Explainable meta-learning weight adjustment method and system

Abstract

The application provides an interpretable meta-learning weight adjustment method and system, wherein the method first pretrains a multi-task historical data set through a meta-learning framework to obtain general initialization parameters; in an online stage, a meta-encoder is used to generate feature representations of users and market states, and an initial weight of an agent is predicted through a weight adapter, and then a small amount of gradient update is performed under the driving of a joint loss function containing a yield loss, a risk loss and an explanation consistency loss ECL to realize millisecond-level personalized adaptation and obtain the final weight. The explanation consistency loss ECL is a regularization term of an explanation function, and the stability of a model attribution vector is constrained in the training stage to solve the inconsistent explanation result problem of the existing interpretable artificial intelligence method. The application can be applied to intelligent transactions, recommendation systems, industrial prediction and other scenes requiring high real-time performance and regulatory transparency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and financial technology, in particular to an explainable meta-learning weight adjustment method and system. BACKGROUND

[0002] With the wide application of artificial intelligence technology in high-risk and high-transparency fields such as financial transactions and recommendation systems, the explainability of models (Explainable AI, XAI) has become a core requirement for regulatory compliance and user trust. Existing explainability tools (such as SHAP, LRP, etc.) can perform post-hoc attribution analysis on trained models to reveal the importance of features relied on by model decisions. However, these methods have a significant drawback: when input features change slightly, the explanation results are prone to dramatic fluctuations, lack stability, and are difficult to pass the "decision consistency" review requirements of financial regulatory agencies.

[0003] On the other hand, in multi-agent systems (such as clusters of trading robots), how to adjust the weights of different agents in real time for different users or market environments is a highly challenging problem. Traditional methods often rely on global model fine-tuning, which is time-consuming and computationally expensive, and cannot meet the needs of millisecond-level real-time decision-making, especially in new user or new market scenarios. In addition, existing methods often only focus on maximizing returns, lacking joint optimization of risk control and explanation consistency, leading to performance and compliance pressures in actual deployment.

[0004] Through the search of patent documents, the invention patent with publication number CN114463576A is found, which discloses a network training method based on reweighting strategy. It adopts a class imbalance processing scheme based on a phased progressive learning strategy, which can alleviate the problem of imbalance between head and tail classes in class imbalance datasets, effectively reducing the risk of overfitting of tail class data and underfitting of head class data. At the same time, it can smoothly transition from the general pattern of learning representation to the training of upper-level classifiers, preserving the original learned deep representation while learning the classifier. In addition, it has good robustness, and can adapt to datasets with different degrees of imbalance or datasets with different sample sizes by controlling phased hyperparameters and progressive hyperparameters. This patent focuses on network training for class imbalance datasets and does not address AI model explainability, which cannot solve the problem of unstable explanation results. It also does not mention agent weight adjustment, which cannot meet real-time requirements.

[0005] In summary, in view of the problems of the existing technology, it is a key task to develop an explainable meta-learning weight adjustment method and system. SUMMARY

[0006] Aiming at the defects in the prior art, the present application aims to provide an explainable meta-learning weight adjustment method and system.

[0007] According to the present application, an explainable meta-learning weight adjustment method is provided, which comprises the following steps:

[0008] Step S1, based on a multi-task historical data set, using MAML algorithm or Reptile algorithm for meta-learning training, obtaining a general initialization parameter;

[0009] Step S2, obtaining the user portrait feature and market state feature of the new task, inputting the user portrait feature and market state feature into the meta-encoder initialized by the general initialization parameter for encoding, obtaining a semantic representation vector;

[0010] Step S3, inputting the semantic representation vector into the weight adapter initialized by the general initialization parameter, generating an initial agent weight;

[0011] Step S4, based on a joint loss function including yield loss, risk loss and explanation consistency loss, performing K times gradient descent update on the initial agent weight, obtaining the final weight;

[0012] Step S5, using the final weight to drive decision and recording process log.

[0013] Preferably, in step S1, the input data of the multi-task historical data set includes user portrait feature, market state feature and task label, and the task label includes target yield type.

[0014] Preferably, in step S2, the user portrait feature includes transaction preference and risk level; the market state feature includes volatility, trading volume and industry rotation intensity; and the meta-encoder is a neural network combining attention mechanism and full connection layer.

[0015] Preferably, in step S2, the user portrait feature and market state feature are encoded by the meta-encoder, and the output semantic representation vector z is represented as:

[0016]

[0017] Wherein,

[0018] The mathematical function of the meta-encoder is represented as:

[0019] The user portrait feature is represented as:

[0020] The market state feature is represented as:

[0021] A semantic representation vector representing the fusion of user and market information.

