the target loss value, so the data processing device can determine the model to be confirmed corresponding to the target loss value, and then determine the parameters of the model to be confirmed corresponding to the target loss value as the target parameters. Since the target loss value is the first loss value corresponding to the second loss value that meets the preset condition, the parameters of the model to be confirmed that obtains the target loss value are used as the parameters of the target model, which can make the group prediction feedback result obtained based on the result output by the target model more accurate. Therefore, the data processing device uses the parameters of the model to be confirmed corresponding to the target loss value as the target parameters.

[0128] In the embodiment of the present application, m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, n is an integer greater than 1, the model structures of the m models to be confirmed are the same as the model structure of the target model, and the parameters of different models to be confirmed are different. The first loss value is obtained based on n first differences, the first difference corresponds to the training feature one by one, the first difference is the difference between the individual prediction feedback result of sending the target service to the training object and the individual actual feedback result of sending the target service to the training object, and the individual prediction feedback result is the feedback result of sending the target service to the training object in the training object set predicted by the model to be confirmed. After obtaining m first loss values, the data processing device determines the target relationship based on the m first loss values, the target relationship is the relationship between the first loss value and the second loss value, the second loss value is obtained based on the second difference, wherein the second difference corresponds to the model to be confirmed one by one, the second difference is the difference between the group prediction feedback result of sending the target service to the target group and the group actual feedback result of sending the target service to the target group, the target group is a group including n training objects, and the group prediction feedback result is obtained based on the n individual prediction feedback results of the n training objects. Based on the target relationship, the first loss value corresponding to the second loss value that meets the preset condition is determined as the target loss value. The parameters of the to-be-confirmed model corresponding to the target loss value are used as the target parameters. In this way, when the first loss value represents the error of the individual prediction feedback result and the second loss value represents the error of the group prediction feedback result, the target parameter for predicting the group feedback result can be determined based on the first loss value.

[0129] In one possible implementation, the data processing device implements step 102 and step 103 by using a Bayesian optimization method. Specifically, the data processing device determines the target relationship and the target loss value using m first loss values ​​based on the Bayesian optimization method. Since the relationship between the first loss value and the second loss value is difficult to model, the Bayesian optimization method can be used to optimize the relationship as a black box to obtain the target parameter, thereby improving the efficiency of determining the target parameter and the accuracy of the target parameter.

[0130] See also Figure 2 , Figure 2 A flow chart of another data processing method provided in an embodiment of the present application. Figure 2 As shown, the data processing method is used to implement object machine learning modeling. Specifically, the parameters of the LGBM model can be determined through the data processing method, wherein LGBM is the target model described above. When the parameters of the LGBM model are the determined parameters, the accuracy of the individual prediction feedback results output by LGBM is high, but the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by LGBM is low. Figure 2 In the data processing method shown, firstly, the first and second level channel characteristics, the attributes of the training object, the behavioral characteristics of the training object in the previous 7 days, the layout, the material category, and the delivery account characteristics are obtained, wherein the first and second level channel characteristics include the characteristics of the first level channel for delivering advertisements and the characteristics of the second level channel for delivering target business, the behavioral characteristics of the training object in the previous 7 days include the behavioral characteristics of the training object in the first 7 days after becoming a user of the target platform, the layout is the display layout of the target business in the target channel, the material category is the category of the material used to display the target business, and the delivery account characteristics are the characteristics of the account for delivering the target business. The first and second level channel characteristics, the attributes of the training object, the behavioral characteristics of the training object in the previous 7 days, the layout, the material category, and the delivery account characteristics are input into the LGBM model to obtain the first loss value, and then perform Bayesian optimization parameter adjustment. Specifically, the parameters of the LGBM model are determined based on the Bayesian optimization parameter adjustment and the first loss value. The target parameters of the LGBM model are obtained based on the data processing method provided in the embodiment of the present application, and the accuracy of the group prediction feedback results obtained based on the individual prediction feedback results output by LGBM can be improved when the parameters of the LGBM model are the target parameters.

