An Incremental Federated Causal Structure Learning Method for Dynamic Scenarios

Abstract

This invention discloses an incremental federated causal structure learning method in dynamic scenarios, involving the interdisciplinary field of computer causal inference and federated learning. The method includes the following steps: S1: Historical federated causal structure learning; S2: Adaptive selection of new clients; S3: Weighted federated causal structure learning. A weighted federated causal structure learning mechanism is designed to achieve efficient collaborative incremental learning between the server and high-quality, non-redundant new clients. This invention uses an improved clustering method and a multi-dimensional quality assessment strategy to select high-quality nodes from new clients. Among these high-quality clients, redundant new clients with historical causal structures similar to those on the server are identified to reduce redundant computation. Weighted federated aggregation learning is then performed on the selected new clients based on the historical causal structure, eliminating the need for full learning of both historical and new clients, effectively reducing communication overhead while maintaining learning accuracy.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computer causal inference and federated learning, specifically to an incremental federated causal structure learning method in dynamic scenarios. Background Technology

[0002] Causal structure learning aims to reveal the causal relationships and dependencies among a set of random variables, providing a structured foundation for causal inference. Federated causal structure learning, based on the federated learning paradigm, accurately infers the causal structure graph (usually a directed acyclic graph, DAG) between variables through interactive learning with the server without disclosing the original data of each party (client). Existing federated causal structure learning methods have made significant progress, but all studies are based on the static client assumption, requiring all participants to maintain a synchronous learning process. In real-world scenarios such as medical collaboration networks, new clients may continuously join the learning process, forming a typical dynamic scenario of "dynamic client access." Furthermore, client selection in federated learning directly affects the communication efficiency and performance stability of model training. Some existing federated learning frameworks propose selecting clients through random sampling, MD sampling, and cluster sampling. However, in scenarios with new clients, traditional federated causal structure learning and static client selection methods still face the following challenges: (1) Communication cost problem: In the face of this dynamic scenario, the traditional federated causal structure learning method requires repeated restarting of global model training, that is, rebuilding the common model on the basis of all participants including the newly added client in order to integrate the data of the new client. This not only wastes computing resources, but also leads to a sharp increase in communication costs.

[0003] (2) Static client selection methods are difficult to adapt to the diverse needs of different dynamic scenarios: The existing static client selection methods have not yet established a correlation analysis framework for the quality of historical and new client data, and the comprehensive evaluation strategy for client data quality is incomplete, which makes it difficult for the selection strategy to adapt to the diverse needs of different dynamic scenarios.

[0004] (3) The static client selection strategy is not comprehensive enough for the comprehensive evaluation of client data quality: In the scenario of new clients, the data characteristics of different new clients are significantly different. A single dimension will amplify the individual quality differences and cannot accurately reflect their true quality, thus becoming a key bottleneck restricting the learning effect of incremental federated causal structure.

[0005] (4) Federated causal learning lacks research on dynamic client access scenarios: Existing research on federated incremental learning focuses on expanding data samples or task categories, and cannot handle the dual quality constraints between new and existing clients when accessing dynamically, resulting in poor transfer performance of existing algorithms in complex dynamic environments.

[0006] Therefore, an incremental federated causal structure learning method for dynamic scenarios is needed to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide an incremental federated causal structure learning method in dynamic scenarios to solve the problems existing in the prior art mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An incremental federated causal structure learning method for dynamic scenarios includes the following steps: S1: Learning the historical federal causal structure: (1) Constructing historical federated causal structure DAGs: Based on the efficient causal structure learning algorithm F2SL, a federated F2SL is designed to learn the federated causal structure of historical clients and construct historical federated causal structure DAGs to provide a benchmark for subsequent incremental learning; (2) Calculate historical client ratings; S2: Added client-adaptive selection: (1) Adaptive clustering of new clients: The adaptive clustering strategy of new clients is adopted to solidify the sample quality distribution pattern of historical clients into a reference frame, so that new clients can complete feature matching and position anchoring within the frame, breaking through the limitations of the initial number of clusters, the scale of historical clients and the number of new clients. (2) Selection of high-quality clients: A multi-factor cross-client evaluation strategy is adopted to select high-quality clients within the cluster; (3) Redundant client removal: By evaluating the similarity between the local causal structure of high-quality clients and the historical causal structure of the server, the redundant clients can be accurately identified and removed; S3: Weighted Federated Causal Structure Learning: Design a weighted federated causal structure learning mechanism to achieve efficient collaborative incremental learning between the server and non-redundant clients; (1) Client-side federated causal structure learning: Selected high-quality, non-redundant new clients perform local-to-global federated causal structure learning based on the federated F2SL method to generate new federated causal structures. ; (2) Dynamic weight allocation: to accurately measure the newly added federal causal structure The relative influence of the server's historical causal structure model is used to assign an overall weight to the new client group. ; (3) Incremental update of federal causal structure.

