Layered continuous learning method and device, equipment and storage medium

Abstract

The embodiment of the invention provides a hierarchical continuous learning method and device, equipment and a storage medium, and is applied to the field of financial insurance. According to the method, the unified semantic representation can be generated according to the multi-source financial insurance data after the multi-source financial insurance data is acquired. And training the low-rank matrix by using the unified semantic representation to perform low-rank increment updating on the model parameters based on the low-rank matrix. The current task information is acquired, and the target expert sub-model is selected through the gating network according to the current task information, so that the decision result is output through continuous learning at multiple learning levels. According to the method, catastrophic forgetting can be relieved based on a hierarchical continuous learning strategy, so that the decision model can keep long-term stable performance under the condition that tasks are continuously increased. According to the method, low-rank adaptation and expert mixing can be utilized, the calculation efficiency is improved, and personalized results are generated according to different insurance types, regions and customer groups.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial insurance machine learning, and in particular to a hierarchical continuous learning method, device, equipment and storage medium. BACKGROUND

[0002] In the field of financial insurance, risk assessment, claim settlement, anti-fraud detection and customer service of financial insurance data can be performed through an artificial intelligence (AI) model. The artificial intelligence model refers to an algorithm model that learns rules and patterns from a large amount of data through machine learning, deep learning and other technologies, and can make predictions, classifications and decisions according to input data.

[0003] Taking the risk assessment process as an example, in the field of financial insurance, the application form of the insurance applicant, personal or enterprise historical records, vehicle information, health data, behavior patterns and other multi-dimensional data can be obtained through multi-dimensional data collection. Then, the AI algorithm model extracts features related to risk description and risk prediction from the multi-dimensional data, and selects the most relevant features according to the requirements of the model and the business goals, and optimizes the features to improve the performance of the model. A large amount of historical data is used to train the model, and after the trained model is obtained, the extracted features are input into the trained AI model to obtain a risk assessment result through the AI model.

[0004] Since the AI model applied in the field of financial insurance is based on a static training model or a single task deep learning method, the AI model lacks continuous learning ability in the field of financial insurance, and the task heterogeneity is insufficient, and the model updating cost is high. In addition, the field of financial insurance also lacks personalized modeling, which cannot achieve fine adaptation and cannot meet the urgent needs of the financial insurance industry for dynamic, robust and intelligent. SUMMARY

[0005] Therefore, the embodiments of the present application provide a hierarchical continuous learning method, device, equipment and storage medium to solve the problem of lack of continuous learning ability of the AI model in the field of financial insurance.

[0006] According to a first aspect of the present application, a hierarchical continuous learning method is provided, the method comprising:

[0007] obtaining multi-source financial insurance data, the multi-source financial insurance data comprising one or more combinations of historical claim settlement records, user portrait data, external economic indicators, natural environment data and policy data;

[0008] generating a unified semantic representation according to the multi-source financial insurance data;

[0009] train a low-rank matrix using the unified semantic representation, to perform a low-rank incremental update on model parameters based on the low-rank matrix; the low-rank matrix is a matrix obtained by performing low-rank decomposition on a key weight matrix in a basic transformation encoder;

[0010] obtain current task information, and select a target expert sub-model through a gating network according to the current task information; the target expert sub-model is an independent neural network model obtained by performing special optimization on a current task;

[0011] output a decision result by performing continuous learning at multiple learning levels based on the model parameters after the low-rank incremental update and the target expert sub-model; the multiple learning levels include a task level, a domain level, and a customer level.

[0012] According to a second aspect of the present application, a layered continuous learning device is provided, and the device comprises:

[0013] a data acquisition module configured to acquire multi-source financial insurance data, the multi-source financial insurance data comprising one or more combinations of historical claim records, user portrait data, external economic indicators, natural environment data, and policy data;

[0014] a semantic representation module configured to generate a unified semantic representation based on the multi-source financial insurance data;

[0015] an incremental update module configured to train a low-rank matrix using the unified semantic representation, to perform a low-rank incremental update on model parameters based on the low-rank matrix; the low-rank matrix is a matrix obtained by performing low-rank decomposition on a key weight matrix in a basic transformation encoder;

[0016] a mixed expert module configured to obtain current task information, and select a target expert sub-model through a gating network according to the current task information; the target expert sub-model is an independent neural network model obtained by performing special optimization on a current task;

[0017] a layered continuous learning module configured to output a decision result by performing continuous learning at multiple learning levels based on the model parameters after the low-rank incremental update and the target expert sub-model; the multiple learning levels include a task level, a domain level, and a customer level.

[0018] According to a third aspect of the present application, a computer device is provided, comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, and the processor implements the layered continuous learning method described above when executing the program.

[0019] According to a fourth aspect of the present application, a storage medium is provided, and the storage medium stores a computer program, and the program is executed by a processor to implement the layered continuous learning method described above.

