A graph neural network forgetting system and method considering forgetting and performance preservation
By optimizing local and global forgetting loss constraints and the Pareto balance algorithm, the problem of performance preservation in graph neural network forgetting is solved, achieving high prediction performance in the forgetting process, which is suitable for graph neural network forgetting systems in medical scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO ARTIFICIAL INTELLIGENCE RES INST OF SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-19
AI Technical Summary
Existing techniques in graph neural networks have failed to effectively account for the impact of forgetting operations on model performance, resulting in poor performance retention and an easy over-elimination of useful features, which affects the model's predictive performance.
By constraining forgetting loss at both local and global levels and combining it with the Pareto balancing algorithm for adaptive weight allocation, the forgetting system of the graph neural network is optimized to maintain high predictive performance during the forgetting process.
It achieves the preservation and improvement of model performance after forgetting operations, reduces the consumption of computing resources, and is applicable to a wide range of medical scenarios such as anesthesia system decision-making.
Smart Images
Figure CN122242641A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and privacy protection technology, and in particular to a graph neural network forgetting system and method that balances forgetting and performance retention. Background Technology
[0002] With the advancement of technology and social development, artificial intelligence (AI) technology is increasingly being applied in daily life. Graph Neural Network (GNN) algorithms, as a method capable of integrating multiple types of information, are also increasingly being used in various AI algorithm models. GNNs are widely used in scenarios such as social networks, recommender systems, financial risk control, and medical decision-making for tasks such as node classification, edge prediction, and graph-level prediction. In medical scenarios such as anesthesia decision-making, patient physiological indicators, medication records, and anesthesia procedures naturally form non-Euclidean relational networks. GNNs are often used for ASA (American Society of Anesthesiologists) classification prediction and complication risk prediction. At the same time, people's awareness of protecting personal information is growing stronger, namely the "right to be forgotten." The "right to be forgotten" requires models to effectively remove the influence of data from the trained model when a user withdraws data authorization, giving rise to "machine unlearning," which in the field of graph neural networks has become "graph unlearning" research. In the current field of anesthesia system decision-making, there is still a lack of a unified, multi-objective, and controllable framework for "how to systematically constrain the predictive performance of remaining nodes while performing machine forgetting in graph neural networks".
[0003] The invention patent "System, method, and computer program product for machine unlearning on identity graph neural networks", publication number WO2025090089A1, discloses a machine forgetting method and apparatus for identity graph neural networks. The method includes: acquiring an identity graph composed of multiple subgraphs, each subgraph including multiple edges and multiple nodes associated with these edges, where all edges and nodes in each subgraph correspond to the same identity; applying at least one edge enhancement algorithm to the identity graph to transform it into a biconnected identity graph; dividing the biconnected identity graph into multiple biconnected components, ensuring that each component has no cut vertices; training a graph neural network on each component to generate a graph embedding and a local minimum for that component; training an ensemble neural network based on the graph embeddings and local minimums of all components; and finally providing the trained ensemble neural network. The premise of this method is that after deleting nodes and edges related to the "identity to be forgotten" from the identity graph, the impact of the deletion is localized through structural enhancement (such as adding edges to form biconnectivity) and block training strategies. This method focuses on eliminating the effects of deletion without considering improving the model's predictive performance, and is somewhat lacking in maintaining the model's performance.
[0004] The invention patent "A Method for Predicting Anesthesia Effect Based on Graph Neural Networks", publication number CN120932937A, discloses a method for predicting anesthesia effects based on graph neural networks. The method includes: obtaining a standardized synchronized time-series data set; forming a time-sliding window sequence; generating an initial topology for a dynamic physiological indicator-pharmacological variable graph structure; generating an initial feature representation set for each graph node, and simultaneously initializing the graph neural network model parameter set; obtaining an updated dynamic physiological indicator-pharmacological variable graph structure; inputting the pharmacodynamic variable graph structure into the graph neural network model corresponding to the graph neural network model parameter set to obtain a node embedding representation set and an edge embedding representation set; outputting a set of predicted anesthesia effects during the induction period; and generating an interpretable output result set, which includes key variable links, dominant influencing factors, and dosing adjustment suggestions, thus improving the ability to rapidly detect abnormal states during the induction period. However, this method only considers the predictive effect of graph neural networks in the field of anesthesia and does not consider the performance preservation of the model when deletion operations are required.
[0005] Therefore, those skilled in the art are dedicated to developing a graph neural network forgetting system and method that balances forgetting and performance retention. Summary of the Invention
[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to eliminate the impact of the deleted part while maintaining performance.
