Federal feature selection method and device based on evolutionary multi-task optimization and storage medium

Abstract

The invention relates to the field of federal learning, and provides a federal feature selection method based on evolutionary multi-task optimization. The method comprises the following steps: each client generates an initial feature selection population based on local data of the client, and uploads a current optimal feature subset to a server; the server carries out client layering through cross evaluation based on the current optimal feature subsets uploaded by the clients and calculates the similarity between tasks; each client updates a local population based on a hierarchical strategy of the server, and in a local optimization process, according to a task similarity and a migration strategy guided by the server, a knowledge migration operation perceived by the similarity is executed, and meanwhile, population diversity is dynamically maintained; and each client uploads an optimized result to the server, and the server screens a global excellent solution and updates a system state for a next round of iteration.

Description

Technical Field

[0001] This application relates to the field of federated learning, and in particular to a federated feature selection method, apparatus and storage medium based on evolutionary multi-task optimization. Background Technology

[0002] With the advent of the big data era, data, as a key driver of artificial intelligence development, is increasingly valuable. However, due to privacy regulations and data security requirements, data is often scattered across different institutions or terminal devices in the form of "islands," making it difficult to directly and centrally train models. Federated Learning (FL), as an emerging distributed machine learning paradigm, allows multiple participants to collaboratively train a global model without exchanging their original local data, thereby achieving knowledge sharing and model performance improvement while protecting data privacy.

[0003] Feature selection is an important preprocessing technique in machine learning, aiming to select a subset of features from the original feature set that significantly contributes to model performance and has low redundancy. In federated learning scenarios, achieving efficient, accurate, and privacy-preserving joint feature selection under the condition that client data is non-independent and identically distributed (non-IID) has become a key challenge. Existing technical solutions can be mainly divided into two categories: First, federated feature selection based on filtering or embedded methods. The core idea is that each client independently calculates feature importance, and then the server performs simple aggregation. While this method has low communication overhead, it ignores the correlation between tasks among clients. In non-IID scenarios, the global feature subset obtained by simple aggregation may not be suitable for the local data distribution of all clients, leading to a decline in model performance. Second, it introduces optimization algorithms such as evolutionary computation into the federated framework, modeling it as a distributed optimization problem. However, existing methods typically treat feature selection for each client as an independent or homogeneous task, employing a uniform initialization, update, and aggregation strategy. This lacks personalized guidance for the heterogeneity of data distribution across different clients and fails to effectively utilize the potential similarities between tasks for guided knowledge transfer. This results in inefficient optimization processes, a tendency to get stuck in local optima, and an inability to fully leverage the advantages of federated collaboration to improve the generalization ability of each local model under its own data distribution.

[0004] Therefore, the existing technology has the following defects: 1) It lacks effective measurement and utilization of task similarity between clients, the knowledge sharing mechanism is rigid, it is difficult to achieve efficient cross-task knowledge transfer, and the optimization effect is not good in non-IID scenarios; 2) In the optimization process, a "one-size-fits-all" strategy is adopted for clients, and it fails to provide differentiated guidance based on the model performance and data characteristics of the clients themselves, resulting in slow convergence or poor performance of some clients and low overall collaborative efficiency. Summary of the Invention

[0005] This application provides a federated feature selection method, apparatus, and storage medium based on evolutionary multi-task optimization. Through co-evolution and hierarchical knowledge transfer mechanisms, it effectively improves the generalization ability and optimization efficiency of federated feature selection under non-independent and identically distributed data.

[0006] On the one hand, this application provides a federated feature selection method based on evolutionary multi-task optimization, in which the server and multiple clients collaborate to perform iterative optimization, and a single iteration includes the following steps: Each client generates an initial feature selection population based on its local data and uploads the current optimal feature subset to the server. The server performs client stratification and calculates task similarity based on the current best feature subset uploaded by each client through cross-evaluation. Each client updates its local population based on the server's hierarchical strategy, and during the local optimization process, performs similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. Each client uploads the optimized results to the server, which then selects the best global solutions and updates the system status for the next iteration.

[0007] On the other hand, this application provides a federated feature selection device based on evolutionary multi-task optimization, in which a server and multiple clients collaborate to perform iterative optimization, and a single iteration is performed by the following modules included in the device: An initialization module is used by each client to generate an initial feature selection population based on its local data and upload the current optimal feature subset to the server. The hierarchical module is used by the server to perform client hierarchical analysis and calculate the similarity between tasks based on the current optimal feature subset uploaded by each client through cross-evaluation. The optimization module is used by each client to update the local population based on the server's hierarchical strategy. During the local optimization process, it performs similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. The update module is used by each client to upload the optimized results to the server. The server selects the best solutions globally and updates the system status for the next iteration.

[0008] Thirdly, this application provides an apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described federated feature selection method based on evolutionary multitasking optimization.