[0020] By means of the technical solutions, the application provides a layered continuous learning method, device, equipment and storage medium. The method can generate a unified semantic representation according to multi-source financial insurance data after obtaining the multi-source financial insurance data. The unified semantic representation is used to train a low-rank matrix, and low-rank incremental updating is performed on model parameters based on the low-rank matrix. By obtaining current task information, a target expert submodel is selected through a gating network according to the current task information, so that the model parameters after low-rank incremental updating and the target expert submodel are used to output a decision result through multi-level continuous learning. The method can alleviate catastrophic forgetting based on the layered continuous learning strategy, so that the decision model can maintain long-term stable performance as the tasks increase. The method can also use low-rank adaptation and expert mixing to improve the calculation efficiency and generate personalized results according to different insurance types, regions and customer groups.

[0021] The above description is only a summary of the technical solutions of the application. In order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0022] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The illustrative embodiments of the application and their descriptions are used to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0023] Figure 1 A layered continuous learning method flowchart is provided for the embodiments of the application;

[0024] Figure 2 A decision model structure diagram is provided for the embodiments of the application;

[0025] Figure 3 A mixed output generation flowchart is provided for the embodiments of the application;

[0026] Figure 4 A layered continuous learning flowchart is provided for the embodiments of the application;

[0027] Figure 5 An incremental updating and knowledge distillation flowchart is provided for the embodiments of the application;

[0028] Figure 6 A layered continuous learning device structure diagram is provided for the embodiments of the application. DETAILED DESCRIPTION

[0029] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0030] In this embodiment, the hierarchical continuous learning method is a model training method used to obtain a decision model through hierarchical continuous learning training, and to output decision results using the decision model. The decision model is an artificial intelligence (AI) model applied in the financial and insurance field. In the financial and insurance field, the decision model can perform risk assessment, claims review, fraud detection, and customer service based on financial and insurance data. The decision model can learn patterns and rules from large amounts of data in the financial and insurance field through machine learning, deep learning, and other technologies, and is an algorithmic model capable of prediction, classification, and decision-making based on input data.

[0031] Taking risk assessment as an example, in the financial insurance field, multi-dimensional data collection can be conducted first, obtaining data such as insurance applicants' application forms, personal or corporate historical records, vehicle information, health data, and behavioral patterns. Then, AI algorithms are used to extract features relevant to risk description and prediction from this multi-dimensional data. Based on the model's needs and business objectives, the most relevant features are selected and optimized to improve model performance. After training the model with a large amount of historical data, the extracted features can be input into the trained AI model to obtain risk assessment results.

[0032] Because decision-making models used in the financial and insurance sectors are based on statically trained models or single-task deep learning methods, they lack continuous learning capabilities. For example, insurance business data continuously evolves with policy adjustments, market fluctuations, and user behavior, while statically trained decision-making models struggle to adapt to new data distributions, resulting in catastrophic forgetting problems.

[0033] Furthermore, due to the significant differences between tasks such as risk pricing, claims prediction, and fraud detection, statically trained models rely on a single model or full parameter fine-tuning, leading to high training costs, poor generalization ability, and insufficient task heterogeneity. Moreover, requiring complete retraining every time a new task or data arrives results in enormous computational and time overhead for model training, hindering financial institutions' ability to quickly respond to business changes and incurring high model update costs.

[0034] Furthermore, different types of insurance, different regions, and different customer groups have vastly different model requirements, and a single model cannot be finely adapted. As a result, the financial insurance field still lacks personalized modeling, which cannot achieve fine-grained adaptation and cannot meet the urgent needs of the financial insurance industry for dynamism, robustness, and intelligence.

[0035] To address the lack of continuous learning capabilities in AI models within the financial and insurance sector, this application provides a hierarchical continuous learning method in some embodiments. This method, based on incremental low-rank adaptation (LoRA) and a mixture of experts (MoE), enables risk modeling and intelligent decision-making in the insurance and financial fields. By combining low-rank adaptation, expert mixture networks, and hierarchical continuous learning mechanisms, this method achieves dynamic adaptation across tasks and timeframes while maintaining computational efficiency.

[0036] The hierarchical continuous learning method can be applied to electronic devices with data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control machines. For ease of description, this embodiment uses an electronic device as the execution subject of the method. It should be understood that the method can also be applied to other types of execution subjects, which are not illustrated in this embodiment. Figure 1 As shown, the method includes:

[0037] S101. Obtain multi-source financial and insurance data.

[0038] In this embodiment, hierarchical continuous learning is an internal data processing step within the decision model during model training or application. Through hierarchical continuous learning, the decision model can understand business data such as customer data, policy records, and disaster data from the financial and insurance industry, and output decisions and train models based on this business data.

[0039] Among them, such as Figure 2 As shown, the decision-making model includes an input layer, a representation layer, an adaptation layer, an expert hybrid layer, a hierarchical continuous learning layer, and an output layer. The input layer is used for data collection and preprocessing; that is, the input layer of the decision-making model can collect multi-source financial and insurance data such as historical claims records, customer profiles, external macroeconomic indicators, natural disaster data, and policy texts.

[0040] The representation layer employs a multimodal pre-trained model, which can vectorize data from various modalities such as text, image, structured, and time-series data, and generate a unified semantic representation through the basic Transformer encoder.

[0041] The adaptation layer is used to introduce a low-rank adaptation (LoRA) module on the key weight matrix of the Transformer to achieve low-rank incremental updates of parameters.