[0007] The applicant argues that existing technologies only focus on eliminating the impact of deletion operations on model performance, without optimizing for performance preservation or improvement. Furthermore, over-elimination during the elimination process can lead to the removal of useful features, causing model performance to collapse. Graph neural networks aggregate neighborhood information through multi-layer message passing mechanisms. When a node or edge is deleted, its impact propagates along the graph structure in multi-hop neighborhoods, causing changes in the embeddings and predictions of many non-deleted nodes. In medical scenarios, models must not only meet privacy regulations' requirements for "complete forgetting" but also maintain sufficiently high predictive performance on the remaining data. The applicant's proposed graph neural network, which balances forgetting and performance preservation, addresses this by using two levels of forgetting loss (local and global) to constrain the representation of deleted elements and their neighborhoods, effectively fading them out of the model. Simultaneously, it applies a prediction preservation loss to the target node set and combines it with a Pareto balancing algorithm for adaptive weight allocation. This achieves a dynamic balance between forgetting intensity and performance preservation within a multi-objective optimization framework, simultaneously addressing both "eliminating the impact of deleted data" and "preserving remaining predictive performance."
[0008] In one embodiment of the present invention, a graph neural network forgetting system that balances forgetting and performance retention is provided, comprising: The data processing module collects patients' raw data and preprocesses it to obtain standard data; The original model pre-training module generates patient-patient similarity graphs based on standard data as the original graphs. In the original image Build a graph neural network model and select the set of nodes to be deleted according to a predetermined method. Deleting edge set and the feature matrix of deleted nodes A complete graph is formed, and a graph neural network model is trained using the complete graph to obtain a baseline graph neural network model. Delete operation module, delete original image Delete node set The feature matrix of the nodes and their corresponding deleted nodes. and the set of deleted edges The edges in the graph are used to obtain the updated graph. ,in, To update the node set, To update the edge set, To update the node feature matrix; The high-influence region selection module calculates the change in node embedding and selects the updated node set. The third-order neighborhood subgraphs are used as candidate regions, and clustering methods are employed to determine high-influence regions. The graph forgetting and prediction balancing module constructs local forgetting loss functions, global forgetting loss functions, and prediction retention loss functions around the target node set, determines the total loss function, balances the prediction and forgetting losses through the Pareto balancing algorithm, and completes the training of the baseline graph neural network model through a specified number of iterations to obtain the graph forgetting learning model. The results output module deploys a graph forgetting learning model, constructs the original graph and extracts features from the new patient data, inputs the corresponding graph structure into the graph forgetting learning model, and outputs the prediction results. The data processing module, the original model pre-training module, the deletion operation module, the high-influence region selection module, the graph forgetting and prediction balance module, and the result output module are connected in sequence.
[0009] Optionally, in the graph neural network forgetting system that balances forgetting and performance retention in the above embodiments, the patient's original data includes the patient's basic characteristics, perioperative physiological indicators, medication records, and surgery type.
[0010] Optionally, in the graph neural network forgetting system that balances forgetting and performance preservation in any of the above embodiments, preprocessing includes processing missing values and outliers, and standardizing and synchronizing features of different dimensions.
[0011] Optionally, in the graph neural network forgetting system that balances forgetting and performance preservation in any of the above embodiments, the predetermined method is either randomly selected or manually specified.
[0012] Optionally, in the graph neural network forgetting system that balances forgetting and performance preservation in any of the above embodiments, the method for determining high-influence regions includes: S410. Calculate the node embedding change in the original graph. Deleted node feature matrix Inverting the matrix while keeping the other features unchanged yields the inverse feature matrix. The formula is as follows: , inverse eigenma Compared with the original image The adjacency matrix K Multiply by powers to obtain the baseline graph neural network model. K Layer message passing, the result is Then compare the node feature matrix X with the original graph. The adjacency matrix K Multiplying by powers gives the result of ,calculate and The cosine similarity is used to obtain the node embedding change. S420. Determine the changes in node embeddings in high-influence regions and select the updated node set. The third-order neighborhood subgraph is used as the candidate region, and the node embedding change of the candidate region is selected as the node embedding change of the high-influence region. S430. Select cluster centers. Select the maximum value of the node embedding variation as the high-impact cluster center and the minimum value of the node embedding variation as the low-impact cluster center. S440. Node classification: Traverse each node in the candidate region and subtract its node embedding change from the high-influence cluster center and the low-influence cluster center respectively. When the difference between the node embedding change and the high-influence cluster center is less than the difference between the node embedding change and the low-influence cluster center, classify the node as a high-influence class; otherwise, classify it as a low-influence class. S450, Iterative calculation: take the average value of the embedding change of nodes in the high-impact class as the new high-impact cluster center, and take the average value of the embedding change of nodes in the low-impact class as the new low-impact cluster center. Return to S440 for iteration until the high-impact cluster center and the low-impact cluster center no longer change. S460. Determine high-influence regions by taking the average of the high-influence cluster centers and the low-influence cluster centers as a threshold. The set of nodes whose node embedding changes exceed the threshold is defined as a high-influence region.
