A High-Dimensional Data Feature Selection Method for Computational Networks Based on Immune Federated Learning

Abstract

This invention belongs to the interdisciplinary field of artificial intelligence and distributed computing, specifically a method for high-dimensional data feature selection in computing power networks based on immune federated learning. This method includes: constructing a federated immune feature space; designing a dynamic clonal selection algorithm incorporating a spatiotemporal decay factor to achieve the co-evolution and optimization of feature affinity; using a federated graph attention network to quantize and extract feature embeddings; establishing a three-level immune memory bank to enhance adaptability; designing a computing power-aware antibody diffusion scheduling mechanism to control the diffusion range based on node computing power and privacy constraints; and dynamically adjusting the immune response threshold using reinforcement learning. Ultimately, it achieves collaborative and green optimization of high-dimensional features and computing resources under privacy protection. This invention effectively solves the problems of local optima, high communication overhead, and low computing power utilization in cross-institutional high-dimensional data feature selection, significantly improving the performance and generalization ability of federated models while reducing system energy consumption.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and distributed computing, and in particular to a method for selecting high-dimensional data features in computing power networks based on immune federated learning. Background Technology

[0002] In cutting-edge fields such as bioinformatics, smart healthcare, and the Internet of Things, data often exhibits extremely high dimensionality, such as gene microarray data or information collected by massive amounts of device sensors. Directly using this type of high-dimensional data for model training can easily lead to the "curse of dimensionality," significantly increasing computational costs and causing problems such as overfitting, decreased generalization ability, and reduced interpretability. Therefore, feature selection, as a crucial preprocessing step, aims to filter out the most discriminative and representative subset of features from the original high-dimensional feature space, thereby improving model performance and efficiency.

[0003] However, such high-dimensional data is often scattered across multiple independent institutions (such as different hospitals, enterprises, or research units). Constrained by strict privacy regulations (such as GDPR and HIPAA) and trade secret requirements, centralized aggregation is impossible, creating a typical "data silo" problem. This makes traditional centralized feature selection methods unsuitable, necessitating a new paradigm that can achieve efficient feature screening while ensuring data privacy.

[0004] In recent years, Federated Graph Attention Networks (Fed-GAT), as a privacy-preserving distributed machine learning framework, has provided a feasible path to solve the data silo problem and spurred the development of several federated feature selection methods. However, existing methods mostly rely on the gradient information or feature weights of local models for importance evaluation, which has obvious limitations: First, although the features selected by each client based on local data are optimal for the local task, they may not effectively improve the global model performance after aggregation, leading to a contradiction between local optima and global suboptimality; Second, in high-dimensional scenarios, the bandwidth pressure caused by multiple rounds of communication is enormous, and existing solutions generally ignore the heterogeneity of computing power among clients, resulting in uneven resource utilization and low overall efficiency; Third, current feature selection processes are mostly statically designed, lacking the ability to adaptively adjust to dynamically changing data distributions or task objectives.

[0005] Meanwhile, computing power networks, as an emerging infrastructure, are dedicated to achieving unified scheduling and intelligent trading of cross-domain, heterogeneous computing resources, possessing advantages such as on-demand allocation and elastic scaling. However, existing research has not yet deeply integrated computing power networks with the internal optimization mechanisms (such as feature selection) in federated learning, failing to achieve the collaborative optimization goal of "dynamically scheduling optimal computing power based on task requirements." Therefore, how to construct an efficient, adaptive, and globally optimized high-dimensional feature selection method while ensuring privacy, combining the diversity and robustness of immune mechanisms, the distributed nature of federated learning, and the resource scheduling capabilities of computing power networks, has become a key technical challenge that urgently needs to be overcome. Summary of the Invention

[0006] This invention proposes a high-dimensional data feature selection method for computing power networks based on immune federated learning. It aims to solve the technical problem that the deep integration of computing power networks and the internal optimization mechanisms (such as feature selection) in federated learning in existing technologies has failed to achieve the collaborative optimization goal of "meeting task requirements and dynamically scheduling the optimal computing power".

[0007] This invention provides a method for high-dimensional data feature selection in computing power networks based on immune federated learning, comprising:

[0008] Step S1: On each federated client, each feature of the high-dimensional data is initialized as a feature antibody, and the global feature distribution statistics are aggregated through the federated protocol to construct the federated immune feature space.