[0009] Fourthly, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described federated feature selection method based on evolutionary multi-task optimization.

[0010] As can be seen from the technical solution provided in this application, on the one hand, in each iteration cycle, the server calculates the inter-task similarity based on the current best feature subset uploaded by each client and guides the client to perform similarity-aware knowledge transfer. This mechanism enables targeted knowledge exchange between clients with similar data distributions but excellent feature subsets, rather than simple global averaging. This similar task-assisted model effectively utilizes the task relevance in the federated system, allowing each client to absorb successful experiences from similar tasks when optimizing its own feature subset. This overcomes the challenge of distribution differences brought about by non-IID data, resulting in the final selected feature subset not only performing well on its own client data but also possessing stronger cross-distribution adaptability and generalization potential. On the other hand, in each iteration cycle, the server stratifies clients through cross-evaluation and formulates different population update strategies for clients at different levels. This differentiated stratified initialization strategy changes the "one-size-fits-all" approach in traditional federated optimization. The processing method allows high-performance clients to reuse and consolidate their excellent experience, medium-performance clients to receive more accurate guidance, and low-performance clients to avoid stagnation in low-quality areas. This "tailored" collaborative mechanism enables all clients to search in directions more suitable to their current state, thereby accelerating the collective convergence process of the entire federated network and improving optimization efficiency. Thirdly, the local optimization step of the client explicitly includes the operation of dynamically maintaining population diversity. This operation removes redundant solutions by calculating the dissimilarity between individuals in the population and dynamically adjusting the threshold based on the population state. This mechanism ensures that even under the influence of knowledge transfer and global guidance, the search population of each client can maintain sufficient diversity, preventing the population from converging prematurely to a local optimum due to convergence. Therefore, this mechanism, combined with knowledge transfer and hierarchical guidance, forms an effective balance between "exploration" and "utilization," ensuring that the algorithm can continuously evolve towards a better feature subset, ultimately improving the global quality of the obtained feature subset. In summary, the technical solution of this application effectively improves the generalization ability and optimization efficiency of federated feature selection under non-independent and identically distributed data through co-evolution and hierarchical knowledge transfer mechanisms. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a flowchart of the federated feature selection method based on evolutionary multi-task optimization provided in the embodiments of this application; Figure 2 This is the process by which the client, according to the cross-evaluation results, is divided into different performance levels and a differentiated population initialization strategy is adopted accordingly, as provided in the embodiments of this application. Figure 3 The embodiments provided in this application are for the current client ( Select the most similar task. A diagram illustrating the role of a knowledge transfer partner; Figure 4 This is a schematic diagram of the structure of the federated feature selection device based on evolutionary multi-task optimization provided in the embodiments of this application; Figure 5 This is a schematic diagram of the device provided in the embodiments of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In this specification, adjectives such as "first" and "second" are used only to distinguish one element or action from another, without necessarily requiring or implying any actual such relationship or order. Where circumstances permit, reference to an element or component or step (etc.) should not be construed as being limited to only one of the elements, components, or steps, but may be one or more of the elements, components, or steps, etc.

[0015] For ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale.

[0016] Feature selection is an important preprocessing technique in machine learning, aiming to select a subset of features from the original feature set that significantly contributes to model performance and has low redundancy. In federated learning scenarios, achieving efficient, accurate, and privacy-preserving joint feature selection under the condition that client data is not independently and identically distributed (non-IID) has become a key challenge. However, existing solutions have the following drawbacks: 1) They lack effective measurement and utilization of task similarity among clients, resulting in rigid knowledge-sharing mechanisms that hinder efficient cross-task knowledge transfer and poor optimization performance in non-IID scenarios; 2) During optimization, a "one-size-fits-all" strategy is adopted for each client, failing to differentiate guidance based on the client's model performance and data characteristics, leading to slow convergence or poor performance for some clients and low overall collaborative efficiency.

[0017] To address the aforementioned problems in existing technologies, this application proposes a federated feature selection method based on evolutionary multi-task optimization. This method involves a server and multiple clients collaborating to perform iterative optimization in a single iteration. The flowchart of this method is attached. Figure 1 As shown, the main steps include S101 to S104, which are detailed below: Step S101: Initial Population Generation and Knowledge Transfer Preparation Steps: Each client generates an initial feature selection population based on its local data and uploads the current optimal feature subset to the server.

[0018] Step S101 aims to prepare a high-quality initial solution set for subsequent collaborative optimization and to initiate information exchange. In existing federated feature selection methods, clients typically initialize their search population randomly, which can lead to the algorithm starting from a poor starting point and converging slowly in a high-dimensional feature space. To address this issue, this application introduces a feature pre-screening mechanism to improve the quality of the initial population in a data-driven manner. Specifically, each client generates its initial feature selection population based on its local data, which can be achieved through steps S1011 to S1013, as detailed below: Step S1011: Calculate the first evaluation index and the second evaluation index of the local data characteristics using a filter.

