Intelligent clustering and feature extraction method for multi-dimensional data

Abstract

The invention relates to an intelligent clustering and feature extraction method for multi-dimensional data. The method comprises the following steps: acquiring a multi-dimensional original data set, and performing feature extraction and dimension internal clustering analysis processing on data of each dimension in the multi-dimensional original data set to generate a clustering result of each dimension; constructing a cross-dimension incidence matrix based on all clustering results; performing joint optimization on the feature representation, the clustering result and the cross-dimension incidence matrix by combining a multi-task loss function based on construction to obtain optimized feature representation, optimized clustering result and optimized cross-dimension incidence matrix; and generating a hierarchical clustering result and cross-dimension fusion feature representation based on the optimized feature representation, the clustering result and the cross-dimension incidence matrix. By adopting the method, the problems of inconsistent structure and insufficient information fusion in multi-dimensional data clustering can be effectively solved, and high-quality structured results and feature support are provided for multi-dimensional data analysis downstream tasks.

Description

Technical Field

[0001] This invention belongs to the field of multidimensional data analysis technology, and in particular relates to an intelligent clustering and feature extraction method for multidimensional data. Background Technology

[0002] With the widespread application of big data and artificial intelligence technologies, the comprehensive analysis of multi-dimensional data has become a key task in many fields, such as social media content understanding, precision medical diagnosis, and intelligent product recommendation. This type of data typically contains information from different dimensions, including images, text, and sequences. Effectively integrating these heterogeneous data and uncovering their inherent consistency is fundamental to achieving accurate cognition and decision-making. Currently, mainstream technical approaches focus on mapping data from different dimensions into a unified semantic space for subsequent analysis through cross-modal representation learning.

[0003] In traditional techniques, methods for handling multi-dimensional data clustering problems typically rely on feature-level or decision-level fusion strategies. For example, a shared latent representation space is learned through complex neural networks (such as cross-modal attention mechanisms) to eliminate superficial semantic differences between different dimensions. Standard clustering algorithms (such as K-means or spectral clustering) are then executed within this unified space to obtain a globally consistent data grouping result.

[0004] However, current fusion clustering methods have a fundamental limitation. These methods implicitly assume that all dimensions of data share a globally optimal clustering structure, attempting to force semantic alignment across different dimensions through technical means. But in reality, data from different dimensions may naturally possess their own reasonable and independent internal grouping logic. For example, the text dimension of a news article may naturally cluster by topic, while its accompanying images may be more reasonably categorized by visual style (such as close-ups of people). Forcibly merging them into a unified clustering framework will inevitably distort the clear data structure of at least one dimension, leading to biased clustering results and the loss of unique structured information for each dimension. This information distortion problem caused by "structural inconsistency" is a flaw that existing unified fusion methods struggle to overcome. Summary of the Invention

[0005] Therefore, it is necessary to provide a new multidimensional data analysis method that can respect the inherent clustering structure of data in each dimension and effectively model the relationships between dimensions, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides an intelligent clustering and feature extraction method for multi-dimensional data, including:

[0007] Obtain the multi-dimensional raw dataset, and extract features from the data of each dimension in the multi-dimensional raw dataset to obtain the feature representation of each dimension;

[0008] For each feature representation, intra-dimensional clustering analysis is performed to generate clustering results for each dimension; the clustering results include the cluster label for each sample and the cluster center for each cluster.

[0009] Based on all clustering results, a cross-dimensional association matrix is ​​constructed; the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions.

[0010] A multi-task loss function is constructed, and based on the multi-task loss function, the feature representation, clustering results, and cross-dimensional association matrix are jointly optimized to obtain the optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix.

[0011] Based on the optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix, hierarchical clustering results and cross-dimensional fused feature representations are generated; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the association relationship between each dimension.

[0012] Furthermore, for each feature representation, intra-dimensional clustering analysis is performed to generate clustering results for each dimension; wherein, the clustering results include the cluster label of each sample and the cluster center of each cluster, including:

[0013] For each dimension, set the search range for the number of clusters based on the feature representation;

[0014] For each candidate cluster number within the search range, a clustering algorithm is applied to group the feature representations to obtain a temporary cluster partitioning result;

[0015] Based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated, and the candidate cluster number with the largest clustering evaluation index value is determined as the optimal cluster number for the dimension.

[0016] Using the optimal number of clusters, cluster analysis is performed on the feature representation to obtain the cluster labels and cluster centers of the dimension;

[0017] The cluster labels and cluster centers of each dimension are summarized to obtain the clustering results for each dimension.

[0018] Furthermore, based on all clustering results, a cross-dimensional association matrix is ​​constructed; this matrix quantifies the degree of association between clusters in different dimensions, including:

[0019] Construct a dimension pair from any two different dimensions, and based on the clustering results of the two dimensions in the dimension pair, count the co-occurrence of clustering labels of all samples under the two dimensions in the dimension pair.

[0020] Based on the co-occurrence of cluster labels, the initial pairwise cross-dimensional association matrix of the dimension pairs is calculated; where each element of the initial pairwise cross-dimensional association matrix represents the initial joint probability that a sample belongs to two specific clusters in two dimensions at the same time.

[0021] Based on the feature vectors of each sample in different dimensions, the association confidence weight of each sample with respect to the dimension pair is calculated through an attention network;

[0022] The initial pairwise cross-dimensional association matrix is ​​weighted and corrected using all association confidence weights to obtain the refined pairwise association matrix.

[0023] Iterate through all possible dimension pairs, repeatedly perform the step of statistically analyzing the co-occurrence of cluster labels until a refined pairwise association matrix is ​​obtained, resulting in a refined pairwise association matrix corresponding to the number of dimension pairs. Combine all refined pairwise association matrices to generate a cross-dimensional association matrix.

[0024] Furthermore, a multi-task loss function is constructed, and based on this function, the feature representation, clustering results, and cross-dimensional association matrix are jointly optimized to obtain optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix, including:

[0025] Based on the feature reconstruction loss of each dimension and the intra-cluster sum of squares of the clustering results of each dimension, an intra-dimensional clustering loss term is constructed.

[0026] A cross-dimensional association loss term is constructed based on the cross-dimensional association matrix and the true association matrix based on the sample distribution.

[0027] Based on the feature representations of each dimension, construct a regularization loss term between dimensions;

[0028] The intra-dimensional clustering loss term, cross-dimensional association loss term, and inter-dimensional regularization loss term are weighted and summed according to preset weights to form a multi-task loss function;

[0029] The feature representation, clustering results, and cross-dimensional correlation matrix are iteratively optimized using the backpropagation algorithm and multi-task loss function until the multi-task loss function converges or the preset number of iterations is reached. The iteration is then stopped, resulting in the optimized feature representation, optimized clustering results, and optimized cross-dimensional correlation matrix.

[0030] Furthermore, based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated, including:

[0031] Based on the provisional cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated using the following formula:

[0032]

[0033] in, Here, α represents the clustering evaluation index, Silhouette(k) is the silhouette coefficient of the temporary cluster partitioning result when the number of clusters is k, CH(k) is the Calinski-Harabasz exponent of the temporary cluster partitioning result when the number of clusters is k, α is the weighting coefficient of the silhouette coefficient, and β is the weighting coefficient of the Calinski-Harabasz exponent.