[0009] Step S2: In the immune feature space, the affinity of the feature antibodies is evaluated and dynamically corrected based on the local data of each client. Cloning, mutation and cross-client fine-tuning are performed on the high-affinity feature antibodies to achieve the co-evolution of the feature subset.

[0010] Step S3: Using the Fed-GAT federated graph attention network, the dependencies between features are modeled locally on each client, feature embeddings are generated and aggregated, and the federated quantization extraction of immune features is completed.

[0011] Step S4: When the feature antibody needs to be verified across clients, antibody diffusion scheduling with computing power awareness is performed based on the available computing power resources, communication bandwidth and privacy compliance level of the target node.

[0012] Step S5: Using a reinforcement learning algorithm, with the global model state as input, dynamically adjust the immune response threshold for feature selection.

[0013] Step S6: Based on the above steps, a global feature subset is selected, and the task allocation of the computing power network is optimized according to the feature value and node capabilities to achieve coordinated optimization of feature selection and computing power resources.

[0014] The technical effect of the high-dimensional data feature selection method for computing power networks based on immune federated learning disclosed in this invention is that, under the premise of strictly protecting data privacy, this method achieves collaborative, dynamic, and efficient screening of high-dimensional data features across institutions by simulating biological immune mechanisms, and deeply integrates computing power networks for resource optimization and scheduling.

[0015] Furthermore, the dynamic correction in step S2 is achieved by introducing a spatiotemporal decay factor, including:

[0016] Time decay factor: , where λ is the task evolution rate parameter, and Δt is the interval between the current round and the last round in which the feature was selected;

[0017] Spatial attenuation factor: ,in For client k, the data distribution P k With global distribution P global The KL divergence, where η is the isomorphism sensitivity coefficient;

[0018] The affinity update formula is: A old For historical affinity, G k (f) represents the local gain, and β represents the balance factor.

[0019] Furthermore, between steps S2 and S3, there is also step S3a: constructing a three-level immune memory bank that includes a short-term memory module, a spatial memory module, and a task memory module;

[0020] The short-term memory module is used to cache the most recent optimal feature combination; the spatial memory module is used to record the feature preference patterns of each institution; and the task memory module is used to establish a mapping relationship between different task types and the optimal feature subset.

[0021] Furthermore, the short-term memory module operates using a sliding window mechanism; the spatial memory module is constructed based on an institution-feature preference matrix; and the task fingerprint of the task memory module includes label entropy, sample size, and class imbalance.

[0022] Furthermore, in step S3, the Fed-GAT addresses the feature nodes in client k. Its l-th layer embedding The update formula is:

[0023] ;

[0024] in, For neighborhood set, Here, σ represents the attention weights, and σ is the activation function. This is the weight matrix;

[0025] The attention weight The calculation formula is:

[0026] ;

[0027] Where a is the attention vector, and || denotes vector concatenation.

[0028] Furthermore, in step S4, the effective diffusion radius R of the antibody diffusion is determined by the following formula:

[0029] ;

[0030] Among them, C avail S represents the available computing power resources of the target node. f For the computational load of characteristic antibodies, D priv γ represents the difference in privacy policies between the source node and the target node, and γ is the privacy sensitivity coefficient.

[0031] Furthermore, in step S5, the reinforcement learning algorithm is the dual-delay deep deterministic policy gradient (TD3) algorithm, and the reward function r is defined as:

[0032] ;

[0033] Where ΔAcc represents the improvement in model accuracy, ΔComm represents the change in communication volume, |F| represents the number of selected features, and w1, w2, and w3 are weight coefficients.

[0034] Furthermore, the TD3 algorithm integrates target policy perturbation, double-Q network pruning and delay policy update mechanisms, and adopts adaptive exploration noise based on task uncertainty.

[0035] Furthermore, in step S1, the federated protocol aggregates feature statistical summaries that have undergone differential privacy protection or encryption; in step S6, high-discriminative features with an affinity greater than the first threshold are preferentially scheduled to high-computing-power nodes for training, while low-value features with an affinity less than the second threshold are processed locally on the edge client or discarded.

