Data processing method and device based on meta-gain model, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to financial or medical related business scenarios, and discloses a data processing method and device based on a meta-gain model, equipment and a storage medium, and the method comprises the steps: obtaining a historical task data set and new task cold start data; performing sampling processing on the historical task data set to obtain a sampling task; performing parameter proxy on the base model to obtain meta-parameters; in combination with a sampling task, an inner circulation single-step updating mechanism and an outer circulation gradient calculation method, updating the basic model by using the updated element parameters to obtain an element gain model; and performing data processing on the cold start data of the new task by utilizing the element gain model. According to the embodiment of the invention, on the basis of a model-independent meta-learning architecture, the meta-parameters are updated by utilizing an internal circulation single-step updating mechanism and an external circulation gradient calculation method, so that the model can quickly adapt to a new task, the cold start data of the new task can be accurately processed, and the data processing effect is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to data processing methods, apparatus, devices, and storage media based on meta-gain models. Background Technology

[0002] Uplift models are machine learning methods based on causal inference, used to assess the incremental effects of interventions on individuals. While some research and products exist in the fields of Uplift modeling and meta-learning, none have yet formed a complete technical solution to the cold-start problem. From an academic perspective, the MAML (Model-Agnostic Meta-Learning) approach pioneered the application of meta-learning in few-shot learning. This method utilizes a two-layer optimization framework of inner and outer loops, allowing the model to quickly adapt to new tasks. However, MAML was initially designed for supervised learning tasks such as image classification and was not optimized for the specificities of Uplift modeling. In recent years, some research has attempted to apply meta-learning ideas to causal inference, such as exploring the application of deep learning in counterfactual prediction. However, these studies mainly focus on performance improvements within a single task, lacking systematic research on cross-task knowledge transfer. Regarding Uplift models, while existing response models have become industry standards, they rely on large amounts of historical data, making them difficult to handle cold-start scenarios such as new markets and new products. Similar limitations exist in the medical and financial fields.

[0003] Existing technologies have several shortcomings and problems. Regarding samples, when launching new markets, products, or activities, only 100-500 samples are typically collected, while traditional Uplift models (such as causal forests and dual-model methods) require at least 5000 samples to obtain stable estimates. Under small sample conditions, these models are prone to overfitting, misinterpreting the effects of confounding variables as treatment effects. For example, in the financial sector, during credit card "double points" promotions, high-value customers are more likely to be selected and have higher retention rates. If confounding variables are not adequately controlled, traditional models may overestimate the effectiveness of the promotion, leading to a misallocation of marketing resources. Similarly, in the medical field, when a new targeted drug undergoes small-scale clinical trials, data from only a few hundred patients may be collected. Traditional Uplift models, when assessing the drug's effect on patients' conditions, may misinterpret the effects of confounding variables such as the patient's own recovery ability as the drug's therapeutic effect, leading to incorrect assessments of the drug's efficacy and impacting subsequent clinical application and promotion.

[0004] In terms of cross-scenario transfer, existing technologies lack the ability to transfer knowledge across different scenarios. Each new task (such as a new market promotion) requires training the model from scratch, making it impossible to utilize knowledge from previous tasks (such as other marketing activities). This not only prolongs the model development cycle but also leads to a sharp drop in performance when feature distribution shifts (such as different user behavior patterns in new regions). For example, in the financial sector, when a bank wants to launch a microloan business in a new city, the traditional Uplift model cannot utilize loan business data from other cities for knowledge transfer because the consumption habits and credit concepts of users in that city differ from those in other cities. It can only collect data again to train the model, which is not only time-consuming but also results in a significant drop in model performance when faced with user behavior patterns unique to the new city. Another example is in the medical field, when promoting a disease treatment plan that has proven effective in hospitals in large cities to hospitals in remote areas, the traditional Uplift model cannot draw on the treatment effect data from hospitals in large cities because the living environment, dietary habits, and other characteristics of patients in the two locations differ. It needs to retrain the model, leading to a longer promotion cycle for the treatment plan and potentially affecting treatment effectiveness due to model performance issues.

[0005] In terms of data processing, existing technologies lack targeted optimization for data preprocessing and feature engineering. Since Uplift modeling involves differences between treatment and control groups, and identifying the true treatment effect, traditional data processing methods often fail to consider these specificities, potentially retaining a large amount of irrelevant features or noisy data, affecting the model's accurate estimation of the treatment effect. Furthermore, when faced with Uplift modeling data from different sources and in different formats, existing technologies lack efficient integration and cleaning mechanisms, resulting in low data quality and further exacerbating performance issues in small sample sizes and cross-scenario transfer. For example, in the financial field, Uplift modeling for personal credit risk assessment involves data from various sources and formats, such as bank statements, credit reports, and social media data. Traditional data processing methods struggle to efficiently integrate this data, potentially retaining social noise unrelated to credit risk, affecting the model's accurate judgment of the treatment effect of credit activities. For example, in the medical field, when evaluating the uplift effect of a new rehabilitation training program, patient medical records, rehabilitation training records, daily dietary data, etc. are collected. These data come from various sources and have different formats. Traditional data processing methods cannot clean and integrate them in a targeted manner, and may retain some dietary details and noise that are irrelevant to the rehabilitation effect, causing the model to be unable to accurately estimate the actual effect of the rehabilitation training program. Summary of the Invention

[0006] This invention provides a data processing method, apparatus, computer device, and storage medium based on a meta-gain model, aiming to solve the cold start problem in the field of Uplift modeling and meta-learning in the prior art and improve the data processing effect.

[0007] In a first aspect, embodiments of the present invention provide a data processing method based on a meta-gain model, comprising:

[0008] Obtain historical task datasets and cold start data for new tasks;

[0009] The historical task dataset is sampled to obtain the corresponding sampled tasks;

[0010] The base model for data processing is parametrically proxied to obtain meta-parameters;

[0011] The meta-parameters are updated by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and the updated meta-parameters are used to update the base model to obtain the meta-gain model.

[0012] The meta-gain model is used to process the cold start data of the new task.

