Interpolation evolution method for processing two-target feature selection problem in high-dimensional data classification
Through the interpolation evolution method, the initial population is created on the high-dimensional dataset and iterative operations are solved, and the problem of feature selection of two targets on the high-dimensional dataset is improved, search ability and diversity are improved, and better feature selection and classification performance are achieved. , and reduces computational complexity.
Patent Information
- Application Number
- CN202411036562.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-26
AI Technical Summary
The existing multi-objective evolutionary method faces the difficulty of search space expansion, local optimality and performance balance in the two-objective feature selection problems in high-dimensional data classification, and it is difficult to maintain population diversity and algorithm convergence.
Using the interpolation evolution method, create initial populations through interpolation initialization strategy, and use offspring reproduction and environment selection operations until the preset number of iterations is reached, and multiple interests in the target space are explored using interpolation points and special subpopulations. area.
It improves the search capability and diversity on high-dimensional data sets, overcomes the limitations of traditional MOEAs, achieves better feature selection and classification performance, and has high computing efficiency.
Smart Images

Figure CN119025959A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of high-dimensional data classification processing, and in particular to an interpolation evolution method for processing a two-target feature selection problem in high-dimensional data classification. Background Art
[0002] In high-dimensional data classification tasks, feature selection is a key step in data preprocessing, and its purpose is to select the subset that has the greatest impact on the classification results from a large number of features of the original data. This process can not only reduce the data dimension that needs to be processed during the classification process, thereby greatly reducing the computational complexity of data classification, but also avoid overfitting and improve the generalization ability of the model. At present, multi-objective evolutionary methods (MOEAs) are widely used in feature selection problems, especially when minimizing the selected feature rate and minimizing the classification error rate are the two optimization goals of feature selection. These methods can provide a solution set in population units by considering multiple optimization goals at the same time. The non-dominated solution set contained in it can help decision makers or practitioners make the best decision according to actual needs from more perspectives and in a wider range.
[0003] Currently, MOEAs on the market can generally be divided into methods based on non-dominated sorting, methods based on weight decomposition, and methods based on performance indicators, etc. However, although these methods are effective in continuous optimization environments, when dealing with discrete optimization problems such as two-objective feature selection, especially when the decision space is rapidly expanded due to high-dimensional classification data sets, their search performance is often greatly affected and hindered. As mentioned above, the two-objective feature selection problem usually takes minimizing the selected feature rate and minimizing the classification error rate as two goals for collaborative optimization. Let the former be the objective function f1(x) and the latter be the objective function f2(x), then the two can be defined as follows:
[0004]
[0005] Among them, D represents the dimension of the feature space, that is, the dimension size of the classification data set, and x represents the solution vector.
[0006]
[0007] Among them, TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives in the classification results, respectively.
[0008] Existing MOEAs face many challenges when dealing with the two-objective feature selection problem in high-dimensional classification datasets. First, the high dimensionality of the feature space increases the size of the search space exponentially, making it difficult for traditional MOEAs to have sufficient search capabilities to fully explore as many non-dominated solutions as possible and maintain the diversity of the population. Second, the nature of discrete optimization problems leads to local optimal problems in the optimization process, which may cause traditional MOEAs to fall into the trap of suboptimal solutions or premature maturity, thereby affecting the convergence of the method. In addition, existing evolutionary methods often need to make trade-offs between low-dimensional and high-dimensional datasets with completely different optimization environments, and how to balance the performance of the two is also a technical challenge. For example, classic MOEAs based on non-dominated sorting (such as NSGA-II) use non-dominated sorting and crowding distance to make environmental selections, but in high-dimensional feature spaces, the simple use of non-dominated sorting and crowding distance methods is no longer sufficient to maintain the diversity of the population and the convergence of the algorithm. Classical MOEAs based on weight decomposition (such as MOEA / D) simplify the problem by decomposing the multi-objective problem into a series of single-objective problems. However, this method may have difficulty balancing the weights between different objectives when dealing with high-dimensional data sets, resulting in the inability to search for a sufficient number of non-dominated solution sets in large-scale sparse feature spaces.
[0009] Therefore, the existing data preprocessing field urgently needs a new evolutionary method that can effectively handle the two-objective feature selection problem in high-dimensional data classification, while overcoming the limitations and shortcomings of traditional MOEAs on high-dimensional data sets.