[0131] Optionally, the data processing device implements the Bayesian optimization method by running the hyperparameter optimization framework (optuna). Specifically, the process of obtaining the target parameters by implementing the Bayesian optimization method through optuna can be called a learning process, wherein the learning process includes multiple trial processes, and the trial process is the sampling process described above. Before running optuna, the data processing device first determines the hyperparameter search space. Specifically, by calling the application programming interface (API) of optuna, the search range of the parameters of the target model is determined. Optionally, the search range of the parameters of the target model is the search range of the hyperparameters of the LGBM model. For example, the hyperparameters of the LGBM model include the maximum number of leaf nodes of the tree, then the search range of the hyperparameters of the LGBM model includes the search range of the maximum number of leaf nodes of the tree. After determining the search range of the parameters of the target model, the second loss value is determined. Specifically, the second loss value is defined as the difference between the group prediction feedback result and the group actual feedback result. By calling optuna's API and using optuna's create_study and optimize methods, the second loss value that meets the preset conditions can be determined, and then the first loss value corresponding to the second loss value that meets the preset conditions can be determined as the target loss value, so that the parameters of the model to be confirmed corresponding to the target loss value can be determined as the target parameters.

[0132] See also Figure 3 , Figure 3 A flowchart of a training method provided in an embodiment of the present application.

[0133] 301. Obtain the model to be trained.

[0134] In an embodiment of the present application, the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, wherein the target parameters are obtained according to the data processing method described above.

[0135] 302. Obtain n training features and n labels of n training objects.

[0136] In the embodiment of the present application, the meaning of the n training features of the n training objects can be found in step 101, which will not be described in detail here. The label represents the individual actual feedback result of sending the target service to the training object.

[0137] 303. Input the n training features into the model to be trained, so that the model to be trained predicts the feedback result of sending the target service to the training object, and obtains n training feedback results.

[0138] 304. Based on the difference between the n training feedback results and the n labels, obtain the training loss of the model to be trained.

[0139] Optionally, the training device determines the MSE of the n training feedback results and the n individual actual feedback results based on the n training feedback results and the n labels as the training loss of the model to be trained.

[0140] 305. Based on the training loss, update the parameters of the model to be trained to obtain a prediction model.

[0141] In one possible implementation, the training device determines the gradient of back propagation of the model to be trained based on the training loss, back propagates the model to be trained based on the gradient, updates the parameters of the model to be trained through back propagation until the training loss converges, and obtains a prediction model.

[0142] In the embodiment of the present application, the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters. Since the target parameters are obtained according to the data processing method described above, when the parameters of the target model are target parameters, the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the target model is high, so the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the model to be trained is high. So after the training device obtains the model to be trained, n training features of n training objects and n labels, the n training features are input into the model to be trained, so that the model to be trained predicts the feedback results of sending the target business to the training object, and obtains n training feedback results. Based on the difference between the n training feedback results and the n labels, the training loss of the model to be trained is obtained. Based on the training loss, the parameters of the model to be trained are updated to obtain a prediction model. Thus, a prediction model can be obtained by training the model to be trained, so that the accuracy of the individual prediction feedback results output by the prediction model is high, so that the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the prediction model is high.

[0143] See also Figure 4 , Figure 4 A flowchart of a prediction method provided in an embodiment of the present application.

[0144] 401. Obtain a feature set to be predicted.

[0145] In an embodiment of the present application, the set of features to be predicted is a set of features to be predicted, wherein the features to be predicted are features of objects to be predicted in the group to be predicted. In one possible implementation, the objects to be predicted in the group to be predicted are objects with the same attributes, for example, the objects to be predicted are objects under the age of 35. In another possible implementation, the objects to be predicted in the group to be predicted are objects that send target services through the same sending method, for example, the target service is advertising, and the advertiser places the advertisement through the target channel, and the objects to be predicted in the group to be predicted are all objects that receive the advertisement through the target channel.

