Knowledge sharing multi-task feature selection method based on particle swarm optimization

Abstract

The invention discloses a knowledge sharing multi-task feature selection method based on particle swarm optimization, which belongs to the field of machine learning and high-dimensional data feature selection, and comprises the following steps: S1, an irrelevant feature deletion step: removing irrelevant features in an original feature set by using a threshold deletion strategy to generate a relevant feature subset; s2, a multi-task generation step: constructing two sub-tasks with different problem search spaces, namely a sub-task 1 and a sub-task 2; and S3, a multi-task optimization step: processing a search space task by adopting a particle swarm optimization algorithm, and executing a knowledge transfer strategy based on feature cluster mapping in the optimization process to realize complementation of global search and local focusing. According to the knowledge sharing multi-task feature selection method based on particle swarm optimization, the calculation cost of the algorithm is reduced by reducing the search space and the individual evaluation cost, and the efficiency and the subset representativeness are improved; knowledge complementation is realized through knowledge transfer, and the accuracy of feature selection is improved.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and high-dimensional data feature selection, and in particular, to a knowledge-sharing multi-task feature selection method based on particle swarm optimization. Background Art

[0002] Feature selection (FS) plays a key role in solving high-dimensional classification problems by identifying relevant features that contribute significantly to the performance of the model. Among them, evolutionary multi-tasking (EMT) has made remarkable progress in this field in recent years because it can effectively utilize the correlation between tasks to share knowledge and optimize the feature selection process. However, the existing FS methods based on EMT have limitations in search strategies and evaluation mechanisms, and the search space exploration is insufficient, resulting in insufficient exploration and utilization of knowledge.

[0003] The EC-based FS method is simply referred to as evolutionary FS. So far, many typical evolutionary algorithms have been applied to the FS problem, including particle swarm algorithm, genetic algorithm, ant colony algorithm, artificial bee colony algorithm, and differential evolution, etc. In the past few decades, a large number of PSO-based FS methods have been designed for high-dimensional classification problems. In the existing methods, PSO combines different search strategies to enhance the discriminative ability of the selected feature subset for high-dimensional classification data. Reducing the search space before using PSO has become a popular practice in high-dimensional classification problems in FS. Although improvement strategies such as variable length and variable scale have been proposed in the prior art to optimize the application of the particle swarm algorithm in high-dimensional feature selection, the problem of the curse of dimensionality of high-dimensional data itself still exists. When dealing with ultra-high-dimensional data, whether it is the update calculation of particles or the evaluation of fitness functions, the computational amount is still large. In addition, when reducing the feature space, only simple statistical information is used to determine which part of the search space to reduce, without considering the information interaction between different features. It may lead to the omission of some feature combinations that have an important impact on the final classification performance, thus reducing the quality of feature selection. And multi-task methods can often transfer and utilize information between different tasks to make up for the information omission caused by reducing the feature space in a single task.

[0004] Although many Particle Swarm Optimization (PSO)-based methods have been proposed and shown some potential in dealing with high-dimensional classification problems, feature selection (FS) under high-dimensional data remains a challenge due to the high computational cost. Evolutionary multi-tasking is a relatively new research topic in the fields of optimization and EC, which is proposed to improve the global search performance by optimizing multiple related tasks simultaneously during the evolutionary process. In the multi-task optimization framework, each task does not exist in isolation, and solving one problem can often help solve other problems. This is because there is complementary or shared knowledge between different problems, which links these tasks closely. Multi-task optimization has achieved initial success in many challenging complex problems and shown great application potential. In view of this, some researchers focus on the collaborative relationship between multi-tasks, construct multi-task models, fully explore and utilize the shared knowledge between tasks, thus effectively improving the efficiency and accuracy of algorithms, opening up new ideas and directions for research in related fields.

