Joint multi-objective optimization strategy generation method and device and electronic equipment

Abstract

The invention provides a combined multi-objective optimization strategy generation method and device and electronic equipment. The method comprises the steps of obtaining original input data and performing feature extraction on the original input data to obtain feature flows associated with a plurality of optimization targets such as a risk control target and a value control target, then calling corresponding feature scoring models for different feature flows to perform scoring, and generating optimization target scores corresponding to the optimization targets. And acquiring external environment input data, adjusting a target weight coefficient corresponding to the feature flow associated with each optimization target according to the calculated optimization target score and the corresponding external environment input data, and generating a corresponding target operation strategy according to the target weight coefficient. According to the method, the risk feature flow and the value feature flow can be completely isolated in the input layer, and the output weights corresponding to the risk feature flow and the value feature flow can be adjusted in real time according to the externally input environmental data, so that the self-adaptive balance between the risk security and the value profit is realized.

Description

Technical Field

[0001] This application relates to the field of data analysis and processing technology, and in particular to a method, apparatus and electronic device for generating joint multi-objective optimization strategies. Background Technology

[0002] In data-driven intelligent decision-making, businesses and institutions often need to balance multiple objectives in complex environments. For example, financial institutions need to assess a customer's risk level and conversion potential when acquiring new customers. E-commerce platforms must prevent fraud and manage return risks when recommending products. Advertising agencies need to maximize click-through rates and conversion rates while controlling their advertising budgets. Medical diagnostics needs to balance the risk of misdiagnosis with the effectiveness of treatment. In these scenarios, how to simultaneously optimize the inherently conflicting objectives of risk control and value enhancement has become a common challenge for businesses and institutions.

[0003] With the development of artificial intelligence technology, some enterprises or institutions have begun to explore the use of artificial intelligence models to simultaneously optimize multiple conflicting objectives. However, existing solutions for using artificial intelligence models to simultaneously optimize multiple conflicting objectives still have significant limitations in adaptability when dealing with complex and ever-changing environments, and cannot generate operational strategies that take into account multiple optimization objectives at the same time. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, and electronic device for generating joint multi-objective optimization strategies, so as to improve the ability of artificial intelligence models to generate operational strategies that simultaneously take into account multiple optimization objectives in complex environments.

[0005] In a first aspect, embodiments of this application provide a method for generating a joint multi-objective optimization strategy, wherein the method includes: Obtain raw input data and perform feature extraction on the raw input data to obtain feature streams associated with each optimization objective, wherein the optimization objectives include at least: risk control objective and value control objective; The feature scoring model corresponding to each optimization objective is invoked to score the feature stream associated with each optimization objective, thereby generating an optimization objective score corresponding to each optimization objective. Obtain external environment input data, and adjust the target weight coefficients of the feature flows associated with each of the optimization targets based on the scores of each optimization target and the external environment input data; Based on the target weight coefficients, generate and output the target operation strategy.

[0006] Secondly, embodiments of this application provide a joint multi-objective optimization strategy generation apparatus, wherein the apparatus includes: The feature decoupling module is used to acquire raw input data and extract features from the raw input data to obtain feature streams associated with each optimization objective, wherein the optimization objectives include at least: risk control objective and value control objective; Multiple target feature optimization modules are provided, each of which is used to call the feature scoring model corresponding to each optimization target, score the feature stream associated with each optimization target, and generate an optimization target score corresponding to each optimization target. The dynamic gating fusion module is used to acquire external environment input data and adjust the target weight coefficient of the feature stream associated with each optimization target based on the scores of each optimization target and the external environment input data. The strategy output module is used to generate and output the target operation strategy based on the target weight coefficients.

[0007] Thirdly, embodiments of this application provide an electronic device, wherein the electronic device includes: a processor; and a memory storing a program; wherein the program includes instructions, which, when executed by the processor, cause the processor to perform the joint multi-objective optimization strategy generation method described in the first aspect.

[0008] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the joint multi-objective optimization strategy generation method described in the first aspect.

[0009] The beneficial effects of this application are: This application provides a method, apparatus, and electronic device for generating joint multi-objective optimization strategies. It acquires raw input data and extracts features from it to obtain feature streams associated with multiple optimization objectives, such as risk control and value control. Then, it calls corresponding feature scoring models to score different feature streams, generating optimization objective scores for each objective. Simultaneously, it acquires external environmental input data and adjusts the objective weight coefficients corresponding to the feature streams associated with each optimization objective based on the calculated optimization objective scores and the corresponding external environmental input data. Finally, it generates the corresponding objective operation strategy based on these objective weight coefficients.

