Training Method, Device, Equipment and Storage Medium for Object Selection Model

Abstract

An embodiment of the present application discloses a training method, device, equipment and storage medium for an object selection model, belonging to the field of artificial intelligence technology. The method includes: determining a sample object, where the sample object is set with a target label and n associated labels, and there is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels; inputting the sample object information of the sample object into an MMOE model to obtain a first prediction result and n second prediction results output by the MMOE model; training the MMOE model based on the target label, n associated labels, the first prediction result, and n second prediction results; and generating a target object selection model corresponding to the target selection task based on the trained MMOE model.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular, to a training method, device, equipment and storage medium for an object selection model. Background Art

[0002] Through the method of object selection, it is possible to select target objects that meet the characteristics from sample objects, so as to make appropriate recommendations for the target objects.

[0003] In the related art, the label rule selection method is adopted, and by labeling the sample objects, the target objects that meet the characteristics are selected from them; or the tree model object selection method is adopted, and a corresponding function is fitted according to the existing target objects and object characteristics, and the target objects that meet the characteristics are determined from the sample objects according to the fitted function.

[0004] However, the label rule selection method is limited to the sample objects that have been labeled, and the tree model object selection method uses single-model technology, which cannot meet the accuracy requirements of object selection when the number of sample objects is large. Summary of the Invention

[0005] The embodiments of the present application provide a training method, device, equipment and storage medium for an object selection model. The technical solutions are as follows:

[0006] On the one hand, the embodiments of the present application provide a training method for an object selection model, and the method includes:

[0007] Determine sample objects, where the sample objects are set with a target label and n associated labels, and there is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels, and n is a positive integer;

[0008] Input the sample object information of the sample objects into a multi-task learning (MMOE) model to obtain a first prediction result and n second prediction results output by the MMOE model. The first prediction result is the prediction result corresponding to the target selection task, and the second prediction results are the prediction results corresponding to the associated selection tasks. And the expert network in the MMOE model includes a shared expert network, a first expert network and a second expert network. The first expert network is an expert network unique to the target selection task, the second expert network is an expert network unique to the associated selection tasks, and the shared expert network is an expert network shared by the target selection task and the associated selection tasks;

[0009] Train the MMOE model based on the target label, the n associated labels, the first prediction result, and the n second prediction results.

[0010] Generate a target object selection model corresponding to the target selection task based on the trained MMOE model.

[0011] On the other hand, an embodiment of the present application provides a training device for an object selection model. The device includes:

[0012] A sample determination module, configured to determine a sample object, where the sample object is set with a target label and n associated labels, and there is a correlation between the target selection task corresponding to the target label and the associated selection task corresponding to the associated label, and n is a positive integer;

[0013] A result prediction module, configured to input the sample object information of the sample object into the MMOE model to obtain a first prediction result and n second prediction results output by the MMOE model. The first prediction result is the prediction result corresponding to the target selection task, and the second prediction result is the prediction result corresponding to the associated selection task. Moreover, the expert network in the MMOE model includes a shared expert network, a first expert network, and a second expert network. The first expert network is an expert network unique to the target selection task, the second expert network is an expert network unique to the associated selection task, and the shared expert network is an expert network shared by the target selection task and the associated selection task;

[0014] A model training module, configured to train the MMOE model based on the target label, the n associated labels, the first prediction result, and the n second prediction results;

[0015] A model generation module, configured to generate a target object selection model corresponding to the target selection task based on the trained MMOE model.

[0016] On the other hand, an embodiment of the present application provides a computer device, which includes a processor and a memory; the memory stores at least one instruction, and the at least one instruction is used to be executed by the processor to implement the training method of the object selection model as described in the above aspect.

[0017] On the other hand, an embodiment of the present application provides a computer-readable storage medium, in which at least one program code is stored, and the program code is loaded and executed by the processor to implement the training method of the object selection model as described in the above aspect.

[0018] On the other hand, an embodiment of the present application provides a computer program product, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the training method of the object selection model provided in various alternative implementations of the above aspects.

[0019] In the embodiment of the present application, the computer device sets a target label and n associated labels for a determined sample object, and based on the MMOE model, simultaneously predicts the target selection task and n associated selection tasks, and then trains the MMOE model with the target label, n associated labels, the first prediction result, and n second prediction results, so as to generate a target object selection model corresponding to the target selection task. By adopting the solution provided in the embodiment of the present application, the computer device can add associated selection tasks related to the target selection task, and on the basis of the MMOE model provided with a shared expert network, set a unique expert network for different selection tasks, and predict different selection tasks through the shared expert network and each unique expert network, so as to train the MMOE model and generate a target object selection model corresponding to the target selection task, improving the quality of the target object selection model and further improving the accuracy of the target selection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0021] Figure 1 shows a flowchart of a training method of an object selection model provided by an exemplary embodiment of the present application;

[0022] Figure 2 is an implementation schematic diagram of result prediction by different expert networks shown in an exemplary embodiment of the present application;

[0023] Figure 3 shows a flowchart of a training method of an object selection model provided by another exemplary embodiment of the present application;

[0024] Figure 4 is an implementation schematic diagram of the training process of the MMOE model shown in another exemplary embodiment of the present application;

[0025] Figure 5 is an implementation schematic diagram of the target object selection model shown in another exemplary embodiment of the present application;

[0026] Figure 6 The flowchart of the training method of the object selection model provided by another exemplary embodiment of the present application is shown;

[0027] Figure 7 It is the structural block diagram of the training device of the object selection model provided by an exemplary embodiment of the present application;

[0028] Figure 8 The structural block diagram of the computer device provided by an exemplary embodiment of the present application is shown. Detailed implementation manners

[0029] To make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0030] As used herein, "a plurality of" means two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0031] In the related art, the label rule selection method is adopted. First, the computer device marks the feature labels of the sample objects, and based on the target label, selects the corresponding target objects from the sample objects. Since this method can only select the target objects that meet the target label from the limited sample objects with label marks, for the sample objects without label marks, the computer device cannot determine whether they meet the target label. Therefore, the label rule selection method is only applicable to selecting target objects from limited sample objects.