[0022] Preferably, in step S3, the semantic representation vector is input to a weight adapter to generate initial agent weights for the current task , denoted as:

[0023]

[0024] wherein,

[0025] is a gated fully connected neural network with task transfer generalization capability.

[0026] Preferably, in step S4, the joint loss function L is in the form of a weighted sum, denoted as:

[0027]

[0028] wherein,

[0029] , and denote weight coefficients;

[0030] denotes a return loss for maximizing investment returns;

[0031] denotes a risk loss for constraining volatility;

[0032] denotes an explanation consistency loss for constraining the stability of the model's explanation results for similar inputs.

[0033] Preferably, in step S4, the explanation consistency loss is defined as: for a pair of input samples and , if - ||<δ, then:

[0034]

[0035] wherein,

[0036] denotes an explanation vector of the input sample calculated by the SHAP attribution method;

[0037] denotes an explanation vector of the input sample calculated by the SHAP attribution method;

[0038] a cosine similarity function representing an explanation vector;

[0039] a preset similarity threshold value for screening sample pairs of stability comparison.

[0040] Preferably, in step S4, the number of gradient descent updates K≤5.

[0041] Preferably, in step S5, the optimal weight is output to a decision system, the decision system uses the optimal weight to drive an agent to generate a decision behavior, and the user portrait feature, the market state feature, the explanation vector and the decision behavior are recorded to a compliance log system.

[0042] The application also provides an explainable meta-learning weight adjustment system, comprising:

[0043] Module M1, based on a multi-task historical data set, performs meta-learning training using a MAML algorithm or a Reptile algorithm to obtain a universal initialization parameter;

[0044] Module M2, obtaining a user portrait feature and a market state feature of a new task, inputting the user portrait feature and the market state feature to a meta-encoder initialized by the universal initialization parameter for encoding to obtain a semantic representation vector;

[0045] Module M3, inputting the semantic representation vector to a weight adapter initialized by the universal initialization parameter to generate an initial agent weight;

[0046] Module M4, based on a joint loss function, performing gradient descent update on the initial agent weight not more than K times to obtain an optimal weight;

[0047] Module M5, using the optimal weight to drive decision and recording process log.

[0048] Compared with the prior art, the application has the following beneficial effects:

[0049] 1. The application can quickly generate personalized initial weights in a new user or new market environment through the synergistic effect of Meta-Encoder and Weight Adapter, significantly improving the adaptation speed and performance of the model in a cold start scenario.

[0050] 2. The application introduces a joint loss function to effectively control risk exposure while optimizing returns, and uses an explanation consistency loss to improve the stability of the explanation result, solving the problem of unstable explanation results of the explainable method of the existing model, and meeting the requirements of financial supervision on decision consistency and auditability.

[0051] 3. This invention completes weight fine-tuning in a very short time through a few-step gradient update mechanism, making it suitable for real-time decision-making scenarios that are sensitive to latency.

[0052] 4. All decision-making processes in this invention are recorded in the compliance log system, supporting post-audit, stress testing, and consistency playback, providing a complete and transparent decision-making trajectory for regulatory review. Attached Figure Description

[0053] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0054] Figure 1 This is a flowchart of an interpretable meta-learning weight adjustment method in an embodiment of the present invention. Detailed Implementation

[0055] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0056] This invention discloses an interpretable meta-learning weight adjustment method and system. The method first pre-trains a multi-task historical dataset using a meta-learning framework to obtain general initialization parameters. In the online phase, a meta-encoder generates feature representations of user and market states, and a weight adapter predicts the initial weights of the agent. Subsequently, a small number of gradient updates are performed under the drive of a joint loss function including profit loss, risk loss, and interpretability consistency loss (ECL), achieving millisecond-level personalized adaptation to obtain the final weights. The interpretability consistency loss (ECL), as a regularization term of the interpretability function, constrains the stability of the model's attribution vector during the training phase, thus addressing the problem of inconsistent interpretation results in existing interpretable artificial intelligence methods. This invention can be applied to scenarios requiring high real-time performance and regulatory transparency, such as intelligent trading, recommendation systems, and industrial forecasting.

[0057] Example 1:

[0058] This embodiment provides an interpretable meta-learning weight adjustment method, including the following steps:

[0059] Step S1: Based on the multi-task historical dataset, perform meta-learning training using the MAML (Model-Agnostic Meta-Learning) algorithm or the Reptile algorithm to obtain general initialization parameters. These general initialization parameters serve as the initial values ​​for the model parameters and are used for rapid adaptation in subsequent new tasks.