[0010] By using the embodiments of this application, complete isolation between the risk feature stream and the value feature stream can be achieved at the input layer, significantly reducing interference between features. Furthermore, the output weights of the risk feature stream and the value feature stream can be adjusted in real time based on externally input environmental data, thereby achieving an adaptive balance between risk safety and value profitability.

[0011] This technical solution is not only applicable to financial risk control, but can also be widely used in predictive scenarios involving multi-objective optimization, such as e-commerce recommendations, advertising, medical diagnosis, and supply chain optimization, providing a unified dynamic optimization framework for complex decision-making problems. Attached Figure Description

[0012] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This paper presents a flowchart illustrating a method for generating joint multi-objective optimization strategies provided in this application. Figure 2 This paper illustrates another flowchart of the joint multi-objective optimization strategy generation method provided in this application; Figure 3 This paper shows a schematic diagram of a joint multi-objective optimization strategy generation device provided in this application; Figure 4 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of this application is shown. Detailed Implementation

[0013] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0015] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0016] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] As described in the background section, in data-driven intelligent decision-making, existing solutions that utilize artificial intelligence models to simultaneously optimize multiple conflicting objectives are unable to cope with complex and ever-changing environments. Common existing solutions for simultaneously optimizing multiple objectives include: Option 1) End-to-end joint modeling. The core idea is to directly map input features to the final joint prediction result through a single model. For example, risk features and value features are input into a deep network, relying on feature interactions to capture nonlinear relationships. This method has a simple modeling process, but because the two types of feature signals interact in the same representation space, interference often occurs, making it difficult for the model to take both objectives into account.

[0018] To overcome the drawbacks of Solution 1), the industry has proposed Solution 2: a two-stage cascaded prediction. The core idea is to train a risk prediction model on the full sample, selecting individuals that meet the risk criteria, and then train a value prediction model on the filtered sample to assess their value potential. The final result is a combination of the probabilities from both stages. This method aligns with the logical chain of business decision-making and has good interpretability. However, Solution 2) suffers from a typical "survivorship bias" problem: value prediction relies solely on the risk-passing sample, thus ignoring potentially high-value individuals within the filtered group, leading to a bias in the model's overall ranking.

[0019] Related technologies propose overcoming the biases of Scheme 2) by constructing a multi-task learning architecture through Scheme 3). Scheme 3) achieves joint optimization of the two objectives by sharing a feature encoding layer and connecting the risk prediction tower and the value prediction tower in parallel at the upper layer. In recent years, the industry has also proposed a federated multi-task framework to achieve full-sample training in cross-institutional scenarios using an encrypted feature sharing mechanism. However, in practice, risk features and value features have deep conflicts in the shared representation space, which can easily lead to gradient competition, causing the risk objective to dominate training and weakening the value prediction capability.

[0020] Furthermore, regardless of which of the three approaches mentioned above, the decision-making logic is statically set. Typically, during the training phase, parameters such as the corresponding loss function weights, the probability fusion coefficients of the cascaded model, and the task balancing parameters in the multi-task architecture are fixed and do not adjust with changes in the business environment. For example, when switching between scenarios with low risk tolerance (such as medical diagnosis and financial risk control) and scenarios prioritizing value (such as e-commerce promotions and advertising), statically designed approaches often fail to meet real-time requirements. Especially at critical junctures such as market fluctuations, policy adjustments, or changes in task objectives, models lacking dynamic adjustment mechanisms are prone to misjudgments, leading to losses in efficiency.

[0021] In view of this, this application provides a method, apparatus, and electronic device for generating joint multi-objective optimization strategies. In a first aspect, this application provides a method for generating joint multi-objective optimization strategies, applicable to any electronic device capable of generating joint multi-objective optimization strategies, including but not limited to personal mobile terminals, computers, or servers. Figure 1 As shown, the method includes the following steps: S11. Obtain raw input data and perform feature extraction on the raw input data to obtain feature streams associated with each optimization objective, wherein the optimization objectives include at least: risk control objective and value control objective; S12. Call the feature scoring model corresponding to each optimization objective, score the feature stream associated with each optimization objective, and generate the optimization objective score corresponding to each optimization objective. S13. Obtain external environment input data, and adjust the target weight coefficient of the feature stream associated with each of the optimization targets based on the scores of each optimization target and the external environment input data. S14. Generate and output the target operation strategy based on the target weight coefficients.