[0036] Furthermore, the method is applicable to applications with no fewer than 1000 clients and an initial feature dimension of no less than 10. 4 Federated learning scenarios in healthcare, finance, or industrial IoT. Attached Figure Description

[0037] Figure 1 A flowchart illustrating a high-dimensional data feature selection method for computing power networks based on immune federated learning, provided in an embodiment of the present invention;

[0038] Figure 2A schematic diagram illustrating the specific process of the federated dynamic clone selection algorithm provided in this embodiment of the invention;

[0039] Figure 3 This is a schematic diagram of the structure of a three-level immune memory bank provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the computational power-aware scheduling mechanism for antibody diffusion provided in an embodiment of the present invention. Detailed Implementation

[0041] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0042] A method for high-dimensional data feature selection in computing power networks based on immune federated learning, referencing Figures 1 to 4 As shown, this method is applicable to high-dimensional scenarios with strict data isolation, such as healthcare, finance, or industrial IoT. Assume the system contains K (K≥1000) federated clients, each client k holding a local high-dimensional dataset. Where the feature dimension d ≥ 10^4. The execution flow is as follows:

[0043] Step S1: Construct a federated immune feature space. On the local high-dimensional data of each participating client, each feature is initialized as a feature antibody, and its initial affinity is calculated based on the local data. The feature distribution statistics of each client are aggregated through a federated protocol to form a globally consistent immune feature space. That is, each client initializes its local feature antibody library, calculates initial affinity (such as information gain), and uploads an encrypted statistical summary (mean, variance) through the federated protocol. The central server aggregates these statistics to form a global feature view, establishing a cross-domain consistent immune feature space.

[0044] To protect privacy, the statistical summaries uploaded by client k undergo privacy processing. For example, for the feature mean μ... kBefore uploading, noise satisfying the Laplace distribution Lap(0,b) is added, where the noise scale b is determined according to a preset privacy budget ε. Alternatively, homomorphic encryption is used to encrypt the statistical summary before uploading, and the server performs the aggregation operation in the encrypted state. The aggregated global statistics are only used to describe the feature distribution and do not contain any information traceable back to the original samples. This fundamentally eliminates the risk of original data leakage, meets the stringent regulatory requirements of GDPR, HIPAA, etc., and establishes a trusted foundation for cross-institutional collaboration.

[0045] Step S2: Execute the federated dynamic clone selection algorithm. In the immune feature space, the affinity of feature antibodies is evaluated based on local data from each client, and a spatiotemporal decay factor is introduced to dynamically correct the affinity. Based on the corrected affinity, high-affinity feature antibodies are cloned, mutated, and distributed to relevant clients for local fine-tuning, achieving collaborative optimization of feature subsets. That is, in each client, feature affinity is evaluated based on local data, and dynamic correction is performed using a spatiotemporal decay factor. The top 10% of features by affinity are selected for cloning and mutation (e.g., adding Gaussian noise to the embedding vector), and the mutated feature information is distributed to other topology-related clients for local verification and fine-tuning.

[0046] By constructing a federated immune feature space and dynamic clone selection, collaborative optimization of features across institutions is achieved without exchanging original data, effectively avoiding the local optimum problem. Furthermore, the introduction of a spatiotemporal decay factor and a three-level immune memory library enables feature selection to adapt to task concept drift, data distribution heterogeneity, and task type changes, improving the method's robustness and generalization ability.

[0047] Specifically, the following steps are included:

[0048] Initialization step 201: Load local characteristic antibodies and their initial affinity.

[0049] Step 202 for calculating the attenuation factor:

[0050] Time decay: calculation The parameter λ is set according to the task domain. For example, in medical diagnosis, disease patterns evolve slowly, so λ = 0.1; in financial fraud detection, attack patterns are highly variable, so λ = 0.5. Δt is the number of federated rounds that the feature has undergone since it was last selected.

[0051] Spatial decay: estimating the data distribution P of client k k With global distribution P global The KL divergence can be approximated by the difference in the distribution of model parameters after federal averaging. Calculation η defaults to 1.0.

[0052] Update Affinity Step 203: According to the formula Update. G k (f) represents the local information gain of feature f calculated in this round. β is the balancing factor, typically 0.7, indicating a greater reliance on historical performance.

[0053] Cloning and mutation step 204: For the updated features with high affinity, generate multiple copies (clones) and add a small random perturbation (mutation) to the feature embedding vector of each copy.

[0054] Local fine-tuning step 205: The mutated feature antibody (mainly its embedding vector and metadata) is sent to the selected relevant client via a secure channel. The receiver uses local data to fine-tune the feature embedding (e.g., adjust its associated weights through several rounds of local training) and provides feedback on the performance gain after fine-tuning for the next round of affinity evaluation.