[0009] Preferably, in step S1, the specific steps for calculating the historical client rating are as follows: Node-level similarity calculation: For each variable node Calculate the client SU vector and server SU vector The cosine similarity is calculated using the following formula:

[0010] in Represents the dot product of vectors, ∥ ∥ represents the L2 norm, and this similarity reflects the degree of consistency between the client's causal structure and the server's aggregated causal structure; Client-side rating weighted aggregation: Based on node weight set The client's overall evaluation score is obtained by weighted summation of the similarity values ​​of all nodes in the client. H ck :

[0011] This score provides a quantitative basis for the dynamic adjustment of client participation strategies during the incremental learning phase by quantifying the consistency between historical clients and historical federated causal structures.

[0012] Preferably, in step S2, the specific steps of the newly added client adaptive clustering are as follows: Pre-clustering of historical clients: A client selection algorithm based on sample size clustering is used to determine the clustering index of each client through the parameters a and b in the multinomial distribution ax+b; New Client-Side Clustering Matching: Based on historical client-side pre-clustering results, an adaptive expansion of the cluster structure is achieved through a dynamic range matching mechanism, proposing the following three matching rules: Range matching: If the sample quality of the newly added client is SQ(C) t + If a sample falls within the quality range of any historical cluster, it is assigned to the corresponding cluster. Cluster split: If the sample quality of a new client exceeds the maximum sample quality SQ of the historical clusters. max If the maximum quality of all historical client samples is reached, the cluster splitting mechanism is triggered to create a new independent cluster for it. Cluster merging: If the sample quality of a new client is lower than the minimum sample quality SQ of the historical clusters. min If the quality of all historical client samples is the minimum, then it is merged into the newly created low-quality client cluster.

[0013] Preferably, in step S2, the specific steps for selecting high-quality clients are as follows: Evaluation Dimension Design: A new client evaluation vector is designed within the cluster to achieve accurate client selection. The difference between the data distribution and the BLC is quantified by KL divergence. The smaller the KL divergence, the more similar the distributions are. Multidimensional similarity fusion calculation: Based on three-dimensional features, the cosine similarity between the client and BLC feature vectors within each cluster is calculated using the following formula:

[0014] in and They represent the client respectively. And the data sample size weights of BLC, and They represent the client respectively. And the random missing rate of BLC samples, Indicates the client The distribution difference between BLC and its own KL divergence is 0, meaning that the distribution is completely consistent. The closer the cosine similarity is to 1, the closer the client data features are to the historical federated causal structure. Selection strategy design: Based on the similarity evaluation results, the following scenario-based selection rules are used to determine the candidate high-quality client set cah_clients. If a cluster contains only 1 new client, it is directly selected into cah_clients; if the number of new clients in a cluster is greater than 1, they are sorted in descending order of cosine similarity, and the new client with the highest similarity in the cluster is selected to be added to the candidate client set cah_clients.

[0015] Preferably, in step S2, the specific steps for filtering redundant clients are as follows: The similarity evaluation index SK is calculated based on the federated F2SL framework. Symmetric uncertainty (SU) is used as the metric to characterize the degree of association between nodes in the causal structure, and the similarity evaluation index SK is constructed as follows:

[0016] in, This is the set of important nodes selected by the set of node weights. V represents the local importance of the client. i The cosine similarity between the SU value and the corresponding node SU value in the server federated model; Define rules for removing redundant clients: if the similarity evaluation metric SK of a client exceeds a preset threshold, it is identified as a redundant client and removed; if all candidate new clients are removed after redundancy screening, the server immediately terminates the current incremental learning, outputs the current federated model results and complete time statistics, and avoids invalid communication and aggregation operations.

