A gene expression profile classification method and system based on feature dependency and multi-objective particle swarm optimization feature selection

Abstract

The application provides a gene expression profile classification method and system based on feature dependence and multi-objective particle swarm optimization feature selection, comprising the following steps: calculating the dependence score of each feature of the gene expression profile data, quantifying the dependence relationship between features by information entropy and mutual information; initializing the particle swarm based on the dependence score, and preferentially selecting features with high dependence scores; adopting a multi-objective particle swarm optimization algorithm for feature selection, combining the classification error rate and the feature selection rate as the objective function, and realizing the balance between low error rate and feature subset selection in the optimization process; adopting an improved particle optimal position updating strategy, generating new solutions through a non-dominated solution updating mechanism based on cubic spline interpolation, and improving the global search ability of the particle swarm; and obtaining the optimal feature subset through the multi-objective optimization process. Through the optimization of the feature selection process and the improvement of the particle swarm optimization strategy, the application improves the classification performance and the efficiency of feature selection, and can effectively capture the dependence relationship between genes.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection, which can be widely used in tasks such as high-dimensional data analysis, machine learning, and gene data classification. Background Technology

[0002] Human capabilities for analyzing gene expression data have advanced significantly in recent years. However, in today's society, with the dramatic increase in biomedical data, relying solely on manual methods for gene data analysis is not only time-consuming and labor-intensive but also inefficient. With the rapid development of computer technology and machine learning algorithms, gene data classification has gradually become an important technology in the life sciences. Various complex and precise mathematical models have emerged, enabling computers to help researchers process large amounts of complex gene expression data, providing support for disease diagnosis and personalized medicine.

[0003] In the field of bioinformatics, gene expression profiling is a crucial component. With the advent of the big data era, gene expression profiling has become an important resource for studying disease mechanisms and gene function. Researchers are demanding higher accuracy in gene expression data classification. However, due to the characteristic of gene data being typically "high-dimensional and small-sample," traditional classification algorithms struggle to effectively handle redundant features and noise in high-dimensional data, resulting in poor classification performance. Therefore, developing efficient feature selection methods is of significant practical importance for gene data classification research.

[0004] Gene expression data analysis is a typical high-dimensional data processing problem. Redundant features and noise not only increase computational complexity but also reduce the accuracy and generalization ability of classification models. Therefore, how to efficiently select key features while reducing computational overhead has become a core issue in the field of gene expression data classification. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a gene expression profile classification method and system based on feature dependence and multi-target particle swarm optimization of feature selection, aiming to optimize feature selection, improve classification accuracy, and capture the dependencies between genes.

[0006] Note that the description of these objectives does not preclude the existence of other objectives. One aspect of the invention does not require achieving all of the above objectives. Objectives other than those described above can be extracted from the description, drawings, and claims.

[0007] In feature selection and classification problems, traditional methods often fail to adequately consider the dependencies between features, leading to redundant features that negatively impact optimization results. Furthermore, it is difficult to balance feature selection scale and classification performance during optimization. To address these issues, this invention proposes a feature selection method based on feature dependencies using multi-objective particle swarm optimization. This method optimizes population initialization through feature dependencies and incorporates an improved particle position update strategy, significantly improving both feature selection efficiency and classification performance. Particle swarm optimization (PSO) is a heuristic search algorithm that finds the optimal solution in the solution space through the collaborative interaction of individuals and the population. However, traditional PSO algorithms, such as MOEA / D-PSO, suffer from premature convergence and uneven solution distribution when handling high-dimensional data and complex optimization problems, limiting their optimization performance. To address these issues, this invention employs a multi-objective particle swarm optimization algorithm, combining a feature-dependent population initialization strategy with a particle individual optimal position update mechanism based on cubic spline interpolation. Through dual optimization of classification error rate and feature selection rate, global optimization of the feature selection problem is achieved.

[0008] This invention presents a gene expression profile data classification method based on feature dependency and multi-objective particle swarm optimization (PSO) for feature selection optimization. This method improves classification accuracy and captures inter-gene dependencies. First, it introduces a feature dependency score calculation method based on information entropy and mutual information to quantify the dependencies between features, thereby guiding the population initialization of the PSO algorithm, reducing redundant features, and improving the quality of the initial population. Subsequently, an improved particle optimal position update strategy is employed, generating new non-dominated solutions based on cubic spline interpolation, thus enhancing the global search capability of the particles and the uniformity of solution distribution. Finally, a multi-objective PSO algorithm is used to obtain a final method with low classification error rate and feature selection rate, capable of capturing inter-gene dependencies.