[0032] Adopting the tree model object selection method, the computer device fits the corresponding function based on the mapping relationship between the determined target object and the target label, and determines the target objects that meet the target label from the sample objects according to this function. However, this method is a single model technology, and the existing mapping relationship between the determined target object and the target label does not have universality and accuracy. In the case of a large number of sample objects, the target objects obtained only according to the fitted function do not have accuracy.

[0033] In the embodiments of the present application, the computer device trains the MMOE model by setting n associated selection tasks related to the target selection task, and using the target label, n associated labels, the first prediction result corresponding to the target selection task, and the n second prediction results corresponding to the n associated selection tasks, so as to generate the target object selection model corresponding to the target selection task according to the trained MMOE model, improving the accuracy of the target selection task.

[0034] Please refer to Figure 1 , which shows a flowchart of a method for training an object selection model provided by an exemplary embodiment of the present application. The method may include the following steps:

[0035] Step 101: Determine a sample object. The sample object is set with a target label and n associated labels. There is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels, where n is a positive integer.

[0036] First, the computer device determines the sample object according to the target selection task. To determine whether the sample object meets the selection target, the computer device sets a target label for the sample object, and this target label corresponds to the target selection task.

[0037] Meanwhile, to improve the accuracy of the target selection task, the computer device sets n associated selection tasks that are related to the target selection task, and correspondingly sets n associated labels for the sample object, where n is a positive integer.

[0038] In a possible implementation manner, the computer device sets one associated selection task that is related according to the target selection task. Meanwhile, the computer device sets a target label and one associated label for the determined sample object.

[0039] In an illustrative example, the target selection task is a task of selecting potential purchasers of goods. The computer device sets the associated selection task as a task of selecting potential clickers of goods advertisements. Meanwhile, the computer device determines the potential purchasers of goods as the sample object, and sets a potential purchase label for the potential purchasers of goods and a potential click label for goods advertisements.

[0040] Step 102: Input the sample object information of the sample object into the MMOE model to obtain a first prediction result and n second prediction results output by the MMOE model. The first prediction result is the prediction result corresponding to the target selection task, and the second prediction results are the prediction results corresponding to the associated selection tasks. Moreover, the expert network in the MMOE model includes a shared expert network, a first expert network, and a second expert network. The first expert network is an expert network specific to the target selection task, the second expert network is an expert network specific to the associated selection tasks, and the shared expert network is an expert network shared by the target selection task and the associated selection tasks.

[0041] In a possible implementation manner, the computer device inputs the sample object information of the sample object into the MMOE model. Thus, the MMOE model outputs a first prediction result based on the target selection task and n corresponding second prediction results based on the n associated selection tasks.

[0042] Optionally, the sample object information can be collected based on the target selection task and the associated selection task, and may include sample portrait information, application usage records, search term records, and other sample object information that can reflect the characteristics of the sample object. The embodiments of the present application do not limit this.

[0043] In a possible implementation manner, the expert network in the MMOE model includes a shared expert network, a first expert network, and a second expert network. Among them, the first expert network is an expert network unique to the target selection task and is used to learn the target selection task; the second expert network is an expert network unique to the associated selection task and is used to learn the associated selection task; the shared expert network is an expert network shared by the target selection task and the associated selection task and is used to learn the common parts of the target selection task and the associated selection task. Moreover, the shared expert network includes at least two expert networks and is used to learn different subtasks in the common tasks. Thus, the first prediction result output by the MMOE model is obtained based on the learning of the first expert network and the shared expert network, and the second prediction result is obtained based on the learning of the second expert network and the shared expert network.

[0044] Schematically, as Figure 2 shown, in the case where there is one associated selection task, the computer device inputs the sample object information 201 into the MMOE model. The shared expert network 202 and the first expert network 203 learn the target selection task to obtain the first prediction result 205, and the shared expert network 202 and the second expert network 204 learn the associated selection task to obtain the second prediction result 206.

[0045] Step 103: Train the MMOE model based on the target label, n associated labels, the first prediction result, and the n second prediction results.

[0046] To reduce the selection error of the target selection task, the computer device trains the MMOE model based on the target label, n associated labels, the first prediction result, and the n second prediction results.

[0047] In a possible implementation manner, during the training of the MMOE model, the computer device sets a loss function and, based on this loss function, repeatedly trains the MMOE model to improve the accuracy of the output of the MMOE model.

[0048] Step 104: Generate a target object selection model corresponding to the target selection task based on the trained MMOE model.

[0049] Furthermore, the computer device generates a target object selection model corresponding to the target selection task based on the trained MMOE model.

[0050] In a possible implementation, the computer device outputs the selection results corresponding to the target selection task and n associated selection tasks through the MMOE model. To further implement the selection of the target object, the computer device generates a corresponding target object selection model based on the part of the network corresponding to the target selection task in the MMOE model.

[0051] In summary, in the embodiments of the present application, the computer device sets a target label and n associated labels for the determined sample object, and based on the MMOE model, simultaneously predicts the target selection task and n associated selection tasks, and then trains the MMOE model with the target label, n associated labels, the first prediction result, and n second prediction results, so as to generate a target object selection model corresponding to the target selection task. By adopting the solution provided in the embodiments of the present application, the computer device can add associated selection tasks related to the target selection task, and based on the MMOE model with a shared expert network, set a unique expert network for different selection tasks, and predict different selection tasks through the shared expert network and each unique expert network, so as to train the MMOE model and generate a target object selection model corresponding to the target selection task, improving the quality of the target object selection model and further improving the accuracy of the target selection task.

[0052] To improve the selection accuracy of the target object selection model, the computer device needs to extract features from the sample object information of the sample object through the MMOE model and generate prediction results, and then repeatedly train the MMOE model. The following uses specific embodiments to detail the process of obtaining the sample object information and using the sample object information to train the MMOE model.

[0053] Please refer to Figure 3 , which shows a flowchart of a method for training an object selection model provided by another exemplary embodiment of the present application. The method may include the following steps:

[0054] Step 301, determine a sample object. The sample object is set with a target label and n associated labels. There is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels, and n is a positive integer.

