Feature selection method, system and storage medium based on quantum forensic investigation mechanism

Through the feature selection method based on the quantum forensic investigation mechanism, the iterative optimization of the K nearest neighbor classifier and quantum police group is solved, and the problems of low classification accuracy and high time complexity in the existing feature selection method are achieved, and efficient feature sub- Set selection, suitable for datasets in different dimensions.

CN117992807BActive Publication Date: 2024-11-22HARBIN ENG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410172704.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-07
Publication Date
2024-11-22
Estimated Expiration
2044-02-07

AI Technical Summary

Technical Problem

The existing feature selection methods have low classification accuracy and high time complexity, and the selected feature subset accuracy is insufficient, making it difficult to achieve the effects of fast convergence speed, high convergence accuracy and low time complexity at the same time.

Method used

Using a feature selection method based on the quantum forensic investigation mechanism, the data set is normalized and the use of K nearest neighbor classifiers are randomly generated, and quantum police groups are divided into investigation teams and pursuit teams to perform quantum position updates and fitness calculations. , iteratively optimize to select the optimal subset of features.

Benefits of technology

It significantly reduces the time overhead of feature selection, improves the accuracy and classification accuracy of feature subsets, has faster convergence speed, higher convergence accuracy and lower time complexity, and is suitable for low-dimensional and high-dimensional datasets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117992807B_ABST
    Figure CN117992807B_ABST
Patent Text Reader

Abstract

A feature selection method, system and storage medium based on quantum forensic investigation mechanism, which relates to the field of data processing, is to solve the problems of low classification accuracy and high time complexity of existing feature selection methods, and insufficient accuracy of selected feature subsets. It includes: step 1: normalize the data set, cluster the training set samples using K nearest neighbor classifier, classify the test set samples, and construct the objective function of feature selection; step 2: initialize the quantum position of each member in the quantum police group, calculate the fitness of the quantum police individual, and determine the initial global optimal quantum position; step 3: update the quantum position of each member of the investigation group and the pursuit group respectively; step 4: update the global optimal quantum position of the investigation group, the pursuit group and the quantum police group; step 5: exchange some members in the investigation group and the pursuit group, and continue to evolve; step 6: finally iterate to obtain the selected optimal feature subset.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a feature selection method, system and storage medium based on a quantum forensic investigation mechanism. Background Art

[0002] Feature selection technology is a method of selecting the most effective feature subset from the original features to reduce the dimension of the data set and optimize the specific indicators of the system. It is an important means to improve the performance of the learning algorithm and a key data preprocessing step in pattern recognition. Existing feature selection methods can be divided into embedded feature selection methods, filtering feature selection methods and encapsulated feature selection methods according to whether the evaluation criteria of the feature subset are combined with the subsequent learning algorithm. Among them, the encapsulated feature selection method is widely used because it combines the feature selection problem with the subsequent learning method and uses the classifier to evaluate the advantages and disadvantages of the selected feature subset, which makes the performance of the selected subset have a greater advantage. At the same time, encapsulated feature selection is often combined with various swarm intelligence algorithms to improve the efficiency of feature selection, reduce time complexity, and increase the accuracy of the selected feature subset. However, due to the difficulty and complexity of engineering problems, the optimization ability of the existing swarm intelligence algorithm is difficult to achieve the effects of good convergence, high convergence accuracy and low time complexity at the same time. Therefore, it is of great theoretical value and research significance to design a feature selection method with short feature selection time, good performance of the selected feature subset and strong robustness.

