An intelligent update system for advertising production feature database

Abstract

The present invention relates to the field of data update technology, and specifically to an intelligent update system for advertising feature library data, including the following modules: a data collection module: real-time collection of scarce data including market segments from multiple channels; a data enhancement module: an adaptive migration learning engine based on a deep neural network, using relevant data to enrich domain data to enhance scarce data, and improving the feature understanding of the scarce data part by migrating the learned knowledge; an evaluation module: using an advertising effect evaluation model, using the enhanced scarce data as input to evaluate the advertising effect, and deciding whether to update the scarce data; a data update module: when determining that an update is required, automatically adjusting the relevant data items in the feature library to ensure the timeliness and accuracy of the advertising feature library data. The present invention improves the coverage and expression of scarce data in the advertising feature library, and also ensures that the advertising feature library can timely reflect the latest market trends and changes in user behavior.

Description

[0001] This application is a divisional application of the application filed on July 1, 2024, with application number 202410869474.3 and invention name “A system for intelligently updating advertising production feature library data”. Technical Field

[0002] The present invention relates to the technical field of data updating, and in particular to an intelligent updating system for advertisement production feature library data. Background Art

[0003] In the field of digital advertising, the production of advertising content and the optimization of delivery strategies are crucial. Traditionally, these strategies rely on ad production feature libraries, which store a large amount of data on target audiences, advertising content, delivery channels, etc. These data are used to guide the customization and delivery of advertisements, aiming to maximize the attractiveness and conversion rate of advertisements. However, with the rapid changes in the market and the diversity of user needs, the existing ad production feature libraries face several major challenges:

[0004] Data timeliness and accuracy issues: Traditional advertising feature library update mechanisms often fail to reflect the latest market trends and user preferences in a timely manner, causing advertising content and strategies to lag behind actual market demand.

[0005] Scarce data problem: For emerging markets or specific market segments, relevant data is often scarce and incomplete, which limits the effectiveness and personalization of advertising strategies.

[0006] Knowledge transfer and adaptability issues: Traditional methods often lack effective knowledge transfer and rapid adaptation mechanisms when dealing with new markets or rapidly changing market environments, making it difficult to use existing data and knowledge to support new advertising decisions.

[0007] Therefore, there is an urgent need for a new solution that can improve the timeliness and accuracy of the advertising feature library data, effectively solve the problem of scarce data, and enhance the adaptability and response speed of the advertising system to new markets. Summary of the invention

[0008] Based on the above objectives, the present invention provides an intelligent updating system for advertisement production feature library data.

[0009] An advertisement production feature library data intelligent update system, comprising the following modules:

[0010] Data collection module: collects scarce data including market segments from multiple channels in real time;

[0011] Data enhancement module: An adaptive transfer learning engine based on deep neural networks that uses relevant data to enrich domain data to enhance scarce data and improves feature understanding of scarce data by transferring learned knowledge;

[0012] Evaluation module: Use the advertising effect evaluation model to evaluate the advertising effect with the enhanced scarce data as input and decide whether to update the scarce data;

[0013] Data update module: When it is determined that an update is needed, the relevant data items in the feature library are automatically adjusted to ensure the timeliness and accuracy of the advertising feature library data.

[0014] Furthermore, the data collection module specifically includes:

[0015] Establish connections with different data sources through multi-channel data interfaces, including social media platforms, online forums, e-commerce websites, and news portals;

[0016] Based on market segment data identification, scarce data for specific market segments are identified and extracted from the collected data. The market segments are defined based on geographic location, interest groups, and industry segmentation standards. First, natural language processing technology is used to extract keywords and phrases from the collected text data, and topic modeling technology is used to identify themes in the text data. According to predefined market segment standards (such as geographic location, age group, interests and hobbies, etc.), the identified keywords and topics are mapped to corresponding market segments.

[0017] Furthermore, the adaptive transfer learning engine based on deep neural network in the data enhancement module specifically includes:

[0018] Domain Adaptation Layer: Introducing a domain adaptation layer to automatically adjust the feature distribution differences between the source domain and the target domain to narrow the distance between different domains and improve the effect of transfer learning. The source domain is a data-rich domain, and the target domain is a segmented market with scarce data.