[0042] The expert hybrid layer can schedule different "expert sub-models" through the MoE framework, with each expert sub-model optimized for a specific task. For example, it can be specifically optimized for tasks such as risk prediction, claims review, and fraud detection, forming expert sub-models for risk prediction, claims review, and fraud detection.

[0043] The tiered continuous learning layer allows for the design of continuous learning strategies at three levels: task level, domain level, and client level, to prevent forgetting and achieve personalization.

[0044] The output layer can generate decision-making results for insurance and financial applications, including risk scoring, claims recommendations, and fraud warnings.

[0045] To achieve hierarchical and continuous machine learning, electronic devices can first acquire business data for learning. Since the decision model can be applied to the financial and insurance sector, the business data is essentially financial and insurance data within the application domain. Electronic devices can acquire financial and insurance data from multiple data sources, i.e., acquire multi-source financial and insurance data. In some embodiments, the multi-source financial and insurance data includes one or more combinations of historical claims records, user profile data, external economic indicators, natural environmental data, and policy data.

[0046] Historical claims records are business data used by financial and insurance departments to record the insurance claims process. User profile data is a set of user characteristics constructed by collecting and integrating multi-dimensional user information through big data, user profile analysis models, and other technologies. External economic indicators refer to external macroeconomic indicators, determined through data analysis, reflecting the current external macroeconomic situation. Natural environmental data refers to environmentally relevant data that impacts the financial and insurance industry, such as natural disaster data and trends in ecological and environmental changes. Policy data refers to relevant policy information that guides or influences the financial and insurance industry.

[0047] Different types of financial and insurance data can have different data sources. For example, historical claims records can come from insurance business databases, user profile data can come from user databases, external economic indicators can come from publicly available information on the Internet, natural environment data can come from dedicated databases or third-party databases related to the natural environment, and policy data can come from relevant information platforms that specify or release policies.

[0048] A data source capable of providing multi-source financial and insurance data may include one or more data source devices. In some embodiments, to acquire multi-source financial data, an electronic device, while performing multi-layer continuous learning, may first generate a data acquisition request based on the continuous learning needs, and then determine at least one target data source device based on the data acquisition request. By sending the data acquisition request to the target data source device, the target data source device can feed back multi-source financial and insurance data to the electronic device performing machine learning.

[0049] After acquiring multi-source financial and insurance data, preprocessing can be performed on the data. Specifically, in some embodiments, when acquiring multi-source financial and insurance data, original multi-source business data can be acquired first, and then preprocessed according to preset preprocessing items to obtain multi-source financial and insurance data that can be used for machine learning.

[0050] The preset preprocessing items can include one or more combinations of data cleaning, data integration, data transformation, feature selection, data segmentation, time series processing, text data preprocessing, privacy protection and compliance processing, data augmentation, anomaly detection, and feature scaling. Taking text data preprocessing as an example, after obtaining the original business data containing text data, the text data can be preprocessed by word segmentation, stop word removal, stemming, and lemmatization, and the text data can be converted into numerical features using Term Frequency-Inverse Document Frequency (TF-IDF) or word embedding.

[0051] S102. Generate a unified semantic representation based on multi-source financial and insurance data.

[0052] After acquiring multi-source financial and insurance data, a unified semantic representation can be generated based on this data. This unified semantic representation maps multi-source financial and insurance data from different modalities, languages, and formats to a shared, unified vector space and transforms it into a unified, standardized format or structure. Through this unified semantic representation, cross-domain information integration, analysis, and understanding become possible.

[0053] After acquiring multi-source financial and insurance data, we can transform the multi-source financial and insurance data in the form of text, images, audio, video, and knowledge graphs into a unified semantic representation through prompting engineering, large language models, precise alignment of multimodal embedding spaces, and structured representation of knowledge graphs.

[0054] In some embodiments, when generating a unified semantic representation from multi-source financial and insurance data, a decision model can be invoked first, and the multi-source financial and insurance data can be input into the representation layer through the input layer. Since the representation layer contains a Transformer encoder, after the multi-source financial and insurance data is input into the representation layer, the basic Transformer encoder in the representation layer can map the multi-source financial and insurance data into multiple semantic vectors in the same vector space. Then, semantic consistency alignment is performed on the multiple semantic vectors to generate a unified semantic representation.

[0055] For example, after preprocessing multi-source financial insurance data through the input layer of a decision model, the preprocessed data can be input into the representation layer of the decision model. The Transformer encoder in the representation layer can process sequential data and capture long-distance dependencies in the data through a self-attention mechanism, thereby generating a vector representation that represents the semantics of the input data. This achieves encoding of the preprocessed data and generates a unified semantic representation. In the process of generating a unified semantic representation, alignment techniques can also be applied to ensure that data from different modalities remain semantically consistent, thus achieving a more accurate semantic representation.

[0056] After generating a unified semantic representation, the same representation can be used to train and optimize the model, ensuring its high performance in decision analysis tasks within the financial and insurance domain. This involves evaluating the generated unified semantic representation to ensure its effective support for financial and insurance decision analysis. Based on the evaluation results, iterative optimization of the model may be necessary to improve accuracy and reliability.

[0057] S103. Train the low-rank matrix using a unified semantic representation to perform low-rank incremental updates on the model parameters based on the low-rank matrix.

