Marine data-based feature selection method and system

Abstract

The invention discloses a feature selection method and system based on ocean data. The method comprises the steps that original ocean data are acquired and integrated into a data set matrix, and a feature selection matrix, a clustering center matrix and an auxiliary matrix are initialized to be in a preset initial state; updating a fuzzy membership matrix based on the data set matrix, the feature selection matrix and the clustering center matrix; updating a weight vector according to the data set matrix, the feature selection matrix, the clustering center matrix and the fuzzy membership matrix; and updating the feature selection matrix through the data set matrix, the auxiliary matrix, the clustering center matrix, the fuzzy membership matrix and the weight vector. According to the method, the fuzzy membership degree is introduced in the unsupervised feature selection process, so that the data can be shared among multiple clusters, and the complex structure and uncertainty of the data are better reflected.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean data processing, and in particular to a feature selection method and system based on ocean data. Background Art

[0002] Data in many application areas, such as image processing, pattern recognition, and machine learning, is often represented as high-dimensional vectors. High-dimensional data increases storage requirements, and the redundant features and noise information present in it not only complicate data processing but also impact the performance of subsequent learning algorithms. Data dimensionality reduction helps address the curse of dimensionality and improve the robustness of related learning algorithms. An increasing number of scholars are devoting themselves to the research of data dimensionality reduction techniques, driving the continued development of this field. In the field of ocean data, with the advancement of modern measurement technology, the dimensionality of ocean data continues to increase. Dimensionality reduction techniques can not only effectively reduce data complexity and storage costs, but also extract key marine environmental variables, remove redundant information and noise, and significantly improve the accuracy and efficiency of research in areas such as marine ecosystem monitoring and climate change analysis. Therefore, dimensionality reduction techniques are widely used and crucial in ocean data analysis.

[0003] Feature selection and feature extraction are two commonly used data dimensionality reduction methods. Feature selection involves selecting the most representative subset of features from the original feature set, while feature extraction involves mapping the original feature set into a new, lower-dimensional space through mathematical transformations. Both methods can effectively reduce computational costs and data storage requirements. The difference lies in that feature selection retains only a subset of the original features, resulting in better interpretability, while feature extraction, due to the mathematical transformation of features, often generates new features lacking clear physical meaning. For marine data, dimensionality reduction requires particularly high interpretability to deepen understanding and research on the mechanisms and mechanisms of marine systems. Marine data often involve complex environmental variables and ecological processes. Therefore, retaining key features with physical meaning through feature selection helps researchers better analyze the relationships between these variables and uncover the driving factors behind marine phenomena. This makes feature selection particularly important in marine data dimensionality reduction. Generally speaking, data can be roughly categorized as fully labeled, partially labeled, or unlabeled based on the degree of annotation. Accordingly, feature selection techniques can be divided into three categories based on the availability of class label information in the data: supervised, semi-supervised, and unsupervised. Supervised feature selection is commonly used for classification and regression tasks, using fully labeled data to assess feature importance and relevance. Semi-supervised feature selection aims to improve the performance of unsupervised techniques by leveraging minimal annotation of partially labeled data. Unsupervised feature selection leverages the inherent structure of unlabeled data to assess feature importance and relevance. Obtaining label information for ocean data is expensive, as accurately labeling each data sample often requires significant manpower, time, and resources. Therefore, unsupervised feature selection methods are particularly important in the field of ocean data and have a wider range of applications.

[0004] Ambiguity is widespread in nature and human society, and the marine environment is no exception. The marine system is complex and changeable, and environmental variables such as temperature, salinity, and ocean currents in different regions and depths often exhibit a high degree of ambiguity. However, most existing unsupervised feature selection algorithms only consider the situation where the sample completely belongs to a certain category, and fail to fully consider the fuzzy relationship between different variables in the ocean data and their environment. In the past few decades, researchers have proposed some fuzzy-based feature selection algorithms in an attempt to better deal with this ambiguity. However, these fuzzy-based feature selection algorithms fail to distinguish between normal ocean observation samples and noise samples. In the case of low-quality ocean data, if feature selection is performed directly on the original dataset, ocean noise data and outlier sample points generated by occasional equipment errors or abnormal environmental conditions may interfere with the feature selection algorithm, resulting in the inability to extract the most effective feature subset, thereby affecting the overall performance of the algorithm.

[0005] In recent years, feature selection algorithms based on sparse learning have garnered widespread attention due to their exceptional robustness and interpretability. Marine data is often high-dimensional, complex, and noisy. Sparse learning-based algorithms can reduce redundant features through sparse representation, improving model robustness and computational efficiency. Therefore, combining fuzzy learning with sparse learning, particularly in the context of marine data, can not only better handle ambiguity and uncertainty but also effectively enhance model performance, thereby providing more accurate analytical results for applications such as marine ecosystem monitoring and climate change prediction. Summary of the Invention

[0006] In view of this, the present invention proposes a feature selection method and system based on ocean data.