[0003] In existing research, Jia Heming et al. published "Synchronous Optimization Feature Selection Based on Improved Vulture Search Algorithm" in Control and Decision (2022, Vol.37, No.02, pp.445-454), which proposed a synchronous optimization feature selection model that uses an algorithm to simultaneously optimize support vector machine parameters and perform feature selection. The improved vulture algorithm that combines the vulture algorithm with the simulated annealing algorithm has improved convergence speed and convergence accuracy to a certain extent, but because the vulture algorithm is a swarm intelligence algorithm with low convergence accuracy, it is difficult to break through the limitations of the algorithm itself by using traditional methods to improve the algorithm accordingly. Therefore, this method has not achieved very ideal results in the selection of feature subsets, and the training overhead of using support vector machines to implement multi-classification problems is large, and feature selection takes a long time. In “A hybrid Harris Hawks optimization algorithm with simulated annealing for feature selection” published in Artificial Intelligence Review (2020, vol. 54, pp. 1-45), Mohamed Abdel-Basset et al. proposed a Harris Hawk algorithm based on simulated annealing and applied the algorithm to the encapsulated feature selection using the K nearest neighbor method as the evaluation indicator of the learning algorithm. This method improves the Harris Hawk algorithm and improves the convergence of the Harris Hawk algorithm. However, since the time advantage of the improved Harris Hawk algorithm is still not obvious, the encapsulated feature selection model designed by utilizing the simplicity and high timeliness of the K nearest neighbor method cannot be better utilized. The advantage of low time complexity. In “Quantum based Whale Optimization Algorithm for wrapper feature selection” published in Artificial Intelligence Review (2020, vol. 89), RKAgrawal et al. proposed a double-chain coded quantum whale algorithm and applied it to encapsulated feature selection. The quantum coding of the whale algorithm greatly improves the convergence speed and convergence accuracy of the algorithm, and can take into account the time taken. The encapsulated feature selection model that combines the quantum whale algorithm with four traditional classification methods has achieved a high classification accuracy and a low average number of features. However, since the whale algorithm is an earlier proposed swarm intelligence algorithm, its convergence is obviously insufficient, and the quantum rotation angle update formula of the designed quantum whale algorithm is too simple, so the accuracy of the algorithm needs to be further improved.

[0004] It can be seen that the existing encapsulated feature selection method will face problems such as low convergence accuracy and high time cost when applying traditional optimization methods to feature selection. It is difficult to take into account the convergence speed, convergence accuracy and time used when improving the swarm intelligence method using common improvement strategies. In order to achieve the goal of rapid convergence of intelligent optimization methods and high superiority of the selected feature subset, how to improve the convergence speed, convergence accuracy and reduce the time used is a technical problem that needs to be solved urgently. Summary of the invention

[0005] The technical problems to be solved by the present invention are:

[0006] The existing feature selection methods have low classification accuracy and high time complexity, and the selected feature subsets are not accurate enough.

[0007] The present invention adopts the following technical solutions to solve the above technical problems:

[0008] The present invention provides a feature selection method based on quantum forensic investigation mechanism, comprising the following steps:

[0009] Step 1: Normalize the data set, randomly divide the processed data set into training set and test set, cluster the training set samples using K nearest neighbor classifier, classify the test set samples, construct the objective function of feature selection based on the classification accuracy of K nearest neighbor classifier and the number of selected features, and randomly generate a group of quantum policemen containing multiple quantum policemen representing different feature subsets;

[0010] Step 2: Initialize the quantum position of each member of the quantum police group, use the objective function to calculate the fitness of the individual quantum police, divide the quantum police group into two sub-groups, the investigation group and the pursuit group, and determine the initial global optimal quantum position of each sub-group and the quantum police group;

[0011] Step 3: Update the quantum position of each member of the investigation team and the pursuit team simultaneously;

[0012] Step 4: Calculate the fitness value of each individual quantum policeman after the quantum position is updated, and update the global optimal quantum position of the investigation group, the pursuit group and the quantum police group;

[0013] Step 5: Exchange some members of the investigation team and the pursuit team to form a new investigation team and pursuit team to continue to perform steps 3 to 4;

[0014] Step 6: Increase the number of iterations by 1, and repeat steps 3 to 5 until the maximum number of iterations is reached; finally, the optimal feature subset is selected.

[0015] Furthermore, step one includes the following steps:

[0016] (1) For the data set I = [(z1,y1), (z2,y2), ..., (z L ,y L )], where Z = [z1,z2,...,z L ] is the data sample in the data set, Y=[y1,y2,...,y L ] is the category label in the data set, L is the total number of data samples in the data set, and each data sample has D feature elements in the feature vector, i.e. z i =[z i1 ,z i2 ,...,z iD ], i = 1, 2, ..., L, D is the number of features contained in the data set;

[0017] Normalize the data set and randomly select data samples with a proportion of α1 and their category labels as the training set The rest of the data is the test set Traverse [k min ,k max ] all odd values ​​in the range are the values ​​of k, and the 0-1 loss function is used as the cross-validation loss function to determine [k min ,k max ], where k is the number of samples contained in each neighborhood of the K nearest neighbor method, k min is the lower bound of the k value search, k max The upper bound for the k value search;

[0018] (2) Let the jth data sample in I′2 be z′ j , initialize j = α1L+1;

[0019] (3) Calculate z′ j Euclidean distance to all data samples in the training set where z′ i is the i-th data sample in the training set, i=1,2,...,α1L;

[0020] (4) Find the value in the training set corresponding to z′ j The k data samples with the smallest Euclidean distance will cover the z′ of these k data samples. j The neighborhood of k (z′ j );

[0021] (5) Determine z′ by majority voting j Category label where y iis the category label corresponding to z′i, i=1,2,...,α1L, h=1,2,...,u, is the indicator function, that is, when hour The value of is 1; otherwise it is 0;

[0022] (6) If j = L, end the determination of the test set data category label and complete the test set classification; otherwise, let j = j + 1 and repeat steps (3) to (5) until the conditions for the end of classification are met.