[0034] Furthermore, a multi-dimensional original dataset is obtained, and features are extracted from the data of each dimension in the multi-dimensional original dataset to obtain the feature representation of each dimension, including:

[0035] Obtain the multi-dimensional raw dataset; the multi-dimensional raw dataset includes image dimension data, text dimension data, and numerical sequence dimension data;

[0036] Feature extraction is performed on the image dimension data, text dimension data, and numerical sequence dimension data in the multi-dimensional original dataset to obtain image feature vectors, semantic feature vectors, and sequence feature vectors.

[0037] The image feature vector, semantic feature vector, and sequence feature vector are normalized to obtain a standardized feature representation for each dimension.

[0038] Secondly, this application also provides an intelligent clustering and feature extraction device for multi-dimensional data, comprising:

[0039] The feature extraction module is used to acquire a multi-dimensional raw dataset and extract features from the data of each dimension in the multi-dimensional raw dataset to obtain the feature representation of each dimension.

[0040] The clustering analysis module is used to perform intra-dimensional clustering analysis on each feature representation and generate clustering results for each dimension; the clustering results include the cluster label of each sample and the cluster center of each cluster;

[0041] The matrix building module is used to construct a cross-dimensional association matrix based on all clustering results; the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions.

[0042] The data optimization module is used to construct a multi-task loss function and, based on the multi-task loss function, jointly optimize the feature representation, clustering results, and cross-dimensional association matrix to obtain optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix.

[0043] The clustering result and fusion feature generation module is used to generate hierarchical clustering results and cross-dimensional fusion feature representations based on the optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the relationship between each dimension.

[0044] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement any of the personalized learning path generation methods based on learning behavior profiles described in the embodiments of this application.

[0045] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the personalized learning path generation method based on learning behavior profiles as described in any of the embodiments of this application.

[0046] The aforementioned intelligent clustering and feature extraction method for multi-dimensional data extracts features from the original multi-dimensional data and performs intra-dimensional clustering; it constructs a cross-dimensional association matrix based on the clustering results to quantify the degree of cluster association between dimensions; and it jointly optimizes the features, clustering results, and association matrix through a multi-task loss function to achieve synergistic improvement of the three, ultimately generating hierarchical clustering results and cross-dimensional fused feature representations. This method effectively solves the problems of structural inconsistency and insufficient information fusion in multi-dimensional data clustering, providing high-quality structured results and feature support for downstream tasks in multi-dimensional data analysis. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating an intelligent clustering and feature extraction method for multi-dimensional data in one embodiment.

[0049] Figure 2 This is a flowchart illustrating the steps involved in performing intra-dimensional clustering analysis on each feature representation to generate clustering results for each dimension in one embodiment.

[0050] Figure 3This is a schematic diagram of the structure of an intelligent clustering and feature extraction device for multi-dimensional data in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] In one embodiment, an intelligent clustering and feature extraction method for multi-dimensional data is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and can be implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0053] Step S101: Obtain the multi-dimensional original dataset, and extract features from the data of each dimension in the multi-dimensional original dataset to obtain the feature representation of each dimension.

[0054] Feature extraction is a core step in data processing, especially in fields such as machine learning, computer vision, and natural language processing. Essentially, it involves filtering and extracting key information that is valuable to the target task from raw data and removing redundant and irrelevant data.

[0055] For example, a multi-dimensional raw dataset containing various types of data is obtained, typically including image data (such as medical images), text data (such as diagnostic reports), and numerical sequence data (such as user behavior time series). Feature extraction processing is performed on the data of each dimension to obtain a standardized feature representation for each dimension.

[0056] Step S102: Perform intra-dimensional clustering analysis on each feature representation to generate clustering results for each dimension; wherein, the clustering results include the clustering label of each sample and the cluster center of each cluster.

[0057] Cluster analysis is an unsupervised machine learning method that aims to automatically identify and divide groups of people with similar characteristics from an unlabeled dataset, i.e., clusters, to achieve the effect of grouping similar things together. Data points within the same cluster have high similarity, while data points between different clusters have low similarity.

[0058] For example, based on the obtained standardized feature representation, cluster analysis is performed independently on each dimension to generate cluster results for each dimension, which include cluster labels for each sample and cluster centers for each cluster.

[0059] Step S103: Based on all clustering results, construct a cross-dimensional association matrix; wherein, the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions.

[0060] For example, based on the clustering results (including cluster labels and cluster centers) across all dimensions, all dimensions are traversed, and different dimensions are paired (e.g., dimension A and dimension B). The co-occurrence of cluster labels for all samples under each dimension pair is then counted. For instance, the number of samples simultaneously belonging to cluster p of dimension A and cluster q of dimension B is counted, forming a co-occurrence statistics matrix. The co-occurrence statistics matrix is ​​then normalized by dividing each element (number of co-occurring samples) by the total number of samples, resulting in paired cross-dimensional association matrices. These matrices are then combined in dimensional order to obtain the cross-dimensional association matrix.

[0061] Step S104: Construct a multi-task loss function, and based on the multi-task loss function, jointly optimize the feature representation, clustering results, and cross-dimensional correlation matrix to obtain optimized feature representation, optimized clustering results, and optimized cross-dimensional correlation matrix.

[0062] Joint optimization is an optimization method that achieves overall optimization rather than local optimization in scenarios involving multiple objectives, multiple variables, or multiple systems by integrating related elements and coordinating constraints. Its core logic is to break the limitations of single-dimensional optimization, that is, to avoid the problem that optimizing only one objective (such as cost) leads to a decrease in the efficiency of other related objectives (such as quality).

[0063] For example, a multi-task loss function is constructed, comprising an intra-dimensional clustering loss term, a cross-dimensional association loss term, and an inter-dimensional regularization loss term. By minimizing this multi-task loss function, the feature representation, clustering results, and cross-dimensional association matrix are jointly optimized, resulting in optimized feature representations, optimized clustering results, and optimized cross-dimensional association matrices. Minimization is a frequently occurring concept in multiple domains; it involves adjusting, optimizing, or manipulating the target object to minimize a key attribute such as its size, value, or range.

[0064] Step S105: Based on the optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix, generate hierarchical clustering results and cross-dimensional fused feature representation; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the association relationship between each dimension.

[0065] For example, when generating hierarchical clustering results, the optimized clustering results of each dimension are used as the bottom-level information to reflect the structured grouping of samples within a single dimension. This optimized clustering result still includes the cluster label of each sample and the cluster center of each cluster. The optimized cross-dimensional association matrix is ​​used as the top-level information to quantify the association strength between clusters in different dimensions. For example, the higher the value of a matrix element, the stronger the association between the two clusters. The bottom-level and top-level information together constitute the hierarchical clustering result, fully presenting the clustering structure and inter-dimensional associations of the multi-dimensional data. When generating cross-dimensional fused feature representations, a weighted fusion strategy is adopted based on the optimized feature representations and the cross-dimensional association matrix: according to the association weights of each dimension cluster in the cross-dimensional association matrix, the optimized feature representations of samples in different dimensions are weighted and summed to finally obtain a cross-dimensional fused feature representation containing multi-dimensional information. Among them, the weighted fusion strategy is a method that combines multiple inputs or model outputs according to pre-set weights. By assigning different weights to data or models from different sources, this strategy can highlight more reliable or relevant information, thereby improving the accuracy and robustness of the overall results. The weights are usually determined based on performance evaluation, prior knowledge, or optimization objectives during training.