[0017] Preferably, in step S3, the dynamic weight allocation... The calculation formula is as follows:

[0018] in, The number of new, high-quality, non-redundant clients is defined by n, which represents the initial total number of new clients. This weight reflects the proportion of contribution of the new client group to the federated model, avoiding excessive impact of data quality fluctuations from a single client on the federated structure.

[0019] Preferably, in step S3, the specific steps for incremental updating the federated causal structure are as follows: The server is based on historical federated causal structure DAGs and newly added federated causal structures. The federal causal model is updated using the following weighted aggregation formula:

[0020] in, This indicates the node V in the updated federated DAG. i With V j The strength of the causal relationship is defined, and its value ranges from [0,1]. and These represent the correlation strengths in the newly added client-side aggregated DAG and the original server-side aggregated DAG, respectively. This formula achieves efficient and accurate expansion of the federated causal structure by dynamically balancing the stability of the historical structure with the incremental value of the new data.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. In the client selection stage, the incremental federated causal structure learning framework of the present invention selects high-quality clients from the new clients through an improved clustering method and a multi-dimensional quality evaluation strategy. Furthermore, redundant new clients with similar historical causal structures to the server are identified among these candidate high-quality clients to reduce redundant computation. In the incremental weighted aggregation learning stage, weighted federated aggregation learning is performed on the selected high-quality non-redundant new clients based on the historical causal structure, so that the system does not need to learn all historical and new clients, effectively reducing communication overhead while maintaining learning accuracy.

[0022] 2. The adaptive client selection strategy proposed in this invention achieves adaptive clustering of new clients by introducing an improved clustering algorithm, enabling the framework to adaptively cluster in different dynamic scenarios. It combines multi-dimensional data quality assessment results to carry out intra-cluster client screening, adapting to the diverse needs of different dynamic scenarios.

[0023] 3. The client quality assessment strategy proposed in this invention comprehensively incorporates multi-dimensional assessment indicators such as sample size, data missing rate, and distribution heterogeneity. It can comprehensively and accurately quantify the data quality of new clients. At the same time, it combines the quality benchmark information of historical clients to screen high-quality new clients, enabling them to participate in the weighted federated aggregation learning process in collaboration with the server, effectively improving the efficiency and accuracy of learning.

[0024] 4. This invention optimizes the entire incremental learning process by strengthening the intrinsic correlation between the three stages of "historical federated causal structure learning, new client selection, and weighted federated causal structure learning," accurately weakens the overall difference in data quality between historical and new clients, reduces the interference of individual client quality differences on the federated learning process, and ultimately achieves dual optimization of the federated causal structure in terms of accuracy and efficiency. Attached Figure Description

[0025] Figure 1 This is an architectural diagram of the method of the present invention.

[0026] Figure 2 This invention adds a client-side adaptive selection core stage architecture diagram. Detailed Implementation

[0027] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0028] Please see Figure 1-2 The present invention provides the following technical solutions: An incremental federated causal structure learning method for dynamic scenarios includes the following steps: S1: Learning the historical federal causal structure: (1) Constructing historical federated causal structure DAGs: Based on the efficient causal structure learning algorithm F2SL, a federated F2SL is designed to learn the federated causal structure of historical clients and construct historical federated causal structure DAGs to provide a benchmark for subsequent incremental learning.

[0029] (2) Calculate historical client ratings. The specific steps for calculating historical client ratings are as follows: Node-level similarity calculation: For each variable node Calculate the client SU (i.e., symmetric uncertainty) vector. and server SU vector The cosine similarity is calculated using the following formula:

[0030] in Represents the dot product of vectors, ∥ ∥ represents the L2 norm, and this similarity reflects the degree of consistency between the client-side causal structure and the server-side aggregated causal structure.