[0019] Filter-based methods are independent of subsequent learning models, evaluating features based on the inherent properties of the data. They are fast but may neglect interaction with the model. To balance efficiency and effectiveness, in this embodiment, optionally, the first evaluation metric for the filter can be based on the Fisher Score, which measures the discriminative power of a feature against different categories. It calculates the ratio of the variance of each feature across different categories to the variance within each category; a higher ratio indicates higher discriminative power. The second evaluation metric for the filter can be based on mutual information (MI), which measures the statistical dependence between a feature and the category label. A higher mutual information value indicates that the feature contains more information about the category.

[0020] Step S1012: Combine the first evaluation index of the filter type and the second evaluation index of the filter type to obtain the feature importance score.

[0021] To comprehensively utilize the advantages of both indicators, they need to be normalized to eliminate the influence of dimensions, and then weighted and fused. For example, the Fisher Score and mutual information MI are normalized to the [0, 1] interval, and then a weighted sum formula is used. Calculate the overall importance score for each feature j. ,in, and These are the weighting coefficients, and This can be adjusted according to the specific scenario; for example, all values ​​can be set to 0.5 to show that they are treated equally.

[0022] Step S1013: Select an important feature subset based on the feature importance score, and use this subset to guide population initialization so that important features appear with a higher probability in the initial solution.

[0023] Based on the calculated comprehensive importance score Sort all features. Select the top-ranked features. K The features (e.g., the top 50% of features) constitute an important feature subset. When initializing the population using algorithms such as particle swarm optimization, it is no longer completely randomized. For each individual in the population (i.e., a potential subset of features), its initial value (representing the probability or weight of that feature being selected) on each feature dimension will be affected by... The impact. One specific implementation is to add a weighted term for the importance score to the initialization formula, for example, ,in, It is a random number. This is an adjustment coefficient. It should be noted that to prevent the probability from exceeding the [0, 1] interval, the calculated probability needs to be truncated. Furthermore, the initial weighted value undergoes boundary processing: if the calculated result is greater than 1, it is set to 1; if it is less than 0, it is set to 0, to ensure that the feature selection probability is within an effective range. This makes it possible to select features from a subset of important features. The features of a given solution have higher initial weights, thus guiding the algorithm to begin exploring from a more promising region of the search space.

[0024] After completing the local population initialization, each client submits its current best feature subset obtained in the current round (or after initial optimization if it is the first round). The local performance metrics (e.g., classification accuracy) are uploaded to the server. Step S101 is now complete, providing the necessary input for global planning on the server side.

[0025] Step S102: Global Knowledge Integration and Task Planning Step: Based on the current best feature subset uploaded by each client, the server performs client stratification through cross-evaluation and calculates the similarity between tasks.

[0026] To address the issue of federated systems' inability to intelligently perceive task relevance among clients and thus hinder effective collaborative planning in the context of non-independent identically distributed (non-IID) data, an existing technical solution is simple averaging or weighted aggregation. This involves the server averaging or weighting the performance metrics (e.g., accuracy) of a subset of features uploaded by clients, then selecting a globally optimal value or directly applying the average. However, this approach is essentially a blind aggregation strategy, completely ignoring the heterogeneity of client data distribution. In non-IID scenarios, a feature subset optimal for one client may be extremely inefficient for another, and forced averaging can lead to negative transfer, severely impairing generalization ability. Using random or round-robin pairing—where the server randomly assigns knowledge transfer partners to clients or employs a fixed round-robin mechanism—lacks intelligence. Random pairing cannot guarantee efficient knowledge flow between similar tasks, resulting in inefficiency and also easily leading to negative transfer. In view of this, this application adopts a global knowledge integration and task planning scheme, namely: the server performs client stratification and calculates task similarity based on the current optimal feature subset uploaded by each client through cross-evaluation. On the one hand, through cross-evaluation, the server can not only know "how well each client is performing" but also "how well its solution adapts to other clients," thus enabling more reasonable stratification based on global adaptability rather than local performance, providing a scientific basis for subsequent differentiated processing. On the other hand, by calculating similarity (e.g., Jaccard coefficient), the server can identify which clients have similar tasks (i.e., data distribution and optimization goals), which is equivalent to equipping the federated system with radar, enabling proactive discovery of potential, efficient collaboration opportunities rather than blind collaboration. As an embodiment of this application, the server's client stratification based on the current optimal feature subset uploaded by each client through cross-evaluation can be implemented through steps S1021 to S1023, as detailed below: Step S1021: The server organizes clients to perform cross-validation on each other's optimal feature subsets.

