Data processing method and device, computer device, storage medium and computer program product

Abstract

The application relates to a data processing method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring a first sample set and a second sample set; performing feature extraction on first feature data of the samples based on a first network of an initial multi-task model to obtain first features; performing feature encoding on second feature data of the samples based on a second network of the initial multi-task model to obtain second initial features; performing feature fusion on the first features and the second initial features to obtain second features; determining prediction results corresponding to each task of the initial multi-task model based on the first features and the second features; and iteratively adjusting network parameters of the initial multi-task model based on differences between the prediction results corresponding to each task and sample label results. The application can balance learning efficiency and accuracy among each task in multi-task learning.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a data processing method, apparatus, computer equipment, storage medium, and computer program product. Background Technology

[0002] In resource recommendation scenarios on resource display platforms, it is often necessary to predict multi-dimensional interactive information of target resources based on artificial intelligence models. However, some dimensions of interactive information suffer from sparse sample data. An effective approach is to use multi-task model training. During the training process of multi-task models, it is possible that some tasks may predict more accurately, while others may not, or for the same task, some display scenarios may predict more accurately, while others may not. Summary of the Invention

[0003] Therefore, it is necessary to provide a data processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can balance the learning efficiency and accuracy of various tasks in multi-task learning, in order to address the above-mentioned technical problems.

[0004] Firstly, this application provides a data processing method. The method includes:

[0005] Obtain a first sample set and a second sample set; the first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain.

[0006] Based on the first network of the initial multi-task model, feature extraction is performed on the first feature data of the sample to obtain the first feature;

[0007] Based on the second network of the initial multi-task model, the second feature data of the sample is feature-encoded to obtain the second initial feature. The first feature and the second initial feature are then fused to obtain the second feature.

[0008] Based on the first feature and the second feature, the prediction results corresponding to each task of the initial multi-task model are determined respectively;

[0009] Based on the differences between the prediction results and the sample labeling results for each task, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model.

[0010] Secondly, this application also provides a data processing method, the method comprising:

[0011] Obtain the first feature data of the target resource corresponding to the original feature domain of the target, and the second feature data of the target resource corresponding to the updated feature domain of the target; the updated feature domain of the target includes the original feature domain of the target and the target feature domain.

[0012] The first feature data is input into the first network in the target multi-task model, and feature extraction is performed on the first feature data based on the first network to obtain the first feature;

[0013] The second feature data is input into the second network in the target multi-task model. Based on the second network, the second feature data is feature encoded to obtain the second initial feature. The first feature and the second feature are then fused to obtain the second feature.

[0014] Based on the first feature and the second feature, the prediction results corresponding to each task of the target multi-task model are determined respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

[0015] Thirdly, this application also provides a data processing apparatus. The apparatus includes:

[0016] The first acquisition module is used to acquire a first sample set and a second sample set; the first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain.

[0017] The feature extraction module is used to extract features from the first feature data of the sample based on the first network of the initial multi-task model to obtain the first feature;

[0018] The feature fusion module is used to encode the second feature data of the sample based on the second network of the initial multi-task model to obtain the second initial feature, and to fuse the first feature and the second initial feature to obtain the second feature;

[0019] The determining module is used to determine the prediction results corresponding to each task of the initial multi-task model based on the first feature and the second feature;

[0020] The processing module is used to iteratively adjust the network parameters of the initial multi-task model based on the difference between the prediction results and the sample labeling results for each task, so as to obtain the corresponding target multi-task model.

[0021] In one embodiment, the first acquisition module is further configured to:

[0022] Obtain characteristic sample feature data corresponding to multiple original resource display scenarios, as well as common sample feature data among different original resource display scenarios;

[0023] Based on the sample features of the aforementioned characteristic sample feature data, a target resource display scenario is determined from the multiple original resource display scenarios;

[0024] Based on the multiple original resource display scenarios, the characteristic sample feature data, and the common sample feature data, the first sample set is determined;

[0025] Acquire new characteristic sample feature data of the target resource display scenario and new common sample feature data between the target display scenario and other original resource display scenarios;

[0026] Based on the target resource display scenario, the first feature data of the sample, the newly added characteristic sample feature data, and the newly added common sample feature data, the second sample set is determined.

[0027] In one embodiment, the first acquisition module is further configured to:

[0028] Obtain the number of sample features and the number of samples in the characteristic sample feature data;

[0029] Based on the ratio of the number of sample features to the corresponding number of samples, the feature proportions corresponding to each original resource display scenario are determined.

[0030] The target resource display scenario is determined based on the feature proportions corresponding to each of the original resource display scenarios.

[0031] In one embodiment, the target resource display scenario includes a subroutine scenario of an application, and the first acquisition module is further configured to:

[0032] Obtain target swipe operation data for a subroutine of the application; wherein the interval between two adjacent swipe operations in the target swipe operation data is less than a preset threshold.

[0033] Based on the target sliding operation data, new personalized feature data for the target resource display scene are determined;

[0034] Obtain interaction data for the target resource landing page of the subroutine of the application;

[0035] Based on the interaction data of the target resource landing page, new common sample feature data are determined between the updated display scenario and other original resource display scenarios.

[0036] In one embodiment, the first acquisition module is further configured to:

[0037] Obtain the importance index of each sample feature corresponding to the original feature domain to the prediction result, and update the importance index of each sample feature corresponding to the feature domain to the prediction result.

[0038] Based on the importance index of each sample feature corresponding to the original feature domain, the sample features corresponding to the original feature domain are sorted to obtain the first feature data;

[0039] Based on the importance index of each sample feature corresponding to the updated feature domain, the sample features corresponding to the updated feature domain are sorted to obtain the second feature data;

[0040] Based on the first feature data, a first sample set is obtained, and based on the second feature data, a second sample set is obtained.

[0041] In one embodiment, the first acquisition module is further configured to:

[0042] Based on the features of each sample corresponding to the original feature domain, a first multi-task model is trained, the accuracy index of the first multi-task model is determined, and the accuracy index of the first multi-task model is used as the benchmark accuracy index.

[0043] Target sample features are obtained from each of the sample features. Based on the other sample features besides the target sample features, a second task model is trained, and the accuracy index of the second task model is determined.

[0044] Based on the difference between the accuracy index of the second task model and the benchmark accuracy index, the importance index of the target sample features is determined.

[0045] In one embodiment, the importance index includes a responsiveness index, and the first acquisition module is further configured to:

[0046] Obtain the number of positive samples and the total number of samples for each sample feature in the original feature domain;

[0047] The grouping response rate of each sample feature is obtained by calculating the ratio between the number of positive samples and the total number of samples.

[0048] The baseline response rate is obtained by proportionally calculating the number of positive samples in each sample feature to the total number of samples corresponding to each sample feature.

[0049] Based on the group response rate and the baseline response rate, the response index corresponding to each sample feature is determined.

[0050] In one embodiment, the first acquisition module is further configured to:

[0051] Obtain the resource interaction amount corresponding to each task when it occurs, and sort the samples corresponding to each task based on the resource interaction amount to obtain the first sample sequence;

[0052] Obtain the ratio of the initial sample quantity corresponding to each task to the total initial sample quantity corresponding to each task. Based on the ratio, sort the samples corresponding to each task to obtain the second sample sequence.

[0053] If the first sorting position of the target sample corresponding to the target task in the first sample sequence does not match the second sorting position in the second sample sequence, the target sample is updated so that the second sorting position of the updated target sample is consistent with the first sorting position.

[0054] Based on the updated target sample's first feature data and the original feature domain where the target sample is located, a first sample set is determined.

[0055] In one embodiment, the feature extraction module is further configured to:

[0056] Based on the first encoding network, the first feature data of the sample is encoded to obtain the first initial feature;

[0057] Based on the first feature fusion network and the interaction weights between the original feature domains to which each of the first initial features belongs, the first initial features are fused to obtain the first feature.

[0058] In one embodiment, the feature fusion module is further configured to:

[0059] By concatenating the first feature and the second initial feature, a second intermediate feature is obtained;

[0060] Based on the cross weights between each of the second intermediate features, the second intermediate features are fused to obtain the second feature.

[0061] In one embodiment, the feature fusion module is further configured to:

[0062] Based on the cross weights between each of the second intermediate features, the features in each of the second intermediate features are fused to obtain the first fused feature;

[0063] Based on nonlinear weights, each of the second intermediate features is fused to obtain the second fused feature;

[0064] Based on the first fusion feature and the second fusion feature, a second feature is obtained.

[0065] In one embodiment, the feature fusion module is further configured to:

[0066] Based on the cross weights in each of the second intermediate features, the second intermediate features are fused to obtain the first fused feature;

[0067] Based on nonlinear weights, each of the first fusion features is fused to obtain the third fusion feature;

[0068] Based on the first fusion feature and the third fusion feature, a second feature is obtained.

[0069] In one embodiment, the determining module is further configured to:

[0070] Based on the gating network corresponding to the current task, feature filtering is performed on the first feature and the first feature to obtain the filtered features;

[0071] Based on the filtered features, predict the prediction result corresponding to the current task;

[0072] The filtered features corresponding to other tasks besides the current task are obtained in each task, and the prediction results corresponding to the other tasks are predicted based on the filtered features corresponding to the other tasks.

[0073] In one embodiment, the processing module is further configured to:

[0074] Based on the prediction results and annotation results corresponding to each task, the first loss corresponding to each task is determined;

[0075] Obtain the noise weights corresponding to each task, and perform fusion processing on the first loss corresponding to each task based on the noise weights to obtain the difference between the prediction results and the labeled results corresponding to each task.

[0076] Fourthly, this application also provides a data processing apparatus. The apparatus includes:

[0077] The second acquisition module is used to acquire the first feature data corresponding to the original feature domain and the second feature data corresponding to the updated feature domain; the updated feature domain includes the original feature domain and the target feature domain.

