A multi-view financial data clustering integration method and device

Abstract

The present invention relates to the technical field of data clustering and integration, and discloses a multi-view financial data clustering and integration method and device. The method comprises: processing financial multi-view data using a K-Means clustering method to generate an ensemble pool, and then using a fuzzy membership function to generate a binary partition matrix; learning the global structure of the label space through the binary partition matrix to obtain an initial co-correlation matrix, and optimizing it using an LSR model with a Frobenius norm to obtain a label co-correlation matrix; learning a subspace projection matrix, an affinity matrix, and a sample adjacency matrix of each view, optimizing the co-correlation matrix from a feature space, and obtaining a feature co-correlation matrix; combining the label co-correlation matrix and the feature co-correlation matrix to obtain an optimized co-correlation matrix; projecting the optimized co-correlation matrix into a constraint space for constructing an undirected graph; and partitioning the undirected graph using a graph cut algorithm to obtain a final clustering integration result of the financial multi-view data.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a multi-view financial data clustering integration method and device. Background Art

[0002] Ensemble clustering techniques can classify financial data into distinct clusters reflecting different market behaviors and investment characteristics. These characteristics are associated with diverse financial performance and can influence investment decisions and risk management effectiveness. Therefore, identifying similarities and differences in financial data is crucial for financial analysis and investment strategy formulation.

[0003] Multi-view cluster analysis combines information from different financial data sources, each of which potentially offers unique insights, contributing to a comprehensive understanding of market dynamics and financial products. However, this approach faces multiple challenges: balancing the information from various financial data sources, assessing their value, and coping with the noise and complexity inherent in high-dimensional financial data.

[0004] Current financial data clustering techniques often have limited co-correlation matrix representation capabilities when processing multi-source heterogeneous data. This makes it difficult to effectively integrate the complementary information between the label and feature spaces within the data, limiting the application of multi-view clustering in areas such as financial risk assessment, portfolio optimization, and market segmentation. The paper "Multi-view Ensemble Clustering based on BLS-Autoencoder" uses the BLS-Autoencoder for feature extraction, but fails to effectively integrate the complementary information between the label and feature spaces. Furthermore, the constructed co-correlation matrix is ​​more suitable for multimedia data with a distinct hierarchical structure, limiting its performance when processing financial data. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve the technical problem is to provide a multi-view financial data clustering integration method, including:

[0007] A K-Means clustering method is used to process each subspace of the financial multi-view data to generate an ensemble pool, which is then processed using a fuzzy membership function to generate a binary partition matrix. The global structure of the label space of the financial multi-view data is learned using the binary partition matrix to obtain an initial co-correlation matrix. The initial co-correlation matrix is ​​optimized using an LSR model with a Frobenius norm to obtain a label co-correlation matrix. The LSR model represents a least squares model.

[0008] For each view of the financial multi-view data, the subspace projection matrix, affinity matrix, and sample adjacency matrix are learned, and adaptive weights are added according to the learning error to optimize the co-correlation matrix from the feature space to obtain the feature co-correlation matrix;

[0009] Combining the label co-correlation matrix and the feature co-correlation matrix, an optimized co-correlation matrix is ​​obtained; the optimized co-correlation matrix is ​​projected into a constraint space, and an undirected graph is constructed based on the projected constraint space; the undirected graph is partitioned using a graph cut algorithm to obtain the final clustering integration result of the financial multi-view data.

[0010] Preferably, the LSR model with Frobenius norm is used to optimize the initial co-correlation matrix to obtain a label co-correlation matrix, which is expressed as:

[0011] ;

[0012] in, The training data set is a binary partition matrix generated by K-Means clustering and fuzzy membership function. represents the initial co-correlation matrix, represents the Frobenius norm, represents the trade-off parameter of the co-correlation matrix, represents the clustering ensemble result, Representation based on the co-correlation matrix The constructed Laplace matrix, represents the transpose of the matrix, represents the rank of the matrix; v represents the view, and V represents the total number of views.

[0013] Preferably, the feature co-correlation matrix is ​​obtained by optimizing the co-correlation matrix from the feature space, which is expressed as:

[0014] ;

[0015] in, represents the subspace projection matrix, Representing multiple views of financial data, represents the row-2 norm and column-1 norm, represents the adaptive weight; represents the parameter that controls the weight distribution, Represents the affinity matrix.