[0010] In view of this, this application is filed. Summary of the invention
[0011] The present invention provides an interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification, which can at least partially improve the above-mentioned problem.
[0012] To achieve the above object, the present invention adopts the following technical solutions:
[0013] An interpolation evolutionary method for processing a two-objective feature selection problem in high-dimensional data classification, comprising:
[0014] Get the difference to be evolved, set the population size N, the total number of features D and the total number of evolution iterations T, and set the current number of iterations C to 0;
[0015] The difference is preprocessed using an interpolation initialization strategy IBI to generate an initial population, wherein the preprocessing includes deploying a plurality of evenly distributed interpolation points in the target space, creating interpolation subpopulations corresponding to the interpolation points, analyzing the distribution of solutions of all interpolation subpopulations to obtain a local search area of interest, and creating four special subpopulations in the local search area of interest to merge into an initial population;
[0016] Determine whether the current iteration has reached the loop termination condition, and output the final evolved population Pop or continue the evolutionary iteration process of the current population based on the judgment result, wherein the loop termination condition is whether the current iteration number C is greater than or equal to the total number of evolutionary iterations T.
[0017] In summary, the English name of the interpolation evolutionary method used to deal with the two-objective feature selection problem in high-dimensional data classification is Interpolation Based Evolutionary Method, abbreviated as IBEA. Specifically, IBEA (interpolation evolutionary method) first uses an adaptive initialization strategy based on interpolation points (Interpolation Based Initialization, abbreviated as IBI) to create a competitive initial population, and then repeatedly reproduces offspring and selects the environment for the initial population until the evolution termination condition (accumulated to a preset number of iterations) is reached and the final evolved population is output. Each solution vector contained in the population corresponds to a D-dimensional row vector, in which the element value 1 represents the selected feature and 0 represents the unselected feature. Each solution vector corresponds to a two-dimensional target vector, that is, the selected feature rate corresponding to the optimization target f1 direction is minimized, and the classification error rate corresponding to the optimization target f2 direction is minimized. This overcomes the limitations and various shortcomings of traditional MOEAs in dealing with two-objective feature selection problems on high-dimensional data sets. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a flow chart of the interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification provided by the first aspect of the present invention;
[0019] Figure 2 It is a flow chart of the interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification provided by the second aspect of the present invention;
[0020] Figure 3 It is a schematic diagram of distribution curves of non-dominated solution sets finally obtained by all algorithms provided in the embodiments of the present invention on various data sets in the target space. DETAILED DESCRIPTION
[0021] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0022] refer to Figure 1 to Figure 2 As shown, the first embodiment of the present invention discloses an interpolation evolution method for processing a two-objective feature selection problem in high-dimensional data classification, which can be executed by an interpolation evolution device for processing a two-objective feature selection problem in high-dimensional data classification (hereinafter referred to as an evolution device), and in particular, executed by one or more processors in the evolution device to implement the following method:
[0023] S101, obtain the difference to be evolved, set the population size N, the total number of features D and the total number of evolution iterations T, and set the current number of iterations C to 0;
[0024] S102, using an interpolation initialization strategy IBI to preprocess the difference to generate an initial population, wherein the preprocessing includes deploying a plurality of evenly distributed interpolation points in the target space, creating interpolation subpopulations corresponding to the interpolation points, analyzing the distribution of solutions of all interpolation subpopulations to obtain a local search area of interest, and creating four special subpopulations in the local search area of interest to merge into an initial population;
[0025] Specifically, step S102 includes:
[0026] Set the number of interpolations NIP to The rounded integer value of , the average interval H of interpolation in the target space f1 direction is set to 1 / (NIP+1), the standard size NSP of the interpolation subpopulation is set to the rounded integer value of N / NIP, the best f2 target value variable BF is initialized to infinity, and the cumulative number of interpolation subpopulations k created is initialized to 0;
[0027] The cumulative creation number k is judged to determine whether the cumulative creation number k is greater than or equal to the interpolation number NIP, and according to the judgment result, a new interpolation subpopulation is continuously created in a loop to form a new interpolation subpopulation subPop, or the interpolation subpopulation loop creation process is terminated to form a special subpopulation Pop2;
[0028] After terminating the interpolation subpopulation loop creation process according to the judgment result, a special subpopulation Pop3 based on the forward search idea and a special subpopulation Pop4 based on the backward search idea are generated;
[0029] The four special sub-populations Pop1, Pop2, Pop3 and Pop4 are merged into a mixed population Hop, and the N optimal solutions are selected from the mixed population Hop to form the final initial population Pop using the non-dominated sorting and crowding distance methods.