[0066] In this embodiment, features are extracted from multi-dimensional raw data and intra-dimensional clustering is performed. A cross-dimensional association matrix is ​​constructed based on the clustering results to quantify the degree of cluster association between dimensions. By jointly optimizing features, clustering results, and the association matrix through a multi-task loss function, the synergistic improvement of these three elements is achieved, ultimately generating hierarchical clustering results and cross-dimensional fused feature representations. This effectively solves the problems of structural inconsistency and insufficient information fusion in multi-dimensional data clustering, providing high-quality structured results and feature support for downstream tasks in multi-dimensional data analysis.

[0067] In one embodiment, such as Figure 2 As shown, for each feature representation, intra-dimensional clustering analysis is performed to generate clustering results for each dimension; wherein, the clustering results include the cluster label of each sample and the cluster center of each cluster, including:

[0068] Step S201: For each dimension, set the search range for the number of clusters based on the feature representation.

[0069] For example, a standardized feature representation based on a single dimension is used. This feature representation is in matrix form and includes key information such as the number of samples and the feature dimension. The total number of samples and the feature dimension data are extracted from the feature representation. Combined with domain experience in multi-dimensional data clustering, such as avoiding meaningless clustering due to too few samples within a cluster and structural fragmentation due to too many clusters, boundary conditions for the number of clusters are determined. The minimum number of clusters for multi-dimensional data clustering is typically set to 2 to ensure discriminative clustering. The maximum number of clusters needs to refer to the ratio between the total number of samples and the feature dimension, such as not exceeding a preset ratio of the total number of samples or not exceeding an empirical threshold to ensure the statistical validity of samples within the cluster. For example, based on the boundary conditions, a continuous set of cluster numbers is generated, i.e., the search range for the number of clusters in that dimension. The preset ratio of the total number of samples refers to a pre-set reasonable ratio threshold.

[0070] Step S202: For each candidate cluster number within the search range, a clustering algorithm is applied to group the feature representations to obtain a temporary cluster partitioning result.

[0071] The features are represented in matrix form, with each row of the matrix corresponding to a feature vector of a sample in that dimension, and each column corresponding to a feature dimension of that dimension. The sample and feature are represented in a one-to-one correspondence between "raw data unit - structured numerical description".

[0072] For example, for each candidate cluster number within the search range, a suitable clustering algorithm is selected to perform the grouping operation. Taking the K-Means algorithm as an example: First, the cluster centers corresponding to the candidate cluster numbers are randomly initialized. These cluster centers are numerical vectors with the same dimension as the sample feature representation. The initialization method can be to randomly select sample feature vectors or generate random vectors that conform to the feature distribution. Then, the distance from each sample to each cluster center (such as Euclidean distance) is calculated, and the sample is assigned to the nearest cluster. The sample mean of each cluster is recalculated as the new cluster center, and the distance calculation and sample assignment process is repeated until the cluster center positions are stable, i.e., the change is less than a preset change threshold. After each clustering is completed for a candidate cluster number, a temporary cluster partitioning result corresponding to that candidate cluster number is generated, containing a temporary cluster label for each sample to identify the cluster number to which the sample belongs, and the mapping relationship between all candidate cluster numbers and the corresponding temporary cluster partitioning result is output. Clustering algorithms are one of the core technologies of unsupervised machine learning. Without the need for manual labeling, they automatically divide a dataset into several clusters by analyzing the similarity of the data's own features. Data features within the same cluster are highly similar, while data features between different clusters are significantly different. K-Means is an unsupervised partitioning clustering algorithm. Its core objective is to divide a given dataset into a predetermined number (K) of clusters, maximizing the similarity of samples within the same cluster and minimizing the similarity of samples between different clusters. K samples are randomly selected as initial cluster centers. Then, all samples are traversed, and the distance from each sample to the K cluster centers (usually Euclidean distance) is calculated. The samples are then assigned to the nearest cluster. Subsequently, the mean of features of all samples in each cluster is recalculated and updated as new cluster centers. The iterative process of "sample assignment - cluster center update" is repeated until the cluster centers no longer change significantly (convergence) or the predetermined number of iterations is reached.

[0073] Step S203: Based on the temporary cluster partitioning results, calculate the clustering evaluation index value corresponding to each candidate cluster number, and determine the candidate cluster number with the largest clustering evaluation index value as the optimal cluster number for the dimension.

[0074] For example, based on the temporary clustering results and feature representations, two core clustering evaluation metrics are calculated: First, the silhouette coefficient, which measures clustering quality by calculating the intra-cluster compactness (i.e., the average distance between a sample and other samples within its own cluster) and inter-cluster separation (i.e., the average distance between a sample and the nearest sample from a different cluster), with the ratio of the difference between these two values ​​to the maximum value. The value ranges from -1 to 1, with a closer value to 1 indicating better clustering performance. Second, the Calinski-Harabasz index (CH index), which quantifies cluster discriminative power by comparing inter-cluster dispersion (i.e., the sum of squared distances between each cluster center and the total sample center) to intra-cluster dispersion (i.e., the sum of squared distances between each sample and the center of its own cluster). A higher value indicates better clustering performance. Based on the obtained silhouette coefficient and CH index, a comprehensive evaluation metric value for each candidate cluster number is calculated. All metric values ​​are compared, and the candidate cluster number corresponding to the maximum value is determined as the optimal cluster number for that dimension.

[0075] Step S204: Using the optimal number of clusters, perform cluster analysis on the feature representation to obtain the cluster labels and cluster centers of the dimensions.

[0076] For example, the same clustering algorithm as in step S202 is used to execute the complete clustering process with the optimal number of clusters as the parameter: Cluster centers are initialized based on the optimal number of clusters, and through multiple rounds of iterative optimization, including calculating the distance from samples to cluster centers, assigning sample cluster affiliations, and updating cluster centers, until the cluster centers converge (the positional changes meet the preset accuracy requirements). After clustering, two core results are generated: one is the cluster label, i.e., the final cluster number corresponding to each sample, forming a label vector; the other is the cluster center, i.e., the mean vector of all sample features within each cluster, forming a cluster center matrix. Validation of these two results is performed, such as confirming that each cluster contains samples and that the cluster center matches the feature distribution of samples within the cluster. After successful validation, the cluster label and cluster center for that dimension are output. The preset accuracy requirement is a quantitative standard or judgment criterion pre-set before execution in scenarios such as cluster analysis, used to measure the accuracy or acceptable error of the results; validity validation is a key step in data processing and other fields used to verify whether information / operations conform to preset rules or target requirements, with the aim of filtering invalid, erroneous, or non-compliant content.

[0077] Step S205: Summarize the cluster labels and cluster centers for each dimension to obtain the clustering results for each dimension.