[0019] Multi-task learning combination: In order to make full use of various related tasks in the advertising field (such as click-through rate prediction, user behavior classification, etc.), a multi-task learning framework is adopted to learn common features across tasks by sharing the underlying representation while maintaining the output layer of the task to enhance the understanding and adaptability of scarce data;

[0020] Meta-learning strategy: Integrate meta-learning mechanism, which is used to adapt to scarce data in new market segments. By training on multiple market segment tasks, the adaptive transfer learning engine can quickly adjust parameters when encountering new scarce data to achieve knowledge transfer;

[0021] Enhanced feature distillation: In order to extract richer and more useful features from the source domain and transfer them to the target domain, the enhanced feature distillation technology is used to enhance the feature expression ability of scarce data by guiding the target model to learn the intermediate layer feature representation of the source model;

[0022] Dynamic transfer weight adjustment: According to the specific requirements of the target task and the characteristics of scarce data, the transfer weights of different layers and different features in the transfer learning process are dynamically adjusted. This dynamic adjustment mechanism ensures that the model can adapt to various scarce data situations more flexibly.

[0023] Furthermore, the domain adaptation layer specifically includes:

[0024] Feature extraction: The data in the source domain (data-rich domain) and the target domain (segmented market data-scarce domain) are extracted through the initial layer of the deep neural network. Let f(·) be the feature extraction function. For the source domain data X s and target domain data X t The extracted features are f(X s ) and f(X t );

[0025] The domain adaptation layer adjusts the feature distribution by minimizing the inter-domain difference metric to make the feature distribution of the source domain and the target domain close, based on the maximum mean difference (MMD), calculated as:

[0026]

[0027] Among them, N s and N t are the number of samples in the source domain and the target domain, respectively. is the mapping function to the Reproducing Kernel Hilbert Space (RKHS), and are samples from the source domain and the target domain respectively;

[0028] Adaptive layer training: During the training process, in addition to the traditional task-related loss, a domain adaptation loss term, namely MMD loss, is added to encourage the model to minimize the feature distribution difference between the source domain and the target domain. The total loss function is expressed as:

[0029] L1=L task +λ·MMD 2 (f(X s ), f(X t )), where L task is the task-related loss (e.g., cross entropy loss for classification tasks), and λ is a balancing factor used to adjust the importance between task loss and domain adaptation loss.

[0030] Furthermore, the multi-task learning combination specifically includes:

[0031] Bottom-layer shared representation: In the multi-task learning framework, the bottom-layer network layer is designed as a shared layer to extract common features from the data of all tasks. Suppose there are T related tasks, f shared (·) represents the feature extraction function of the shared layer. For any task t, the input data E t , shared as f shared (E t );

[0032] Task-specific output layer: For each task t, there is a corresponding output layer It is used to predict the corresponding task based on the shared features. The prediction output of task t is represented as f task (f shared (E t ));

[0033] Loss function of multi-task learning: In the multi-task learning framework, the total loss function is the weighted sum of all task loss functions. For each task t, there is a loss function corresponding to the task. Used to evaluate the prediction performance of this task, therefore, the total loss function is expressed as:

[0034] Among them, α t is the weight of task t, which is used to balance the importance of different tasks in the total loss, Y t is the true label of task t.

[0035] Furthermore, the meta-learning strategy specifically includes:

[0036] Initialization: Choose an initial parameter θ;

[0037] Task sampling: From the task distribution Randomly sample a batch of tasks

[0038] Adaptive update for each task: For each task Using the initial parameters θ and the task The training data set is updated by gradient descent to obtain the adapted parameters θ i ′, the adapted parameters are calculated as follows: Among them, α is the learning rate, It's a task The loss function of

[0039] Meta-update: Use the adapted parameters θ from all sampled tasks i ′ calculates the loss on their respective validation datasets and meta-updates the initial parameters θ. The meta-update is calculated as follows:

[0040] Here, β is the meta-learning rate.

