A neighborhood rough set attribute reduction method for eye disease image diagnosis

Abstract

The application provides a neighborhood rough set attribute reduction method for eye disease image diagnosis, and comprises the following steps: step 1, reading and processing eye disease image data sets; step 2, performing distributed processing on the eye disease image data sets, and building a distributed framework; step 3, respectively performing spectral clustering on a feature set and a label set, and finding out label main clusters and feature clusters; step 4, performing feature selection through neighborhood rough set selection, and selecting a locally optimal feature subset; and step 5, performing feature fusion on the locally optimal feature subset, so that the fused feature subset is a globally optimal feature subset. The method can reduce the detection time of eye disease images, improve the detection efficiency, and assist doctors in diagnosing eye diseases.

Description

Technical Field

[0001] This invention belongs to the field of auxiliary medical processing, and particularly relates to a neighborhood rough set attribute reduction method for image diagnosis of eye diseases. Background Technology

[0002] Eye diseases are a global public health issue, and their causes are influenced by a variety of factors, including population aging and lifestyle changes. With the increasing aging of societies, the prevalence of retinal degenerative diseases is rising. There are many types of eye diseases, including cataracts, glaucoma, dry eye syndrome, fundus diseases, and age-related macular degeneration, which significantly impact patients' quality of life. Therefore, regular eye exams and appropriate preventative measures are crucial for maintaining eye health.

[0003] Currently, there is no efficient diagnostic technology available for diagnosing eye diseases; diagnosis relies solely on doctors' experience. With changing lifestyles, eye diseases have become a global health concern, creating an urgent need for an effective method to extract relevant regions from eye disease data for diagnosis. Feature selection can help doctors analyze eye diseases and improve diagnostic efficiency.

[0004] Rough set theory is primarily used to handle uncertainty and fuzziness in data. Its core idea is to divide a dataset into different equivalence classes using equivalence relations, thereby analyzing the information. Rough sets can handle incomplete or fuzzy information, describing data boundaries through upper and lower approximations. Unlike probability theory and fuzzy set theory, rough sets do not require prior knowledge, relying solely on the data itself. Rough set theory uses attribute reduction techniques to extract key features from data, reducing redundant information. Classical rough sets also have limitations; they can only handle discrete data. Hu Qinghua et al. extended classical rough sets to neighborhood rough sets. Sun Lin et al., based on classical rough sets, proposed introducing neighborhood rough sets into multi-label data in their paper "A Multi-Label Specific Feature Selection Method Based on Neighborhood Rough Sets" (Computer Science 2018, 173-178). However, they neglected the complementary relationship between features; some features, when viewed individually, may have limited contribution to classification, but when combined, they can provide more valuable information. Similarly, in multi-label data, there are correlations between labels. For example, in medicine, patients with hypertension are more likely to suffer from cardiovascular diseases than others. Therefore, it is necessary to consider the correlations between features and the correlations between labels in multi-label learning. Thus, how to consider the correlations between features and the correlations between labels in multi-label learning is the focus of this research. Summary of the Invention

[0005] Objective of the Invention: The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a neighborhood rough set attribute reduction method for image diagnosis of eye diseases, comprising the following steps:

[0006] Step 1: Read and process the image dataset of eye diseases;

[0007] Step 2: Perform distributed processing on the eye disease image dataset and build a distributed framework;

[0008] Step 3: Perform spectral clustering on the feature set and label set respectively to find the main label cluster and feature cluster;

[0009] Step 4: Select the best local feature subset by using neighborhood rough set selection;

[0010] Step 5: Perform feature fusion on the local best feature subset so that the fused feature subset becomes the global best feature subset.

[0011] Step 1 includes the following steps:

[0012] Step 1-1: Read the eye disease image dataset, determine the feature set and decision classification, and establish a multi-label decision system MNDT.<U,C∪D,Δ,δ> Where U = {x1, x2, ..., x} N} represents the set of objects to be detected in the eye disease image dataset, where N is the number of objects to be detected, and x N Let C = {a1, a2, ..., a...} represent the Nth object to be checked. n} represents the feature set of data attributes, where n is the number of sample features, and a n Let D = {l1, l2, ..., ln} be the nth feature value of the sample. d} represents the decision class of the sample, where d is the number of sample labels, l d Let d represent the label of the d-th sample, Δ represent the distance function, and δ represent the neighborhood radius.