[0030] Preferably, the cumulative creation quantity k is judged and processed as follows:
[0031] When it is determined that the cumulative creation number k is greater than or equal to the interpolation number NIP, the process of cyclic creation of the interpolation subpopulation is terminated, and N solutions are randomly generated according to the best interpolation position BIP obtained in the cycle as the probability threshold to form a special subpopulation Pop2;
[0032] When it is determined that the cumulative creation quantity k is less than the interpolation quantity NIP, a new interpolation sub-population is continuously created in a loop.
[0033] Preferably, a new interpolation subpopulation is created in a loop according to the judgment result, specifically:
[0034] The actual size t1 of the interpolation subpopulation to be created is set to the smaller value between the standard size NSP of the interpolation subpopulation and (N-NSP*k), and the interpolation position t2 in the direction of the target space f1 is set to H*(k+1);
[0035] Using t2 as the probability threshold, randomly generate t1 solutions to form a new interpolation subpopulation subPop;
[0036] Get the minimum value of all solutions in subPop in the direction of the f2 target value and set it as t3;
[0037] When it is determined that the target value variable BF is greater than t3, the target value variable BF is set to t3 and the best interpolation position BIP is set to the current interpolation position t2, and subPop is merged into the special subpopulation Pop1, wherein the special subpopulation Pop1 is initially empty;
[0038] The cumulative creation number k is added by 1, and the cumulative creation number k is judged to determine whether the cumulative creation number k is greater than or equal to the interpolation number NIP, and the process of looping to create new interpolation sub-populations or terminating the interpolation sub-population loop creation process is continued according to the judgment result.
[0039] Preferably, t1 solutions are randomly generated with t2 as the probability threshold to form a new interpolation subpopulation subPop, specifically: a t1*D zero matrix is generated, each column of each row of the zero matrix is processed, and a random decimal p between 0 and 1 is generated. When it is judged that p is less than t2, the value of the current column of the current row is set to 1.
[0040] Preferably, a special subpopulation Pop3 based on the forward search concept is generated, specifically:
[0041] Set the size h1 of the special subpopulation Pop3 to be created to the minimum value between the population size N and the total number of features D;
[0042] Generate an h1*D zero matrix, process each row of the zero matrix, and set the value of a random column of the current row to 1, where the random column must be different from the previous column.
[0043] Preferably, a special subpopulation Pop4 based on the backward search concept is generated, specifically:
[0044] Set the size h2 of the special subpopulation Pop4 to be created to the minimum value between the population size N and the total number of features D;
[0045] Generate an h2*D matrix full of 1s, process each row of the matrix, and set the value of a random column of the current row to 0, where the random column must be different from the previous column.
[0046] In this embodiment, the interpolation initialization strategy (IBI) is the core component of the proposed interpolation evolution method (IBEA). Specifically, the IBI method first adaptively deploys a series of uniformly distributed interpolation points in the target space, and then creates corresponding interpolation subpopulations based on these interpolation points, and then obtains a most promising and potential local search area of interest by analyzing the distribution of solutions of all interpolation subpopulations. Then, a new special subpopulation Pop2 is created in the area, and two special subpopulations Pop3 and Pop4 based on the forward search idea and the backward search idea are created at the extreme points at both ends of the target space. All the interpolation subpopulations created previously can also form a special subpopulation Pop1, and then these four special subpopulations Pop1, Pop2, Pop3 and Pop4 are merged together to form a large mixed population. Finally, N solutions that need to be retained are selected from the mixed population through the classical non-dominated sorting method and the crowding distance method (assuming that N is the required initial population size), ultimately forming a competitive initial population that can be used for subsequent evolutionary iterations.