[0023] Furthermore, the objective function in step 1 is specifically: Among them, χ is the classification accuracy, which represents the proportion of classification accuracy in the fitness function, and γ R is the classification accuracy obtained by the K-nearest neighbor classifier, ξ is the importance of the selected features, which represents the weight of the number of selected features in the fitness function, ξ=1-χ, is the number of selected feature subsets.

[0024] Furthermore, step 2 includes the following process:

[0025] Initialize the quantum position of the d-th dimension of the n-th quantum policeman of the first generation to a random number between [0,1], and the quantum position of the n-th quantum policeman of the t-th generation is N is the number of quantum police groups, D is the maximum dimension of the search space; the quantum position is measured according to the measurement rules to obtain the position of each quantum police The d-dimensional quantum position of the n-th quantum policeman of the t-th generation The corresponding measurement position is The measurement rule is in is a random number that satisfies uniform distribution in the interval;

[0026] The quantum police group is divided into two subgroups: the investigation group and the pursuit group. For the investigation group, the quantum police are labeled For the pursuit team, the quantum police are labeled According to the fitness function Calculate the initial fitness value of each individual quantum policeman, sort the initial fitness values ​​of the members of the investigation group and the pursuit group respectively, set the quantum positions of the members with the best initial fitness values ​​in the two subgroups as the initial global optimal quantum positions of the two subgroups respectively, compare the optimal fitness values ​​of the two subgroups, and set the quantum positions of the members with better fitness values ​​in the two subgroups as the initial global optimal quantum positions of the quantum police group.

[0027] Furthermore, step three includes the following steps:

[0028] The quantum revolving gate update process of the d-dimensional quantum position of the n-th member of the investigation team is: in The quantum rotation angle when the investigation team was investigating and collecting evidence, and is a random number between (0,1), is the optimal quantum position of the d-th dimension of the n-th member of the investigation group during the iteration process, is the quantum position of the d-th dimension of the global optimal solution of the investigation group during the first t iterations, is a random number uniformly distributed between [0,1], and β1 is the probability of mutation during the investigation of the survey group;

[0029] The quantum revolving gate update process of the d-dimensional quantum position of the n-th member of the pursuit group is: in is the quantum rotation angle of the pursuit group during the action phase, and is a standard normal random number between (0,1), is the d-th dimension quantum position of the global optimal solution of the pursuit group during the first t iterations, is the d-th dimension average quantum position of the hunting group, is a random number uniformly distributed between [0,1], and β2 is the mutation probability of the pursuit group during the pursuit process.

[0030] Furthermore, the probability that the nth member of the investigation team is selected to join the pursuit team in step 5 is The probability that the nth member of the pursuit team is selected to join the investigation team is

[0031] Furthermore, it is used in the fields of data classification, biomedicine, image processing and communication signal processing.

[0032] A feature selection system based on quantum forensic investigation mechanism, the system has a program module corresponding to the steps of any one of the above technical solutions, and executes the steps in the feature selection method based on quantum forensic investigation mechanism when running.

[0033] A computer-readable storage medium storing a computer program, wherein the computer program is configured to implement the steps of the feature selection method based on the quantum forensic investigation mechanism described in any one of the above technical solutions when called by a processor.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The present invention designs a single-chain coded quantum forensic investigation mechanism, discretizes the forensic investigation mechanism of the continuous optimization problem, and combines it with a K-nearest neighbor classifier without an obvious learning process, which greatly reduces the time required for encapsulated feature selection and reduces time overhead. At the same time, compared with the filtering feature selection method, the selected feature subset has higher accuracy.

[0036] 2. The quantum forensic investigation mechanism method proposed in the invention, which is based on quantum coding and simulated quantum revolving door evolution, has better global convergence and convergence speed than other traditional intelligent optimization methods, and can be used to solve discrete optimization problems. It breaks through the application limitations of the original forensic investigation mechanism and can select feature subsets with higher accuracy in feature selection problems.