[0078] The feature extraction module is used to input the first feature data into the first network in the target multi-task model, and extract features from the first feature data based on the first network to obtain the first feature;

[0079] The feature fusion module is used to input the second feature data into the second network in the target multi-task model, perform feature encoding on the second feature data based on the second network to obtain the second initial feature, and perform feature fusion on the first feature and the second feature to obtain the second feature;

[0080] The determination module is used to determine the prediction results corresponding to each task of the target multi-task model based on the first feature and the second feature, respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

[0081] In one embodiment, the determining module is further configured to: input the first feature and the second feature into the tower unit corresponding to each task respectively, and output the corresponding prediction result through each tower unit.

[0082] Fifthly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement embodiments of the methods described in any embodiment of this disclosure.

[0083] Sixthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements an embodiment of the method described in any embodiment of this disclosure.

[0084] In a seventh aspect, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements an embodiment of the method described in any embodiment of this disclosure.

[0085] The aforementioned data processing methods, apparatus, computer equipment, storage media, and computer program products determine the target feature domain based on sample features of the original feature domain. For example, the original feature domain may contain a large number of samples but a small number of sample features, or the existing features in the original feature domain may have a minimal impact on the learning efficiency or accuracy of the task. An updated feature domain is determined based on the original and target feature domains, and a second feature data sample for the updated feature domain is obtained. Compared to the first feature data sample, the second feature data sample has richer sample features in the target feature domain. This is beneficial for balancing the learning efficiency and accuracy among various tasks in multi-task learning. Furthermore, the first feature of the first feature data sample is extracted through a first network, and the second feature data is encoded through samples from a second network. The encoded first features of the second initial feature domain are then fused to obtain the second feature. This facilitates feature interaction between the original and updated feature domains, improves the expressive power of the features, and enables the target multi-task model to have higher generalization ability. Attached Figure Description

[0086] Figure 1 This is a diagram illustrating the application environment of a data processing method in one embodiment.

[0087] Figure 2 This is a flowchart illustrating a data processing method in one embodiment;

[0088] Figure 3 This is a schematic diagram of the network architecture of a target multi-task model in one embodiment;

[0089] Figure 4 This is a flowchart illustrating the data processing method in another embodiment;

[0090] Figure 5 This is a flowchart illustrating the data processing method in another embodiment;

[0091] Figure 6 This is a flowchart illustrating the data processing method in another embodiment;

[0092] Figure 7 This is a schematic diagram of the network architecture of the target multi-task model in another embodiment;

[0093] Figure 8 This is a schematic diagram of the network architecture of the target multi-task model in another embodiment;

[0094] Figure 9 This is a schematic diagram of the network architecture of the target multi-task model in another embodiment;

[0095] Figure 10 This is a flowchart illustrating the data processing method in another embodiment;

[0096] Figure 11 This is a schematic diagram illustrating the prediction performance of the same task in different feature domains in another embodiment;

[0097] Figure 12 This is a schematic block diagram of the advertising retrieval platform structure in one embodiment;

[0098] Figure 13 This is a structural block diagram of a data processing device in one embodiment;

[0099] Figure 14 This is a structural block diagram of the data processing apparatus in another embodiment;

[0100] Figure 15 This is an internal structural diagram of a computer device in one embodiment;

[0101] Figure 16 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0102] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0103] The data processing method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store data that server 104 needs to process, such as a first sample set and a second sample set, which can originate from data obtained from terminal 102. The first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain, which is determined from sample features based on the original feature domain. The data storage system can be integrated on server 104 or placed in the cloud or on another server. Terminal 102 or server 104 can extract features from the sample first feature data based on the first network of the initial multi-task model to obtain a first feature vector. Based on the second network of the initial multi-task model, feature encoding is performed on the sample second feature data to obtain a second initial feature vector. The first feature vector and the second initial feature vector are then fused to obtain a second feature vector. Based on the first feature vector and the second feature vector, the prediction results corresponding to each task of the initial multi-task model are determined. Based on the differences between the prediction results and sample labeling results for each task, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model. The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0104] In one embodiment, such as Figure 2 As shown, a data processing method is provided, which can be applied to... Figure 1 Taking terminal 102 or server 104 as an example, the following steps are included:

[0105] Step S201: Obtain the first sample set and the second sample set.

[0106] The first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and the target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain.

[0107] The feature domain can be a classification or grouping of feature data. Feature data can be classified according to its business meaning. In an exemplary embodiment, in a resource recommendation scenario, the feature domain can include an account domain, a target resource domain, and a context domain. The account domain can include account identifier features, age features, and gender features; the target resource domain can include target resource identifier features, category features, and price features; and the context domain can include time features, location features, and device features. In an exemplary embodiment, for a target resource, such as an advertising resource, the feature domain can be matched with a resource display scenario that describes the display location of the advertising resource. Examples include public service account scenarios, application subroutine scenarios, text resource display scenarios, and personal account resource sharing scenarios. Optionally, each scenario can be further subdivided into more specific sub-scenarios. For example, a public service account scenario may have multiple sub-scenarios such as in-text, bottom, and subscription accounts, and each sub-scenarios can also correspond to a feature domain.

[0108] In this embodiment of the disclosure, the original feature domain may include the feature domain obtained by classifying the feature data according to the business meaning of the feature data in the above embodiments. The target feature domain may include the sample features determined from the original feature domain based on the original feature domain. Optionally, by obtaining the gradient of each task in the initial multi-task model, where the gradient is used to describe the learning speed, the target task with a learning speed lower than a preset threshold is determined, and the discriminative power of the target sample features on the target task is analyzed. Based on this discriminative power, the target feature domain in which the target sample features are located is determined. For example, if sample features a, b, and c have low discriminative power on the task and all belong to the original feature domain W, then the original feature domain W is taken as the target feature domain. Optionally, by using the number of sample features and the number of samples in each of the multiple original feature domains, the feature ratio of the number of sample features in each original feature domain to the total number of samples is determined, and the original feature domain with the smallest feature ratio is selected as the target feature domain. For example, the sample size corresponding to original feature domain 1 is M1, and the number of sample features is m1; the sample size corresponding to original feature domain 2 is M2, and the number of sample features is m2; the sample size corresponding to original feature domain 3 is M3, and the number of sample features is m3. For example, if m2 / M2 is less than m1 / M2 and m2 / M2 is less than m3 / M3, then the original feature domain 2 is used as the target feature domain. In another possible implementation, the target feature domain may include additional feature domains besides the original feature domain.

[0109] In the specific implementation method, refer to Figure 3As shown, taking the original feature domain 301 as a public service account scenario and an application subroutine scenario as examples, the first feature data of the sample corresponding to the public server scenario may include: the characteristic display feature f1 of advertising resources in this scenario, and the feature f2 of access account a. The first feature data of the sample corresponding to the application subroutine scenario may include: the characteristic display feature f3 of advertising resources in this scenario, and the feature f2 of access account a. The target feature domain is the application subroutine 303, and the second feature data of the sample corresponding to the application subroutine 303 may include newly added sample feature data as well as the first feature data f3 and f2 of the sample corresponding to the above application subroutine scenario. The newly added sample feature data is, for example, the duration feature and number of times of the fast swipe operation f4. Then the second feature data corresponding to the updated feature domain may include: f1, f2, f3, and f4. That is, the second feature data corresponding to the updated feature domain includes the first feature data of the sample corresponding to the original feature domain and the newly added feature data corresponding to the target feature domain.

[0110] Step S203: Based on the first network of the initial multi-task model, feature extraction is performed on the first feature data of the sample to obtain the first feature.

[0111] Among them, reference Figure 3 As shown, in one exemplary embodiment, the first network may include a feature extraction network that matches the first feature data, such as a Convolutional Neural Network (CNN), a Transformer network, a CLIP multimodal network, and a Recurrent Neural Network (RNN). In another exemplary embodiment, the first network may include a first encoding network and a first feature fusion network. The first encoding network can map discrete feature data to a low-dimensional continuous vector space, such as a Word2Vec network, a BERT network, or a One-Hot network. The first feature fusion network may include various factorization machine models, such as FM networks, FFM networks, and FwFMs networks.

[0112] Step S205: Based on the second network of the initial multi-task model, feature encoding is performed on the second feature data of the sample to obtain the second initial feature. The first feature and the second initial feature are then fused to obtain the second feature.

[0113] In one exemplary embodiment, reference is made to Figure 3As shown, the second feature data of the sample can be encoded using a second encoding network to obtain a second initial feature 305. The second encoding network can have the same structure as the first encoding network, such as a Word2Vec network, BERT network, or One-Hot network. In an exemplary embodiment, the first feature and the second initial feature can be concatenated to obtain the second feature. Optionally, the first feature and the second initial feature can be fused using a second feature fusion network to obtain the second feature. The second feature fusion network can include an attention mechanism network, an FPN network, a DCN network, and a DCN-v2 network, etc.

[0114] Step S207: Based on the first feature and the second feature, determine the prediction results corresponding to each task of the initial multi-task model.

[0115] The initial multi-task model can be a model that learns multiple tasks simultaneously. For example, in the field of computer vision, multi-task can include tasks such as simultaneously predicting the classification, detection, and segmentation of target objects in an image. Another example is in resource display scenarios, where tasks include calculating the click-through rate, conversion rate, and like rate of a target advertising resource.

[0116] In one exemplary embodiment, the first feature and the second feature can be used together to determine the prediction results corresponding to each task of the initial multi-task model. In another exemplary embodiment, the first feature X1 and the second feature X2 can be weighted and summed, for example: a×X1+b×X2, to determine the prediction results corresponding to each task of the initial multi-task model. In yet another exemplary embodiment, the first feature and the second feature can be selected using a gating network, and the prediction results corresponding to each task of the initial multi-task model can be determined based on the selected features.

[0117] Step S209: Based on the difference between the prediction results and sample labeling results for each task, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model.