[0146] 402. Input the feature set to be predicted into a prediction model, so that the prediction model predicts and sends individual prediction feedback results of the target business to the objects to be predicted in the group to be predicted, thereby obtaining an individual prediction feedback result set.

[0147] 403. Based on the individual prediction feedback results in the individual prediction feedback result set, obtain a group actual feedback result of sending the target service to the group to be predicted.

[0148] In the embodiment of the present application, the prediction model is trained according to the training method described above. The predicted features of the object to be predicted are input into the prediction model, so that the prediction model can predict the individual prediction feedback results of the target business sent to the object to be predicted. Moreover, the group prediction feedback results determined based on the individual prediction feedback results output by the prediction model have high accuracy, wherein the group prediction feedback results are group prediction feedback results including the group to be predicted.

[0149] In one possible scenario, the target business is a promotional advertisement for a mobile software (application, APP), and the delivery channel of the promotional advertisement is a target channel. The object to be predicted in the group to be predicted is the conversion object of the promotional advertisement, that is, the object to be predicted becomes a user of the APP through the promotional advertisement. Based on steps 401 to 403, the group prediction feedback results of the group to be predicted can be predicted to obtain the LTV of the group to be predicted. Based on the LTV of the group to be predicted, the LTV that can be obtained by delivering the promotional advertisement of the APP through the target channel can be evaluated, and then a promotion strategy for the promotional advertisement of the APP can be formulated based on the LTV of the group to be predicted. For example, based on the LTV of the group to be predicted, the return on investment (ROI) of delivering the promotional advertisement through the target channel can be determined, and then based on the ROI, it can be determined whether to deliver the promotional advertisement through the target channel.

[0150] In another possible scenario, the target business is promotional advertising for an APP, the group to be predicted is the target of the promotional advertising for the APP, and the objects to be predicted in the group to be predicted are all objects with target attributes. The objects to be predicted in the group to be predicted are the conversion objects of the promotional advertising, that is, the objects to be predicted become users of the APP through the promotional advertising. Based on steps 401 to 403, the group prediction feedback results of the group to be predicted can be predicted to obtain the LTV of the group to be predicted. Based on the LTV of the group to be predicted, the LTV that can be obtained by placing promotional advertising on objects with target attributes can be evaluated, and then a promotion strategy for the promotional advertising of the APP can be formulated based on the LTV of the group to be predicted. For example, based on the LTV of the group to be predicted, the ROI of placing promotional advertising on objects with target attributes can be determined, and then based on the ROI, it can be determined whether to place promotional advertising on objects with target attributes.

[0151] Optionally, the error rate of LTV prediction of the target model obtained based on the training method described above for the above two scenarios is less than 7%.

[0152] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0153] The method of the embodiment of the present application is described in detail above, and the device of the embodiment of the present application is provided below.

[0154] See also Figure 5 , Figure 5 A structural schematic diagram of a data processing device provided in an embodiment of the present application, wherein the data processing device 1 is used to determine a target parameter of a target model, wherein the target model is used to predict a target business feedback result based on a target feature of a target object in a target object set, wherein the target business feedback result is a feedback result of a target business sent to the target object in the target object set, wherein the data processing device 1 comprises: an acquisition unit 11, a determination unit 12, and a processing unit 13, specifically:

[0155] An acquisition unit 11 is used to acquire m first loss values, where m is an integer greater than 1, and the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, where n is an integer greater than 1, and the model structures of the m models to be confirmed are the same as the model structure of the target model, and different parameters of the models to be confirmed are different; the first loss value is obtained based on n first differences, and the first differences correspond to the training features one by one, and the first differences are the difference between the individual predicted feedback results of sending the target business to the training objects and the individual actual feedback results of sending the target business to the training objects, and the individual predicted feedback results are the feedback results of sending the target business to the training objects in the training object set predicted by the model to be confirmed;