[0058] After generating a unified semantic representation, the decision model can use this unified semantic representation corresponding to multi-source financial insurance data as incremental data for model training. Since model parameters are adjusted during training, and adjusting all parameters requires a comprehensive update of the large model, the training cost and efficiency of the decision model are high. Therefore, an incremental low-rank adaptation (LoRA) module can be introduced on top of the transformation encoder. This LoRA module performs low-rank incremental updates to the decision model, reducing training costs and improving training efficiency.

[0059] Therefore, after generating a unified semantic representation, a low-rank matrix can be trained using the unified semantic representation to perform low-rank incremental updates on the model parameters based on the low-rank matrix. The low-rank matrix is ​​obtained by performing a low-rank decomposition on the key weight matrix in the basic transformation encoder.

[0060] In some embodiments, to achieve low-rank incremental updates, when training the low-rank matrix using the unified semantic representation, the key weight matrix of the unified semantic representation can be obtained first, and low-rank decomposition can be performed on the key weight matrix to obtain the low-rank matrix. The low-rank matrix includes a first low-rank matrix and a second low-rank matrix.

[0061] Then, a unified semantic representation is used to train the first low-rank matrix and the second low-rank matrix separately. Through multiple iterations of training, the model parameters corresponding to the first and second low-rank matrices are adjusted until the iterative parameters during the training process meet the preset parameter threshold, at which point the iterative training ends. Finally, the trained first and second low-rank matrices are multiplied together and merged into the weight matrix corresponding to the model parameters.

[0062] For example, by introducing a LoRA module onto the key weight matrix of the Transformer, low-rank incremental updates of the parameters can be achieved. The LoRA module can update the weight matrix according to the following formula:

[0063] W′=W+ΔW;ΔW=AB T ;A∈R d×r ,B∈R d×r ;

[0064] Where W is the original weight matrix; W' is the updated weight matrix; ΔW is the incremental update of the weights; A is the first low-rank matrix, B is the second low-rank matrix, and both the first and second low-rank matrices have dimensions d×r, where d is the dimension of the original weight matrix W, and r is a rank much smaller than d, representing the dimension of the low-rank approximation, i.e., r << d; AB T This indicates that ΔW is obtained by transposing B through matrix multiplication of A and B; R d×r This indicates that matrices A and B belong to the real number field of d rows and r columns.

[0065] For example, when performing low-rank incremental updates, the key weight matrix can be selected first. That is, in the model architecture corresponding to the Transformer encoder, the projection matrices corresponding to the query (Q), key (K), and value (V) in the attention module, as well as the output projection matrix (O), can be selected as the target module of LoRA. Alternatively, the weight matrices (such as gate, up, down, etc.) in the feedforward network (FFN) can be selected as the target module of LoRA.

[0066] After selecting the key weight matrix, you can first freeze the original weights, i.e., freeze the pre-trained weights. At the start of fine-tuning, freeze all the original weight matrices in the Transformer model so that these frozen weights are not updated during training, thus preserving the knowledge already learned in the pre-trained model and reducing the computational and storage costs during training.

[0067] Then, a low-rank decomposition is performed on the key weight matrix using the LoRA module. That is, for each selected original weight matrix W, a first low-rank matrix A and a second low-rank matrix B are introduced, such that ΔW = AB. T The first low-rank matrix A has a dimension of d×r, the second low-rank matrix B has a dimension of k×r, and the transpose of the second low-rank matrix B has a dimension of r×k.

[0068] The first low-rank matrix A is initialized using a random Gaussian distribution, and the second low-rank matrix B is initialized using a zero matrix. Then, the low-rank matrices are trained using a unified semantic representation corresponding to multi-source financial and insurance data, enabling fine-tuning of the model parameters corresponding to the low-rank matrices. During fine-tuning, since only the parameters of low-rank matrices A and B are trained, while the original weight matrix W remains unchanged, the decision model can learn incremental updates for a specific task without needing to update a large number of original parameters.

[0069] During training, appropriate training parameters such as the learning rate and number of iterations can be set according to the specific task and dataset. For example, different learning rates can be used for training different task layers (such as general layers, domain layers, and task layers). When the training parameters meet the preset training completion conditions, the first low-rank matrix A and the second low-rank matrix B can be multiplied and merged into the weight matrix of the original model, i.e., W' = W + AB. T By using low-rank incremental training, the structure of the decision model can remain consistent with the original model during inference, ensuring that inference speed is not affected. Furthermore, when multi-task support or dynamic updates are required, the structure of the LoRA module can be maintained, and task switching can be achieved by loading different LoRA weights.

[0070] Therefore, by introducing the LoRA module, only the low-rank matrix corresponding to the LoRA module needs to be updated when incremental data or new tasks arrive, significantly reducing training overhead. Furthermore, considering the heterogeneity of tasks in the insurance domain, each task can have its own independent set of LoRA modules. This allows for efficient scaling during incremental learning, where only the LoRA weights for the corresponding task need to be stored and loaded, while the underlying model parameters are shared.

[0071] S104. Obtain current task information and select the target expert sub-model through the gating network based on the current task information.