[0007] According to one aspect of the present invention, a feature selection method based on ocean data is provided, the method comprising:

[0008] Obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state;

[0009] Updating the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix;

[0010] Updating a weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix;

[0011] Updating the feature selection matrix using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix, and the weight vector;

[0012] Updating the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix and the weight vector;

[0013] Determine whether a termination condition is met. If so, sort the feature vectors in the feature selection matrix and select the target feature vector to construct an optimal feature subset.

[0014] Feature selection is performed on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

[0015] In the above technical solution, the method aims to select the optimal feature subset from the original ocean data to achieve data dimensionality reduction. It mainly optimizes the feature selection results by constructing multiple matrices and performing a series of update operations. It is an iterative feature selection solution that integrates clustering ideas. Specifically:

[0016] First, the raw ocean data is acquired and integrated into a dataset matrix, and the state is initialized. Next, the fuzzy membership matrix is updated based on the dataset matrix, feature selection matrix, and cluster center matrix. The fuzzy membership matrix reflects the degree to which each data point belongs to a different cluster center. This update process effectively measures the correlation between data points and cluster centers, providing a basis for subsequent operations such as determining the importance weights of each feature. This considers the fuzziness and uncertainty of the data and better reflects the complexities that can exist in real-world ocean data. Next, the weight vector is updated by combining the dataset matrix, feature selection matrix, cluster center matrix, and the recently updated fuzzy membership matrix. The weight vector represents the importance of each feature. By dynamically adjusting the weights in this way, we can emphasize features that are important for classification or data differentiation based on the current data distribution and cluster structure, while reducing the weights of less important features. Finally, the feature selection matrix is updated using the dataset matrix, auxiliary matrix, cluster center matrix, fuzzy membership matrix, and weight vector. This is a core step that integrates multiple pieces of information to reevaluate and adjust the features to be selected, gradually aligning the feature selection matrix toward a more optimal feature combination. The auxiliary matrix plays a role in assisting optimization or constraining the selection process, helping to better determine which features to consider. Subsequently, the cluster center matrix is updated based on the dataset matrix, feature selection matrix, fuzzy membership matrix, and weight vector. As feature selection and weights change, the actual clustering structure of the data also changes. Updating the cluster center matrix better reflects the current clustering of the data within the selected feature space, providing more accurate cluster center locations for the next round of iterative updates, thereby further optimizing the entire feature selection and clustering process. Finally, a termination criterion is determined. This can include reaching a preset maximum number of iterations, the change in the feature selection matrix falling below a certain threshold, or the clustering results reaching a certain level of stability. If the termination criterion is met, the eigenvectors in the feature selection matrix are sorted, and the target eigenvectors are selected based on the sorted results to construct the optimal feature subset. This sorting process can be based on factors such as the weights of the features. The resulting optimal feature subset is the key feature set used for dimensionality reduction of the raw ocean data.

[0017] This method iteratively updates multiple factors, including the dataset matrix, feature selection matrix, cluster center matrix, auxiliary matrix, fuzzy membership matrix, and weight vector, by linking them together. This method comprehensively considers the characteristics of ocean data and the interactions between different features, helping to discover more representative and discriminative features. By introducing the cluster center matrix and fuzzy membership matrix and using clustering to assist in feature selection, the selected features not only better represent the data itself but also reflect the data's inherent structure and distribution to a certain extent, which is beneficial for improving the performance of subsequent analysis or modeling based on the reduced dimensionality data.

[0018] In some embodiments, updating the fuzzy membership matrix based on the data set matrix, the feature selection matrix, and the cluster center matrix includes:

[0019] After the sample data in the data set matrix is transformed by the feature selection matrix, the membership value of the sample to each data cluster is calculated with the data cluster center in the cluster center matrix; the fuzzy membership matrix is updated according to the membership value, and the updated fuzzy membership matrix is used as input data for subsequent updating of the weight vector.

[0020] In the above technical solution, the feature selection matrix plays the role of feature screening and weight adjustment for the original data set matrix. It selects and weights each feature in the original ocean data according to certain rules, so that the transformed data can better reflect the essential characteristics of the data and its relationship with the cluster center. For example, assuming that the original ocean data contains multiple features (such as temperature, salinity, depth, etc.), the feature selection matrix will give higher weights to certain features that are closely related to the clustering target (such as dividing different types of ocean areas), while reducing the weight or even discarding the features with less relationship, so that the subsequent cluster analysis can focus more on effective features. After the sample data in the data set matrix is transformed by the feature selection matrix, the result will be compared with the data cluster center in the cluster center matrix to calculate the sample's membership value to each data cluster.

[0021] In some embodiments, updating the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix includes:

[0022] The reconstruction error of the sample is calculated by multiplying the mth power of the membership degree of each sample to each cluster center by the square of the Euclidean distance from the sample to each cluster center.

[0023] For the reference sample with the minimum reconstruction error, a weight vector is calculated and updated, and the updated weight vector is used as an input parameter for subsequent updating of the feature selection matrix and the cluster center matrix.