[0156] A determination unit 12 is used to determine a target relationship based on the m first loss values, wherein the target relationship is a relationship between the first loss value and a second loss value, wherein the second loss value is obtained based on a second difference, wherein the second difference corresponds one-to-one to the model to be confirmed, and wherein the second difference is a difference between a group prediction feedback result of sending the target service to a target group and a group actual feedback result of sending the target service to the target group, wherein the target group is a group including the n training objects; and the group prediction feedback result is obtained based on the n individual prediction feedback results of the n training objects;

[0157] The determining unit 12 is used to determine, based on the target relationship, that the first loss value corresponding to the second loss value that meets a preset condition is a target loss value;

[0158] The processing unit 13 is configured to use the parameters of the to-be-confirmed model corresponding to the target loss value as the target parameters.

[0159] In combination with any implementation manner of the present application, the determining unit 12 is specifically configured to:

[0160] Sampling from the m first loss values ​​to obtain p sampled loss values, where p is an integer less than m;

[0161] For each of the p sampled loss values, respectively determine n individual predicted feedback results corresponding to the sampled loss value and n individual actual feedback results corresponding to the sampled loss value;

[0162] Determine the group prediction feedback result of the target group based on the n individual prediction feedback results corresponding to the same sampling loss value, and obtain p group prediction feedback results;

[0163] Determine the group actual feedback result of the target group based on the n individual actual feedback results corresponding to the same sampling loss value, and obtain p group actual feedback results;

[0164] Determine the second loss value based on the difference between the group prediction feedback result and the group actual feedback result corresponding to the same sampled loss value, and obtain p second loss values;

[0165] The target relationship is obtained based on the p second loss values ​​and the p sampled loss values.

[0166] In combination with any implementation manner of the present application, the determining unit 12 is specifically configured to:

[0167] Obtaining a priori distribution of the first loss value and the second loss value, where the prior distribution represents a distribution of the second loss value corresponding to the first loss value;

[0168] The prior distribution is updated based on the p second loss values ​​and the p sampled loss values ​​to obtain a posterior distribution as the target relationship.

[0169] In combination with any embodiment of the present application, the loss values ​​other than the p sampled loss values ​​among the m first loss values ​​are loss values ​​to be sampled, and the determining unit 12 is specifically configured to:

[0170] Based on the target relationship, determining an acquisition function, the acquisition function being used to estimate a value of the first loss value being sampled;

[0171] Based on the acquisition function, determine the sampled value of the loss value to be sampled, and obtain (mp) sampling values;

[0172] Based on the (mp) sampling values, q sampling loss values ​​are obtained by sampling from the (mp) loss values ​​to be sampled, where q is an integer less than or equal to (mp);

[0173] Determine the second loss value that meets the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as a reference loss value;

[0174] The first loss value corresponding to the reference loss value is determined as the target loss value.

[0175] In combination with any embodiment of the present application, the acquisition function satisfies at least one of the following conditions: the mean of the second loss value corresponding to the first loss value in the target relationship is negatively correlated with the value of the first loss value sampled, the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value sampled, and the value of the sampled loss value in the first loss value is lower than the value of the loss value in the first loss value other than the sampled loss value.

[0176] In combination with any implementation manner of the present application, the determining unit 12 is specifically configured to:

[0177] When the sum of p and q is greater than or equal to the sampling threshold, the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​is determined as the reference loss value.

[0178] In combination with any implementation manner of the present application, the first loss value is obtained based on the sum of the squares of n first differences.