[0047] S103, judging whether the current iteration reaches the loop termination condition, and outputting the final evolved population Pop or continuing the evolutionary iteration process of the current population according to the judgment result, wherein the loop termination condition is whether the current iteration number C is greater than or equal to the total number of evolutionary iterations T.
[0048] Specifically, step S103 includes: when it is determined that the current number of iterations C is greater than or equal to the total number of evolution iterations T, terminating the entire evolution iteration process and outputting the final evolved population Pop;
[0049] When it is determined that the current number of iterations C is less than the total number of evolutionary iterations T, the evolutionary iteration process of the current population continues.
[0050] Preferably, the evolutionary iteration process of the current population is continued according to the judgment result, specifically:
[0051] Randomly select N pairs of parent solutions from the current population Pop, and perform a crossover operation and a mutation operation with a probability of 1 / D on each pair of parent solutions to generate a set Sop consisting of N offspring solutions;
[0052] Merge the current population Pop with the offspring solution set Sop to form a joint population Uop;
[0053] Remove duplicate solutions in the joint population Uop until there are no duplicate solutions or there are only N solutions left in the Uop;
[0054] Through the classic non-dominated sorting and crowding distance methods, N optimal solutions are selected from the joint population Uop and retained, and all the solutions in Pop are replaced with the solutions in Uop to form the latest evolved population;
[0055] The current iteration count C is added by 1 to determine whether the current iteration has reached the loop termination condition, and the final evolved population Pop is output based on the determination result or the evolutionary iteration process of the current population is continued.
[0056] Preferably, an effective crossover operation and a mutation operation with a probability of 1 / D are performed on each pair of parent solutions to generate a set Sop consisting of N offspring solutions, specifically:
[0057] Randomly crossover the equivalent gene positions with different values in the two parent solutions;
[0058] Determine whether it mutates bit by bit with a probability of 1 / D;
[0059] If so, flip the value of the gene bit from 1 to 0 or from 0 to 1.
[0060] In this embodiment, the core innovation of IBEA (Interpolation Evolution Method), i.e., the interpolation evolution method for processing the two-target feature selection problem in high-dimensional data classification, is that it designs and adopts an interpolation-based adaptive initialization strategy (Interpolation Based Initialization, referred to as IBI) to create an initial population for iterative evolution. The initialization strategy adopted by traditional MOEAs usually randomly generates an initial population directly in the middle of the target space by randomly sampling from the decision space. This may be feasible for the feature space of low-dimensional classification data sets, but for the feature space of high-dimensional classification data sets, because the range of feature space that needs to be searched is exponentially expanded, the feasible solutions become more sparse, so it is difficult to obtain sufficient diversity for the initial population created by traditional methods, and it is also difficult to cover the fast search area of real interest in the target space (the so-called region of interest).
[0061] Compared with the prior art, the interpolation evolutionary method for processing the two-target feature selection problem in high-dimensional data classification is different in that the interpolation subpopulations adaptively created by the interpolation initialization strategy (IBI) adopted by IBEA evenly cover the entire target space over a large range, and the three special subpopulations created afterwards deeply explore multiple local interest search areas in the target space to a certain extent. Among them, the two special subpopulations Pop3 and Pop4 created based on the forward search idea and the backward search idea explore the local interest areas at the first and last extreme positions of the target space, while the other special subpopulation Pop2 explores a certain intermediate interest search area obtained by analyzing the distribution of all previous interpolation subpopulations. Therefore, the IBI strategy adopted by the IBEA method takes into account both diversity and convergence. At the same time, its adaptive exploration characteristics also make IBEA more universal and robust to classification data of different dimensions. As the saying goes, well begun is half done. The initial population created by IBI (interpolation initialization strategy) is the cornerstone and main reason why IBEA (interpolation evolution method) can achieve excellent search performance and optimization effects.
[0062] In this embodiment, in order to more comprehensively and accurately verify the technical effect of the IBEA method, a total of 20 open source classification data sets derived from real-life data were used to test the optimization performance of each control method in solving the two-objective feature selection problem. The detailed information of these data sets is shown in Table 1, and the total number of features of the data set, the total number of classification samples, and the total number of classification categories are respectively shown in different columns of Table 1. It can be seen that these classification data sets used for testing cover a very wide range of feature dimensions. Although several low-dimensional data are also included as controls, they are mainly focused on testing high-dimensional data. More specific data set information can be queried from the following website https: / / archive.ics.uci.edu / . For example, the "Sonar" data set contains various modes of sonar signals reflected from metal cylinders at different angles.