[0027] The server coordinates with each client to process the current optimal feature subset uploaded by other clients without exchanging any original local data. Verification is performed. Specifically, the client... Use your own local test dataset to validate the client. Feature subset The performance, for example, using only The selected features are used for training and evaluation, and the accuracy is calculated.

[0028] Step S1022: Rank the clients based on their performance according to the average verification accuracy.

[0029] For each client The server calculates its feature subset. The average cross-validation accuracy is calculated across all other client datasets. This average cross-validation accuracy reflects the global generalization ability of this feature subset. The server then ranks all clients based on this average accuracy.

[0030] Step S1023: Determine the client with the highest ranking as the optimal client, and divide the remaining clients into high-performance group, medium-performance group, and low-performance group.

[0031] The client ranked first The feature subset corresponding to the client identified as the optimal client in this round. As candidates for the globally optimal solution, the remaining clients are then divided into three tiers based on their ranking: the top 25% are in the high-performance group, those ranked 25% to 50% are in the medium-performance group, and the bottom 50% are in the low-performance group. This tiered approach provides the basis for subsequent differentiated processing; a strategy diagram can be found in [reference needed]. Figure 2 It visually demonstrates the process by which clients are divided into different performance levels based on cross-evaluation results, and a corresponding differentiated population initialization strategy is adopted.

[0032] To facilitate effective knowledge transfer, the server needs to quantify the correlation between client tasks. This application uses the Jaccard similarity coefficient to compare the optimal feature subsets between clients to calculate task similarity. For any two clients... and Their task similarity By calculating their current optimal feature subset and The Jaccard index is obtained as follows: The above In the calculation formula, the numerator is the cardinality of the intersection of two feature subsets, and the denominator is the cardinality of the union. The Jaccard coefficient ranges from 0 to 1; a larger value indicates a higher degree of overlap in the feature spaces that the two clients are interested in during the current iteration, and a greater similarity in their tasks. The server calculates the similarity of each client to all other clients, and typically selects one or more clients with the highest similarity as potential knowledge transfer partners.

[0033] In step S102, the server completes the "perception" and "planning" of the global state, generating crucial information to guide the next stage of optimization: the client-side hierarchical strategy and the task similarity matrix. This step is the core of achieving intelligent and personalized federated optimization. Step S103: Co-evolutionary optimization step: Each client updates its local population based on the server's hierarchical strategy, and during the local optimization process, performs similarity-aware knowledge transfer operations according to the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. It should be noted that the co-evolutionary optimization in this application adopts different execution logic in the first iteration (i.e., cold start) and subsequent iterations (update phase). That is, in the first iteration, since the server has not yet obtained the performance status of each client, each client mainly relies on local data-driven initialization (i.e., step S101) to perform independent search. However, in the second and subsequent iterations, each client performs co-optimization entirely based on the hierarchical strategy and migration guidance fed back by the server.

[0034] To achieve efficient personalized optimization and address the problem of premature convergence in the evolutionary process, ensuring the sustainability and effectiveness of the search, one existing approach is independent local optimization. This involves each client running its optimization algorithm solely based on its local data without external interaction. This approach completely negates the collaborative advantages of federated learning, failing to leverage the knowledge of other clients to improve the generalization ability of its own model, especially performing poorly on clients with small datasets. Another approach, unified strategy federated optimization, employs the same initialization strategy, the same knowledge transfer rules (e.g., fixed probabilities), and the same population management method across all clients. While attempting collaboration, this approach suffers from rigidity and a lack of specificity. It cannot adapt to the personalized needs of each client on non-IID data, and the experience of high-performance clients may mislead low-performance clients (negative transfer). Furthermore, the lack of diversity maintenance mechanisms can easily cause all clients to converge prematurely to local optima. Therefore, this application adopts a co-evolutionary optimization approach: each client updates its local population based on the server's hierarchical strategy. During local optimization, it performs similarity-aware knowledge transfer operations based on the task similarity and transfer strategy guided by the server, while dynamically maintaining population diversity.

[0035] Specifically, as an embodiment of this application, each client updates its local population based on the server's hierarchical strategy, and during the local optimization process, performs similarity-aware knowledge transfer operations according to the task similarity and migration strategy guided by the server. At the same time, dynamically maintaining population diversity can be achieved through steps S1031 to S1033, as detailed below: Step S1031: Each client updates its local population based on the server's hierarchical strategy.

