CN107845407A - Based on filtering type and improve the human body physiological characteristics selection algorithm for clustering and being combined - Google Patents

Abstract

The invention discloses a kind of human body physiological characteristics selection algorithm being combined based on filtering type and improvement cluster, including：S1：Impedance model is selected, collects fisrt feature parameter and second feature supplemental characteristic structure initial characteristicses collection and final optimal subset；S2：Filter algorithm is introduced, for each feature in the data that are collected into；S3：Feature set is ranked up from big to small according to HSIC value；S4：The feature of K before ranking is added in feature set, parameter uncorrelated to body composition is filtered off using Filter algorithms, builds initial data set；S5：According to clustering algorithm by dataset construction feature sparse graph；S6：Redundancy feature in cluster is screened using improved clustering algorithm；The human body physiological characteristics selection algorithm that the application establishes can improve human body composition precision of prediction, and more efficiently detection means is provided for body composition Study and clinical practice.

Description

Human Physiological Feature Selection Algorithm Based on the Combination of Filtering and Improved Clustering

technical field

The invention belongs to the field of bioinformatics, in particular to a human physiological feature selection algorithm based on the combination of filtering and improved clustering.

Background technique

The balance of body composition plays an important role in maintaining the stability of the internal environment of the body, and is an important factor affecting human health. When disease occurs, changes in body composition often precede clinical symptoms of the disease. Therefore, changes in body composition can be used to predict the correlation of diseases such as hypertension, dyslipidemia, and metabolic syndrome. However, there are many relevant parameters that affect human body composition, and there are characteristics such as high nonlinearity, redundancy, and irrelevance among the parameters.

The existing Wrapper algorithm removes redundant features. This method can obtain better general performance, but due to the high complexity of the algorithm, it is not suitable for large-scale data sets; the Filter algorithm assigns a weight value to each feature according to the calculation results of the criteria, and calculates The efficiency is high, but this method does not fully consider the redundancy between features, and the selected feature subset is likely to have a large amount of redundancy; the clustering method divides the body component parameter data into multiple groups or clusters, making the cluster The inner objects have a high similarity. Judgment is made according to the distance between each cluster and the center point, and redundant features are effectively screened out, but irrelevant features cannot be effectively screened out. In view of this, before analyzing high-dimensional data of body composition, it is necessary to propose a new method for data dimensionality reduction.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a human physiological feature selection algorithm based on the combination of filtering and improved clustering. First, the Filter feature selection algorithm is used to remove features that are not related to the body composition category, and then M-Chameleon feature clustering is used to The class method removes redundant features, so that the advantages of Filter feature selection algorithm and feature clustering can be maximized. The human body composition prediction model established in this way can improve the prediction accuracy of human body composition, and provide more effective detection means for human body composition research and clinical application.

In order to achieve the above object, the present invention provides a human body physiological feature selection algorithm based on the combination of filtering formula and improved clustering, including:

S1: Select the impedance model, collect the first characteristic parameter and the second characteristic parameter data to construct the initial feature set and the final optimal subset, and initialize the initial feature set and the final optimal subset to an empty set;

Further, the body composition data measured by the human body composition analyzer (INBODY) is used as the data set, which is recorded as T=(O, F, C), where O is the data sample set, F is the selected feature set, and C is the body Composition categories; the parameter sets that have an important impact on human body composition, such as weight, height, age, gender, and the impedance value of each segment of the human body, are used as the first characteristic parameter, and the reciprocal 1/R _i , square R _i ² , R _i R _j is used as the second characteristic parameter. Among them, the impedance value selects the 1KHZ impedance parameter in INBODY, the first characteristic parameter (R1, R2, R3, R4, R5, A, H, W, where A is age, H is height, W is weight) and the second characteristic parameter (1/R1, 1/R2, 1/R3, 1/R4, 1/R5, R1R2, R1R3, R1R4, R1R5, R2R3, R2R4, R2R5, R3R4, R3R5, R5R4, etc.) as the original feature parameter set, denoted as F={f ₁ ,f ₂ ,...,f _m }; the body composition category set C includes body fat mass (BFM) and total water mass (TBW).