[0179] In the embodiment of the present application, m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, n is an integer greater than 1, the model structures of the m models to be confirmed are the same as the model structure of the target model, and the parameters of different models to be confirmed are different. The first loss value is obtained based on n first differences, the first difference corresponds to the training feature one by one, the first difference is the difference between the individual prediction feedback result of sending the target service to the training object and the individual actual feedback result of sending the target service to the training object, and the individual prediction feedback result is the feedback result of sending the target service to the training object in the training object set predicted by the model to be confirmed. After obtaining m first loss values, the data processing device determines the target relationship based on the m first loss values, the target relationship is the relationship between the first loss value and the second loss value, the second loss value is obtained based on the second difference, wherein the second difference corresponds to the model to be confirmed one by one, the second difference is the difference between the group prediction feedback result of sending the target service to the target group and the group actual feedback result of sending the target service to the target group, the target group is a group including n training objects, and the group prediction feedback result is obtained based on the n individual prediction feedback results of the n training objects. Based on the target relationship, the first loss value corresponding to the second loss value that meets the preset condition is determined as the target loss value. The parameters of the to-be-confirmed model corresponding to the target loss value are used as the target parameters. In this way, when the first loss value represents the error of the individual prediction feedback result and the second loss value represents the error of the group prediction feedback result, the target parameter for predicting the group feedback result can be determined based on the first loss value.

[0180] See also Figure 6 , Figure 6 A training device is provided in an embodiment of the present application. The training device 2 includes: an acquisition unit 21, a prediction unit 22, a processing unit 23, and an update unit 24. Specifically:

[0181] An acquisition unit 21 is used to acquire a model to be trained, wherein the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, which are obtained according to the first aspect and any embodiment thereof;

[0182] The acquisition unit 21 is used to acquire n training features and n labels of n training objects, wherein the labels represent individual actual feedback results of sending the target service to the training objects;

[0183] A prediction unit 22, configured to input the n training features into the model to be trained, so that the model to be trained predicts a feedback result of sending the target service to the training object, and obtains n training feedback results;

[0184] A processing unit 23, configured to obtain a training loss of the model to be trained based on a difference between the n training feedback results and the n labels;

[0185] The updating unit 24 is used to update the parameters of the model to be trained based on the training loss to obtain a prediction model.

[0186] In the embodiment of the present application, the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters. Since the target parameters are obtained according to the data processing method described above, when the parameters of the target model are target parameters, the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the target model is high, so the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the model to be trained is high. So after the training device obtains the model to be trained, n training features of n training objects and n labels, the n training features are input into the model to be trained, so that the model to be trained predicts the feedback results of sending the target business to the training object, and obtains n training feedback results. Based on the difference between the n training feedback results and the n labels, the training loss of the model to be trained is obtained. Based on the training loss, the parameters of the model to be trained are updated to obtain a prediction model. Thus, a prediction model can be obtained by training the model to be trained, so that the accuracy of the individual prediction feedback results output by the prediction model is high, so that the accuracy of the group prediction feedback results determined based on the individual prediction feedback results output by the prediction model is high.

[0187] See also Figure 7 , Figure 7A prediction device is provided in an embodiment of the present application. The prediction device 3 includes: an acquisition unit 31, a prediction unit 32, and a processing unit 33. Specifically:

[0188] An acquisition unit 31 is used to acquire a set of features to be predicted, where the set of features to be predicted is a set of features to be predicted, and the features to be predicted are features of objects to be predicted in a group to be predicted;

[0189] A prediction unit 32, configured to input the feature set to be predicted into a prediction model, so that the prediction model predicts the individual prediction feedback result of sending the target service to the object to be predicted in the group to be predicted, and obtains an individual prediction feedback result set, wherein the prediction model is trained according to the method of the second aspect;

[0190] The processing unit 33 is configured to obtain a group actual feedback result of sending the target service to the group to be predicted based on the individual prediction feedback results in the individual prediction feedback result set.

[0191] In the embodiment of the present application, the prediction model is trained according to the training method described above. The predicted features of the object to be predicted are input into the prediction model, so that the prediction model can predict the individual prediction feedback results of the target business sent to the object to be predicted. Moreover, the group prediction feedback results determined based on the individual prediction feedback results output by the prediction model have high accuracy, wherein the group prediction feedback results are group prediction feedback results including the group to be predicted.