[0013] Steps 1-2 involve normalizing the features in the eye disease data. The normalization formula is as follows:

[0014]

[0015] Where x refers to the value of a feature in the original sample, x normalized This represents the value of the feature of the original sample after normalization, x max This refers to the maximum value of all samples in a single attribute, x. min This represents the minimum value of all samples in a single attribute.

[0016] Step 2 includes: setting up a Spark distributed framework to divide the eye disease image dataset S' into m subsets of eye disease information S' = {S'1, S'2, S'3, ..., S'...} m}, where S' m Let m represent the m-th subset of eye disease information. Each of the m subsets of eye disease information is distributed to the m working nodes in parallel using a random sending strategy.

[0017] Step 3 includes the following steps:

[0018] Step 3-1: At each node, perform spectral clustering on the feature set and label set in the subset of eye disease information in the node to find the main label cluster and sub-label cluster in each node. The main label cluster is the cluster with the most labels in the cluster, and the sub-label cluster is the cluster other than the main label cluster.

[0019] The feature clusters formed by spectral clustering show high correlation among features within the clusters and weak similarity among features between the clusters.

[0020] Step 3-2, for label spectrum clustering (feature spectrum clustering is similar): construct a similarity matrix W∈R d×d Where R represents the real space, and the element W in the i-th row and j-th column of the similarity matrix W is... ij Represented as:

[0021]

[0022] Where the intermediate parameter σ = 1.0, exp represents the natural exponential function, and i and j take values ​​from 1 to d; the distance formula is defined as:

[0023]

[0024] Among them l iN Indicates sample X N In label l i The value of l jN Represents sample X N In label l j The value on;

[0025] Step 3-3: Construct the diagonal matrix Z;

[0026] Steps 3-4: Let matrix A = ZW, and normalize the Laplace matrix to obtain the normalized Laplace matrix A'.

[0027] Steps 3-5: Select the eigenvectors v1, v2, ..., v corresponding to the first T smallest non-zero eigenvalues. T T is the preset number of clusters, v T The eigenvectors v1, v2, ..., v3 represent the T-th smallest non-zero eigenvalues.T The matrix that forms d×T;

[0028] Steps 3-6: For the selected feature vectors v1, v2, ..., v T The resulting matrix is ​​subjected to K-means clustering to obtain label clusters. The clusters are sorted by the number of labels within each cluster, with the cluster containing the most labels being the primary cluster and the remaining clusters being the secondary clusters.

[0029] In step 3-3, the diagonal matrix Z is constructed using the following formula:

[0030]

[0031] Z(i,i) represents the sum of all elements in the i-th row of the similarity matrix W, and is called the degree matrix.

[0032] In steps 3-4, the normalized Laplace matrix A' is expressed as:

[0033] A'=IZ -1 W (4),

[0034] Where I is the identity matrix.

[0035] Step 4 includes the following steps:

[0036] Step 4-1, for any sample x i ∈U and B∈C, sample x i The neighborhood δ under the conditional attribute subset B B (x i )for:

[0037] δ B (x i )={x j |x j ∈U,dis B (x i ,x j )≤δ} (5),

[0038] Where δ is the neighborhood size, if the i-th sample x i With the j-th sample x j If x satisfies relation (5), then x is called j In x i In the neighborhood of , dis B (x i ,x j ) represents the i-th sample x i With the j-th sample x j The Euclidean distance under the conditional attribute subset B is defined as:

[0039]

[0040] Where, f(x) i ,a t ) represents the i-th sample x i The value corresponding to the t-th feature;

[0041] Step 4-2, D j Indicates having label l j The sample set, Indicates in label cluster L k The following set of tags to consider. Indicates in label cluster L k The next sample x i The set of labels possessed, in the subset of conditional features Below, the rough set of a multi-label neighborhood is approximated as:

[0042]

[0043] Among them, the lower approximation The description is in the tag cluster L k The next sample x i The neighborhood of the label is entirely within the label cluster L k The sample set;

[0044] upper approximation The description is in the tag cluster L k The next sample x i The neighborhood may belong to the tag cluster L k The sample set, the boundary region This indicates that it is impossible to determine whether it belongs to tag cluster L. k The sample set;