[0041] Furthermore, the enhanced characteristic distillation technology specifically includes:

[0042] Pre-training of source model: The source model is pre-trained on a rich dataset to learn complex feature representations. Let the source model be F s (·);

[0043] Intermediate layer feature extraction: In the source model F s Select one or more intermediate layers to extract feature representation. For a given input data X, the intermediate layer feature representation of the source model is

[0044] Target model training and feature distillation: Target model F t (·) When training on scarce data, in addition to the traditional task loss, feature distillation loss is introduced to make the intermediate layer feature representation of the target model close to the corresponding layer feature representation of the source model. The feature distillation loss is calculated as follows: in, and denote the feature representations of the target model and the source model at the i-th intermediate layer, respectively, ‖·‖ 2 represents the squared L2 norm;

[0045] Total loss function: The total loss function of the target model combines the task-related loss and feature distillation loss, that is: L3 = L task (F t (X), Y) + λ·L distill , where L task is the task-related loss function, Y is the true label, and λ is the weight parameter used to balance the two losses.

[0046] Furthermore, the dynamic migration weight adjustment specifically includes:

[0047] Setting of weight adjustment mechanism: Setting a learnable weight parameter α for each layer l and each feature f l,f , the weight parameters will be used to adjust the contribution of each layer and each feature to the final transfer learning performance;

[0048] Importance evaluation at the layer and feature level: During the training process, an additional loss term is introduced to evaluate the contribution of each layer and feature to the performance of the target task. This is achieved by calculating the gradient size of each layer and feature. The gradient size reflects the sensitivity of the parameter to the loss. For feature f in layer l, its importance I l,f By the following calculation:

[0049] Among them, Ltask is the task-related loss function, F t (X; α) is the output of the weighted target model, Y is the true label, Represents the weight parameter α l,f The gradient of

[0050] Dynamic weight update: Based on the importance evaluation of each layer and feature l,f , dynamically adjust the weight parameter a l,f , the weight update is performed as follows: Where η is the update step size, sign(I l,f ) is a sign function evaluated based on importance, indicating whether the weight increases or decreases.

[0051] Furthermore, the advertising effect evaluation model adopts a gradient boosting decision tree, including receiving scarce data processed by a data enhancement module, using a gradient boosting decision tree to construct an advertising effect prediction model, and predicting the potential effect of the advertisement based on the enhanced data, including click-through rate, conversion rate or user engagement, as follows:

[0052] Initialization: Use the most frequent class in the data as the initial prediction value;

[0053] Iterative training: For each iteration n: calculate the residual of the current model, that is, the difference between the true value and the current predicted value; train a decision tree based on the current residual to try to predict the residual; add the decision tree to the model and update the model's predicted value;

[0054] Model update: The model’s prediction is the weighted sum of the predictions of all decision trees, and the contribution of each tree is scaled by the learning rate (η);

[0055] For the nth iteration, the contribution of the mth tree is expressed as:

[0056] F m (x) = F m-1 (x)+η·h m (x), where F m-1 (x) is the predicted value of the previous model, h m (x) is the prediction value of the residual by the mth tree, and η is the learning rate, which controls the contribution of each tree to the final prediction;

[0057] According to the advertising effect prediction results output by the model, combined with the preset effect threshold or improvement index, it is judged whether the advertising effect under the current enhanced data meets the expectations. If the effect does not meet the standard or there is room for improvement, the data update module is triggered to update and optimize the relevant scarce data in the feature library. Otherwise, the existing data status is maintained.

[0058] Beneficial effects of the present invention:

[0059] The present invention, by introducing an adaptive transfer learning engine based on deep neural networks, can effectively utilize pre-trained models and knowledge in rich fields to enhance and optimize scarce data, which not only improves the coverage and expressiveness of scarce data in the advertising feature library, but also ensures that the advertising feature library can timely reflect the latest market trends and changes in user behavior, thereby improving the accuracy and timeliness of advertising strategies.

[0060] The present invention utilizes enhanced data processed by an adaptive transfer learning engine, combined with an efficient advertising effectiveness evaluation model, to more accurately predict and evaluate the potential effects of advertising content and strategies, help identify and optimize inefficient advertising features, and continuously improve the ROI of advertising through continuous data updates and feature optimization. In addition, the dynamic migration weight adjustment mechanism further enhances the adaptability and flexibility of the model in different market segments and data environments.