[0031] Client-side rating weighted aggregation: Based on node weight set ( The client's overall evaluation score is obtained by weighted summing of the similarity values ​​of all nodes in the client. H ck :

[0032] This score provides a quantitative basis for the dynamic adjustment of client participation strategies during the incremental learning phase by quantifying the consistency between historical clients and historical federated causal structures.

[0033] S2: Added client-adaptive selection: (1) Adaptive Clustering for New Clients: In the scenario of new client access, if the selection operation is directly performed on the new client without associating it with the features of historical clients, the overall distribution difference between the new and old client groups may be overlooked, leading to a decrease in the stability of incremental learning of causal structure. Therefore, an adaptive clustering strategy for new clients is adopted to solidify the sample quality distribution pattern of historical clients into a reference framework, so that the new client can complete feature matching and position anchoring within this framework, thus overcoming the limitations of the initial cluster size, the size of historical clients, and the number of new clients.

[0034] The specific steps for adding client-side adaptive clustering are as follows: Pre-clustering of historical clients: Inspired by cluster sampling client selection algorithms in the federated learning field, a sample-size-based cluster sampling client selection algorithm is used to determine the clustering index of each client through the parameters a and b in the multinomial distribution ax+b.

[0035] New Client-Side Clustering Matching: Based on historical client-side pre-clustering results, an adaptive expansion of the cluster structure is achieved through a dynamic range matching mechanism, proposing the following three matching rules: Range matching: If the sample quality of the newly added client is SQ(C) t + If a sample falls within the quality range of any historical cluster, it is assigned to the corresponding cluster. Cluster split: If the sample quality of a new client exceeds the maximum sample quality SQ of the historical clusters. max If the maximum quality of all historical client samples is reached, the cluster splitting mechanism is triggered to create a new independent cluster for it. Cluster merging: If the sample quality of a new client is lower than the minimum sample quality SQ of the historical clusters. minIf the quality of all historical client samples is the minimum, then it is merged into the newly created low-quality client cluster.

[0036] (2) Selection of high-quality clients: A multi-factor cross-client evaluation strategy is adopted for the selection of high-quality clients. The specific steps for the selection of high-quality clients are as follows: Evaluation Dimension Design: The new client evaluation vector (sample size, missing rate, KL divergence) within the cluster is designed to achieve accurate client selection. The KL divergence (Kullback-Leibler divergence) is used to quantify the difference in data distribution between the client and the BLC (benchmark historical client with the highest H score). The smaller the KL divergence, the more similar the distribution.

[0037] Multidimensional similarity fusion calculation: Based on three-dimensional features, the cosine similarity between the client and BLC feature vectors within each cluster is calculated using the following formula:

[0038] in and They represent the client respectively. And the data sample size weights of BLC, and They represent the client respectively. And the random missing rate of BLC samples, Indicates the client The distribution difference between BLC and its own KL divergence is 0, meaning the distribution is completely consistent. The closer the cosine similarity is to 1, the closer the client data features are to the historical federated causal structure.

[0039] Selection strategy design: Based on the similarity evaluation results, the following scenario-based selection rules are used to determine the candidate high-quality client set cah_clients. If a cluster contains only 1 new client, it is directly selected into cah_clients; if the number of new clients in a cluster is greater than 1, they are sorted in descending order of cosine similarity, and the new client with the highest similarity in the cluster is selected to be added to the candidate client set cah_clients.

[0040] (3) Redundant Client Removal: The local causal structure of some candidate high-quality new clients highly overlaps with the existing federated causal structure of the server. Such clients participating in aggregation not only fail to provide effective incremental information but also generate redundant communication overhead. Therefore, this stage achieves accurate identification and removal of redundant clients by evaluating the similarity between the client's local causal structure and the server's historical causal structure; the specific steps for redundant client removal are as follows: The similarity evaluation index SK is calculated based on the federated F2SL framework. Symmetric uncertainty (SU) is used as the metric to characterize the degree of association between nodes in the causal structure, and the similarity evaluation index SK is constructed as follows:

[0041] in, This is the set of important nodes selected by the set of node weights. V represents the local importance of the client. i The cosine similarity between the SU value and the corresponding node SU value in the server federated model; Define rules for removing redundant clients: if the similarity evaluation metric SK of a client exceeds a preset threshold, it is identified as a redundant client and removed; if all candidate new clients are removed after redundancy screening, the server immediately terminates the current incremental learning, outputs the current federated model results and complete time statistics, and avoids invalid communication and aggregation operations.