[0013] Secondly, embodiments of the present invention provide a data processing apparatus based on a meta-gain model, comprising:

[0014] The data acquisition unit is used to acquire historical task datasets and cold start data for new tasks;

[0015] The task sampling unit is used to sample the historical task dataset to obtain the corresponding sampling task;

[0016] The parameter proxy unit is used to perform parameter proxying on the base model of data processing to obtain meta-parameters;

[0017] The model update unit is used to update the meta-parameters by combining the sampling task, the inner loop single-step update mechanism and the outer loop gradient calculation method, and to update the base model using the updated meta-parameters to obtain the meta-gain model.

[0018] The data processing unit is used to process the cold start data of the new task using the meta-gain model.

[0019] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the data processing method based on the meta-gain model as described in the first aspect.

[0020] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the data processing method based on the meta-gain model as described in the first aspect.

[0021] This invention provides a data processing method, apparatus, computer device, and storage medium based on a meta-gain model. The method includes: acquiring historical task datasets and new task cold-start data; sampling the historical task dataset to obtain corresponding sampling tasks; performing parameter proxying on the base model of the data processing to obtain meta-parameters; updating the meta-parameters by combining the sampling tasks, an inner loop single-step update mechanism, and an outer loop gradient calculation method, and updating the base model using the updated meta-parameters to obtain a meta-gain model; and processing the new task cold-start data using the meta-gain model. This invention effectively solves the cold-start problem in Uplift modeling and meta-learning by utilizing the meta-gain model, improving data processing performance. During the cold-start phase, due to the limited amount of data for new tasks, traditional Uplift models struggle to obtain stable estimation results. However, this invention, by combining knowledge from historical task datasets and using an inner loop single-step update mechanism and an outer loop gradient calculation method to update the meta-parameters, enables the model to quickly adapt to new tasks and achieve accurate processing of new task cold-start data. This approach not only improves the efficiency of data processing but also shortens the model development cycle and reduces reliance on large amounts of historical data, enabling the Uplift model to be better applied to the promotion of new markets, new products, or new activities. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a data processing method based on a meta-gain model provided in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of a sub-process of a data processing method based on a meta-gain model provided in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of another sub-process of a data processing method based on a meta-gain model provided in an embodiment of the present invention;

[0026] Figure 4 A schematic block diagram of a data processing device based on a meta-gain model provided in an embodiment of the present invention;

[0027] Figure 5A schematic block diagram of a data processing device based on a meta-gain model provided in an embodiment of the present invention;

[0028] Figure 6 Another schematic block diagram of a data processing device based on a meta-gain model provided in an embodiment of the present invention;

[0029] Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0032] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0033] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0034] Please see below. Figure 1 This invention provides a data processing method based on a meta-gain model, specifically including steps S101 to S105.

[0035] Step S101: Obtain historical task datasets and new task cold start data;

[0036] Step S102: Perform sampling processing on the historical task dataset to obtain the corresponding sampling tasks;

[0037] Step S103: Perform parameter proxying on the base model of data processing to obtain meta-parameters;

[0038] Step S104: Combine the sampling task, the inner loop single-step update mechanism and the outer loop gradient calculation method to update the meta-parameters, and use the updated meta-parameters to update the base model to obtain the meta-gain model;

[0039] Step S105: Use the meta-gain model to process the cold start data of the new task.

[0040] In this embodiment, the historical task dataset and the cold start data of the new task are first obtained; then the historical task dataset is sampled to obtain the sampled task; then the data processing base model parameters are proxied to obtain the meta-parameters; then the sampled task, the inner loop single-step update mechanism and the outer loop gradient calculation method are combined to update the meta-parameters, and the base model is updated accordingly to obtain the meta-gain model; finally, the meta-gain model is used to process the cold start data of the new task.

[0041] This embodiment effectively addresses the cold-start problem in Uplift modeling and meta-learning by utilizing a meta-gain model, thereby improving data processing performance. During the cold-start phase, traditional Uplift models struggle to obtain stable estimation results due to the limited amount of data available for new tasks. This embodiment, however, combines knowledge from historical task datasets with an inner-loop single-step update mechanism and an outer-loop gradient calculation method to update meta-parameters, enabling the model to quickly adapt to new tasks and accurately process cold-start data. This approach not only improves data processing efficiency but also shortens the model development cycle, reduces reliance on large amounts of historical data, and allows Uplift models to be better applied to the promotion of new markets, new products, or new activities.

[0042] The data processing method based on the meta-gain model provided in this embodiment is applicable to a wide range of scenarios that require accurate assessment of the causal effects of interventions and face the challenge of data scarcity. For example, in the financial industry, it is suitable for scenarios such as credit card marketing, loan product design, and customer churn prevention, solving the cold start problem when expanding into new markets or launching new products, while meeting the stringent requirements of the financial industry for model interpretability. In the e-commerce and retail industry, it is suitable for new product promotion, personalized recommendations, and promotional strategy optimization, helping e-commerce platforms quickly identify target customer groups and achieve precise marketing in the early stages when data is scarce. In the digital advertising and media industry, it is suitable for programmatic advertising optimization and content recommendation systems, solving the data scarcity problem when new advertisers or new advertising formats are launched, improving the accuracy and ROI of advertising. In the healthcare industry, it is suitable for personalized treatment plan recommendations and health intervention effect evaluation, helping medical institutions quickly determine the optimal target population using limited patient data in the early stages of implementing new therapies or interventions.

[0043] Based on the aforementioned application scenarios, the historical task datasets and new task cold start data mentioned in this embodiment can specifically include historical marketing campaign data, user behavior logs, historical advertising records, etc. These data contain samples from the processing and control groups, as well as the corresponding response results. New task cold start data can be user behavior data from new market promotions, patient data from the initial stages of implementing new therapies, etc. While these data are relatively small in volume, they contain crucial information used to train and optimize the meta-gain model, enabling accurate evaluation and processing of new tasks. By utilizing these historical task datasets and new task cold start data, the data processing method based on the meta-gain model provided in this embodiment can achieve rapid adaptation and accurate processing of new tasks, providing strong support for applications in various fields.