[0013] Preferably, in the graph neural network forgetting system that balances forgetting and performance preservation in the above embodiments, K =5.
[0014] Optionally, in the graph neural network forgetting system that balances forgetting and performance preservation in any of the above embodiments, the number of rounds is specified as 1000.
[0015] Optionally, in the graph neural network forgetting system that balances forgetting and performance retention in any of the above embodiments, the prediction results include ASA level, risk of complications, and warning of abnormal blood pressure.
[0016] Based on any of the above embodiments, in another embodiment of the present invention, a graph neural network forgetting method that balances forgetting and performance preservation is provided, comprising the following steps: S100. Obtain the patient's raw data, collect the patient's raw data, preprocess it, and construct a patient-patient similarity graph as the original graph. , obtain the node set V Edge set E and node feature matrix X ; S200, Training baseline graph neural network model, in the original graph A graph neural network model is built on top of the data, the dataset is divided according to a specified ratio, and the anesthesia risk rating in the dataset is used as the prediction label for supervised learning to train the graph neural network model and obtain the baseline graph neural network model. S300, Calculate and update the graph, in the original graph The set of nodes to be deleted is selected according to a predetermined method. In response to the deletion operation, the baseline graph neural network model deletes the original graph. Delete node set The feature matrix of the nodes and their corresponding deleted nodes. and the set of deleted edges The edges in the graph are used to obtain the updated graph. ,in, To update the node set, To update the edge set, To update the node feature matrix; S400: Calculate high-influence regions, calculate node embedding changes, and select the updated node set. The third-order neighborhood subgraphs are used as candidate regions, and clustering methods are employed to determine high-influence regions. S500, graph forgetting and prediction balance optimization, calculate local forgetting loss function, global forgetting loss function and prediction retention loss function, determine the total loss function, complete the training of the baseline graph neural network model through a specified number of iterations, and obtain the graph forgetting learning model; S600: Output prediction results. Deploy the graph forgetting learning model, construct the original graph and extract features from the new patient data, input the corresponding graph structure into the graph forgetting learning model, and output the prediction results.
[0017] Optionally, the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments further includes: S700, Model Evaluation: Record the changes in F1-Score, AUC (Area Under the Curve), AUP (Area Under the Precision Curve), and ACC (Accuracy) of the baseline graph neural network model and graph forgetting learning model to evaluate the graph forgetting learning model.
[0018] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the preprocessing includes processing missing values and outliers, and standardizing and synchronizing features of different dimensions.
[0019] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the method for constructing patient-patient similarity graphs includes clinical prior methods or similarity measurement methods.
[0020] Furthermore, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the similarity measurement method includes calculating the cosine similarity using the physiological characteristics of each pair of patients. When the cosine similarity exceeds the similarity threshold, it is considered that there is an edge connection between the pair of patients.
[0021] Preferably, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the similarity threshold is 0.6.
[0022] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, step S200 includes: S210. Build a graph neural network model, in the original graph... Build a graph neural network model, set the hidden dimensions, and determine the activation function; S220. Divide the dataset into training set, test set and validation set according to the specified ratio; S230. Train the graph neural network model by using the anesthesia risk rating in the dataset as the prediction label for supervised learning. After a set number of rounds, obtain the baseline graph neural network model.
[0023] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the graph neural network has a three-layer structure.
[0024] Furthermore, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the hidden dimensions of the graph neural network are 128, 64, and 64.
[0025] Furthermore, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the activation function is the sigmoid function.
[0026] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the graph neural network uses GCN (Graph Convolutional Network) or GAT (Graph Attention Network).
[0027] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the specified ratio is 8:1:1.
[0028] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the number of rounds is set to 2000.
[0029] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, the predetermined method is randomly selected or manually specified.