[0016] Preferably, the label co-correlation matrix and the feature co-correlation matrix are combined to obtain an optimized co-correlation matrix, which is expressed as:

[0017] .

[0018] Preferably, the optimized co-correlation matrix is ​​projected into the constraint space and expressed as:

[0019] ;

[0020] in, Represents the elements projected into the constraint space; represents the optimized co-correlation matrix The elements in , i represents the row index, and j represents the column index.

[0021] Preferably, the undirected graph is constructed based on the projection constraint space, which is expressed as:

[0022] ;

[0023] ;

[0024] in, represents an undirected bipartite graph, represents the edge weight of the undirected bipartite graph node, represents a set of sample nodes, represents the optimized co-correlation matrix, Represents the fuzzy partition matrix The elements in, k represents the row index, l represents the column index; fuzzy partition matrix is the optimized co-correlation matrix The results obtained after spectral clustering and normalization; represents a node consisting of samples and clusters, Representation based on fuzzy partition matrix The connection weight, E represents the set of all valid edge connections of the undirected bipartite graph, Indicates that there is an edge connection between sample node k and cluster node l.

[0025] Preferably, the graph cutting algorithm is used to divide the undirected graph to obtain the final clustering integration result of the financial multi-view data, specifically: the normalized cut algorithm is used to cut the undirected bipartite graph into several non-intersecting subgraphs, each subgraph contains the sample node set and a subset of nodes corresponding to the co-correlation matrix, and the properties of the undirected bipartite graph are maintained. The sample nodes belonging to the same subgraph are divided into the same cluster, thereby forming the final consensus division. , where each sample node is uniquely assigned to a cluster, and the clustering integration result of the multi-view financial data is obtained.

[0026] The present invention also provides a multi-view financial data clustering and integration device, comprising:

[0027] The label space construction module uses the K-Means clustering method to process each subspace of the financial multi-view data to generate an ensemble pool. The ensemble pool is then processed using a fuzzy membership function to generate a binary partition matrix. The global structure of the label space of the financial multi-view data is learned through the binary partition matrix to obtain an initial co-correlation matrix. The initial co-correlation matrix is ​​optimized using the LSR model with the Frobenius norm to obtain a label co-correlation matrix.

[0028] The feature space construction module learns the subspace projection matrix, affinity matrix, and sample adjacency matrix for each view of the financial multi-view data, adds adaptive weights based on the learning error, and optimizes the co-correlation matrix from the feature space to obtain the feature co-correlation matrix;

[0029] The data clustering integration module combines the label co-correlation matrix and the feature co-correlation matrix to obtain an optimized co-correlation matrix; the optimized co-correlation matrix is ​​projected into a constraint space, and an undirected graph is constructed based on the projected constraint space; the undirected graph is partitioned using a graph cut algorithm to obtain the final clustering integration result of the financial multi-view data.

[0030] The present invention has the following beneficial effects:

[0031] (1) The multi-view clustering ensemble method based on co-correlation matrix optimization constructed in this paper obtains a high-quality co-correlation matrix by combining the LSR model with feature space optimization. The trained optimization model is used to perform feature extraction and relationship mining on the input financial multi-view data, which significantly improves the representation ability of the co-correlation matrix and efficiently completes the multi-view clustering analysis of financial data.

[0032] (2) This paper optimizes the features of the original data through a subspace projection matrix and an adaptive weight mechanism, and simultaneously learns the sample adjacency matrix to capture the intrinsic relationship between samples. It runs a standardized cut algorithm on the constructed undirected graph to obtain high-quality consensus partitions, and then completes the multi-view clustering integration of financial data, significantly optimizing the accuracy and stability of the model and enhancing its application value in complex financial scenarios.

[0033] (3) This method breaks through the limitation of traditional multi-view clustering algorithms that only learn from a single space and fully integrates the complementary information of the label space and feature space. In the co-correlation matrix construction stage, this method innovatively captures global structural information through the LSR model of the Frobenius norm. In the matrix optimization stage, the interference of noise features is effectively reduced by simultaneously learning the subspace projection matrix, affinity matrix and adaptive weights. In the integration stage, the undirected graph structure is constructed through the precisely optimized co-correlation matrix, achieving high-quality clustering partitioning, thereby enhancing the clustering effect of financial multi-view data.