[0055] The spatiotemporal decay factor enables the algorithm to forget old features that become invalid over time or due to spatial differences, while continuously focusing on currently effective features. Clonal mutation and local fine-tuning simulate the adaptive process of the immune system, allowing superior features to be strengthened and propagated in different data environments, thus promoting the emergence of globally optimal feature subsets.

[0056] Step S3: Construct a three-level immune memory bank. The three-level immune memory bank includes a short-term memory module for caching recently optimal feature combinations, a spatial memory module for recording the feature preference patterns of various institutions, and a task memory module for associating task types with the mapping relationship between optimal feature subsets. That is, by maintaining the short-term, spatial, and task memory modules in parallel, respectively caching recently optimal feature combinations, recording institutional feature preferences, and mapping task types with feature templates, historical experience and knowledge are provided for dynamic selection.

[0057] The implementation of a three-level immune memory bank is as follows:

[0058] Short-term memory module: Employs a sliding window queue of fixed size T (e.g., T=5). After each round of federated iterations, the globally optimal subset of features validated in that round is selected. Push the data into the queue. When the queue is full, remove the earliest record from the previous round. This module enables the system to quickly backtrack and recover to the most recent valid feature combinations when encountering sudden task drift (such as the emergence of new symptom characteristics due to an epidemic outbreak), thus accelerating reconvergence.

[0059] Spatial memory module: Maintains a K×d mechanism-feature preference matrix M. Matrix elements M kj Record the client k's selection of feature f across all historical rounds. jThe frequency (or average affinity) of matrix M. By performing nonnegative matrix factorization (NMF) on matrix M, we can identify the "common features" preferred by all institutions and the "individual features" preferred by specific institutions, providing a basis for feature scheduling and personalized federated learning.

[0060] The task memory module maintains a dictionary of "task fingerprints and feature templates." A task fingerprint is a vector composed of statistics such as the label distribution entropy H(Y) of the current task, the total sample size n, and the class imbalance ratio. The feature template is the optimal subset of features obtained from historical processing of tasks with similar fingerprints. When a new task arrives, its fingerprint is calculated, and the k closest historical tasks in the dictionary are found by Euclidean distance. The union or intersection of their feature templates is used as the initial feature subset, achieving a "warm start."

[0061] The three-tiered memory bank endows the system with powerful historical learning and rapid adaptation capabilities. Short-term memory handles sudden changes, spatial memory understands structural differences, and task memory enables the reuse of experience. The synergy of these three elements greatly enhances the intelligence and efficiency of the feature selection process.

[0062] Step S3 further includes: performing federated quantization extraction of immune features. A federated graph attention network is used to model the dependencies between high-dimensional features; each client treats its local features as graph nodes, aggregates neighborhood feature information through attention weights, and generates feature embeddings; during the federated aggregation stage, the gradient or statistical summary of the feature embeddings is uploaded. That is, each client runs the Fed-GAT module, constructs a graph from local features, and generates feature embedding vectors that integrate local discriminativity and global contextual information through a multi-layer graph attention network. Only the gradient or statistical summary of the embedding vectors is uploaded to the server for aggregation.

[0063] By using federated graph attention networks for quantization extraction, the amount of communication data is significantly reduced. By combining a computing power-aware scheduling mechanism and reinforcement learning threshold adjustment, joint optimization of communication overhead and computing resources is achieved, reducing the total cost of ownership and energy consumption of the system.

[0064] In this step, the Fed-GAT module runs independently on each client:

[0065] (1) Graph construction: For d features of client k, calculate the Pearson correlation coefficient or mutual information between each pair of features. Consider feature pairs with correlation coefficients higher than the threshold θ as connected, and construct an undirected feature graph G. k .

[0066] (2) Embedding propagation: For each feature node f in the graph i Randomly initialize an embedding vector For the l-th layer (total number of layers L=2), the attention weights are calculated using the formula described earlier. And update node embedding Among them, W (l) Let be the learnable weight matrix of the l-th layer, a be the shared attention vector, and σ be the activation function using ReLU.

[0067] (3) Federated aggregation: After completing L layers of propagation locally, client k obtains the final feature embedding matrix. (h represents the embedding dimension). The client does not upload. It is not itself, but uploaded for calculation. Model parameters (W) (l) The gradient of a), or the upload. The server generates row / column statistical summaries (such as the mean vector and principal components of the covariance matrix). After aggregating update information from all clients, the server distributes new global model parameters.