S2: Introduce a filtering algorithm (Filter), and calculate the HSIC value under the body composition category C for each feature in the collected data, which represents the correlation between physiological characteristics and body composition categories;

Further, for each feature {f ₁ ,f ₂ ,…,f _m }∈F, define a nonlinear feature map φ: This mapping can map the feature points f ₁ , f ₂ ,...,f _m to the regenerated kernel Hilbert space , the kernel function is: In the formula: space inner product on . Similarly, define a body component category map ψ: Map the C space of the body composition index to the Hilbert space of the regenerating nucleus and write it as , the kernel function is: In addition, the cross-covariance operator that defines the feature and body component categories is: in the formula represents the tensor product, and Express expectations. For each feature {f ₁ ,f ₂ ,…,f _m }∈F, calculate the HSIC value under the body composition category c (HSIC is a kernel-based independence measure method, defined by regenerating kernel Hilbert space Cross-covariance operator, and the independence judgment criterion is obtained through the empirical estimation of the operator norm, which can be used to measure the similarity between two data distributions, and is widely used in feature selection and dimensionality reduction), this value represents Correlation magnitude of physiological characteristics and body composition categories:

For a certain feature f and body composition category c, the larger the value of HSIC, the stronger the dependence of c on f.

S3: Sort the feature set according to the value of HSIC from large to small;

S4: Add the top K features to the feature set, use the Filter algorithm to filter out parameters not related to body composition, and construct an initial data set;

S5: According to the clustering algorithm (M-chameleon), the data set is constructed into a feature sparse map. RI is the edge set connected to each other between features, RC is the similarity between features, and the number k of expected clusters is initialized;

Furthermore, Chameleon uses the agglomerative hierarchical clustering method to construct a feature sparse graph according to the K-nearest neighbor graph method. Each vertex in the graph represents a data object, and there is an edge between these two vertices. Using the edge Weighting can reflect the similarity of objects, and the principle of the algorithm is shown in Figure 1. The similarity of feature subclusters is evaluated based on two points: 1) the interconnection of objects in the cluster; 2) the proximity of the clusters. If the interconnectivity of two feature clusters is high and the distance is very close, the feature clusters that are farther away will be merged and replaced. The similarity between the two features is determined according to the relative interconnection RI and relative proximity RC of the two feature clusters. Given a normalized and filtered feature dataset F={f ₁ ,f ₂ ,…,f _m }, the dataset cluster F is divided into subclusters f ₁ and f ₂ , and F is divided into f ₁ and f 2 f ₂ and the cut edge has the smallest weight, the greater the relative interconnectivity between the feature subclusters f ₁ and f ₂ . The relative interconnection degree RI(f ₁ , f ₂ ) of two feature clusters f ₁ and f ₂ is defined as the relative interconnection degree between feature clusters f ₁ and f ₂ , and the internal interconnection degree of two clusters f ₁ and f ₂ Normalization, that is:

in, is the edge cut of the cluster containing f ₁ and f ₂ , similarly, or is the minimum sum of edge cuts that divide f ₁ (or f ₂ ) into two approximately equal parts.

The relative approximation RC(f ₁ , f ₂ ) of two feature clusters f ₁ and f ₂ is defined as the absolute approximation between f ₁ and f ₂ , with respect to the internal approximation of the two feature clusters f ₁ and f ₂ Normalization, that is:

in, is the average weight _of the edges connecting vertices f1 and vertices _f2 , (or ) is the average weight of the edges of the smallest bipartite cluster f ₁ (or f ₂ ). The similarity between _two subclusters is determined by the relative interconnectivity and relative proximity _of the feature subclusters f1 and f2.