[0060] Specifically, the input data of the multi-task historical data set includes user portrait features, market state features and task labels, and the task labels include target return types.

[0061] Through meta-learning training, the model can quickly converge through a few gradient update steps when facing new tasks.

[0062] Step S2, obtain the user portrait features and market state features of the new task, input the user portrait features and market state features into the meta-encoder initialized by the general initialization parameters for encoding to obtain the semantic representation vector, which is used for weight prediction.

[0063] Specifically, the user portrait features include transaction preferences and risk levels; the market state features include volatility, trading volume and industry rotation intensity; and the meta-encoder is a neural network that integrates attention mechanisms and fully connected layers.

[0064] Further, the user portrait features and market state features are encoded by the meta-encoder, and the output semantic representation vector z is represented as:

[0065]

[0066] Wherein,

[0067] represents the user portrait features;

[0068] represents the market state features;

[0069] represents the semantic representation vector that integrates user and market information;

[0070] Step S3, input the semantic representation vector into the weight adapter initialized by the general initialization parameters to generate the initial agent weight.

[0071] Specifically, the semantic representation vector is input into the weight adapter to generate the initial agent weight of the current task , which is represented as:

[0072]

[0073] Wherein,

[0074] is a gated fully connected neural network with task transfer generalization ability.

[0075] As the initial point of subsequent fine-tuning, to avoid instability caused by direct fine-tuning from the general initialization parameters.

[0076] Step S4, performing K times of gradient descent updates on the initial agent weights based on a joint loss function comprising a return loss, a risk loss, and an explanation consistency loss, to obtain final weights.

[0077] Specifically, the joint loss function L is in the form of a weighted sum, denoted as:

[0078]

[0079] wherein,

[0080] , and denote weight coefficients;

[0081] denotes a return loss for maximizing investment returns;

[0082] denotes a risk loss for constraining volatility;

[0083] denotes an explanation consistency loss for constraining the stability of the model's explanation results for similar inputs.

[0084] Explanation consistency loss is measured by calculating the cosine similarity between explanation vectors in different time windows.

[0085] Further, the explanation consistency loss is defined as: for a pair of input samples and , if - ||<δ, then:

[0086]

[0087] wherein,

[0088] denotes an explanation vector of the input sample calculated by the SHAP attribution method, reflecting the features of interest to the model;

[0089] denotes an explanation vector of the input sample calculated by the SHAP attribution method;

[0090] denotes a cosine similarity function of the explanation vectors;

[0091] denotes a preset similarity threshold for screening sample pairs for stability comparison.

[0092] The final weight and the explanation vector are used to achieve personalized adaptation in a new user or a new market environment, and the adaptation process is completed within milliseconds, while the stability of the model performance and the explanation result is considered.

[0093] Further, in step S4, the number of gradient descent updates K≤5, and in the embodiment .

[0094] In step S5, the final weight is used to drive decision-making and process logs are recorded.

[0095] Specifically, the final weight is output to a decision system, the final weight is used by the decision system to drive an agent to generate a decision behavior, and the user portrait features, market state features, explanation vector and decision behavior are recorded to a compliance log system to meet the requirements of financial supervision on operation traceability, responsibility attribution and decision explainability, and to provide data support for stress testing and consistency playback analysis.

[0096] In the embodiment, the decision system is a multi-agent decision module known in the art.

[0097] Embodiment 2:

[0098] The application also provides an interpretable meta-learning weight adjustment system, which can be realized by executing the process steps of the interpretable meta-learning weight adjustment method, that is, the interpretable meta-learning weight adjustment method can be understood by those skilled in the art as a preferred implementation of the interpretable meta-learning weight adjustment system.

[0099] Specifically, the interpretable meta-learning weight adjustment system comprises:

[0100] Module M1, based on a multi-task historical data set, performs meta-learning training using a MAML algorithm or a Reptile algorithm to obtain a universal initialization parameter;

[0101] Module M2, acquires user portrait features and market state features of a new task, and inputs the user portrait features and the market state features to a meta-encoder initialized by the universal initialization parameter for encoding to obtain a semantic representation vector;

[0102] Module M3, inputs the semantic representation vector to a weight adapter initialized by the universal initialization parameter to generate an initial agent weight;

[0103] Module M4, based on a joint loss function, performs gradient descent updates on the initial agent weight not more than K times to obtain a final weight;

[0104] Module M5, uses the final weight to drive decision-making and records process logs.