[0118] In an exemplary embodiment, the difference between the prediction results for each task and the sample labeling results can be obtained by linearly weighting the losses of each task. For example:

[0119] (1)

[0120] in, Indicates the first The losses of each task (such as cross-entropy or mean squared error). Indicates the task weight.

[0121] In another exemplary embodiment, task noise can be used as the aforementioned task weights. Task noise characterizes the uncertainty of the task; the greater the uncertainty, the smaller the task loss. Optionally, the gradient norm of each task can be calculated, and the task weights can be adjusted to minimize the difference between the gradient norm of each task and the average value.

[0122] In the above data processing method, the target feature domain is determined based on the sample features of the original feature domain. For example, the original feature domain may contain a large number of samples but a small number of sample features, or the existing features in the original feature domain may have a small impact on the learning efficiency or accuracy of the task. The updated feature domain is then determined based on the original and target feature domains, and second feature data samples are obtained for the updated feature domain. Compared to the first feature data samples, the second feature data samples have richer sample features in the target feature domain. This is beneficial for balancing the learning efficiency and accuracy among various tasks in multi-task learning. Furthermore, the first feature of the first feature data is extracted through the first network, and the second feature data is encoded through the second network samples. The encoded first features of the second initial feature domain are then fused to obtain the second feature. This facilitates feature interaction between the original and updated feature domains, improves the expressive power of the features, and enables the target multi-task model to have higher generalization ability.

[0123] In one embodiment, reference Figure 4 As shown, obtaining the first sample set and the second sample set includes:

[0124] Step S401: Obtain the characteristic sample feature data corresponding to multiple original resource display scenarios and the common sample feature data corresponding to different original resource display scenarios.

[0125] The resources displayed in the resource display scenarios can include various resource formats, such as: video resource display scenarios in video accounts; text and image resource display scenarios in public service accounts; functional resource display scenarios in application subroutines; and text, image, and video resource display scenarios in personal account resource sharing, etc.

[0126] The characteristic sample feature data can be sample feature data unique to a specific resource display scenario, and not covered by other resource display scenarios. For example, a theme specific to a resource display scenario, such as a sports theme, might use the content of advertising resources, such as sports equipment, sports drinks, and sportswear, as sample feature data. Another example is the feature data of the first screen display of advertising resources in a personal account resource sharing scenario. Yet another example is the widget component of advertising resources on a video account in a video resource display scenario, which could be used as the aforementioned characteristic sample feature data.

[0127] The common sample feature data can be sample feature data that exists in multiple different resource display scenarios. The more common sample feature data there is, the greater the likelihood that training will yield good results for both resource display scenarios when the two resource display scenarios are trained together. For example, sample features that the same account has in a target number of resource display scenarios include: account A's attribute features, basic features, preference features, operational behavior features, and cross-features across different advertising resources.

[0128] Step S403: Based on the sample features of the characteristic sample feature data, determine the target resource display scenario from the multiple original resource display scenarios.

[0129] For example, by obtaining the gradient of each task in the initial multi-task model (the gradient describes the learning speed), a target task with a learning speed below a preset threshold is identified. The discriminative power of the target sample features on the target task is analyzed, and based on this discriminative power, the target resource display scenario in which the target sample features reside is determined. For example, if sample features a, b, and c have low discriminative power on the task and all belong to the original resource display scenario R, then the original resource display scenario R is taken as the target resource display scenario. Optionally, by using the number of sample features and the number of samples in each of the multiple original feature domains, the feature ratio of the number of sample features in each original feature domain to the total number of samples is determined. For example, a public service account has 10,000 samples and 200 sample features; an application subroutine has 20,000 samples and 100 sample features; and a video resource display scenario in a video account has 5,000 samples and 50 sample features. Since the application subroutine has insufficient sample features, it can be taken as the target resource display scenario.

[0130] Step S405: Based on the multiple original resource display scenarios, the characteristic sample feature data, and the common sample feature data, determine the first sample set.

[0131] Specifically, multiple original resource display scenarios are used as multiple original feature domains in the first sample set. The characteristic sample feature data corresponding to each of the multiple original resource display scenarios, as well as the common sample feature data corresponding to different original resource display scenarios, are used as the first feature data of the samples.

[0132] Step S407: Obtain the newly added characteristic sample feature data of the target resource display scene and the newly added common sample feature data between the target display scene and other original resource display scenes.

[0133] Specifically, for example, taking a subroutine of an application as the target resource display scenario, the duration and number of quick swipes in the subroutine are characteristic sample data of the application's subroutine, and this characteristic sample data does not exist in the first sample set, thus it is considered newly added characteristic sample data. Another example is the interaction data of the advertising resources corresponding to the application's subroutine, which is common sample feature data between the application's subroutine and other original resource display scenarios, and does not exist in the first sample set, thus it is considered newly added common sample feature data. Yet another example is a multi-image / text link advertising resource, including multiple images and multiple paragraphs of text, where users can view more content by swiping and jump to the corresponding website or subroutine by clicking the advertising resource. This multi-image / text link advertising resource has sample feature data present in both personal account resource sharing scenarios and target resource display scenarios, but does not exist in the first sample set, thus it is considered newly added common sample feature data.

[0134] Step S409: Based on the target resource display scenario, the first feature data of the sample, the newly added personalized sample feature data, and the newly added common sample feature data, determine the second sample set.

[0135] Specifically, the target resource display scenario is taken as the target feature domain, and the first feature data of the sample, the newly added characteristic sample feature data, and the newly added common sample feature data are taken as the second feature data of the sample.

[0136] In the above embodiments, the use of characteristic sample feature data avoids feature dimension explosion or negative learning effects. For example, characteristic sample feature data of one resource display scenario may not provide any benefit to another resource display scenario. By using common sample feature data from different resource display scenarios, it is beneficial to train different resource display scenarios together, and all different resource display scenarios have good training results.

[0137] In one embodiment, determining the target resource display scenario from the plurality of original resource display scenarios based on the sample features of the characteristic sample feature data includes:

[0138] Obtain the number of sample features and the number of samples in the characteristic sample feature data.

[0139] Based on the ratio of the number of sample features to the corresponding number of samples, the feature proportions corresponding to each original resource display scenario are determined.

[0140] Based on the feature proportions corresponding to each of the original resource display scenarios, the target resource display scenario for updating is determined.

[0141] In the specific implementation process, the original resource display scenario A includes: Sample 1 {sample feature a1, sample feature b1, sample feature c1}; Sample 2 {sample feature b1, sample feature b2, sample feature c1}; Sample 3 {sample feature c1, sample feature b2, sample feature c2}. The original resource display scenario B includes: Sample 4 {sample feature b1, sample feature b1, sample feature c1}; Sample 5 {sample feature b1, sample feature b2, sample feature b3}; Sample 6 {sample feature b1, sample feature b2, sample feature c2}. The original resource display scenario C includes: Sample 7 {sample feature c1, sample feature b1, sample feature c2}; Sample 8 {sample feature c1, sample feature c2, sample feature c1}; Sample 9 {sample feature b1, sample feature c2, sample feature c2}.

[0142] In the above embodiments, it is assumed that ai, i=1, 2, 3… belongs to the characteristic sample data of the original resource display scenario A. bi, i=1, 2, 3… belongs to the characteristic sample data of the original resource display scenario B. ci, i=1, 2, 3… belongs to the characteristic sample data of the original resource display scenario C. Then the feature proportion of the original resource display scenario A is represented as 1 / 3; the feature proportion of the original resource display scenario B is represented as 6 / 3; and the feature proportion of the original resource display scenario C is 7 / 3. It should be noted that the number of samples and the number of sample features corresponding to different original resource display scenarios can be the same or different, and this disclosure does not impose any restrictions on this.

[0143] In the above embodiments, the feature domain with a large number of samples but a small number of sample features in the original feature domain characteristic sample feature data is used as the target feature domain. The sample features of the target feature domain are then supplemented. This helps to reduce the prediction difference between different resource display scenarios under the same training task and improves the prediction accuracy and stability of the target multi-task model.

[0144] In one embodiment, the target updated resource display scenario includes a subroutine scenario of the application, and the step of obtaining newly added personalized sample feature data of the target resource display scenario and newly added common sample feature data between the display scenario and other original resource display scenarios includes:

[0145] Obtain target swipe operation data for a subroutine of the application; wherein the interval between two adjacent swipe operations in the target swipe operation data is less than a preset threshold.

[0146] Based on the target sliding operation data, new personalized feature data for the updated resource display scenario is determined.

[0147] Obtain interaction data for the target resource landing page performed by a subroutine of the application.

[0148] Based on the interaction data of the target resource landing page, new common sample feature data are determined between the updated display scenario and other original resource display scenarios.

[0149] The application's subroutines, also known as mini-programs, are referred to as subroutines. Swiping is a human-computer interaction operation, such as continuously touching different display areas of a page. In this embodiment, the target swipe operation may include a swipe where the interval between two adjacent swipe operations is less than a preset threshold, also known as a fast swipe operation. The target swipe operation data may include the fast swipe duration and the number of fast swipes. For example, the fast swipe duration may be n fast swipe operations performed within a preset range, where n is the number of fast swipes. The start time of these n swipe operations is t1, the end time is t2, and the fast swipe duration is t2-t1. Optionally, the fast swipe duration may also include the time required to perform one fast swipe operation. The aforementioned target swipe operation data belongs to the subroutine's newly added personalized feature data, and is not involved in other resource display scenarios.

[0150] In an exemplary embodiment, the target resource may be an advertising resource displayed in a subroutine. The landing page of the target resource in the subroutine may include the advertising resource displayed after the subroutine is launched, with clicking the corresponding link of the advertising resource leading to the landing page. Optionally, the interaction data of the landing page may include the number of clicks, conversions, duration, etc. The aforementioned interaction data of the target resource landing page is also included in other resource display scenarios; this embodiment adds it to the application scenario of mini-programs, and it belongs to newly added common sample feature data.