S6: Use the improved clustering algorithm to screen redundant features in the cluster;

S61: Calculate the distance between clusters and sort them, and judge whether the number of sample sub-clusters h is equal to the number of initialization expected clusters k; S62: If not, select the two sub-clusters with the largest similarity function value to merge, if End if equal; S63: Recalculate the relative approximation RC of the new subcluster, traverse all subclusters, whether all subclusters try to merge between each other; S64: If all subclusters try to merge, return to S61; otherwise, similar The two sub-clusters with the smallest degree function are merged and returned to S63; S65: Select the feature with the largest HSIC value to combine.

S7: Select a feature with the largest HSIC value from each feature cluster to form an optimal feature set.

Due to the adoption of the above technical scheme, the present invention can obtain the following technical effects: according to the characteristics of human physiological characteristic parameters, a human characteristic parameter selection algorithm based on the combination of Filter and clustering is proposed, and the characteristic filtering method using the Hilbert-Schmidt dependence criterion The features that are not related to the category are eliminated, and the improved Chameleon clustering is used in feature selection and optimized to improve, which removes redundant features well and effectively selects the optimal feature parameter set for constructing the body composition model , solve the problem of many and redundant human physiological characteristic parameters, and provide more effective detection means for human body composition research and clinical application.

Description of drawings

Figure 1 is a schematic diagram of the Chameleon clustering algorithm;

Figure 2 is a schematic diagram of the improved Chameleon algorithm;

Fig. 3 is the selection process of human body feature parameters;

Figure 4 shows the correlation between the characteristic parameters and BFM obtained by using the filtering algorithm in the 1KHZ frequency band;

Figure 5 shows the correlation between the characteristic parameters and BFM obtained by using the filtering algorithm in the 250KHZ frequency band;

Figure 6 shows the correlation between the characteristic parameters and BFM obtained by using the filtering algorithm in the 500KHZ frequency band

Figure 7 is an analysis of the number of parameter clusters after using the filtering algorithm;

Figure 8 shows the distance between characteristic parameters and BFM indicators when different sample sizes are clustered into four categories;

Figure 9 is the comparison between the predicted value of the BFM model and the actual value;

Figure 10 shows the comparison of the relative error of the predicted value of the BFM model.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

The body composition data measured by INBODY is used as the data set, which is recorded as T=(O, F, C); the parameter sets that have an important impact on the body composition of the human body, such as weight, height, age, gender, and the impedance value of each segment of the human body, are used as The first characteristic parameter, the reciprocal 1/R _i , square R _i ² , and R _i R _j of the impedance of each segment are used as the second characteristic parameter. There are three frequency bands for INBODY measurement: 1KHZ, 250KHZ, and 500KHZ. This paper studies the relationship between body composition and characteristic parameters in the above three frequency bands and different sample sizes. Among them, select the first characteristic parameter (R ₁ , R ₂ , R ₃ , R ₄ , R ₅ , A, H, W) and the second characteristic parameter 1/R ₁ , 1/R ₂ , 1/R ₃ , 1 /R ₄ , 1/R ₅ , R ₁ R ₂ , R ₁ R ₃ , R ₁ R ₄ , R ₁ R ₅ , R ₂ R ₃ , R ₂ R ₄ , R ₂ R ₅ , R ₃ R ₄ , R ₃ R ₅ , R ₄ R ₅ are used as the original feature parameter set, recorded as F={f ₁ ,f ₂ ,…,f _m }; body composition category set C includes body fat mass (BFM), total water mass (TBW) ; Body composition category set C includes body fat mass (BFM) and total water mass (TBW). The following table lists some sample datasets.

However, there are many relevant parameters that affect human body composition, and there are characteristics such as high nonlinearity, redundancy, and irrelevance among the parameters. In view of the above problems, it is necessary to propose a method for dimensionality reduction of data to solve the problem of redundant and irrelevant characteristic parameters above. The clustering method divides the body composition parameter data object into multiple groups or clusters, so that the objects in the cluster have a high similarity, and judge according to the distance between each cluster and the center point, and effectively screen out redundant features. At the same time, before analyzing the high-dimensional data of volume components, the attributes that are not related to the required features are eliminated by reducing the number of features;

Therefore, the filtering algorithm should be applied to the data first. Given the original feature set F={f ₁ ,f ₂ ,…,f _m }, data sample set O={o ₁ ,o ₂ ,…,o _n }, human body composition BFM, TBW, for the first 100 human samples Run the filtering algorithm in the three frequency bands of 1KHZ, 250KHZ, and 500KHZ. The following figure lists the correlation of the filtered characteristic parameters after running the algorithm on the body composition BFM.