[0044] In one embodiment, step S101 includes:

[0045] The historical task dataset and the new task cold start data are obtained according to the following formulas:

[0046] D hist ={(T1,D1),(T2,D2),...,(T N D N )};

[0047] D new =(X new ,T new ,Y new M new );

[0048] Among them, D hist Let T1 represent the first historical task, D1 represent the dataset corresponding to the first historical task, T2 represent the dataset corresponding to the second historical task, and D1 represent the dataset corresponding to the third historical task. N Indicates the first historical mission, D N Let D represent the dataset corresponding to the first historical task, and N represent the total number of historical tasks; new This represents the cold start data for a new task, X. new T represents the feature matrix of the new task. new Indicating a new task, Y new M represents the result vector of the new task. new This represents the metadata corresponding to the new task.

[0049] In this embodiment, the historical task dataset is D. hist ={(T1,D1),(T2,D2),...,(T N D N )}, where T i(i = 1, 2, ..., N) represents a historical task, and each historical task T i Specifically, it can include a feature matrix. Processing vectors Result vector and task metadata M i (Such as market, channel, product type, etc.). The characteristic of historical task datasets is that there are distributional differences between tasks, but they share underlying causal structures.

[0050] The new task cold start data is D. new =(X new ,T new ,Y new M new ), where X new T represents the feature matrix of the new task. new Indicates the specific type or definition of the new task, Y new This represents the expected predicted outcome vector for the new task, such as user response or purchasing behavior, while M... new This includes additional metadata about the new task, such as task description and target group characteristics. Together, this data constitutes the complete information set needed to evaluate the uplift effect of the new task. In practical applications, the size of the cold start data for a new task can be n. new The sample size is ∈[100,500], which is much smaller than the sample size required by the traditional Uplift model. The characteristics of cold start data for new tasks are that it may contain distribution shifts, such as different user behavior patterns in new markets.

[0051] In one embodiment, step S102 includes:

[0052] The historical task dataset is randomly sampled in batches based on a finite strategy for difficult tasks to obtain the sampled tasks.

[0053] In this embodiment, a finite strategy for difficult tasks can be adopted when performing task sampling. Randomly sample task batch B = {T} from Dhist i1 ,...T iB}. Among them, AUUC(T i ) represents historical task T i The area under the uplift curve (AUUC) is used to measure the performance of the uplift model on historical tasks. The difficult task finite strategy means that during sampling, higher sampling probabilities are given to historical tasks with lower AUUC values ​​(i.e., relatively more difficult tasks), in order to improve the generalization ability and adaptability of the meta-gain model through these more challenging tasks. The sampling task batch B obtained in this way is B = {T}.i1 ,...T iB This not only includes diverse historical mission information, but also pays more attention to those difficult-to-handle scenarios, which helps in the optimization of meta-parameters and the construction of meta-gain models in subsequent steps.

[0054] In one embodiment, such as Figure 2 As shown, step S103 includes steps S201 to S202.

[0055] Step S201: Obtain the hyperparameters or initial parameters of the base model;

[0056] Step S202: Use the hyperparameters or initial parameters as the meta-parameters.

[0057] In this embodiment, when acquiring meta-parameters, the hyperparameters or initial state of the base model can be used as learnable meta-parameters θ. For example, when the base model is an XGBoost model, the meta-parameter θ = {initial base_margin value, feature weights}; as another example, when the base model is a causal tree model, the meta-parameter θ = {feature splitting priority, leaf node uplift prior}. These meta-parameters θ serve as key information for model initialization or tuning, and will be optimized in subsequent steps through a specific update mechanism to adapt to the needs of different historical and new tasks. By proxying the parameters of the base model into meta-parameters, this embodiment provides a solid foundation for the rapid construction and optimization of the meta-gain model. This parameter proxying method not only simplifies the complexity of model training but also enhances the flexibility and adaptability of the model, enabling the meta-gain model to better capture the potential causal structure between different tasks, thereby improving the overall performance of the model.

[0058] In one embodiment, such as Figure 3 As shown, step S104 includes steps S301 to S302.

[0059] Step S301: Update the meta-parameters using the inner loop single-step update mechanism according to the following formula:

[0060]

[0061] Where, θ′ i Here, θ represents the updated meta-parameters, and α represents the inner loop learning rate. L represents the operator. Ti This represents the loss of the i-th historical task;

[0062] Step S302: Estimate the gradient of the meta-parameters using the outer loop gradient calculation method according to the following formula:

[0063]

[0064] in, L represents the result of gradient estimation. val e represents the validation loss. j Let represent the unit vector of the j-th dimension, and ε be a small perturbation.

[0065] In this embodiment, when updating the meta-parameters using the inner loop single-step update mechanism, specifically for each task T... i ∈B, starting from the global parameters, perform a single-step gradient update: This embodiment uses an inner loop single-step update mechanism to perform proxy updates on non-differentiable models (such as XGBoost models) and obtain updated task-specific parameters.

[0066] In a specific embodiment, a lightweight single-step inner loop optimization mechanism is used for parameter updates. Specifically, considering the scarcity of samples in cold-start scenarios, this embodiment employs a single-step inner loop design (K=1), using meta-training to make the initial parameters the "optimal learning starting point," ensuring a significant performance improvement with a single update. The optimized meta-objective function is:

[0067] Unlike traditional multi-step updates, in this embodiment, a single-step update can obtain a near-optimal θ when the meta-parameter θ satisfies the following conditions. i ′:

[0068]

[0069] Among them, H i For task T i The Hessian matrix means that θ lies in the "flat region of the valley of loss for multiple tasks", and the gradient direction points directly to the optimal solution.

[0070] Next, the loss L is verified using the outer loop. val The gradient with respect to θ is used as the optimization objective, and the gradient estimate is calculated. This is used to update the global meta-parameter θ. The gradient estimation here employs a finite difference method based on small perturbations. Specifically, a small perturbation ε is added to each meta-parameter dimension j, and then the change in the validation loss Lval is calculated to approximate the gradient of that dimension. This approach effectively handles non-differentiable models, making the update of meta-parameters more accurate and efficient.

[0071] By combining the inner loop single-step update mechanism and the outer loop gradient calculation method, this embodiment achieves dual optimization of the meta-parameters, which not only improves the generalization ability of the meta-gain model, but also enhances its adaptability to new tasks.