[0063] Table 1 Information of various classification datasets used for test problems
[0064] Serial number Dataset name Number of features Number of samples Number of categories 1 Sonar 60 208 2 2 HillValley 100 606 2 3 Synthetic 100 200 5 4 Arrhythmia 278 452 13 5 Yale 1024 165 15 6 Colon 2000 62 2 7 SRBCT 2308 83 4 8 AR10P 2400 130 10 9 PIE10P 2420 210 10 10 Leukemia1 5327 72 3 11 Tumor9 5726 60 9 12 Brain1 5920 90 5 13 Leukemia2 7070 72 2 14 CNS 7129 60 2 15 ALLAML 7129 72 2 16 Nci9 9712 60 9 17 Pixraw10P 10000 100 10 18 Orlraws10P 10304 100 10 19 Brain2 10367 50 4 20 Prostate 10509 102 2
[0065] In order to verify the technical effect of the IBEA method, four classic excellent MOEAs were introduced to participate in the same test experiment as IBEA. These MOEAs used for comparison are the classic method NSGA-II based on non-dominated sorting, the classic method MOEA / D based on decomposition, the latest method MOEA / HD based on hierarchical decomposition, and the latest method PMEA based on extreme index. To ensure fairness, these methods will use the same internal parameter settings as stated in their respective literatures, but share other common experimental parameters. Specifically, all participating methods, including IBEA, use the same population size (N=100), the same evolutionary iteration termination condition (T=100), and the same experimental equipment environment. In the classification process, the KNN classifier (K=5) will be used to randomly divide the classification data into training set and test set according to the ratio of 7 / 3, and perform 10-fold cross validation. Each experiment will be run 20 times independently and randomly to obtain average statistical results with less fluctuation.
[0066] In terms of the evaluation index of method performance, two evaluation indexes, hypervolume HV and inverse distance IGD, which are widely used in the field of MOEAs, are adopted. Both indicators can take into account the diversity of optimization methods and the evaluation of convergence performance to a certain extent, and both require setting reference points. Specifically, the HV indicator only needs to set a common reference point as (1, 1), while the IGD indicator needs to set a whole set of reference points as the set of all final non-dominated solutions obtained by the current test method and test problem feedback. It should be noted that the HV evaluation index requires the larger the value, the better, while the IGD evaluation index is just the opposite, requiring the smaller the value, the better. The average HV and IGD performance statistics of each method after performing 20 independent random experiments on each test problem (classification data set) are shown in Tables 2 and 3, respectively, including the statistical mean and the corresponding positive and negative standard deviations.
[0067] Table 2 Average HV performance statistics of each method on each data set (the best value is marked in gray)
[0068]
[0069]
[0070] Table 3 Average IGD performance statistics of each method on each data set (the best value is marked in gray)
[0071]
[0072]
[0073] From the statistical results of the average HV and IGD performance in Tables 2 and 3, it can be seen that the proposed interpolation evolution method IBEA shows the best HV and IGD average performance on all tested classification data sets, whether for low-dimensional data sets or high-dimensional data sets. It is worth mentioning that the performance of IBEA on high-dimensional data sets is even more significant than that on low-dimensional data sets compared with other control methods. This is mainly because the interpolation initialization strategy IBI adopted by IBEA shows a more obvious search ability advantage on high-dimensional data sets, both in diversity and convergence. Considering that HV and IGD are comprehensive performance indicators of multiple objectives, they also take into account the evaluation of classification accuracy and classification efficiency. Therefore, it can be inferred that the classification results fed back by the final evolutionary population obtained by IBEA show the best performance in both classification accuracy and classification efficiency.
[0074] In addition, to further verify the computational complexity of the IBEA method, the experiment also counted the average operation time (in seconds) of each method running 20 times on each test data set in Table 4. As can be seen from Table 4, the average operation time of the proposed method IBEA on all test data sets is the smallest and best among all control methods, and the average operation time advantage on high-dimensional data sets is even more obvious. This means that the computational efficiency of IBEA in dealing with the two-target feature selection problem in high-dimensional classification data is far superior to other traditional MOEAs, and it can achieve better classification performance with less computing time.