In the formula: space inner product on . Similarly, define a body component category map ψ: Map the C space of the body composition index to the Hilbert space of the regenerating nucleus and write it as , the corresponding kernel function is:

The kernel function can calculate the inner product between two feature points in the feature space projection without explicitly calculating the specific mapping There is no computational cost implied by the dimensionality. Therefore, the cross-covariance operator of feature and body component category can be defined as:

in the above formula represents the tensor product, and Indicates the expectation ^[16] , the norm of the square of the covariance can be Called HSIC: its expression is ^[14] :

After using the filter algorithm to run the BFM of different impedances, the correlation can be obtained, as shown in Figure 4, Figure 5, and Figure 6. From the above three figures, it can be seen that when the impedance frequency band gradually increases, the value of the impedance also increases. The amount of BFM information contained in each feature parameter is gradually decreasing. According to the confidence interval of 80% as a filter, select the characteristic parameters, and summarize the characteristics after running the filter algorithm in different frequency bands, as shown in Table 2 below:

Table 2: Features after running the filter algorithm in different frequency bands

It can be seen from Table 2 that the algorithm in this paper greatly reduces the number of original feature sets, and the 250KHZ frequency band has more features. Therefore, the filtered features of the middle impedance frequency band 250KHZ are selected for cluster analysis to screen out redundant information.

Before clustering, it is first necessary to determine how many clusters are clustered, and to obtain the number of information contained in different clustering situations from the filtered feature parameters, as shown in Figure 7, the analysis shows that the feature parameters and body components are divided into The four categories can better represent the selected feature information. When the number of samples is 20, 40, 60, 80, and 100, as shown in Figure 8, the clustering changes little, 1/R ₄ , 1/R ₅ , clustered into one class, A, H, W, R ₅ , R ₄ are grouped into one group, R ₄ R ₅ , R ₁ R ₂ , R ₂ ² , R ₁ ² , R ₅ ² are grouped into one group, R ₂ R ₃ , R ₁ R ₃ are grouped into one group. The characteristic parameters obtained after the Filter algorithm can be clustered into 4 categories to remove 1/R ₄ , R ₄ , R ₁ ² , and R ₁ R ₃ that are far away from the cluster center BFM. Table 3 lists the selection of feature parameters after Filter and clustering algorithm.

Table 3: Feature parameters after Filter and clustering algorithm

Table 4 lists the use of three feature selection methods to obtain candidate feature sets and time complexity for body composition BFM prediction;

Table 4: Optimal feature set and complexity comparison

As can be seen from Table 4, under the same situation of data set dimension, use the number of the candidate feature set that the algorithm of the present invention obtains and its time complexity are all less than Filter and Wrapper, mRMR feature selection algorithm;

In order to verify the performance of this feature selection algorithm, for body composition (BFM), use mRMR, Filter and Wrapper combined feature selection algorithm and this feature selection algorithm for feature selection, in order to accurately measure the above candidate feature sets in a given The pros and cons of body composition under BFM, the first 80 samples in the sample set are used as the training sample set, recorded as T ₁ ={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x ₈₀ ,y ₈₀ )}, the last 20 as the test sample set

T ₂ ＝{(x ₈₁ ,y ₈₁ ),(x ₈₂ ,y ₈₂ ),…,(x ₁₀₀ ,y ₁₀₀ )}, where x _i ∈ R ^l is the input characteristic parameter value, as an independent variable, y _i ∈R is the actual body composition value as the dependent variable _; T1 is trained using multiple linear regression in SPSS software. Table 5 shows the model summary obtained by regression modeling of BFM using the above feature sets:

Table 5: Model summary (change)

a. Predictor variables: (constant), W, S, A, R ₃ , 1/R ₂ , 1/R ₁ , 1/R ₃ , R ₄ ² , R ₄ R ₅ , R ₅ ²

b. Predictor variables: (constant), 1/R ₃ ,W,S,R ₂ ² ,R ₄ ² ,R ₄ R ₅ ,R ₅ ² ,1/R ₁ ,R ₅

c. Predictor variables: (constant), A, H, W, R ₅ , R ₁ R ₂ , R ₂ R ₃ , R ₄ R ₅ , 1/R ₅ , R ₂ ² , R ₅ ² ,

According to Table 5, the correlations between the physiological feature sets and BFM in models 1, 2, and 3 are 0.927, 0.906, and 0.978, respectively. Therefore, the feature sets obtained by using the algorithm in this paper have the strongest correlation with body composition;

According to the obtained regression coefficients of each model, the prediction equation is listed:

BFM ₁ ＝0.041*W+0.126*S+0.523*A-0.212*R ₃ +0.171*1/R ₁ +0.126*1/R ₂ +0.179*1/R ₃ +0.132R ² ₄ +0.13R ₄ R ₅ +0.127R ² ₅ -8.56(1)

BFM ₂ ＝0.313*W-0.044*S-0.125*1/R ₃ +0.108*1/R ₁ +0.016*R ₄ ² -0.01R ₂ ² +0.071R ₅ ² +0.072R ₄ R ₅ -0.526R ₅ +5.674 (2)

BFM ₃ ＝-0.464*A-0.15*H+0.122*W-0.143*R ₅ +0.129*R ₁ R ₂ +0.122*R ₂ R ₃ -0.134*R ₄ R ₅ +0.145*1/R ₅ +0.129 *R ₂ ² -0.141*R ₅ ² (3)

Use the obtained prediction model to predict the test set T ₂ and compare it with the actual value to obtain the comparison between the predicted value and the actual value of the BFM model in Figure 9 and the error analysis Figure 10. It can be seen from Figure 10 that the prediction model constructed using the features obtained by the feature selection algorithm in this paper has a high accuracy, and its prediction relative error is less than 0.12. The results show that the feature set obtained by the human physiological feature selection algorithm based on filter and clustering shows a good correlation with body composition, which can improve the fitting accuracy of the body composition prediction model and reduce the prediction error.

Compared with the prior art, the present invention provides a human physiological feature selection algorithm based on the combination of Filter and clustering. The feature filtering method using the Hilbert-Schmidt dependency criterion eliminates features that are not related to the category, and uses the improved Chameleon clustering for feature selection and optimizes the improvement, which removes redundant features well and effectively selects them. Based on constructing the optimal characteristic parameter set of the body composition model, it solves the problem of many and redundant human physiological characteristic parameters; the human body composition prediction model established in this way can improve the prediction accuracy of human body composition, and provide more information for human body composition research and clinical application. as an effective means of detection.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technical field within the technical scope disclosed in the present invention, according to the technical solution of the present invention Any equivalent replacement or change of the inventive concepts thereof shall fall within the protection scope of the present invention.

Claims (8)