[0151] In the above embodiments, the target swiping operation data and the interaction data of the target resource landing page have a high degree of distinguishability for the prediction results of the subprogram. Therefore, the target swiping operation data is used as new personalized feature data of the subprogram scenario, and the interaction data of the target resource landing page is used as new common sample feature data of the subprogram scenario. This helps to reduce the difference in prediction results between the mini-program scenario and other resource display scenarios on the same task.

[0152] In one embodiment, reference Figure 5 As shown, obtaining the first sample set and the second sample set includes:

[0153] Step S501: Obtain the importance index of each sample feature corresponding to the original feature domain to the prediction result, and update the importance index of each sample feature corresponding to the feature domain to the prediction result.

[0154] For example, in the original feature domain R, account preference feature a corresponds to an importance of 0.8, account statistics feature b corresponds to an importance of 0.3, and account behavior feature c corresponds to an importance of 0.6. In the updated feature domain, account preference feature a corresponds to an importance of 0.7, account statistics feature b corresponds to an importance of 0.4, account behavior feature c corresponds to an importance of 0.5, and the number of clicks on the target resource associated with the account, d, corresponds to an importance of 0.9.

[0155] Step S503: Based on the importance index of each sample feature corresponding to the original feature domain, sort the sample features corresponding to the original feature domain to obtain the first feature data.

[0156] In the above embodiments, based on the importance index of each sample feature, the sample features corresponding to the original feature domain are sorted. Optionally, they can be sorted in descending order of importance. Therefore, the first feature data can be represented as: account preference feature a, account behavior feature c, and account statistical feature b.

[0157] Step S505: Based on the importance index of each sample feature corresponding to the updated feature domain, sort the sample features corresponding to the updated feature domain to obtain the second feature data.

[0158] Similarly, in the above embodiments, based on the importance index of each sample feature, the sample features corresponding to the updated feature domain are sorted. Optionally, they can be sorted in descending order of importance. Therefore, the second feature data can be represented as: the number of clicks on the target resource associated with the account (d), the account preference feature (a), the account behavior feature (c), and the account statistical feature (b).

[0159] Step S507: Based on the first feature data, obtain a first sample set; based on the second feature data, obtain a second sample set.

[0160] Specifically, the first sample set includes the first feature data of the samples corresponding to the original feature domain, and the second sample set includes the second feature data of the samples corresponding to the updated feature domain.

[0161] In the above embodiments, sorting the sample features in the first feature data and the second feature data according to their importance helps to discover hidden patterns in the sample features, reduce computational complexity and cost, and improve the performance of the target multi-task model.

[0162] In one embodiment, obtaining the importance index of each sample feature corresponding to the original feature domain to the prediction result includes:

[0163] Based on the features of each sample corresponding to the original feature domain, a first multi-task model is trained, and the accuracy index of the first multi-task model is determined. The accuracy index of the first multi-task model is used as the benchmark accuracy index.

[0164] Target sample features are obtained from each of the sample features. Based on the other sample features besides the target sample features, a second task model is trained, and the accuracy index of the second task model is determined.

[0165] Based on the difference between the accuracy index of the second task model and the benchmark accuracy index, the importance index of the target sample features is determined.

[0166] Among them, accuracy metrics can include the AUC metric, where AUC (Area Under the ROC Curve) is a core metric used for model performance in machine learning, reflecting the model's overall ability to distinguish between positive and negative samples.

[0167] In an exemplary embodiment, a first multi-task model is trained based on the sample features corresponding to the original feature domain, for example, by a supervised machine learning algorithm. The accuracy metric of the first multi-task model is determined, which can be denoted as AUC_base. Target sample features are obtained from the sample features, for example, each sample feature includes sample feature A, sample feature B, and sample feature C. Target sample feature A is obtained, and a second task model is trained based on sample feature B and sample feature C, and the accuracy metric of the second task model is determined as AUC_A. Similarly, another second task model is trained based on sample feature A and sample feature C, and the accuracy metric of the other second task model is determined as AUC_B. Similarly, yet another second task model is trained based on sample feature A and sample feature B, and the accuracy metric of the yet another second task model is determined as AUC_C.

[0168] The importance index of sample feature A is represented as AUC_base-AUC_A; the importance index of sample feature B is represented as AUC_base-AUC_B; and the importance index of sample feature C is represented as AUC_base-AUC_C.

[0169] In the above embodiments, a second task model is trained using sample features other than the target sample features among the sample features, and an accuracy index of the second task model is determined. Based on the difference between this accuracy index and the benchmark accuracy index, the importance of the target sample features can be accurately characterized.

[0170] In one embodiment, the importance index includes a responsiveness index, and the step of obtaining the importance index of each sample feature corresponding to the original feature domain to the prediction result includes:

[0171] Obtain the number of positive samples and the total number of samples for each sample feature in the original feature domain.

[0172] The group response rate of each sample feature is obtained by performing a ratio calculation based on the number of positive samples and the total number of samples.

[0173] The baseline response rate is obtained by proportionally calculating the number of positive samples in each sample feature to the total number of samples corresponding to each sample feature.

[0174] Based on the group response rate and the baseline response rate, the response index corresponding to each sample feature is determined.

[0175] Positive samples can include samples labeled as having performed a specific task, such as an account sample clicking on a target advertising resource, or the target advertising resource sample achieving a conversion. In an exemplary embodiment, the number of positive samples in sample feature D is denoted as d1, and the total number of samples is denoted as d2. The grouped response rate of sample feature D can then be expressed as d1 / d2. Similarly, the number of positive samples in sample feature E is denoted as e1, and the total number of samples is denoted as e2. The grouped response rate of sample feature E can then be expressed as e1 / e2. Likewise, the number of positive samples in sample feature F is denoted as f1, and the total number of samples is denoted as f2. The grouped response rate of sample feature F can then be expressed as f1 / f2.

[0176] In an exemplary embodiment, the number of positive samples for each sample feature can be represented as G1, the total number of samples corresponding to each sample feature can be represented as G2, and the baseline response rate can be represented as G1 / G2. The response index of a sample feature is obtained by proportionally calculating the grouped response rate of the sample feature to the baseline response rate. Taking sample feature D as an example, the response index of sample feature D can be represented as (d1 / d2) / (G1 / G2). Similarly, the response indices of sample features E and F can be obtained.

[0177] In the above embodiments, by using the grouped response rate of each sample feature and the baseline response rate, the response index corresponding to each sample feature is determined, which can measure the improvement factor of the model's predictive ability for the target task of that sample feature compared to random selection.

[0178] In one embodiment, reference Figure 6 As shown, each task in the initial multi-task model includes interactive operations targeting a target resource; obtaining the first sample set includes:

[0179] Step S601: Obtain the resource interaction amount corresponding to each task when it occurs, and sort the samples corresponding to each task based on the resource interaction amount to obtain the first sample sequence.

[0180] Specifically, taking click-through rate (CTR), conversion rate (CTR), and like rate as examples, the resource interaction volume corresponding to each task can be the resource consumption corresponding to CTR, the resource consumption corresponding to conversion rate, and the resource consumption corresponding to like rate. Optionally, it can be the cost-per-click (CPC) corresponding to CTR, the cost-per-conversion (CPC) corresponding to conversion rate, and the cost-per-like (CPC) corresponding to like rate.

[0181] For example, the first sample sequence is obtained by ordering the resource interaction amounts corresponding to tasks a, b, and c from largest to smallest, and is represented as: task b, task c, and task a.

[0182] Step S603: Obtain the ratio of the initial sample quantity corresponding to each task to the total initial sample quantity corresponding to each task. Based on the ratio, sort the samples corresponding to each task to obtain the second sample sequence.

[0183] For example, let the initial sample size for task a be denoted as a1, and the total initial sample size for all tasks be denoted as T. Then, the proportion of task a is represented as a1 / T. Similarly, let the initial sample size for task b be denoted as b1, and the total initial sample size for all tasks be denoted as T. Then, the proportion of task b is represented as b1 / T. Similarly, let the initial sample size for task c be denoted as c1, and the total initial sample size for all tasks be denoted as T. Then, the proportion of task c is represented as c1 / T. Assuming a1 / T is greater than b1 / T, which is greater than c1 / T, the second sample sequence can be represented as: task a, task b, and task c.

[0184] Step S605: If the first sorting position of the target sample corresponding to the target task in the first sample sequence does not match the second sorting position in the second sample sequence, update the target sample so that the second sorting position of the updated target sample is consistent with the first sorting position.

[0185] In the above embodiment, task b ranks first in the first sample sequence and has a large amount of resource interaction, while it ranks second in the second sample sequence. Optionally, the number of samples corresponding to task b can be increased to b2. Optionally, it can be increased according to the ratio of positive and negative samples in the initial number of samples b1, that is, the ratio of positive and negative samples in the number of samples b2 is the same as that in the initial number of samples b1. In this way, the second sample sequence is updated to task b, task a, and task c. Among them, task c ranks second in the first sample sequence. Unlike in the second sample sequence, the number of samples corresponding to task a can be increased to obtain the second sample sequence of task b, task c, and task a.

[0186] Step S607: Determine the first sample set based on the updated first feature data of the target sample and the original feature domain in which the target sample is located.

[0187] In this embodiment of the disclosure, the updated task b corresponds to b2 target samples, the updated task c corresponds to c2 target samples, and task a is not updated, corresponding to a1 samples. Based on the updated target samples and their original feature domains, a first sample set is determined.

[0188] In the above embodiments, by resampling the samples corresponding to each task according to the resource interaction amount corresponding to each task when it occurs, the samples become balanced, which is beneficial to the overall consistency of the prediction effect of each task in the target multi-task model.

[0189] In one embodiment, the first network includes a first encoding network and a first feature fusion network, wherein the feature extraction of the sample first feature data based on the first network to obtain the first feature includes:

[0190] Based on the first encoding network, the first feature data of the sample is encoded to obtain the first initial feature.