[0072] After training the low-rank matrix using a unified semantic representation, new model parameters can be obtained based on the training results. Therefore, the output of the adaptation layer can be obtained based on the trained model parameters, and the output of the adaptation layer can be used as the input data for the expert hybrid layer for further calculation. To address the variability of decision-making tasks in the financial insurance field, the decision model can use an expert hybrid layer to produce mixed outputs. That is, the expert hybrid layer adopts the MoE framework, which improves the performance and flexibility of the decision model by combining multiple expert sub-models.

[0073] The expert hybrid layer can utilize a dynamic routing mechanism to determine which expert sub-models should be activated or used given input. Therefore, the decision model can obtain current task information based on the output of the adaptation layer, and select the target expert sub-model through a gating network based on this information. The target expert sub-model is an independent neural network model obtained through task-specific optimization.

[0074] The target expert sub-models, as small neural networks, can run in parallel to improve computational efficiency and model flexibility. In the MoE architecture, an expert refers to a set of independent sub-models, each of which is responsible for processing a specific portion or feature of the input data based on the current task type.

[0075] In a hybrid expert (MoE) architecture, a dynamic routing sub-model can also be implemented using a gating network, which determines the activation weight of each expert based on the input data. For example, a gating network can dynamically select which experts should be activated and the weight of each expert based on the characteristics of the input data, enabling the decision model to dynamically adjust its behavior according to the characteristics of the input data.

[0076] The final output of the expert hybrid layer is obtained by weighted combination of the outputs of all target expert sub-models. That is, through weighted combination, the decision model can utilize the outputs of multiple target expert sub-models, thereby improving overall performance and flexibility.

[0077] like Figure 3 As shown, in some embodiments, when performing tasks such as obtaining current task information and selecting a target expert sub-model through a gating network based on the current task information, input data can be obtained first. The input data is formed from the output data of the adaptation layer or the input layer, that is, the input data includes a unified semantic representation.

[0078] Next, task-related features are extracted from the input data, and the current task information is determined based on these features. Then, the target expert sub-model is activated based on the current task information, and the hybrid weights of the target expert sub-model are calculated using a gating network.

[0079] For example, various specialized expert sub-models can be pre-trained according to the actual task requirements. For instance, using training data related to geographical regions, a regional perception expert can be pre-trained for areas such as regional natural disasters and differences in medical security. Similarly, fraud detection experts can be pre-trained using training data related to fraudulent behavior. These experts can accurately capture fine-grained features of fraudulent behavior, obtaining feature data related to fraudulent activities.

[0080] After multi-source financial and insurance data passes through the representation and adaptation layers of the decision model to output feature data, the gating network of the expert hybrid layer can extract features that reflect the current task from the feature data, forming the current task features. If the feature data contains many region-related features, the current task type can be determined to be a region-aware task. In this case, current task information can be generated based on the region-related features to characterize the current task type. Then, the region-aware expert sub-model is activated based on the current task information, serving as the target expert sub-model.

[0081] Similarly, when the feature data contains a large number of fraud-related features, the current task type can be determined to be a fraud detection task. In this case, fraud-related features can be extracted from the feature data to form the current task information. Then, based on the current task information, a dedicated fraud detection expert sub-model is activated as the target expert sub-model.

[0082] In the MoE architecture, a dynamic routing mechanism can be used, allowing the decision model to dynamically select expert sub-models based on the characteristics of the input data. Specifically, a gating network is used to calculate the activation weights of each expert based on the input data, and then the outputs of the selected experts are combined according to these weights.

[0083] In some embodiments, by inputting input data into a target expert sub-model, the target expert sub-model performs calculations on the input data, and the expert output result output by the target expert sub-model can be obtained. Then, a mixed output is generated based on the mixed weights and the expert output result, wherein the mixed output is the weighted sum of the expert output results according to the mixed weights.

[0084] For example, to achieve dynamic routing in expert hybrid (MoE) mode, the most relevant expert, i.e., the target expert sub-model, can be selected based on input features. The input features are then fed into the target expert sub-model, allowing it to process the resulting data and obtain the output of each sub-model. Therefore, based on the MoE framework, the hybrid output can be generated according to the following formula:

[0085]

[0086] Where y represents the hybrid output, i.e., the final output of the expert hybrid network; x represents the input data, i.e., the feature data output by the input layer and the adaptation layer; E i G represents the expert output of the i-th expert sub-model. i (x) is the gating function, which can be activated using softmax; k represents the total number of expert sub-models.

[0087] As can be seen, through expert hybrid (MoE) dynamic routing, task-aware gating can be constructed to automatically select the appropriate expert sub-model in different task scenarios such as risk prediction and claims review. For example, for regional task differences such as regional natural disasters and medical security, regionally aware experts can provide regional output results, thereby improving prediction accuracy. For fraud detection experts, fine-grained features of fraudulent behavior can be captured, reducing false positives.

[0088] S105. Based on the model parameters updated by low-rank incremental updates and the target expert sub-model, the decision results are output through continuous learning at multiple learning levels.

[0089] The LoRA plugin is loaded by triggering the adaptation layer of the decision model through new tasks or new data. After selecting the target expert sub-model through the gating network of the decision model, hierarchical continuous learning can be performed based on the loaded LoRA plugin and the selected target expert sub-model. That is, based on the model parameters and target expert sub-model updated by low-rank increments, continuous learning is performed through multiple learning levels to output decision results.