[0078] For example, based on the current dimension's cluster labels and cluster centers, the cluster labels and cluster centers of other dimensions that have already been processed are simultaneously retrieved. The cluster labels and cluster centers of a single dimension are structurally integrated to establish a correspondence of "dimensional identifier - cluster label - cluster center," ensuring the uniqueness and integrity of the data association. Then, for all dimensions, according to a preset dimension order (e.g., image dimension, text dimension, numerical sequence dimension), the structured data of each dimension are summarized sequentially, forming a set containing the unique cluster labels and cluster centers for each dimension. During the summarization process, data consistency checks are required, such as confirming that the sample size for each dimension is consistent and the cluster label numbering rules are unified, to avoid subsequent processing anomalies due to data format differences. Finally, independent clustering results for each dimension are output. Among these, the data consistency check is a crucial operation to ensure that the logic, content, and format of the data conform to preset rules throughout its entire lifecycle, including storage, transmission, and processing, avoiding contradictions, errors, or inconsistencies.

[0079] In this embodiment, a reasonable range for the number of clusters is first set; candidate clusters are traversed to generate temporary results, which are then evaluated using a combination of silhouette coefficient and CH index; clustering is performed with the optimal number of clusters, and the results are summarized. This approach respects the inherent grouping logic of data across different dimensions and avoids information loss caused by forcibly unifying clustering structures across different dimensions.

[0080] In one embodiment, a cross-dimensional association matrix is ​​constructed based on all clustering results; wherein, the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions, including:

[0081] Step S301: Construct a dimension pair from any two different dimensions, and based on the clustering results of the two dimensions in the dimension pair, count the co-occurrence of clustering labels of all samples under the two dimensions in the dimension pair.

[0082] For example, all dimensions are traversed, and any two different dimensions are paired according to the combination rules to ensure that all heterogeneous dimension combinations are covered without repetition or omission. For each dimension pair, the cluster label data of the two dimensions in the pair is extracted, and a label mapping relationship is established on a sample-by-sample basis: for each sample, its cluster number in the first dimension and the cluster number in the second dimension are recorded, forming a sample-double-cluster label correspondence. Based on this correspondence, the co-occurrence of cluster labels is statistically analyzed, that is, the number of samples that simultaneously belong to a cluster in the first dimension and a cluster in the second dimension is counted, forming a co-occurrence statistical matrix. The rows of the matrix correspond to the cluster numbers in the first dimension, the columns correspond to the cluster numbers in the second dimension, and the matrix elements are the number of co-occurring samples of the corresponding cluster combination.

[0083] Step S302: Based on the co-occurrence of cluster labels, calculate the initial pairwise cross-dimensional association matrix of dimension pairs; wherein each element of the initial pairwise cross-dimensional association matrix represents the initial joint probability that a sample belongs to two specific clusters in two dimensions.

[0084] The core definition of the initial pairwise cross-dimensional association matrix is ​​that each element in the matrix is ​​the initial joint probability that a sample belongs to a specific cluster in both dimensions of the pair, which is used to initially quantify the strength of the association between clusters.

[0085] For example, for each element in the co-occurrence statistics matrix, i.e., the number of co-occurring samples corresponding to the cluster combination, normalization is performed, and the normalized value is divided by the total number of samples to obtain the joint probability value of the cluster combination. All normalized joint probability values ​​are then sequentially filled in according to the row and column order of the co-occurrence statistics matrix to form an initial pairwise cross-dimensional correlation matrix, with the matrix dimensions consistent with the co-occurrence statistics matrix. The normalization process transforms the sample counts into a probability distribution, ensuring that the matrix elements are in the [0,1] interval, facilitating subsequent unified quantitative analysis.

[0086] Step S303: Based on the feature vectors of each sample in different dimensions, calculate the association confidence weight of each sample with respect to the dimension pair through an attention network.

[0087] For example, feature vectors of all samples in both dimensions of the dimension pair are extracted. For each sample, its feature vectors in the two dimensions are concatenated to form a joint feature vector for that sample in relation to the dimension pair. The purpose of concatenation is to integrate the feature information of the samples in heterogeneous dimensions, providing a complete basis for association confidence assessment. The joint feature vector is then input into an attention network, which learns the correlation between the features of samples in heterogeneous dimensions and dynamically allocates confidence weights: the network performs a non-linear transformation on the joint features through a fully connected layer, and then outputs normalized weight values ​​through a Sigmoid activation function. The weight values ​​are in the range [0,1], and the larger the value, the stronger the feature correlation and the higher the information confidence of the sample in the two dimensions of the current dimension pair. This calculation process is repeated for all samples to obtain the association confidence weight for each sample in relation to the current dimension pair. Attention networks are a type of deep learning model architecture. The idea behind them is to enable the model to focus on or notice the most important parts when processing input data. This mechanism allows the model to dynamically select the information most useful for the current task by assigning different weights to different parts of the input. Nonlinear transformation is a data or signal processing method relative to linear transformation. Its characteristic is that the input and output do not satisfy a proportional and superimposed linear relationship. That is, when the input changes, the magnitude and trend of the output change are not fixed and cannot be described by a linear formula such as "output = input × fixed coefficient + constant".

[0088] Step S304: Use all association confidence weights to perform weighted correction on the initial pairwise cross-dimensional association matrix to obtain the refined pairwise association matrix.

[0089] For example, for each element in the pairwise cross-dimensional association matrix, i.e., the initial joint probability of a certain cluster combination, all samples belonging to that cluster combination are selected through the sample-double-cluster label correspondence to form a sample subset. The average association confidence weight of this sample subset is calculated, which is the sum of the association confidence weights of all samples in the subset divided by the number of samples in the sample subset. Each element in the initial pairwise cross-dimensional association matrix is ​​multiplied by the average association confidence weight of the corresponding cluster combination to obtain the corrected element value. All corrected elements are arranged in the original matrix row and column order to form a refined pairwise association matrix.

[0090] Step S305: Traverse all possible dimension pairs and repeat the step of statistically analyzing the co-occurrence of cluster labels until a refined pairwise association matrix is ​​obtained. Obtain a refined pairwise association matrix corresponding to the number of dimension pairs, and combine all refined pairwise association matrices to generate a cross-dimensional association matrix.

[0091] For example, following the dimension pair formation rules in step S301, all possible dimension pairs are traversed. For each dimension pair, the co-occurrence statistics in step S301, the initial matrix calculation in step S302, the confidence weight acquisition in step S303, and the matrix correction in step S304 are repeated sequentially to obtain a set of refined pairwise association matrices corresponding one-to-one with all dimension pairs. Then, the overall dimension of the cross-dimensional association matrix is ​​determined: the total number of rows and columns are both the sum of the number of clusters in all dimensions, and the clusters in each dimension are arranged in a preset order to form a unified cluster index. Each refined pairwise association matrix is ​​treated as a submatrix, and its corresponding cluster index range for each dimension pair is filled into the corresponding position in the cross-dimensional association matrix. For example, if the cluster index range for a certain dimension pair is row ab and column cd, then the refined matrix of that dimension pair is filled into row ab and column cd of the overall matrix. After all submatrices are filled, a cross-dimensional association matrix covering all dimension cluster relationships is obtained.

[0092] In this embodiment, all dimension pairs are constructed and the co-occurrence of cluster labels is statistically analyzed. Based on the co-occurrence, an initial pairwise cross-dimensional association matrix is ​​calculated for each dimension pair. An attention network is then used to dynamically calculate the sample association confidence weights, and the initial matrix is ​​weighted and corrected to obtain a refined pairwise association matrix. Finally, by traversing all dimension pairs and piecing together the refined pairwise association matrix in blocks according to cluster indices, a complete cross-dimensional association matrix is ​​generated. This effectively solves the problems of one-sided association relationships and low reliability in traditional cross-dimensional association modeling, ensuring the scientific rigor and effectiveness of multi-dimensional data association analysis.