[0055] The implementation of this step may refer to step 101 above, and this embodiment will not be elaborated here.

[0056] Step 302, obtain the original object information of the sample object.

[0057] The computer device obtains the original object information of the sample object, which is a direct manifestation of the characteristics of the sample object.

[0058] Optionally, the original object information may be portrait information of a sample object, application usage records, web browsing records, etc., and the embodiments of the present application do not limit this.

[0059] Step 303: Preprocess the original object information to obtain sample object information, where the preprocessing includes at least one of information aggregation processing and information gain processing.

[0060] Since the number of sample objects is large, the sample object features between individual sample objects are also diverse, and the obtained original object information is also different. Therefore, in order to classify the sample objects to different degrees, the computer device needs to preprocess the original object information to obtain sample object information, so that the computer device can perform different feature classifications on the sample objects according to the sample object information. The preprocessing includes at least one of information aggregation processing and information gain processing.

[0061] In a possible implementation manner, the original object information includes application usage records, and the application usage records include application package names.

[0062] Since the application package names of applications downloaded through different channels may be different, if the application package names are directly used as the classification criteria for sample objects to divide the sample objects, then the sample objects using the same application will be divided into different categories due to the differences in the application package names, and using the application package names as the classification criteria to divide the sample objects, the obtained classification categories are too detailed and the distinctions are not obvious. Therefore, the computer device processes the application package names through information aggregation processing, so as to divide the sample objects with the application names as the sample object information.

[0063] In a possible implementation manner, the information aggregation processing of the application package names may include the following steps:

[0064] 1. Obtain the application package names included in the original object information.

[0065] First, the computer device obtains the application package names included in the original object information, and the application package names are the application package names corresponding to the applications actually used by the sample objects.

[0066] 2. Cluster the application package names based on the package name similarity between the application package names to obtain application clusters.

[0067] For the different application package names, the computer device clusters the application package names based on the package name similarity between the application package names to obtain application clusters.

[0068] Optionally, since clustering application package names based on package name similarity requires the premise that the same application is represented by the package name, the computer device only clusters application package names with relatively high package name similarity.

[0069] 3. Replace the application package name in the original object information with the application name corresponding to the application cluster to which the application package name belongs.

[0070] Further, the computer device replaces the application package name in the original object information with the application name corresponding to the application cluster to which the application package name belongs, so that the application name represents the sample object characteristics as the sample object information.

[0071] In a possible implementation, the original object information includes portrait information, and the portrait information contains data to be mapped, and the data to be mapped is used to be mapped to a data interval.

[0072] Optionally, the data to be mapped in the portrait information may include height, weight, age, etc., and the embodiments of the present application do not limit this.

[0073] In the related art, the computer device maps the data to be mapped with a fixed data interval, which increases the uncertainty of feature selection in the data interval. In the embodiments of the present application, by performing information gain processing on the data to be mapped, a data interval that conforms to the characteristics of the data to be mapped is determined, and the data to be mapped with similar features can be divided into the same data interval, improving the rationality of data interval division.

[0074] In a possible implementation, performing information gain processing on the data to be mapped may include the following steps:

[0075] 1. Obtain the data to be mapped included in the original object information.

[0076] First, the computer device obtains the data to be mapped included in the original object information, and the data to be mapped directly represents the portrait characteristics of the sample object.

[0077] 2. Determine the data to be mapped and the sample label of the sample object to which the data to be mapped belongs, and the sample label includes a target label and an associated label.

[0078] The computer device confirms the data to be mapped and obtains the sample label of the sample object to which the data to be mapped belongs, and the sample label includes a target label and an associated label.

[0079] Illustratively, taking the data to be mapped as age and the target label as purchasing goods as an example, the computer device determines the age data of the sample object and the situation of purchasing goods, and obtains the sample data as (25, 0), (27, 1), (30, 0), (39, 1), (42, 1), where 0 indicates not purchasing goods and 1 indicates purchasing goods.

[0080] 3. Determine the original information entropy based on the data to be mapped and the sample labels.

[0081] In order to determine the optimal interval division point during the interval division process, the computer device determines the original information entropy based on the data to be mapped and the sample labels, and calculates the information gain with the original information entropy as the reference.

[0082] Illustratively, taking the data to be mapped as age and the sample label as purchasing goods as an example, the original information entropy can be expressed as:

[0083]

[0084] 4. Map the data to be mapped to the candidate data intervals based on the candidate interval division methods.

[0085] In order to find the optimal interval division point, the computer device uses each data to be mapped as a candidate interval division point to divide the data to be mapped, so that the computer device maps the data to be mapped to the candidate data intervals based on the candidate interval division methods.

[0086] 5. Determine the candidate information entropy based on the candidate data intervals and the sample labels.

[0087] Furthermore, the computer device determines the current candidate information label based on the current candidate data intervals and the sample labels.

[0088] Illustratively, taking age 30 as the interval division point to divide the data to be mapped, the obtained candidate information entropy is: Info(x≤30) = P1Entropy(x≤30) + P2Entropy(x>30)

[0089]

[0090]

[0091] Info(x≤30) = 0.551

[0092] 6. Determine the information gain value corresponding to the candidate interval division method based on the original information entropy and the candidate information entropy.

[0093] Based on the calculated original information entropy and candidate information entropy, the computer device determines the information gain value corresponding to the current candidate interval division method, where the information gain value is the difference between the original information entropy and the candidate information entropy.

[0094] Illustratively, taking age 30 as the interval division point, the obtained information gain value can be expressed as:

[0095] Entropy(y) - Info(x≤30) = 0.97 - 0.551 = 0.419

[0096] 7. Determine the target interval partitioning method based on the information gain value.

[0097] By taking each data to be mapped as a candidate interval partitioning point and calculating the information gain value, according to the rule that the larger the information gain value, the more stable the candidate interval partitioning state, the computer device determines the target interval partitioning method.

[0098] Step 304, perform embedding processing on the sample object information of the sample object through the embedding network of the MMOE model to obtain embedded features.