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A diagram showing the steps of a multi-view financial data clustering and integration method according to an embodiment of the present invention;

[0036] Figure 2 This is a flow chart of a multi-view financial data clustering and integration method according to an embodiment of the present invention;

[0037] Figure 3 This is a comparison chart of the results of an ablation experiment on a multi-view financial data clustering and integration method according to an embodiment of the present invention;

[0038] Figure 4 The figure is a schematic structural diagram of a multi-view financial data clustering and integration device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] See also Figure 1 and Figure 2 FIG. 1 is a method step diagram and a flow chart of a multi-view financial data clustering and integration method according to an embodiment of the present invention, which includes the following steps:

[0040] S101: Use the K-Means clustering method to process each subspace of the financial multi-view data to generate an ensemble pool, then use the fuzzy membership function to process the ensemble pool to generate a binary partition matrix; use the binary partition matrix to learn the global structure of the label space of the financial multi-view data to obtain an initial co-correlation matrix; use the LSR model with the Frobenius norm to optimize the initial co-correlation matrix to obtain a label co-correlation matrix;

[0041] S102, for each view of the financial multi-view data, learning a subspace projection matrix, an affinity matrix, and a sample adjacency matrix, and adding adaptive weights according to the learning error, optimizing the co-correlation matrix from the feature space, and obtaining a feature co-correlation matrix;

[0042] S103, combining the label co-correlation matrix and the feature co-correlation matrix to obtain an optimized co-correlation matrix; projecting the optimized co-correlation matrix into the constraint space, and constructing an undirected graph based on the projected constraint space; using the standardized cut algorithm to partition the undirected graph to obtain the final clustering integration result of the financial multi-view data.

[0043] Specifically, in S101, K-Means clustering is used to extract integrated clustering results from different views to form a diverse ensemble pool; the fuzzy membership function is used to quantify the probabilistic association between samples and clusters, generating a binary partition matrix with rich information. The LSR model is optimized using a training data set, and by minimizing the reconstruction error and regularization term, an accurate estimation of the co-correlation matrix is ​​achieved, effectively reducing noise interference and the impact of outliers, so that the co-correlation matrix can accurately reflect the inherent global structural characteristics of the data in the label space. The training data set is a data set constructed from multi-view financial data. The co-correlation matrix obtained by learning the global structure of the label space of the data through the binary partition matrix is ​​as follows:

[0044] ;

[0045] in, The training data set is a binary partition matrix generated by K-Means clustering and fuzzy membership function. represents the co-correlation matrix obtained by training the global structure of the data, represents the Frobenius norm, represents the trade-off parameter of the co-correlation matrix, represents the clustering ensemble result, Representation based on the co-correlation matrix The constructed Laplace matrix, represents the transpose of the matrix, Represents the rank of the matrix.

[0046] Specifically, in S102, the adjacency relationship of samples is further mined from the feature layer of the original data. Since the original feature space of each view is noisy, the subspace projection matrix, affinity matrix and sample adjacency matrix are learned simultaneously, and adaptive weights are added according to the learning error to better optimize the co-correlation matrix from the feature space. The co-correlation matrix obtained by feature space learning is as follows:

[0047]

[0048] in, represents the subspace projection matrix, Representing multiple views of financial data, represents the row-2 norm and column-1 norm, Adaptive weights; represents the parameter that controls the weight distribution, represents the affinity matrix; represents the sample adjacency matrix.

[0049] Specifically, the step S103 includes the following steps:

[0050] S1031, the label space and feature space of the data are integrated to obtain the overall optimization objective function, which is expressed as:

[0051] .

[0052] The co-correlation matrix projected into the constraint space is as follows:

[0053] ;

[0054] represents the optimized co-correlation matrix The elements in , i represents the row index, and j represents the column index.