[0072] In a specific embodiment, the verification loss can be calculated according to the following formula:

[0073]

[0074] Here, QiniLoss represents the loss on the initial task set, used to evaluate the model's performance on unseen initial tasks. The validation loss L... val As the target of the outer loop optimization, it can guide the update direction of the meta-parameters, making them more adaptable to new, unseen tasks. By minimizing the validation loss, the meta-gain model in this embodiment can improve its adaptability to new tasks while maintaining good generalization ability to historical tasks.

[0075] In addition, after estimating the gradient of the meta-parameters using the outer loop gradient calculation method, gradient aggregation can be performed according to the following formula:

[0076]

[0077] Among them, g meta This represents the result of gradient aggregation. By weighted aggregation of gradients across different dimensions, the final gradient update direction is obtained, which is used to update the global meta-parameter θ. This gradient aggregation method comprehensively considers the importance of gradients in different dimensions, making the update of meta-parameters more comprehensive and accurate. In practical applications, aggregation coefficients can be set based on the relevance or similarity of historical tasks to the new task. The higher the relevance or similarity of historical tasks, the greater the weight assigned to the gradient of the corresponding dimension during aggregation, thereby ensuring that the update of meta-parameters can draw more on information from historical tasks that are closer to the new task. In this way, this embodiment can further improve the adaptability of the meta-gain model to new tasks.

[0078] In one embodiment, step S104 further includes:

[0079] The base model is globally updated using the updated meta-parameters, and a task diversity penalty is added to the base model based on a regularization mechanism.

[0080] The loss function is constructed according to the following formula, and the loss function is used to optimize the loss basis model to obtain the meta-gain model:

[0081]

[0082] Among them, L total Let B represent the loss function, B represent the sampling task, i represent the i-th sampling task, and θ′ represent the loss function. i Let L represent the updated meta-parameters, λ represent the regularization strength parameter, and L represent the regularization strength parameter. val represents the validation loss, and Var represents the diversity task.

[0083] After optimizing the meta-parameters using both the inner loop single-step update mechanism and the outer loop gradient calculation method, this embodiment further introduces a regularization mechanism to enhance the model's generalization ability. Specifically, during the global parameter update of the base model using the updated meta-parameters, not only is the minimization of the validation loss considered, but a task diversity penalty is also added to the base model based on the regularization mechanism. This diversity penalty aims to encourage the model to learn the potential differences between different historical tasks, thereby improving its adaptability to new tasks. The global parameter update is specifically: θ←θ-βg meta , where β is the learning rate.

[0084] Constructed loss function L total The model comprehensively considers the loss, regularization term, and diversity task penalty for each task in the sampled task set B. The regularization strength parameter λ is used to balance the weights between the validation loss and the task diversity penalty, ensuring that the model neither overfits to specific historical tasks during optimization nor fails to effectively capture the diversity between tasks. This is achieved by adjusting the loss function L... total Through optimization, this embodiment yields a more robust and adaptable meta-gain model. Furthermore, during the optimization process, appropriate optimization algorithms (such as stochastic gradient descent, Adam, etc.) can be used to iteratively update the loss function until convergence or a preset number of iterations is reached. Through this process, the parameters of the meta-gain model will gradually be adjusted to their optimal state, enabling it to perform well on cold-start data for new tasks.

[0085] Therefore, this embodiment successfully achieves efficient and accurate processing of cold-start data for new tasks by comprehensively integrating the meta-gain model, the inner loop single-step update mechanism, the outer loop gradient calculation method, and the regularization mechanism. This embodiment not only significantly improves the effectiveness and efficiency of data processing but also greatly reduces the dependence on large amounts of historical data, thus providing solid and powerful support for practical applications in various fields. Furthermore, this embodiment is compatible with and supports traditional models such as XGBoost as base learners. With the help of the surrogate gradient mechanism, even models that are not originally differentiable can participate in the meta-learning process. After the meta-training phase is completed, the meta-parameter θ is optimized and evolved, ultimately forming the "optimal learning starting point." This starting point exhibits extremely high sensitivity and adaptability to gradient updates, further improving the model's training effect and generalization ability.

[0086] In one embodiment, step S105 includes:

[0087] Load the meta-gain model and perform a single-step update on the cold start data of the new task to obtain the corresponding update parameters;

[0088] Gain prediction is performed on the cold start data of the new task based on the updated parameters.

[0089] In this embodiment, when processing cold-start data for a new task based on the meta-gain model, the optimized and trained meta-gain model is first loaded. This model already possesses the latent causal structure learned from historical tasks and the ability to adapt to new tasks. Subsequently, a single-step update operation is performed on the cold-start data Tnew for the new task. Utilizing the meta-gain model's "optimal learning starting point" and surrogate gradient mechanism, the specific parameters for the new task, namely θ, are obtained quickly and accurately. new =θ-α▽ θ L Tnew (θ). This process fully utilizes the generalization and adaptability of the meta-gain model, enabling high-quality parameter updates even when new task data is scarce and distribution bias may exist. It also avoids the computational overhead and overfitting risk of multi-step updates, significantly improving training speed.

[0090] Based on the updated parameters, this embodiment further performs gain prediction on the cold start data of new tasks. Gain prediction aims to evaluate the response probability or effectiveness improvement of different users or samples under specific marketing strategies or processes, which is of great significance for application scenarios such as precision marketing and personalized recommendations. Through the gain prediction function of the meta-gain model, enterprises can more effectively identify high-value user groups and formulate more targeted marketing strategies, thereby improving overall business performance and user experience.

[0091] In a specific embodiment, this embodiment is based on meta-parameter update training, which can obtain an uplift model adapted to the new task, namely M. new =MetaUplift(D hist D new ,Λ), where D hist Let D represent the historical task dataset, Dnew represent the cold start data for new tasks, and Λ represent the set of business constraint parameters, Λ∈{α, β, ε, λ}. Simultaneously, through gain prediction, we can ultimately obtain the customer UpliftScore prediction, which is the potential gain value for each customer under a specific treatment. This value reflects the probability of a customer's response to a marketing strategy or treatment, or the magnitude of the improvement in effectiveness. Enterprises can use customer UpliftScores for targeted marketing or personalized recommendations, concentrating resources on user groups most likely to respond or improve effectiveness, thereby improving overall business performance and user experience.