[0030] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, step S400 includes: S410. Calculate the node embedding change in the original graph. Deleted node feature matrix Inverting the matrix while keeping the other features unchanged yields the inverse feature matrix. The formula is as follows: , inverse eigenma Compared with the original image The adjacency matrix K Multiply by powers to obtain the baseline graph neural network model. K Layer message passing, the result is Then compare the node feature matrix X with the original graph. The adjacency matrix K Multiplying by powers gives the result of ,calculate and The cosine similarity is used to obtain the node embedding change. S420. Determine the changes in node embeddings in high-influence regions and select the updated node set. The third-order neighborhood subgraph is used as the candidate region, and the node embedding change of the candidate region is selected as the node embedding change of the high-influence region. S430. Select cluster centers. Select the maximum value of the node embedding variation as the high-impact cluster center and the minimum value of the node embedding variation as the low-impact cluster center. S440. Node classification: Traverse each node in the candidate region and subtract its node embedding change from the high-influence cluster center and the low-influence cluster center respectively. When the difference between the node embedding change and the high-influence cluster center is less than the difference between the node embedding change and the low-influence cluster center, classify the node as a high-influence class; otherwise, classify it as a low-influence class. S450, Iterative calculation: take the average value of the embedding change of nodes in the high-impact class as the new high-impact cluster center, and take the average value of the embedding change of nodes in the low-impact class as the new low-impact cluster center. Return to S440 for iteration until the high-impact cluster center and the low-impact cluster center no longer change. S460. Determine high-influence regions by taking the average of the high-influence cluster centers and the low-influence cluster centers as a threshold. The set of nodes whose node embedding changes exceed the threshold is defined as a high-influence region.
[0031] Preferably, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, K =5.
[0032] Optionally, in the graph neural network forgetting method that balances forgetting and performance preservation in any of the above embodiments, step S500 includes: S510. Determine the local forgetting loss function, and use the cosine similarity between the features of the two nodes connected by the deleted edge and the features of any two nodes in the original graph that are not connected by an edge as the local forgetting loss function. S520. Determine the global forgetting loss function. Use the cosine similarity of the features of all nodes in the second-order neighborhood of the two nodes connected by the deleted edge before and after the edge deletion operation as the global forgetting loss function to measure the degree of shift in the embedding distribution of the target node set before and after deletion. S530. Determine the prediction preservation loss function, calculate the task supervision loss on the target node set, and use the cross-entropy between the predicted and actual values of the node labels in high-influence regions as the prediction loss function. S540. Determine the total loss function, sum the local forgetting loss function and the global forgetting loss function by weight to obtain the forgetting loss function, adaptively determine the optimal weights of the forgetting loss function and the prediction loss function, and then sum them by weight to obtain the total loss function. S550, Model Training: Iterate through a specified number of rounds, backpropagate and update the model parameters, complete the training of the baseline graph neural network model through a specified number of iterations, achieve multi-objective optimization of forgetting and prediction, and obtain the graph forgetting learning model.
[0033] Furthermore, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the task supervision loss includes cross-entropy or mean squared error.
[0034] Preferably, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the number of rounds is specified as 1000.
[0035] Furthermore, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, step S540 includes: S541. Calculate the forgetting loss function, determine the weights of the local forgetting loss function and the global forgetting loss function, and then perform a weighted summation to obtain the forgetting loss function; S542. Calculate the weights. Adaptively determine the weights of the forgetting loss function and the prediction loss function using the Pareto balancing algorithm. The weights of the forgetting loss function are: a The weights of the prediction loss function are 1-a The calculation formula is as follows: , in, G1 The gradient matrix of the forgetting loss function and the weight parameters of the baseline graph neural network model is given. G2 The gradient matrix of the weight parameters of the baseline graph neural network model for predicting the loss function; S543. Calculate the total loss function by weighted summation of the forgetting loss function and the prediction loss function to obtain the total loss function.
[0036] Preferably, in the graph neural network forgetting method that balances forgetting and performance preservation in the above embodiments, the weights of both the local forgetting loss function and the global forgetting loss function are 0.5.
[0037] This application approximates the message passing of nonlinear graph neural networks by using linear adjacency matrix multiplication, thereby reducing the consumption of computational resources. It combines the two objectives of eliminating the impact of deletion and maintaining the performance of the remaining parts. Based on the Pareto balance method, it balances the loss function of the two, which greatly reduces the performance degradation of the model after deletion. It can maintain the model performance with very few computational resources, avoids retraining the entire model, reduces the consumption of computational resources, greatly improves the preservation effect of model performance, and has wide applicability.
[0038] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of the structure of a graph neural network forgetting system that balances forgetting and performance retention, as an exemplary embodiment. Figure 2 This is a flowchart of a graph neural network forgetting method that balances forgetting and performance preservation, as an exemplary embodiment. Detailed Implementation
[0040] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.