[0024] In the above technical solution, the reconstruction error of each sample is calculated by multiplying the mth power of its membership degree with the square of the Euclidean distance from the sample to each cluster center, and then summing the results. This calculation method comprehensively considers the sample's membership degree to each cluster center and the distance from the sample to the cluster center. By raising the membership degree to the mth power and multiplying it by the square of the distance, the calculation of the reconstruction error is more consistent with the concept of fuzzy clustering. The mth power of the membership degree can be understood as a weighted weighting of the degree of ambiguity of the sample's belonging to each cluster center, while the square of the distance measures the actual difference between the sample and the cluster center in the feature space. For example, if a sample has a high membership degree to a cluster center and a low distance from the cluster center, then the cluster center's contribution to the sample's reconstruction error will be small. Conversely, if the sample has a low membership degree to a cluster center and a long distance from the cluster center, then the cluster center's contribution to the sample's reconstruction error will be large. Summing the contributions of all cluster centers yields the overall reconstruction error for the sample.

[0025] The weight vector is calculated and updated for the reference sample with the minimum reconstruction error. The reference sample with the minimum reconstruction error is selected to find the sample that best represents the current data distribution and feature selection situation. The weight vector is updated based on this sample so that the weight vector can better reflect the effectiveness of each feature in distinguishing different cluster centers. The specific method of updating the weight vector is to adjust the weight of each feature based on the eigenvalue of the reference sample and its relationship with each cluster center. For example, gradient descent or other optimization methods can be used to update the weight vector so that the reconstruction error is further reduced under this reference sample, thereby optimizing feature selection and weight distribution. The updated weight vector will serve as the input parameter for subsequent updates of the feature selection matrix and cluster center matrix. The new weight vector reflects the importance of each feature in the current feature selection and clustering situation. By using it as input, the feature selection matrix and cluster center matrix can be further optimized, so that the entire feature selection and clustering process can continuously improve accuracy and robustness.

[0026] In some embodiments, updating the feature selection matrix using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix, and the weight vector includes setting:

[0027]

[0028] E=blkdiag(U1,U2,…,U c )

[0029]

[0030] according to Update the feature selection matrix W;

[0031] Where W is the feature selection matrix, Represents the new matrix after the data set matrix is upgraded; Indicates that the membership values of the samples to each data cluster are constructed as a diagonal matrix; E = blkdiag (U1, U2, ..., U c ) indicates that the fuzzy membership matrix corresponding to each sample is constructed as a block diagonal matrix; It is expressed as a diagonal matrix constructed after uniform scaling of the weight vector; β represents the regularization parameter; D is the identity matrix, which is used as an auxiliary matrix; V is the cluster center matrix.

[0032] In the above technical solution, the original data set matrix X is subjected to a dimensionality-increasing operation to form a matrix This dimensionality increase is achieved by copying the dataset matrix itself. This operation can provide a richer information basis for subsequent feature selection and cluster analysis, so that the characteristic performance of each sample in different categories can be fully considered. The membership value of the sample to each data cluster is constructed as a diagonal matrix The diagonal elements of the matrix are the membership values of each sample to different data clusters, and the off-diagonal elements are zero. This structure can easily extract and utilize membership information in matrix operations, so that the relationship between each sample and each data cluster can be expressed separately and clearly. The fuzzy membership matrix corresponding to each sample is constructed as a block diagonal matrix E. Each block of the block diagonal matrix corresponds to the fuzzy membership matrix of a sample. Such a structure can keep the membership information between different samples independent of each other in matrix operations, which facilitates the separate processing and optimization of feature selection for each sample. The weight vector is uniformly scaled and constructed into a diagonal matrix Scaling adjusts the scale of the weight vector so that the weights of each feature can function more appropriately in subsequent matrix operations, preventing excessive or insufficient weights for certain features from unduly impacting the results. The updated feature selection matrix W will be used in subsequent steps to further optimize other related matrices, such as the cluster center matrix and auxiliary matrix, thereby continuously improving and optimizing the entire feature selection and clustering process. Through iterative updates, the feature selection matrix W will gradually converge to a stable state, effectively extracting the most valuable feature subset for the clustering task.

[0033] In some embodiments, updating the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix, and the weight vector includes:

[0034] right Perform singular value decomposition to obtain the left singular matrix and right singular matrices according to Update the cluster center matrix V; where I(m,c) represents the truncated identity matrix.

[0035] In the above technical solution, the matrix Perform singular value decomposition. Singular value decomposition is a commonly used matrix decomposition technique that can decompose the matrix into a left singular matrix L V , singular value matrix (diagonal matrix) and right singular matrix R V In this process, the left singular matrix L V and the right singular matrix R V Contains the orthogonal basis vectors of the matrix in row space and column space respectively. In this step, the matrix can be extracted by singular value decomposition. The important information is used to update the cluster center matrix V. Update formula: Among them I m,c Represents the truncated identity matrix. The purpose of truncating the identity matrix is to maintain the dimensional consistency and structural rationality of the cluster center matrix. By replacing the right singular matrix R V , truncated identity matrix I m,c and the transpose of the left singular matrix Multiply them together to get the new cluster center matrix V. This update method combines the results of singular value decomposition and can effectively use the feature information and membership information of the samples in the data set to adjust the position of the cluster center, so that the cluster center can better reflect the intrinsic structure and feature distribution of the data.

[0036] In some embodiments, the termination condition includes: the number of iterations reaches a set number, or the difference between the target values obtained from two adjacent iterations is less than a set threshold.