[0005] Chen et al. proposed an evolutionary multi-task feature selection method based on Particle Swarm Optimization (PSO) specifically for high-dimensional classification problems. By sharing and transferring knowledge between multiple related tasks, it effectively improves the efficiency and accuracy of feature selection. Zhang et al. proposed a dynamic transformation strategy based on the similarity between tasks to achieve a unified representation of knowledge in different dimensions, including dimension supplementation and dimension reduction. At the same time, this method utilizes the search characteristics of Particle Swarm Optimization to promote knowledge transfer between tasks in different dimensions. Chen et al. proposed an evolutionary multi-task learning method that transforms the high-dimensional feature selection problem into multiple related low-dimensional feature selection tasks, and then finds the optimal feature subset through knowledge transfer between these low-dimensional tasks. Wang et al. proposed a multi-task evolutionary learning method (MEL) that conducts co-evolutionary search based on two sub-populations, and each sub-population independently uses different methods to search for the optimal feature subset. Wu et al. proposed a diverse knowledge transfer strategy for the multi-task Particle Swarm Optimization algorithm. According to the state of population evolution, an adaptive task selection mechanism is introduced to manage the source tasks that contribute to the target task, thereby achieving knowledge transfer to improve model performance.

[0006] Although these multi-task methods aim to transfer knowledge between different tasks to improve the accuracy of feature selection, in practical applications, the dimensionality of the search space has not been significantly reduced, and they still face the problem of the "curse of dimensionality". Summary of the Invention

[0007] The object of the present invention is to provide a knowledge sharing multi-task feature selection method based on particle swarm optimization, which reduces the computational cost of the algorithm by reducing the search space and the cost of individual evaluation, improves the efficiency and the representativeness of the subset, and realizes the complementarity of knowledge through knowledge transfer to improve the accuracy of feature selection.

[0008] To achieve the above object, the present invention provides a knowledge sharing multi-task feature selection method based on particle swarm optimization, which includes the following steps:

[0009] S1. Step of deleting irrelevant features: Using a threshold deletion strategy to remove the irrelevant features in the original feature set to generate a relevant feature subset;

[0010] S2. Step of generating multi-tasks: Constructing two sub-tasks with different problem search spaces, namely sub-task 1 and sub-task 2;

[0011] S3. Step of optimizing multi-tasks: Using the particle swarm optimization algorithm to process the search space tasks, and implementing a knowledge transfer strategy based on feature cluster mapping during the optimization process to achieve the complementarity of global search and local focus.

[0012] Preferably, in sub-task 1 in S2, feature selection is directly performed in the feature space after the step of deleting irrelevant features; for sub-task 2, the feature space is first clustered to obtain a feature clustering space, and then representative features are selected from each cluster.

[0013] Preferably, in S3, the optimization framework includes a feature selection optimization model, a solution evaluation mechanism, and a knowledge transfer strategy.

[0014] Preferably, the feature selection optimization model includes a traditional optimization model and an integer optimization model; in the traditional optimization model, the candidate solutions are represented by 0 and 1 or real numbers between 0 and 1, and their values represent the probability of the feature being selected; in the integer optimization model, the candidate solutions are represented by integers, and their values represent the selected feature numbers in the feature cluster.

[0015] Preferably, in the traditional optimization model, for Task1 with the feature space as the problem search space, the traditional optimization model is adopted; for Task2 with the feature clustering space as the problem search space, whose objective is to generate an optimal feature subset by selecting representative features from the feature clusters, an overall optimization model is introduced to search for the optimal feature subset from different feature clusters. Assuming that the high-dimensional feature space is clustered into Q clusters, denoted as Cluster k , where k = 1, 2, …, Q; the objective of the overall optimization model is to select Q 1 representative features (Q 1 ≤Q) from the Q feature clusters to make the objective function F(·) reach an extreme value:

[0016]

[0017] Among them, X = (x1, x2, …, x Q ) represents the candidate solution of the current clustering space, and its length is the number of feature clusters; the value range of each dimension is an integer between 0 and |Cluster i |. If x i = a, it means that the a-th feature of Cluster i is selected. If x i = 0, it means that no feature in the feature cluster Cluster i is selected. L represents the number of instances, and Φ represents the empty set.