[0099] Since the sample object information is presented in the form of text data, and the MMOE model can only process numerical vectors, when inputting the sample object information into the MMOE model, the computer device performs embedding processing on the sample object information of the sample object through the embedding network of the MMOE model, performs text vectorization processing on the sample object information, and thus obtains the embedded features corresponding to the sample object information.

[0100] Schematically, as Figure 4 shown, the computer device inputs the sample object information into the MMOE model, performs embedding processing on the sample object information through the embedding network 401, and thus obtains the embedded features corresponding to the sample object information for the MMOE model to perform corresponding feature extraction.

[0101] Step 305, perform feature extraction on the embedded features through the shared expert network, the first expert network, and the first gate network to obtain first intermediate features.

[0102] Furthermore, in order to obtain the correspondence between the target circle selection task and the sample object information, the computer device performs feature extraction on the embedded features through the shared expert network, the first expert network, and the first gate network, and thus obtains first intermediate features.

[0103] In a possible implementation manner, this step may include the following sub-steps:

[0104] Step 305A, input the embedded features into the shared expert network to obtain the shared features output by each shared expert network.

[0105] In a possible implementation manner, the computer device inputs the embedded features into the shared expert network. The shared expert network performs feature extraction on the embedded features and learns the embedded features, and then each shared expert network outputs the extracted shared features.

[0106] Schematically, as Figure 4As shown, the computer device extracts features from the embedded features through the first shared expert network 4021, the second shared expert network 4022, and the third shared expert network 4023 in the shared expert network. Among them, each shared expert network extracts features from different aspects of the embedded features, thereby obtaining shared features.

[0107] Step 305B: Input the embedded features into the first gating network to obtain the first weight set output by the first gating network. The first weight set contains the weights of each shared feature.

[0108] Since each shared expert network is responsible for extracting features from different aspects of the embedded features, for different circled selection tasks, the weights of the corresponding shared features are also different. Therefore, the computer device needs to obtain the weights of each shared feature through the gating network corresponding to different circled selection tasks.

[0109] In a possible implementation, the computer device inputs the embedded features into the first gating network corresponding to the target circled selection task. The first gating network extracts features from the embedded features to obtain the first weight set output by the first gating network. Among them, the first weight set contains the weights of each shared feature.

[0110] Schematically, as Figure 4 shown, the computer device inputs the embedded features into the first gating network 405. The first gating network 405 extracts features from the embedded features, thereby obtaining the first weight set.

[0111] Step 305C: Input the embedded features into the first expert network to obtain the first unique feature output by the first expert network.

[0112] To obtain the unique features of the target circled selection task, the computer device inputs the embedded features into the first expert network corresponding to the target circled selection task. The first expert network extracts features from the embedded features and learns the embedded features, and then obtains the first unique feature output by the first expert network.

[0113] Schematically, as Figure 4 shown, the computer device inputs the embedded features into the first expert network 403 corresponding to the target circled selection task. The first expert network 403 extracts features from the embedded features and learns the embedded features, and then obtains the first unique feature output by the first expert network 403.

[0114] Step 305D: Based on the first weight set, perform feature fusion on the shared features to obtain the first fused feature.

[0115] Since the weights of the respective shared features are different in different circle selection tasks, the computer device performs feature fusion on the shared features output by the shared expert network based on the first weight set to obtain the first fused feature corresponding to the target circle selection task.

[0116] Schematically, as Figure 4 shown, the computer device performs feature fusion on the shared features output by the first shared expert network 4021, the second shared expert network 4022, and the third shared expert network 4023 based on the first weight set output by the first gate network 405, thereby obtaining the first fused feature.

[0117] Step 305E, generating a first intermediate feature based on the first fused feature and the first specific feature.

[0118] Furthermore, the computer device generates a first intermediate feature corresponding to the target circle selection task based on the first fused feature and the first specific feature.

[0119] Step 306, inputting the first intermediate feature into the first tower network to obtain a first prediction result output by the first tower network.

[0120] To enable the gate networks corresponding to different circle selection tasks to learn different shared expert network combination patterns and obtain the relationship between different circle selection tasks and the shared expert network, the computer device inputs the first intermediate feature into the first tower network, and the first tower network corresponds to the target circle selection task, thereby obtaining a first prediction result output by the first tower network.

[0121] Schematically, as Figure 4 shown, the computer device inputs the first intermediate feature into the first tower network 407 corresponding to the target circle selection task, thereby obtaining a first prediction result 409 output by the first tower network 407.

[0122] Step 307, performing feature extraction on the embedded feature through the shared expert network, n second expert networks, and n second gate networks to obtain n second intermediate features.

[0123] Meanwhile, to obtain the correspondence between each associated circle selection task and the sample object information, the computer device performs feature extraction on the embedded feature through the shared expert network, n second expert networks, and n second gate networks, thereby obtaining n second intermediate features.

[0124] In a possible implementation manner, for each associated circle selection task, this step may include the following sub-steps:

[0125] Step 307A, inputting the embedded feature into the shared expert network to obtain the shared features output by each shared expert network.

[0126] In a possible implementation, the computer device inputs the embedded features into the shared expert network. The shared expert network extracts features from the embedded features and learns the embedded features. Then, each shared expert network outputs the extracted shared features.

[0127] Schematically, as Figure 4 shown, the computer device extracts features from the embedded features through the first shared expert network 4021, the second shared expert network 4022, and the third shared expert network 4023 in the shared expert network. Among them, each shared expert network extracts features from different aspects of the embedded features, so as to obtain the shared features.

[0128] Step 307B: Input the embedded features into the second gate network to obtain the second weight set output by the second gate network. The second weight set contains the weights of each shared feature.

[0129] In order to obtain the weights of each shared feature corresponding to each associated circle selection task, the computer device inputs the embedded features into the second gate network. The second gate network extracts features from the embedded features to obtain the second weight set corresponding to the associated circle selection task output by the second gate network.

[0130] Schematically, as Figure 4 shown, the computer device inputs the embedded features into the second gate network 406. The second gate network 406 extracts features from the embedded features, so as to obtain the second weight set.

[0131] Step 307C: Input the embedded features into the second expert network to obtain the second unique feature output by the second expert network.