[0041] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of components is schematically exaggerated in some places in the drawings.
[0042] The inventors designed a graph neural network forgetting system that balances forgetting and performance retention, such as... Figure 1 As shown, it includes: The data processing module collects patients' raw data, including patients' basic characteristics, perioperative physiological indicators, medication records, and surgical types. It performs preprocessing, including handling missing and outlier values, standardizing and synchronizing features of different dimensions to obtain standardized data. The original model pre-training module generates patient-patient similarity graphs based on standard data as the original graphs. In the original image Build a graph neural network model and select the set of nodes to be deleted according to human specifications. Deleting edge set and the feature matrix of deleted nodes A complete graph is formed, and the graph neural network model is trained using the complete graph to obtain the baseline graph neural network model.
[0043] Delete operation module, delete original image Delete node set The feature matrix of the nodes and their corresponding deleted nodes. and the set of deleted edges The edges in the graph are used to obtain the updated graph. ,in, To update the node set, To update the edge set, To update the node feature matrix.
[0044] The high-influence region selection module calculates the change in node embedding and selects the updated node set. The third-order neighborhood subgraph is used as a candidate region, and clustering methods are employed to determine high-influence regions. Methods for determining high-influence regions include: S410. Calculate the node embedding change in the original graph. Deleted node feature matrix Inverting the matrix while keeping the other features unchanged yields the inverse feature matrix. The formula is as follows: , inverse eigenma Compared with the original image The adjacency matrix K Multiply by powers to obtain the baseline graph neural network model. K Layered message passing, K =5, the result is Then compare the node feature matrix X with the original graph. The adjacency matrix K Multiplying by powers gives the result of ,calculate and The cosine similarity is used to obtain the node embedding change. S420. Determine the changes in node embeddings in high-influence regions and select the updated node set. The third-order neighborhood subgraph is used as the candidate region, and the node embedding change of the candidate region is selected as the node embedding change of the high-influence region. S430. Select cluster centers. Select the maximum value of the node embedding variation as the high-impact cluster center and the minimum value of the node embedding variation as the low-impact cluster center. S440. Node classification: Traverse each node in the candidate region and subtract its node embedding change from the high-influence cluster center and the low-influence cluster center respectively. When the difference between the node embedding change and the high-influence cluster center is less than the difference between the node embedding change and the low-influence cluster center, classify the node as a high-influence class; otherwise, classify it as a low-influence class. S450, Iterative calculation: take the average value of the embedding change of nodes in the high-impact class as the new high-impact cluster center, and take the average value of the embedding change of nodes in the low-impact class as the new low-impact cluster center. Return to S440 for iteration until the high-impact cluster center and the low-impact cluster center no longer change. S460. Determine high-influence regions by taking the average of the high-influence cluster centers and the low-influence cluster centers as a threshold. The set of nodes whose node embedding changes exceed the threshold is defined as a high-influence region.
[0045] The graph forgetting and prediction balancing module constructs local forgetting loss functions, global forgetting loss functions, and prediction retention loss functions around the target node set, determines the total loss function, balances prediction and forgetting losses using the Pareto balancing algorithm, and completes the training of the baseline graph neural network model through 1000 iterations to obtain the graph forgetting learning model.
[0046] The results output module deploys a graph forgetting learning model, constructs the original graph and extracts features from new patient data, inputs the corresponding graph structure into the graph forgetting learning model, and outputs prediction results, including ASA level, complication risk, and blood pressure abnormality warning.
[0047] The data processing module, the original model pre-training module, the deletion operation module, the high-influence region selection module, the graph forgetting and prediction balance module, and the result output module are connected in sequence.
[0048] Based on the above embodiments, the inventors provide a method-based approach, such as... Figure 2 As shown, it includes the following steps: A graph neural network forgetting method that balances forgetting and performance preservation is provided, comprising the following steps: S100. Obtain raw patient data, collect raw patient data, and perform preprocessing, including handling missing and outlier values, standardizing and synchronizing features of different dimensions to obtain standardized data, and constructing a patient-patient similarity graph as the original graph using a similarity metric. , obtain the node set V Edge set E and node feature matrix X The similarity measurement method includes calculating the cosine similarity using the physiological characteristics of each pair of patients. When the cosine similarity exceeds the similarity threshold of 0.6, it is considered that there is an edge connection between the pair of patients.