[0009] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0010] A gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection includes the following steps:

[0011] Step S1: Download gene expression profile data, calculate the dependency score for each feature, and quantify the dependency relationship between features using information entropy and mutual information methods;

[0012] Step S2: Based on the dependency scores obtained in Step S1, the first generation of particle swarm is initialized using a population initialization strategy based on feature dependency, prioritizing features with high dependency scores to generate the initial positions of particles.

[0013] Step S3: Based on the initial particle positions generated in step S2, the feature selection problem is optimized using a multi-objective particle swarm optimization algorithm. The classification error rate and feature selection rate are combined as objective functions, and the fitness value is calculated using the objective function.

[0014] Step S4: By calculating and comparing the fitness values ​​generated in step S3, select the individual optimal and global optimal for each population; when updating the individual optimal, use a particle optimal position update strategy based on cubic spline interpolation to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front.

[0015] Step S5: Through a multi-objective optimization process, select the optimal feature subset on the Pareto front generated in step S4;

[0016] Step S6: Input the optimal feature subset selected in step S5 into multiple classifiers for training and testing to verify the effectiveness of the method;

[0017] Step S7: If the termination condition is met, proceed to step S8; otherwise, jump to step S4 and loop.

[0018] Step S8: End.

[0019] In the above scheme, step S1 includes the following steps:

[0020] Step S1.1: Calculate the univariate information content of each feature using information entropy;

[0021] Step S1.2: Calculate the mutual information value between each feature and the other features, and quantify the dependency of the features using the dependency score formula, whereby the dependency score calculation formula is as follows:

[0022]

[0023] Among them, f j and f k Representing the j-th and k-th features respectively, I(f j ;f k ) represents feature f j and feature f k Mutual information values ​​between them;

[0024] Step S1.3: Normalize the dependency scores and use them as the basis for weights in the subsequent population initialization process.

[0025] Furthermore, step S1.3 also includes the following steps:

[0026] The selection probability p for each feature is calculated for subsequent initialization, as shown in the following formula:

[0027]

[0028] Where DS(j) represents the dependency score of the j-th feature, λ = 0.8, σ = 10.

[0029] In the above scheme, step S2 includes the following steps:

[0030] Step S2.1: Initialize the parameters of the particle swarm optimization algorithm PSO, including the number of iterations T (maxi-mumiterations), inertia weight ω, random variables r1 and r2 between [0,1], determine the upper bound of particle velocity v_max, self-learning factor c1 and global learning factor c2;

[0031] Step S2.2: Generate an initial particle swarm based on dependency scores and transfer functions. A probabilistic selection strategy based on dependency scores is adopted, which makes features with high dependency scores more likely to be selected by particles. Each particle represents a subset of features, i.e., n populations are generated. The position, velocity and fitness of the particles are randomly initialized.

[0032] Furthermore, in step S2.2, the position vector of each particle represents a feature subset, and the value of each dimension is 0 or 1, where 1 indicates that the feature is selected and 0 indicates that it is not selected.

[0033]

[0034] Among them, s i,j The position vector of the particle, x i,j ∈[0,1] represents the j-th feature of the i-th particle, and p(j) is the selection probability of the j-th feature.

[0035] In the above scheme, step S3 includes the following steps:

[0036] Step S3.1: Set the first objective function as the feature selection rate. The calculation formula is as follows:

[0037]

[0038] Where D is the total number of features, i is the i-th particle, and s ij (t) represents the position information of the j-th feature of the i-th particle in the t-th iteration;

[0039] Step S3.2: Set the second objective function as follows: The formula for calculating the classification error rate is as follows:

[0040]

[0041] Among them, FP, FN, TP, and TN represent false positive, false negative, true positive, and true negative, respectively.