[0132] In order to obtain the unique features of the associated circle selection task, the computer device inputs the embedded features into the second expert network. The second expert network extracts features from the embedded features and learns the embedded features. Then, the second unique feature corresponding to the associated circle selection task output by the second expert network is obtained.

[0133] Schematically, as Figure 4 shown, the computer device inputs the embedded features into the second expert network 404 corresponding to the associated circle selection task. The second expert network 404 extracts features from the embedded features and learns the embedded features. Then, the second unique feature output by the second expert network 404 is obtained.

[0134] Step 307D: Based on the second weight set, perform feature fusion on the shared features to obtain the second fused feature.

[0135] Since the weights of the respective shared features are different in different circle selection tasks, the computer device performs feature fusion on the shared features output by the shared expert network based on the second weight set to obtain the second fusion feature corresponding to the associated circle selection task.

[0136] Illustratively, as Figure 4 shown, the computer device performs feature fusion on the shared features output by the first shared expert network 4021, the second shared expert network 4022, and the third shared expert network 4023 based on the second weight set output by the second gate network 406, thereby obtaining the second fusion feature.

[0137] Step 307E, generating a second intermediate feature based on the second fusion feature and the second specific feature.

[0138] Furthermore, the computer device generates a second intermediate feature corresponding to the associated circle selection task based on the second fusion feature and the second specific feature.

[0139] Step 308, inputting the n second intermediate features into n second tower networks to obtain n second prediction results output by the n second tower networks.

[0140] Similarly, the computer device inputs the n second intermediate features into the corresponding n second tower networks, and the second tower networks correspond to n associated circle selection tasks, thereby obtaining n second prediction results output by the n second tower networks.

[0141] Illustratively, as Figure 4 shown, the computer device inputs the second intermediate feature into the second tower network 408 corresponding to the associated circle selection task, thereby obtaining the second prediction result 410 output by the second tower network 408.

[0142] Step 309, training the MMOE model based on the target label, the n associated labels, the first prediction result, and the n second prediction results.

[0143] The implementation manner of this step may refer to the above step 103, and this embodiment will not be elaborated herein.

[0144] Step 310, generating a target object circle selection model based on the shared expert network, the first expert network, the first gate network, and the first tower network in the trained MMOE model.

[0145] Furthermore, the computer device generates a target object circle selection model based on the shared expert network, the first expert network, the first gate network, and the first tower network in the trained MMOE model.

[0146] Illustratively, as Figure 5As shown in the figure, the computer device generates a target object selection model corresponding to the target selection task based on the embedded network 501, the first shared expert network 5021, the second shared expert network 5022, the third shared expert network 5023, the first expert network 503, the first gate network 504, and the first tower network 505.

[0147] In the above embodiment, the computer device preprocesses the original object information to obtain the sample object information, and can obtain more unified and regular features of the sample object, making the feature mapping between the sample object and the sample object information more standardized. At the same time, the computer device extracts and learns features for different selection tasks based on the shared expert network, the first expert network, and the second expert network in the MMOE model, and uses the weight set output by the gate network corresponding to different selection tasks to perform weight fusion on the shared features output by the shared expert network, obtaining the fusion features of different selection tasks, thereby improving the accuracy of the prediction results of the MMOE model for different selection tasks.

[0148] In a possible implementation manner, the computer device selects the target object based on the generated target object selection model. This process may include the following steps:

[0149] 1. Obtain the target object information of the target object.

[0150] In a possible implementation manner, the computer device obtains the target object information of the target object, and the target object information may include target object portrait information, application usage records, search and browsing records, etc., and the embodiments of the present application do not limit this.

[0151] 2. Input the target object information into the target object selection model to obtain the selection result output by the target object selection model.

[0152] Further, the computer device inputs the target object information into the target object selection model, and outputs the selection results corresponding to each target object through the target object selection model. Thus, based on this selection result, the computer device selects the selected object that meets the target selection result from the target objects according to the set target selection result.

[0153] Combined with the above embodiment, taking the target selection task as the task of selecting potential purchase objects of goods, and the associated selection tasks at least including the task of selecting potential click objects of goods advertisements as an example for illustration.

[0154] Step 601, determine the sample object, and the sample object is set with a goods purchase label and a goods advertisement click label.

[0155] First, the computer device determines sample objects, and sets a product purchase label and a product advertisement click label for the sample objects based on the task of selecting potential product purchase objects and the task of selecting potential click objects for product advertisements.

[0156] In a possible implementation, the sample objects can be determined as objects that have purchased products in the last month, objects that have clicked on product advertisements but have not purchased products in the last month, and objects that have not purchased products and have not clicked on product advertisements in the last month. Thus, the computer device sets a product purchase label and a product advertisement click label for the sample objects. The product purchase label corresponding to the object that has purchased products in the last month is 1, the product purchase label corresponding to the object that has not purchased products in the last month is 0, the product advertisement click label corresponding to the object that has clicked on product advertisements in the last month is 1, and the product advertisement click label corresponding to the object that has not clicked on product advertisements in the last month is 0.

[0157] Step 602: Obtain the original object information of the sample objects. The original object information includes at least one of original application usage information, original historical search information, and original portrait information.

[0158] The computer device obtains the original object information of the sample objects. The original object information includes at least one of original application usage information, original historical search information, and original portrait information. Among them, the original portrait information can include age, gender, etc., the original application usage information can include application package names, application usage durations, etc., and the original historical search information can include search term information, search web page browsing durations, etc.

[0159] Step 603: Preprocess the original object information to obtain sample object information.

[0160] The computer device preprocesses the original object information to obtain sample object information. The sample object information includes at least one of application usage information, historical search information, and portrait information.

[0161] In an illustrative example, the computer device first obtains the original application usage information of the sample objects, filters out applications with an installation volume less than 1000 according to the application installation volume, and after performing information aggregation processing on the application package names, obtains the application names as the sample object information. Furthermore, the computer device establishes an index based on the application names, numbers the application names in sequence, the size of the index is 99999, and then counts the application usage records of the sample objects in the last month, sorts the order of the sample objects using applications according to the relationship of the time intervals from near to far from the information acquisition time, intercepts the first 100 usage records of the sample object application usage records, and for the part less than 100, supplements it with the index value 100000.