[0049] S200, Training baseline graph neural network model, in the original graph A graph neural network model is built by dividing the dataset according to a specified ratio, using the anesthesia risk rating in the dataset as the prediction label for supervised learning, training the graph neural network model, and obtaining the baseline graph neural network model; specifically including: S210. Build a graph neural network model, in the original graph... A graph neural network model was built on top of this. The graph neural network is a three-layer GCN (graph convolutional network). The hidden dimensions were set to 128, 64, and 64, and the activation function was determined to be the sigmoid function. S220. Divide the dataset into training set, test set and validation set according to the specified ratio of 8:1:1. S230. Train the graph neural network model. Use the anesthesia risk rating in the dataset as the prediction label for supervised learning. Train the graph neural network model and obtain the baseline graph neural network model after a set number of 2000 rounds.
[0050] S300, Calculate and update the graph, in the original graph The set of nodes to be deleted is selected according to human specifications. In response to the deletion operation, the baseline graph neural network model deletes the original graph. Delete node set The feature matrix of the nodes and their corresponding deleted nodes. and the set of deleted edges The edges in the graph are used to obtain the updated graph. ,in, To update the node set, To update the edge set, To update the node feature matrix; S400: Calculate high-influence regions, calculate node embedding changes, and select the updated node set. Using the third-order neighborhood subgraph as candidate regions, a clustering method is employed to determine high-influence regions; specifically including: S410. Calculate the node embedding change in the original graph. Deleted node feature matrix Inverting the matrix while keeping the other features unchanged yields the inverse feature matrix. The formula is as follows: , inverse eigenma Compared with the original image The adjacency matrix K Multiply by powers to obtain the baseline graph neural network model. K Layered message passing, K =5, the result is Then compare the node feature matrix X with the original graph. The adjacency matrix K Multiplying by powers gives the result of ,calculate and The cosine similarity is used to obtain the node embedding change. S420. Determine the changes in node embeddings in high-influence regions and select the updated node set. The third-order neighborhood subgraph is used as the candidate region, and the node embedding change of the candidate region is selected as the node embedding change of the high-influence region. S430. Select cluster centers. Select the maximum value of the node embedding variation as the high-impact cluster center and the minimum value of the node embedding variation as the low-impact cluster center. S440. Node classification: Traverse each node in the candidate region and subtract its node embedding change from the high-influence cluster center and the low-influence cluster center respectively. When the difference between the node embedding change and the high-influence cluster center is less than the difference between the node embedding change and the low-influence cluster center, classify the node as a high-influence class; otherwise, classify it as a low-influence class. S450, Iterative calculation: take the average value of the embedding change of nodes in the high-impact class as the new high-impact cluster center, and take the average value of the embedding change of nodes in the low-impact class as the new low-impact cluster center. Return to S440 for iteration until the high-impact cluster center and the low-impact cluster center no longer change. S460. Determine high-influence regions by taking the average of the high-influence cluster centers and the low-influence cluster centers as a threshold. The set of nodes whose node embedding changes exceed the threshold is defined as a high-influence region.
[0051] S500, graph forgetting and prediction balance optimization, calculates the local forgetting loss function, global forgetting loss function, and prediction retention loss function, determines the total loss function, and completes the training of the baseline graph neural network model through a specified number of iterations to obtain the graph forgetting learning model; specifically including: S510. Determine the local forgetting loss function, and use the cosine similarity between the features of the two nodes connected by the deleted edge and the features of any two nodes in the original graph that are not connected by an edge as the local forgetting loss function. S520. Determine the global forgetting loss function. Use the cosine similarity of the features of all nodes in the second-order neighborhood of the two nodes connected by the deleted edge before and after the edge deletion operation as the global forgetting loss function to measure the degree of shift in the embedding distribution of the target node set before and after deletion. S530. Determine the prediction preservation loss function, calculate the task supervision loss on the target node set, i.e., cross-entropy or mean squared error, and use the cross-entropy between the predicted value and the true value of the node label in the high-influence region as the prediction loss function. S540. Determine the total loss function. Summate the local forgetting loss function and the global forgetting loss function with weights to obtain the total forgetting loss function. Adaptively determine the optimal weights for the forgetting loss function and the prediction loss function. The weights for both the local and global forgetting loss functions are 0.5. Then, sum them with weights to obtain the total loss function. Specifically, this includes: S541. Calculate the forgetting loss function, determine the weights of the local forgetting loss function and the global forgetting loss function, and then perform a weighted summation to obtain the forgetting loss function; S542. Calculate the weights. Adaptively determine the weights of the forgetting loss function and the prediction loss function using the Pareto balancing algorithm. The weights of the forgetting loss function are: a The weights of the prediction loss function are 1-a The calculation formula is as follows: , in, G1 The gradient matrix of the forgetting loss function and the weight parameters of the baseline graph neural network model is given. G2 The gradient matrix of the weight parameters of the baseline graph neural network model for predicting the loss function; S543. Calculate the total loss function by weighted summation of the forgetting loss function and the prediction loss function to obtain the total loss function.