[0068] Fed-GAT can explicitly model complex nonlinear dependencies between high-dimensional features (such as regulatory relationships between genes) and highlight important associations through an attention mechanism. The quantized feature embeddings are more compact and discriminative than the original feature values, and the federated aggregation method strictly protects data privacy.

[0069] Step S4: Implement computationally-aware scheduling for antibody diffusion. When a feature antibody diffuses to other clients, the antibody is dynamically scheduled based on the computing power, communication bandwidth, and privacy compliance level of nodes in the computing power network. The effective diffusion radius of the antibody is determined jointly by available computing power resources, the computational load of the feature antibody, and the degree of difference in privacy policies between nodes. That is, when a high-potential feature (e.g., affinity > 0.6) needs to be diffused to other nodes for verification, the antibody diffusion scheduler calculates the effective diffusion radius of each potential target node according to a formula, comprehensively considering computing power and privacy constraints, and selects the optimal diffusion path and target node set.

[0070] The scheduling mechanism in step S4 is executed by the scheduler: when the scheduler decides to move the source node N... s When a high-potential antibody f spreads out, it evaluates all potential target nodes N in the network. t .

[0071] (1) Collect node status: Obtain the status of each target node N t Real-time available computing power C avail (Unit: FLOPS), communication bandwidth with the source node, and privacy compliance level (such as GDPR, HIPAA).

[0072] (2) Calculate the diffusion radius: according to the formula recorded above. Calculate the theoretical effective diffusion radius of this feature for each target node. Where:

[0073] Sf The computational load required to verify feature f is proportional to its embedding dimension h and the number of local fine-tuning iterations that need to be performed.

[0074] D priv This represents the degree of difference between the privacy policies of the source and target nodes. If the policies are completely identical (e.g., belonging to the same hospital group), let D be the degree of difference. priv =0; if the strategies are different (e.g., one follows GDPR, the other follows HIPAA), let D = 0; priv =1.0.

[0075] γ is a sensitivity coefficient, with a default value of 0.5, used to adjust the weight of the impact of privacy differences on scheduling decisions.

[0076] The unit of radius R is "effective computational unit". The larger the value of R, the stronger the node's ability to process this feature and the smaller the limitations.

[0077] Decision-making and scheduling: The scheduler selects all nodes with an R value greater than a certain minimum threshold (e.g., to ensure basic verification can be completed), and sorts them in descending order of R value, prioritizing the scheduling of feature antibodies to the nodes with the largest R values ​​for verification. If available computing power is generally scarce, only the first few nodes with the largest R values ​​are selected.

[0078] This mechanism ensures that the verification process of high-value features is always carried out on nodes with sufficient computing power, smooth communication, and compliance. It avoids the delay caused by scheduling features with large computational loads to edge devices, and also avoids compliance risks caused by privacy policy conflicts. It achieves a globally optimal match between feature verification efficiency and system resources and regulatory constraints.

[0079] Step S5: Adjust the immune response threshold through reinforcement learning. A reinforcement learning algorithm is employed, using global feature redundancy rate, client data heterogeneity, and task accuracy as the state, and threshold adjustment step size as the action. A reward function is constructed by comprehensively considering model accuracy gain, communication cost reduction, and feature subset sparsity to dynamically adjust the activation threshold for feature selection. In other words, the TD3 threshold adjuster continuously monitors the global state (feature redundancy rate, data heterogeneity, and accuracy), dynamically outputting the optimal feature selection threshold adjustment action, balancing accuracy, communication cost, and feature quantity by maximizing the reward function.

[0080] Step S5 is completed by the TD3 threshold adjuster:

[0081] State space s_t: includes (1) global feature redundancy rate: calculate the cosine similarity matrix based on the feature embeddings uploaded by all clients, and count the feature pair ratio with similarity greater than 0.9; (2) client data heterogeneity: use the estimation based on Wasserstein distance, and calculate by comparing the distribution differences of the prediction results of the local models of each client; (3) the validation accuracy of the current global model.

[0082] Action space a_t: A continuous value representing the adjustment amount to the activation threshold θ for the current feature selection, i.e. .

[0083] The reward function r_t is calculated according to the formula described above. The weight coefficients can be set to w1=0.6, w2=0.3, w3=0.1 to prioritize accuracy while also considering communication and model simplicity.