[0162] In a schematic example, the computer device obtains the original historical search information of the sample object, creates an index based on the search terms, numbers the search terms in sequence, the size of the index is 99999, and then counts the search term records of the sample object in the most recent month, sorts them according to the number of occurrences of the search terms, and intercepts the first 200 search terms of the search term records of the sample object. For the part less than 200, the index value 200000 is used as a supplement.

[0163] In a schematic example, the computer device obtains the original portrait information of the sample object, including the gender and age of the sample object. Among them, the male gender is numbered 200010, the female gender is numbered 200011, and the null value is supplemented with 200012. Then, information gain processing is performed on the age to obtain the optimal age division intervals as 0 - 17, 18 - 22, 23 - 28, 29 - 32, 33 - 40, 41 - 45, 46 - 59, 60 and above. Thus, the computer device numbers the eight age groups as 200001 - 200008 respectively, and the null value is supplemented with 200009.

[0164] Step 604: Input the sample object information of the sample object into the MMOE model to obtain the first prediction result and the second prediction result output by the MMOE model.

[0165] The computer device inputs the processed sample object information into the MMOE model, and through the MMOE model, obtains the first prediction result corresponding to the task of selecting potential purchasers of goods, and the second prediction result corresponding to the task of selecting potential clickers of goods advertisements.

[0166] Step 605: Train the MMOE model based on the product purchase label, product advertisement click label, first prediction result, and second prediction result.

[0167] In a possible implementation manner, the computer device trains the MMOE model based on the product purchase label, product advertisement click label, first prediction result corresponding to the task of selecting potential purchasers of goods, and second prediction result corresponding to the task of selecting potential clickers of goods advertisements.

[0168] Step 606: Generate a model for selecting potential purchasers of goods based on the shared expert network, first expert network, first gate network, and first tower network in the trained MMOE model.

[0169] Furthermore, the computer device generates a model for selecting potential purchasers of goods based on the shared expert network, first expert network, first gate network, and first tower network in the trained MMOE model.

[0170] Step 607, obtain the target object information of the target object.

[0171] In a possible implementation manner, the computer device obtains the target object information of the target object, and the target object information includes at least one of application usage information, historical search information, and portrait information.

[0172] Step 608, input the target object information into the commodity potential purchase object selection model, and obtain the selection result output by the commodity potential purchase object selection model.

[0173] Furthermore, the computer device inputs the target object information into the commodity potential purchase object selection model, and outputs the selection result of the target object purchasing the commodity through this model. Thus, the computer device selects the target object that reaches the target selection result, and the selected target object is the commodity potential purchase object.

[0174] In a schematic example, the computer device sets the target selection result to 0.75. When the target object selection result output by the commodity potential purchase object selection model is 0.8, the computer device selects this target object.

[0175] In the above embodiment, when the target selection task is the commodity potential purchase object selection task and the associated selection task is the commodity advertisement potential click object selection task, the computer device processes different original object information respectively, improving the correspondence between the sample object and the sample object information.

[0176] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.), and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions. For example, the portrait information, application usage information, historical search information, etc. involved in this application are all obtained under full authorization.

[0177] Please refer to Figure 7 , which shows the structural block diagram of the training device of the object selection model provided by an exemplary embodiment of this application. The device may include the following structures:

[0178] The sample determination module 701 is used to determine a sample object, and the sample object is set with a target label and n associated labels. There is a correlation between the target selection task corresponding to the target label and the associated selection task corresponding to the associated label, and n is a positive integer;

[0179] A result prediction module 702, configured to input the sample object information of the sample object into an MMOE model, and obtain a first prediction result and n second prediction results output by the MMOE model. The first prediction result is the prediction result corresponding to the target selection task, the second prediction results are the prediction results corresponding to the associated selection tasks, and the expert network in the MMOE model includes a shared expert network, a first expert network, and a second expert network. The first expert network is an expert network unique to the target selection task, the second expert network is an expert network unique to the associated selection tasks, and the shared expert network is an expert network shared by the target selection task and the associated selection tasks;

[0180] A model training module 703, configured to train the MMOE model based on the target label, n associated labels, the first prediction result, and n second prediction results;

[0181] A model generation module 704, configured to generate a target object selection model corresponding to the target selection task based on the trained MMOE model.

[0182] Optionally, the MMOE model further includes a first gate network and a first tower network corresponding to the target selection task, and a second gate network and a second tower network corresponding to each of the associated selection tasks;

[0183] The result prediction module 702 includes:

[0184] An embedding processing unit, configured to perform embedding processing on the sample object information of the sample object through an embedding network of the MMOE model to obtain embedded features;

[0185] A first feature extraction unit, configured to perform feature extraction on the embedded features through the shared expert network, the first expert network, and the first gate network to obtain first intermediate features;

[0186] A first result prediction unit, configured to input the first intermediate features into the first tower network to obtain the first prediction result output by the first tower network;

[0187] A second feature extraction unit, configured to perform feature extraction on the embedded features through the shared expert network, n second expert networks, and n second gate networks to obtain n second intermediate features;

[0188] A second result prediction unit, configured to input n second intermediate features into n second tower networks to obtain n second prediction results output by the n second tower networks.

[0189] Optionally, the first feature extraction unit is configured to:

[0190] Input the embedded feature into the shared expert network to obtain the shared features output by each of the shared expert networks;

[0191] Input the embedded feature into the first gate network to obtain the first weight set output by the first gate network, where the first weight set includes the weights of each of the shared features;

[0192] Input the embedded feature into the first expert network to obtain the first unique feature output by the first expert network;

[0193] Perform feature fusion on the shared features based on the first weight set to obtain a first fused feature;

[0194] Generate the first intermediate feature based on the first fused feature and the first unique feature.