[0191] Based on the first feature fusion network and the interaction weights between the original feature domains to which each of the first initial features belongs, the first initial features are fused to obtain the first feature.

[0192] Specifically, refer to Figure 7 As shown, the first network may include a first encoding network and a first feature fusion network 703, wherein the first encoding network encodes the first feature data of the sample to obtain a first initial feature vector 701. Optionally, the first encoding network may include the Word2Vec network, BERT network, One-Hot network, etc. described in the above embodiments.

[0193] In this embodiment of the disclosure, the first feature fusion network 703 can be a FwFMs model (Field-weighted Factorization Machines). For example, the first feature... It can be represented as follows:

[0194] (2)

[0195] in, Represents the first initial feature The original feature domain, Represents the first initial feature The original feature domain, Represents the original feature domain and the original feature domain The interaction weights between them are learnable parameters, and their initial values ​​can be randomly generated. and Representing the first initial features and the first initial feature The implicit vector. and This represents the learning weight.

[0196] In the above embodiments, the first initial features are fused using a first feature fusion network and the interaction weights between the original feature domains to which each first initial feature belongs, to obtain the first feature. Explicitly modeling the interaction weights between the original feature domains to which each first initial feature belongs reduces parameter complexity, decreases dependence on sparse feature interactions, and improves generalization ability.

[0197] In one embodiment, the feature fusion of the first feature and the second initial feature to obtain the second feature includes:

[0198] By concatenating the first feature and the second initial feature, a second intermediate feature is obtained;

[0199] Based on the cross weights between each of the second intermediate features, the second intermediate feature vectors are fused to obtain the second feature.

[0200] In the specific implementation process, the second feature fusion network may include a cross-network, which includes multiple cross layers. Based on the cross-network, the first feature and the second initial feature are concatenated to obtain the second intermediate feature. Further, the cross-network inputs the second intermediate feature... With the previous layer fusion results Multiply by the transpose, then multiply by the weight matrix. Specifically, the current layer fusion results It is expressed as follows:

[0201] (3)

[0202] in, This represents a biased vector.

[0203] Finally, based on the fusion result of the last layer, the second feature is obtained.

[0204] In the above embodiments, the first feature and the second initial feature are automatically cross-referenced through a cross-network, avoiding more manual feature combinations based on business understanding. Feature cross-references can be explicitly applied at each layer, efficiently learning cross-referenced features with low complexity.

[0205] In one embodiment, reference Figure 8 As shown, the process of fusing the second intermediate features based on the cross weights between them to obtain the second feature includes:

[0206] Based on the cross weights between each of the second intermediate features, the features in each of the second intermediate features are fused to obtain the first fused feature;

[0207] Based on nonlinear weights, each of the second intermediate features is fused to obtain the second fused feature;

[0208] Based on the first fusion feature and the second fusion feature, a second feature is obtained.

[0209] In this embodiment of the disclosure, reference is made to Figure 8 The cross network shown, based on the cross weights between each of the second intermediate features, performs fusion processing on the features in each of the second intermediate features to obtain the first fused feature. The specific process can be referred to Equation (3) in the above embodiment, and will not be repeated here.

[0210] In one exemplary embodiment, reference is made to Figure 8 The deep network shown has the output of the previous layer represented as: The fusion features of the current layer It can be represented as follows:

[0211] (4)

[0212] in, Represents the weight vector (learnable parameters). Indicates the bias term. This represents a non-linear weight. For example, non-linear weights can be achieved using functions such as ReLU, Sigmoid, and Softmax.

[0213] Next, based on the output of the last layer, the second fused feature is obtained. Finally, based on the first fused feature and the second fused feature, the second feature is obtained, for example, by concatenating the first fused feature and the second fused feature, or by performing a weighted summation of the first fused feature and the second fused feature.

[0214] In the above embodiments, based on the cross-fusion of the second intermediate features, the nonlinear fusion of the second intermediate features is further added in parallel, which strengthens the expressive power of the first fused features and improves the prediction accuracy of the target multi-task model.

[0215] In one embodiment, the step of fusing the second intermediate features based on the cross weights between the second intermediate features to obtain the second feature includes:

[0216] Based on the cross weights in each of the second intermediate features, the second intermediate features are fused to obtain the first fused feature.

[0217] Based on nonlinear weights, the first fusion features are fused to obtain the third fusion feature.

[0218] Based on the first fusion feature and the third fusion feature, a second feature is obtained.

[0219] Specifically, refer to Figure 9 The cross-network shown in this embodiment of the disclosure, based on the cross-network and the cross weights in each of the second intermediate features, performs a fusion process on each of the second intermediate features to obtain a first fused feature. The specific process can be referred to equation (3) in the above embodiment, and will not be repeated here.

[0220] In one exemplary embodiment, reference is made to Figure 9 The deep network shown in this embodiment of the disclosure, based on the deep network and nonlinear weights, performs fusion processing on each of the first fusion features to obtain a third fusion feature vector. The specific process can be referred to equation (4) in the above embodiment, and will not be repeated here.

[0221] Finally, based on the first fusion feature and the third fusion feature, a second feature is obtained. For example, the first fusion feature and the third fusion feature can be concatenated to obtain the second feature. Alternatively, the first fusion feature and the third fusion feature can be weighted and fused to obtain the second feature.

[0222] In the above embodiments, based on the cross-fusion of the second intermediate features, a nonlinear fusion of the second intermediate features is further added in a series form, which strengthens the expressive power of the first fused features and improves the prediction accuracy of the target multi-task model.

[0223] In one embodiment, determining the prediction results for each task of the initial multi-task model based on the first feature and the second feature includes:

[0224] Based on the gating network corresponding to the current task, feature filtering is performed on the first feature and the second feature to obtain the filtered features;

[0225] Based on the filtered features, predict the prediction result corresponding to the current task;

[0226] The filtered features corresponding to other tasks besides the current task are obtained in each task, and the prediction results corresponding to the other tasks are predicted based on the filtered features corresponding to the other tasks.

[0227] Specifically, a gating network can be a neural network structure that uses gating mechanisms to control the flow of information, thereby enhancing the model's ability to capture hierarchical features. (Reference) Figure 3 As shown, the gating network 307 can perform feature filtering on the first and second features. For example, the gating network determines the passing ratio of features by using learnable parameters, such as values ​​between 0 and 1 from the output of the sigmoid function: 0: completely block information; 1: completely retain information; 0~1: filter information proportionally. The opening and closing of the gate is automatically adjusted according to the current input and context to adapt to different tasks.

[0228] After passing through a gating network, different tasks may correspond to different first and second features. For example, task 1 corresponds to the first feature, task 2 corresponds to the second feature, and task 3 corresponds to (a×first feature+b×second feature).

[0229] The above embodiments adaptively filter the first and second features through a gating network to predict different tasks, thereby improving the feature representation of each task in the multi-task prediction model and improving the accuracy of the multi-task prediction model.

[0230] In one embodiment, the step of iteratively adjusting the network parameters of the initial multi-task model based on the difference between the prediction results and sample labeling results for each task to obtain the corresponding target multi-task model includes:

[0231] Based on the prediction and annotation results for each task, the first loss for each task is determined.

[0232] Obtain the noise weights corresponding to each task, and perform fusion processing on the first loss corresponding to each task based on the noise weights to obtain the difference between the prediction results and the labeled results corresponding to each task.

[0233] Specifically, the task The corresponding first loss can be expressed as The noise weights are used to characterize the uncertainty of the model. Uncertainty may include noise in the data or the model's lack of knowledge about certain data. In an exemplary embodiment, by incorporating the noise weights into the loss function, the difference between the prediction results and the labeled results for each task can be represented as follows:

[0234] (5)

[0235] in, Indicates task Uncertainty These are learnable parameters.

[0236] In the above embodiments, noise weights are added to the loss of each task. For tasks with high noise weights, the loss weights are reduced to avoid dominating the training process. Through the design of the loss function, the model learns the prediction results and their uncertainties during the training process, thereby improving the robustness and reliability of different tasks.

[0237] Based on the same inventive concept, this application proposes a data processing method, the method comprising:

[0238] Obtain the first feature data corresponding to the original feature domain and the second feature data corresponding to the updated feature domain; the updated feature domain includes the original feature domain and the target feature domain.

[0239] The first feature data is input into the first network in the target multi-task model, and feature extraction is performed on the first feature data based on the first network to obtain the first feature;

[0240] The second feature data is input into the second network in the target multi-task model. Based on the second network, the second feature data is feature encoded to obtain the second initial feature. The first feature and the second feature are then fused to obtain the second feature.

[0241] Based on the first feature and the second feature, the prediction results corresponding to each task of the target multi-task model are determined respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

[0242] In one embodiment, reference Figure 10 As shown, this application provides a data processing method, the method comprising:

[0243] Step S1001: Obtain the first feature data of the target resource corresponding to the original feature domain of the target, and the second feature data of the target resource corresponding to the updated feature domain of the target; the updated feature domain of the target includes the original feature domain of the target and the target feature domain.

[0244] Specifically, in an exemplary embodiment, the feature domain is for a target resource, such as an advertising resource. The feature domain can be matched with a resource display scenario that describes the display location of the advertising resource. Examples include public service account scenarios, application subroutine scenarios, text resource display scenarios, and personal account resource sharing scenarios. Optionally, each scenario can be further subdivided into more specific sub-scenarios. For example, a public service account scenario may have multiple sub-scenarios such as in-text, bottom, and subscription accounts, and each sub-scenarios can also correspond to a feature domain. The original feature domain can include the feature domain obtained by classifying the feature data according to the business meaning of the feature data in the above embodiments. The target feature domain can include the feature domain determined from sample features based on the original feature domain.