[0037] 3. The feature selection method based on the quantum forensic investigation mechanism of the present invention has the advantages of high classification accuracy and low number of features in the selected feature subsets under low-dimensional and high-dimensional data sets. For data sets with higher dimensions, it can obviously achieve the effect of dimensionality reduction of the data set. At the same time, with less prior knowledge, it can still select high-precision feature subsets. Therefore, the method of the present invention has faster convergence speed, higher convergence accuracy, lower time complexity, stronger robustness and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of a feature selection method based on a quantum forensic investigation mechanism in an embodiment of the present invention;

[0039] Figure 2 It is a curve diagram of fitness curve simulation results of the same low-dimensional data set under the same initial value in an embodiment of the present invention;

[0040] Figure 3 It is a curve diagram of fitness curve simulation results of the same high-dimensional data set under the same initial value in an embodiment of the present invention;

[0041] Figure 4 It is a comparison diagram of fitness value simulation results of the same low-dimensional data set under the same initial value in an embodiment of the present invention;

[0042] Figure 5 It is a comparison diagram of fitness value simulation results of the same high-dimensional data set under the same initial value in an embodiment of the present invention;

[0043] Figure 6 This is a comparison chart of the average values ​​of the feature numbers selected by different methods under different data sets in the embodiment of the present invention. From left to right under each data set are QFBI, DBI, JS and PSO;

[0044] Figure 7This is a comparison chart of classification accuracy obtained by different methods under different data sets in an embodiment of the present invention. From left to right in each data set are QFBI, DBI, JS and PSO. DETAILED DESCRIPTION

[0045] In the description of the present invention, it should be noted that the terms "first", "second", and "third" mentioned in the embodiments of the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second", and "third" may explicitly or implicitly include one or more of the features.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Specific implementation plan 1: Figure 1 As shown, the present invention provides a feature selection method based on quantum forensic investigation mechanism, comprising the following steps:

[0048] Step 1: Normalize the data set, randomly divide the processed data set into training set and test set, cluster the training set samples using K nearest neighbor classifier, classify the test set samples, construct the objective function of feature selection based on the classification accuracy of K nearest neighbor classifier and the number of selected features, and randomly generate a group of quantum policemen containing multiple quantum policemen representing different feature subsets;

[0049] Step 2: Initialize the quantum position of each member of the quantum police group, use the objective function to calculate the fitness of the individual quantum police, divide the quantum police group into two sub-groups, the investigation group and the pursuit group, and determine the initial global optimal quantum position of each sub-group and the quantum police group;

[0050] Step 3: Update the quantum position of each member of the investigation team and the pursuit team simultaneously;

[0051] Step 4: Calculate the fitness value of each individual quantum policeman after the quantum position is updated, and update the global optimal quantum position of the investigation group, the pursuit group and the quantum police group;

[0052] Step 5: Exchange some members of the investigation team and the pursuit team to form a new investigation team and pursuit team to continue evolving;

[0053] Step 6: Add 1 to the number of iterations, and repeat steps 3 to 5 until the maximum number of iterations is reached; output the global optimal quantum position and the corresponding mapping position of the quantum police group. The set of features with a median value of 1 in the mapping position is the selected optimal feature subset, and the classification accuracy and classification results are obtained at the same time.

[0054] Specific implementation plan 2: Step 1 includes the following steps:

[0055] (1) For the data set I = [(z1,y1), (z2,y2), ..., (z L ,y L )], where Z = [z1,z2,...,z L ] is the data sample in the data set, Y=[y1,y2,...,y L ] is the category label in the data set, L is the total number of data samples in the data set, and each data sample has D feature elements in the feature vector, i.e. z i =[z i1 ,z i2 ,...,z iD ], i = 1, 2, ..., L, D is the number of features contained in the data set;

[0056] Normalize the data set and convert all data samples and category labels in the data set into numbers. The data samples are The maximum value of the data sample is z max =[z 1,max ,z 2,max ,...,z n,max ], the minimum value is z min =[z 1,min ,z 2,min ,...,z n,min ],in For the The maximum value of the characteristic elements, For the The minimum value of the characteristic elements, Using the normalization formula Normalize all data samples in the data set to obtain the normalized data set I′=[(z′1,y1),(z′2,y2),...,(z′ L ,y L )], where z′ i =[z′ i1 ,z′ i2 ,...,z′ in ], i = 1, 2, ..., L; let the category to which the data sample belongs be ω = [ω1, ω2, ..., ω u ],in Indicates kind, u is the number of categories contained in the data sample, and the values ​​of the category labels corresponding to different categories are in Indicates The category label corresponding to the class data sample,