[0018] Preferably, in the evaluation mechanism of the solution, for the neighborhood component-based feature selection method NCFS, a fitness function inspired by neighborhood component analysis NCA is adopted and used as the fitness function of Task1. Its expression is as follows:

[0019]

[0020] S i is the candidate solution vector in the Task1 population. w represents the feature weight vector obtained by the neighborhood component-based feature selection method NCFS. s i represents the search positions in S i that are greater than 0.5. w j represents the feature weight value. size(s i ) and size(S i ) represent the number of selected features and the number of all features respectively; β is a control parameter that maintains the balance between the subset weight and the subset size, and its value is 0.1;

[0021] Use the index based on the classification performance of the classifier as the fitness function of Task2, specifically as follows:

[0022]

[0023] Among them, acc i is the classification accuracy of the i-th candidate solution. |D i | is the number of selected features in the i-th candidate solution. D is the total number of features. α is a parameter that reflects the role of the classification accuracy and the number of the selected feature subset, and its value is 0.9.

[0024] Preferably, in the knowledge transfer strategy, different subpopulations explore in their respective unique search spaces. Subpopulation 1 explores the feature space, and subpopulation 2 explores the feature clustering space.

[0025] Therefore, the present invention adopts the above-mentioned knowledge sharing multi-task feature selection method based on particle swarm optimization, which has the following advantages:

[0026] 1) Improve the dynamic cooperation mechanism of the dual search space: The global exploration of the complete feature space ensures the comprehensiveness of feature associations, and the local focus of the clustering feature space narrows the search range, improving efficiency and the representativeness of subsets.

[0027] 2) Differential solution evaluation mechanism: The rapid screening in the complete space and the precise evaluation in the clustering space complement each other, breaking through the efficiency bottleneck of traditional single evaluation, and balancing the calculation speed and solution quality.

[0028] 3) Cross-space knowledge transfer strategy: The global feature association guides local search to avoid the optimal trap, and the local feature characteristics feedback to optimize the global direction, achieving deep complementarity.

[0029] Next, through the accompanying drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0030] Figure 1 is the overall framework diagram of an embodiment of the knowledge sharing multi-task feature selection method based on particle swarm optimization of the present invention;

[0031] Figure 2 is the schematic diagram of the top-down algorithm search space of an embodiment of the knowledge sharing multi-task feature selection method based on particle swarm optimization of the present invention;

[0032] Figure 3 is the search space conversion diagram from Task1 to Task2 of an embodiment of the knowledge sharing multi-task feature selection method based on particle swarm optimization of the present invention;

[0033] Figure 4 is the search space conversion diagram from Task2 to Task1 of an embodiment of the knowledge sharing multi-task feature selection method based on particle swarm optimization of the present invention. Detailed Embodiments

[0034] The following further illustrates the technical solutions of the present invention through the accompanying drawings and embodiments.

[0035] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the field to which the present invention belongs.

[0036] Embodiment 1

[0037] The purpose of feature selection is to identify and select a subset of features with the greatest predictive power for the target variable from the complete feature set to improve the performance of the machine learning model. Assuming that the dataset has D features and L instances, and its complete feature set is F, the FS problem in classification can be expressed as: select as few features as possible to optimize the given performance indicator H(·). The mathematical model is as follows:

[0038] max H(x)

[0039] stX=(x1,x2,…,x D )

[0040] x i ∈rand(0,1),i=0,1,…,D

[0041] x i >θ, then X i =1; otherwise, X i =0(1);

[0042] Among them, X i =1 indicates that the feature is selected into the feature subset X; otherwise, the feature is not selected.

[0043] The standard particle swarm optimization (PSO) contains several control parameters, including inertia weight and acceleration coefficient. Improper parameter settings may reduce the performance of PSO. In view of this, simplified particle swarm optimization (BBPSO) is used. For a D-dimensional optimization problem, the position of the i-th particle is X i =(x i,1 ,…,x i,D The best position experienced by a particle (i.e., the individual best position) is denoted as Pbest i =(pbest i,1 ,…,pbest i,D ); and the best position experienced by the group (i.e., the global best position) is expressed as Gbest = (gbest1,…,gbest D ). Then, its update rule is as follows:

[0044]

[0045] Among them, G(μ,δ) is a Gaussian distribution function with mean μ and variance σ.