[0045] Step 4-3, under conditional feature subset B, conditional feature subset B for label cluster L k Dependence for:

[0046]

[0047] Step 4-4, for the multi-label decision system S =<U,C∪D,Δ,δ> , The importance of attribute a, SIG(a, B, D), is defined as follows:

[0048] SIG(a,B,D)=λ B∪{a} (D)-λ B (D)(11),

[0049] Where λ B∪{a}(D) indicates the change in dependency caused by adding feature a to the conditional feature subset B. If SIG(a,B,D)>0, it means that feature a is a useful feature for feature subset B. If SIG(a,B,D)<0, it means that feature a is a redundant feature for feature subset B.

[0050] Steps 4-5: For the feature clusters derived from spectral clustering, calculate the dependency of each feature cluster on the label principal cluster L1, and find the feature set F with the highest dependency. i Remove feature set F according to formula (11) i Redundant features in the data, and the feature set F after removing redundant features. i The selected feature clusters are combined to form a subset of candidate features.

[0051] Steps 4-6: For the label sub-cluster, the features in the candidate feature subset are added to the selected feature set in turn. The dependency of the change caused by the update condition feature subset is calculated according to formula (11), and the corresponding feature with the largest dependency is added to the selected feature set until the dependency no longer increases. The selected feature is the feature subset corresponding to the label sub-cluster.

[0052] Steps 4-7: For the two feature subsequences selected from the main label cluster and the sub-label cluster respectively, the features are selected iteratively to generate the final local optimal feature subset (the feature subset corresponding to the child node).

[0053] Step 5 includes: In the Spark distributed framework, the best feature subsets of each worker node are aggregated through the reduce operation of the master node to form the global best feature subset.

[0054] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0055] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention uses the distributed Spark framework, which distributes the dataset to multiple nodes for parallel processing. Distributed processing can significantly improve the data processing speed, achieve faster query response time and data analysis. Through parallel processing, the distributed system can execute multiple tasks at the same time, make full use of the cluster's computing resources, improve the overall processing capacity, and optimize data locality by allocating computing tasks to the nodes where the data is located, reducing data transmission and improving processing efficiency.

[0057] (2) This invention uses spectral clustering technology. By performing spectral clustering on the label set, the main label cluster is identified, which plays a decisive role relative to the entire label set. Similarly, there is a strong correlation between features. When performing spectral clustering on the feature set, features in the same cluster are highly correlated, while features in different clusters are weakly correlated. Therefore, performing spectral clustering on the feature set is beneficial for subsequent feature selection.

[0058] (3) Introducing neighborhood rough set technology, by combining neighborhood rough set with multi-label feature selection, we can handle the uncertainty in multi-label data, consider the correlation between labels, avoid information loss caused by processing each label independently, preserve the local structure of the data through neighborhood relationship, and help select more discriminative features. Attached Figure Description

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0060] Figure 1 This is a flowchart of the method of the present invention.

[0061] Figure 2 This is a detailed architectural diagram of the method of the present invention.

[0062] Figure 3 This is a flowchart of the data processing method of the present invention. Detailed Implementation

[0063] See Figure 1 , Figure 2 and Figure 3 This invention provides a neighborhood rough set attribute reduction method for image diagnosis of eye diseases, comprising the following steps:

[0064] S10. Read and process the eye disease image dataset;

[0065] S20. Perform distributed processing on the image dataset of eye diseases and build a distributed framework;

[0066] S30. Perform spectral clustering on the feature set and label set respectively to find the main label cluster and the highly correlated feature clusters within the cluster;

[0067] S40. Select the local best feature subset by performing feature selection through neighborhood rough set;

[0068] S50. Perform feature fusion on local feature subsets so that the fused feature subset is the best feature subset globally.