[0092] In summary, the data processing method based on the meta-gain model provided in this embodiment has the following characteristics:

[0093] (1) Model-independent meta-learning architecture: The meta-learning framework is innovatively extended to non-deep learning models, enabling traditional Uplift models with strong interpretability, such as XGBoost and causal trees, to also have meta-learning capabilities. By using a surrogate gradient mechanism, the gradient calculation problem of non-differentiable models is solved, realizing the organic combination of model interpretability and meta-learning capabilities required by highly regulated industries.

[0094] (2) Lightweight single-step inner loop optimization: A single-step inner loop design (K=1) is adopted, and the initial parameters are made into the "optimal learning starting point" through meta-training, ensuring that a high-performance model can be obtained with a single step update under the condition of a small number of samples. This design significantly reduces the computational cost and effectively avoids the risk of overfitting under the condition of multi-step updates with small samples.

[0095] (3) Task performance equalization regularization: A loss variance regularization term is introduced to solve the "negative transfer" problem in meta-learning. This mechanism automatically balances the optimization intensity of each task to ensure that the learned meta-knowledge has a stronger generalization ability, thereby improving the transfer effect of the model between different tasks.

[0096] Figure 4 This is a schematic block diagram of a data processing device 400 based on a meta-gain model provided in this embodiment. The device 400 includes:

[0097] Data acquisition unit 401 is used to acquire historical task datasets and new task cold start data;

[0098] The task sampling unit 402 is used to sample the historical task dataset to obtain the corresponding sampling task.

[0099] The parameter proxy unit 403 is used to perform parameter proxying on the base model of data processing to obtain meta-parameters;

[0100] The model update unit 404 is used to update the meta-parameters by combining the sampling task, the inner loop single-step update mechanism and the outer loop gradient calculation method, and to update the base model using the updated meta-parameters to obtain the meta-gain model.

[0101] The data processing unit 405 is used to process the cold start data of the new task using the meta-gain model.

[0102] In this embodiment, the historical task dataset and the cold start data of the new task are first obtained; then the historical task dataset is sampled to obtain the sampled task; then the data processing base model parameters are proxied to obtain the meta-parameters; then the sampled task, the inner loop single-step update mechanism and the outer loop gradient calculation method are combined to update the meta-parameters, and the base model is updated accordingly to obtain the meta-gain model; finally, the meta-gain model is used to process the cold start data of the new task.

[0103] This embodiment effectively addresses the cold-start problem in Uplift modeling and meta-learning by utilizing a meta-gain model, thereby improving data processing performance. During the cold-start phase, traditional Uplift models struggle to obtain stable estimation results due to the limited amount of data available for new tasks. This embodiment, however, combines knowledge from historical task datasets with an inner-loop single-step update mechanism and an outer-loop gradient calculation method to update meta-parameters, enabling the model to quickly adapt to new tasks and accurately process cold-start data. This approach not only improves data processing efficiency but also shortens the model development cycle, reduces reliance on large amounts of historical data, and allows Uplift models to be better applied to the promotion of new markets, new products, or new activities.

[0104] The data processing device based on the meta-gain model provided in this embodiment is applicable to a wide range of scenarios that require accurate assessment of the causal effects of interventions and face the challenge of data scarcity. For example, in the financial industry, it is suitable for scenarios such as credit card marketing, loan product design, and customer churn prevention, solving the cold start problem when expanding into new markets or launching new products, while meeting the stringent requirements of the financial industry for model interpretability. In the e-commerce and retail industry, it is suitable for new product promotion, personalized recommendations, and promotional strategy optimization, helping e-commerce platforms quickly identify target customer groups and achieve precise marketing in the early stages when data is scarce. In the digital advertising and media industry, it is suitable for programmatic advertising optimization and content recommendation systems, solving the data scarcity problem when new advertisers or new advertising formats are launched, improving the accuracy and ROI of advertising. In the healthcare industry, it is suitable for personalized treatment plan recommendations and health intervention effect evaluation, helping medical institutions quickly determine the optimal target population using limited patient data in the early stages of implementing new therapies or interventions.

[0105] Based on the aforementioned application scenarios, the historical task datasets and new task cold start data mentioned in this embodiment can specifically include historical marketing campaign data, user behavior logs, historical advertising records, etc. These data contain samples from the processing and control groups, as well as the corresponding response results. New task cold start data can be user behavior data from new market promotions, patient data from the initial stages of implementing new therapies, etc. This data is relatively small in volume but contains key information used to train and optimize the meta-gain model, enabling accurate evaluation and processing of new tasks. By utilizing these historical task datasets and new task cold start data, the data processing device based on the meta-gain model provided in this embodiment can achieve rapid adaptation and accurate processing of new tasks, providing strong support for applications in various fields.

[0106] In one embodiment, the data acquisition unit 401 includes:

[0107] The task acquisition unit is used to acquire the historical task dataset and the new task cold start data according to the following formulas:

[0108] D hist ={(T1,D1),(T2,D2),...,(T N D N )};

[0109] D new =(X new ,T new ,Y new M new );

[0110] Among them, D hist Let T1 represent the first historical task, D1 represent the dataset corresponding to the first historical task, T2 represent the dataset corresponding to the second historical task, and D1 represent the dataset corresponding to the third historical task. N Indicates the first historical mission, D N Let D represent the dataset corresponding to the first historical task, and N represent the total number of historical tasks; new This represents the cold start data for a new task, X. new T represents the feature matrix of the new task. new Indicating a new task, Y new M represents the result vector of the new task. new This represents the metadata corresponding to the new task.

[0111] In this embodiment, the historical task dataset is D. hist ={(T1,D1),(T2,D2),...,(T N D N )}, where T i (i = 1, 2, ..., N) represents a historical task, and each historical task T i Specifically, it can include a feature matrix. Processing vectors Result vector and task metadata M i (Such as market, channel, product type, etc.). The characteristic of historical task datasets is that there are distributional differences between tasks, but they share underlying causal structures.

[0112] The new task cold start data is D. new =(X new ,T new ,Y new M new ), where X new T represents the feature matrix of the new task. new Indicates the specific type or definition of the new task, Ynew This represents the expected predicted outcome vector for the new task, such as user response or purchasing behavior, while M... new This includes additional metadata about the new task, such as task description and target group characteristics. Together, this data constitutes the complete information set needed to evaluate the uplift effect of the new task. In practical applications, the size of the cold start data for a new task can be n. new The sample size is ∈[100,500], which is much smaller than the sample size required by the traditional Uplift model. The characteristics of cold start data for new tasks are that it may contain distribution shifts, such as different user behavior patterns in new markets.