[0046] Previous researchers extended BBPSO to the FS problem and proposed an improved BBPSO algorithm called HFS-CP. In HFS-CP, an improved particle update rule was developed as follows:

[0047]

[0048] Among them, G(0,1) is a standard Gaussian distribution; is the floor function.

[0049] As Figure 1 shown, the present invention includes three key steps:

[0050] S1. Step of deleting irrelevant features;

[0051] The input of the step of deleting irrelevant features is the source feature set. By removing irrelevant or weakly relevant features, a set of relevant features is obtained, the high-dimensional feature space is initially reduced, and a feature space shared by subtasks is formed.

[0052] S2. Step of multi-task generation;

[0053] The goal of the step of multi-task generation is to construct two feature selection subtasks with different problem search spaces. Specifically, the problem search space of one subtask is the above-mentioned feature space, which is based on individual features, that is, each feature is regarded as an independent decision variable, and the value of the decision variable determines whether to select the feature. The other subtask first clusters the feature space, and divides similar features into the same feature cluster. The feature clustering space serves as the problem search space, and each feature cluster is regarded as an independent decision variable, and the value of the decision variable determines which feature to select from the feature cluster. This way can significantly reduce the number of decision variables.

[0054] As Figure 2 shown, in the high-dimensional feature selection problem, the classification performance of the evolutionary multi-task feature selection (FS) method is affected by the correlation between tasks. In the multi-task framework, each subtask should have a certain degree of commonality or complementarity in terms of computational efficiency and accuracy. The present invention fully considers the characteristics of different search spaces, and proposes a clustering-guided multi-task generation strategy to construct two subtasks with different problem search spaces. Subtask 1 directly performs feature selection in the feature space processed in the first step. Subtask 2 is more complex. First, it clusters the feature space to obtain the feature clustering space, and then selects representative features from each cluster. For high-dimensional data, after deleting irrelevant features, the search space is still large. By feature clustering, similar features can be divided into the same cluster, and regarding each feature cluster as an independent decision variable can significantly reduce the search space of Subtask 2.

[0055] S3. Step of multi-task optimization;

[0056] The multi-task optimization step simultaneously processes the above-mentioned subtasks through multi-task optimization based on a multi-type feature selection method. First, different feature selection optimization models are constructed for each subtask. In the method proposed in the present invention, two models are considered: a traditional optimization model and an integer optimization model. In the traditional optimization model, candidate solutions are usually represented by 0 and 1, or by real numbers between 0 and 1, where the value of a certain dimension represents the probability that the feature is selected. In the integer optimization model, candidate solutions are represented by integers, and the value of a certain dimension represents the number of the selected feature in the feature cluster. Second, in order to solve the above two optimization models, this framework adopts different types of evolutionary feature selection methods: a filter method based on an evolutionary algorithm and a wrapper method based on an evolutionary algorithm. The former uses a correlation-redundancy metric that does not depend on the classifier as the fitness function to evaluate the individuals (i.e., candidate solutions) in the subpopulation, and can quickly screen out feature subsets with greater potential, thereby reducing the computational overhead and improving the optimization efficiency. The latter evaluates the individuals in the subpopulation by calling the classification performance metric of the classifier as the fitness function to more accurately measure the contribution of each feature subset to the performance of the final classification model. By combining these two strategies, it is possible to ensure the superiority of the optimization result in terms of classification performance while guaranteeing the computational efficiency. Finally, a global search ability and are introduced to solve the two subtasks respectively, and a knowledge transfer strategy is executed between the tasks throughout the evolution process to enhance the overall effectiveness and efficiency of the algorithm.

[0057] In the proposed framework, in order to adapt to different task requirements, two types of feature selection optimization models are constructed: a traditional optimization model and an integer optimization model. As shown in Equation (1), in the traditional optimization model, the number of elements included in the candidate solution is equal to the dimension of the feature, and the value range of each dimension is a real number between 0 and 1, representing the probability that the feature of this dimension is selected. For Task1 with the feature space as the problem search space, it is appropriate to adopt the traditional optimization model. However, after the second step of implementing the proposed framework, for Task2 with the feature clustering space as the problem search space, whose goal is to generate an optimal feature subset by selecting representative features from these feature clusters, the traditional optimization model is no longer applicable. In order to search for the optimal feature subset from different feature clusters, an integer optimization model is introduced. Assume that the high-dimensional feature space is clustered into Q clusters, denoted as Cluster k , where k = 1, 2,..., Q. The goal of this optimization model is to select Q 1 representative features (Q 1 ≤Q) from Q feature clusters to optimize the objective function F(·).