[0195] Optionally, for each of the associated circle selection tasks, the second feature extraction unit is configured to:

[0196] Input the embedded feature into the shared expert network to obtain the shared features output by each of the shared expert networks;

[0197] Input the embedded feature into the second gate network to obtain the second weight set output by the second gate network, where the second weight set includes the weights of each of the shared features;

[0198] Input the embedded feature into the second expert network to obtain the second unique feature output by the second expert network;

[0199] Perform feature fusion on the shared features based on the second weight set to obtain a second fused feature;

[0200] Generate the second intermediate feature based on the second fused feature and the second unique feature.

[0201] Optionally, the model generation module 704 is configured to:

[0202] Generate the target object circle selection model based on the shared expert network, the first expert network, the first gate network, and the first tower network in the trained MMOE model.

[0203] Optionally, the apparatus further includes:

[0204] An information acquisition module, configured to acquire the original object information of the sample object;

[0205] An information processing module, configured to preprocess the original object information to obtain the sample object information, where the preprocessing includes at least one of information aggregation processing and information gain processing.

[0206] Optionally, the original object information includes application usage records, and the application usage records include application package names;

[0207] The information processing module is configured to:

[0208] Obtain the application package names included in the original object information;

[0209] Cluster the application package names based on the package name similarity between the application package names to obtain application clusters;

[0210] Replace the application package names in the original object information with the application names corresponding to the application clusters to which the application package names belong.

[0211] Optionally, the original object information includes portrait information, and the portrait information contains data to be mapped, and the data to be mapped is used to be mapped to a data interval;

[0212] The information processing module is configured to:

[0213] Obtain the data to be mapped included in the original object information;

[0214] Determine the data to be mapped and the sample labels of the sample objects to which the data to be mapped belongs, where the sample labels include the target labels and the associated labels;

[0215] Determine the original information entropy based on the data to be mapped and the sample labels;

[0216] Map the data to be mapped to a candidate data interval based on a candidate interval partitioning method;

[0217] Determine the candidate information entropy based on the candidate data interval and the sample labels;

[0218] Determine the information gain value corresponding to the candidate interval partitioning method based on the original information entropy and the candidate information entropy;

[0219] Determine the target interval partitioning method based on the information gain value.

[0220] Optionally, the apparatus further includes:

[0221] An information acquisition module, configured to acquire the target object information of a target object;

[0222] A result output module, configured to input the target object information into the target object selection model, and obtain a selection result output by the target object selection model.

[0223] Optionally, the target selection task is a task of selecting potential purchase objects of a commodity, and the associated selection tasks at least include a task of selecting potential click objects of a commodity advertisement;

[0224] The sample object information includes at least one of application usage information, historical search information, and portrait information.

[0225] In summary, in the embodiment of the present application, a computer device sets a target label and n associated labels for a determined sample object, and based on the MMOE model, simultaneously predicts a target selection task and n associated selection tasks, and then trains the MMOE model with the target label, n associated labels, a first prediction result, and n second prediction results, so as to generate a target object selection model corresponding to the target selection task. By adopting the solution provided in the embodiment of the present application, the computer device can add an associated selection task related to the target selection task, and on the basis of the MMOE model provided with a shared expert network, set a unique expert network for different selection tasks, and predict different selection tasks through the shared expert network and each unique expert network, so as to train the MMOE model, generate a target object selection model corresponding to the target selection task, improve the quality of the target object selection model, and further improve the accuracy of the target selection task.

[0226] It should be noted that: for the device provided in the above embodiment, only the above division of each functional module is used for illustration. In actual application, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device provided in the above embodiment and the method embodiment belong to the same concept, and the implementation process is detailed in the method embodiment, which will not be repeated here.

[0227] Please refer to Figure 8 , which shows a structural block diagram of a computer device provided in an exemplary embodiment of the present application. The computer device 1200 may include one or more of the following components: a processor 1210 and a memory 1220.

[0228] The processor 1210 may include one or more processing cores. The processor 1210 connects various parts within the entire computer device 1200 through various interfaces and circuits. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 1220, and by invoking data stored in the memory 1220, it performs various functions of the computer device 1200 and processes data. Optionally, the processor 1210 may be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). The processor 1210 may integrate one or a combination of several of a central processing unit (CPU), a graphics processing unit (GPU), a neural-network processing unit (NPU), and a modem, etc. Among them, the CPU mainly processes the operating system, user interface, application programs, etc.; the GPU is responsible for rendering and drawing the content to be displayed on the touch display screen; the NPU is used to implement artificial intelligence (AI) functions; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 1210 and may be implemented separately by a single chip.

[0229] The memory 1220 may include random access memory (RAM) and may also include read-only memory (ROM). Optionally, the memory 1220 includes a non-transitory computer-readable storage medium. The memory 1220 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1220 may include a program storage area and a data storage area. Among them, the program storage area may store instructions for implementing the operating system, instructions for at least one function (such as touch control function, sound playback function, image playback function, etc.), instructions for implementing the following various method embodiments, etc.; the data storage area may store data created according to the use of the computer device 1200 (such as audio data, phone book), etc.

[0230] In addition, those skilled in the art can understand that the structure of the computer device 1200 shown in the above drawings does not constitute a limitation on the computer device. The computer device may include more or fewer components than shown in the drawings, or combine some components, or have different component arrangements. For example, the computer device 1200 also includes components such as a display screen, a camera, a microphone, a speaker, a radio frequency circuit, an input unit, sensors (such as an acceleration sensor, an angular velocity sensor, a light sensor, etc.), an audio circuit, a WiFi module, a power supply, a Bluetooth module, etc., which will not be elaborated here.

[0231] The embodiment of the present application also provides a computer-readable storage medium, which stores at least one piece of program code, and the program code is loaded and executed by a processor to implement the training method of the object selection model as described in each of the above embodiments.

[0232] The embodiment of the present application provides a computer program product, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the training method of the object selection model provided in various alternative implementation manners of the above aspects.

[0233] It should be understood that "a plurality of" mentioned herein refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after. In addition, the step numbers described herein only exemplarily show a possible execution sequence between steps. In some other embodiments, the above steps may not be executed in the order of the numbers. For example, two steps with different numbers are executed simultaneously, or two steps with different numbers are executed in the reverse order of the illustration. The embodiments of the present application do not limit this.