[0113] In one embodiment, the task sampling unit 402 includes:

[0114] The random sampling unit is used to perform random sampling of the historical task dataset based on a finite strategy for difficult tasks, thereby obtaining the sampled tasks.

[0115] In this embodiment, a finite strategy for difficult tasks can be adopted when performing task sampling. Randomly sample task batch B = {T} from Dhist i1 ,...T iB}. Among them, AUUC(T i ) represents historical task T i The area under the uplift curve (AUUC) is used to measure the performance of the uplift model on historical tasks. The difficult task finite strategy means that during sampling, higher sampling probabilities are given to historical tasks with lower AUUC values ​​(i.e., relatively more difficult tasks), in order to improve the generalization ability and adaptability of the meta-gain model through these more challenging tasks. The sampling task batch B obtained in this way is B = {T}. i1 ,...T iB This not only includes diverse historical mission information, but also pays more attention to those difficult-to-handle scenarios, which helps in the optimization of meta-parameters and the construction of meta-gain models in subsequent steps.

[0116] In one embodiment, such as Figure 5 As shown, the parameter proxy unit 403 includes:

[0117] The parameter acquisition unit 501 is used to acquire the hyperparameters or initial parameters of the base model;

[0118] The parameter setting unit 502 is used to set the hyperparameters or initial parameters as the metaparameters.

[0119] In this embodiment, when acquiring meta-parameters, the hyperparameters or initial state of the base model can be used as learnable meta-parameters θ. For example, when the base model is an XGBoost model, the meta-parameter θ = {initial base_margin value, feature weights}; as another example, when the base model is a causal tree model, the meta-parameter θ = {feature splitting priority, leaf node uplift prior}. These meta-parameters θ serve as key information for model initialization or tuning, and will be optimized in subsequent steps through a specific update mechanism to adapt to the needs of different historical and new tasks. By proxying the parameters of the base model into meta-parameters, this embodiment provides a solid foundation for the rapid construction and optimization of the meta-gain model. This parameter proxying method not only simplifies the complexity of model training but also enhances the flexibility and adaptability of the model, enabling the meta-gain model to better capture the potential causal structure between different tasks, thereby improving the overall performance of the model.

[0120] In one embodiment, such as Figure 6 As shown, the model update unit 404 includes:

[0121] The inner loop unit 601 is used to update the meta-parameters according to the following formula via an inner loop single-step update mechanism:

[0122]

[0123] Where, θ′ i Here, θ represents the updated meta-parameters, and α represents the inner loop learning rate. L represents the operator. Ti This represents the loss of the i-th historical task;

[0124] The outer loop unit 602 is used to estimate the gradient of the meta-parameters according to the following formula using the outer loop gradient calculation method:

[0125]

[0126] in, Let Lval represent the result of gradient estimation, ej represent the unit vector of the j-th dimension, and ε represent the small perturbation.

[0127] In this embodiment, when updating the meta-parameters using the inner loop single-step update mechanism, specifically for each task T... i ∈B, starting from the global parameters, perform a single-step gradient update: This embodiment uses an inner loop single-step update mechanism to perform proxy updates on non-differentiable models (such as XGBoost models) and obtain updated task-specific parameters.

[0128] In a specific embodiment, a lightweight single-step inner loop optimization mechanism is used for parameter updates. Specifically, considering the scarcity of samples in cold-start scenarios, this embodiment employs a single-step inner loop design (K=1), using meta-training to make the initial parameters the "optimal learning starting point," ensuring a significant performance improvement with a single update. The optimized meta-objective function is:

[0129] Unlike traditional multi-step updates, in this embodiment, a single-step update can obtain a near-optimal θ when the meta-parameter θ satisfies the following conditions. i ′:

[0130]

[0131] Among them, H i For task T i The Hessian matrix means that θ lies in the "flat region of the valley of loss for multiple tasks", and the gradient direction points directly to the optimal solution.

[0132] Next, the loss L is verified using the outer loop. val The gradient with respect to θ is used as the optimization objective, and the gradient estimate is calculated. This is used to update the global meta-parameter θ. The gradient estimation here employs a finite difference method based on small perturbations. Specifically, a small perturbation ε is added to each meta-parameter dimension j, and then the change in the validation loss Lval is calculated to approximate the gradient of that dimension. This approach effectively handles non-differentiable models, making the update of meta-parameters more accurate and efficient.

[0133] By combining the inner loop single-step update mechanism and the outer loop gradient calculation method, this embodiment achieves dual optimization of the meta-parameters, which not only improves the generalization ability of the meta-gain model, but also enhances its adaptability to new tasks.

[0134] In a specific embodiment, the verification loss can be calculated according to the following formula:

[0135]

[0136] Here, QiniLoss represents the loss on the initial task set, used to evaluate the model's performance on unseen initial tasks. The validation loss L... val As the target of the outer loop optimization, it can guide the update direction of the meta-parameters, making them more adaptable to new, unseen tasks. By minimizing the validation loss, the meta-gain model in this embodiment can improve its adaptability to new tasks while maintaining good generalization ability to historical tasks.

[0137] In addition, after estimating the gradient of the meta-parameters using the outer loop gradient calculation method, gradient aggregation can be performed according to the following formula:

[0138]

[0139] Among them, g meta This represents the result of gradient aggregation. By weighted aggregation of gradients across different dimensions, the final gradient update direction is obtained, which is used to update the global meta-parameter θ. This gradient aggregation method comprehensively considers the importance of gradients in different dimensions, making the update of meta-parameters more comprehensive and accurate. In practical applications, aggregation coefficients can be set based on the relevance or similarity of historical tasks to the new task. The higher the relevance or similarity of historical tasks, the greater the weight assigned to the gradient of the corresponding dimension during aggregation, thereby ensuring that the update of meta-parameters can draw more on information from historical tasks that are closer to the new task. In this way, this embodiment can further improve the adaptability of the meta-gain model to new tasks.