[0036] In the first iteration (i.e., the cold start phase), since the server has not yet obtained the performance data of each client, each client directly generates an initial population using the strategy described in step S1013 above, which is based on feature importance scoring (combining Fisher Score and mutual information), without involving the server's hierarchical guidance. When entering the second round and subsequent iterations, the server sends hierarchical information (as described in step S1023 of the aforementioned embodiment) to each client. Each client initializes or resets its population for this round of search using different strategies according to its hierarchical level: for high-performance clients, a mixed initialization is performed using their local historical best solution and the global historical best solution. For example, a high proportion of individuals in the new population are derived from the local best solution of the previous round. and the global optimal solution Genetically generated through crossover, mutation, and other genetic operations, the system aims to consolidate and integrate existing advantages. For clients in the medium-performance group, a hybrid initialization is performed using their local and global historical best solutions with similar partners. The server assigns one or more partner clients with the highest similarity (calculated by comparing the optimal feature subsets among clients using the Jaccard similarity coefficient) to clients within this group, and the population initialization incorporates the historical best solutions of these partner clients. This aims to benefit from the successful experiences of similar tasks. For low-performance clients, a combination of global historical best-solution guidance and suboptimal solution guidance is used for initialization. This population is partly composed of... One group of individuals is guided to quickly gravitate towards superior regions; another group is initialized using a hybrid approach, combining the globally suboptimal solution (e.g., the second-ranked feature subset) with the local historical best solution. This differentiated strategy is as follows: Figure 2 As shown, this ensures that all clients can start their search from a starting point suitable for their current state.

[0037] Step S1032: During the local optimization process, perform similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server.

[0038] During the iteration of local optimization algorithms (such as Particle Swarm Optimization, PSO), the client intermittently attempts to absorb knowledge from other clients. This operation is not random or fixed, but intelligent and similarity-based. Specifically, based on the task similarity and transfer strategy guided by the server, similarity-aware knowledge transfer operations can be performed through steps S1032a to S1032c, as detailed below: Step S1032a: Set an adaptive migration probability for each client.

[0039] The Random Mating Probability (RMP) is not a fixed value. Its initial value can be set to 0.3, and then dynamically adjusted based on the effectiveness of knowledge transfer in improving individual quality during historical iterations. Specifically, a learning rate parameter can be introduced (…). The cross-task migration success rate and local optimization success rate are calculated. If the migration success rate is high, the migration probability for the next round is increased based on the learning rate. The calculation logic is as follows: the new migration probability is obtained by weighted summation of the old migration probability and the increment based on the success rate ratio, i.e.: This makes the knowledge transfer behavior adaptive to the current optimization environment.

[0040] Step S1032b: Based on the adaptive migration probability, extract feature selection information of the shared feature domain from the selected migration partner.

[0041] At each possible moment of migration, generate a random number. .like If this happens, a migration operation is triggered. The server will assign a knowledge migration partner to each client (the most similar client calculated based on S1024), such as... Figure 3 As shown, by calculating the Jaccard similarity between the intersection and union of feature subsets, it specifically demonstrates how to provide the current client ( Select the most similar task. The two clients, acting as knowledge transfer partners, define the shared feature domain as the intersection of their respective important feature subsets. During transfer, the entire solution from the partner is not copied verbatim; instead, the focus is on the selection information within this intersection feature domain. For example, weights or selection states corresponding to these shared features are extracted from the partner's high-performing individuals. Specifically, extracting feature selection information from the selected transfer partner can involve the transfer partner client using a linear autoencoder based on Principal Component Analysis (PCA) to compress and encode the local population, generating a low-dimensional representation, which is then uploaded. The current client receives and decodes this representation to reconstruct the feature selection information. By transmitting only the compressed representation instead of the original feature matrix, data privacy is protected and communication overhead is reduced.

[0042] Step S1032c: Using task similarity as weight, integrate feature selection information into the individuals to be optimized in the current client.

[0043] After extracting the information, it is injected into the individual being optimized by the current client using a weighted fusion formula. The fusion weight is precisely the task similarity between the two clients. For example, the feature weight of a new individual = (1- *Current individual weight + *Individual partner weight. This means that the more similar the tasks, the greater the degree of borrowing. The principle behind this mechanism can be found in [link to relevant documentation]. Figure 3 .

[0044] Step S1033: Dynamically maintain population diversity.

[0045] During the iteration process, evolutionary algorithms may cause the population to converge, losing the ability to explore new regions and leading to premature convergence. This invention actively maintains diversity in each iteration. This operation can be achieved through steps S1033a to S1033c, as detailed below: Step S1033a: Calculate the dissimilarity among individuals in the population.

[0046] The dissimilarity among individuals in the population is obtained by calculating the normalized Manhattan distance between the individual feature selection vectors. For two individuals, i.e., feature selection vectors... and their dissimilarity , where D is the total dimension of the features. The value is between 0 and 1, and the smaller the value, the more similar the two solutions are.

[0047] Step S1033b: Set the dynamic dissimilarity threshold.