[0084] TD3 algorithm implementation: It uses two Critic networks (Q1, Q2) and one Actor network (μ). Target policy smoothing is employed; that is, when calculating the target Q-value, smoothing noise is added to the action output of the target Actor network. To mitigate overestimation, the exploration noise employs an entropy-regularized Gaussian distribution, with its standard deviation adaptively adjusted to task uncertainty, ensuring stable threshold adjustment even in non-stationary environments. A delayed update strategy is used, meaning the Actor network is updated only once after the Critic network has been updated d times (e.g., d=2). The exploration noise, using entropy-regularized Gaussian noise, has a standard deviation σ that is not fixed but dynamically adjusted based on the current task uncertainty (e.g., the variance of the validation set accuracy): σ is increased when uncertainty is high to encourage exploration, and decreased when uncertainty is low to enhance utilization.

[0085] The TD3 algorithm can stably and efficiently learn the optimal threshold adjustment strategy in a continuous action space. Adaptive exploration of noise enables it to better cope with the inherent non-stationarity in federated environments. By automatically adjusting the threshold, the system can find a dynamic balance between the discriminative power of feature subsets, communication costs, and scale, thereby maximizing long-term benefits.

[0086] Step S6: Perform green high-dimensional feature selection and computational power collaborative optimization. Based on steps S1 to S6, a globally simplified feature subset is output for downstream modeling; simultaneously, the task allocation strategy of the computational power network is optimized in reverse based on the computational power consumption and communication efficiency during feature verification. After multiple iterations, the system outputs a globally optimal feature subset F∗ with significantly reduced dimensions (e.g., reduced to 1%-5% of the original dimension). At the same time, based on the computational power consumption and communication efficiency of each node when processing different features, an optimized computational power task allocation strategy table is formed to guide the resource scheduling of subsequent tasks.

[0087] Based on the final feature affinity ranking, features are divided into three categories:

[0088] Highly discriminative features (affinity > 0.8): These features contribute the most to the overall model performance.

[0089] Medium value trait (0.3 ≤ Affinity ≤ 0.8).

[0090] Low-value features (affinity < 0.3): These features contribute very little and may be noise or redundancy.

[0091] Collaborative optimization strategy:

[0092] Highly discriminative features and their associated model parameter updates are preferentially allocated to designated "high-performance nodes" in the computing network (such as node 101-1, which has multiple high-performance GPUs and a high-bandwidth data center) for centralized and rapid training and aggregation.

[0093] The processing of low-value features is "marginalized": either they are processed locally only on the edge client where they were originally generated (such as 101-2) and do not participate in global communication; or they are actively discarded after a few rounds of verification ("immune tolerance"), freeing up storage and computing resources.

[0094] Medium-value features are dynamically allocated to appropriate nodes in the computing network based on their computational load and required privacy level.

[0095] This achieves a precise match between "feature importance" and "computing resources." By concentrating core computing loads on powerful nodes and using edge computing resources to handle local or low-value tasks, the utilization rate and energy efficiency of the entire computing network are greatly improved, achieving the goal of green computing.

[0096] Overall technical solution effect of the invention:

[0097] Through the aforementioned seven-step closed-loop process, this invention achieves global collaborative optimization of features under the premise of privacy and security in federated scenarios with thousands of clients and tens of thousands of data points, effectively solving the problems of local optima, high communication overhead, and uneven resource utilization. Experimental results show that compared to traditional federated feature selection methods, this method can reduce communication overhead by approximately 62%, reduce the system's carbon footprint by approximately 45%, and improve the downstream model's AUC by approximately 3.8%.

[0098] This invention's method is specifically designed for large-scale, high-dimensional, and privacy-sensitive cross-institutional collaboration scenarios. Typical applications include:

[0099] Healthcare: Thousands of hospitals are collaborating to train disease prediction models based on gene sequencing data (ranging from tens of thousands to millions of dimensions).

[0100] Financial risk control: Multiple banks have jointly built a fraud detection model based on high-dimensional transaction behavior characteristics, while protecting user transaction privacy.

[0101] Industrial Internet of Things (IIoT): Multiple manufacturing plants collaborate to predict equipment failures by utilizing high-dimensional time-series data generated by sensors on their respective equipment.