[0140] In one embodiment, the model update unit 404 further includes:

[0141] The parameter update unit is used to perform global parameter updates on the base model using the updated meta-parameters, and to add task diversity penalties to the base model based on a regularization mechanism.

[0142] The model optimization unit is used to construct a loss function according to the following formula, and to optimize the loss basis model using the loss function to obtain the meta-gain model:

[0143]

[0144] Among them, L total Let B represent the loss function, B represent the sampling task, i represent the i-th sampling task, and θ′ represent the loss function. i Let L represent the updated meta-parameters, λ represent the regularization strength parameter, and L represent the regularization strength parameter. val represents the validation loss, and Var represents the diversity task.

[0145] After optimizing the meta-parameters using both the inner loop single-step update mechanism and the outer loop gradient calculation method, this embodiment further introduces a regularization mechanism to enhance the model's generalization ability. Specifically, during the global parameter update of the base model using the updated meta-parameters, not only is the minimization of the validation loss considered, but a task diversity penalty is also added to the base model based on the regularization mechanism. This diversity penalty aims to encourage the model to learn the potential differences between different historical tasks, thereby improving its adaptability to new tasks. The global parameter update is specifically: θ←θ-βg meta , where β is the learning rate.

[0146] Constructed loss function L totalThe model comprehensively considers the loss, regularization term, and diversity task penalty for each task in the sampled task set B. The regularization strength parameter λ is used to balance the weights between the validation loss and the task diversity penalty, ensuring that the model neither overfits to specific historical tasks during optimization nor fails to effectively capture the diversity between tasks. This is achieved by adjusting the loss function L... total Through optimization, this embodiment yields a more robust and adaptable meta-gain model. Furthermore, during the optimization process, appropriate optimization algorithms (such as stochastic gradient descent, Adam, etc.) can be used to iteratively update the loss function until convergence or a preset number of iterations is reached. Through this process, the parameters of the meta-gain model will gradually be adjusted to their optimal state, enabling it to perform well on cold-start data for new tasks.

[0147] Therefore, this embodiment successfully achieves efficient and accurate processing of cold-start data for new tasks by comprehensively integrating the meta-gain model, the inner loop single-step update mechanism, the outer loop gradient calculation method, and the regularization mechanism. This embodiment not only significantly improves the effectiveness and efficiency of data processing but also greatly reduces the dependence on large amounts of historical data, thus providing solid and powerful support for practical applications in various fields. Furthermore, this embodiment is compatible with and supports traditional models such as XGBoost as base learners. With the help of the surrogate gradient mechanism, even models that are not originally differentiable can participate in the meta-learning process. After the meta-training phase is completed, the meta-parameter θ is optimized and evolved, ultimately forming the "optimal learning starting point." This starting point exhibits extremely high sensitivity and adaptability to gradient updates, further improving the model's training effect and generalization ability.

[0148] In one embodiment, the data processing unit 405 includes:

[0149] A single-step update unit is used to load the meta-gain model and perform a single-step update on the cold start data of the new task to obtain the corresponding update parameters.

[0150] A gain prediction unit is used to predict the gain of the new task cold start data based on the updated parameters.

[0151] In this embodiment, when processing cold-start data for a new task based on the meta-gain model, the optimized and trained meta-gain model is first loaded. This model already possesses the latent causal structure learned from historical tasks and the ability to adapt to new tasks. Subsequently, a single-step update operation is performed on the cold-start data Tnew for the new task. Utilizing the meta-gain model's "optimal learning starting point" and surrogate gradient mechanism, the specific parameters for the new task, namely θ, are obtained quickly and accurately. new =θ-α▽ θ L Tnew(θ). This process fully utilizes the generalization and adaptability of the meta-gain model, enabling high-quality parameter updates even when new task data is scarce and distribution bias may exist. It also avoids the computational overhead and overfitting risk of multi-step updates, significantly improving training speed.

[0152] Based on the updated parameters, this embodiment further performs gain prediction on the cold start data of new tasks. Gain prediction aims to evaluate the response probability or effectiveness improvement of different users or samples under specific marketing strategies or processes, which is of great significance for application scenarios such as precision marketing and personalized recommendations. Through the gain prediction function of the meta-gain model, enterprises can more effectively identify high-value user groups and formulate more targeted marketing strategies, thereby improving overall business performance and user experience.

[0153] In a specific embodiment, this embodiment is based on meta-parameter update training, which can obtain an uplift model adapted to the new task, namely M. new =MetaUplift(D hist D new ,Λ), where D hist Let D represent the historical task dataset, Dnew represent the cold start data for new tasks, and Λ represent the set of business constraint parameters, Λ∈{α, β, ε, λ}. Simultaneously, through gain prediction, we can ultimately obtain the customer UpliftScore prediction, which is the potential gain value for each customer under a specific treatment. This value reflects the probability of a customer's response to a marketing strategy or treatment, or the magnitude of the improvement in effectiveness. Enterprises can use customer UpliftScores for targeted marketing or personalized recommendations, concentrating resources on user groups most likely to respond or improve effectiveness, thereby improving overall business performance and user experience.

[0154] In summary, the data processing device based on the meta-gain model provided in this embodiment has the following characteristics:

[0155] (1) Model-independent meta-learning architecture: The meta-learning framework is innovatively extended to non-deep learning models, enabling traditional Uplift models with strong interpretability, such as XGBoost and causal trees, to also have meta-learning capabilities. By using a surrogate gradient mechanism, the gradient calculation problem of non-differentiable models is solved, realizing the organic combination of model interpretability and meta-learning capabilities required by highly regulated industries.

[0156] (2) Lightweight single-step inner loop optimization: A single-step inner loop design (K=1) is adopted, and the initial parameters are made into the "optimal learning starting point" through meta-training, ensuring that a high-performance model can be obtained with a single step update under the condition of a small number of samples. This design significantly reduces the computational cost and effectively avoids the risk of overfitting under the condition of multi-step updates with small samples.

[0157] (3) Task performance equalization regularization: A loss variance regularization term is introduced to solve the "negative transfer" problem in meta-learning. This mechanism automatically balances the optimization intensity of each task to ensure that the learned meta-knowledge has a stronger generalization ability, thereby improving the transfer effect of the model between different tasks.