[0069] Step S10 includes the following steps:

[0070] S11. Read the image dataset of eye diseases, determine its feature set and decision classification, and establish a multi-label decision system MNDT.<U,C∪D,Δ,δ> Where U = {x1, x2, ... x} N} represents the set of objects to be detected in the eye disease image dataset, where N is the number of objects to be detected, and C = {a1, a2, ... a n} represents the feature set of data attributes, where n is the number of sample features, and a n Let D = {l1, l2, ... ln} be the nth feature value of the sample. d} represents the decision class of the sample, where d is the number of sample labels, Δ represents the distance function, and δ represents the neighborhood radius;

[0071] S12. Normalize the features in the eye disease data. The normalization formula is as follows:

[0072]

[0073] Where x refers to the value of a certain feature in the original sample, x normalized This represents the value of a certain feature in the original sample after normalization, x max It refers to the maximum value of all samples in a certain attribute, x. min This represents the minimum value of all samples in a certain attribute;

[0074] Step S20 includes the following steps:

[0075] S21. Build a distributed framework to divide the eye disease image dataset S' into m subsets of eye disease information S' = {S'1, S'2, S'3, ..., S'...} m}, where S' m This represents the m-th subset of eye disease information. These m subsets are then distributed in parallel to the m working nodes using a random distribution strategy, as follows: Figure 2 The left side of the framework diagram shows the allocation nodes. Assume that a data subset S1 obtained at a certain node is as shown in Table 1.

[0076] Table 1

[0077] U <![CDATA[a1]]> <![CDATA[a2]]> <![CDATA[a3]]> <![CDATA[a4]]> <![CDATA[a5]]> <![CDATA[l1]]> <![CDATA[l2]]> <![CDATA[l3]]> <![CDATA[l4]]> <![CDATA[l5]]> <![CDATA[x1]]> 0.5710 0.6129 0.2120 0.3210 0.4200 1 1 0 1 0 <![CDATA[x2]]> 0.4811 0.4240 0.3330 0.0180 0.7120 1 1 0 0 0 <![CDATA[x3]]> 0.3590 0.4240 0.2700 0.5910 0.6380 0 0 1 0 1 <![CDATA[x4]]> 0.2989 0.5730 0.6030 0.9800 0.4210 0 0 1 1 1 <![CDATA[x5]]> 0.1883 0.7060 0.4720 0.5150 0.5720 1 1 0 1 0 <![CDATA[x6]]> 0.3521 0.8500 0.4360 0.5980 0.9080 1 1 1 0 1 <![CDATA[x7]]> 0.2684 0.2040 0.3140 0.9370 0.5630 1 0 1 0 1

[0078] Step S30 includes the following steps:

[0079] S31. At each node, perform spectral clustering on the feature set and label set within the subset of eye disease information in that node. Identify the primary label cluster (the cluster with the most labels within it) and secondary label clusters (clusters other than the primary label cluster) in each node. The feature clusters formed by spectral clustering exhibit high correlation within clusters and weak similarity between clusters. Figure 2 The diagram of mid-spectral clustering is shown.

[0080] S32. Taking label spectrum clustering as an example, feature spectrum clustering is similar. Construct a similarity matrix W∈R. d×d Where R represents the real space, and the element W in the i-th row and j-th column of the similarity matrix W is... ij Represented as:

[0081]

[0082] Where the intermediate parameter σ = 1.0, exp represents the natural exponential function, and the distance formula is defined as:

[0083]

[0084] Among them l iN Indicates sample X N In label l i The value of l jN Represents sample X N In label l j The value on.

[0085] The label similarity matrix is ​​as follows:

[0086]

[0087] S33. Construct the diagonal matrix Z. Z(i,i) represents the sum of all elements in the i-th row of the similarity matrix W, and is called the degree matrix:

[0088]

[0089] S34. Let matrix A = ZW, and normalize the Laplace matrix to obtain the normalized Laplace matrix A':

[0090] A'=IZ -1 W(4),

[0091]

[0092]

[0093] Where I is the identity matrix;

[0094] S35. Select the eigenvectors e1 and e2 corresponding to the first two smallest non-zero eigenvalues.

[0095] After calculation, the eigenvalues ​​of A' are as follows:

[0096] e1=0.283833107, e2=0.144247828, e3=-1.47790718×10 -16 e4 = 0.028290718

[0097] e5 = 1.00000000.

[0098] The eigenvectors corresponding to the eigenvalues ​​are:

[0099]

[0100] Calculations show that vectors v3 and v4 are the eigenvectors corresponding to the first two smallest non-zero eigenvalues:

[0101]

[0102] S36. Perform K-means clustering on the selected feature vectors.

[0103] K-means clustering was performed on the feature vector v, k=2, and the result was [0 0 1 0 1].