[0061] The present invention integrates a meta-learning mechanism and enhanced feature distillation technology, which enables rapid capture and adaptation to newly emerging market segments and user needs. The meta-learning mechanism finds model parameter initialization that can quickly adapt to new tasks by training on multiple market segment tasks, while the feature distillation technology helps the target model learn the high-level feature representation of the source model, thereby accelerating the knowledge transfer and learning process. These mechanisms work together to improve the responsiveness of the advertising system to new markets and the effectiveness of advertising strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0063] Figure 1 A schematic diagram of system function modules according to an embodiment of the present invention;

[0064] Figure 2 Schematic diagram of an adaptive transfer learning engine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0066] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0067] like Figure 1-2 As shown, a system for intelligently updating advertisement production feature library data includes the following modules:

[0068] Data collection module: collects scarce data including market segments from multiple channels in real time;

[0069] Data enhancement module: An adaptive transfer learning engine based on deep neural networks that uses relevant data to enrich domain data to enhance scarce data and improves feature understanding of scarce data by transferring learned knowledge;

[0070] Evaluation module: Use the advertising effect evaluation model to evaluate the advertising effect with the enhanced scarce data as input and decide whether to update the scarce data;

[0071] Data update module: When it is determined that an update is needed, the relevant data items in the feature library are automatically adjusted to ensure the timeliness and accuracy of the advertising feature library data.

[0072] The data collection module specifically includes:

[0073] Establish connections with different data sources through multi-channel data interfaces, including social media platforms, online forums, e-commerce websites, and news portals;

[0074] Based on market segment data identification, scarce data for specific market segments are identified and extracted from the collected data. Market segments are defined based on geographic location, interest groups, and industry segmentation standards. First, natural language processing technology is used to extract keywords and phrases from the collected text data, and topic modeling technology is used to identify themes in the text data. According to predefined market segment standards (such as geographic location, age group, interests and hobbies, etc.), the identified keywords and topics are mapped to the corresponding market segments.

[0075] The deep neural network-based adaptive transfer learning engine in the data enhancement module specifically includes:

[0076] Domain Adaptation Layer: The domain adaptation layer is introduced to automatically adjust the feature distribution difference between the source domain and the target domain to narrow the distance between different domains and improve the effect of transfer learning. The source domain is a data-rich domain, and the target domain is a segmented market with scarce data.

[0077] Multi-task learning combination: In order to make full use of various related tasks in the advertising field (such as click-through rate prediction, user behavior classification, etc.), a multi-task learning framework is adopted to learn common features across tasks by sharing the underlying representation while maintaining the output layer of the task to enhance the understanding and adaptability of scarce data;

[0078] Meta-learning strategy: Integrate meta-learning mechanism, which is used to adapt to scarce data in new market segments. By training on multiple market segment tasks, the adaptive transfer learning engine can quickly adjust parameters when encountering new scarce data to achieve knowledge transfer;

[0079] Enhanced feature distillation: In order to extract richer and more useful features from the source domain and transfer them to the target domain, the enhanced feature distillation technology is used to enhance the feature expression ability of scarce data by guiding the target model to learn the intermediate layer feature representation of the source model;

[0080] Dynamic transfer weight adjustment: According to the specific requirements of the target task and the characteristics of scarce data, the transfer weights of different layers and different features in the transfer learning process are dynamically adjusted. This dynamic adjustment mechanism ensures that the model can adapt to various scarce data situations more flexibly.

[0081] The domain adaptation layer specifically includes:

[0082] Feature extraction: The data in the source domain (data-rich domain) and the target domain (segmented market data-scarce domain) are extracted through the initial layer of the deep neural network. Let f(·) be the feature extraction function. For the source domain data X s and target domain data X t The extracted features are f(X s ) and f(X t );

[0083] The domain adaptation layer adjusts the feature distribution by minimizing the inter-domain difference metric to make the feature distribution of the source domain and the target domain close, based on the maximum mean difference (MMD), calculated as:

[0084]

[0085] Among them, N s and Nt are the number of samples in the source domain and the target domain, respectively. is the mapping function to the Reproducing Kernel Hilbert Space (RKHS), and are samples from the source domain and the target domain respectively;

[0086] Adaptive layer training: During the training process, in addition to the traditional task-related loss, a domain adaptation loss term, namely MMD loss, is added to encourage the model to minimize the feature distribution difference between the source domain and the target domain. The total loss function is expressed as:

[0087] L1=L task +λ·MMD 2 (f(X s ), f(X t )), where L task is the task-related loss (e.g., cross entropy loss for classification tasks), and λ is a balancing factor used to adjust the importance between task loss and domain adaptation loss.