[0048] The dissimilarity threshold is not fixed. It is adjusted based on the statistical characteristics of the number of features selected by individuals in the current population. For example, it can be negatively correlated with the average number of features selected by the current population. When the average number of selected features is low, it means the solution is relatively "sparse," and the dissimilarity threshold should be increased. To encourage the exploration of more features, the dissimilarity threshold should be appropriately relaxed when the average number of selected features is large. A certain degree of convergence is allowed.

[0049] Step S1033c: Remove redundant solutions and add new individuals.

[0050] Traverse the population, and for each individual, calculate the minimum dissimilarity between it and all other individuals in the population. .like If an individual is deemed redundant, it is removed from the population. Subsequently, new individuals are generated to fill the vacancy, either through random generation or by mutation of superior individuals, thus introducing new diversity.

[0051] Through the synergistic effect of S1031 to S1033, the client completed a local optimized search in this iteration that is both globally guided and personalized and robust.

[0052] Step S104: Result Feedback and Global Update Step: Each client uploads the optimized results to the server. The server selects the best global solutions and updates the system status for the next iteration.

[0053] Results feedback and global updates conclude a single iteration and prepare for the next. After completing local optimization, each client uploads its current best feature subset found in this round, along with its performance metrics (e.g., accuracy) on its local validation set, to the server. The server collects all information and then selects the globally best solution. This process can be similar to the cross-validation in S1021; the server can either conduct another rapid cross-validation or directly select the best-performing feature subset as the new globally best solution based on the local performance metrics reported by each client. Simultaneously, the server updates information such as performance rankings for client-side tiering. These updated system states (including new globally excellent solutions, client performance rankings, etc.) will serve as input for step S102 in the next iteration, thus forming a closed loop of continuous optimization. Figure 1 The example method, after feature selection, also includes a step of training a federated model based on the final feature subset. Once the feature selection iterations meet the termination conditions (such as reaching the maximum number of iterations or performance convergence), each client processes its local data using the finally determined optimal feature subset and participates in standard federated learning model training (e.g., the FedAvg algorithm) to ultimately obtain a high-performance global model.

[0054] From the above appendix Figure 1As illustrated by the example of the federated feature selection method based on evolutionary multi-task optimization, on the one hand, in each iteration cycle, the server calculates inter-task similarity based on the current best feature subset uploaded by each client and guides the clients to perform similarity-aware knowledge transfer. This mechanism enables targeted knowledge exchange between clients with similar data distributions but excellent feature subsets, rather than simple global averaging. This cooperative model of similar tasks effectively utilizes the task relevance in the federated system, allowing each client to absorb successful experiences from similar tasks when optimizing its own feature subset. This overcomes the challenge of distribution differences brought about by non-IID data, resulting in the final selected feature subset not only performing well on its own client data but also possessing stronger cross-distribution adaptability and generalization potential. On the other hand, in each iteration cycle, the server stratifies clients through cross-evaluation and formulates different population update strategies for clients at different levels. This differentiated stratified initialization strategy changes the traditional "one-size-fits-all" approach in federated optimization. The "treating all clients equally" approach allows for the reuse and consolidation of the best practices of high-performance clients, provides more precise guidance to medium-performance clients, and prevents low-performance clients from stagnating in low-quality areas. This collaborative mechanism of "teaching according to aptitude" enables all clients to search in directions more suitable to their current state, thereby accelerating the collective convergence process of the entire federated network and improving optimization efficiency. Thirdly, the local optimization steps of the clients explicitly include operations to dynamically maintain population diversity. This operation removes redundant solutions by calculating the dissimilarity between individuals in the population and dynamically adjusting the threshold based on the population state. This mechanism ensures that even under the influence of knowledge transfer and global guidance, the search population of each client can maintain sufficient diversity, preventing the population from converging prematurely to a local optimum due to convergence. Therefore, this mechanism, combined with knowledge transfer and hierarchical guidance, forms an effective balance between "exploration" and "utilization," ensuring that the algorithm can continuously evolve towards a better feature subset, ultimately improving the global quality of the obtained feature subset. In summary, the technical solution of this application effectively improves the generalization ability and optimization efficiency of federated feature selection under non-independent and identically distributed data through co-evolution and hierarchical knowledge transfer mechanisms.

[0055] Please see the appendix Figure 4 This application provides a federated feature selection device based on evolutionary multi-task optimization, in which a server and multiple clients collaborate to perform iterative optimization. Each iteration is executed by the following modules included in the device: initialization module 401, hierarchical module 402, optimization module 403, and update module 404, as detailed below: Initialization module 401 is used by each client to generate an initial feature selection population based on its local data and upload the current optimal feature subset to the server. The hierarchical module 402 is used by the server to perform client hierarchical analysis and calculate the similarity between tasks based on the current best feature subset uploaded by each client through cross-evaluation. The optimization module 403 is used by each client to update the local population based on the server's hierarchical strategy. During the local optimization process, it performs similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. The 404 update module is used by each client to upload the optimized results to the server. The server then selects the best solutions globally and updates the system status for the next iteration.