[0075] Table 4 Statistics of average operation time of each method on each data set (the optimal value is marked in gray)
[0076]
[0077]
[0078] In order to more intuitively verify the diversity and convergence of the IBEA method, the experiment is also Figure 3 The distribution curves of the final non-dominated solution set obtained by all control methods on each test data set in the target space are shown. It should be noted that, for the sake of fairness, each distribution curve selects the result of the median performance of the method in 20 random independent tests to remove the interference of the best or worst performance. Figure 3 As shown in the figure, the 20 subgraphs correspond to 20 test data sets. Overall, IBEA performs significantly better than all the control methods, and its non-dominated solution set has more diverse solution distributions and more convergent target values in the f1 and f2 directions. Specifically, Figure 3 In the first four subgraphs, IBEA's solution set distribution advantage is not obvious enough compared with other methods, and the solution set distribution curves of multiple methods are close to each other. This is because IBEA's search performance advantage cannot be fully exerted on low-dimensional data sets. However, starting from high-dimensional data sets, IBEA's solution set distribution in all subgraphs has opened up a very obvious gap between all other control methods, leaving the solution set distribution curves of other algorithms far behind.
[0079] In summary, the interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification first uses an adaptive initialization strategy based on interpolation points to create a competitive initial population, and then repeatedly performs offspring reproduction and environmental selection operations on the initial population until the evolution termination condition is reached (accumulated to a certain preset number of iterations) and then outputs the final evolved population. Each solution vector contained in the population corresponds to a D-dimensional row vector, in which the element value 1 represents the selected feature and 0 represents the unselected feature. Each solution vector corresponds to a two-dimensional target vector, that is, minimizing the selected feature rate corresponding to the optimization target f1 direction, and minimizing the classification error rate corresponding to the optimization target f2 direction. This overcomes the limitations and various shortcomings of traditional MOEAs in processing two-objective feature selection problems on high-dimensional data sets.
[0080] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. An interpolation evolutionary method for processing the two-objective feature selection problem in high-dimensional data classification, characterized in that: include: Get the difference to be evolved, set the population size N, the total number of features D and the total number of evolution iterations T, and set the current number of iterations C to 0; The difference is preprocessed using an interpolation initialization strategy IBI to generate an initial population, wherein the preprocessing includes deploying a plurality of evenly distributed interpolation points in the target space, creating interpolation subpopulations corresponding to the interpolation points, analyzing the distribution of solutions of all interpolation subpopulations to obtain a local search area of interest, and creating four special subpopulations in the local search area of interest to merge into an initial population; Determine whether the current iteration has reached the loop termination condition, and output the final evolved population Pop or continue the evolutionary iteration process of the current population based on the judgment result, wherein the loop termination condition is whether the current iteration number C is greater than or equal to the total number of evolutionary iterations T.
2. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 1 is characterized in that: The interpolation initialization strategy IBI is used to preprocess the difference to generate an initial population, specifically: Set the number of interpolations NIP to The rounded integer value of , the average interval H of interpolation in the target space f1 direction is set to 1 / (NIP+1), the standard size NSP of the interpolation subpopulation is set to the rounded integer value of N / NIP, the best f2 target value variable BF is initialized to infinity, and the cumulative number of interpolation subpopulations k created is initialized to 0; The cumulative creation number k is judged to determine whether the cumulative creation number k is greater than or equal to the interpolation number NIP, and according to the judgment result, a new interpolation subpopulation is continuously created in a loop to form a new interpolation subpopulation subPop, or the interpolation subpopulation loop creation process is terminated to form a special subpopulation Pop2; After terminating the interpolation subpopulation loop creation process according to the judgment result, a special subpopulation Pop3 based on the forward search idea and a special subpopulation Pop4 based on the backward search idea are generated; The four special sub-populations Pop1, Pop2, Pop3 and Pop4 are merged into a mixed population Hop, and the N optimal solutions are selected from the mixed population Hop to form the final initial population Pop using the non-dominated sorting and crowding distance methods.
3. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 2 is characterized in that: The cumulative creation quantity k is judged and processed as follows: When it is determined that the cumulative creation number k is greater than or equal to the interpolation number NIP, the process of cyclic creation of the interpolation subpopulation is terminated, and N solutions are randomly generated according to the best interpolation position BIP obtained in the cycle as the probability threshold to form a special subpopulation Pop2; When it is determined that the cumulative creation quantity k is less than the interpolation quantity NIP, a new interpolation sub-population is continuously created in a loop.
4. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 2 is characterized in that: Continue to create new interpolation sub-populations in a loop according to the judgment results, specifically: The actual size t1 of the interpolation subpopulation to be created is set to the smaller value between the standard size NSP of the interpolation subpopulation and (N-NSP*k), and the interpolation position t2 in the direction of the target space f1 is set to H*(k+1); Using t2 as the probability threshold, randomly generate t1 solutions to form a new interpolation subpopulation subPop; Get the minimum value of all solutions in subPop in the direction of the f2 target value and set it as t3; When it is determined that the target value variable BF is greater than t3, the target value variable BF is set to t3 and the best interpolation position BIP is set to the current interpolation position t2, and subPop is merged into the special subpopulation Pop1, wherein the special subpopulation Pop1 is initially empty; The cumulative creation number k is added by 1, and the cumulative creation number k is judged to determine whether the cumulative creation number k is greater than or equal to the interpolation number NIP, and the process of looping to create new interpolation sub-populations or terminating the interpolation sub-population loop creation process is continued according to the judgment result.
5. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 4 is characterized in that: With t2 as the probability threshold, t1 solutions are randomly generated to form a new interpolation subpopulation subPop, specifically: Generate a t1*D zero matrix, process each column of each row of the zero matrix, generate a random decimal p between 0 and 1, and when it is judged that p is less than t2, set the value of the current column of the current row to 1.
6. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 2, characterized in that: Generate a special subpopulation Pop3 based on the forward search idea, specifically: Set the size h1 of the special subpopulation Pop3 to be created to the minimum value between the population size N and the total number of features D; Generate an h1*D zero matrix, process each row of the zero matrix, and set the value of a random column of the current row to 1, where the random column must be different from the previous column.
7. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 2 is characterized in that: Generate a special subpopulation Pop4 based on the backward search idea, specifically: Set the size h2 of the special subpopulation Pop4 to be created to the minimum value between the population size N and the total number of features D; Generate an h2*D matrix full of 1s, process each row of the matrix, and set the value of a random column of the current row to 0, where the random column must be different from the previous column.
8. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 1 is characterized in that: Determine whether the current iteration has reached the loop termination condition, specifically: When it is determined that the current number of iterations C is greater than or equal to the total number of evolutionary iterations T, the entire evolutionary iteration process is terminated and the final evolved population Pop is output; When it is determined that the current number of iterations C is less than the total number of evolutionary iterations T, the evolutionary iteration process of the current population continues.
9. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 1, characterized in that: The evolutionary iteration process of the current population continues according to the judgment result, specifically: Randomly select N pairs of parent solutions from the current population Pop, and perform a crossover operation and a mutation operation with a probability of 1 / D on each pair of parent solutions to generate a set Sop consisting of N offspring solutions; Merge the current population Pop with the offspring solution set Sop to form a joint population Uop; Remove duplicate solutions in the joint population Uop until there are no duplicate solutions or there are only N solutions left in the Uop; Through the classic non-dominated sorting and crowding distance methods, N optimal solutions are selected from the joint population Uop and retained, and all the solutions in Pop are replaced with the solutions in Uop to form the latest evolved population; The current iteration count C is added by 1 to determine whether the current iteration has reached the loop termination condition, and the final evolved population Pop is output based on the determination result or the evolutionary iteration process of the current population is continued.
10. The interpolation evolution method for processing the two-objective feature selection problem in high-dimensional data classification according to claim 9, characterized in that: Perform an effective crossover operation and a mutation operation with a probability of 1 / D on each pair of parent solutions to generate a set Sop consisting of N offspring solutions, specifically: Randomly crossover the equivalent gene positions with different values in the two parent solutions; Determine whether it mutates bit by bit with a probability of 1 / D; If so, flip the value of the gene bit from 1 to 0 or from 0 to 1.