[0022] This application embodiment obtains raw input data and extracts features from it to obtain feature streams associated with multiple optimization objectives, such as risk control objectives and value control objectives. Then, it calls the corresponding feature scoring model to score different feature streams, generating optimization objective scores for each optimization objective. At the same time, it obtains external environment input data, and adjusts the target weight coefficients corresponding to the feature streams associated with each optimization objective based on the calculated optimization objective scores and the corresponding external environment input data. Finally, it generates the corresponding target operation strategy based on the target weight coefficients.

[0023] By using the embodiments of this application, complete isolation between the risk feature stream and the value feature stream can be achieved at the input layer, significantly reducing interference between features. Furthermore, the output weights of the risk feature stream and the value feature stream can be adjusted in real time based on externally input environmental data, thereby achieving an adaptive balance between risk safety and value profitability.

[0024] The technical solution provided in this application is not only applicable to financial risk control, but can also be widely applied to prediction scenarios with multiple objectives such as e-commerce recommendation, advertising, medical diagnosis, and supply chain optimization, providing a unified dynamic optimization framework for complex decision-making problems.

[0025] The following will provide a detailed explanation of steps S11 to S14 with specific examples: In this application, multiple objectives are flexibly set according to actual business needs. An objective refers to a goal that an enterprise or institution needs to achieve based on its business requirements, and the specific number of objectives can be flexibly set according to actual needs. For example, in the aforementioned financial risk control scenario, multiple objectives could be risk control, value control, and customer conversion, etc. In the e-commerce recommendation scenario, multiple objectives could be fraud prevention, return risk prevention, increasing repurchase rate, increasing conversion rate, etc. As a preferred implementation method, the joint multi-objective optimization strategy generation method provided in this application is mainly applied to financial risk control scenarios, and the optimization objectives in step S11 include at least: risk control objectives and value control objectives.

[0026] In this context, risk control objectives refer to the expected risks set based on business development needs, while value control objectives refer to the expected values ​​set based on business operation value needs. Since risk control objectives and value control objectives often inherently conflict—for example, in the end-to-end joint modeling of Scheme 1) and the multi-task learning architecture of Scheme 3)—risk features and value features inevitably compete when sharing the model's representation layer. Specifically, risk features often possess high data completeness and strong statistical significance; therefore, during the backpropagation of feature learning, risk features will dominate the direction of model gradient updates. Value features, on the other hand, are often sparse and highly dynamic, and their learning signals are gradually weakened during optimization. This imbalance in competition will cause the model's output operational strategies to increasingly favor risk control objectives, ultimately undermining the synergistic optimization of risk and value.

[0027] To overcome this limitation, this application, when performing step S11, performs feature extraction through the following steps S11-1 and S11-2 to obtain the feature streams associated with risk control objectives and value control objectives, i.e., to obtain the risk feature stream and the value feature stream, thereby achieving the following... Figure 2The flowchart shown separates the risk characteristics and value characteristics of the original input data at a physical level. The feature stream refers to streaming data composed of several features. Step S11-1: Call the preset feature extraction model to map the risk features associated with the risk control target and the value features associated with the value control target in the original input data to different feature spaces, thereby generating the risk control target feature space and the value control target feature space respectively.

[0028] The original input data depends on the application scenario of the method. If the method is applied to the field of financial risk control, the original input data includes basic information of financial product users, behavioral data, environmental variables, and contextual features, etc. The pre-set feature extraction model is an algorithmic model that maps the input to different feature spaces according to specified mapping rules. Specifically, a binary classification model can be selected. In this embodiment, the pre-set feature extraction model can use a learnable orthogonal projection matrix or a contrastive decoupling mechanism to map the information in the original input data that is related to the risk control objective as risk features to the risk control objective feature space, and the information in the original input data that is related to the value control objective as value features to the value control objective feature space. That is, the original input data is mapped to two independent feature spaces, thereby generating the corresponding risk control objective feature space and value control objective feature space respectively.

[0029] Specifically, the information in the raw input data related to risk control objectives refers to information related to security, stability, and compliance, such as credit scores and transaction anomalies in the lending field, or misdiagnosis probability and quality stability coefficients in the medical diagnostics field. The information in the raw input data related to value control objectives specifically refers to information focusing on potential returns, conversion tendencies, or benefits, such as purchase frequency and click behavior in e-commerce promotion, or treatment response rates or drug production turnover rates in the medical diagnostics field.

[0030] The learnable orthogonal projection matrix refers to the matrix corresponding to the risk control target feature space, which is orthogonal to the matrix corresponding to the value control target feature space. In other words, the matrices corresponding to the risk control target feature space and the value control target feature space are linearly independent. The parameters of this orthogonal projection matrix are not fixed but are obtained through continuous training and optimization of the feature extraction model. The feature extraction model can automatically adjust the matrices corresponding to the risk control target feature space and the value control target feature space based on the true patterns of risk and value in the original input data, so that the two mapped feature matrices can more accurately reflect risk and value.