[0037] In the above technical solution, stopping after reaching a certain number of iterations is primarily to prevent the algorithm from falling into an infinite loop and ensure that the algorithm completes within a finite time. This setting ensures that the algorithm has a worst-case time limit. Even if the algorithm does not reach the optimal solution during the iteration process, it can be guaranteed to complete within a reasonable time frame, providing protection for subsequent data analysis and processing.

[0038] For example, for large-scale ocean data, it may be difficult to quickly converge to a stable state. By setting an upper limit on the number of iterations, the algorithm can be prevented from running for a long time and being unable to terminate, ensuring the efficiency and controllability of the entire data processing process.

[0039] The difference between the target values obtained from two consecutive iterations is less than a set threshold. This condition determines whether to terminate the algorithm based on its convergence. When the difference between the target values (such as reconstruction error and clustering performance indicators) obtained from two consecutive iterations is less than the set threshold, it can be considered that the algorithm has converged to a relatively stable state. Further iterations may not significantly improve the quality of the results and may even lead to overfitting.

[0040] For example, if the target value is reconstruction error, the reconstruction error will gradually decrease as the feature selection and clustering process progresses. When the difference in reconstruction error between two iterations is less than the set threshold, it means that the algorithm has found a good combination of feature selection and clustering centers, and further iterations may not significantly reduce the error. At this point, the algorithm can be terminated.

[0041] According to another aspect of the present invention, there is provided a feature selection system based on ocean data, based on the above method;

[0042] The acquisition module is used to obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state;

[0043] A first updating module is used to update the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix;

[0044] A second updating module is used to update the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix and the fuzzy membership matrix;

[0045] A third updating module is used to update the feature selection matrix through the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector;

[0046] A fourth updating module, configured to update the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix, and the weight vector;

[0047] A determination module is used to determine whether a termination condition is met. If the termination condition is met, the eigenvectors in the feature selection matrix are sorted and a target eigenvector is selected to construct an optimal feature subset;

[0048] The output module is used to perform feature selection on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

[0049] In the above technical solution, in order to better use the above method, this application proposes a feature selection system based on ocean data, and each module corresponds to each step of the above method. The specific principles have been described above and will not be repeated here.

[0050] According to another aspect of the present invention, there is provided a feature selection device based on ocean data, comprising:

[0051] at least one processor and a memory communicatively coupled to the at least one processor;

[0052] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above method.

[0053] In the above technical solution, in order to better run and process the method, the above method is stored in a memory and a processor is used to execute the stored method. It should be noted that the principle and effect of each step have been described above and will not be further explained here.

[0054] According to another aspect of the present invention, a computer-readable storage medium is provided, storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0055] In the above technical solution, in order to better run and use the method, the above method is stored in a computer-readable storage medium and implemented by a processor. It should be noted that the principle and effect of each step have been described above and will not be further explained here. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 This is a flow chart of an embodiment of a feature selection method based on ocean data according to the present invention;

[0058] Figure 2 This is a schematic structural diagram of an embodiment of a feature selection system based on ocean data according to the present invention;

[0059] Figure 3 This is a schematic diagram of the importance scores of different features in a marine sand dataset according to an embodiment of a feature selection method based on marine data of the present invention;

[0060] Figure 4 This is a graph showing the change in recognition accuracy and projection dimension after feature selection of a marine sand dataset according to an embodiment of a feature selection method based on marine data of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0062] Example 1

[0063] See also Figure 1 A feature selection method based on ocean data. The unsupervised feature selection method disclosed in the present invention has better accuracy and robustness, can reduce the dimensionality of ocean data, and has better results on noisy data sets. The method includes:

[0064] S1. Obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state;

[0065] In this embodiment, raw data is integrated into a sample dataset matrix X. Raw ocean data is acquired and integrated into a dataset matrix containing multiple sample vectors, each representing an observed value of an ocean environmental variable. Raw ocean data is collected using ocean observation equipment to obtain a set of observations of various environmental variables, such as temperature, salinity, and ocean currents. Based on the collected raw ocean data, a standardized processing method is used to unify the data format, resulting in a structured data table. From the data table, the environmental variable values for each observation point are obtained, resulting in a dataset matrix X consisting of multiple sample vectors. If there are missing values in the dataset matrix X, an interpolation algorithm is used to fill in the missing values to obtain a complete sample dataset matrix.

[0066] For example, raw ocean data is collected through ocean observation equipment to obtain a set of observation values for various environmental variables such as temperature, salinity, and ocean currents. For example, ocean environmental data obtained from buoys, satellites, and ship sensors have a temperature range of -2°C to 30°C, a salinity range of 30‰ to 38‰, and an ocean current speed of 0.1m / s to 2.5m / s. Based on the collected raw ocean data, a standardization processing method is used to unify the data format to obtain a structured data table. For example, the temperature, salinity, and ocean current data are normalized to the interval [0,1] to eliminate dimensional differences. Through the data table, the environmental variable values of each observation point are obtained to obtain a data set matrix X consisting of multiple sample vectors. For example, each row of the matrix X represents an observation point, and each column represents an environmental variable. If there are missing values in the data set matrix X, an interpolation algorithm is used to fill in the missing values to obtain a complete sample data set matrix. For example, linear interpolation is used to fill in missing temperature data.