1. Based on the combination of filtering and improved clustering, the human physiological feature selection algorithm is characterized in that, comprising: S1: Select the impedance model, collect the first characteristic parameter and the second characteristic parameter data to construct the initial feature set and the final optimal subset, and initialize the initial feature set and the final optimal subset as empty sets; S2: Introduce a filtering algorithm to calculate the HSIC value under the body composition category for each feature in the collected data, which represents the correlation between physiological characteristics and body composition categories; S3: Sort the feature set according to the value of HSIC from large to small; S4: Add the top K features to the feature set, use the filtering algorithm to filter out parameters that are not related to the body composition, and construct the initial data set; S5: According to the clustering algorithm, the data set is constructed into a feature sparse graph, RI is the edge set connected to each other between features, RC is the similarity between features, and the number k of expected clusters is initialized; S6: Use the improved clustering algorithm to screen redundant features in the cluster; S7: Select a feature with the largest HSIC value from each feature cluster to form an optimal feature set. 2. according to claim 1 based on the human body physiological feature selection algorithm that filtering formula combines with improved clustering, it is characterized in that, adopt the body composition data that human body composition analyzer measures as data set, be denoted as T=(O, F, C), where O is the data sample set, F is the selected feature set, and C is the body composition category; the parameter set that has an important influence on the body composition of the human body is used as the first feature parameter, and the reciprocal of the impedance of each segment is 1/R _i , square R _i ² , R _i R _j as the second characteristic parameter; wherein, the impedance value is selected from the 1KHZ impedance parameter in the human body composition analyzer, and the first characteristic parameter R1, R2, R3, R4, R5, A, H, W And the second feature parameter 1/R1, 1/R2, 1/R3, 1/R4, 1/R5, R1R2, R1R3, R1R4, R1R5, R2R3, R2R4, R2R5, R3R4, R3R5, R5R4 as the original feature parameter set, Recorded as F={f ₁ ,f ₂ ,···,f _m }, f ₁ ,f ₂ ,···,f _m are feature points, where A is age, H is height, W is weight, body composition Category C includes body fat mass and total water mass. 3. According to claim 1 or 2, based on the human body physiological feature selection algorithm combined with filtering and improved clustering, it is characterized in that, for each feature in the collected data, calculate the HSIC value under the body composition category C ,Specifically: For each feature {f ₁ ,f ₂ ,···,f _m }∈F, define a nonlinear feature map This mapping maps feature points f ₁ , f ₂ ,···,f _m to the regenerating kernel Hilbert space , the body composition index C space is mapped to the regenerating kernel Hilbert space as , the kernel function is: For each feature {f ₁ , f ₂ ,···,f _m }∈F, the HSIC value under the body composition category C is calculated. 4. according to claim 3 based on the human body physiological feature selection algorithm that filtering formula combines with improved clustering, it is characterized in that, the mutual covariance operator of definition feature and body composition category is: in the formula represents the tensor product, and Indicates expectation; HSIC value characterizes the correlation between physiological characteristics and body composition category: For a certain feature f and body composition category c, the larger the value of HSIC, the stronger the dependence of c on f. 5. according to claim 1 based on the human body physiological feature selection algorithm that filtering formula combines with improved clustering, it is characterized in that, use Chameleon agglomerative hierarchical clustering method, construct feature sparse graph according to the method of K-nearest neighbor graph, Each vertex in the graph represents a data object, and there is an edge between the two vertices. The weight of the edge can reflect the similarity of the object. The similarity of the feature sub-cluster is evaluated based on two points: 1) the Interconnection; 2) Proximity of clusters: According to the relative interconnection RI and relative proximity RC of two feature clusters, the similarity between their two features is determined. 