[0031] In this application, adaptive dimensionality learning can also be introduced through a learnable orthogonal projection matrix, which allows the feature extraction model to automatically adjust the dimensionality of the risk control target feature space according to the data complexity, so as to avoid overfitting due to excessive dimensionality or information loss due to excessively low dimensionality.

[0032] The contrast decoupling mechanism decomposes the information affecting risk and value in the original input data into two independent feature factors. By designing a reasonable loss function, it forces the correlation between the risk features and value features extracted by the feature extraction model to be minimized. For example, the mutual information loss function can be used to make the correlation between risk features and value features 0 or close to 0.

[0033] As one implementation method, a contrastive decoupled representation learning framework can be used to replace the orthogonal projection matrix. Specifically, the orthogonality and independence of the feature space are achieved through a contrastive learning mechanism. The feature extraction model constructs negative sample pairs from risk-related and value-related features of the same object, and treats similar features of different objects as interference terms. By optimizing the contrastive loss function, the cosine similarity between similar feature vectors is minimized, so that the risk flow and value flow gradually approach an orthogonal relationship in the embedding space. This method can achieve feature isolation without explicitly constructing a projection matrix, making it suitable for high-dimensional unstructured data scenarios, such as user profile features, text vectors, or medical image embeddings. Further, in step S11-2, the feature space of the risk control target is determined as the feature flow associated with the risk control target, and the feature space of the value control target is determined as the feature flow associated with the value control target.

[0034] As one implementation method, during the execution of step S11-1, the method provided in this application may further include the following steps: A1. Using a pre-set adversarial constraint mechanism, monitor the risk characteristics to determine whether the risk characteristics carry value information associated with the value control target; A2. If the risk feature carries valuable information, a dynamic penalty constraint mechanism is initiated for the target risk feature carrying valuable information until the risk feature no longer carries valuable information.

[0035] Ideally, the risk feature space and the value feature space should be independent. However, if risk features can predict a user's value, it indicates that risk features and value features are not completely isolated, and subsequent use of risk features and value features will still cause mutual interference. This application pre-designs a dynamic monitoring and constraint logic to control the secondary entanglement of already separated risk features and value features. This dynamic monitoring and constraint logic is a pre-defined adversarial constraint mechanism. As one implementation method, this dynamic penalty constraint mechanism includes: a gradient penalty mechanism and a constraint mutual information mechanism. Step A2 above can be implemented through the following steps: A21. Dynamically activate the gradient penalty mechanism to penalize the gradient of the feature extraction model, change the gradient of the model, until the gradient of the model reaches a preset gradient threshold. A22. Dynamically activate the constraint mutual information mechanism, and adjust the loss function of the feature extraction model according to the mutual information between the risk feature and the value feature until the mutual information between the risk feature and the value feature is reduced to a preset mutual information threshold.

[0036] In this embodiment, the gradient penalty mechanism calculates the AUC (Area Under ROC Curve, the area under the ROC curve and the coordinate axes) of the feature extraction model. The AUC reflects the predictive ability of risk features for value-based behavior. In this gradient penalty mechanism, if the AUC exceeds 0.55, it means that the risk features still retain valuable information, i.e., the risk features still carry value information associated with the value control objective. For example... Figure 2 As shown, if the AUC exceeds 0.55, feature isolation constraints are required, automatically triggering the gradient penalty mechanism and the constraint mutual information mechanism.

[0037] Specifically, by penalizing the gradient of the feature extraction model, the predictive ability of the feature extraction model to learn risk features for value behavior is limited, thereby causing the feature extraction model to ignore the correlation between risk features and value features. A dynamic mutual information constraint mechanism is used to calculate the mutual information between risk features and value features, and then the loss function of the feature extraction model is adjusted based on the mutual information until the mutual information between risk features and value features decreases to a set mutual information threshold. As a preferred implementation, this mutual information threshold can be 0 or close to 0. If the mutual information is 0 or close to 0, it indicates that the risk feature and value feature remain independent of each other.

[0038] This approach, by decoupling the risk feature stream and the value feature stream, can block gradient interference between risk signals and value signals at the source, ensuring that the risk feature stream and the value feature stream maintain independence and complementarity in the dual-objective optimization process of risk control and value control objectives.