[0192] In some embodiments, the functions or modules included in the device provided in the embodiments of the present application can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0193] Figure 8 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. The electronic device 4 includes a processor 41 and a memory 42. Optionally, the electronic device 4 also includes an input device 43 and an output device 44. The processor 41, the memory 42, the input device 43 and the output device 44 are coupled via a connector, and the connector includes various interfaces, transmission lines or buses, etc., which are not limited in the embodiments of the present application. It should be understood that in each embodiment of the present application, coupling refers to mutual connection in a specific manner, including direct connection or indirect connection through other devices, for example, it can be connected through various interfaces, transmission lines, buses, etc.

[0194] The processor 41 may include one or more processors, for example, one or more central processing units (CPUs). When the processor is a CPU, the CPU may be a single-core CPU or a multi-core CPU. Optionally, the processor 41 may be a processor group consisting of multiple CPUs, and the multiple processors are coupled to each other through one or more buses. Optionally, the processor may also be other types of processors, etc., which are not limited in the embodiments of the present application.

[0195] The memory 42 can be used to store computer program instructions and various computer program codes including the program code for executing the program code of the present application. Optionally, the memory includes but is not limited to random access memory (RAM), read-only memory (ROM), erasable programmable read only memory (EPROM), or portable read only memory (CD-ROM), which is used for related instructions and data.

[0196] The input device 43 is used to input data and / or signals, and the output device 44 is used to output data and / or signals. The input device 43 and the output device 44 can be independent devices or an integrated device.

[0197] It can be understood that in the embodiment of the present application, the memory 42 can be used not only to store relevant instructions, but also to store relevant data. The embodiment of the present application does not limit the specific data stored in the memory.

[0198] Understandably, Figure 8 Only a simplified design of an electronic device is shown. In practical applications, the electronic device may also include other necessary components, including but not limited to any number of input / output devices, processors, memories, etc., and all electronic devices that can implement the embodiments of the present application are within the protection scope of the present application.

[0199] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0200] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. Those skilled in the art can also clearly understand that the descriptions of the various embodiments of the present application have different focuses. For the convenience and brevity of description, the same or similar parts may not be repeated in different embodiments. Therefore, for parts not described or not described in detail in a certain embodiment, refer to the records of other embodiments.

[0201] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0202] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0203] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0204] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital versatile disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).

[0205] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by a computer program to instruct the relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned method embodiments. The aforementioned storage medium includes: a read-only memory (ROM) or a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

Claims

1. A data processing method, characterized in that: The method is used to determine a target parameter of a target model, the target model is used to predict a target service feedback result based on a target feature of a target object in a target object set, the target service feedback result is a feedback result of a target service sent to the target object in the target object set, and the method includes: Obtain m first loss values, where m is an integer greater than 1, and the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, where n is an integer greater than 1, and the model structures of the m models to be confirmed are the same as the model structure of the target model, and different parameters of the models to be confirmed are different; the first loss value is obtained based on n first differences, and the first differences correspond to the training features one by one, and the first differences are the difference between the individual predicted feedback results of sending the target business to the training objects and the individual actual feedback results of sending the target business to the training objects, and the individual predicted feedback results are the feedback results of sending the target business to the training objects in the training object set predicted by the model to be confirmed; Determine a target relationship based on the m first loss values, the target relationship being a relationship between the first loss value and a second loss value, the second loss value being obtained based on a second difference, the second difference corresponding one-to-one to the model to be confirmed, the second difference being a difference between a group prediction feedback result of sending the target service to a target group and a group actual feedback result of sending the target service to the target group, the target group being a group including the n training objects; the group prediction feedback result being obtained based on the n individual prediction feedback results of the n training objects; Based on the target relationship, determining that the first loss value corresponding to the second loss value that meets a preset condition is a target loss value; The parameters of the model to be confirmed corresponding to the target loss value are used as the target parameters.