[0104] Based on the clustering results, the label clusters are clustered into two clusters, L1 and L2, where L1 = {l1, l2, l4} and L2 = {l3, l5}. Therefore, label cluster L1 is the primary label cluster of node S1, and L2 is the secondary label cluster of node S1. The primary label cluster has a more significant impact on the entire dataset than the label clusters themselves. Therefore, when considering feature selection, the primary label cluster should be prioritized. Similarly, spectral clustering is performed on the features, resulting in [1 0 1 2 0]. Based on the feature clustering results, the labels are clustered into three clusters: F1 = {a1, a3}, F2 = {a2, a5}, and F3 = {a4}. The features within clusters are highly correlated, while the features between clusters are weakly correlated. Obviously, the feature clusters obtained from spectral clustering are beneficial for subsequent feature selection.

[0105] Step S40 includes the following steps:

[0106] S41. For any sample x i ∈U and B∈C, sample x i The neighborhood δ under the conditional attribute subset B B (x i )for:

[0107] δ B (x i )={x j |x j ∈U,dis B (x i ,x j )≤δ} (5),

[0108] Where δ is the neighborhood size, if sample x i With sample x j If x satisfies relation (5), then x is called j In x i In the neighborhood of , dis B (x i ,x j ) is the sample x i With sample x j The Euclidean distance under the conditional attribute subset B is defined as:

[0109]

[0110] Where, f(x) i ,a t ) represents sample x i The value corresponding to the t-th feature;

[0111] S42, D j Indicates having label l j The sample set, Indicates in label cluster L k The following set of tags to consider. Indicates in label cluster L k The next sample x i The set of labels possessed, in the subset of conditional features Below, the rough set of a multi-label neighborhood is approximated as:

[0112]

[0113] Among them, the lower approximation The description is in the tag cluster L k The next sample x i The neighborhood of the label is entirely within the label cluster L k The sample set;

[0114] upper approximation The description is in the tag cluster L k The next sample x i The neighborhood may belong to the tag cluster L k The sample set, the boundary region This indicates that it is impossible to determine whether it belongs to tag cluster L. k The sample set;

[0115] Given the label cluster L1 = {l1, l2, l4}, find the feature cluster that is most closely associated with the label cluster. Taking feature cluster F1 as an example...

[0116] Under the primary label cluster L1, L1 = {l1, l2, l4}, therefore Considering only the main label cluster, taking sample x1 as an example, we have: D 2 ={x1,x2,x5,x6}, D 4 ={x1,x4,x5}, according to formula (20) we can get If the constraints are satisfied, we get Similarly, y2 = {l1, l2}, so Similarly

[0117] S43. Under conditional feature subset B, the dependency of conditional feature subset B on label cluster L1 is:

[0118]

[0119] For the example above, according to formula (23), under feature cluster F1, the dependency of the label principal cluster L1={l1,l2,l4} is:

[0120] Under the tag primary cluster L1={l1,l2,l4} Taking the feature cluster F2={a2,a5} as an example, the following neighborhood is obtained according to formula (19):

[0121] Under the label principal cluster L1, L1 = {l1, l2, l4}. Considering only the label principal cluster, taking sample x1 as an example, we know that y1 = {l1, l2, l4}, therefore we have... D 1 ={x1,x2,x5,x6,x7}, D 2 ={x1,x2,x5,x6}, D 4 ={x1,x4,x5}, according to formula (20) we can get so Similarly, y2 = {l1, l2}, Therefore, we can obtain Similarly

[0122] According to formula (23), under feature cluster F2 = {a2, a5}, the dependency of feature cluster F2 = {a2, a5} on the label principal cluster L1 = {l1, l2, l4} is:

[0123] Similarly, under label cluster F3 = {a4}, the dependency on the primary cluster label is...

[0124] Under the main label cluster, the final feature cluster rankings are: F1 = {a1, a3}, F2 = {a2, a5}, F3 = {a4}.

[0125] S44. For decision set S =<U,C∪D> , If a∈B, the importance of attribute a is defined as:

[0126] SIG(a,B,D)=λ B∪{a} (D)-λ B (D) (11),

[0127] λ B∪{a} (D) indicates the change in dependency caused by adding feature a to the conditional feature subset B. If SIG(a,B,D)>0, it means that feature a is a useful feature for feature subset B. If SIG(a,B,D)<0, it means that feature a is a redundant feature for feature subset B.