[0058]

[0059] Among them, X = (x1, x2,..., x Q) represents the candidate solution of the current clustering space, and its length is the number of feature clusters. The value range of each dimension is an integer between 0 and |Cluster i |. If x i = a indicates that the a-th feature of Cluster i is selected. If x i = 0, it means that no feature in Cluster i is selected. L represents the number of instances, and Φ represents the empty set. Different from traditional optimization models, the selection of features in the overall optimization model is not based on probability or binary decision, but is achieved by explicitly selecting representative features in the feature clusters. Since the length of the candidate solution is the number of feature clusters, this optimization model can improve the efficiency of the algorithm by reducing the search space.

[0060] In the multi-task feature selection framework, the solution evaluation mechanism is crucial because it directly affects the search efficiency and overall performance of the algorithm. In the solution evaluation mechanism, the most critical is the design of the fitness function. In the proposed framework, the main role of Task1 is to quickly identify potential feature subsets in a large feature space. The NCFS method adopts a fitness function inspired by Neighborhood Component Analysis (NCA). When performing distance measurement or dimensionality reduction, NCA does not require complex matrix operations and does not make specific assumptions about the distribution of the sample space. More importantly, NCA not only focuses on the correlation between features and class labels, but also considers the mutual relationship between features through the Mahalanobis Distance. Therefore, taking the fitness function as the fitness function of Task1, its expression is as follows:

[0061]

[0062] S i is the candidate solution vector in the Task1 population, w represents the feature weight vector obtained by NCFS, s i represents the search positions in S i that are greater than 0.5, w j represents the feature weight value, size(s i ) and size(S i ) represent the number of selected features and all features respectively. β is a control parameter that maintains the balance between subset weight and subset size, and according to existing research, its value is 0.1.

[0063] Since the above fitness function does not require calling a classifier, although the computational cost is small, it may not be able to fully capture the complex relationships between features and may ignore other potential high-quality feature subsets. Task2 expects to search for the optimal feature subset more precisely in a smaller search space. Therefore, an index based on the classification performance of the classifier is used as the fitness function for Task2, as follows:

[0064]

[0065] where acc i is the classification accuracy of the i-th candidate solution, |D i | is the number of selected features in the i-th candidate solution, D is the total number of features, and α is a parameter reflecting the roles of classification accuracy and the number of the selected feature subset, with a value of 0.9.

[0066] The proposed evaluation mechanism based on multiple fitness functions enables the proposed framework to achieve a balance between computational efficiency and evaluation accuracy by using fitness functions with different computational costs and accuracies in two subtasks respectively. The fitness function with low computational cost helps the algorithm quickly determine the search direction, while the fitness function with high accuracy ensures in-depth exploration in the region of potential optimal solutions.

[0067] The core goal of multi-task optimization is not only to solve the optimization problem of a single task, but to optimize in multiple problem domains by considering multiple related tasks simultaneously, so as to improve the overall performance. Since Task1 and Task2 in the proposed framework set the search spaces of the problem from different perspectives, the common feature selection methods based on multi-task optimization with a unified solution representation and a single evolutionary mechanism to solve multiple tasks are no longer applicable. Therefore, the present invention assigns solvers with specific optimization models, solution representations, and solution evaluation mechanisms to each task. In Task1, BBPSO with enhanced global exploration ability by Gaussian distribution is adopted, while in Task2, an improved version HFS-C-P of BBPSO applicable to integer form is adopted. It should be noted that BBPSO in the optimization framework can be extended to other evolutionary algorithms.