[0052] S550, Model Training: Iterate through a specified number of rounds, backpropagate and update the model parameters, and complete the training of the baseline graph neural network model through 1000 iterations. This achieves multi-objective optimization of forgetting and prediction, resulting in a graph forgetting learning model.
[0053] S600: Output prediction results. Deploy the graph forgetting learning model, construct the original graph and extract features from the new patient data, input the corresponding graph structure into the graph forgetting learning model, and output the prediction results.
[0054] S700, Model Evaluation: Record the changes in F1-Score, AUC (Area Under the Curve), AUP (Area Under the Precision Curve), and ACC (Accuracy) of the baseline graph neural network model and graph forgetting learning model to evaluate the graph forgetting learning model.
[0055] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A graph neural network forgetting system that balances forgetting and performance retention, characterized in that, include: The data processing module collects patients' raw data and preprocesses it to obtain standard data; The original model pre-training module generates patient-patient similarity maps as the original maps based on the standard data. Build a graph neural network model and select the set of nodes to be deleted according to a predetermined method. Deleting edge set and the feature matrix of deleted nodes A complete graph is formed, and the graph neural network model is trained to obtain a baseline graph neural network model. The deletion operation module deletes the set of nodes to be deleted. The nodes in the matrix and their corresponding feature matrices of the deleted nodes and the set of deleted edges The edges in the graph are used to obtain the updated graph. ; The high-influence region selection module calculates the node embedding change and selects the updated node set. The third-order neighborhood subgraphs are used as candidate regions, and clustering methods are employed to determine high-influence regions. The graph forgetting and prediction balancing module determines the total loss function, balances the prediction and forgetting losses using the Pareto balancing algorithm, and completes the training of the baseline graph neural network model through a specified number of iterations to obtain the graph forgetting learning model. The results output module deploys the graph forgetting learning model, constructs the original graph and extracts features from the new patient data, inputs the corresponding graph structure into the graph forgetting learning model, and outputs the prediction results. The data processing module, the original model pre-training module, the deletion operation module, the high-influence region selection module, the graph forgetting and prediction balance module, and the result output module are sequentially connected in communication.
2. The graph neural network forgetting system that balances forgetting and performance retention as described in claim 1, characterized in that, The patient's raw data includes the patient's basic characteristics, perioperative physiological indicators, medication records, and surgical type.
3. The graph neural network forgetting system that balances forgetting and performance retention as described in claim 1, characterized in that, The predetermined method can be either randomly selected or manually specified.
4. The graph neural network forgetting system that balances forgetting and performance retention as described in claim 1, characterized in that, The preprocessing includes handling missing values and outliers, and standardizing and synchronizing features of different dimensions over time.
5. A graph neural network forgetting method that balances forgetting and performance preservation, using the graph neural network forgetting system that balances forgetting and performance preservation as described in any one of claims 1-4, characterized in that, Includes the following steps: S100. Obtain the patient's raw data, collect the patient's raw data, preprocess it, and construct a patient-patient similarity graph as the original graph. , obtain the node set V Edge set E and node feature matrix X ; S200, Training the baseline graph neural network model, in the original graph A graph neural network model is built on the dataset, which is divided into a specified proportion. The anesthesia risk rating in the dataset is used as a prediction label for supervised learning to train the graph neural network model and obtain a baseline graph neural network model. S300, Calculate and update the graph in the original graph. The set of nodes to be deleted is selected according to a predetermined method. In response to the deletion operation, delete the set of nodes to be deleted. The feature matrix of the nodes and their corresponding deleted nodes. and the set of deleted edges The edges in the graph are used to obtain the updated graph. ; S400. Calculate the high-influence region, calculate the node embedding change, and select the updated node set. The third-order neighborhood subgraphs are used as candidate regions, and clustering methods are employed to determine high-influence regions. S500, graph forgetting and prediction balance optimization, determine the total loss function, and complete the training of the baseline graph neural network model through a specified number of iterations to obtain the graph forgetting learning model; S600. Output the prediction result. Deploy the graph forgetting learning model, construct the original graph and extract features from the new patient data, input the corresponding graph structure into the graph forgetting learning model, and output the prediction result.