[0105] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module and unit thereof in the form of pure computer readable program code, the system provided by the present application and each device, module and unit thereof can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers, etc. by logically programming the method steps to achieve the same functions. Therefore, the system provided by the present application and each device, module and unit thereof can be considered as a hardware component, and the devices, modules and units included therein for achieving various functions can also be considered as structures within the hardware component; the devices, modules and units for achieving various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0106] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other in any manner without conflict.

Claims

1. An explainable meta-learning weight adjustment method, characterized in that, The method comprises the following steps: Step S1, based on a multi-task historical data set, using a MAML algorithm or a Reptile algorithm for meta-learning training to obtain a universal initialization parameter, the input data of the multi-task historical data set comprising user portrait features, market state features and task labels, the task labels comprising target return types; Step S2, obtaining user portrait features and market state features of a new task, inputting the user portrait features and the market state features into a meta-encoder initialized by the universal initialization parameter for encoding to obtain a semantic representation vector, the user portrait features comprising transaction preferences and risk levels, and the market state features comprising volatility, trading volume and industry rotation intensity; Step S3, inputting the semantic representation vector into a weight adapter initialized by the universal initialization parameter to generate initial agent weights; Step S4, performing K times of gradient descent updates on the initial agent weight based on a joint loss function comprising a reward loss, a risk loss and an explanation consistency loss to obtain final weights; the explanation consistency loss is defined as: for a pair of input samples and , if - ||<δ, then: Wherein, an explanatory vector of the input sample calculated by the SHAP attribution method ; represents an explanation vector of the input sample calculated by the SHAP attribution method ; denotes a cosine similarity function that interprets the vector; represents a preset similarity threshold value for screening the sample pairs of stability comparison; Step S5, using the optimal weights to drive decision-making and recording process logs.

2. The explainable meta-learning weight adjustment method of claim 1, wherein, In the step S2, the meta-encoder is a neural network combining attention mechanisms and fully connected layers.

3. The explainable meta-learning weight adjustment method of claim 2, wherein, In the step S2, the semantic representation vector z output by the meta-encoder is represented as: Wherein, mathematical function representing the meta-encoder; representing a user profile characteristic; representing market state characteristics; represent semantic representation vectors that fuse user and market information.

4. The explainable meta-learning weight adjustment method of claim 3, wherein, In the step S3, the semantic representation vector is input to a weight adapter to generate an initial agent weight of the current task is represented as: Wherein, A gated fully connected neural network with task migration generalization capability.

5. The interpretable meta-learning weight adjustment method of claim 1, wherein, In the step S4, the joint loss function L is in the form of a weighted sum, represented as: Wherein, , and denote weight coefficients; represents a loss of yield, for maximizing return on investment; represents a risk loss, used to constrain volatility; represents an explanation consistency loss used to constrain the stability of the explanation results of the model for similar inputs.

6. The interpretable meta-learning weight adjustment method of claim 1, wherein, In the step S4, the number of gradient descent updates K≤5.

7. The interpretable meta-learning weight adjustment method of claim 1, wherein, In the step S5, the optimal weights are output to a decision system, the decision system uses the optimal weights to drive an agent to generate decision-making behaviors, and the user portrait features, the market state features, the explanation vectors and the decision-making behaviors are recorded to a compliance log system.

8. An interpretable meta-learning weight adjustment system employing the interpretable meta-learning weight adjustment method of any one of claims 1 to 7. Comprise: Module M1, based on a multi-task historical data set, using a MAML algorithm or a Reptile algorithm for meta-learning training to obtain a universal initialization parameter, the input data of the multi-task historical data set comprising user portrait features, market state features and task labels, the task labels comprising target return types; Module M2, obtaining user portrait features and market state features of a new task, inputting the user portrait features and the market state features into a meta-encoder initialized by the universal initialization parameter for encoding to obtain a semantic representation vector, the user portrait features comprising transaction preferences and risk levels, and the market state features comprising volatility, trading volume and industry rotation intensity; Module M3, inputting the semantic representation vector into a weight adapter initialized by the universal initialization parameter to generate initial agent weights; The module M4 performs gradient descent updates on the initial agent weights based on a joint loss function comprising a reward loss, a risk loss, and an explanation consistency loss, for at most K times, to obtain final weights; the explanation consistency loss is defined as: for a pair of input samples and , if - || < δ, then: Wherein, an explanation vector for the input sample calculated by the SHAP attribution method ; an input sample for which an explanation vector is computed by the SHAP attribution method the explanation vector; denotes a cosine similarity function that interprets the vector; represents a preset similarity threshold value for screening the sample pairs of stability comparison; Module M5, using the optimal weights to drive decision-making and recording process logs.

Images