[0068] The optimization framework includes three key components: First is the feature selection optimization model, which is used to characterize the feature selection problems of different tasks and design appropriate representations of candidate solutions to cope with diverse search spaces. Second is the solution evaluation mechanism, whose main role is to evaluate the quality of the feature subsets obtained during the multi-task optimization process. Finally, the knowledge transfer strategy is an important part of this method, which is responsible for sharing information between related feature selection tasks, thereby promoting collaborative optimization and knowledge transfer between different tasks and improving the overall optimization effect.

[0069] In the multi-task optimization algorithm, each sub-population not only has the ability to learn independently, but also can further enhance its search ability through the knowledge transfer strategy. Based on this, a knowledge transfer strategy based on space transformation is proposed. This strategy allows different sub-populations to explore in their respective unique search spaces: for example, sub-population 1 focuses on the feature space, while sub-population 2 focuses on the feature clustering space. Given the differences in the solution representations in the two sub-populations, the present invention first selects the optimal individuals from the two sub-populations respectively, and maps them from their respective problem search spaces to a unified feature space; knowledge interaction and sharing are carried out within this space; then, the shared information is mapped back to their respective problem search spaces in the reverse direction to guide the update of the individual positions in each sub-population, thereby promoting the efficient exploration and optimization of the algorithm in different search spaces. Specifically, the knowledge transfer between sub-population 1 and sub-population 2 can be achieved in the following way:

[0070] Case 1: The individuals in sub-population 2 usually update their positions according to formula (3). When knowledge transfer is satisfied, they will be guided by the dominant individual gbest (task1) in sub-population 1, and the update process is as follows:

[0071]

[0072] where a, b, and c are learning factors and satisfy a + b + c = 1. z_best (task1) is the representation of gbest (task1) transformed into the individuals in sub-population 2 using the feature space. is the individual optimum for task2, is the global optimum for task2, is the global optimum individual transferred from task1 to task2.

[0073] The process of obtaining z_best (task1) is as follows: First, select the top p individuals in sub-population 1. Then, re-evaluate these individuals using the high-accuracy fitness function fitness2, and select the individual with the largest fitness value as the dominant individual gbest (task1) . Finally, use the feature space information to transform gbest (task1) into the individual representation z_best (task1) applicable to Task2. Here, p is set to 20%. Among them, the transformation process is as Figure 3 shown. Assume that gbest (task1) selects {f1, f2, f4, f6, f7}. In the feature clustering of Task1, {f1, f2} belongs to Cluster1, and let f1 be the feature of SU max in {f1, f2}, so is 1. {f4} belongs to Cluster2, gbest (task1) only f4 in it belongs to Cluster2, so is 1. {f6, f7} belongs to Cluster3, let f7 be the SU in {f6, f7} max characteristics, so is 2. Since gbest (task1) has no characteristics belonging to Cluster4, so is 0. Then z_best (task1) is represented as [1, 1, 2, 0].

[0074] Case 2: Individuals in Sub-population 1 are usually updated in position according to formula (2). When knowledge transfer is satisfied, they will be guided by the optimal individual gbest (task2) in Sub-population 2, and the update process is as follows:

[0075]

[0076] where a, b, and c are learning factors and satisfy a + b + c = 1. z_gest (task2) is gbest (task2) converted into the representation form of individuals in Sub-population 1 using the feature space, is the individual optimal for task1, is the global optimal for task1, is the global optimal individual transferred from task2 to task1. The conversion process is as Figure 4 shown.

[0077] z_best (task2) is obtained as follows: Assume the individual representation of gbest (task2) is [3, 1, 2, 1], then the selected features from Cluster are {f3, f4, f7, f8}. In the search space of Task1, to increase the diversity of particle positions, {f3, f4, f7, f8} takes values of rand(θ, 1) in the position of z_best (task2) , where θ is the feature selection threshold in Task1. The remaining z_best (task2) is equal to gbest (task2) . Finally, z_best (task2) is represented as Figure 4 shown.

[0078] To avoid negative transfer as much as possible, the conditions for knowledge transfer are set. When the classification performance of the dominant individual gbest (task1) in Sub-population 1 is better than the optimal individual gbest in Sub-population 2(task2) When, that is, it satisfies fitness2(gbest (task1) )≥fitness2(gbest (task2) ), the individuals in sub-population 1 are updated using the conventional formula (2), and the individuals in sub-population 2 are updated using formula (7); otherwise, the individuals in sub-population 1 are updated using formula (8), and the individuals in sub-population 2 are updated using the conventional formula (3).