[0128] S45. For the conditional attribute subsets obtained from spectral clustering, calculate the dependency of each conditional attribute subset on the label principal cluster L1, and find the feature set F with the highest dependency. i Remove feature set F according to formula (11) i Redundant features in the data, and the feature set F after removing redundant features. i The selected feature clusters are combined to form a subset of candidate features.

[0129] Taking F1 = {a1, a3} as an example, according to formula (24), After removing a1, the calculation showed that... The decrease indicates that a1 is a useful feature. This indicates that a2 is also a useful feature, so the feature cluster F1 = {a1, a3} is indispensable to the main label cluster L1 = {l1, l2, l4}. Therefore, cluster F1 = {a1, a3} is a locally optimal subset of the main label cluster.

[0130] S46. For the label sub-cluster, the features in the candidate feature subset are added to the selected feature set in turn. The dependency of the change caused by the update condition feature subset is calculated according to formula (11). The corresponding feature with the largest dependency is added to the selected feature set until the dependency no longer increases. The selected feature is the feature subset corresponding to the label sub-cluster.

[0131] The label sub-cluster is L2={l3,l5}, and the candidate feature subset is F4={a2,a4,a5}. According to formula (24), the label sub-cluster feature subset is {a2,a5}.

[0132] S47. For the two feature subsequences selected from the main label cluster and the sub-label cluster respectively, the features are selected iteratively to generate the final local optimal feature subset (the feature subset corresponding to the child node).

[0133] Therefore, based on the above example, the final feature subsequence is: {a1,a2,a3,a5}. The average precision of the multi-label evaluation metric is defined as follows:

[0134]

[0135] Where Y i Indicates sample x i The set of true labels (i.e., the correct labels in a multi-label system), |Y i | represents sample x i The number of true labels, l represents a true label currently being calculated, and l' is a temporary variable used to iterate through all labels that meet the above conditions. rank(x) i ,l) For sample x i The predicted ranking of the label l (sorted from high to low by predicted probability or score, starting from 1).

[0136] The average accuracy of this example is calculated to be 81.23%, which is higher than some mainstream models, indicating that the present invention has practical significance.

[0137] Step S50 includes: fusing the locally optimal feature subsets selected from each node within the Spark distributed framework, so that the fused feature subset becomes the globally optimal subset, such as... Figure 2 The feature subsets of each node in the framework diagram are shown in the fusion diagram.

[0138] This invention provides a neighborhood rough set attribute reduction method for image diagnosis of eye diseases. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A neighborhood rough set attribute reduction method for image diagnosis of eye diseases, characterized in that, Includes the following steps: Step 1: Read and process the image dataset of eye diseases; Step 1 includes: Step 1-1: Read the image dataset of eye diseases, determine the feature set and decision classification, and establish a multi-label decision system. ,in N represents the set of objects to be detected in the eye disease image dataset, where N is the number of objects to be detected. This represents the Nth inspection object. The feature set represents the data attributes, where n is the number of sample features. For the nth feature value of the sample, Let represent the decision class of the sample, where d is the number of sample labels. This represents the label of the d-th sample. Represents the distance function. Represents the size of the neighborhood radius; Step 2: Perform distributed processing on the eye disease image dataset and build a distributed framework; Step 3: Perform spectral clustering on the feature set and label set respectively to find the main label cluster and feature cluster; Step 4: Select the best local feature subset by using neighborhood rough set selection; Step 5: Perform feature fusion on the local best feature subset so that the fused feature subset becomes the global best feature subset.

2. The method as described in claim 1, characterized in that, Step 1 also includes the following steps: Steps 1-2 involve normalizing the features in the eye disease data. The normalization formula is as follows: （1）， Where x refers to the value of a feature in the original sample. This represents the feature values ​​of the original sample after normalization. This refers to the maximum value of all samples in a single attribute. This represents the minimum value of all samples in a single attribute.

3. The method as described in claim 2, characterized in that, Step 2 includes: setting up a Spark distributed framework to process the eye disease image dataset. Divided into m subsets of eye disease information ,in Let m represent the m-th subset of eye disease information. Each of the m subsets of eye disease information is distributed to the m working nodes in parallel using a random sending strategy.