[0042] S3: Weighted Federated Causal Structure Learning: After redundant client screening, directly applying equal weights for federated incremental learning makes it difficult to accurately match the differentiated influence of new clients and server models, easily leading to insufficient accuracy and stability of the federated causal structure aggregation. Therefore, this stage designs a weighted federated causal structure learning mechanism to achieve efficient collaborative incremental learning between the server and non-redundant clients. (1) Client-side federated causal structure learning: Selected high-quality, non-redundant new clients perform local-to-global federated causal structure learning based on the federated F2SL method to generate new federated causal structures. .

[0043] (2) Dynamic weight allocation: to accurately measure the newly added federal causal structure The relative influence of the server's causal structure model is used to assign an overall weight to the new client group. In dynamic weight allocation The calculation formula is as follows:

[0044] in, The number of new, high-quality, non-redundant clients is defined by n, which represents the initial total number of new clients. This weight reflects the proportion of contribution of the new client group to the federated model, avoiding excessive impact of data quality fluctuations from a single client on the federated structure.

[0045] (3) Incremental update of the federated causal structure. The specific steps of incremental update of the federated causal structure are as follows: The server is based on historical federated causal structure DAGs and newly added federated causal structures. The federal causal model is updated using the following weighted aggregation formula:

[0046] in, This indicates the node V in the updated federated DAG. i With V j The strength of the causal relationship is defined, and its value ranges from [0,1]. and These represent the correlation strengths in the newly added client-side aggregated DAG and the original server-side aggregated DAG, respectively. This formula achieves efficient and accurate expansion of the federated causal structure by dynamically balancing the stability of the historical structure with the incremental value of the new data.

[0047] Example 1. Learning about the causal structure of historical federation: (1) Federated F2SL Feature Selection (Skeleton Construction): The client calculates the local SU value (i.e., symmetric uncertainty) of the feature node and reports it. The server aggregates the SU values ​​of each client and calculates the set of federated PCs (i.e., parent and child nodes). The client constructs a node correlation matrix based on the server information. The server identifies and removes redundant features through the aggregation matrix to generate the final undirected causal skeleton.

[0048] (2) Federation F2SL Conflict Resolution (Direction): Based on the local skeleton, each client uses the Bayesian climbing algorithm to perform heuristic direction and generate a directed graph. The server collects the direction results from all parties, performs weighted voting to unify the edge direction, and outputs the final historical federated causal structure graph DAG. s .

[0049] Input: Multiple Sachs datasets (different structured values ​​represent different client data, with data randomly missing). This dataset represents a baseline graphical model of the protein signaling network in human cells, containing 11 nodes (cell types) and 17 edges, used to verify the applicability of the method of this invention in real-world scenarios.

[0050] Output: Historical Federation Causal Structure Graph (DAG) s Historical client ratings

[0051] Evaluation metrics compared to a true-graph: F1 = 0.7059 (the higher the better) SHD = 8 (the smaller the better) Client evaluation metric scores (descending order): 0.9983 0.9973 0.9971 0.9956 0.9956 0.9955 0.9954 0.9950 0.9938 0.9936 0.9933 0.9931 Client score sorting: C 6， C 3， C 7， C 4， C 8， C 11， C 1， C 10， C 5， C 9， C 12， C2 2. Added client-side adaptive selection 2.1 Added client-side adaptive clustering (1) Pre-clustering of historical clients: The client selection algorithm based on sample size is used to determine the clustering index of each client by using parameters a and b in the multinomial distribution ax+b. This method first traverses all historical clients and accumulates the total number of samples, and then decomposes the number of samples of each client into the sum of integer multiples of the total number of samples and the remainder. The clustering index of each client is determined by using parameters a and b in the multinomial distribution ax+b.