[0090] The learning layers include task-level, domain-level, and customer-level. Task-level refers to the model computation process of continuously learning from feature data at one or more task levels, driven by specific tasks. Domain-level refers to comprehensive feature identification and learning across the entire business domain to output decision results at a macro level. Customer-level refers to personalized continuous learning for specific users or user groups to output personalized decision results based on customer usage preferences.

[0091] In some embodiments, when executing the model parameters and target expert sub-model based on low-rank incremental updates, and continuously learning and outputting decision results through multiple learning levels, each learning level in multiple learning levels can be read first. That is, by determining multiple learning levels, the decision model can perform hierarchical continuous learning at levels such as task level, domain level, and customer level.

[0092] Once the learning level that requires tiered continuous learning is determined, target learning units can be loaded according to the learning level. The target learning unit is used to represent the plugins, units, and sub-models that the decision model needs to load or call at the current learning level.

[0093] Clearly, the target learning units identified and loaded differ for different learning levels. That is, for example... Figure 4 As shown, when the learning level is task-level, the loaded target learning units include regularization units and low-rank adaptation plugins corresponding to the task; when the learning level is domain-level, the loaded target learning units include shared expert sub-models and regional expert sub-models; when the learning level is client-level, the loaded target learning units include personalized cue vectors and client-level fine-tuned low-rank adaptation plugins.

[0094] After loading the target learning unit according to the learning hierarchy, the decision model can use the target learning unit to perform hierarchical continuous learning updates to obtain decision results. For example, to solve problems such as catastrophic forgetting, a three-layer continuous learning strategy can be adopted. For task-level continuous learning, regularization units can be used to maintain old knowledge between different insurance tasks through regularization methods (EWC, L2-SP), while a low-rank adaptation plugin can be used to achieve lightweight updates. For domain-level continuous learning, a shared expert sub-model can be used for comprehensive output, while referring to the output of regional expert sub-models to realize cross-regional or cross-population differences, and knowledge conflicts can be avoided through expert sharing and regional experts. For customer-level personalized learning, a personalized prompt vector plugin can be used to provide personalized prompt vectors for high-value customers, and a LoRA plugin can be used to achieve customer-level fine-tuning to reduce the impact on the global model.

[0095] By applying the technical solutions of the above embodiments, the hierarchical continuous learning method described in the embodiments can collect and preprocess data at the input layer of the decision model to obtain multi-source financial insurance data such as customer data, policy records, and disaster data. Then, it performs Transformer multimodal encoding through the representation layer to achieve feature extraction and unified semantic representation. The unified semantic representation is then used as a new task or new data to trigger the adaptation layer to load the LoRA plugin and to select the target expert sub-model through the gating network of the expert hybrid layer. This allows for hierarchical continuous learning updates at multiple learning levels, including task-level, domain-level, and customer-level, outputting decision results such as insurance risk classification, claims recommendations, and fraud warnings.

[0096] The method described above can achieve low-rank updates and efficient incremental learning through LoRA, significantly reducing computation and storage costs. It can also be combined with a MoE architecture and personalized LoRA to achieve customized adaptation, meeting diverse customer needs. A hierarchical continuous learning mechanism ensures long-term task stability and improves the decision model's resistance to forgetting. Furthermore, cross-task collaboration enables the same decision model to support multi-task decisions such as risk prediction, claims processing, and fraud detection, and incremental learning allows for rapid iteration, enabling the decision model to be quickly updated in response to business and policy changes.

[0097] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and in order to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a hierarchical continuous learning method, such as... Figure 5 As shown, the method includes:

[0098] S201. Obtain incremental data;

[0099] S202. Quickly adapt incremental data using a low-rank adaptation plugin;

[0100] S203. Using knowledge distillation techniques, retain the soft labels output by the historical expert sub-model;

[0101] S204. Perform incremental updates on model parameters based on soft labels, and save the weights of important parameters during the incremental updates.

[0102] To ensure a balance between old and new knowledge, incremental updates and knowledge distillation techniques can be used to improve the adaptability and performance of decision-making models. Incremental updates refer to the ability of a pre-trained decision-making model to learn from new data over time without retraining from scratch. Incremental updates enable decision-making models to adapt to environments with constantly changing data, such as those in finance and insurance. Knowledge distillation is a model compression technique that transfers knowledge from a large, complex model (the teacher model) to a smaller, simpler model (the student model). By comparing the outputs of the teacher and student models (e.g., softmax probability distributions), the student model is trained to mimic the behavior of the teacher model, providing richer intermediate gradient information through soft objectives.

[0103] Therefore, decision-making models can acquire incremental data before application. This incremental data is also a type of multi-source financial and insurance data, which can include one or more combinations of historical claims records, user profile data, external economic indicators, natural environmental data, and policy data.

[0104] Then, a low-rank adaptation plugin is loaded to quickly adapt the incremental data. Simultaneously, following knowledge distillation techniques, the soft labels output by the historical expert sub-models are preserved, allowing for incremental updates of model parameters based on these soft labels, while simultaneously saving the weights of important parameters during these updates.