[0067] In this embodiment, the initialization feature selection matrix W is a unit vector matrix, the c mutually independent vertical coordinate basis matrices V with dimensions m in the clustering process are the unit matrix, and the auxiliary matrix D introduced to relax the optimization problem is the unit matrix. Initialize the feature selection matrix, cluster center matrix and auxiliary matrix as the unit matrix. The feature selection matrix is used to screen key features, the cluster center matrix is used to determine the center position of the data cluster, and the auxiliary matrix is used to optimize the calculation process.

[0068] Exemplarily, the feature selection matrix W is initialized to a unit vector matrix. For example, when the dimension is 5, W is a 5×5 unit matrix, where the diagonal elements are 1 and the remaining elements are 0, which is used to determine the initial screening weights of the features of each dimension. The cluster center matrix V is initialized to the unit matrix. For example, when the number of data clusters is 3, V is a 3×3 unit matrix, which is used to determine the initial center position of the data cluster. The auxiliary matrix D is initialized to the unit matrix. For example, in the optimization calculation, D is a 3×3 unit matrix, which is used to obtain the initial state of the optimization calculation.

[0069] S2. updating the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix;

[0070] In this embodiment, S2, updating the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix, includes: calculating the membership value of the sample to each data cluster with the data cluster center in the cluster center matrix after the sample data in the data set matrix is transformed by the feature selection matrix; updating the fuzzy membership matrix according to the membership value, and using the updated fuzzy membership matrix as input data for subsequent updating of the weight vector.

[0071] Specifically,

[0072] Updating the fuzzy membership matrix U based on the data set matrix X, the feature selection matrix W and the cluster center matrix V specifically includes: updating the fuzzy membership matrix U according to the following formula;

[0073]

[0074] Among them, u ij is the membership of the jth sample to the i-th data cluster; W is the feature selection matrix; x j is the jth sample; v i is the i-th cluster center; v i is the qth cluster center; q is the index of each item traversed in the summation process, from 1 to c. c represents the number of cluster centers.

[0075] Exemplarily, based on the data set matrix X, the feature selection matrix W, and the cluster center matrix V, the distance from each sample to the center of each data cluster is calculated to obtain the Euclidean distance matrix between the sample and the cluster center. Based on the Euclidean distance matrix, the membership formula in the fuzzy clustering algorithm is used to update the fuzzy membership matrix U and determine the membership value of each sample to the i-th data cluster. If the sum of the membership values of a sample in the fuzzy membership matrix U is not 1, the membership value of the sample is normalized to obtain a normalized fuzzy membership matrix. Based on the data set matrix X, the feature selection matrix W, and the cluster center matrix V, the Euclidean distance formula is used to calculate the distance from the sample to the center of each cluster.

[0076] S3, updating the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix and the fuzzy membership matrix;

[0077] In this embodiment, S3, updating the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix, includes:

[0078] S31, multiplying the mth power of the degree of membership of each sample to each cluster center by the square of the Euclidean distance from the sample to each cluster center, and then summing the results to calculate the reconstruction error of the sample;

[0079] In this embodiment, let Use g j represents g(x j ), g=[g1,g2,…,g N ], and assume that g1≤g2≤…≤g N ;

[0080] S32. Calculate and update a weight vector for the reference sample with the minimum reconstruction error, and use the updated weight vector as an input parameter for subsequent updating of the feature selection matrix and the cluster center matrix.

[0081] In this embodiment, when only the k reference samples with the minimum reconstruction error are considered, according to Update updates the weight vector p.

[0082] Specifically, a weight vector is updated based on the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix. The weight vector is used to assign different importance weights to samples and screen representative reference samples by minimizing the reconstruction error. Based on the data set matrix X, the feature selection matrix W, the cluster center matrix V, and the fuzzy membership matrix U, the reconstruction error of each sample is calculated to obtain a reconstruction error vector. If there are elements in the reconstruction error vector that are greater than a preset threshold, the error vector is normalized to obtain a normalized error weight. Based on the normalized error weight, the weight vector p is updated to determine the importance weight distribution of the sample.

[0083] S4, updating the feature selection matrix using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector;

[0084] In this embodiment, S4, updating the feature selection matrix using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector, includes setting:

[0085]

[0086]

[0087] E=blkdiag(U1,U2,…,U c )

[0088]

[0089] according to Update the feature selection matrix W;

[0090] Where W is the feature selection matrix, Represents the new matrix after the data set matrix is upgraded; Indicates that the membership values of the samples to each data cluster are constructed as a diagonal matrix; E = blkdiag (U1, U2, ..., U c ) indicates that the fuzzy membership matrix corresponding to each sample is constructed as a block diagonal matrix; It is expressed as a diagonal matrix constructed by uniformly scaling the weight vector; β represents the regularization parameter; D is the unit matrix, which is used as an auxiliary matrix; V is the cluster center matrix. It should be noted that the regularization parameter is a hyperparameter used to control the degree of influence of the regularization term in the model, and is widely used in scenarios to prevent model overfitting and improve generalization ability. It mainly plays the following roles: (1) Control model complexity: By penalizing excessive model weights, the model is simpler and the parameters are smaller or sparser. (2) Prevent overfitting: Moderately increasing the regularization term can effectively reduce the gap between training error and test error. (3) Balance bias and variance: When the regularization parameter is small, the model is complex, prone to overfitting, and has a large variance; when the regularization parameter is large, the model is simple, prone to underfitting, and has a large bias. Common examples include ridge regression (L2 regularization) and Lasso regression (L1 regularization). The regularization parameter is usually selected through methods such as cross-validation to achieve the best balance between model complexity and fitting ability.