[0052] Input: Historical client sample quality set Output: Historical client cluster set Cluster1: C 3， C 8， C6 Cluster2: C1, C7 Cluster3: C5 Cluster4: C 9， C 10 Cluster5: C 2， C 5, C 11， C 12 (2) Add client-side adaptive clustering: Design three matching rules: "range matching, cluster splitting, and cluster merging" for matching.

[0053] Code snippet: Input: Historical client cluster set Cluster={Cluster1,...,Cluster...} p Historical and new client sample quality SQ and SQ + Output: New client cluster set (Cluster) + ={Cluster + 1,...,Cluster + q} #Calculate the maximum and minimum sample quality across all historical clients for k = 1 to m do SQ max ← Max(SQ max ,SQ(C k )) SQ min ← Min(SQ max ,SQ(C k )) end for Cluster + min ← []; for t = 1 to n do #The quality of newly added client samples is greater than the maximum historical client sample quality (cluster split) if SQ(C t ) + >SQ max then Cluster + j ← []; / / j=1...q Cluster + j ← c + t ; #The quality of newly added client samples is less than the lowest historical client sample quality (cluster merging) else if SQ(C t ) + <SQ min then Cluster + min ←c + t ; / / 1 <min<p #Sample quality matches historical cluster range (range matching) else if SQ min <SQ(C t ) + <SQ max then SQ hmax =max c∈Clusterh SQ(C) SQ h min =min c∈Clusterh SQ(C) if SQ h min <SQ(C t ) + <SQ h max then Cluster h ←c + t ; / / 1 <h<p Cluster + j ← Cluster h ; end if end if end for Return Cluster + ={Cluster + 1,...,Cluster + q} 2.2 High-quality client selection: Based on the evaluation vector (sample size, missing rate, KL divergence), the similarity between the BLC (benchmark historical clients) and the new clients is calculated. For each cluster... + ={Cluster + 1,...,Cluster + q} =The client selects and obtains a candidate set of new clients, cah_clients.

[0054] Code snippet: Input: Add a new client cluster set (Cluster) + ={Cluster + 1,...,Cluster + q Add a new client dataset D+={ Historical and new client sample size Q; historical and new client missing rate M; baseline historical client BLC Output: A set of candidate high-quality clients, cah_clients for i = 1 to |Cluster + | do #If there is only one client in the cluster, add it to the candidate high-quality client set. if |Cluster + i | == 1 then cah_clients ← D + t else #Calculate the similarity between clients within the cluster and BLC across three dimensions, and then filter them. for t = 1 to n do Calculate the probability distribution of c + t pd_ct Calculate the probability distribution of BLC: pd_BLC Calculate KL betweenc + t and baseline client:

[0055] Calculate Similarity Score Sim(c + t , BLC) in (Q,1 M,KL): end for c + best ← arg max(Sim(c + BLC)) cah_clients ← c + best end if end for return cah_clients Input: Add a new client cluster collection (Cluster) + ={Cluster + 1,...,Cluster + q Add a new client dataset D+={ Historical and new client sample size Q; historical and new client missing rate M; baseline historical client BLC Output: A set of candidate high-quality clients, cah_clients Cluster + 1:C + 11 Cluster + 2:C + 6 Cluster + 3:C + 1,C + 2,C + 7,C + 9,C + 14 (Sort C) + 14 C + 1,C + 2,C + 7,C + 9) Cluster + 4:C + 3,C + 4,C + 5,C + 8,C + 10 C + 12 C + 13 C + 15 (Sort C) + 15 C + 12 C + 3,C + 14 C + 4,C + 10 C + 8,C + 5C + 13 ) cah_clients:C + 11、 C + 6、 C + 14、 C + 15 2.3 Redundant Client Removal: Redundancy determination is performed on candidate high-quality clients: the similarity between their local structure and the server's historical aggregation structure is calculated, and redundant clients with excessive similarity are removed (their information is already contained in the existing global structure, and the gain is limited), reducing communication and computing overhead, and finally obtaining a set of high-quality non-redundant new clients.