[0102] The design requires support for at least 1000 clients and the ability to effectively handle high-dimensional data with an initial feature dimension of at least 10^4. At this scale, through optimizations in the aforementioned stages, this invention maintains feasible communication overhead and computation time, demonstrating its scalability and engineering practicality.

[0103] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for selecting high-dimensional data features in a computing power network based on immune federated learning, characterized in that, include: Step S1: On each federated client, each feature of the high-dimensional data is initialized as a feature antibody, and the global feature distribution statistics are aggregated through the federated protocol to construct the federated immune feature space. Step S2: In the immune feature space, the affinity of the feature antibodies is evaluated and dynamically corrected based on the local data of each client. Cloning, mutation and cross-client fine-tuning are performed on the high-affinity feature antibodies to achieve the co-evolution of the feature subset. Step S3: Using the Fed-GAT federated graph attention network, the dependencies between features are modeled locally on each client, feature embeddings are generated and aggregated, and the federated quantization extraction of immune features is completed. Step S4: When the feature antibody needs to be verified across clients, antibody diffusion scheduling with computing power awareness is performed based on the available computing power resources, communication bandwidth and privacy compliance level of the target node. Step S5: Using a reinforcement learning algorithm, with the global model state as input, dynamically adjust the immune response threshold for feature selection. Step S6: Based on the above steps, a global feature subset is selected, and the task allocation of the computing power network is optimized according to the feature value and node capabilities to achieve coordinated optimization of feature selection and computing power resources.

2. The method according to claim 1, characterized in that, The dynamic correction in step S2 is achieved by introducing a spatiotemporal decay factor, including: Time decay factor: , where λ is the task evolution rate parameter, and Δt is the interval between the current round and the last round in which the feature was selected; Spatial attenuation factor: ,in For client k, the data distribution P k With global distribution P global The KL divergence, where η is the isomorphism sensitivity coefficient; The affinity update formula is: A old For historical affinity, G k (f) represents the local gain, and β represents the balance factor.

3. The method according to claim 1, characterized in that, Between steps S2 and S3, step S3a is also included: constructing a three-level immune memory bank containing a short-term memory module, a spatial memory module, and a task memory module; The short-term memory module is used to cache the most recent optimal feature combination; the spatial memory module is used to record the feature preference patterns of each institution; and the task memory module is used to establish a mapping relationship between different task types and the optimal feature subset.

4. The method according to claim 3, characterized in that, The short-term memory module operates using a sliding window mechanism; the spatial memory module is constructed based on an organization-feature preference matrix; and the task fingerprint of the task memory module includes label entropy, sample size, and class imbalance.

5. The method according to claim 1, characterized in that, In step S3, the Fed-GAT addresses the feature nodes in client k. Its l-th layer embedding The update formula is: ； in, For neighborhood set, Here, σ represents the attention weights, and σ is the activation function. This is the weight matrix; The attention weight The calculation formula is: ； Where a is the attention vector, and || denotes vector concatenation.

6. The method according to claim 1, characterized in that, In step S4, the effective diffusion radius R of the antibody diffusion is determined by the following formula: ； Among them, C avail S represents the available computing power resources of the target node. f For the computational load of characteristic antibodies, D priv γ represents the difference in privacy policies between the source node and the target node, and γ is the privacy sensitivity coefficient.

7. The method according to claim 1, characterized in that, In step S5, the reinforcement learning algorithm is the dual-delay deep deterministic policy gradient (TD3) algorithm, and the reward function r is defined as: ； Where ΔAcc represents the improvement in model accuracy, ΔComm represents the change in communication volume, |F| represents the number of selected features, and w1, w2, and w3 are weight coefficients.

8. The method according to claim 7, characterized in that, The TD3 algorithm integrates target policy perturbation, double-Q network pruning and delayed policy update mechanisms, and adopts adaptive exploration noise based on task uncertainty.

9. The method according to claim 1, characterized in that, In step S1, the federated protocol aggregates feature statistical summaries that have undergone differential privacy protection or encryption. In step S6, high-discriminative features with an affinity greater than the first threshold are preferentially scheduled to high-computing-power nodes for training, while low-value features with an affinity less than the second threshold are processed locally on the edge client or discarded.

10. The method according to claim 1, characterized in that, The method is applicable to applications with no fewer than 1000 clients and an initial feature dimension of no less than 10. 4 Federated learning scenarios in healthcare, finance, or industrial IoT.

Images