[0057] For the features l1,l2,...,l in the dataset n , if the value is 1, it means that the feature is selected; otherwise, the value is 0 and the feature is not selected. Set all feature data with corresponding values ​​of 1 in the data set as the feature subset selected by feature selection, and set the number of features contained in the feature subset to be

[0058] In the feature subset, randomly select data samples and their category labels with a proportion of α1 as the training set, and the rest of the data as the test set. Renumber the data samples and category labels in the training set and the test set. Let the numbered training set be The rest of the data is the test set In I′1, randomly select data samples with a proportion of α2 and the corresponding category labels as the training subset, and the remaining data in the training set as the validation set. min ,k max ] all odd values ​​in the range are the values ​​of k, and the 0-1 loss function is used as the cross-validation loss function to determine [k min ,k max ], where k is the number of samples contained in each neighborhood of the K nearest neighbor method, k min is the lower bound of the k value search, k max The upper bound for the k value search;

[0059] (2) Let the jth data sample in I′2 be z′ j , initialize j = α1L+1;

[0060] (3) Calculate z′ j Euclidean distance to all data samples in the training set where z′ i is the i-th data sample in the training set, i=1,2,...,α1L;

[0061] (4) Find the value in the training set corresponding to z′ j The k data samples with the smallest Euclidean distance will cover the z′ of these k data samples. j The neighborhood of k (z′ j );

[0062] (5) Determine z′ by majority voting j Category label where yi is z′ iThe corresponding category labels are i=1,2,...,α1L, h=1,2,...,u, is the indicator function, that is, when hour The value of is 1; otherwise it is 0;

[0063] (6) If j = L, the determination of the test set data category label is completed and the test set classification is completed; otherwise, let j = j + 1 and repeat steps (3) to (5) until the classification end condition is met. The rest of this implementation is the same as the specific implementation plan 1.

[0064] This implementation scheme compares the category labels obtained from the test with the correct category labels. If the two are the same, the classification is correct, otherwise the classification is wrong. The percentage of correctly classified data samples to the total number of test set data samples is calculated, and the calculation result is set as the classification accuracy of the K nearest neighbor method.

[0065] Specific implementation scheme three: The objective function described in step one is specifically: Among them, χ is the classification accuracy, which represents the proportion of classification accuracy in the fitness function, and γ R is the classification accuracy obtained by the K-nearest neighbor classifier, ξ is the importance of the selected features, which represents the weight of the number of selected features in the fitness function, ξ=1-χ, The other aspects of this embodiment are the same as those of the second embodiment.

[0066] Specific implementation plan 4: Step 2 includes the following process:

[0067] Initialize the quantum position of the d-th dimension of the n-th quantum policeman of the first generation to a random number between [0,1], and the quantum position of the n-th quantum policeman of the t-th generation is N is the number of quantum police groups, D is the maximum dimension of the search space; the quantum position is measured according to the measurement rules to obtain the position of each quantum police The d-dimensional quantum position of the n-th quantum policeman of the t-th generation The corresponding measurement position is The measurement rule is in is a random number that satisfies a uniform distribution in the interval.

[0068] The quantum police group is divided into two subgroups: the investigation group and the pursuit group. For the investigation group, the quantum police are labeled For the pursuit team, the quantum police are labeled According to the fitness function The classification accuracy obtained in the K-nearest neighbor classifier model calculates the initial fitness value of each individual quantum policeman, sorts the initial fitness values ​​of the members of the investigation group and the pursuit group respectively, sets the quantum positions of the members with the best initial fitness values ​​in the two subgroups as the initial global optimal quantum positions of the two subgroups respectively, compares the optimal fitness values ​​of the two subgroups, and sets the quantum positions of the members with better fitness values ​​in the two subgroups as the initial global optimal quantum positions of the quantum police group. The rest of this implementation plan is the same as the specific implementation plan one.