[0158] Please see Figure 7 , Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. The computer device is equipped with both wireless and wired communication capabilities.

[0159] The computer device includes a processor 702, a memory, and a network interface 705 connected via a system bus 701. The memory may include a non-volatile storage medium 703 and internal memory 704.

[0160] The non-volatile storage medium 703 can store an operating system 7031 and a computer program 7032. When the computer program 7032 is executed, it causes the processor 702 to perform a data processing method based on a meta-gain model.

[0161] The processor 702 provides computing and control capabilities to support the operation of the entire computer device.

[0162] The internal memory 704 provides an environment for the execution of the computer program 7032 in the non-volatile storage medium 703. When the computer program 7032 is executed by the processor 702, the processor 702 can execute a data processing method based on the meta-gain model.

[0163] This network interface 705 is used for network communication with other devices. Those skilled in the art will understand that... Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0164] The processor 702 is used to run a computer program 7032 stored in a memory to implement any embodiment of the data processing method based on the meta-gain model described above.

[0165] It should be understood that, in this embodiment of the invention, the processor 702 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0166] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0167] Obtain historical task datasets and cold start data for new tasks;

[0168] The historical task dataset is sampled to obtain the corresponding sampled tasks;

[0169] The base model for data processing is parametrically proxied to obtain meta-parameters;

[0170] The meta-parameters are updated by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and the updated meta-parameters are used to update the base model to obtain the meta-gain model.

[0171] The meta-gain model is used to process the cold start data of the new task.

[0172] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0173] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0174] Obtain historical task datasets and cold start data for new tasks;

[0175] The historical task dataset is sampled to obtain the corresponding sampled tasks;

[0176] The base model for data processing is parametrically proxied to obtain meta-parameters;

[0177] The meta-parameters are updated by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and the updated meta-parameters are used to update the base model to obtain the meta-gain model.

[0178] The meta-gain model is used to process the cold start data of the new task.

[0179] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0180] Those skilled in the art will understand that all or part of the processes in the methods of 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 of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0181] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0182] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A data processing method based on a meta-gain model, characterized in that, include: Obtain historical task datasets and cold start data for new tasks; The historical task dataset is sampled to obtain the corresponding sampled tasks; The base model for data processing is parametrically proxied to obtain meta-parameters; The meta-parameters are updated by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and the updated meta-parameters are used to update the base model to obtain the meta-gain model. The meta-gain model is used to process the cold start data of the new task.

2. The data processing method based on the meta-gain model according to claim 1, characterized in that, The acquisition of historical task datasets and new task cold start data includes: The historical task dataset and the new task cold start data are obtained according to the following formulas: D hist ={(T1,D1),(T2,D2),...,(T N ,D N )}; D new =(X new ,T new ,Y new ,M new ); Among them, D hist Let T1 represent the first historical task, D1 represent the dataset corresponding to the first historical task, T2 represent the dataset corresponding to the second historical task, and D1 represent the dataset corresponding to the third historical task. N Indicates the first historical mission, D N Let D represent the dataset corresponding to the first historical task, and N represent the total number of historical tasks; new This represents the cold start data for a new task, X. new T represents the feature matrix of the new task. new Indicating a new task, Y new M represents the result vector of the new task. new This represents the metadata corresponding to the new task.

3. The data processing method based on the meta-gain model according to claim 1, characterized in that, The step of sampling the historical task dataset to obtain the corresponding sampling task includes: The historical task dataset is randomly sampled in batches based on a finite strategy for difficult tasks to obtain the sampled tasks.

4. The data processing method based on the meta-gain model according to claim 1, characterized in that, The parameter proxying of the base model for data processing yields meta-parameters, including: Obtain the hyperparameters or initial parameters of the base model; The hyperparameters or initial parameters are used as the meta-parameters.

5. The data processing method based on the meta-gain model according to claim 4, characterized in that, The process of updating the meta-parameters by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and then using the updated meta-parameters to update the base model to obtain the meta-gain model, includes: The meta-parameters are updated using a single-step update mechanism within an inner loop, according to the following formula: Where, θ′ i Here, θ represents the updated meta-parameters, and α represents the inner loop learning rate. L represents the operator. Ti This represents the loss of the i-th historical task; The gradient of the meta-parameters is estimated using the outer loop gradient calculation method according to the following formula: in, Let Lval represent the result of gradient estimation, ej represent the unit vector of the j-th dimension, and ε represent the small perturbation.

6. The data processing method based on the meta-gain model according to claim 4, characterized in that, The process of updating the meta-parameters by combining the sampling task, the inner loop single-step update mechanism, and the outer loop gradient calculation method, and then using the updated meta-parameters to update the base model to obtain the meta-gain model, further includes: The base model is globally updated using the updated meta-parameters, and a task diversity penalty is added to the base model based on a regularization mechanism. The loss function is constructed according to the following formula, and the loss function is used to optimize the loss basis model to obtain the meta-gain model: Among them, L total Let B represent the loss function, B represent the sampling task, i represent the i-th sampling task, and θ′ represent the loss function. i Let L represent the updated meta-parameters, λ represent the regularization strength parameter, and L represent the regularization strength parameter. val represents the validation loss, and Var represents the diversity task.

7. The data processing method based on the meta-gain model according to claim 1, characterized in that, The data processing of the new task cold start data using the meta-gain model includes: Load the meta-gain model and perform a single-step update on the cold start data of the new task to obtain the corresponding update parameters; Gain prediction is performed on the cold start data of the new task based on the updated parameters.

8. A data processing device based on a meta-gain model, characterized in that, include: The data acquisition unit is used to acquire historical task datasets and cold start data for new tasks; The task sampling unit is used to sample the historical task dataset to obtain the corresponding sampling task; The parameter proxy unit is used to perform parameter proxying on the base model of data processing to obtain meta-parameters; The model update unit is used to update the meta-parameters by combining the sampling task, the inner loop single-step update mechanism and the outer loop gradient calculation method, and to update the base model using the updated meta-parameters to obtain the meta-gain model. The data processing unit is used to process the cold start data of the new task using the meta-gain model.

9. A computer device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the data processing method based on the meta-gain model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the data processing method based on the meta-gain model as described in any one of claims 1 to 7.

Images