[0056] Input: A set of candidate high-quality new clients, cah_clients: C + 11、 C + 6、 C + 14、 C + 15 Output: High-quality collection of newly added clients h_clients: C + 11、 C + 6、 C + 15 3. Weighted Federated Causal Structure Learning The server is based on historical federated causal structure DAGs and newly added federated causal structure DAGs. + The federated causal model is updated by setting weights based on the number of new clients participating in the learning process and performing weighted aggregation.

[0057] Input: High-quality set of newly added clients h_client, historical federated causal DAGs, weight of newly added clients: 0.2 (3 / 15) Output: Incremental Federated Causal Model DAG = [ 0 0 0 0 1 0 0 0 0 0 0; 1 0 0 0 0 0 0 0 0 0; 0 0 0 1 0 0 0 0 0 0 0; 1 1 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 1 1 1 1 1 0 0 0 0 0 0; 0 0 1 1 1 0 0 1 0 0 0; 0 0 0 0 0 1 1 0 0 0 0; 0 0 0 1 0 0 0 1 1 0 0 ]; F1 = 0.7222 SHD= 7 Time consuming = 0.6418s Results of traditional non-incremental causal structure learning (full-scale learning): DAG = [ 0 0 0 0 1 0 0 0 0 0 0; 1 0 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 1 1 1 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 1 0 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0 0 0 0; 1 1 1 1 1 0 0 0 0 0 0; 1 0 1 1 1 0 0 1 0 0 0; 0 0 0 0 0 1 1 0 0 0 0; 0 1 0 1 0 0 0 1 1 0 0 ]; F1 = 0.6154 SHD= 11 Time consuming = 1.3899s Compared with traditional non-incremental causal structure learning, the incremental federated causal structure learning framework of this invention improves both accuracy and communication time.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An incremental federated causal structure learning method for dynamic scenarios, characterized in that, Includes the following steps: S1: Learning the historical federal causal structure: (1) Constructing historical federated causal structure DAGs: Based on the efficient causal structure learning algorithm F2SL, a federated F2SL is designed to learn the federated causal structure of historical clients and construct historical federated causal structure DAGs to provide a benchmark for subsequent incremental learning; (2) Calculate historical client ratings; S2: Added client-adaptive selection: (1) Adaptive clustering of new clients: The adaptive clustering strategy of new clients is adopted to solidify the sample quality distribution pattern of historical clients into a reference frame, so that new clients can complete feature matching and position anchoring within the frame, breaking through the limitations of the initial number of clusters, the scale of historical clients and the number of new clients. (2) Selection of high-quality clients: A multi-factor cross-client evaluation strategy is adopted to select high-quality clients within the cluster; (3) Redundant client removal: By evaluating the similarity between the local causal structure of high-quality clients and the historical causal structure of the server, the redundant clients can be accurately identified and removed; S3: Weighted Federated Causal Structure Learning: Design a weighted federated causal structure learning mechanism to achieve efficient collaborative incremental learning between the server and non-redundant clients; (1) Client-side federated causal structure learning: Selected high-quality, non-redundant new clients perform local-to-global federated causal structure learning based on the federated F2SL method to generate new federated causal structures. ; (2) Dynamic weight allocation: to accurately measure the newly added federal causal structure The relative influence of the server's historical causal structure model is used to assign an overall weight to the new client group. ; (3) Incremental update of federal causal structure.

2. The incremental federated causal structure learning method in a dynamic scenario according to claim 1, characterized in that, In step S1, the specific steps for calculating historical client ratings are as follows: Node-level similarity calculation: For each variable node Calculate the client SU vector and server SU vector The cosine similarity is calculated using the following formula: in Represents the dot product of vectors, ∥ ∥ represents the L2 norm, and this similarity reflects the degree of consistency between the client's causal structure and the server's aggregated causal structure; Client-side rating weighted aggregation: Based on node weight set The client's overall evaluation score is obtained by weighted summation of the similarity values ​​of all nodes in the client. H ck : This score provides a quantitative basis for the dynamic adjustment of client participation strategies during the incremental learning phase by quantifying the consistency between historical clients and historical federated causal structures.