[0093] In one embodiment, a multi-task loss function is constructed, and based on the multi-task loss function, the feature representation, clustering results, and cross-dimensional association matrix are jointly optimized to obtain optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix, including:

[0094] Step S401: Based on the feature reconstruction loss of each dimension and the intra-cluster sum of squares of the clustering results of each dimension, construct the intra-dimensional clustering loss term.

[0095] For example, the feature representation of each dimension is input into the autoencoder, which maps the features to the latent space. The decoder then reconstructs the reconstructed features with the same dimensions as the original features. The difference between the original features and the reconstructed features, such as the mean squared error (MSE), is calculated to obtain the feature reconstruction loss for each dimension. This loss is used to ensure that the features retain their original semantic information during the clustering process.

[0096]

[0097] in, The loss is the feature reconstruction loss, where n is the number of samples. For feature dimension, For the m-th sample feature, The features output by the reconstructor. For example, for each dimension, the cluster center of each sample is found based on the cluster label. The sum of squared Euclidean distances between the sample feature vector and the corresponding cluster center vector is calculated to obtain the intra-cluster sum of squares loss. This loss is used to constrain the compactness of samples within the cluster and reduce intra-cluster dispersion.

[0098]

[0099] in, For the sum of squares loss within the cluster, For each corresponding cluster center, the feature reconstruction loss of each dimension is weighted and summed with the intra-cluster sum of squares loss. The weights can be preset according to the importance of the dimensions to obtain the intra-dimensional clustering loss of each dimension, and then the sums are used to form the global intra-dimensional clustering loss term. An autoencoder (AE) is an unsupervised artificial neural network model used to learn an efficient representation (i.e., encoding) of data and reconstruct the original data using this representation. Essentially, it completes a closed-loop process of "input → compression → reconstruction," comprising an encoder and a decoder. The encoder, the first part of the autoencoder, typically consists of multiple layers of neural networks (such as fully connected layers and convolutional layers). By progressively reducing the feature dimension, it maps the high-dimensional feature representation of the input to a low-dimensional latent space, generating a compact latent feature vector. The decoder, the second part, uses a network structure symmetrical to or adapted to the encoder. Taking the latent feature vector as input, it progressively increases the dimension, mapping back to the same dimension as the original feature representation, and finally outputs the reconstructed features. The reconstructed features must approximate the original features as closely as possible. The difference between the reconstructed features and the original features (e.g., measured by mean squared error) serves as a loss signal, used to iteratively optimize the network parameters of the encoder and decoder, ensuring that the latent features accurately retain the key information of the original features. The output of the entire process is a reconstructed feature with the same dimension as the original features.

[0100] Step S402: Construct a cross-dimensional association loss term based on the cross-dimensional association matrix and the true association matrix based on the sample distribution.

[0101] The true association matrix is ​​generated by analyzing the inherent statistical regularities of multi-dimensional raw data. For example, it quantifies the true association strength between different dimensional clusters based on indicators such as mutual information and Pearson correlation coefficient of sample features in different dimensions, and serves as a benchmark for evaluating the accuracy of the predicted association matrix.

[0102] For example, Kullback-Leibler divergence (KL divergence) is used to calculate the difference between the predicted association matrix (i.e., the cross-dimensional association matrix) and the true association matrix. KL divergence is a metric that measures the asymmetry between two probability distributions, and its calculation formula is as follows:

[0103]

[0104] For cross-dimensional correlation matrices, For the true correlation matrix, p and q are cluster indices. p represents the row index of the matrix, corresponding to a certain cluster in a certain dimension, and q represents the column index of the matrix, corresponding to a certain cluster in another dimension. The range of values ​​for p and q covers all clusters in all dimensions, i.e., p,q∈[1,C], where C is the total number of clusters. For the true correlation matrix The element value in the p-th row and q-th column represents the association probability between the "p-th cluster" and the "q-th cluster" in the real scenario; For the prediction of cross-dimensional correlation matrix The element value in the p-th row and q-th column represents the predicted association probability between the "p-th cluster" and the "q-th cluster". For example, for each dimensional pair, the KL divergence is calculated, and then summed over all dimensional pairs to obtain the cross-dimensional association loss term.

[0105] Step S403: Based on the feature representations of each dimension, construct the inter-dimensional regularization loss term.

[0106] For example, a linear projection layer maps feature representations of each dimension to a latent space of the same dimension, ensuring comparability of features across different dimensions. Subsequently, a distribution alignment strategy is used to calculate the loss, aligning the mean and variance. The mean and variance of latent features in each dimension are calculated, constraining the mean and variance of features across different dimensions to tend towards consistency, resulting in the inter-dimensional regularization loss term.

[0107]

[0108] in, The inter-dimensional regularization loss term is used to constrain the consistency of the distribution of features of different dimensions in the latent space, reducing the interference caused by dimensional heterogeneity. The smaller the loss value, the closer the distribution of features of different dimensions is. i and j are dimension indices, with i ranging from [1, k−1] and j ranging from [i+1, k]. The combination of the two is used to traverse all unordered dimension pairs. For example, when k=3, it traverses (i=1, j=2), (i=1, j=3), and (i=2, j=3) to avoid repeatedly calculating the distribution differences of the same dimension pair. and These are the latent spatial feature matrices of the i-th and j-th dimensions, respectively, obtained by mapping the original feature matrices of the i-th and j-th dimensions through a linear projection layer; and The mean of the latent feature matrices in the i-th and j-th dimensions is calculated separately. and The elements of all sample feature vectors are obtained, which are used to reflect the overall distribution center of the potential features of this dimension; and The variances of the latent feature matrices in the i-th and j-th dimensions are calculated separately. and The variance of the elements of the feature vectors of all samples is used to reflect the degree of dispersion of the potential features in that dimension. The square of the Euclidean distance is a measure of the difference between two vectors.

[0109] Step S404: The intra-dimensional clustering loss term, cross-dimensional association loss term, and inter-dimensional regularization loss term are weighted and summed according to preset weights to form a multi-task loss function.

[0110] The preset weights are set according to the task priority of multi-dimensional data clustering and feature extraction. For example, if more attention is paid to the accuracy of single-dimensional clustering, the weight of the intra-dimensional clustering loss term can be increased; if the focus is on the accuracy of cross-dimensional association, the weight of the cross-dimensional association loss term can be increased, and the sum of the three weight coefficients is 1.

[0111] For example, each loss term is multiplied by its corresponding preset weight and then summed to form a multi-task loss function. This function is used to comprehensively balance the quality of single-dimensional clustering, the accuracy of cross-dimensional association, and the consistency of features between dimensions.

[0112] Step S405: The feature representation, clustering results, and cross-dimensional correlation matrix are iteratively optimized using the backpropagation algorithm and multi-task loss function until the multi-task loss function converges or the preset number of iterations is reached. Then, the iteration is stopped, and the optimized feature representation, optimized clustering results, and optimized cross-dimensional correlation matrix are obtained.