2. The method according to claim 1, characterized in that: The determining the target relationship based on the m first loss values ​​comprises: Sampling from the m first loss values ​​to obtain p sampled loss values, where p is an integer less than m; For each of the p sampled loss values, respectively determine n individual predicted feedback results corresponding to the sampled loss value and n individual actual feedback results corresponding to the sampled loss value; Determine the group prediction feedback result of the target group based on the n individual prediction feedback results corresponding to the same sampling loss value, and obtain p group prediction feedback results; Determine the group actual feedback result of the target group based on the n individual actual feedback results corresponding to the same sampling loss value, and obtain p group actual feedback results; Determine the second loss value based on the difference between the group prediction feedback result and the group actual feedback result corresponding to the same sampled loss value, and obtain p second loss values; The target relationship is obtained based on the p second loss values ​​and the p sampled loss values.

3. The method according to claim 2, characterized in that The obtaining the target relationship based on the p second loss values ​​and the p sampling loss values ​​comprises: Obtaining a priori distribution of the first loss value and the second loss value, where the prior distribution represents a distribution of the second loss value corresponding to the first loss value; The prior distribution is updated based on the p second loss values ​​and the p sampled loss values ​​to obtain a posterior distribution as the target relationship.

4. The method according to claim 3, characterized in that The loss values ​​other than the p sampled loss values ​​among the m first loss values ​​are loss values ​​to be sampled, and determining, based on the target relationship, the first loss value corresponding to the second loss value satisfying a preset condition as a target loss value comprises: Based on the target relationship, determining an acquisition function, the acquisition function being used to estimate a value of the first loss value being sampled; Based on the acquisition function, determine the sampled value of the loss value to be sampled, and obtain (mp) sampling values; Based on the (mp) sampling values, q sampling loss values ​​are obtained by sampling from the (mp) loss values ​​to be sampled, where q is an integer less than or equal to (mp); Determine the second loss value that meets the preset condition from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values ​​as a reference loss value; The first loss value corresponding to the reference loss value is determined as the target loss value.

5. The method according to claim 4, characterized in that The acquisition function satisfies at least one of the following conditions: the mean of the second loss value corresponding to the first loss value in the target relationship is negatively correlated with the value of the first loss value sampled, the variance of the second loss value corresponding to the first loss value in the target relationship is positively correlated with the value of the first loss value sampled, and the value of the sampled loss value in the first loss value is lower than the value of the loss value other than the sampled loss value in the first loss value.

6. The method according to claim 4 or 5, characterized in that: The determining, from the second loss values ​​corresponding to the p sampled loss values ​​and the second loss values ​​corresponding to the q sampled loss values, the second loss value satisfying the preset condition as a reference loss value comprises: When the sum of p and q is greater than or equal to the sampling threshold, the minimum value of the second loss value corresponding to the p sampling loss values ​​and the second loss value corresponding to the q sampling loss values ​​is determined as the reference loss value.

7. The method according to any one of claims 1 to 5, characterized in that The first loss value is obtained based on the sum of squares of n first differences.

8. A training method, characterized in that: The method comprises: Obtain a model to be trained, wherein the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, and the target parameters are obtained according to the method according to any one of claims 1 to 7; Obtaining n training features and n labels of n training objects, wherein the labels represent individual actual feedback results of sending the target services to the training objects; Inputting the n training features into the model to be trained, so that the model to be trained predicts a feedback result of sending the target service to the training object, and obtaining n training feedback results; Based on the difference between the n training feedback results and the n labels, obtaining the training loss of the model to be trained; Based on the training loss, the parameters of the model to be trained are updated to obtain a prediction model.

9. A prediction method, characterized in that: The method comprises: Obtaining a set of features to be predicted, wherein the set of features to be predicted is a set of features to be predicted, and the features to be predicted are features of objects to be predicted in a group to be predicted; Inputting the feature set to be predicted into a prediction model so that the prediction model predicts sending individual prediction feedback results of the target business to the objects to be predicted in the group to be predicted, thereby obtaining an individual prediction feedback result set, wherein the prediction model is trained according to the method of claim 8; Based on the individual prediction feedback results in the individual prediction feedback result set, a group actual feedback result of sending the target service to the group to be predicted is obtained.