3. The incremental federated causal structure learning method in a dynamic scenario according to claim 2, characterized in that, In step S2, the specific steps for the newly added client adaptive clustering are as follows: Pre-clustering of historical clients: A client selection algorithm based on sample size clustering is used to determine the clustering index of each client through the parameters a and b in the multinomial distribution ax+b; New Client-Side Clustering Matching: Based on historical client-side pre-clustering results, an adaptive expansion of the cluster structure is achieved through a dynamic range matching mechanism, proposing the following three matching rules: Range matching: If the sample quality of the newly added client is SQ(C) t + If a sample falls within the quality range of any historical cluster, it is assigned to the corresponding cluster. Cluster split: If the sample quality of a new client exceeds the maximum sample quality SQ of the historical clusters. max If the maximum quality of all historical client samples is reached, the cluster splitting mechanism is triggered to create a new independent cluster for it. Cluster merging: If the sample quality of a new client is lower than the minimum sample quality SQ of the historical clusters. min If the quality of all historical client samples is the minimum, then it is merged into the newly created low-quality client cluster.

4. The incremental federated causal structure learning method in a dynamic scenario according to claim 1, characterized in that, In step S2, the specific steps for selecting high-quality clients are as follows: Evaluation Dimension Design: A new client evaluation vector is designed within the cluster to achieve accurate client selection. The difference between the data distribution and the BLC is quantified by KL divergence. The smaller the KL divergence, the more similar the distributions are. Multidimensional similarity fusion calculation: Based on three-dimensional features, the cosine similarity between the client and BLC feature vectors within each cluster is calculated using the following formula: in and They represent the client respectively. And the data sample size weights of BLC, and They represent the client respectively. And the random missing rate of BLC samples, Indicates the client The distribution difference between BLC and its own KL divergence is 0, meaning that the distribution is completely consistent. The closer the cosine similarity is to 1, the closer the client data features are to the historical federated causal structure. Selection strategy design: Based on the similarity evaluation results, the following scenario-based selection rules are used to determine the candidate high-quality client set cah_clients. If a cluster contains only 1 new client, it is directly selected into cah_clients; if the number of new clients in a cluster is greater than 1, they are sorted in descending order of cosine similarity, and the new client with the highest similarity in the cluster is selected to be added to the candidate client set cah_clients.

5. The incremental federated causal structure learning method in a dynamic scenario according to claim 1, characterized in that, In step S2, the specific steps for filtering out redundant clients are as follows: The similarity evaluation index SK is calculated based on the federated F2SL framework. Symmetric uncertainty (SU) is used as the metric to characterize the degree of association between nodes in the causal structure, and the similarity evaluation index SK is constructed as follows: in, This is the set of important nodes selected by the set of node weights. V represents the local importance of the client. i The cosine similarity between the SU value and the corresponding node SU value in the server federated model; Define rules for removing redundant clients: if the similarity evaluation metric SK of a client exceeds a preset threshold, it is identified as a redundant client and removed; if all candidate new clients are removed after redundancy screening, the server immediately terminates the current incremental learning, outputs the current federated model results and complete time statistics, and avoids invalid communication and aggregation operations.

6. The incremental federated causal structure learning method in a dynamic scenario according to claim 1, characterized in that, In step S3, the dynamic weight allocation... The calculation formula is as follows: in, The number of new, high-quality, non-redundant clients is defined by n, which represents the initial total number of new clients. This weight reflects the proportion of contribution of the new client group to the federated model, avoiding excessive impact of data quality fluctuations from a single client on the federated structure.

7. The incremental federated causal structure learning method in a dynamic scenario according to claim 1, characterized in that, In step S3, the specific steps for incremental updating the federated causal structure are as follows: The server is based on historical federated causal structure DAGs and newly added federated causal structures. The federal causal model is updated using the following weighted aggregation formula: in, This indicates the node V in the updated federated DAG. i With V j The strength of the causal relationship is defined, and its value ranges from [0,1]. and These represent the correlation strengths in the newly added client-side aggregated DAG and the original server-side aggregated DAG, respectively. This formula achieves efficient and accurate expansion of the federated causal structure by dynamically balancing the stability of the historical structure with the incremental value of the new data.

Images