[0113] Among them, the backpropagation algorithm (BP algorithm) is the core algorithm for deep learning and neural network training. It is used to adjust the weights and biases of each layer in the network by calculating the error in reverse, so that the prediction results of the neural network are closer to the real labels and the model learning is realized. The preset number of iterations is the upper limit of the loop update based on the parameters set in advance according to the task requirements before performing backpropagation optimization.

[0114] For example, based on the constructed multi-task loss function and the set of core parameters to be optimized, including network weights in the feature extraction process (such as parameters of CNN and BERT), cluster center parameters in the clustering results, element parameters of the cross-dimensional association matrix, and attention network weights, the backpropagation algorithm can be used to calculate the gradient of the multi-task loss function with respect to each parameter to be optimized, i.e., the rate and direction of change of the loss value with respect to the parameters. Subsequently, the optimizer updates the parameters according to the gradient information: adjusting the feature extraction network weights to optimize the feature representation, updating the cluster centers to improve cluster compactness, and correcting the association matrix parameters to approximate the true association relationship. The above gradient calculation and parameter update process is repeated, and the change of the multi-task loss function is monitored after each iteration. When the decrease in the loss value is less than a preset convergence threshold, or the number of iterations reaches a preset upper limit, the optimization stops. Finally, the optimized feature representation, the optimized clustering results (including optimized cluster labels and cluster centers), and the optimized cross-dimensional association matrix are output. Among them, the optimizer is the core tool in deep learning used to update model parameters such as weights and biases. By iteratively adjusting the parameters, it minimizes the model's loss function, which measures the difference between the predicted result and the true value, allowing the model to learn to fit the data and improve task performance. The preset convergence threshold is a pre-set judgment standard value in numerical calculation and algorithm iteration, used to determine whether the iteration process is close enough to the target result, thereby stopping the iteration.

[0115] In this embodiment, three types of loss terms are constructed to precisely constrain the quality of single-dimensional clustering, the accuracy of cross-dimensional association, and the consistency of features between dimensions. The various loss terms are then weighted and summed to form a multi-task loss function. Iterative optimization is performed using the backpropagation algorithm, allowing feature representation, clustering results, and cross-dimensional association matrices to mutually promote and improve simultaneously. This effectively solves the problem of result bias caused by the disconnect between feature extraction, clustering, and association modeling in traditional methods. The optimized data retains a reasonable independent clustering structure for each dimension and possesses accurate cross-dimensional association quantification relationships.

[0116] In one embodiment, based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated, including:

[0117] Step S501: Based on the temporary cluster partitioning results, calculate the clustering evaluation index value corresponding to each candidate cluster number using the following formula:

[0118]

[0119] in, Here, α represents the clustering evaluation index, Silhouette(k) is the silhouette coefficient of the temporary cluster partitioning result when the number of clusters is k, CH(k) is the Calinski-Harabasz exponent of the temporary cluster partitioning result when the number of clusters is k, α is the weighting coefficient of the silhouette coefficient, and β is the weighting coefficient of the Calinski-Harabasz exponent.

[0120] Wherein, S(k) is the clustering evaluation index value corresponding to the number of candidate clusters k, which is used to comprehensively quantify the quality of temporary cluster division; α is the weight coefficient of the silhouette coefficient, and β is the weight coefficient of the Calinski-Harabasz exponent. Both are predefined non-negative constants and satisfy α+β=1, and their values ​​can be adjusted according to the clustering task requirements of multi-dimensional data.

[0121] For example, for each candidate cluster number k, Silhouette(k) is first calculated based on its corresponding temporary cluster partitioning result and standardized feature representation: For each sample, the mean feature distance between the sample and all samples in the same cluster (intra-cluster distance) and the mean feature distance between the sample and all samples in the nearest different cluster (inter-cluster distance) are calculated. The silhouette value of a single sample is calculated using the silhouette coefficient formula, and then the mean of the silhouette values ​​of all samples is taken as Silhouette(k). Then, CH(k) is calculated: Sample clusters are partitioned according to the temporary clustering labels. The sum of squared feature distances between each cluster center and the total sample center (inter-cluster dispersion) and the sum of squared feature distances between each sample and its cluster center (intra-cluster dispersion) are calculated. CH(k) is obtained using the Calinski-Harabasz exponent formula. Finally, Silhouette(k) and CH(k) are multiplied by their corresponding weight coefficients α and β, respectively, and summed to obtain the clustering evaluation index value S(k) corresponding to the candidate cluster number k. The mapping relationship between all candidate cluster numbers and their corresponding S(k) is then output.

[0122] In this embodiment, a comprehensive evaluation index is constructed by integrating the silhouette coefficient and the Calinski-Harabasz index. This makes the evaluation results of each candidate cluster number more comprehensive and objective, effectively improving the reliability of intra-dimensional clustering results.

[0123] In one embodiment, a multi-dimensional raw dataset is obtained, and features are extracted from the data of each dimension in the multi-dimensional raw dataset to obtain a feature representation for each dimension, including:

[0124] Step S601: Obtain the multi-dimensional raw dataset; wherein, the multi-dimensional raw dataset includes image dimension data, text dimension data, and numerical sequence dimension data.

[0125] For example, a multi-dimensional raw dataset is acquired, and the core feature of this dataset is that it contains at least three heterogeneous dimensions of data, with each dimension corresponding to a sample. Image dimension data refers to raw data carrying visual information, such as product display images, medical diagnostic images, and scene monitoring footage; text dimension data refers to structured or unstructured data composed of natural language, such as product descriptions, medical records, and news reports; and numerical sequence dimension data refers to numerical data arranged in chronological or logical order, such as continuous user consumption records, real-time sensor monitoring data, and time-series data of device operating status. After acquiring the data, a data integrity check is performed to remove invalid samples lacking key dimension information, ensuring that each sample contains complete image, text, and numerical sequence dimension data. The final output is a well-structured and complete multi-dimensional raw dataset. Integrity check is a technical means used to verify whether data maintains its original state during transmission, storage, or copying, and has not been accidentally damaged, tampered with, or lost, thus ensuring data integrity.

[0126] Step S602: Perform feature extraction processing on the image dimension data, text dimension data and numerical sequence dimension data in the multi-dimensional original dataset to obtain image feature vector, semantic feature vector and sequence feature vector.

[0127] For example, based on multi-dimensional original datasets, appropriate feature extraction techniques are adopted for the characteristics of data in different dimensions. For image-dimensional data, a Convolutional Neural Network (CNN) can be used for feature extraction. Convolutional layers capture local image features, pooling layers compress dimensions, and fully connected layers integrate high-order semantics, transforming a two-dimensional pixel matrix into a one-dimensional numerical vector, i.e., an image feature vector. This vector contains key information such as image edges, textures, and semantics. For text-dimensional data, a pre-trained language model (such as Bidirectional Encoder Representations from Transformers, BERT) can be used. The text is first segmented and word-embedded into a vector sequence, and then the Transformer encoder captures contextual semantic relationships, outputting a semantic feature vector containing the core meaning of the text. For numerical sequence-dimensional data, a Long Short-Term Memory (LSTM) network can be used, leveraging its ability to capture temporal dependencies to extract trend features and key node information from the sequence data, generating a sequence feature vector. The above three types of feature vectors are output, with each sample corresponding to a set of image, semantic, and sequence feature vectors. Among them, Convolutional Neural Networks (CNNs) are deep learning models inspired by biological visual systems, capable of automatically extracting spatial features and widely used in image recognition, video analysis, medical imaging, and other fields; Pre-trained language models refer to a type of natural language processing model that first completes general language ability learning on large-scale general text data, and then fine-tunes and adapts for specific tasks such as text classification and question answering. Through the "pre-training + fine-tuning" mode, it solves the problems of insufficient data for small tasks and weak model generalization ability. BERT is a representative model among them; Transformer encoders are one of the core components of Transformer models, mainly used to extract deep features of input data, and are widely used in tasks such as text understanding (such as BERT). The characteristic is that parallel processing is achieved based on the self-attention mechanism; Long Short-Term Memory (LSTM) networks are a special type of Recurrent Neural Network (RNN) used to solve the gradient vanishing / gradient explosion problem of traditional RNNs when processing long sequence data. That is, traditional RNNs have difficulty remembering the key information in the early stages of the sequence, such as the core subject at the beginning of a sentence, while LSTM achieves selective memory and forgetting of information through a unique gating mechanism.