[0091] Ridge regression formula: λ controls the L2-norm penalty on the weights.

[0092] Lasso regression formula: λ controls the L1-norm penalty on the weights.

[0093] S5. Update the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix and the weight vector;

[0094] In this embodiment, S5, updating the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix and the weight vector, includes:

[0095] right Perform singular value decomposition to obtain the left singular matrix and right singular matrices according to Update the cluster center matrix V; where I(m,c) represents the truncated identity matrix. Note that truncating the identity matrix means taking a large identity matrix (e.g., n×n) and retaining only the first m rows and c columns to form a new m×c matrix, with the diagonal still having 1s as much as possible and the rest being 0s. For example, the standard 4×4 identity matrix is: If you want to get a truncated identity matrix 3×2, for example, take the first 3 rows and first 2 columns from above, it is:

[0096] Obtain the dataset matrix X, feature selection matrix W, fuzzy membership matrix U, and weight vector p, and initialize the cluster center matrix V. Based on the dataset matrix X and feature selection matrix W, calculate the weighted data distribution to obtain a weighted data matrix. Perform singular value decomposition on the weighted data matrix to obtain left and right singular matrices. Based on the left and right singular matrices, extract the eigenvectors corresponding to the first k singular values to determine the direction of cluster center adjustment. Use the eigenvectors and fuzzy membership matrix U to update the cluster center matrix V, obtaining the initial adjusted cluster center positions.

[0097] Specifically, for example, a dataset matrix X, a feature selection matrix W, a fuzzy membership matrix U, and a weight vector p are obtained to initialize the cluster center matrix V. The dataset matrix X contains 1000 samples, each with 20 features. The feature selection matrix W is a 20×20 diagonal matrix. The fuzzy membership matrix U is a 1000×5 matrix. The weight vector p is a 1000-dimensional vector. Based on the dataset matrix X and the feature selection matrix W, a weighted data distribution is calculated to obtain a weighted data matrix. Specifically, each column of the dataset matrix X is multiplied by the corresponding weight of the feature selection matrix W to generate a weighted data matrix. Singular value decomposition is performed on the weighted data matrix to obtain a left singular matrix and a right singular matrix. The left singular matrix is a 1000×1000 matrix, and the right singular matrix is a 20×20 matrix. Based on the left and right singular matrices, the eigenvectors corresponding to the first k singular values are extracted to determine the cluster center adjustment direction. The value of k is set to 5, and the eigenvectors corresponding to the first five singular values are extracted. The eigenvector and fuzzy membership matrix U are used to update the cluster center matrix V to obtain the preliminary adjusted cluster center position. The specific method is to multiply the eigenvector with each column of the fuzzy membership matrix U to generate a new cluster center matrix V.

[0098] S6. Determine whether a termination condition is met. If the termination condition is met, sort the feature vectors in the feature selection matrix and select the target feature vector to construct an optimal feature subset;

[0099] In this embodiment, if the termination condition is not met, the process returns to S2. If the termination condition is met, the eigenvectors in the updated reconstructed transformation matrix are sorted, and the first d largest eigenvectors are selected to construct the optimal feature subset W'; d is a positive integer greater than or equal to 1.

[0100] In this embodiment, the termination condition includes: the number of iterations reaches a set number, or the difference between the target values obtained from two adjacent iterations is less than a set threshold.

[0101] Exemplarily, the current number of iterations is obtained and compared with a preset iteration threshold, and a determination is made as to whether the preset number of iterations has been reached by comparing the two. If the number of iterations has not reached the preset threshold, the change in the current target value is calculated to obtain target value change data. Based on the comparison of the target value change data with the preset threshold, a determination is made as to whether the target value change is less than the preset threshold. If the target value change is greater than or equal to the preset threshold, the fuzzy membership matrix U calculation module is returned to update the fuzzy membership matrix. A new membership distribution is calculated using the updated fuzzy membership matrix to obtain updated membership data. Based on the updated membership data, the target value is recalculated to determine a new target value. The difference between the new target value and the previous target value is obtained, and the target value change trend is determined by calculating the difference. If the target value change trend meets a preset convergence condition, the eigenvectors of the current fuzzy membership matrix are extracted to obtain an eigenvector set. Using the eigenvector set, an eigenvector selection step is performed to determine a final eigenvector. An updated feature selection matrix W is obtained, and each eigenvector in the matrix is extracted. Based on the extracted eigenvectors, the modulus of each eigenvector is calculated to obtain a modulus value sequence. The modulus value sequence is used to sort the eigenvectors in descending order to determine the list of sorted eigenvectors. If the number of sorted eigenvectors is greater than a preset threshold d, the first d eigenvectors are selected; otherwise, all eigenvectors are selected to obtain the optimal feature subset W′.