[0039] Furthermore, in step S12, the scoring model corresponding to each optimization objective is invoked to score the feature streams associated with each optimization objective, generating an optimization objective score for each objective. Taking a financial risk control scenario as an example, this can be done as follows: Figure 2 As shown, the feature space of the risk control target is processed according to the risk prediction process to output a risk score, and the feature space of the value control target is processed according to the value prediction process to output a value score. Specifically, as one implementation method, the feature scoring model corresponding to the risk control target is an XGBoost model, and the feature scoring model corresponding to the value control target is an intelligent model based on the LTV-Transformer architecture. The above step S12 is implemented through the following steps: S12-1. Call the XGBoost model to score the feature stream associated with the risk control target, and generate and output the risk score.

[0040] The process for calculating risk scores is relatively simple and can utilize models with high interpretability and structured processing capabilities, such as XGBoost, logistic regression, or graph neural networks. As a preferred implementation, XGBoost is chosen. XGBoost is a common tree-based model that learns the association patterns between risk features and risk labels by using features from the risk control target feature space as training data. After training, inputting new features from the risk control target feature space yields the corresponding risk score. This risk score is typically a probability value between 0 and 1 or a standardized score; a higher value corresponds to a higher risk. Thus, depending on the scenario, inputting the appropriate risk feature set outputs the probability of risk occurrence or the risk level. The model retains a complete decision-making path or visualization rules during operation, providing transparent traceability for various fields. For example, in finance, it can be used for credit approval explanations; in medicine, it can be used for risk diagnostic traceability; and in industrial production, it can locate the source of equipment anomalies.

[0041] S12-2. Call the intelligent model based on the LTV-Transformer architecture to dynamically calculate the lifecycle value of the feature stream associated with the value control target, and map the lifecycle value to a value score and output it.

[0042] Because value tags are scarcer than risk tags, the process for calculating value scores is more complex than that for risk scores, and it is specifically divided into several stages: full-sample engine training, pseudo-label generation, and attention value enhancement. Specifically, firstly, a full-sample training engine is trained using user data with value tags. This engine can be understood as a neural network that learns the basic correlation between value features and value tags. Further, for users without value tags, a KNN nearest neighbor search can be used to find users with value tags who are most similar to them. Then, based on the conversion rate of similar users, pseudo-value tags are generated for these users.

[0043] Furthermore, the features in the value control target feature space and their corresponding pseudo-labels are input into the intelligent model based on the LTV-Transformer architecture. This intelligent model automatically captures the correlation information between value features and pseudo-labels, and also captures the dynamic changes in long-term user value. In this process, an attention mechanism is introduced, allowing the intelligent model to automatically identify the feature dimensions that have the greatest impact on the value optimization target, such as purchase frequency, average order value, and repurchase cycle, and strengthen the weights of these feature dimensions to improve the prediction accuracy of the value score. Finally, a value score is output, which can be a predicted value of the user's lifetime value or a standardized score, where a higher value indicates higher value.

[0044] One implementation approach is to employ a joint architecture of graph neural networks (GNNs) and conditional diffusion models to calculate risk and value scores separately. Specifically, the risk propagation path between different entities is modeled using graph structures, and a generative model is introduced to reconstruct the value objective. Risk prediction can be based on heterogeneous graph structures to represent object-resource or user-asset relationships, utilizing graph attention mechanisms to characterize the transmission and aggregation process of risk factors, thereby achieving a visualized risk interpretation. Value prediction can use a conditional diffusion model to generate target labels, injecting noise in the forward phase to disrupt the true label distribution, and gradually denoising and reconstructing based on object behavioral characteristics in the reverse phase, using lifecycle value (LTV) or target return score as conditional weights to adjust the reconstruction loss. This alternative approach can effectively improve model robustness in scenarios with complex nonlinear relationships, incomplete labels, or high sample noise.

[0045] In this embodiment, value prediction focuses on predicting potential returns or target achievement rates, innovatively breaking through the limitation of traditional models relying only on partial sample training and establishing a prediction system covering all individuals. For samples with clear target labels, the model directly uses real values ​​for supervised training; for samples with missing labels, a distributed pseudo-label generation mechanism is introduced. This mechanism extracts behavioral feature vectors of samples (such as interaction frequency, preference features, behavioral trajectories, etc.) and performs a high-precision nearest neighbor search in the existing labeled sample set. If a sufficient number of similar samples are found, a pseudo-label is generated using their target average value. If the matching is insufficient, it falls back to the baseline value of the business channel or scenario. This mechanism effectively uncovers "marginal potential groups" that are ignored in traditional modeling, such as objects with moderate risk indicators but outstanding behavioral value, enabling the model to identify more targets with potential returns.