[0245] In an exemplary embodiment, taking the public service account scenario and the application's subroutine scenario as examples, the first feature data corresponding to the public server scenario may include: the characteristic display feature f1 of the target advertising resource within the scenario, and the feature f2 of the access account a. The first feature data corresponding to the application's subroutine scenario may include: the characteristic display feature f3 of the target advertising resource within the scenario, and the feature f2 of the access account a. The target feature domain is the application's subroutine 303, and the second feature data corresponding to the application's subroutine 303 may include newly added feature data as well as the first feature data f3 and f2 corresponding to the application's subroutine scenario. For example, the newly added feature data may include: the duration and number of swipe operations f4. Then, the second feature data corresponding to the updated feature domain may include: f1, f2, f3, and f4. That is, the second feature data corresponding to the updated feature domain includes the first feature data corresponding to the target original feature domain and the newly added feature data corresponding to the target feature domain.

[0246] Step S1003: Input the first feature data into the first network in the target multi-task model, and extract features from the first feature data based on the first network to obtain the first feature.

[0247] Specifically, you can refer to Figure 3As shown, in one exemplary embodiment, the first network may include a feature extraction network that matches the first feature data, such as a Convolutional Neural Network (CNN), a Transformer network, a CLIP multimodal network, and a Recurrent Neural Network (RNN). In another exemplary embodiment, the first network may include a first encoding network and a first feature fusion network. The first encoding network can map discrete feature data to a low-dimensional continuous vector space, such as a Word2Vec network, a BERT network, or a One-Hot network. The first feature fusion network may include various factorization machine models, such as FM networks, FFM networks, and FwFMs networks.

[0248] Step S1005: Input the second feature data into the second network in the target multi-task model, perform feature encoding on the second feature data based on the second network to obtain the second initial feature, and perform feature fusion on the first feature and the second feature to obtain the second feature.

[0249] Specifically, you can refer to the following: Figure 3 As shown, the second feature data of the sample can be encoded using a second encoding network to obtain a second initial feature 305. The second encoding network can have the same structure as the first encoding network, such as a Word2Vec network, BERT network, or One-Hot network. In an exemplary embodiment, the first feature and the second initial feature can be concatenated to obtain the second feature. Optionally, the first feature and the second initial feature can be fused using a second feature fusion network to obtain the second feature. The second feature fusion network can include an attention mechanism network, an FPN network, a DCN network, and a DCN-v2 network, etc.

[0250] Step S1007: Based on the first feature and the second feature, determine the prediction results corresponding to each task of the target multi-task model respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

[0251] The initial multi-task model can be a model that learns multiple tasks simultaneously, such as click-through rate, conversion rate, and like rate for target advertising resources in a resource display scenario.

[0252] In one exemplary embodiment, the first feature and the second feature can be used together to determine the prediction results corresponding to each task of the initial multi-task model. In another exemplary embodiment, the first feature X1 and the second feature X2 can be weighted and summed, for example: a×X1+b×X2, to determine the prediction results corresponding to each task of the initial multi-task model. In yet another exemplary embodiment, the first feature and the second feature can be selected using a gating network, and the prediction results corresponding to each task of the initial multi-task model can be determined based on the selected features.

[0253] In the above data processing method, the target feature domain is determined based on the sample features of the original feature domain. For example, the original feature domain may contain a large number of samples but a small number of sample features, or the existing features in the original feature domain may have a small impact on the school efficiency or accuracy of the task. An updated feature domain is then determined based on the original and target feature domains, and a second feature data sample is obtained for the updated feature domain. Compared to the first feature data sample, the second feature data sample has richer sample features in the target feature domain, which can improve the accuracy and stability of the multi-task prediction results.

[0254] In one embodiment, the method further includes:

[0255] Based on the prediction results corresponding to each target resource, the target resources are sorted to obtain the sorting results;

[0256] Based on the ranking results, the recommended results for each target resource are determined.

[0257] Specifically, for example, the predicted results for each task of target resource a are click-through rate (CTR) a1, like rate (A2), and conversion rate (A3). The predicted results for each task of target resource b are CTR b1, like rate (B2), and conversion rate (B3). The predicted results for each task of target resource c are CTR c1, like rate (C2), and conversion rate (C3). The ranking of each target resource can be determined by comprehensively considering the above data for each resource. For example, the weighting of each task can be set; for instance, the ranking index for target resource a is w1×a1 + w2×a2 + w3×a3, and the ranking index for target resource b is w1×b1 + w2×b2 + w3×b3.

[0258] Optionally, prediction results for multiple tasks within a specific target original feature domain can be determined for each target resource. For example, the first feature data of target resource a carries identification information of the target original feature domain G, and the first feature data of target resource b carries identification information of the target original feature domain G. This allows for the determination of first prediction results for target resource a corresponding to each of the multiple tasks within the target original feature domain G, and second prediction results for target resource b corresponding to each of the multiple tasks within the target original feature domain G. Optionally, based on the first and second prediction results, a target resource to be recommended within the target original feature domain G is determined.

[0259] In a specific embodiment, the method of this application can be applied to multi-task scenarios such as predicting click-through rate, like rate, and conversion rate for advertising resources. In existing technologies, during the training of multi-task models, there may be instances where predictions for some tasks are more accurate than those for others, or where predictions for the same task are more accurate in some display scenarios than others. For example, referring to… Figure 11 As shown, during the training process of multiple tasks on the official account, it can be seen that official account 1101 and mini-program 1103, in certain tasks, such as mini-games, are activated in the mini-program and activated in the official account's traffic-driving application. These two belong to the same task but different scenarios. The real-time AUC training is shown in the figure below, showing a trend of one increasing while the other decreases. This application provides a data processing method that can improve the learning efficiency and accuracy between various tasks in multi-task learning.

[0260] Specifically, data processing methods may include: sample collection stage, sample processing stage, and model training stage.

[0261] In the sample collection phase, the initial collection includes: Step S401, acquiring characteristic sample feature data corresponding to multiple original resource display scenarios and common sample feature data corresponding to different original resource display scenarios. Step S403, determining the target resource display scenario from the multiple original resource display scenarios based on the sample features of the characteristic sample feature data. Step S405, determining the first sample set based on the multiple original resource display scenarios, the characteristic sample feature data, and the common sample feature data. Step S407, acquiring newly added characteristic sample feature data of the target resource display scenario and newly added common sample feature data between the target display scenario and other original resource display scenarios. Step S409, determining the second sample set based on the target resource display scenario, the first feature data of the samples, the newly added individual sample feature data, and the newly added common sample feature data.

[0262] The data collection and supplementation period includes: Step S601, obtaining the resource interaction volume corresponding to each task when it occurs, and sorting the samples corresponding to each task based on the resource interaction volume to obtain a first sample sequence. Step S603, obtaining the ratio of the initial sample quantity corresponding to each task to the total initial sample quantity corresponding to each task, and sorting the samples corresponding to each task based on the ratio to obtain a second sample sequence. Step S605, if the first sorting position of the target sample corresponding to the target task in the first sample sequence does not match the second sorting position in the second sample sequence, updating the target sample so that the second sorting position of the updated target sample is consistent with the first sorting position. Step S607, determining a first sample set based on the first feature data of the updated target sample and the original feature domain in which the target sample is located.

[0263] In the sample processing stage, step S501 involves obtaining the importance index of each sample feature corresponding to the original feature domain to the prediction result, and the importance index of each sample feature corresponding to the updated feature domain to the prediction result. Step S503 involves sorting the sample features corresponding to the original feature domain based on their importance indexes to obtain first feature data. Step S505 involves sorting the sample features corresponding to the updated feature domain based on their importance indexes to obtain second feature data. Step S507 involves obtaining a first sample set based on the first feature data and a second sample set based on the second feature data.

[0264] During the model training phase, the method includes: Step S203, extracting features from the sample's first feature data using the first network of the initial multi-task model to obtain a first feature. Step S205, encoding features from the sample's second feature data using the second network of the initial multi-task model to obtain a second initial feature, and fusing the first feature and the second initial feature to obtain a second feature. Step S207, determining the prediction results for each task of the initial multi-task model based on the first feature and the second feature. Step S203 may further include: encoding features from the sample's first feature data using the first encoding network to obtain a first initial feature; fusing each first initial feature using the first feature fusion network and the interaction weights between the original feature domains to which each first initial feature belongs to, to obtain a first feature. Step S205 may further include: concatenating the first feature and the second initial feature to obtain a second intermediate feature; fusing each second intermediate feature based on the cross weights between each second intermediate feature to obtain a second feature. Step S207 may further include: based on the gating network corresponding to the current task, performing feature filtering on the first feature and the first feature to obtain filtered features; predicting the prediction result corresponding to the current task based on the filtered features; obtaining the filtered features corresponding to other tasks besides the current task, and predicting the prediction results corresponding to the other tasks based on the filtered features corresponding to the other tasks. Step S209: based on the difference between the prediction results and the sample labeling results corresponding to each task, iteratively adjusting the network parameters of the initial multi-task model to obtain the corresponding target multi-task model. Step S209 may further include: determining the first loss corresponding to each task based on the prediction results and labeling results corresponding to each task; obtaining the noise weights corresponding to each task, and fusing the first losses corresponding to each task based on the noise weights to obtain the difference between the prediction results and the labeling results corresponding to each task.

[0265] In a specific implementation scenario, the embodiments of this disclosure can be applied to multi-task scenarios involving predicting click-through rates, like rates, and conversion rates for advertising resources. It can be applied to, for example... Figure 12The example shown is a fine-ranking module of an advertising retrieval platform. Specifically: A first sample set and a second sample set are obtained. The first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain. For example, the original feature domain may include: video resource display scenarios in video accounts; text and image resource display scenarios in public service accounts; functional resource display scenarios in application subroutines; and text, image, and video resource display scenarios in personal account resource sharing, etc. Based on the first network of the initial multi-task model, features are extracted from the sample first feature data to obtain a first feature. Based on the second network of the initial multi-task model, features are encoded from the sample second feature data to obtain a second initial feature. The first feature and the second initial feature are fused to obtain a second feature. Based on the first feature and the second feature, the prediction results corresponding to each task of the initial multi-task model are determined. The prediction results include, for example, click-through rate, conversion rate, and like rate. Based on the differences between the prediction results and the sample labeling results for each task, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model.