[0042] In the above scheme, step S4 includes the following steps:

[0043] Step S4.1: An update strategy based on cubic spline interpolation is adopted to smoothly interpolate the non-dominated solution set, generating new non-dominated solutions and improving the global performance of particle swarm search. The calculation formula of the interpolation function is as follows:

[0044] I i (x)=a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3

[0045] Where the coefficient a of the cubic polynomial i b i c i d i These are used to ensure the smoothness of the generated data, a i Control the value of the interpolation function at the nodes, b i The slope of the control curve is related to the slope of the interpolation function, c. i The curvature change is related to the second derivative of the curve, d i The influence control curve of the cubic term determines the smooth transition between nodes;

[0046] Step S4.2: Based on the distance between the non-dominated solution and the nearest Pareto solution set, if the GD(X) from the non-dominated solution to the Pareto front is less than the GD(X) from the nearest Pareto solution set, then uniformly cover the Pareto front using a cubic spline interpolation function.

[0047] Furthermore, the distance calculation formula in step S4.2 is as follows:

[0048]

[0049] Where N is the population size, q is the Euclidean distance, and D... i This represents the distance between the i-th solution point in the target space and the nearest solution point on the Pareto front.

[0050] In the above scheme, step S5 includes the following steps:

[0051] Step S5.1: By updating the position and velocity of the particles, guide the particles to move towards the Pareto optimal direction, and update the velocity v of each particle.i and position x i The formula is as follows:

[0052]

[0053] Where i represents the current particle, d represents the d-th dimension of the search space, t represents the t-th iteration of the evolutionary process, and ω represents the inertia weight. This represents the velocity of particle i in the d-th dimension at the t-th iteration. pbest represents the position of particle i in the d-th dimension at the t-th iteration. id gbest represents the best historical position of particle i in the d-th dimension. id Let represent the personal historical best position of particle i in the d-th dimension, c1 and c2 be acceleration constants, and r1 and r2 be random variables between [0,1].

[0054] Step S5.2: Use the fitness values ​​of the particles to filter the non-dominated solutions and archive them;

[0055] Step S5.3: Select the optimal solution from the archive to generate the final Pareto front.

[0056] In the above scheme, step S6 includes the following steps:

[0057] Step S6.1: Input the optimized feature subset into the classifier for training;

[0058] Step S6.2: Evaluate the classifier's classification performance on the test set to verify the effectiveness and classification accuracy of the selected features;

[0059] Step S6.3: To verify the feasibility of the algorithm, repeat the experiments in steps S6.1-S6.2 on the gene expression dataset and record the classification accuracy and the number of selected features.

[0060] A system for implementing the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection includes a feature score calculation module, a particle position initialization module, a fitness value calculation module, an interpolation calculation module, an optimal feature set selection module, and a method verification module.

[0061] The feature scoring calculation module is used to download the gene expression profile data Colon according to the GSE number, calculate the dependency score of each feature, and quantify the dependency relationship between features through information entropy and mutual information methods.

[0062] The particle position initialization module is used to initialize the first generation of particle swarm based on the dependency score obtained by the feature scoring module, and to give priority to features with high dependency scores to generate the initial position of the particles.

[0063] The fitness value calculation module is used to optimize the feature selection problem by using a multi-objective particle swarm optimization algorithm based on the initial particle positions generated by the particle position initialization module. The module combines the classification error rate and the feature selection rate as objective functions and calculates the fitness value using the objective function.

[0064] The interpolation calculation module is used to select the individual optimal and global optimal for each population by calculating and comparing the fitness values ​​generated by the fitness value calculation module; when updating the individual optimal, a particle optimal position update strategy based on cubic spline interpolation is used to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front.

[0065] The optimal feature set selection module is used to search for and select the optimal feature subset on the Pareto front generated by the interpolation calculation module through a multi-objective optimization process.

[0066] The method verification module is used to input the optimal feature subset selected by the optimal feature set selection module into multiple classifiers for training and testing, thereby verifying the effectiveness of the method.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] 1. Following step 1, this invention calculates the dependency score of each feature using information entropy and mutual information methods, proposing a feature-dependent population initialization strategy. This strategy utilizes the correlation information between features to optimize the initialization process of the particle swarm, prioritizing features with high dependency scores for particle position generation. Compared to traditional random initialization methods, this invention significantly improves the initial quality of the population, reduces the impact of redundant features, and lays an efficient foundation for subsequent optimization.