[0088] In the present invention, the domain adaptation layer is implemented through the above mechanism, which can effectively utilize the rich advertisement click-through rate data, user interaction history and other information in the source domain to enhance the scarce data feature expression in the target domain. By reducing the feature distribution difference between the source domain and the target domain (i.e., the scarce data domain in a specific market segment), the deep adaptive transfer learning engine can more effectively transfer the knowledge learned in the source domain, thereby improving the understanding of the advertising characteristics of the market segment, further optimizing the advertising content and delivery strategy, and improving the advertising effect.

[0089] The combination of multi-task learning specifically includes:

[0090] Bottom-layer shared representation: In the multi-task learning framework, the bottom-layer network layer is designed as a shared layer to extract common features from the data of all tasks. Suppose there are T related tasks, f shared (·) represents the feature extraction function of the shared layer. For any task t, the input data E t , shared as f shared (E t );

[0091] Task-specific output layer: For each task t, there is a corresponding output layer It is used to predict the corresponding task based on the shared features. The prediction output of task t is represented as f task (f shared (E t ));

[0092] Loss function of multi-task learning: In the multi-task learning framework, the total loss function is the weighted sum of all task loss functions. For each task t, there is a loss function corresponding to the task. Used to evaluate the prediction performance of this task, therefore, the total loss function is expressed as:

[0093] Among them, α t is the weight of task t, which is used to balance the importance of different tasks in the total loss, Y t is the true label of task t.

[0094] In the context of advertising feature databases, the multi-task learning framework can be used to simultaneously process multiple tasks related to advertising effectiveness, such as click-through rate prediction, user conversion prediction, and advertising content classification. The shared layer can extract common user behavior and advertising response features from the advertising data, while the task-specific output layer is optimized for each task.

[0095] For example, the task-specific layer for click-through rate prediction can focus on identifying which advertising features are most likely to attract user clicks, while the task-specific layer for user conversion prediction can focus on identifying which features are most likely to promote users to complete purchases. In this way, the multi-task learning framework not only improves the model's understanding and adaptability to scarce data, but also promotes cross-learning and transfer of knowledge between different advertising effectiveness tasks, further improving the quality of the advertising feature library data and the effectiveness of advertising strategies.

[0096] The meta-learning mechanism is applied in a multi-task learning environment to train a model so that it can quickly adapt to new tasks or data. This mechanism is suitable for dealing with the scarce data problem of new market segments in the advertising field. The commonly used meta-learning method is model-agnostic meta-learning (MAML). Its basic principle is to find a good model parameter initialization by training on multiple tasks, so that the initialization can quickly adapt to new tasks after a small number of gradient update steps. The details are as follows:

[0097] Initialization: Choose an initial parameter θ;

[0098] Task sampling: From the task distribution Randomly sample a batch of tasks

[0099] Adaptive update for each task: For each task Using the initial parameters θ and the task The training data set is updated by gradient descent to obtain the adapted parameters θ i ′, the adapted parameters are calculated as follows: Among them, α is the learning rate, It's a task The loss function of

[0100] Meta-update: Use the adapted parameters θ from all sampled tasks i ′ calculates the loss on their respective validation datasets and meta-updates the initial parameters θ. The meta-update is calculated as follows:

[0101] Here, β is the meta-learning rate.

[0102] In the advertisement feature database used in the present invention, each market segment can be regarded as a task It contains scarce data unique to the market segment. Through meta-learning mechanisms such as MAML, the adaptive transfer learning engine can be trained on multiple market segment tasks to find a model parameter initialization that can quickly adapt to the scarce data of the new market segment. When encountering a new market segment (i.e., a new task), the good parameter initialization found through meta-learning can quickly adapt to the specific needs of the new market with a small amount of gradient updates, thereby achieving effective knowledge transfer and rapid adaptation.