[0128] Step S603: Normalize the image feature vector, semantic feature vector, and sequence feature vector to obtain a standardized feature representation for each dimension.

[0129] The purpose of normalization is to eliminate the differences in the dimensions and numerical ranges of feature vectors of different dimensions, so as to ensure that the features of each dimension are comparable and computable.

[0130] For example, Z-Score normalization can be used for image feature vectors, semantic feature vectors, and sequence feature vectors. Based on the mean and standard deviation of the feature vectors, the data is transformed into a standard distribution with a mean of 0 and a variance of 1, preserving the relative distribution characteristics of the features. During processing, normalization is performed independently for each dimension to avoid interference between data from different dimensions. All sample feature vectors in each dimension are processed according to the same standard. Finally, the normalized feature representation of each dimension is output, which is in the form of a matrix structure, with rows corresponding to samples and columns corresponding to feature dimensions. The formula for calculating Z-Score normalization is:

[0131]

[0132] σ represents a single original data set; μ is the mean of the data set (i.e., the average of all data); σ is the standard deviation of the data set, which is an indicator of the dispersion of the data.

[0133] In this embodiment, a complete multi-dimensional raw dataset containing images, text, and numerical sequences is acquired. Specialized feature extraction techniques are applied to each dimension to capture core visual, semantic, and temporal information. Normalization is used to eliminate dimensional differences, resulting in feature representations for each dimension. This effectively improves the compatibility of multi-dimensional data and avoids the shortcomings of traditional feature extraction, such as coarse processing of heterogeneous data and poor feature comparability.

[0134] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0135] Based on the same inventive concept, this application also provides an intelligent clustering and feature extraction device for multi-dimensional data, which is a method for intelligent clustering and feature extraction of multi-dimensional data. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent clustering and feature extraction device for multi-dimensional data provided below can be found in the limitations of the intelligent clustering and feature extraction method for multi-dimensional data described above, and will not be repeated here.

[0136] In one exemplary embodiment, such as Figure 3 As shown, an intelligent clustering and feature extraction device 300 for multi-dimensional data is provided, comprising:

[0137] The feature extraction module 301 is used to acquire a multi-dimensional raw dataset and extract features from the data of each dimension in the multi-dimensional raw dataset to obtain the feature representation of each dimension.

[0138] The clustering analysis module 302 is used to perform intra-dimensional clustering analysis on each feature representation and generate clustering results for each dimension; wherein, the clustering results include the cluster label of each sample and the cluster center of each cluster;

[0139] The matrix construction module 303 is used to construct a cross-dimensional association matrix based on all clustering results; wherein, the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions;

[0140] The data optimization module 304 is used to construct a multi-task loss function and, based on the multi-task loss function, jointly optimize the feature representation, clustering results, and cross-dimensional association matrix to obtain optimized feature representation, optimized clustering results, and optimized cross-dimensional association matrix.

[0141] The clustering result and fusion feature generation module 305 is used to generate hierarchical clustering results and cross-dimensional fusion feature representations based on the optimized feature representation, optimized clustering results and optimized cross-dimensional association matrix; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the association relationship between each dimension.

[0142] In one embodiment, the clustering analysis module 302 is further configured to:

[0143] For each dimension, set the search range for the number of clusters based on the feature representation;

[0144] For each candidate cluster number within the search range, a clustering algorithm is applied to group the feature representations to obtain a temporary cluster partitioning result;

[0145] Based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated, and the candidate cluster number with the largest clustering evaluation index value is determined as the optimal cluster number for the dimension.

[0146] Using the optimal number of clusters, cluster analysis is performed on the feature representation to obtain the cluster labels and cluster centers of the dimension;

[0147] The cluster labels and cluster centers of each dimension are summarized to obtain the clustering results for each dimension.

[0148] In one embodiment, the matrix construction module 303 is further configured to:

[0149] Construct a dimension pair from any two different dimensions, and based on the clustering results of the two dimensions in the dimension pair, count the co-occurrence of clustering labels of all samples under the two dimensions in the dimension pair.

[0150] Based on the co-occurrence of cluster labels, the initial pairwise cross-dimensional association matrix of the dimension pairs is calculated; where each element of the initial pairwise cross-dimensional association matrix represents the initial joint probability that a sample belongs to two specific clusters in two dimensions at the same time.

[0151] Based on the feature vectors of each sample in different dimensions, the association confidence weight of each sample with respect to the dimension pair is calculated through an attention network;

[0152] The initial pairwise cross-dimensional association matrix is ​​weighted and corrected using all association confidence weights to obtain the refined pairwise association matrix.

[0153] Iterate through all possible dimension pairs, repeatedly perform the step of statistically analyzing the co-occurrence of cluster labels until a refined pairwise association matrix is ​​obtained, resulting in a refined pairwise association matrix corresponding to the number of dimension pairs. Combine all refined pairwise association matrices to generate a cross-dimensional association matrix.

[0154] In one embodiment, the data optimization module 304 is further configured to:

[0155] Based on the feature reconstruction loss of each dimension and the intra-cluster sum of squares of the clustering results of each dimension, an intra-dimensional clustering loss term is constructed.

[0156] A cross-dimensional association loss term is constructed based on the cross-dimensional association matrix and the true association matrix based on the sample distribution.

[0157] Based on the feature representations of each dimension, construct a regularization loss term between dimensions;

[0158] The intra-dimensional clustering loss term, cross-dimensional association loss term, and inter-dimensional regularization loss term are weighted and summed according to preset weights to form a multi-task loss function;

[0159] The feature representation, clustering results, and cross-dimensional correlation matrix are iteratively optimized using the backpropagation algorithm and multi-task loss function until the multi-task loss function converges or the preset number of iterations is reached. The iteration is then stopped, resulting in the optimized feature representation, optimized clustering results, and optimized cross-dimensional correlation matrix.

[0160] In one embodiment, the clustering analysis module 302 is further configured to:

[0161] Based on the provisional cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated using the following formula:

[0162]

[0163] in, Here, α represents the clustering evaluation index, Silhouette(k) is the silhouette coefficient of the temporary cluster partitioning result when the number of clusters is k, CH(k) is the Calinski-Harabasz exponent of the temporary cluster partitioning result when the number of clusters is k, α is the weighting coefficient of the silhouette coefficient, and β is the weighting coefficient of the Calinski-Harabasz exponent.