[0046] Furthermore, this application employs an LTV-Transformer architecture (or a similar self-attention model) that integrates business value metrics to dynamically calculate an individual's lifetime value (LTV) or overall contribution, mapping it to a value score of 0–100. This score is injected into the attention layer through a learnable network, increasing the weight of features from high-value samples and thus strengthening the model's ability to capture key behavioral patterns. For example, it can identify purchasing behavior during promotional windows in e-commerce scenarios, capture high ROI click paths in advertising scenarios, and identify patients with high recovery potential in medical scenarios. The output layer introduces a quantile calibration mechanism to ensure that the prediction results remain comparable and consistent across different sample distributions.

[0047] Furthermore, in step S13, a pre-defined dynamic gating fusion mechanism can be used to acquire external environmental input data and adaptively fuse risk and value characteristics. External environmental input data can include: data source or channel characteristics (e.g., APP, advertising platform, hospital department, supply node), task or product type characteristics (e.g., financial product category, commodity category, treatment plan, production process), and external environmental variables (e.g., market volatility, policy factors, price / interest rate changes, seasonal cycles, etc.). A lightweight, embedded scene-aware gating network continuously receives external environmental signals, generates a dynamic gating factor through nonlinear transformation, and then adjusts the fusion weight coefficients of the risk score and value score outputs in real time based on this dynamic gating factor as their respective target weight coefficients, thereby automatically adjusting the focus of operational decisions according to the business scenario.

[0048] As one implementation method, risk feature flow weight coefficients and value feature flow weight coefficients that match the application scenario information can be determined based on the application scenario information. For example, in high-risk constraint scenarios (such as medical diagnosis and financial compliance audits), the model automatically increases the weight coefficient of the risk score. In profit-oriented scenarios (such as advertising and e-commerce promotions), the weight coefficient of the value score is increased to enhance the impact of value control objectives. When the market fluctuates or the environment is uncertain, a balancing strategy is adopted to maintain overall robustness. The fusion unit ultimately outputs a joint predicted value, representing the individual's comprehensive risk-value potential in the current environment, providing dynamic and accurate support for intelligent decision-making.

[0049] As one implementation method, a sparse routing system based on Mixture-of-Experts (MoE) can be constructed to replace the aforementioned lightweight neural network gating structure to achieve the effect of adjusting target weight coefficients. Specifically, this coefficient routing system consists of multiple expert networks, each focusing on three typical task scenarios: risk-first, value-first, and balanced strategies. The gating engine intelligently parses contextual information and dynamically allocates samples to the optimal expert subnetwork based on real-time input scenario characteristics (such as source channel, task type, environmental variables, etc.). When a high-risk constraint scenario is identified, the system automatically activates the risk expert; when a profit-driven scenario is detected, it switches to the value expert; and in scenarios with environmental uncertainty or multiple objective trade-offs, it calls the balanced expert to achieve robust fusion. This alternative solution has the advantages of high scalability and parameter sparsity, and is suitable for large-scale dynamic prediction systems with multiple business nodes and across tasks.

[0050] Finally, step S14 is executed to generate and output the target operation strategy based on the target weight coefficients, including: Based on the target weight coefficients, the feature flows associated with each optimization target are fused to obtain a fused target feature vector; The fused target feature vector is input into a preset decision model, which then generates and outputs the target operation strategy based on the input target feature vector.

[0051] The joint prediction value output by the aforementioned dynamic gate control fusion device is essentially the risk-value comprehensive potential of an individual (such as a user) in the current scenario. It can be divided into corresponding types by constructing a two-dimensional hierarchical matrix based on risk score and value score, such as high risk-high value, high risk-medium value, high risk-low value, medium risk-high value, medium risk-medium value, etc. Each layer represents a combination of risk-value features. The features within this combination are fused into a feature vector matrix to obtain the fused target feature vector.

[0052] Furthermore, the fused target feature vector is input into a pre-trained decision model. This model, trained in advance based on business needs, determines the corresponding operational strategy for each layer based on the layer corresponding to the input fused target feature vector. For example, for high-risk users, a risk control strategy is output, such as: Strong intervention strategy (high-risk-low-value stratification): directly restrict risky behaviors, such as refusing service (refusing to lend in credit scenarios), terminating cooperation (eliminating suppliers with high default rates in supply chain scenarios), and removing products from shelves (removing products with high complaint rates in e-commerce scenarios).

[0053] Conditional restriction strategy (high-risk-medium value stratification): Under the premise of controllable risk, limited services are provided, such as reducing credit limits (from 50,000 to 20,000), shortening the cooperation period (from 30 days to 15 days), and adding review steps (adding secondary diagnosis in medical scenarios).