6. The graph neural network forgetting method that balances forgetting and performance preservation as described in claim 5, characterized in that, Also includes: S700. Model evaluation: Record the changes in F1-Score, AUC, AUP, and ACC of the baseline graph neural network model and the graph forgetting learning model, and evaluate the graph forgetting learning model.
7. The graph neural network forgetting method that balances forgetting and performance preservation as described in claim 5, characterized in that, Step S200 includes: S210. Construct a graph neural network model, in the original graph... Build a graph neural network model, set the hidden dimensions, and determine the activation function; S220. Divide the dataset into training set, test set and validation set according to the specified ratio; S230. Train the graph neural network model by using the anesthesia risk rating in the dataset as the prediction label for supervised learning, and obtain the baseline graph neural network model after a set number of rounds.
8. The graph neural network forgetting method that balances forgetting and performance preservation as described in claim 5, characterized in that, Step S400 includes: S410. Calculate the node embedding change in the original graph. The feature matrix of the deleted node Inverting the matrix while keeping the other features unchanged yields the inverse feature matrix. The formula is as follows: ， The inverse feature matrix With the original image The adjacency matrix K Multiply by powers to perform the baseline graph neural network model. K Layer message passing, the result is Then compare the node feature matrix X with the original graph. The adjacency matrix K Multiplying by powers gives the result of ,calculate and The cosine similarity is used to obtain the node embedding change. S420. Determine the node embedding changes in high-influence regions and select the updated node set. The third-order neighborhood subgraph is used as a candidate region, and the node embedding change of the candidate region is selected as the node embedding change of the high-influence region. S430. Select cluster centers. Select the maximum value of the node embedding variation as the high-impact cluster center and the minimum value of the node embedding variation as the low-impact cluster center. S440. Node classification: Traverse each node in the candidate region and subtract its node embedding change from the high-influence cluster center and the low-influence cluster center respectively. When the difference between the node embedding change and the high-influence cluster center is less than the difference between the node embedding change and the low-influence cluster center, the node is classified as a high-influence class; otherwise, it is classified as a low-influence class. S450. Iterative calculation: take the average value of the node embedding change of the high-impact class as the new high-impact cluster center, take the average value of the node embedding change of the low-impact class as the new low-impact cluster center, return to step S440 for iteration, until the high-impact cluster center and the low-impact cluster center no longer change. S460. Determine high-influence regions. Take the average value of the high-influence cluster centers and the low-influence cluster centers as a threshold. The set of nodes whose node embedding changes exceed the threshold is defined as a high-influence region.
9. The graph neural network forgetting method that balances forgetting and performance preservation as described in claim 5, characterized in that, Step S500 includes: S510. Determine the local forgetting loss function, and use the cosine similarity between the features of the two nodes connected by the deleted edge and the features of any two nodes in the original graph that are not connected by an edge as the local forgetting loss function. S520. Determine the global forgetting loss function. Use the cosine similarity of the features of all nodes in the second-order neighborhood of the two nodes connected by the deleted edge before and after the edge deletion operation as the global forgetting loss function to measure the degree of shift in the embedding distribution of the target node set before and after deletion. S530. Determine the prediction preservation loss function, calculate the task supervision loss on the target node set, and use the cross-entropy between the predicted and actual values of the node labels in high-influence regions as the prediction loss function. S540. Determine the total loss function, sum the local forgetting loss function and the global forgetting loss function by weight to obtain the forgetting loss function, adaptively determine the optimal weights of the forgetting loss function and the prediction loss function, and then sum them by weight to obtain the total loss function. S550. Model training: Iterate through a specified number of rounds, backpropagate and update the model parameters, and complete the training of the baseline graph neural network model through a specified number of iterations to achieve multi-objective optimization of forgetting and prediction, thereby obtaining a graph forgetting learning model.
10. The graph neural network forgetting method that balances forgetting and performance preservation as described in claim 9, characterized in that, Step S540 includes: S541. Calculate the forgetting loss function, determine the weights of the local forgetting loss function and the global forgetting loss function, and then perform a weighted summation to obtain the forgetting loss function; S542. Calculate the weights. Adaptively determine the weights of the forgetting loss function and the prediction loss function using the Pareto balancing algorithm. The weights of the forgetting loss function are: a The weights of the prediction loss function are 1-a The calculation formula is as follows: ， in, G1 The gradient matrix of the forgetting loss function and the weight parameters of the baseline graph neural network model is given. G2 The gradient matrix of the prediction loss function and the weight parameters of the baseline graph neural network model; S543. Calculate the total loss function by weighted summation of the forgetting loss function and the prediction loss function to obtain the total loss function.