[0234] The above are only alternative embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A training method for an object selection model, characterized in that, The method includes: Determine a sample object, where the sample object is set with a target label and n associated labels, and there is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels, and n is a positive integer; Perform embedding processing on the sample object information of the sample object through the embedding network of the MMOE model to obtain an embedded feature; perform feature extraction on the embedded feature through the shared expert network, the first expert network, and the first gate network of the MMOE model to obtain a first intermediate feature; input the first intermediate feature into the first tower network of the MMOE model to obtain a first prediction result output by the first tower network; perform feature extraction on the embedded feature through the shared expert network, n second expert networks, and n second gate networks of the MMOE model to obtain n second intermediate features; input the n second intermediate features into n second tower networks to obtain n second prediction results output by the n second tower networks. The first prediction result is the prediction result corresponding to the target selection task, the second prediction result is the prediction result corresponding to the associated selection task, the first expert network is an expert network unique to the target selection task, the second expert network is an expert network unique to the associated selection task, and the shared expert network is an expert network shared by the target selection task and the associated selection task; Train the MMOE model based on the target label, the n associated labels, the first prediction result, and the n second prediction results; Generate a target object selection model corresponding to the target selection task based on the shared expert network, the first expert network, the first gate network, and the first tower network in the trained MMOE model.

2. The method according to claim 1, characterized in that, The performing feature extraction on the embedded feature through the shared expert network, the first expert network, and the first gate network of the MMOE model to obtain a first intermediate feature includes: Input the embedded feature into the shared expert network to obtain shared features output by each of the shared expert networks; Input the embedded feature into the first gate network to obtain a first weight set output by the first gate network, where the first weight set contains the weights of each of the shared features; Input the embedded feature into the first expert network to obtain a first unique feature output by the first expert network; Perform feature fusion on the shared features based on the first weight set to obtain a first fused feature; Generate the first intermediate feature based on the first fused feature and the first unique feature.

3. The method according to claim 1, characterized in that, For each of the associated selection tasks, the performing feature extraction on the embedded feature through the shared expert network, n second expert networks, and n second gate networks of the MMOE model to obtain n second intermediate features includes: Input the embedded feature into the shared expert network to obtain shared features output by each of the shared expert networks; Input the embedded features into the second gating network to obtain a second weight set output by the second gating network, where the second weight set contains the weights of each of the shared features; Input the embedded features into the second expert network to obtain a second unique feature output by the second expert network; Perform feature fusion on the shared features based on the second weight set to obtain a second fused feature; Generate the second intermediate feature based on the second fused feature and the second unique feature.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Obtain the original object information of the sample object; Preprocess the original object information to obtain the sample object information, where the preprocessing includes at least one of information aggregation processing and information gain processing.

5. The method according to claim 4, characterized in that, The original object information includes application usage records, and the application usage records include application package names; The preprocessing the original object information to obtain the sample object information includes: Obtain the application package names included in the original object information; Cluster the application package names based on the package name similarity between the application package names to obtain application clusters; Replace the application package names in the original object information with the application names corresponding to the application clusters to which the application package names belong.

6. The method according to claim 4, characterized in that, The original object information includes portrait information, and the portrait information contains data to be mapped, and the data to be mapped is used to be mapped to a data interval; The preprocessing the original object information to obtain the sample object information includes: Obtain the data to be mapped included in the original object information; Determine the data to be mapped and the sample label of the sample object to which the data to be mapped belongs, where the sample label includes the target label and the associated label; Determine the original information entropy based on the data to be mapped and the sample label; Map the data to be mapped to a candidate data interval based on a candidate interval partitioning method; Determine a candidate information entropy based on the candidate data interval and the sample label; Determine an information gain value corresponding to the candidate interval partitioning method based on the original information entropy and the candidate information entropy; Determine a target interval partitioning method based on the information gain value.

7. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Obtain the target object information of the target object; Input the target object information into the target object selection model to obtain a selection result output by the target object selection model.

8. The method according to any one of claims 1 to 3, characterized in that, The target selection task is a task of selecting potential purchasers of goods, and the associated selection tasks include at least a task of selecting potential click objects for commodity advertisements; The sample object information includes at least one of application usage information, historical search information, and portrait information.

9. A training device for an object selection model, characterized in that, The device includes: A sample determination module, configured to determine a sample object, where the sample object is set with a target label and n associated labels, and there is a correlation between the target selection task corresponding to the target label and the associated selection tasks corresponding to the associated labels, and n is a positive integer; A result prediction module, configured to perform embedding processing on the sample object information of the sample object through the embedding network of the MMOE model to obtain embedded features; perform feature extraction on the embedded features through the shared expert network, the first expert network, and the first gate network of the MMOE model to obtain first intermediate features; input the first intermediate features into the first tower network of the MMOE model to obtain a first prediction result output by the first tower network; perform feature extraction on the embedded features through the shared expert network, n second expert networks, and n second gate networks of the MMOE model to obtain n second intermediate features; input the n second intermediate features into n second tower networks to obtain n second prediction results output by the n second tower networks, where the first prediction result is the prediction result corresponding to the target circle selection task, the second prediction result is the prediction result corresponding to the associated circle selection task, and the first expert network is an expert network unique to the target circle selection task, the second expert network is an expert network unique to the associated circle selection task, and the shared expert network is an expert network shared by the target circle selection task and the associated circle selection task; A model training module, configured to train the MMOE model based on the target label, n associated labels, the first prediction result, and n second prediction results; A model generation module, configured to generate a target object circle selection model corresponding to the target circle selection task based on the shared expert network, the first expert network, the first gate network, and the first tower network in the trained MMOE model.

10. A computer device, characterized in that, The computer device includes a processor and a memory; the memory stores at least one instruction, and the at least one instruction is used to be executed by the processor to implement the training method of the object circle selection model according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, At least one program code is stored in the computer-readable storage medium, and the program code is loaded and executed by the processor to implement the training method of the object circle selection model according to any one of claims 1 to 8.

12. A computer program product, characterized in that, The computer program product includes computer instructions, and the computer instructions are stored in a computer-readable storage medium; the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to enable the computer device to implement the training method of the object circle selection model according to any one of claims 1 to 8.

Images