[0102] S7. Perform feature selection on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

[0103] In this embodiment, the optimal feature subset W′ is used to perform denoising and data dimensionality reduction on the dataset matrix X, eliminating irrelevant information to achieve the goals of feature selection and denoising. Feature selection is performed on the dataset matrix using the optimal feature subset to generate a reduced-dimensional data matrix that retains key marine environmental variable information while removing redundant and noisy data.

[0104] Exemplarily, the raw ocean data is processed by a data integration module to obtain a dataset matrix X. The feature selection matrix W and the cluster center matrix V are initialized based on the dataset matrix X to obtain initialized matrix parameters. The fuzzy membership update algorithm is used to calculate the membership of the dataset matrix X and the cluster center matrix V to obtain a fuzzy membership matrix U. If the convergence of the fuzzy membership matrix U meets a preset threshold, the sample weights are calculated using a robust weight learning framework to obtain a sample weight vector; if not, the fuzzy membership matrix U is returned and updated again. The feature selection matrix W is optimized based on the sample weight vector and the fuzzy membership matrix U to obtain an updated feature selection matrix W. Feature subsets are extracted from the dataset matrix X using the feature selection matrix W to obtain candidate feature subsets. The candidate feature subsets are processed using a denoising algorithm to obtain a denoised feature subset. The dataset matrix X is subjected to a dimensionality reduction operation based on the denoised feature subset to obtain a reduced-dimensionality data matrix. The reduced-dimensionality data matrix is verified using the key marine environmental variable screening rules to obtain a final data matrix that retains the key marine environmental variables.

[0105] Based on the technical solution of one embodiment, this embodiment has the following advantages:

[0106] This paper introduces fuzzy membership into the unsupervised feature selection process, enabling data sharing across multiple clusters and better reflecting the complex structure and uncertainty of the data. Furthermore, through a robust weight learning framework, the present invention assigns different weights to samples, thereby reducing the impact of noisy samples while preserving the role of reliable samples. Compared with the benchmark algorithm AllFea and seven related unsupervised feature selection algorithms: MV, MCFS, UDFS, NDFS, SOCFS, EGCFS, and SPCAFS, the present invention demonstrates superior effectiveness and robustness.

[0107] Example 2

[0108] See also Figure 2 , a feature selection system based on ocean data, based on the method described in one of the embodiments;

[0109] The acquisition module is used to obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state;

[0110] A first updating module is used to update the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix;

[0111] A second updating module is used to update the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix and the fuzzy membership matrix;

[0112] A third updating module is used to update the feature selection matrix through the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector;

[0113] A fourth updating module, configured to update the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix, and the weight vector;

[0114] A determination module is used to determine whether a termination condition is met. If the termination condition is met, the eigenvectors in the feature selection matrix are sorted and a target eigenvector is selected to construct an optimal feature subset;

[0115] The output module is used to perform feature selection on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

[0116] In the above technical solution, in order to better use the method described in one of the embodiments, the present application proposes a feature selection system based on ocean data, and each module corresponds to each step of the above method. The specific principles have been described above and will not be repeated here.

[0117] Embodiment 3

[0118] The implementation scheme of the present invention is illustrated by taking the analysis of marine sand data as an example. The composition and particle size characteristics of marine sediments vary significantly in different marine environments. In order to better study the marine ecosystem and its dynamic changes, it is necessary to conduct in-depth analysis of the sand data collected at each site. Marine sand data were collected at different sites, and the following eight indicators were recorded: sand proportion (%), silt proportion (%), clay proportion (%), median particle size (Mdφ), average particle size (Mzφ), sorting coefficient (σi), skewness (Ski) and peak state (Kg). The characteristics of the sand samples collected at each site are arranged into a sample vector, and all samples constitute the marine sand data set used in the experiment. Feature selection is performed on the sample vector using the method described in Example 1, and the results are as follows. Figure 3 and Figure 4 As shown. Figure 3 The results show that when analyzing the sand types in different marine sedimentary environments, the sand ratio (%), silt ratio (%) and clay ratio (%) are identified as the most important characteristics, which can effectively characterize the particle size distribution and sedimentation process of marine sediments. Figure 4 It can be seen that using redundant features for modeling will cause classification errors and lead to a decrease in model accuracy.

[0119] To validate the effectiveness and accuracy of our method for ocean data analysis, we compared it with classic unsupervised feature selection algorithms (e.g., MV, MCFS, UDFS, NDFS, SOCFS, EGCFS, and SPCAFS) and the benchmark algorithm AllFea. Table 1 shows the ACC values of the above algorithms and our algorithm under optimal parameter conditions on 12 real-world datasets. The best results are in bold, and the suboptimal results are underlined. The last row shows the average performance of each comparison algorithm across all datasets.