4. The method as described in claim 3, characterized in that, Step 3 includes the following steps: Step 3-1: At each node, perform spectral clustering on the feature set and label set in the subset of eye disease information in the node to find the main label cluster and sub-label cluster in each node. The main label cluster is the cluster with the most labels in the cluster, and the sub-label cluster is the cluster other than the main label cluster. Step 3-2, Clustering of Tag Spectrum: Constructing a Similarity Matrix Where R represents the real space, and the element in the i-th row and j-th column of the similarity matrix W is... Represented as: （2）， intermediate parameters exp represents the natural exponential function, where i and j take values ​​from 1 to d; the distance formula is defined as: （3）， in Indicates sample In the label The value on, Represents the sample In the label The value on; Step 3-3: Construct the diagonal matrix Z; Steps 3-4: Let matrix A = ZW, and normalize the Laplace matrix to obtain the normalized Laplace matrix. ; Steps 3-5: Select the eigenvectors corresponding to the first T smallest non-zero eigenvalues. T is the preset number of clusters. The eigenvector represents the T-th smallest non-zero eigenvalue; eigenvector composition Matrix; Steps 3-6: For the selected feature vectors The resulting matrix is ​​subjected to K-means clustering to obtain label clusters. The clusters are sorted by the number of labels within each cluster, with the cluster containing the most labels being the primary cluster and the remaining clusters being the secondary clusters.

5. The method as described in claim 4, characterized in that, In step 3-3, the diagonal matrix Z is constructed using the following formula: ； in The sum of all elements in the i-th row of a similarity matrix W is called the degree matrix.

6. The method as described in claim 5, characterized in that, In steps 3-4, the normalized Laplace matrix Represented as: （4）， Where I is the identity matrix.

7. The method as described in claim 6, characterized in that, Step 4 includes the following steps: Step 4-1, for any sample and ,sample Neighborhood under conditional attribute subset B for: （5）， in It is the neighborhood size, if the i-th sample With the j-th sample If relation (5) is satisfied, then it is called exist In the neighborhood of For the i-th sample With the j-th sample The Euclidean distance under the conditional attribute subset B is defined as: （6）， in, Represents the i-th sample The value corresponding to the t-th feature; Step 4-2, Indicates having a label The sample set, Indicates in the tag cluster The following set of tags to consider. Indicates in the tag cluster Lower sample The set of labels possessed, in the conditional feature subset Below, the rough set of a multi-label neighborhood is approximated as: （7）， （8）， （9）， Among them, the lower approximation The description is in the tag cluster Lower sample The neighborhood of the label is entirely a label cluster. The sample set; upper approximation The description is in the tag cluster Lower sample The neighborhood may belong to a tag cluster. The sample set, the boundary region This indicates that it is impossible to determine whether it belongs to a tag cluster. The sample set; Step 4-3, under conditional feature subset B, conditional feature subset B for label clusters Dependence for: （10）， Step 4-4, for multi-label decision systems , , The importance of attribute a Defined as: （11）， in This represents the change in dependency caused by adding feature a to the conditional feature subset B. >0 indicates that feature a is a useful feature for feature subset B. <0 indicates that feature a is a redundant feature for feature subset B; Steps 4-5: For the feature clusters derived from spectral clustering, calculate the relationship between each feature cluster and the label principal cluster. Determine the dependency ratio and find the feature set with the highest dependency ratio. Remove the feature set according to formula (11) Redundant features in the data, and the feature set after removing redundant features. The selected feature clusters are combined to form a subset of candidate features. Steps 4-6: For the label sub-cluster, the features in the candidate feature subset are added to the selected feature set in turn. The dependency of the change caused by the update condition feature subset is calculated according to formula (11), and the corresponding feature with the largest dependency is added to the selected feature set until the dependency no longer increases. The selected feature is the feature subset corresponding to the label sub-cluster. Steps 4-7 involve iteratively selecting features from the two feature subsequences selected from the main label cluster and the sub-label cluster to generate the final locally optimal feature subset.

8. The method as described in claim 7, characterized in that, Step 5 includes: In the Spark distributed framework, the best feature subsets of each worker node are aggregated through the reduce operation of the master node to form the global best feature subset.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images