[0069] Specific implementation plan five: Step three includes the following steps:

[0070] The investigation team is responsible for investigating the location of the suspect and determining the most suspicious location for search. The quantum revolving door update process of the d-dimensional quantum position of the n-th member of the investigation team is: in The quantum rotation angle when the investigation team was investigating and collecting evidence, and is a random number between (0,1), is the optimal quantum position of the d-th dimension of the n-th member of the investigation team during the iteration process, that is, the quantum position of the d-th dimension when the fitness value of the n-th member of the investigation team is optimal, is the quantum position of the dth dimension of the global optimal solution of the investigation group during the first t iterations, that is, the quantum position of the quantum police individual with the best fitness value in the investigation group during the iteration. If the quantum rotation angle Quantum Bits Using quantum NOT gate to update with a smaller probability, this process can be expressed as Therefore, during the investigation, the update formula for the d-dimensional quantum position of the n-th member of the investigation team is expressed as is a random number uniformly distributed between [0,1], and β1 is the probability of mutation during the investigation of the survey group;

[0071] The pursuit team determines the location of the action through the tracking results and takes action. The quantum revolving gate update process of the d-dimensional quantum position of the n-th member of the pursuit team is: in is the quantum rotation angle of the pursuit group during the action phase, and is a standard normal random number between (0,1), is the d-th dimension quantum position of the global optimal solution of the pursuit group during the first t iterations, that is, the quantum position of the quantum police individual with the best fitness value of the pursuit group during the iteration process. is the d-th dimension average quantum position of the hunting group, If the quantum rotation angle Quantum Bits Using quantum NOT gate to update with a smaller probability, this process can be expressed as Therefore, during the pursuit process, the update formula for the d-dimensional quantum position of the n-th member of the pursuit team is expressed as in is a random number uniformly distributed between [0,1], and β2 is the probability of variation during the pursuit process by the pursuit group. The rest of this implementation is the same as the first implementation.

[0072] Specific implementation plan 6: Step 4 includes the following process: the quantum position of the nth quantum police member in the t+1th iteration Perform measurement and use the measurement rules to measure the position of the nth quantum policeman The selected feature subsets are divided into training sets and test sets, and the training sets and test sets are input into the K nearest neighbor classifier model to obtain the classification accuracy. The obtained classification accuracy and the number of selected feature subsets are substituted into the fitness function to calculate the corresponding fitness value, and the fitness value of the quantum police individual in the t+1th iteration is recorded. The quantum police individuals with the best fitness values ​​in the investigation group and the pursuit group are found and their quantum positions are recorded. If the optimal fitness value of the quantum police in the investigation group and the pursuit group in the t+1th iteration is less than the fitness value corresponding to the global optimal solution of the two subgroups obtained in the previous t iterations, then the quantum positions corresponding to the quantum police individuals with the best fitness values ​​in the two subgroups during the t+1th iteration are set as the global optimal quantum positions of the two subgroups; otherwise, the global optimal quantum positions of the investigation group and the pursuit group in the t+1th iteration are equal to the global optimal quantum positions of the two subgroups in the previous t iterations. Let the global optimal solutions of the investigation group and the pursuit group in the t+1th iteration be and Compare the fitness values ​​of the global optimal quantum positions of the two sub-populations, take the quantum position with the smallest fitness value as the global optimal quantum position of the entire quantum police population, and obtain the global optimal position of the quantum police population through measurement rules. The rest of this implementation plan is the same as the specific implementation plan one.

[0073] Specific implementation plan seven: In step five, in order to ensure the smooth progress of the operation, the investigation team and the pursuit team members should exchange necessary information. The roulette method is used to select U members from the investigation team and the pursuit team to enter the other team according to probability. The probability that the nth member of the investigation team is selected to enter the pursuit team is The probability that the nth member of the pursuit team is selected to join the investigation team is The selected members are transferred between groups to form new investigation groups and pursuit groups to continue the evolution. The rest of this implementation plan is the same as the first implementation plan.

[0074] Specific implementation scheme eight: The application of the feature selection method based on quantum forensic investigation mechanism is used for feature selection in the fields of data classification, biomedicine, image processing and communication signal processing.

[0075] The feature selection method based on quantum forensic investigation mechanism described in this implementation scheme is used to simplify the feature data in the data set, remove redundant features, and obtain the optimal feature subset under specified conditions. The selected data sets are 7 general data sets in the UCI official website.

[0076] Specific implementation scheme nine: A feature selection system based on quantum forensic investigation mechanism, the system has a program module corresponding to the steps of any one of the above implementation schemes one to eight, and executes the steps in the above feature selection method based on quantum forensic investigation mechanism when running.

[0077] Specific implementation scheme ten: A computer-readable storage medium, the computer-readable storage medium storing a computer program, the computer program being configured to implement the steps in the feature selection method based on the quantum forensic investigation mechanism described in any one of implementation schemes one to eight when called by a processor.