[0120] Table 1 ACC values of the proposed algorithm and the comparison algorithm with optimal parameters on 12 real data sets

[0121]

[0122]

[0123] Results show that compared to other unsupervised feature selection algorithms, the proposed algorithm outperforms other algorithms on most datasets, demonstrating its broad applicability, including effective processing of ocean data. This superior performance is due to the introduction of a fuzzy membership mechanism, which allows data to be shared across multiple clusters, thereby better capturing the complex structure and uncertainty of ocean data. Furthermore, by assigning different weights to samples, the proposed algorithm can reduce the impact of noisy samples while retaining representative and reliable samples. Therefore, the algorithm can accurately select the most discriminative features from all available features and can be effectively used for processing and analyzing ocean data.

[0124] Example 4

[0125] A feature selection device based on ocean data, comprising:

[0126] at least one processor and a memory communicatively coupled to the at least one processor;

[0127] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method as described in one of the embodiments.

[0128] In the above technical solution, in order to better run and process the method described in one of the embodiments, the above method is stored in a memory and executed by a processor. It should be noted that the principle and effect of each step have been described above and will not be further explained here.

[0129] Embodiment 5

[0130] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method described in one of the embodiments.

[0131] In the above technical solution, in order to better run and use the method described in one of the embodiments, the above method is stored in a computer-readable storage medium and implemented using a processor. It should be noted that the principles and effects of each step have been described above and will not be further explained here.

[0132] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A feature selection method based on ocean data, characterized in that: The method comprises: Obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state; Updating a fuzzy membership matrix based on the data set matrix, the feature selection matrix, and the cluster center matrix; Updating a weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix; Updating the feature selection matrix using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix, and the weight vector; Updating the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix and the weight vector; Determine whether a termination condition is met. If so, sort the feature vectors in the feature selection matrix and select the target feature vector to construct an optimal feature subset. Feature selection is performed on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

2. A feature selection method based on ocean data according to claim 1, characterized in that: Updating a fuzzy membership matrix based on the data set matrix, the feature selection matrix, and the cluster center matrix includes: After the sample data in the data set matrix is transformed by the feature selection matrix, the membership value of the sample to each data cluster is calculated with the data cluster center in the cluster center matrix; the fuzzy membership matrix is updated according to the membership value, and the updated fuzzy membership matrix is used as input data for subsequent updating of the weight vector.

3. A feature selection method based on ocean data according to claim 1, characterized in that: Updating a weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix, and the fuzzy membership matrix includes: The reconstruction error of the sample is calculated by multiplying the mth power of the membership degree of each sample to each cluster center by the square of the Euclidean distance from the sample to each cluster center. For the reference sample with the minimum reconstruction error, a weight vector is calculated and updated, and the updated weight vector is used as an input parameter for subsequent updating of the feature selection matrix and the cluster center matrix.

4. A feature selection method based on ocean data according to claim 1, characterized in that: The feature selection matrix is updated by using the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector, Including, order: E=blkdiag(U1,U2,…,U c ) according to Update the feature selection matrix W; Where W is the feature selection matrix, Represents the new matrix after the data set matrix is upgraded; Indicates that the membership values of the samples to each data cluster are constructed as a diagonal matrix; E = blkdiag (U1, U2, ..., U c ) indicates that the fuzzy membership matrix corresponding to each sample is constructed as a block diagonal matrix; It is expressed as a diagonal matrix constructed after uniform scaling of the weight vector; β represents the regularization parameter; D is the unit matrix, which is used as an auxiliary matrix; V is the cluster center matrix.

5. The feature selection method based on ocean data according to claim 1, characterized in that: Updating the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix and the weight vector includes: right Perform singular value decomposition to obtain the left singular matrix and right singular matrices according to Update the cluster center matrix V; where I(m,c) represents the truncated identity matrix.

6. A feature selection method based on ocean data according to claim 1, characterized in that: The termination conditions include: the number of iterations reaches a set number, or the difference between the target values obtained from two adjacent iterations is less than a set threshold.

7. A feature selection system based on ocean data, characterized in that: The method according to any one of claims 1 to 6; The acquisition module is used to obtain the original ocean data and integrate it into a data set matrix, and initialize the feature selection matrix, cluster center matrix and auxiliary matrix to the preset initial state; A first updating module is used to update the fuzzy membership matrix based on the data set matrix, the feature selection matrix and the cluster center matrix; A second updating module is used to update the weight vector according to the data set matrix, the feature selection matrix, the cluster center matrix and the fuzzy membership matrix; A third updating module is used to update the feature selection matrix through the data set matrix, the auxiliary matrix, the cluster center matrix, the fuzzy membership matrix and the weight vector; A fourth updating module, configured to update the cluster center matrix based on the data set matrix, the feature selection matrix, the fuzzy membership matrix, and the weight vector; A determination module is used to determine whether a termination condition is met. If the termination condition is met, the eigenvectors in the feature selection matrix are sorted and a target eigenvector is selected to construct an optimal feature subset; The output module is used to perform feature selection on the original ocean data using the optimal feature subset to obtain data after dimensionality reduction.

8. A feature selection device based on ocean data, characterized in that: include: at least one processor and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images