[0054] Risk hedging strategy (high risk - high value tiering): Balance risk and value by adding conditions, such as requiring collateral (financial scenarios), purchasing insurance (supply chain scenarios), and setting up tiered rights (users need to complete risk verification to unlock high-value services).

[0055] In this way, risk-value can be used to accurately locate objects, and corresponding operational strategies can be matched and output according to the pre-set strategy library through the pre-set decision model. Ultimately, a balance between controllable risk and value maximization can be achieved, and the output operational strategies can adapt and iterate with changes in business scenarios and object characteristics, avoiding the limitations of static strategies.

[0056] Based on the method provided in the first aspect, in the second aspect, this application provides a joint multi-objective optimization strategy generation apparatus, wherein, as... Figure 3 As shown, the device 30 includes: The feature decoupling module 31 is used to acquire raw input data and extract features from the raw input data to obtain feature streams associated with each optimization objective, wherein the optimization objectives include at least: risk control objective and value control objective; Multiple target feature optimization modules 32 are provided, each of which is used to call the feature scoring model corresponding to each optimization target, score the feature stream associated with each optimization target, and generate an optimization target score corresponding to each optimization target. The dynamic gating fusion module 33 is used to acquire external environment input data and adjust the target weight coefficient of the feature stream associated with each optimization target based on the scores of each optimization target and the external environment input data. The strategy output module 34 is used to generate and output the target operation strategy based on the target weight coefficients.

[0057] In some possible embodiments, the feature decoupling module 31 is specifically used for: The preset feature extraction model is invoked to map the risk features associated with the risk control target and the value features associated with the value control target in the original input data to different feature spaces, thereby generating the risk control target feature space and the value control target feature space, respectively. The feature space of the risk control target is determined as the feature flow associated with the risk control target, and the feature space of the value control target is determined as the feature flow associated with the value control target.

[0058] In some possible embodiments, during the process of generating the risk control target feature space and the value control target feature space respectively, the feature decoupling module is further configured to: By utilizing a pre-set adversarial constraint mechanism, the risk characteristics are monitored to determine whether the risk characteristics carry value information associated with the value control objective. If the risk feature carries valuable information, a dynamic penalty constraint mechanism is initiated for the target risk feature carrying valuable information until the risk feature no longer carries valuable information.

[0059] In some possible embodiments, the dynamic penalty constraint mechanism includes: a gradient penalty mechanism and a constraint mutual information mechanism; the activation of the dynamic penalty constraint mechanism includes: The gradient penalty mechanism is dynamically activated to penalize the gradient of the feature extraction model, thereby changing the gradient of the model until the gradient of the model reaches a preset gradient threshold. The constraint mutual information mechanism is dynamically activated, and the loss function of the feature extraction model is adjusted according to the mutual information between the risk feature and the value feature until the mutual information between the risk feature and the value feature is reduced to a preset mutual information threshold.

[0060] In some possible embodiments, the feature scoring model corresponding to the risk control objective is an XGBoost model, and the feature scoring model corresponding to the value control objective is an intelligent model based on the LTV-Transformer architecture. Each of the target feature optimization modules is used for: The XGBoost model is invoked to score the feature stream associated with the risk control objective, and a risk score is generated and output. The intelligent model based on the LTV-Transformer architecture is invoked to dynamically calculate the lifecycle value of the feature stream associated with the value control target, and the lifecycle value is mapped to a value score and output.

[0061] In some possible embodiments, the external environment input data includes: application scenario information; the dynamic gating fusion module is specifically used for: Based on the application scenario information, determine the risk feature flow weight coefficient and the value feature flow weight coefficient that match the application scenario information.

[0062] In some possible embodiments, the policy output module 34 is specifically used for: Based on the target weight coefficients, the feature flows associated with each optimization target are fused to obtain a fused target feature vector; The fused target feature vector is input into a preset decision model, which then generates and outputs the target operation strategy based on the input target feature vector.

[0063] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application comply with relevant laws and regulations and do not violate public order and good morals.

[0064] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0065] Thirdly, exemplary embodiments of this application also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the electronic device to perform a method according to an embodiment of this application.

[0066] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of this application.

[0067] An exemplary embodiment of this application also provides a computer program product, including a computer program, wherein, when executed by a computer's processor, the computer program is used to cause the computer to perform a method according to an embodiment of this application.

[0068] refer to Figure 4The present invention describes a structural block diagram of an electronic device 400 that can serve as a server or client of this application, which is an example of a hardware device that can be applied to various aspects of this application. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0069] like Figure 4 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM 402) or a computer program loaded from a storage unit 408 into a random access memory (RAM 403). The RAM 403 may also store various programs and data required for the operation of the electronic device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output interface (I / O interface 405) is also connected to the bus 404.