[0078] Example 1

[0079] In order to verify the effect of the present invention, the method of the present invention is described by using the following examples.

[0080] For the convenience of description, the feature selection method of the quantum forensic investigation mechanism proposed in the present invention is abbreviated as QFBI, and the methods used for comparison are based on the forensic investigation mechanism, artificial jellyfish search mechanism and particle swarm mechanism, which are abbreviated as FBI, JS and PSO respectively.

[0081] Set the parameters of QFBI to α1 = 40%, α2 = 50%, k min =1, k max =15, c=5, χ=0.99, The relevant parameters of JS are: time control constant is 0.5, distribution factor is 3, and motion coefficient is 0.1; the relevant parameters of PSO are: inertia weight is 0.75, cognitive coefficient is 1.8, and social coefficient is 2. The population size and maximum number of iterations of the four mechanisms are the same, and their values ​​are N = 30 and T = 50 respectively. Each method is run independently 30 times, and the average fitness of 30 runs is taken to draw the fitness curve.

[0082] like Figure 2 and Figure 3 As shown, regardless of the dimension of the data set, the QFBI of the present invention has the advantages of fast convergence speed and high convergence accuracy compared with FBI, JS and PSO, which fully reflects the advantages of QFBI.

[0083] like Figure 4 and Figure 5 As shown, the QFBI of the present invention has a smaller variance in fitness value than FBI, JS and PSO, and performs better in system stability.

[0084] In order to more accurately measure the effect of each feature selection method, data sets of 7 different dimensions were selected for feature selection, and three indicators were selected for measurement: average classification accuracy, average number of selected features, and average running time (total running time divided by 30). Average classification accuracy is the ratio of the classification accuracy obtained by the classifier constructed by the K nearest neighbor method to the number of repeated experiments for classifying the selected feature subset. It is used to measure the accuracy of the selected feature subset for the subsequent learning algorithm. The larger the value, the higher the accuracy of the selected feature subset. The average number of selected features is the ratio of the number of features contained in the selected feature subset to the number of repeated experiments. It is used to measure the effect of feature selection on the dimensionality reduction of the data set. The smaller the value, the better the effect of feature selection. Average running time is the ratio of the total time of each optimization method after repeated running to the number of repeated experiments. It is used to measure the time complexity of each optimization method. The lower the value, the lower the time complexity of the optimization method used.

[0085] The results are shown in Table 1 and Figure 6 and Figure 7 As shown, it can be seen that under data sets of different dimensions, the feature subsets selected by QFBI have the highest accuracy and the least number of features.

[0086] Table 1

[0087]

[0088]

[0089] It can be seen from Table 1 that among the 84 groups of data tested under the same conditions, the present invention is superior to 83 groups of data of the comparison method and 1 group of data that is the same as the comparison method. It can be seen that the feature subset selected by the feature selection method based on the quantum forensic investigation mechanism of the present invention is significantly better than the result of feature selection by the comparison mechanism in terms of accuracy and the number of features contained in the subset, and it also has great advantages in time. The method of the present invention can take into account the convergence speed, convergence accuracy and time complexity at the same time, has strong robustness, and is a feasible, effective and high-precision feature selection method.

[0090] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A feature selection method based on quantum forensic investigation mechanism, characterized in that, The steps include: Step 1: Normalize the image data set, randomly divide the processed image data set into a training set and a test set, cluster the training set samples using the K nearest neighbor classifier, and classify the test set samples. Based on the classification accuracy of the K nearest neighbor classifier and the number of selected features, the objective function of feature selection is constructed, and a group of quantum policemen representing different feature subsets is randomly generated; The objective function is specifically: Among them, χ is the classification accuracy, which represents the proportion of classification accuracy in the fitness function, and γ R is the classification accuracy obtained by the K-nearest neighbor classifier, ξ is the importance of the selected features, which represents the weight of the number of selected features in the fitness function, ξ=1-χ, is the number of selected feature subsets; Step 2: Initialize the quantum position of the d-th dimension of the n-th quantum policeman of the first generation to a random number between [0,1], and the quantum position of the n-th quantum policeman of the t-th generation is N is the number of quantum police groups, D is the maximum dimension of the search space; the quantum position is measured according to the measurement rules to obtain the position of each quantum police The d-dimensional quantum position of the n-th quantum policeman of the t-th generation The corresponding measurement position is The measurement rule is in is a random number that satisfies uniform distribution in the interval; The quantum police group is divided into two subgroups: the investigation group and the pursuit group. For the investigation group, the quantum police are labeled For the pursuit team, the quantum police are labeled According to the fitness function Calculate the initial fitness value of each quantum police individual, sort the initial fitness values ​​of the members of the investigation group and the pursuit group respectively, set the quantum positions of the members with the best initial fitness values ​​in the two subgroups as the initial global optimal quantum positions of the two subgroups respectively, compare the optimal fitness values ​​of the two subgroups, and set the quantum positions of the members with better fitness values ​​in the two subgroups as the initial global optimal quantum positions of the quantum police group; Step 3: Update the quantum position of each member of the investigation team and the pursuit team simultaneously; Step 4: Calculate the fitness value of each individual quantum policeman after the quantum position is updated, and update the global optimal quantum position of the investigation group, the pursuit group and the quantum police group; Step 5: Exchange some members of the investigation team and the pursuit team to form a new investigation team and pursuit team to continue to perform steps 3 to 4; Step 6: Increase the number of iterations by 1, and repeat steps 3 to 5 until the maximum number of iterations is reached; finally, the optimal feature subset is selected.