[0103] This meta-learning mechanism enables the adaptive transfer learning engine based on deep neural networks to not only use data from existing market segments to improve the accuracy of ad features and the effectiveness of ad strategies, but also quickly adapt to emerging market segments, even when data in these new markets is very scarce. This is extremely valuable for advertising systems because it ensures that ad content and strategies are always up to date and relevant, thereby maximizing ad effectiveness and ROI.

[0104] The enhanced feature distillation technology specifically includes:

[0105] Pre-training of source model: The source model is pre-trained on a rich dataset to learn complex feature representations. Let the source model be F s (·);

[0106] Intermediate layer feature extraction: In the source model F s Select one or more intermediate layers to extract feature representation. For a given input data X, the intermediate layer feature representation of the source model is

[0107] Target model training and feature distillation: Target model F t (·) When training on scarce data, in addition to the traditional task loss, feature distillation loss is introduced to make the intermediate layer feature representation of the target model close to the corresponding layer feature representation of the source model. The feature distillation loss is calculated as follows: in, and denote the feature representations of the target model and the source model at the i-th intermediate layer, respectively, ‖·‖ 2 represents the squared L2 norm;

[0108] Total loss function: The total loss function of the target model combines the task-related loss and feature distillation loss, that is: L3 = L task (F t (X), Y) + λ·L distill , where L task is the task-related loss function, Y is the true label, and λ is the weight parameter used to balance the two losses.

[0109] In the scenario of an advertising feature database, the source model is a deep model pre-trained on a wide range of advertising data, which has learned a rich representation of advertising features. The target model is a model trained for scarce data in a specific market segment. By guiding the target model to learn the intermediate-layer feature representation of the source model, the target model's ability to understand and express the advertising data of the market segment can be significantly improved, thereby improving the accuracy of advertising target prediction and the relevance of advertising content. Feature distillation technology enables the target model to effectively utilize the complex features learned by the source model on rich data when data is scarce, accelerate the knowledge transfer and learning process, and improve the quality of the advertising feature library and the effectiveness of advertising strategies.

[0110] Dynamic migration weight adjustment specifically includes:

[0111] Setting of weight adjustment mechanism: Setting a learnable weight parameter α for each layer l and each feature f l,f , the weight parameters will be used to adjust the contribution of each layer and each feature to the final transfer learning performance;

[0112] Importance evaluation at the layer and feature level: During the training process, an additional loss term is introduced to evaluate the contribution of each layer and feature to the performance of the target task. This is achieved by calculating the gradient size of each layer and feature. The gradient size reflects the sensitivity of the parameter to the loss. For feature f in layer l, its importance I l,f By the following calculation:

[0113] Among them, L task is the task-related loss function, F t (X; α) is the output of the weighted target model, Y is the true label, Represents the weight parameter α l,f The gradient of

[0114] Dynamic weight update: Based on the importance evaluation of each layer and feature l,f , dynamically adjust the weight parameter α l,f , the weight update is performed as follows: Where η is the update step size, sign(I l,f ) is a sign function evaluated based on importance, indicating whether the weight increases or decreases.

[0115] In the scenario of the advertising feature database, dynamic migration weight adjustment allows the adaptive migration learning engine to flexibly adjust the contribution of each layer and feature in the migration process according to the characteristics of the scarce data in a specific market segment and the specific needs of the target task. For example, for a specific market segment, if a feature layer is found to be particularly critical to improving the accuracy of ad click-through rate prediction, the weight of that layer will be increased to strengthen its role in the final model. This dynamic adjustment mechanism ensures the flexibility and efficiency of the migration learning process, helps improve model performance under scarce data, and then optimizes advertising content and strategies to improve advertising effectiveness.