[0069] According to step 2, this invention optimizes the particle swarm based on population initialization and the priority of dependency scores, making the particles more evenly distributed in the feature space, thereby improving the optimization coverage and population diversity. By prioritizing the processing of highly dependent features, the algorithm's ability to identify key features is further enhanced.

[0070] According to step 3, this invention employs a multi-objective particle swarm optimization algorithm, combining two objective functions—classification error rate and feature selection rate—to comprehensively optimize the feature selection problem. By simultaneously optimizing the scale of feature selection and classification performance, this invention effectively avoids the problem that traditional single-objective methods cannot balance classification accuracy and feature quantity. Experiments show that this optimization process can significantly improve the algorithm's search capability and the quality of the final solution.

[0071] Based on step 4, this invention proposes a particle individual optimal position update strategy based on cubic spline interpolation. During individual optimal position update, cubic spline interpolation is used to generate new non-dominated solutions, smoothing the distribution of optimized solutions and enhancing the uniformity of solutions. Compared to traditional particle update mechanisms, this method can effectively improve the distribution coverage of optimized solutions in the target space, enhance optimization stability and global search performance, and alleviate the situation where the current particle and individual optimal positions are not mutually dominant.

[0072] According to step 5, this invention searches for Pareto front solutions through multi-objective optimization and selects the optimal feature subset in the target space. By screening and archiving Pareto front solutions, this invention achieves diversity and balance in the solution set, effectively solving the problem of uneven solution distribution in high-dimensional data optimization.

[0073] Therefore, this invention effectively solves the problems of redundant features, uneven solution distribution, and low optimization efficiency in gene data feature selection by employing a feature-dependent population initialization strategy and a particle individual optimal position update mechanism based on cubic spline interpolation. Experimental results show that this method exhibits excellent feature selection ability and classification performance on gene expression profile datasets, and is particularly suitable for complex tasks such as high-dimensional data analysis and gene data classification.

[0074] Note that the description of these effects does not preclude the existence of other effects. One aspect of the invention does not necessarily have all the aforementioned effects. Effects other than those described above can be readily observed and extracted from the description, drawings, claims, etc. Attached Figure Description

[0075] Figure 1 This is a flowchart of the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to the present invention;

[0076] Figure 2 This is a comparison chart of the effects of multi-objective optimization feature selection on two objectives in this invention;

[0077] Figure 3 This is a graph comparing the accuracy of gene expression profile data before and after the algorithm.

[0078] Figure 4 This invention relates to a gene expression profile-assisted diagnostic system based on feature selection.

[0079] Figure 5 This is the data analysis and verification interface. Detailed Implementation

[0080] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0081] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "front," "rear," "left," "right," "upper," "lower," "axial," "radial," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0082] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0083] Example 1

[0084] Figure 1 The image shows a preferred embodiment of the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection mainly includes the following steps:

[0085] Step S1: Feature Score Calculation. Download the gene expression profile data Colon based on the GSE number, calculate the dependency score for each feature, and quantify the dependencies between features using information entropy and mutual information methods;

[0086] Step S2: Particle position initialization. Based on the dependency scores obtained in Step S1, the first-generation particle swarm is initialized using a feature-dependent population initialization strategy, prioritizing features with high dependency scores to generate initial particle positions;

[0087] Step S3: Fitness value calculation. Based on the initial particle positions generated in Step S2, the feature selection problem is optimized using a multi-objective particle swarm optimization algorithm. The classification error rate and feature selection rate are combined as objective functions, and the fitness value is calculated using the objective function.

[0088] Step S4: Interpolation calculation. By calculating and comparing the fitness values ​​generated in step S3, the individual optima and global optima of each population are selected; when updating the individual optima, a particle optimal position update strategy based on cubic spline interpolation is used to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front;

[0089] Step S5: Optimal Feature Set Selection. Through a multi-objective optimization process, the optimal feature subset is selected based on the Pareto front generated in step S4.

[0090] Step S6: Method Validation. The optimal feature subset selected in Step S5 is input into multiple classifiers for training and testing to validate the effectiveness of the method.

[0091] Step S7: If the termination condition is met, proceed to step S8; otherwise, jump to step S4 and loop.