[0056] From the above appendix Figure 4As illustrated by the example of a federated feature selection device based on evolutionary multi-task optimization, on the one hand, in each iteration cycle, the server calculates inter-task similarity based on the current best feature subset uploaded by each client and guides the clients to perform similarity-aware knowledge transfer. This mechanism enables targeted knowledge exchange between clients with similar data distributions but excellent feature subsets, rather than simple global averaging. This cooperative mode of similar tasks effectively utilizes the task relevance in the federated system, allowing each client to absorb successful experiences from similar tasks when optimizing its own feature subset. This overcomes the challenge of distribution differences brought about by non-IID data, resulting in the final selected feature subset not only performing well on its own client's data but also possessing stronger cross-distribution adaptability and generalization potential. On the other hand, in each iteration cycle, the server stratifies clients through cross-evaluation and formulates different population update strategies for clients at different levels. This differentiated stratified initialization strategy changes the traditional "one-size-fits-all" approach in federated optimization. The "treating all clients equally" approach allows for the reuse and consolidation of the best practices of high-performance clients, provides more precise guidance to medium-performance clients, and prevents low-performance clients from stagnating in low-quality areas. This collaborative mechanism of "teaching according to aptitude" enables all clients to search in directions more suitable to their current state, thereby accelerating the collective convergence process of the entire federated network and improving optimization efficiency. Thirdly, the local optimization steps of the clients explicitly include operations to dynamically maintain population diversity. This operation removes redundant solutions by calculating the dissimilarity between individuals in the population and dynamically adjusting the threshold based on the population state. This mechanism ensures that even under the influence of knowledge transfer and global guidance, the search population of each client can maintain sufficient diversity, preventing the population from converging prematurely to a local optimum due to convergence. Therefore, this mechanism, combined with knowledge transfer and hierarchical guidance, forms an effective balance between "exploration" and "utilization," ensuring that the algorithm can continuously evolve towards a better feature subset, ultimately improving the global quality of the obtained feature subset. In summary, the technical solution of this application effectively improves the generalization ability and optimization efficiency of federated feature selection under non-independent and identically distributed data through co-evolution and hierarchical knowledge transfer mechanisms.

[0057] Figure 5 This is a schematic diagram of the structure of a device provided in one embodiment of this application. For example... Figure 5 As shown, the device 5 in this embodiment mainly includes: a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and executable on the processor 50, such as a program for a federated feature selection method based on evolutionary multi-task optimization. When the processor 50 executes the computer program 52, it implements the steps described in the above embodiment of the federated feature selection method based on evolutionary multi-task optimization, for example... Figure 1The steps S101 to S104 are shown. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 4 The functions of the initialization module 401, the layering module 402, the optimization module 403, and the update module 404 are shown.

[0058] For example, the computer program 52 based on the federated feature selection method of evolutionary multi-task optimization mainly includes: each client generating an initial feature selection population based on its local data and uploading the current best feature subset to the server; the server performing client stratification and calculating inter-task similarity based on the current best feature subset uploaded by each client through cross-evaluation; each client updating its local population based on the server's stratification strategy, and performing similarity-aware knowledge transfer operations during local optimization according to the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity; each client uploading the optimized results to the server, and the server selecting globally excellent solutions and updating the system state for the next iteration. The computer program 52 can be divided into one or more modules / units, which are stored in memory 51 and executed by processor 50 to complete this application. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 52 in device 5. For example, computer program 52 can be divided into the functions of initialization module 401, hierarchical module 402, optimization module 403, and update module 404 (a module in the virtual device). The specific functions of each module are as follows: Initialization module 401 is used by each client to generate an initial feature selection population based on its local data and upload the current best feature subset to the server; Hierarchical module 402 is used by the server to hierarchize the clients and calculate the task similarity based on the current best feature subset uploaded by each client through cross-evaluation; Optimization module 403 is used by each client to update the local population based on the server's hierarchical strategy, and in the local optimization process, to perform similarity-aware knowledge transfer operations according to the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity; Update module 404 is used by each client to upload the optimized results to the server, and the server selects the globally excellent solution and updates the system state for the next iteration.

[0059] Device 5 may include, but is not limited to, processor 50 and memory 51. Those skilled in the art will understand that... Figure 5 This is merely an example of device 5 and does not constitute a limitation on device 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, the device may also include input / output devices, network access devices, buses, etc.