[0116] The advertising effect evaluation model uses a gradient boosting decision tree, including receiving scarce data processed by the data enhancement module, using the gradient boosting decision tree to build an advertising effect prediction model, and predicting the potential effect of the advertisement based on the enhanced data, including click-through rate, conversion rate or user engagement, as follows:

[0117] Initialization: Use the most frequent class in the data as the initial prediction value;

[0118] Iterative training: For each iteration n: calculate the residual of the current model, that is, the difference between the true value and the current predicted value; train a decision tree based on the current residual to try to predict the residual; add the decision tree to the model and update the model's predicted value;

[0119] Model update: The model’s prediction is the weighted sum of the predictions of all decision trees, and the contribution of each tree is scaled by the learning rate (η);

[0120] For the nth iteration, the contribution of the mth tree is expressed as:

[0121] F m (x) = F m-1 (x)+η·h m (x), where F m-1 (x) is the predicted value of the previous model, h m (x) is the prediction value of the residual by the mth tree, and η is the learning rate, which controls the contribution of each tree to the final prediction;

[0122] According to the advertising effect prediction results output by the model, combined with the preset effect threshold or improvement index, it is judged whether the advertising effect under the current enhanced data meets the expectations. If the effect does not meet the standard or there is room for improvement, the data update module is triggered to update and optimize the relevant scarce data in the feature library. Otherwise, the existing data status is maintained.

[0123] The data update module specifically includes:

[0124] Feature integration preparation: Verify new features before integrating them to ensure compatibility with the existing feature library. Establish a version control mechanism for the feature library to ensure that each feature update is recorded to facilitate change tracking and rollback.

[0125] Database structure update: Evaluate the structure of the current ad feature library to determine whether it needs to be adjusted to accommodate new features. For example, new tables, columns, or indexes need to be added. Update the database structure as needed to ensure that the new feature set can be stored and indexed.

[0126] Feature data integration: Migrate optimized feature data to the ad feature library, including copying data, converting data formats, or mapping data to a new structure;

[0127] Update process automation: Develop automated scripts or tools to handle feature integration and data update processes, including steps such as data validation, structure adjustment, and data migration.

[0128] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0129] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent updating system for advertisement production feature library data, characterized in that: Includes the following modules: Data collection module: collects scarce data including market segments from multiple channels in real time, including: Establish connections with different data sources through multi-channel data interfaces, including social media platforms, online forums, e-commerce websites, and news portals; Based on market segment data identification, identify and extract scarce data for specific market segments from the collected data. The market segments are defined based on geographic location, interest groups, and industry segmentation standards. First, natural language processing technology is used to extract keywords and phrases from the collected text data. The topic modeling technology is used to identify the topics in the text data. According to the predefined market segment standards, the identified keywords and topics are mapped to the corresponding market segments; Data enhancement module: An adaptive transfer learning engine based on deep neural networks. It uses relevant data to enrich domain data to enhance scarce data, and improves the feature understanding of scarce data by transferring the learned knowledge. The adaptive transfer learning engine based on deep neural networks specifically includes: Domain Adaptation Layer: Introducing a domain adaptation layer to automatically adjust the feature distribution differences between the source domain and the target domain to narrow the distance between different domains and improve the effect of transfer learning. The source domain is a data-rich domain, and the target domain is a segmented market with scarce data. Multi-task learning combination: adopt a multi-task learning framework to learn common features across tasks by sharing the underlying representation while maintaining the output layer of the task, enhancing the understanding and adaptability of scarce data; Meta-learning strategy: Integrate meta-learning mechanism, which is used to adapt to scarce data in new market segments. By training on multiple market segment tasks, the adaptive transfer learning engine can quickly adjust parameters when encountering new scarce data to achieve knowledge transfer; Enhanced feature distillation: Enhanced feature distillation technology is used to enhance the feature expression ability of scarce data by guiding the target model to learn the intermediate layer feature representation of the source model; Dynamic transfer weight adjustment: Dynamically adjust the transfer weights of different layers and features in the transfer learning process according to the specific requirements of the target task and the characteristics of scarce data; The dynamic migration weight adjustment specifically includes: Setting of weight adjustment mechanism: for each layer and each feature Set a learnable weight parameter , the weight parameters will be used to adjust the contribution of each layer and each feature to the final transfer learning performance; Importance evaluation at the layer and feature level: During the training process, additional loss terms are introduced to evaluate the contribution of each layer and feature to the performance of the target task. This is achieved by calculating the gradient size of each layer and feature. The gradient size reflects the sensitivity of the parameter to the loss. Features in , its importance By the following calculation: ,in, is the task-related loss function, is the output of the weighted target model, is the true label, Represents the weight parameter The gradient of Dynamic weight update: based on the importance evaluation of each layer and feature , dynamically adjust the weight parameters , the weight update is performed as follows: ,in, is the update step size, is a sign function evaluated according to importance, used to indicate whether the weight increases or decreases; Evaluation module: Use the advertising effect evaluation model to evaluate the advertising effect with the enhanced scarce data as input, and decide whether to update the scarce data. The advertising effect evaluation model adopts a gradient boosting decision tree, including receiving the scarce data processed by the data enhancement module, using the gradient boosting decision tree to build an advertising effect prediction model, and predicting the potential effect of the advertisement based on the enhanced data, including click-through rate, conversion rate or user engagement, as follows: Initialization: Use the most frequent class in the data as the initial prediction value; Iterative training: For each iteration : Calculate the residual of the current model, that is, the difference between the true value and the current predicted value; train a decision tree based on the current residual to try to predict the residual; add the decision tree to the model and update the model's predicted value; Model update: The model's prediction value is obtained by weighting the prediction values ​​of all decision trees, and the contribution of each tree is calculated through the learning rate To scale; For Iteration, The contribution of a tree is expressed as: ,in, is the prediction value of the previous model, It is The predicted value of the residual by the tree, is the learning rate, which controls the contribution of each tree to the final prediction; Based on the advertising effect prediction results output by the model, combined with the preset effect threshold or improvement index, it is determined whether the advertising effect under the current enhanced data meets expectations. If the effect does not meet the standard or there is room for improvement, the data update module is triggered to update and optimize the relevant scarce data in the feature library. Otherwise, the existing data status is maintained. Data update module: When it is determined that an update is needed, the relevant data items in the feature library are automatically adjusted to ensure the timeliness and accuracy of the advertising feature library data.