[0092] Step S8: End.

[0093] The flowchart of the entire process is as follows Figure 1 As shown.

[0094] Step S1 includes the following steps:

[0095] Step S1.1: Calculate the univariate information content of each feature in the Colon dataset using information entropy; information entropy is an indicator of feature uncertainty. For a discrete feature X, its information entropy is defined as:

[0096]

[0097] Where the prior probability of X = x is denoted by p(x), and given another discrete variable Y, the conditional entropy H(X|Y) can be calculated as follows:

[0098]

[0099] Wherein, given Y=y, the conditional probability of X=x is represented by p(x|y), while p(x,y) represents the joint probability of X=x and Y=y. The greater the information entropy, the higher the uncertainty of the feature. Information entropy reflects the "information content" of a single feature and is used to quantify the contribution of the feature to the classification of the sample.

[0100] Step S1.2: Calculate the mutual information value between each feature and the other features, and quantify the dependency of the features using the dependency score formula, whereby the dependency score calculation formula is as follows:

[0101]

[0102] Among them, f j and f k Representing the j-th and k-th features respectively, I(f j ;f k ) represents feature f j and feature f k Mutual information values ​​between them;

[0103] Step S1.3: Normalize the dependency scores and use them as weights in the subsequent population initialization process. Simultaneously, calculate the selection probability p for each feature for subsequent initialization, as shown in the following formula:

[0104]

[0105] Where DS(j) represents the dependency score of the j-th feature, λ = 0.8, σ = 10.

[0106] Step 2 above includes the following steps:

[0107] Step S2.1: Initialize the parameters of the Particle Swarm Optimization (PSO) algorithm, such as the number of iterations T = 100, the inertia weight ω = 0.749, random variables r1 and r2 between [0,1], determine the upper bound of particle velocity v_max = 0.6, and set both the self-learning factor c1 and the global learning factor c2 to 1.46.

[0108] Step S2.2: Generate an initial particle swarm based on dependency scores and transfer functions. A probability selection strategy based on dependency scores is adopted, which makes features with high dependency scores more likely to be selected by particles. Each particle represents a feature subset, that is, n populations are generated. The position vector of each particle represents a feature subset, and the value of each dimension is 0 or 1, where 1 indicates that the feature is selected and 0 indicates that it is not selected.

[0109]

[0110] Among them, s i,j The position vector of the particle, x i,j∈[0,1] represents the j-th feature of the i-th particle, and p(j) is the selection probability of the j-th feature.

[0111] Step S3 includes the following steps:

[0112] Step S3.1: Set the first objective function as the feature selection rate. The calculation formula is as follows:

[0113]

[0114] Where D is the total number of features, i is the i-th particle, and s ij (t) represents the position information of the j-th feature of the i-th particle in the t-th iteration;

[0115] Step S3.2: Set the second objective function as follows: The formula for calculating the classification error rate is as follows:

[0116]

[0117] Among them, FP, FN, TP, and TN represent false positive, false negative, true positive, and true negative, respectively.

[0118] Furthermore, step S4 includes the following steps:

[0119] Step S4.1: Employ an update strategy based on cubic spline interpolation to smoothly interpolate the non-dominated solution set, generating new non-dominated solutions and improving the global performance of particle swarm optimization. First, select all non-dominated solutions in the current particle swarm using a multi-objective particle swarm optimization algorithm. Second, perform cubic spline interpolation on the selected non-dominated solution set. The goal of cubic spline interpolation is to connect these solution points with a smooth curve, thereby generating a new set of solutions. The calculation formula for the interpolation function is as follows:

[0120] I i (x)=a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3

[0121] Where the coefficient a of the cubic polynomial i b i c i d i These are used to ensure the smoothness of the generated data, a i Control the value of the interpolation function at the nodes, b i The slope of the control curve is related to the slope of the interpolation function, c.i The curvature change is related to the second derivative of the curve, d i The influence control curve of the cubic term determines the smooth transition between nodes;

[0122] Step S4.2: Based on the distance between the non-dominated solution and the nearest Pareto solution set, if the GD(X) from the non-dominated solution to the Pareto front is less than the GD(X) from the nearest Pareto solution set, then uniformly cover the Pareto front using a cubic spline interpolation function; the distance calculation formula is as follows:

[0123]

[0124] Where N is the population size, q = 2 (Euclidean distance), and D i This represents the distance between the i-th solution point in the target space and the nearest solution point on the Pareto front.