[0266] In another specific implementation scenario, the embodiments of this disclosure can be applied to tasks such as classification, detection, and segmentation of target objects in images. Specifically, a first sample set and a second sample set are obtained. The original feature domain may include image samples from the Internet, image samples from camera components, and the target feature domain may include image samples from multimedia platforms, etc. Based on the first network of the initial multi-task model, feature extraction is performed on the first feature data of the samples to obtain a first feature. Based on the second network of the initial multi-task model, feature encoding is performed on the second feature data of the samples to obtain a second initial feature. The first feature and the second initial feature are fused to obtain a second feature. Based on the first feature and the second feature, the prediction results corresponding to each task of the initial multi-task model are determined respectively. Based on the difference between the prediction results corresponding to each task and the sample annotation results, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model.

[0267] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0268] Based on the same inventive concept, this application also provides a data processing apparatus for implementing the data processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more data processing apparatus embodiments provided below can be found in the limitations of the data processing method described above, and will not be repeated here.

[0269] In one embodiment, such as Figure 13 As shown, a data processing apparatus 1300 is provided, comprising:

[0270] The first acquisition module 1301 is used to acquire a first sample set and a second sample set; the first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain.

[0271] Feature extraction module 1303 is used to extract features from the first feature data of the sample based on the first network of the initial multi-task model to obtain the first feature;

[0272] The feature fusion module 1305 is used to encode the sample second feature data based on the second network of the initial multi-task model to obtain the second initial feature, and to fuse the first feature and the second initial feature to obtain the second feature.

[0273] The determining module 1307 is used to determine the prediction results corresponding to each task of the initial multi-task model based on the first feature and the second feature;

[0274] The processing module 1309 is used to iteratively adjust the network parameters of the initial multi-task model based on the difference between the prediction results and the sample labeling results corresponding to each task, so as to obtain the corresponding target multi-task model.

[0275] In one embodiment, the first acquisition module is further configured to:

[0276] Obtain characteristic sample feature data corresponding to multiple original resource display scenarios, as well as common sample feature data among different original resource display scenarios;

[0277] Based on the sample features of the aforementioned characteristic sample feature data, a target resource display scenario is determined from the multiple original resource display scenarios;

[0278] Based on the multiple original resource display scenarios, the characteristic sample feature data, and the common sample feature data, the first sample set is determined;

[0279] Acquire new characteristic sample feature data of the target resource display scenario and new common sample feature data between the target display scenario and other original resource display scenarios;

[0280] Based on the target resource display scenario, the first feature data of the sample, the newly added characteristic sample feature data, and the newly added common sample feature data, the second sample set is determined.

[0281] In one embodiment, the first acquisition module is further configured to:

[0282] Obtain the number of sample features and the number of samples in the characteristic sample feature data;

[0283] Based on the ratio of the number of sample features to the corresponding number of samples, the feature proportions corresponding to each original resource display scenario are determined.

[0284] The target resource display scenario is determined based on the feature proportions corresponding to each of the original resource display scenarios.

[0285] In one embodiment, the target resource display scenario includes a subroutine scenario of an application, and the first acquisition module is further configured to:

[0286] Obtain target swipe operation data for a subroutine of the application; wherein the interval between two adjacent swipe operations in the target swipe operation data is less than a preset threshold.

[0287] Based on the target sliding operation data, new personalized feature data for the target resource display scene are determined;

[0288] Obtain interaction data for the target resource landing page performed by a subroutine of the application;

[0289] Based on the interaction data of the target resource landing page, new common sample feature data are determined between the updated display scenario and other original resource display scenarios.

[0290] In one embodiment, the first acquisition module is further configured to:

[0291] Obtain the importance index of each sample feature corresponding to the original feature domain to the prediction result, and update the importance index of each sample feature corresponding to the feature domain to the prediction result.

[0292] Based on the importance index of each sample feature corresponding to the original feature domain, the sample features corresponding to the original feature domain are sorted to obtain the first feature data;

[0293] Based on the importance index of each sample feature corresponding to the updated feature domain, the sample features corresponding to the updated feature domain are sorted to obtain the second feature data;

[0294] Based on the first feature data, a first sample set is obtained, and based on the second feature data, a second sample set is obtained.

[0295] In one embodiment, the first acquisition module is further configured to:

[0296] Based on the features of each sample corresponding to the original feature domain, a first multi-task model is trained, the accuracy index of the first multi-task model is determined, and the accuracy index of the first multi-task model is used as the benchmark accuracy index.

[0297] Target sample features are obtained from each of the sample features. Based on the other sample features besides the target sample features, a second task model is trained, and the accuracy index of the second task model is determined.

[0298] Based on the difference between the accuracy index of the second task model and the benchmark accuracy index, the importance index of the target sample features is determined.

[0299] In one embodiment, the importance index includes a responsiveness index, and the first acquisition module is further configured to:

[0300] Obtain the number of positive samples and the total number of samples for each sample feature in the original feature domain;

[0301] The grouping response rate of each sample feature is obtained by calculating the ratio between the number of positive samples and the total number of samples.

[0302] The baseline response rate is obtained by proportionally calculating the number of positive samples in each sample feature to the total number of samples corresponding to each sample feature.

[0303] Based on the group response rate and the baseline response rate, the response index corresponding to each sample feature is determined.

[0304] In one embodiment, the first acquisition module is further configured to:

[0305] Obtain the resource interaction amount corresponding to each task when it occurs, and sort the samples corresponding to each task based on the resource interaction amount to obtain the first sample sequence;

[0306] Obtain the ratio of the initial sample quantity corresponding to each task to the total initial sample quantity corresponding to each task. Based on the ratio, sort the samples corresponding to each task to obtain the second sample sequence.

[0307] If the first sorting position of the target sample corresponding to the target task in the first sample sequence does not match the second sorting position in the second sample sequence, the target sample is updated so that the second sorting position of the updated target sample is consistent with the first sorting position.

[0308] Based on the updated target sample's first feature data and the original feature domain where the target sample is located, a first sample set is determined.

[0309] In one embodiment, the feature extraction module is further configured to:

[0310] Based on the first encoding network, the first feature data of the sample is encoded to obtain the first initial feature;

[0311] Based on the first feature fusion network and the interaction weights between the original feature domains to which each of the first initial features belongs, the first initial features are fused to obtain the first feature.

[0312] In one embodiment, the feature fusion module is further configured to:

[0313] By concatenating the first feature and the second initial feature, a second intermediate feature is obtained;

[0314] Based on the cross weights between each of the second intermediate features, the second intermediate features are fused to obtain the second feature.

[0315] In one embodiment, the feature fusion module is further configured to:

[0316] Based on the cross weights between each of the second intermediate features, the features in each of the second intermediate features are fused to obtain the first fused feature;

[0317] Based on nonlinear weights, each of the second intermediate features is fused to obtain the second fused feature;

[0318] A second feature is obtained based on the first fusion feature and the second fusion feature.

[0319] In one embodiment, the feature fusion module is further configured to:

[0320] Based on the cross weights in each of the second intermediate features, the second intermediate features are fused to obtain the first fused feature;

[0321] Based on nonlinear weights, each of the first fusion features is fused to obtain the third fusion feature;

[0322] Based on the first fusion feature and the third fusion feature, a second feature is obtained.

[0323] In one embodiment, the determining module is further configured to:

[0324] Based on the gating network corresponding to the current task, feature filtering is performed on the first feature and the first feature to obtain the filtered features;

[0325] Based on the filtered features, predict the prediction result corresponding to the current task;

[0326] The filtered features corresponding to other tasks besides the current task are obtained in each task, and the prediction results corresponding to the other tasks are predicted based on the filtered features corresponding to the other tasks.

[0327] In one embodiment, the processing module is further configured to:

[0328] Based on the prediction results and annotation results corresponding to each task, the first loss corresponding to each task is determined;

[0329] Obtain the noise weights corresponding to each task, and perform fusion processing on the first loss corresponding to each task based on the noise weights to obtain the difference between the prediction results and the labeled results corresponding to each task.

[0330] In one embodiment, such as Figure 14 As shown, a data processing apparatus 1400 is provided, comprising:

[0331] The second acquisition module 1401 is used to acquire the first feature data corresponding to the original feature domain and the second feature data corresponding to the updated feature domain; the updated feature domain includes the original feature domain and the target feature domain.

[0332] The feature extraction module 1303 is used to input the first feature data into the first network in the target multi-task model, and extract features from the first feature data based on the first network to obtain the first feature;

[0333] The feature fusion module 1305 is used to input the second feature data into the second network in the target multi-task model, perform feature encoding on the second feature data based on the second network to obtain the second initial feature, and perform feature fusion on the first feature and the second feature to obtain the second feature.

[0334] The determining module 1307 is used to determine the prediction results corresponding to each task of the target multi-task model based on the first feature and the second feature, respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

[0335] In one embodiment, the determining module is further configured to: input the first feature and the second feature to the tower unit corresponding to each task respectively, and output the corresponding prediction result through each tower unit.

[0336] Each module in the aforementioned data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0337] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 15As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for data processing. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a data processing method.

[0338] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 16 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a data processing method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0339] Those skilled in the art will understand that Figure 16The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0340] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0341] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0342] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0343] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data processing method, characterized in that, The method includes: Obtain a first sample set and a second sample set; the first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain. Based on the first network of the initial multi-task model, feature extraction is performed on the first feature data of the sample to obtain the first feature; Based on the second network of the initial multi-task model, the second feature data of the sample is feature-encoded to obtain the second initial feature. The first feature and the second initial feature are then fused to obtain the second feature. Based on the first feature and the second feature, the prediction results corresponding to each task of the initial multi-task model are determined respectively; Based on the differences between the prediction results and the sample labeling results for each task, the network parameters of the initial multi-task model are iteratively adjusted to obtain the corresponding target multi-task model.