[0079] Therefore, the present invention adopts the above-mentioned method for multi-task feature selection with knowledge sharing based on particle swarm optimization, reduces the computational cost of the algorithm by reducing the search space and the cost of individual evaluation, improves the efficiency and the representativeness of the subset; realizes the complementarity of knowledge through knowledge transfer, and improves the accuracy of feature selection.

[0080] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A knowledge sharing multi-task feature selection method based on particle swarm optimization, characterized in that It includes the following steps: S1. Step of deleting irrelevant features: Using a threshold deletion strategy to remove irrelevant features in the original feature set, generating a relevant feature subset; S2. Step of multi-task generation: Constructing two subtasks with different problem search spaces, namely subtask 1 and subtask 2; S3. Step of multi-task optimization: Using a particle swarm optimization algorithm to process the search space tasks, and implementing a knowledge transfer strategy based on feature cluster mapping during the optimization process to achieve the complementarity of global search and local focusing.

2. The knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 1, characterized in that: The subtask 1 in S2 directly performs feature selection in the feature space processed by the step of deleting irrelevant features; the subtask 2 first clusters the feature space to obtain a feature clustering space, and then selects representative features from each cluster.

3. A knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 1, characterized in that: In S3, the optimization framework includes a feature selection optimization model, a solution evaluation mechanism, and a knowledge transfer strategy.

4. A knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 3, characterized in that: The feature selection optimization model includes a traditional optimization model and an integer optimization model; in the traditional optimization model, candidate solutions are represented by 0 and 1 or real numbers between 0 and 1, and their values represent the probability of the feature being selected; in the integer optimization model, candidate solutions are represented by integers, and their values represent the selected feature numbers in the feature cluster.

5. A knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 4, characterized in that: In the traditional optimization model, the traditional optimization model is adopted for Task1 with the feature space as the problem search space; for Task2 with the feature clustering space as the problem search space, whose goal is to generate an optimal feature subset by selecting representative features from feature clusters, an overall optimization model is introduced to search for the optimal feature subset from different feature clusters. Suppose the high-dimensional feature space is clustered into Q clusters, denoted as Cluster k , where k = 1, 2, …, Q; the goal of the overall optimization model is to select Q 1 representative features from Q feature clusters (Q 1 ≤Q) such that the objective function F(·) reaches an extreme value: where X = (x1, x2, …, x Q ) represents a candidate solution in the current clustering space, and its length is the number of feature clusters; the value range of each dimension is an integer between 0 and |Cluster i |. If x i = a, it means that the a-th feature of Cluster i is selected. If x i = 0, it means that no feature in the feature cluster Cluster i is selected. L represents the number of instances, and Φ represents the empty set.

6. A knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 3, characterized in that: In the solution evaluation mechanism, based on the neighborhood components feature selection method NCFS, a fitness function inspired by neighborhood components analysis NCA is adopted and used as the fitness function of Task1, and its expression is as follows: S i is a candidate solution vector in the Task1 population, w represents the feature weight vector obtained by the neighborhood component-based feature selection method NCFS method, and s i represents the search positions in S i that are greater than 0.5, w j represents the feature weight value, size(s i ) and size(S i ) represent the number of selected features and all features respectively; β is a control parameter that maintains the balance between the subset weight and the subset size, and its value is 0.1; An index based on the classification performance of the classifier is adopted as the fitness function of Task2, specifically as follows: where acc i is the classification accuracy of the i-th candidate solution, |D i | is the number of selected features in the i-th candidate solution, D is the total number of features, and α is a parameter reflecting the role of the classification accuracy and the number of selected feature subsets, with a value of 0.

9.

7. A knowledge sharing multi-task feature selection method based on particle swarm optimization according to claim 3, characterized in that: In the knowledge transfer strategy, different sub-populations explore in their respective unique search spaces. The sub-population 1 explores the feature space, and the sub-population 2 explores the feature clustering space.

Images