[0164] In one embodiment, the feature extraction module 301 is further configured to:

[0165] Obtain the multi-dimensional raw dataset; the multi-dimensional raw dataset includes image dimension data, text dimension data, and numerical sequence dimension data;

[0166] Feature extraction is performed on the image dimension data, text dimension data, and numerical sequence dimension data in the multi-dimensional original dataset to obtain image feature vectors, semantic feature vectors, and sequence feature vectors.

[0167] The image feature vector, semantic feature vector, and sequence feature vector are normalized to obtain a standardized feature representation for each dimension.

[0168] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent clustering and feature extraction method for multi-dimensional data as described above.

[0169] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0170] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0171] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent clustering and feature extraction of multi-dimensional data, characterized in that, The method includes: Obtain a multi-dimensional raw dataset, and extract features from the data of each dimension in the multi-dimensional raw dataset to obtain the feature representation of each dimension; For each of the aforementioned feature representations, intra-dimensional clustering analysis is performed to generate clustering results for each of the aforementioned dimensions; wherein, the clustering results include clustering labels for each sample and cluster centers for each cluster; Based on all the clustering results, a cross-dimensional association matrix is ​​constructed; wherein, the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions; A multi-task loss function is constructed, and based on the multi-task loss function, the feature representation, the clustering result, and the cross-dimensional correlation matrix are jointly optimized to obtain the optimized feature representation, the optimized clustering result, and the optimized cross-dimensional correlation matrix. Based on the optimized feature representation, the optimized clustering results, and the optimized cross-dimensional association matrix, hierarchical clustering results and cross-dimensional fused feature representations are generated; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the association relationship between each dimension.

2. The method according to claim 1, characterized in that, For each of the aforementioned feature representations, intra-dimensional clustering analysis is performed to generate clustering results for each dimension; wherein, the clustering results include cluster labels for each sample and cluster centers for each cluster, including: For each dimension, the search range for the number of clusters is set according to the feature representation; For each candidate cluster number within the search range, a clustering algorithm is applied to group the feature representations to obtain a temporary cluster partitioning result; Based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated, and the candidate cluster number with the largest clustering evaluation index value is determined as the optimal cluster number for the dimension. Using the optimal number of clusters, cluster analysis is performed on the feature representation to obtain the cluster labels and cluster centers for the specified dimension; The clustering labels and cluster centers for each dimension are summarized to obtain the clustering results for each dimension.

3. The method according to claim 1, characterized in that, Based on all the clustering results, a cross-dimensional association matrix is ​​constructed; wherein, the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions, including: Any two different dimensions are combined into a dimension pair, and based on the clustering results of the two dimensions in the dimension pair, the co-occurrence of clustering labels of all samples under the two dimensions in the dimension pair is statistically analyzed. Based on the co-occurrence of the cluster labels, the initial pairwise cross-dimensional association matrix of the dimension pairs is calculated; wherein, each element of the initial pairwise cross-dimensional association matrix represents the initial joint probability that a sample belongs to two specific clusters in two dimensions at the same time; Based on the feature vectors of each sample in different dimensions, the association confidence weight of each sample with respect to the dimension pair is calculated by an attention network; The initial pairwise cross-dimensional association matrix is ​​weighted and corrected using all the aforementioned association confidence weights to obtain a refined pairwise association matrix. Iterate through all possible dimension pairs, repeatedly perform the step of statistically analyzing the co-occurrence of the cluster labels until the refined pairwise association matrix is ​​obtained, obtain the refined pairwise association matrix corresponding to the number of dimension pairs, and combine all the refined pairwise association matrices to generate a cross-dimensional association matrix.

4. The method according to claim 1, characterized in that, The step of constructing a multi-task loss function and, based on the multi-task loss function, jointly optimizing the feature representation, the clustering result, and the cross-dimensional association matrix to obtain optimized feature representation, optimized clustering result, and optimized cross-dimensional association matrix includes: Based on the feature reconstruction loss of each dimension and the intra-cluster sum of squares of the clustering results of each dimension, an intra-dimensional clustering loss term is constructed. Based on the cross-dimensional association matrix and the true association matrix based on the sample distribution, a cross-dimensional association loss term is constructed; Based on the feature representations of each dimension, construct a regularization loss term between dimensions; The intra-dimensional clustering loss term, the cross-dimensional association loss term, and the inter-dimensional regularization loss term are weighted and summed according to preset weights to form a multi-task loss function; The feature representation, clustering result, and cross-dimensional correlation matrix are iteratively optimized using the backpropagation algorithm and the multi-task loss function until the multi-task loss function converges or reaches the preset number of iterations, at which point the iteration stops, resulting in the optimized feature representation, optimized clustering result, and optimized cross-dimensional correlation matrix.

5. The method according to claim 2, characterized in that, The calculation of the clustering evaluation index value corresponding to each candidate cluster number based on the temporary clustering result includes: Based on the temporary cluster partitioning results, the clustering evaluation index value corresponding to each candidate cluster number is calculated using the following formula: in, Here, α represents the clustering evaluation index, Silhouette(k) is the silhouette coefficient of the temporary cluster partitioning result when the number of clusters is k, CH(k) is the Calinski-Harabasz exponent of the temporary cluster partitioning result when the number of clusters is k, α is the weighting coefficient of the silhouette coefficient, and β is the weighting coefficient of the Calinski-Harabasz exponent.

6. The method according to claim 1, characterized in that, The step of obtaining a multi-dimensional raw dataset and extracting features from the data of each dimension in the multi-dimensional raw dataset to obtain a feature representation for each dimension includes: Obtain a multi-dimensional raw dataset; wherein the multi-dimensional raw dataset includes image dimension data, text dimension data, and numerical sequence dimension data; Feature extraction processing is performed on the image dimension data, the text dimension data, and the numerical sequence dimension data in the multi-dimensional original dataset to obtain image feature vectors, semantic feature vectors, and sequence feature vectors. The image feature vector, the semantic feature vector, and the sequence feature vector are normalized to obtain a standardized feature representation for each dimension.

7. An intelligent clustering and feature extraction device for multi-dimensional data, characterized in that, The device includes: The feature extraction module is used to acquire a multi-dimensional raw dataset and extract features from the data of each dimension in the multi-dimensional raw dataset to obtain the feature representation of each dimension. The clustering analysis module is used to perform intra-dimensional clustering analysis on each of the feature representations to generate clustering results for each dimension; wherein, the clustering results include the clustering label of each sample and the cluster center of each cluster; A matrix construction module is used to construct a cross-dimensional association matrix based on all the clustering results; wherein the cross-dimensional association matrix is ​​used to quantify the degree of association between clusters in different dimensions; The data optimization module is used to construct a multi-task loss function and, based on the multi-task loss function, jointly optimize the feature representation, the clustering result, and the cross-dimensional correlation matrix to obtain optimized feature representation, optimized clustering result, and optimized cross-dimensional correlation matrix. The clustering result and fusion feature generation module is used to generate hierarchical clustering results and cross-dimensional fusion feature representations based on the optimized feature representation, the optimized clustering results, and the optimized cross-dimensional association matrix; wherein, the hierarchical clustering results include the clustering results of each dimension and the cross-dimensional association matrix used to describe the association relationship between each dimension.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images