[0105] For example, to ensure a balance between old and new knowledge, when new incremental data arrives, the decision module can first perform rapid adaptation using the LoRA plugin in the adaptation layer. This involves selecting key weight matrices within the Transformer architecture, performing low-rank decomposition on these key weight matrices, and then training the model parameters corresponding to the two low-rank matrices resulting from the low-rank decomposition separately. Since only the model parameters corresponding to the low-rank matrices are trained during fine-tuning, while the original weight matrices remain unchanged, the amount of training data updates can be reduced, improving training efficiency.

[0106] Furthermore, by introducing knowledge distillation techniques, the soft labels output by historical experts are preserved, preventing performance degradation. Specifically, a large, complex model is first trained as the teacher model, which achieves high performance on specific tasks. Then, by comparing the outputs of the teacher and student models (e.g., softmax probability distributions), the student model is trained to mimic the behavior of the teacher model. Finally, a temperature parameter is used to soften the output of the teacher model to obtain soft labels, providing richer intermediate gradient information to the student model.

[0107] During incremental updates, important parameter weights can be preserved to reduce forgetting. This is achieved through regularization methods, such as Elastic Weight Consolidation (EWC) or L2-Sparse (L2-SP), to prevent the learning of new tasks from interfering with the knowledge of older tasks. Furthermore, based on experience replay, samples or features from old data are stored for reuse when training new data. During incremental updates, by preserving the weights of important parameters that significantly impact the performance of older tasks, the interference of new task learning on these important parameters is reduced.

[0108] By applying the technical solutions of the above embodiments, the hierarchical continuous learning method described in the above embodiments can, based on incremental updates and knowledge distillation techniques, quickly adapt the incremental data through a low-rank adaptation plugin after acquisition, and maintain the soft labels output by the historical expert sub-models according to knowledge distillation techniques. Then, incremental updates are performed on the model parameters based on the soft labels, and important parameter weights are preserved during the incremental updates.

[0109] The described method improves computational efficiency. For example, by utilizing LoRA for low-rank adaptation, the number of parameter updates is reduced by more than 90% compared to full-parameter fine-tuning, significantly saving computational costs. Applying this method enables the decision-making model to have continuous learning capabilities, avoiding catastrophic forgetting through a hierarchical continuous learning strategy, allowing the model to maintain long-term stable performance even with increasing tasks. The method also supports personalized and differentiated outputs. Through expert hybridization and customer-level LoRA plugins, the decision-making model can generate personalized results based on different insurance types, regions, and customer groups, improving customer satisfaction. The decision-making model can also achieve multi-task collaboration, simultaneously supporting tasks such as risk assessment, claims review, and fraud detection, enhancing the intelligence level of the entire business process. Furthermore, it has strong industry adaptability, enabling the decision-making model to quickly respond to policy changes and market dynamics, ensuring the stable operation of insurance companies in complex environments. Therefore, the described method makes the decision-making model superior to insurance and financial models in terms of efficiency, continuity, personalization, and intelligence, and has broad application prospects.

[0110] In some embodiments, as a specific implementation of the hierarchical continuous learning method described in the above embodiments, some embodiments of this application also provide a hierarchical continuous learning apparatus, such as... Figure 6 As shown, the device includes:

[0111] The data acquisition module is used to acquire multi-source financial and insurance data, which includes one or more combinations of historical claims records, user profile data, external economic indicators, natural environment data, and policy data.

[0112] The semantic representation module is used to generate a unified semantic representation based on the multi-source financial and insurance data;

[0113] The incremental update module is used to train a low-rank matrix using the unified semantic representation, so as to perform low-rank incremental updates on the model parameters based on the low-rank matrix; the low-rank matrix is ​​the matrix obtained by performing low-rank decomposition on the key weight matrix in the basic transformation encoder.

[0114] A hybrid expert module is used to acquire current task information and select a target expert sub-model through a gating network based on the current task information; the target expert sub-model is an independent neural network model obtained by performing specific optimization on the current task.

[0115] The hierarchical continuous learning module is used to continuously learn at multiple learning levels based on the model parameters updated by low-rank incremental updates and the target expert sub-model, so as to output decision results; the multiple learning levels include task level, domain level and client level.

[0116] It should be noted that other corresponding descriptions of the functional units involved in the hierarchical continuous learning device provided in the embodiments of this application can be found in the corresponding descriptions in the hierarchical continuous learning method provided in the above embodiments, and will not be repeated here.

[0117] By applying the technical solutions of the above embodiments, the hierarchical continuous learning device described in the above embodiments can, after the data acquisition module acquires multi-source financial insurance data, have the semantic representation module generate a unified semantic representation based on the multi-source financial insurance data. The incremental update module then uses the unified semantic representation to train a low-rank matrix, and performs low-rank incremental updates on the model parameters based on the low-rank matrix. The hybrid expert module acquires current task information and selects a target expert sub-model through a gating network based on the current task information, thereby enabling the hierarchical continuous learning module to continuously learn at multiple learning levels based on the low-rank incrementally updated model parameters and the target expert sub-model, in order to output decision results. The device can mitigate catastrophic forgetting based on the hierarchical continuous learning strategy, enabling the decision model to maintain long-term stable performance even with an increasing number of tasks. The device can also utilize low-rank adaptation and expert hybridization to improve computational efficiency and generate personalized results based on different insurance types, regions, and customer groups.