[0125] Furthermore, step S5 includes the following steps:

[0126] Step S5.1: Guide the particles towards the Pareto optimal direction by updating their position and velocity. Each particle updates its own velocity v. i Position x i The formula is as follows:

[0127]

[0128] Where i represents the current particle, d represents the d-th dimension of the search space, t represents the t-th iteration of the evolutionary process, and ω represents the inertia weight. This represents the velocity of particle i in the d-th dimension at the t-th iteration. pbest represents the position of particle i in the d-th dimension at the t-th iteration. id gbest represents the best historical position of particle i in the d-th dimension. id Let represent the personal historical best position of particle i in the d-th dimension, c1 and c2 be acceleration constants, and r1 and r2 be random variables between [0,1].

[0129] Step S5.2: Use the fitness values ​​of the particles to filter the non-dominated solutions and save them to an external archive;

[0130] Step S5.3: Select the optimal solution from the external archive to generate the final Pareto front.

[0131] Furthermore, step S6 includes the following steps:

[0132] Step S6.1: Input the optimized feature subset into the classifier (KNN, SVM) for training;

[0133] Step S6.2: Evaluate the classifier's classification performance on the test set to verify the effectiveness and classification accuracy of the selected features;

[0134] Step S6.3: Finally, to verify the feasibility of the algorithm, this invention repeats steps S6.1-S6.2 of the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection on the Colon gene expression data set, and records the classification accuracy and the selected feature rate.

[0135] The Pareto front of the optimized feature selection method's classification error rate and the selected feature rate is visualized as follows: Figure 2 As shown.

[0136] like Figure 2 As shown, compared with the original traditional multi-objective optimization method MOEA / D-FS, the Pareto front of this invention is closer to the origin of the target on the gene expression data Colon, and the points distributed are also closer to the x-axis and y-axis. It can be seen that the optimized method significantly reduces the classification error rate and the number of selected features, with obvious effects.

[0137] like Figure 3 As shown in the bar chart, compared with the original traditional multi-objective optimization method MOEA / D-FS, our invention significantly improves the classification accuracy on the Colon gene expression data, and achieves an accuracy of 0.9100 (i.e., 91%) on KNN. Figure 5 As shown, the optimized method achieves a 0.3-point improvement over the traditional multi-objective optimization method MOEA / D-FS, and a 0.6800 improvement over the traditional multi-objective optimization method MOEA / D-FS on SVM, which is 0.28 higher. This demonstrates that the optimized method significantly improves the classification accuracy on both KNN and SVM classifiers.

[0138] Example 2

[0139] A system for implementing the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection includes a feature score calculation module, a particle position initialization module, a fitness value calculation module, an interpolation calculation module, an optimal feature set selection module, and a method verification module.

[0140] The feature scoring calculation module is used to download the gene expression profile data Colon according to the GSE number, calculate the dependency score of each feature, and quantify the dependency relationship between features through information entropy and mutual information methods.

[0141] The particle position initialization module is used to initialize the first generation of particle swarm based on the dependency score obtained by the feature scoring module, and to give priority to features with high dependency scores to generate the initial position of the particles.

[0142] The fitness value calculation module is used to optimize the feature selection problem by using a multi-objective particle swarm optimization algorithm based on the initial particle positions generated by the particle position initialization module. The module combines the classification error rate and the feature selection rate as objective functions and calculates the fitness value using the objective function.

[0143] The interpolation calculation module is used to select the individual optimal and global optimal for each population by calculating and comparing the fitness values ​​generated by the fitness value calculation module; when updating the individual optimal, a particle optimal position update strategy based on cubic spline interpolation is used to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front.

[0144] The optimal feature set selection module is used to search for and select the optimal feature subset on the Pareto front generated by the interpolation calculation module through a multi-objective optimization process.

[0145] The method verification module is used to input the optimal feature subset selected by the optimal feature set selection module into multiple classifiers for training and testing, thereby verifying the effectiveness of the method.