2. The feature selection method based on quantum forensic investigation mechanism according to claim 1 is characterized in that, Step 1 includes the following steps: (1) For the data set I = [(z1,y1), (z2,y2), ..., (z L ,y L )], where Z = [z1,z2,...,z L ] is the data sample in the data set, Y=[y1,y2,...,y L ] is the category label in the data set, L is the total number of data samples in the data set, and each data sample has D feature elements in the feature vector, i.e. z i =[z i1 ,z i2 ,...,z iD ], i = 1, 2, ..., L, D is the number of features contained in the data set; Normalize the data set and randomly select data samples with a proportion of α1 and their category labels as the training set The rest of the data is the test set Traverse [k min ,k max ] all odd values ​​in the range are the values ​​of k, and the 0-1 loss function is used as the cross-validation loss function to determine [k min ,k max ], where k is the number of samples contained in each neighborhood of the K nearest neighbor method, k min is the lower bound of the k value search, k max The upper bound for the k value search; (2) Let the jth data sample in I′2 be z′ j , initialize j = α1L+1; (3) Calculate z′ j Euclidean distance to all data samples in the training set where z′ i is the i-th data sample in the training set, i=1,2,...,α1L; (4) Find the value in the training set corresponding to z′ j The k data samples with the smallest Euclidean distance will cover the z′ of these k data samples. j The neighborhood of k (z′ j ); (5) Determine z′ by majority voting j Category label where y i is z′ i The corresponding category labels are i=1,2,...,α1L, h=1,2,...,u, is the indicator function, that is, when hour The value of is 1; otherwise it is 0; (6) If j = L, end the determination of the test set data category label and complete the test set classification; otherwise, let j = j + 1 and repeat steps (3) to (5) until the conditions for the end of classification are met.

3. The feature selection method based on quantum forensic investigation mechanism according to claim 1 is characterized in that, Step three includes the following steps: The quantum revolving gate update process of the d-dimensional quantum position of the n-th member of the investigation team is: in The quantum rotation angle when the investigation team was investigating and collecting evidence, and is a random number between (0,1), is the optimal quantum position of the d-th dimension of the n-th member of the investigation group during the iteration process, is the quantum position of the d-th dimension of the global optimal solution of the investigation group during the first t iterations, is a random number uniformly distributed between [0,1], and β1 is the probability of mutation during the investigation of the survey group; The quantum revolving gate update process of the d-dimensional quantum position of the n-th member of the pursuit group is: in is the quantum rotation angle of the pursuit group during the action phase, and is a standard normal random number between (0,1), is the d-th dimension quantum position of the global optimal solution of the pursuit group during the first t iterations, is the d-th dimension average quantum position of the hunting group, is a random number uniformly distributed between [0,1], and β2 is the mutation probability of the pursuit group during the pursuit process.

4. The feature selection method based on quantum forensic investigation mechanism according to claim 1 is characterized in that, The probability that the nth member of the investigation team is selected to join the pursuit team in step 5 is The probability that the nth member of the pursuit team is selected to join the investigation team is 5. A feature selection system based on quantum forensic investigation mechanism, characterized in that, The system has a program module corresponding to the steps of any one of claims 1 to 4 above, and executes the steps in the feature selection method based on the quantum forensic investigation mechanism when running.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps in the feature selection method based on quantum forensic investigation mechanism according to any one of claims 1 to 4 when called by a processor.