6. According to claim 4, based on the human body physiological feature selection algorithm combined with filtering formula and improved clustering, it is characterized in that, the feature data set F={f ₁ , f given normalization and filtered using filtering algorithm ₂ ,···,f _m }, the data set cluster F is divided into subclusters f ₁ and f ₂ , and the weight of the edge that is cut off when F is divided into f ₁ and f ₂ is the smallest, and the feature subclusters f ₁ and f ₂ The greater the relative interconnectedness between; the relative interconnectedness RI(f ₁ , f ₂ ) of two feature clusters f ₁ and f ₂ is defined as the relative interconnectedness between feature clusters f ₁ and f ₂ , for two clusters The internal interconnectivity of f ₁ and f ₂ is normalized, namely: <mrow><mi>R</mi><mi>I</mi><mrow><mo>(</mo><msub><mi>f</mi><mn>1</mn></msub><mo>,</mo><msub><mi>f</mi><mn>2</mn></msub><mo>)</mo></mrow><mo>=</mo><mfrac><mrow><mo>|</mo><msub><mi>EC</mi><mrow><mo>{</mo><msub><mi>f</mi><mn>1</mn></msub><mo>,</mo><msub><mi>f</mi><mn>2</mn></msub><mo>}</mo></mrow></msub><mo>|</mo></mrow><mrow><mfrac><mn>1</mn><mn>2</mn></mfrac><mrow><mo>(</mo><mo>|</mo><mi>E</mi><mi>C</mi><msub><mo>|</mo><msub><mi>f</mi><mn>1</mn></msub></msub><mo>+</mo><mo>|</mo><mi>E</mi><mi>C</mi><msub><mo>|</mo><msub><mi>f</mi><mn>2</mn></msub></msub><mo>)</mo></mrow></mrow></mfrac></mrow> in, is the edge cut of the cluster containing f ₁ and f ₂ , similarly, or is the minimum sum _of edge cuts that divide f1 or f2 into _two roughly equal parts. 7. according to claim 5 based on the human body physiological feature selection algorithm that filtering formula combines with improved clustering, it is characterized in that, The relative approximation RC(f ₁ , f ₂ ) of two feature clusters f ₁ and f ₂ is defined as the relative approximation between f ₁ and f ₂ , about the internal approximation of two feature clusters f ₁ and f ₂ Normalization, that is: <mrow><mi>R</mi><mi>C</mi><mrow><mo>(</mo><msub><mi>f</mi><mn>1</mn></msub><mo>,</mo><msub><mi>f</mi><mn>2</mn></msub><mo>)</mo></mrow><mo>=</mo><mfrac><msub><mover><mi>S</mi><mo>&amp;OverBar;</mo></mover><mrow><msub><mi>EC</mi><mrow><mo>{</mo><msub><mi>f</mi><mn>1</mn></msub><mo>,</mo><msub><mi>f</mi><mn>2</mn></msub><mo>}</mo></mrow></msub></mrow></msub><mrow><mfrac><mrow><mo>|</mo><msub><mi>f</mi><mn>1</mn></msub><mo>|</mo></mrow><mrow><mo>|</mrow>mo><msub><mi>f</mi><mn>1</mn></msub><mo>|</mo><mo>+</mo><mo>|</mo><msub><mi>f</mi><mn>2</mn></msub><mo>|</mo></mrow></mfrac><msub><mover><mi>S</mi><mo>&amp;OverBar;</mo></mover><mrow><msub><mi>EC</mi><msub><mi>f</mi><mn>1</mn></msub></msub></mrow></msub><mo>+</mo><mfrac><mrow><mo>|</mo><msub><mi>f</mi><mn>1</mn></msub><mo>|</mo></mrow><mrow><mo>|</mo><msub><mi>f</mi><mn>1</mn></msub><mo>|</mo><mo>+</mo><mo>|</mo><msub><mi>f</mi><mn>2</mn></msub><mo>|</mo></mrow></mfrac><msub><mover><mi>S</mi><mo>&amp;OverBar;</mo></mover><mrow><msub><mi>EC</mi><msub><mi>f</mi><mn>2</mn></msub></msub></mrow></msub></mrow></mfrac></mrow> in, is the average weight _of the edges connecting vertices f1 and vertices _f2 , or is the average weight of the edge of the smallest bipartite cluster f ₁ or f ₂ ; The similarity between _two subclusters is determined by the relative interconnectivity and relative proximity _of the feature subclusters f1 and f2. 8. according to claim 1, based on the human body physiological feature selection algorithm combined with filtering and improved clustering, it is characterized in that the improved clustering algorithm is that all feature sub-clusters are traversed and attempted to merge and replace, by merging After subclustering, evaluate the quality of feature selection, and try to merge existing subclusters; the specific steps are: S61: Calculate the distance between clusters and sort them, and judge whether the number of sample sub-clusters h is equal to the number of initialized expected clusters k; S62: If not equal, select the two sub-clusters with the largest similarity function value for merging, and if they are equal, end; S63: Recalculate the relative approximation RC of the new subcluster, traverse all subclusters, and check whether all subclusters are tried to merge in pairs; S64: If all sub-clusters try to merge, return to S61; otherwise, merge the two sub-clusters with the smallest similarity function and return to S63; S65: Select the feature with the largest HSIC value for combination.