[0070] Multiple components in electronic device 400 are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information to electronic device 400. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows electronic device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0071] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the aforementioned joint multi-objective optimization strategy generation method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 400 via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the aforementioned joint multi-objective optimization strategy generation method by any other suitable means (e.g., by means of firmware).

[0072] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0073] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0074] As used in this application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0075] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0076] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0077] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

Claims

1. A method for generating a joint multi-objective optimization strategy, characterized in that, The method comprises: obtaining original input data and performing feature extraction on the original input data to obtain a feature flow associated with each optimization target, wherein the optimization target at least includes a risk control target and a value control target; calling a feature scoring model corresponding to each optimization target to score the feature flow associated with each optimization target to generate an optimization target score corresponding to each optimization target; obtaining external environment input data, and adjusting a target weight coefficient of the feature flow associated with each optimization target based on the optimization target score and the external environment input data; generating and outputting a target operation strategy according to the target weight coefficient.

2. The method of claim 1, wherein, The method of obtaining original input data and performing feature extraction on the original input data to obtain a feature flow associated with each optimization target comprises: calling a preset feature extraction model to map risk features associated with the risk control target in the original input data and value features associated with the value control target in the original input data to different feature spaces respectively to generate a risk control target feature space and a value control target feature space respectively; determining the risk control target feature space as the feature flow associated with the risk control target and determining the value control target feature space as the feature flow associated with the value control target.

3. The method of claim 2, wherein, In the process of generating the risk control target feature space and the value control target feature space respectively, the method further comprises: monitoring the risk features by using a preset adversarial constraint mechanism to determine whether the risk features carry value information associated with the value control target; if the risk features carry value information, starting a dynamic penalty constraint mechanism for the target risk features carrying value information until the risk features do not carry value information.

4. The method of claim 3, wherein, The dynamic penalty constraint mechanism comprises a gradient penalty mechanism and a constraint mutual information mechanism; and the starting of the dynamic penalty constraint mechanism comprises: dynamically starting the gradient penalty mechanism to penalize the gradient of the feature extraction model, changing the gradient of the model until the gradient of the model reaches a preset gradient threshold; dynamically starting the constraint mutual information mechanism to adjust the loss function of the feature extraction model according to the mutual information between the risk features and the value features until the mutual information between the risk features and the value features decreases to a preset mutual information threshold.

5. The method of claim 1, wherein, The feature scoring model corresponding to the risk control target is an XGBoost model, the feature scoring model corresponding to the value control target is an intelligent model based on an LTV-Transformer architecture, and the calling of the feature scoring model corresponding to each optimization target to score the feature flow associated with each optimization target to generate an optimization target score corresponding to each optimization target comprises: calling the XGBoost model to score the feature flow associated with the risk control target to generate and output a risk score; The intelligent model based on the LTV-Transformer architecture is called to dynamically calculate a life cycle value of the feature flow associated with the value control target, and the life cycle value is mapped to a value score and output.

6. The method of claim 1, wherein, The external environment input data includes application scenario information. Based on the optimization target scores and the external environment input data, the target weight coefficients of the feature flows associated with the optimization targets are adjusted, including: According to the application scenario information, a risk feature flow weight coefficient matched with the application scenario information and a value feature flow weight coefficient matched with the application scenario information are determined.

7. The method of claim 6, wherein, The target operation strategy is generated and output according to the target weight coefficients, including: According to the target weight coefficients, the feature flows associated with the optimization targets are fused to obtain a fused target feature vector. The fused target feature vector is input into a preset decision model, and the target operation strategy is generated based on the input target feature vector by the preset decision model.

8. A joint multi-objective optimization strategy generation apparatus characterized by comprising: The device includes: A feature decoupling module is configured to acquire original input data, perform feature extraction on the original input data, and acquire feature flows associated with optimization targets, wherein the optimization targets include at least a risk control target and a value control target. A plurality of target feature optimization modules are configured to respectively call feature score models corresponding to the optimization targets, score the feature flows associated with the optimization targets, and generate optimization target scores corresponding to the optimization targets. A dynamic gating fusion module is configured to acquire external environment input data, adjust target weight coefficients of the feature flows associated with the optimization targets based on the optimization target scores and the external environment input data. A strategy output module is configured to generate and output a target operation strategy according to the target weight coefficients.

9. An electronic device, comprising: The electronic device includes a processor and a memory storing a program, wherein the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images