[0146] According to this embodiment, preferably, Figure 4 The diagram shows a practical application of the present invention, specifically a cross-sectional view of a gene expression profile-assisted diagnostic system based on feature selection. This system uses the gene expression profile data classification method based on feature dependence and multi-objective particle swarm optimization feature selection described in Example 1 to classify gene expression profile data. Figure 5 The interface shown is the classification verification interface using the data from Example 1, which achieved an accuracy of 91% on KNN.

[0147] This invention presents a gene expression profiling data classification method based on feature dependency and multi-objective particle swarm optimization (PSO) for feature selection. It primarily aims to improve the accuracy of gene expression profiling data classification and effectively capture the dependencies between genes. The method first calculates the dependency score of each feature in the gene expression profiling data, quantifying the dependencies between features using information entropy and mutual information. Then, it initializes the PSO based on the dependency scores, prioritizing features with high dependency scores to reduce redundant features and improve the initialization quality. Next, a multi-objective PSO algorithm is used for feature selection, combining classification error rate and feature selection rate as objective functions to achieve a balance between low error rate and feature subset selection during the optimization process. Subsequently, a particle optimal update strategy is employed, generating new solutions through a non-dominated solution update mechanism based on cubic spline interpolation, enhancing the global search capability of the PSO. Finally, through the multi-objective optimization process, the optimal feature subset is obtained, and the selected feature subset is input into multiple classifiers for training and testing to verify the effectiveness of the method. This invention improves classification performance and feature selection efficiency by optimizing the feature selection process and improving the PSO optimization strategy, effectively capturing the dependencies between genes, thus providing a novel optimization method for gene expression profiling data classification.

[0148] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0149] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection, characterized in that, Includes the following steps: Step S1: Download gene expression profile data, calculate the dependency score for each feature, and quantify the dependency relationship between features using information entropy and mutual information methods; Step S2: Based on the dependency scores obtained in Step S1, the first generation of particle swarm is initialized using a population initialization strategy based on feature dependency, prioritizing features with high dependency scores to generate initial particle positions; Step S3: Based on the initial particle positions generated in Step S2, the feature selection problem is optimized using a multi-objective particle swarm optimization algorithm, combining the classification error rate and feature selection rate as objective functions, and the fitness value is calculated using the objective function. Step S4: By calculating and comparing the fitness values ​​generated in S3, select the individual optimal and global optimal for each population; when updating the individual optimal, use a particle optimal position update strategy based on cubic spline interpolation to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front. Step S5: Through a multi-objective optimization process, select the optimal feature subset on the Pareto front generated in step S4; Step S6: Input the optimal feature subset selected in step S5 into multiple classifiers for training and testing to verify the effectiveness of the method; Step S7: If the termination condition is met, proceed to step S8; otherwise, jump to step S4 and loop. Step S8: End.

2. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S1 includes the following steps: Step S1.1: Calculate the univariate information content of each feature using information entropy; Step S1.2: Calculate the mutual information value between each feature and the other features, and quantify the dependency of the features using the dependency score formula, wherein the dependency score calculation formula is as follows: in, and Representing the j-th and k-th features respectively, Representation of features and characteristics The mutual information value between them; Step S1.3: Normalize the dependency scores and use them as the basis for weights in the subsequent population initialization process.

3. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 2, characterized in that, Step S1.3 further includes the following steps: Calculate the selection probability for each feature For subsequent initialization, the formula is as follows: in, Let λ represent the dependency score of the j-th feature, λ=0.8, σ=10.

4. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Initialize the parameters of the Particle Swarm Optimization (PSO) algorithm, including the number of iterations T and the inertia weight. A random variable between [0,1] , Determine the upper bound of the particle's velocity. Self-learning factor and global learning factor Step S2.2: Generate an initial particle swarm based on dependency scores and transfer functions. A probabilistic selection strategy based on dependency scores is adopted, where each particle represents a feature subset, i.e., n populations are generated. The position, velocity, and fitness of the particles are randomly initialized.

5. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 4, characterized in that, In step S2.2, the position vector of each particle represents a feature subset, and the value of each dimension is 0 or 1, where 1 indicates that the feature is selected and 0 indicates that it is not selected. in, Represents the position vector of the particle. This represents the j-th feature of the i-th particle. Let be the selection probability of the j-th feature.

6. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: Set the first objective function as the feature selection rate calculation formula as follows: in, It is the total number of features. It is the first One particle, This represents the position information of the j-th feature of the i-th particle in the t-th iteration; Step S3.2: Set the second objective function as follows: The formula for calculating the classification error rate is as follows: in, These represent false positives, false negatives, true positives, and true negatives, respectively.

7. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: Using an update strategy based on cubic spline interpolation, smooth interpolation is performed on the non-dominated solution set to generate new non-dominated solutions, thereby improving the global performance of particle swarm search. The calculation formula of the interpolation function is as follows: Among them, the coefficients of the cubic polynomial , , , These are used to ensure the smoothness of the generated data. Controlling the value of the interpolation function at the nodes, The slope of the control curve is related to the slope of the interpolation function. The curvature change is related to the second derivative of the curve. The influence control curve of the cubic term is determined to smoothly transition between nodes; Step S4.2: Based on the distance between the non-dominated solution and the nearest Pareto solution set, if the GD(X) from the non-dominated solution to the Pareto front is less than the GD(X) from the nearest Pareto solution set, then the Pareto front is uniformly covered using a cubic spline interpolation function; the distance calculation formula is as follows: Where N is the population size. This represents the distance between the i-th solution point in the target space and the nearest solution point on the Pareto front.

8. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S5 includes the following steps: Step S5.1: By updating the position and velocity of the particles, guide the particles to move towards the Pareto optimal direction, and update the velocity of each particle. and location The formula is as follows: Where i represents the current particle, d represents the d-th dimension of the search space, t represents the t-th iteration of the evolutionary process, and ω represents the inertia weight. This represents the velocity of particle i in the d-th dimension at the t-th iteration. This represents the position of particle i in the d-th dimension during the t-th iteration. This represents the best historical position of particle i in the d-th dimension. Let c1 and c2 represent the global historical optimal position of particle i in the d-th dimension, and let r1 and r2 be acceleration constants and random variables between ∈ [0,1]. Step S5.2: Use the fitness value of the particle to filter the non-dominated solutions and archive them. Step S5.3: Select the optimal solution from the archive and generate the final Pareto front.

9. The gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection according to claim 1, characterized in that, Step S6 includes the following steps: Step S6.1: Input the optimized feature subset into the classifier for training; Step S6.2: Evaluate the classification performance of the classifier on the test set and verify the effectiveness and classification accuracy of the selected features; Step S6.3: To verify the feasibility of the algorithm, repeat the experiments of steps S6.1-S6.2 on the gene expression dataset and record the classification accuracy and the number of selected features.

10. A system for implementing the gene expression profile classification method based on feature dependence and multi-objective particle swarm optimization feature selection as described in any one of claims 1-9, characterized in that, It includes a feature scoring calculation module, a particle position initialization module, a fitness value calculation module, an interpolation calculation module, an optimal feature set selection module, and a method verification module; The feature scoring calculation module is used to download the gene expression profile data Colon according to the GSE number, calculate the dependency score of each feature, and quantify the dependency relationship between features through information entropy and mutual information methods. The particle position initialization module is used to initialize the first generation of particle swarm based on the dependency score obtained by the feature scoring module, and to give priority to features with high dependency scores to generate the initial position of the particles. The fitness value calculation module is used to optimize the feature selection problem by using a multi-objective particle swarm optimization algorithm based on the initial particle positions generated by the particle position initialization module. The module combines the classification error rate and the feature selection rate as objective functions and calculates the fitness value using the objective function. The interpolation calculation module is used to select the individual optimal and global optimal for each population by calculating and comparing the fitness values ​​generated by the fitness value calculation module; when updating the individual optimal, a particle optimal position update strategy based on cubic spline interpolation is used to generate new non-dominated solutions, improve the uniformity of solution distribution, and generate the Pareto front. The optimal feature set selection module is used to search for and select the optimal feature subset on the Pareto front generated by the interpolation calculation module through a multi-objective optimization process. The method verification module is used to input the optimal feature subset selected by the optimal feature set selection module into multiple classifiers for training and testing, thereby verifying the effectiveness of the method.

Images