[0118] This application also provides a computer device, specifically a personal computer, server, network device, etc. The computer device includes a bus, processor, memory, and communication interface, and may also include input / output interfaces and a display device. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores location information. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0119] Those skilled in the art will understand that the structure of the computer device described above is only a partial structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. A specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0120] In one embodiment, a computer-readable storage medium is also provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0121] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0122] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0123] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0124] Any references to memory, database, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc.

[0125] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0126] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0127] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0128] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A hierarchical continual learning method, characterized in that, The method comprises: acquiring multi-source financial insurance data, the multi-source financial insurance data comprising one or more combinations of historical claim records, user portrait data, external economic indicators, natural environment data, and policy data; generating a unified semantic representation according to the multi-source financial insurance data; training a low-rank matrix using the unified semantic representation to perform a low-rank incremental update on model parameters based on the low-rank matrix; the low-rank matrix being a matrix obtained by performing low-rank decomposition on a key weight matrix in a basic conversion encoder; acquiring current task information and selecting a target expert submodel through a gating network according to the current task information; the target expert submodel being an independent neural network model obtained by performing special optimization on a current task; outputting a decision result by performing continuous learning at multiple learning levels based on the model parameters after low-rank incremental update and the target expert submodel; the multiple learning levels comprising a task level, a domain level, and a customer level.

2. The method of claim 1, wherein, Generating a unified semantic representation according to the multi-source financial insurance data comprises: calling a decision model, the decision model comprising an input layer, a representation layer, an adaptation layer, an expert mixing layer, a hierarchical continuous learning layer, and an output layer; inputting the multi-source financial insurance data to the representation layer through the input layer; mapping the multi-source financial insurance data into multiple semantic vectors in the same vector space through a basic conversion encoder in the representation layer; performing semantic consistency alignment on the multiple semantic vectors to generate the unified semantic representation.

3. The method of claim 1, wherein, Training a low-rank matrix using the unified semantic representation comprises: obtaining a key weight matrix of the unified semantic representation; performing low-rank decomposition on the key weight matrix to obtain a low-rank matrix, the low-rank matrix comprising a first low-rank matrix and a second low-rank matrix; training the first low-rank matrix and the second low-rank matrix using the unified semantic representation; multiplying and combining the trained first low-rank matrix and the second low-rank matrix into a weight matrix corresponding to model parameters.

4. The method of claim 1, wherein, Acquiring current task information and selecting a target expert submodel through a gating network according to the current task information comprises: obtaining input data, the input data comprising the unified semantic representation; extracting task-related features from the input data; determining current task information according to the task-related features; activating a target expert submodel based on the current task information and calculating a mixing weight of the target expert submodel using the gating network.

5. The method of claim 4, wherein, The method further comprises: inputting the input data into the target expert submodel; obtaining an expert output result output by the target expert submodel; generating a mixed output based on the mixing weight and the expert output result, the mixed output being a weighted sum result of the expert output result according to the mixing weight.

6. The method of claim 1, wherein, Outputting a decision result by performing continuous learning at multiple learning levels based on the model parameters after low-rank incremental update and the target expert submodel comprises: reading each learning level in the multiple learning levels; The target learning unit is loaded according to the learning level, wherein when the learning level is a task level, the loaded target learning unit comprises a regularization unit and a low-rank adaptive plug-in corresponding to the task; when the learning level is a domain level, the loaded target learning unit comprises a shared expert submodel and a regional expert submodel; and when the learning level is a customer level, the loaded target learning unit comprises a personalized prompt vector and a customer-level fine-tuning low-rank adaptive plug-in. The layered continuous learning update is performed using the target learning unit to obtain a decision result.

7. The method of claim 1, wherein, The method further comprises: Incremental data is obtained, and the incremental data comprises one or more combinations of historical claim records, user portrait data, external economic indicators, natural environment data and policy data; The incremental data is quickly adapted by a low-rank adaptive plug-in; Soft labels output by a historical expert submodel are maintained according to a knowledge distillation technique; Based on the soft labels, incremental updates are performed on model parameters, and important parameter weights are saved while the incremental updates are performed.

8. A layered continual learning device, comprising: The apparatus comprises: A data acquisition module is configured to acquire multi-source financial insurance data, wherein the multi-source financial insurance data comprises one or more combinations of historical claim records, user portrait data, external economic indicators, natural environment data and policy data; A semantic representation module is configured to generate a unified semantic representation according to the multi-source financial insurance data; An incremental update module is configured to train a low-rank matrix using the unified semantic representation, and perform low-rank incremental updates on model parameters based on the low-rank matrix; the low-rank matrix is a matrix obtained by performing low-rank decomposition on a key weight matrix in a basic conversion encoder; A mixed expert module is configured to acquire current task information, and select a target expert submodel through a gating network according to the current task information; the target expert submodel is an independent neural network model obtained by performing special optimization on a current task; A layered continuous learning module is configured to output a decision result by performing continuous learning at multiple learning levels based on the model parameters after the low-rank incremental updates and the target expert submodel; the multiple learning levels comprise a task level, a domain level and a customer level.

9. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, The processor executes the computer program to implement the method in any one of claims 1 to 7.

10. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the method in any one of claims 1 to 7.

Images