2. The advertisement production feature library data intelligent update system according to claim 1, characterized in that: The domain adaptation layer specifically includes: Feature extraction: The data of both the source domain and the target domain are extracted through the initial layer of the deep neural network. is the feature extraction function, for the source domain data and target domain data The extracted features are and ; The domain adaptation layer adjusts the feature distribution by minimizing the inter-domain difference metric to make the feature distribution of the source domain and the target domain close, based on the maximum mean difference (MMD), calculated as: ； in, and are the number of samples in the source domain and the target domain, respectively. is the mapping function to the reproducing kernel Hilbert space, and are samples from the source domain and the target domain respectively; Adaptive layer training: During the training process, in addition to the traditional task-related loss, a domain adaptation loss term, namely MMD loss, is added to encourage the model to minimize the feature distribution difference between the source domain and the target domain. The total loss function is expressed as: ,in, is the task-related loss, is a balance coefficient used to adjust the importance between task loss and domain adaptation loss.

3. The advertisement production feature library data intelligent update system according to claim 1, characterized in that: The multi-task learning combination specifically includes: Bottom-layer shared representation: In the multi-task learning framework, the bottom-layer network layer is designed as a shared layer to extract common features from the data of all tasks. Related tasks, Represents the feature extraction function of the shared layer. For any task Input data , sharing is represented by ; Task-specific output layer: For each task , each has a corresponding output layer , used to predict the corresponding task based on the shared features, task The predicted output is expressed as ; Loss function of multi-task learning: In the multi-task learning framework, the total loss function is the weighted sum of all task loss functions. For each task , each has a loss function corresponding to a task , used to evaluate the prediction performance of this task, so the total loss function is expressed as: ,in, It's a task The weights are used to balance the importance of different tasks in the total loss. It's a task The real label.

4. The advertisement production feature library data intelligent update system according to claim 1, characterized in that: The meta-learning strategy specifically includes: Initialization: Choose an initial parameter ; Task sampling: From the task distribution Randomly sample a batch of tasks ; Adaptive update for each task: For each task , using the initial parameters and tasks The training data set is updated by gradient descent to obtain the adapted parameters , the adapted parameters are calculated as follows: ,in, is the learning rate, It's a task The loss function of Meta-update: Use the adapted parameters of all sampled tasks Calculate the loss on the respective validation datasets and adjust the initial parameters Perform a meta-update, which is calculated as follows: ,in, is the meta-learning rate.

Images