10. A data processing device, characterized in that: The data processing device is used to determine the target parameters of the target model, the target model is used to predict the target service feedback result based on the target characteristics of the target object in the target object set, the target service feedback result is the feedback result of the target service sent to the target object in the target object set, and the data processing device includes: an acquisition unit, for acquiring m first loss values, where m is an integer greater than 1, and the m first loss values ​​are obtained by inputting n training features of n training objects into m models to be confirmed, where n is an integer greater than 1, and the model structures of the m models to be confirmed are the same as the model structure of the target model, and the parameters of different models to be confirmed are different; the first loss value is obtained based on n first differences, and the first differences correspond to the training features one by one, and the first differences are the difference between the individual predicted feedback results of sending the target business to the training objects and the individual actual feedback results of sending the target business to the training objects, and the individual predicted feedback results are the feedback results of sending the target business to the training objects in the training object set predicted by the model to be confirmed; a determination unit, configured to determine a target relationship based on the m first loss values, the target relationship being a relationship between the first loss value and a second loss value, the second loss value being obtained based on a second difference, the second difference corresponding one-to-one to the model to be confirmed, the second difference being a difference between a group prediction feedback result of sending the target service to a target group and a group actual feedback result of sending the target service to the target group, the target group being a group including the n training objects; the group prediction feedback result being obtained based on the n individual prediction feedback results of the n training objects; The determining unit is configured to determine, based on the target relationship, that the first loss value corresponding to the second loss value that satisfies a preset condition is a target loss value; A processing unit is used to use the parameters of the model to be confirmed corresponding to the target loss value as the target parameters.

11. A training device, characterized in that: The training device comprises: An acquisition unit, used for acquiring a model to be trained, wherein the model structure of the model to be trained is the same as the model structure of the target model, and the parameters of the model to be trained are target parameters, and the target parameters are obtained according to the method according to any one of claims 1 to 7; The acquisition unit is used to acquire n training features and n labels of n training objects, wherein the labels represent individual actual feedback results of sending the target service to the training objects; A prediction unit, used for inputting the n training features into the model to be trained, so that the model to be trained predicts the feedback result of sending the target service to the training object, and obtains n training feedback results; A processing unit, configured to obtain a training loss of the model to be trained based on a difference between the n training feedback results and the n labels; An updating unit is used to update the parameters of the model to be trained based on the training loss to obtain a prediction model.

12. A prediction device, characterized in that: The prediction device comprises: An acquisition unit, used for acquiring a set of features to be predicted, wherein the set of features to be predicted is a set of features to be predicted, and the features to be predicted are features of objects to be predicted in a group to be predicted; A prediction unit, configured to input the feature set to be predicted into a prediction model, so that the prediction model predicts the individual prediction feedback result of sending the target business to the object to be predicted in the group to be predicted, and obtains an individual prediction feedback result set, wherein the prediction model is trained according to the method of claim 8; The processing unit is used to obtain a group actual feedback result of sending the target service to the group to be predicted based on the individual prediction feedback results in the individual prediction feedback result set.

13. An electronic device, characterized in that: include: A processor and a memory, wherein the memory is used to store computer program code, wherein the computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes the method as claimed in any one of claims 1 to 7, or the electronic device executes the method as claimed in claim 8, or the electronic device executes the method as claimed in claim 9.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the method described in any one of claims 1 to 7, or the processor executes the method described in claim 8, or the processor executes the method described in claim 9.

15. A computer program product, characterized in that The computer program product includes a computer program or instructions; when the computer program or instructions are run on a computer, the computer is enabled to execute the method described in any one of claims 1 to 7, or the computer is enabled to execute the method described in claim 8, or the computer is enabled to execute the method described in claim 9.

Images