[0060] The processor 50 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0061] The memory 51 can be an internal storage unit of the device 5, such as a hard disk or RAM of the device 5. The memory 51 can also be an external storage device of the device 5, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the device 5. Furthermore, the memory 51 can include both internal and external storage units of the device 5. The memory 51 is used to store computer programs and other programs and data required by the device. The memory 51 can also be used to temporarily store data that has been output or will be output.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed. That is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above-described device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0063] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0064] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0065] In the embodiments provided in this application, it should be understood that the disclosed apparatus / device and method can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0066] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0067] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0068] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a storage medium. Based on this understanding, all or part of the processes in the above-described embodiments can also be implemented by a computer program instructing related hardware. The computer program for the federated feature selection method based on evolutionary multi-task optimization can be stored in a storage medium. When executed by a processor, the computer program can implement the steps of the various method embodiments described above, namely, each client generates an initial feature selection population based on its local data and uploads the current optimal feature subset to the server; the server performs client stratification and calculates task similarity based on the current optimal feature subset uploaded by each client through cross-evaluation; each client updates its local population based on the server's stratification strategy, and during the local optimization process, performs similarity-aware knowledge transfer operations according to the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity; each client uploads the optimized results to the server, and the server selects globally excellent solutions and updates the system state for the next iteration. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. Storage media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of storage media can be appropriately added to or removed according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, storage media may not include electrical carrier signals and telecommunication signals.

[0069] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application. The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the protection scope of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A federated feature selection method based on evolutionary multi-task optimization, characterized in that, Iterative optimization is performed collaboratively by the server and multiple clients. A single iteration includes the following steps: Each client generates an initial feature selection population based on its local data and uploads the current optimal feature subset to the server. The server performs client stratification and calculates task similarity based on the current best feature subset uploaded by each client through cross-evaluation. Each client updates its local population based on the server's hierarchical strategy, and during the local optimization process, performs similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. Each client uploads the optimized results to the server, which then selects the best global solutions and updates the system status for the next iteration.

2. The federated feature selection method based on evolutionary multi-task optimization as described in claim 1, characterized in that, Each client generates an initial feature selection population based on its local data, including: Calculate the first and second evaluation metrics of the filter type for local data characteristics; The feature importance score is obtained by combining the first and second filter-type evaluation metrics. A subset of important features is selected based on the feature importance score, and the population initialization is guided based on this subset, so that important features appear with a higher probability in the initial solution.

3. The federated feature selection method based on evolutionary multi-task optimization as described in claim 1, characterized in that, The server performs client stratification based on the current optimal feature subset uploaded by each client through cross-evaluation, including: The server organizes clients to perform cross-validation on each other's optimal feature subsets; Clients are ranked by performance based on average verification accuracy. The client with the highest ranking is identified as the optimal client, and the remaining clients are divided into high-performance, medium-performance, and low-performance groups.

4. The federated feature selection method based on evolutionary multi-task optimization as described in claim 3, characterized in that, Each client updates its local population based on the server's hierarchical strategy, including: For high-performance group clients, a hybrid initialization is performed using their local historical best solution and the global historical best solution; For clients in the medium-performance group, a hybrid initialization is performed using their local historical best solution and the global historical best solution, along with those of similar partners. For low-performance clients, initialization is performed using a combination of global historical best solution guidance and random exploration.

5. The federated feature selection method based on evolutionary multi-task optimization as described in claim 1, characterized in that, The step of performing similarity-aware knowledge transfer operations based on the task similarity and transfer strategy guided by the server includes: Set an adaptive migration probability for each client; Based on the adaptive migration probability, feature selection information of shared feature domains is extracted from the selected migration partners; Using the task similarity as a weight, the feature selection information is integrated into the optimized individual of the current client.

6. The federated feature selection method based on evolutionary multi-task optimization as described in claim 5, characterized in that, The adaptive transfer probability is dynamically adjusted based on the effectiveness of knowledge transfer in improving individual quality during historical iterations.

7. The federated feature selection method based on evolutionary multi-task optimization as described in claim 1, characterized in that, The dynamic maintenance of population diversity includes: Calculate the dissimilarity among individuals in a population; The dissimilarity threshold is dynamically adjusted based on the statistical characteristics of the number of selected features in the current population. Redundant solutions with dissimilarity levels below the dissimilarity threshold are removed from the population and new individuals are added.

8. A federated feature selection device based on evolutionary multi-task optimization, characterized in that, Iterative optimization is performed collaboratively by the server and multiple clients, with each iteration executed by the following modules included in the device: An initialization module is used by each client to generate an initial feature selection population based on its local data and upload the current optimal feature subset to the server. The hierarchical module is used by the server to perform client hierarchical analysis and calculate the similarity between tasks based on the current optimal feature subset uploaded by each client through cross-evaluation. The optimization module is used by each client to update the local population based on the server's hierarchical strategy. During the local optimization process, it performs similarity-aware knowledge transfer operations based on the task similarity and migration strategy guided by the server, while dynamically maintaining population diversity. The update module is used by each client to upload the optimized results to the server. The server selects the best solutions globally and updates the system status for the next iteration.

9. An apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images