2. The method according to claim 1, characterized in that, The process of obtaining the first sample set and the second sample set includes: Obtain characteristic sample feature data corresponding to multiple original resource display scenarios, as well as common sample feature data among different original resource display scenarios; Based on the sample features of the aforementioned characteristic sample feature data, a target resource display scenario is determined from the multiple original resource display scenarios; Based on the multiple original resource display scenarios, the characteristic sample feature data, and the common sample feature data, the first sample set is determined; Acquire new characteristic sample feature data of the target resource display scenario and new common sample feature data between the target display scenario and other original resource display scenarios; Based on the target resource display scenario, the first feature data of the sample, the newly added characteristic sample feature data, and the newly added common sample feature data, the second sample set is determined.

3. The method according to claim 2, characterized in that, The determination of the target resource display scenario from the plurality of original resource display scenarios based on the sample features of the characteristic sample feature data includes: Obtain the number of sample features and the number of samples in the characteristic sample feature data; Based on the ratio of the number of sample features to the corresponding number of samples, the feature proportions corresponding to each original resource display scenario are determined. The target resource display scenario is determined based on the feature proportions corresponding to each of the original resource display scenarios.

4. The method according to claim 2, characterized in that, The target resource display scenario includes sub-program scenarios of the application. The step of obtaining newly added personalized sample feature data of the target resource display scenario and newly added common sample feature data between the updated display scenario and other original resource display scenarios includes: Obtain target swipe operation data for a subroutine of the application; wherein the interval between two adjacent swipe operations in the target swipe operation data is less than a preset threshold. Based on the target sliding operation data, new personalized feature data for the target resource display scene are determined; Obtain interaction data for the target resource landing page performed by a subroutine of the application; Based on the interaction data of the target resource landing page, new common sample feature data are determined between the updated display scenario and other original resource display scenarios.

5. The method according to claim 1, characterized in that, The process of obtaining the first sample set and the second sample set includes: Obtain the importance index of each sample feature corresponding to the original feature domain to the prediction result, and update the importance index of each sample feature corresponding to the feature domain to the prediction result. Based on the importance index of each sample feature corresponding to the original feature domain, the sample features corresponding to the original feature domain are sorted to obtain the first feature data; Based on the importance index of each sample feature corresponding to the updated feature domain, the sample features corresponding to the updated feature domain are sorted to obtain the second feature data; Based on the first feature data, a first sample set is obtained, and based on the second feature data, a second sample set is obtained.

6. The method according to claim 5, characterized in that, The method of obtaining the importance index of each sample feature corresponding to the original feature domain to the prediction result includes: Based on the features of each sample corresponding to the original feature domain, a first multi-task model is trained, the accuracy index of the first multi-task model is determined, and the accuracy index of the first multi-task model is used as the benchmark accuracy index. Target sample features are obtained from each of the sample features. Based on the other sample features besides the target sample features, a second task model is trained, and the accuracy index of the second task model is determined. Based on the difference between the accuracy index of the second task model and the benchmark accuracy index, the importance index of the target sample features is determined.

7. The method according to claim 5, characterized in that, The importance index includes a responsiveness index, and the importance index of each sample feature corresponding to the original feature domain to the prediction result includes: Obtain the number of positive samples and the total number of samples for each sample feature in the original feature domain; The grouping response rate of each sample feature is obtained by calculating the ratio between the number of positive samples and the total number of samples. The baseline response rate is obtained by proportionally calculating the number of positive samples in each sample feature to the total number of samples corresponding to each sample feature. Based on the group response rate and the baseline response rate, the response index corresponding to each sample feature is determined.

8. The method according to claim 1, characterized in that, Each task in the initial multi-task model includes interactive operations targeting a target resource; obtaining the first sample set includes: Obtain the resource interaction amount corresponding to each task when it occurs, and sort the samples corresponding to each task based on the resource interaction amount to obtain the first sample sequence; Obtain the ratio of the initial sample quantity corresponding to each task to the total initial sample quantity corresponding to each task. Based on the ratio, sort the samples corresponding to each task to obtain the second sample sequence. If the first sorting position of the target sample corresponding to the target task in the first sample sequence does not match the second sorting position in the second sample sequence, the target sample is updated so that the second sorting position of the updated target sample is consistent with the first sorting position. Based on the updated target sample's first feature data and the original feature domain where the target sample is located, a first sample set is determined.

9. The method according to claim 1, characterized in that, The first network includes a first encoding network and a first feature fusion network. The step of extracting features from the sample's first feature data based on the first network to obtain first features includes: Based on the first encoding network, the first feature data of the sample is encoded to obtain the first initial feature; Based on the first feature fusion network and the interaction weights between the original feature domains to which each of the first initial features belongs, the first initial features are fused to obtain the first feature.

10. The method according to claim 1, characterized in that, The step of fusing the first feature and the second initial feature to obtain the second feature includes: By concatenating the first feature and the second initial feature, a second intermediate feature is obtained; Based on the cross weights between each of the second intermediate features, the second intermediate features are fused to obtain the second feature.

11. The method according to claim 10, characterized in that, The step of fusing the second intermediate features based on the cross weights between them to obtain the second feature includes: Based on the cross weights between each of the second intermediate features, the features in each of the second intermediate features are fused to obtain the first fused feature; Based on nonlinear weights, each of the second intermediate features is fused to obtain the second fused feature; Based on the first fusion feature and the second fusion feature, a second feature is obtained.

12. The method according to claim 10, characterized in that, The step of fusing the second intermediate feature vectors based on the cross weights between them to obtain the second feature vector includes: Based on the cross weights in each of the second intermediate features, the second intermediate features are fused to obtain the first fused feature; Based on nonlinear weights, each of the first fusion features is fused to obtain the third fusion feature; Based on the first fusion feature and the third fusion feature, a second feature is obtained.

13. The method according to claim 1, characterized in that, The step of determining the prediction results for each task of the initial multi-task model based on the first feature and the second feature includes: Based on the gating network corresponding to the current task, feature filtering is performed on the first feature and the first feature to obtain the filtered features; Based on the filtered features, predict the prediction result corresponding to the current task; The filtered features corresponding to other tasks besides the current task are obtained in each task, and the prediction results corresponding to the other tasks are predicted based on the filtered features corresponding to the other tasks.

14. The method according to claim 1, characterized in that, The process of iteratively adjusting the network parameters of the initial multi-task model based on the differences between the prediction results and sample labeling results for each task to obtain the corresponding target multi-task model includes: Based on the prediction results and annotation results corresponding to each task, the first loss corresponding to each task is determined; Obtain the noise weights corresponding to each task, and perform fusion processing on the first loss corresponding to each task based on the noise weights to obtain the difference between the prediction results and the labeled results corresponding to each task.

15. A data processing method, characterized in that, The method includes: Acquire the first feature data of the target resource corresponding to the original feature domain of the target, and the second feature data of the target resource corresponding to the updated feature domain of the target; the updated feature domain of the target includes the original feature domain of the target and the target feature domain. The first feature data is input into the first network in the target multi-task model, and feature extraction is performed on the first feature data based on the first network to obtain the first feature; The second feature data is input into the second network in the target multi-task model. Based on the second network, the second feature data is feature encoded to obtain the second initial feature. The first feature and the second feature are then fused to obtain the second feature. Based on the first feature and the second feature, the prediction results corresponding to each task of the target multi-task model are determined respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

16. The method according to claim 15, characterized in that, Also includes: Based on the prediction results corresponding to each target resource, the target resources are sorted to obtain the sorting results; Based on the ranking results, the recommended results for each target resource are determined.

17. A data processing apparatus, characterized in that, The device includes: The first acquisition module is used to acquire a first sample set and a second sample set; the first sample set includes sample first feature data corresponding to the original feature domain, and the second sample set includes sample second feature data corresponding to the updated feature domain. The updated feature domain includes the original feature domain and a target feature domain. The target feature domain includes sample features determined from the original feature domain based on the original feature domain. The feature extraction module is used to extract features from the first feature data of the sample based on the first network of the initial multi-task model to obtain the first feature; The feature fusion module is used to encode the second feature data of the sample based on the second network of the initial multi-task model to obtain the second initial feature, and to fuse the first feature and the second initial feature to obtain the second feature; The determining module is used to determine the prediction results corresponding to each task of the initial multi-task model based on the first feature and the second feature; The processing module is used to iteratively adjust the network parameters of the initial multi-task model based on the difference between the prediction results and the sample labeling results for each task, so as to obtain the corresponding target multi-task model.

18. A data processing apparatus, characterized in that, The device includes: The second acquisition module is used to acquire the first feature data corresponding to the original feature domain and the second feature data corresponding to the updated feature domain; the updated feature domain includes the original feature domain and the target feature domain. The feature extraction module is used to input the first feature data into the first network in the target multi-task model, and extract features from the first feature data based on the first network to obtain the first feature; The feature fusion module is used to input the second feature data into the second network in the target multi-task model, perform feature encoding on the second feature data based on the second network to obtain the second initial feature, and perform feature fusion on the first feature and the second feature to obtain the second feature; The determination module is used to determine the prediction results corresponding to each task of the target multi-task model based on the first feature and the second feature, respectively; wherein, the target multi-task model is obtained by training based on the sample first feature data corresponding to the original feature domain and the sample second feature data corresponding to the updated feature domain, and the target feature domain is determined from the sample features based on the original feature domain.

19. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 14, or the steps of the method according to any one of claims 15 to 16.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 14, or the steps of the method according to any one of claims 15 to 16.

21. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 14, or the steps of the method according to any one of claims 15 to 16.

Images