[0055] S1032, constructing an undirected bipartite graph based on the co-correlation matrix, wherein the nodes in the undirected bipartite graph utilize all samples and clusters in the integrated pool, and the connection weights between the nodes of the undirected bipartite graph correspond to the values ​​in the fuzzy partition matrix, as shown in the following formula:

[0056] ;

[0057] ;

[0058] in, represents an undirected bipartite graph, represents the edge weight of the undirected bipartite graph node, represents a set of sample nodes, represents the optimized co-correlation matrix, Represents the fuzzy partition matrix The elements in, k represents the row index, l represents the column index; fuzzy partition matrix is the optimized co-correlation matrix The results obtained after spectral clustering and normalization; represents a node consisting of samples and clusters, Representation based on fuzzy partition matrix The connection weight, E represents the set of all valid edge connections of the undirected bipartite graph, Indicates that there is an edge connection between sample node k and cluster node l.

[0059] S1033, using a graph cut algorithm to cut the undirected bipartite graph into several non-overlapping subgraphs, each subgraph contains the sample node set and a subset of nodes corresponding to the co-correlation matrix, maintaining the properties of the undirected bipartite graph, and the sample nodes belonging to the same subgraph are divided into the same cluster, thus forming the final consensus partition , where each sample node is uniquely assigned to a cluster, and the clustering integration result of the multi-view financial data is obtained.

[0060] The following examples of the present invention were validated using experimental datasets consisting of a dataset of 100 plant leaves (100 leaves), a facial image dataset (ORL), 474 images (Caltech 7), and seven image categories (MSRC-v1). Clustering performance was evaluated using two widely used metrics: normalized mutual information (NMI) and adjusted Rand index (ARI). The standardized mutual information is normalized, where 0 means the two partitions are completely unrelated and 1 means the two partitions are consistent. It is a cluster evaluation metric obtained by adjusting the original Rand Index based on the consistency of sample pair assignments. The results are shown in Tables 1 and 2.

[0061] Table 1 - Comparative analysis of NMI measurements:

[0062]

[0063] Table 2 - Comparative analysis of ARI measurements:

[0064]

[0065] According to the NMI metric, our method MvCOMOC ranks first on three datasets and second on the MSRC-v1 dataset. In terms of ARI, it ranks first on two datasets.

[0066] At the same time, ablation experiments are used to prove the effectiveness of the model constructed by the present invention. Figure 3 As shown in the figure, the effectiveness of the model of the present invention is demonstrated by comparing three models. Among the three models, MvCO is a feature space model without binary partition feature space optimization; MvCO-SA is a label space model without feature layer label space optimization; MvCOMOC is a complete model including label space and feature space; as can be seen from the figure, MvCOMOC has the highest NMI and ARI on different datasets. This phenomenon can be attributed to two main reasons: (1) The multi-view clustering ensemble method based on co-correlation matrix optimization can synergistically integrate the global structure of the label space of the LSR model and the subspace mapping relationship of the feature space to construct a high-quality co-correlation matrix, thereby effectively reducing the interference of noise features. (2) The co-correlation matrix optimization method based on subspace affinity introduces an adaptive weight mechanism by simultaneously learning the subspace projection matrix and the sample adjacency matrix, accurately capturing the intrinsic relationship between samples and making the integration result more robust.

[0067] In summary, the method of the present invention innovatively combines co-correlation matrix optimization with subspace affinity for application in financial multi-view data analysis. The method first uses K-Means clustering to construct a diverse ensemble pool and generates a binary partitioning matrix through a fuzzy membership function. Then, an LSR model with a Frobenius norm is designed to learn and refine the co-correlation matrix to capture the global structural information in the label space. Next, the feature space level is optimized by introducing a subspace projection matrix and a sample adjacency matrix with L2,1 norm constraints, while the contribution of each view is dynamically adjusted using an adaptive weight mechanism. Finally, the optimized co-correlation matrix is ​​projected into the constraint space and an undirected graph is constructed, and high-quality clustering ensemble results are obtained through the NCUT algorithm. A large number of experiments have verified the excellent performance of this method in financial multi-view data analysis, demonstrated the effectiveness of the combination of the LSR model and subspace affinity optimization, and demonstrated the good application prospects of co-correlation matrix optimization technology in multi-view clustering tasks. The present invention combines the optimization of the co-correlation matrix with the subspace affinity method, which is particularly suitable for the multi-source heterogeneity and time-varying correlation characteristics of financial data. It can more accurately identify market sector rotation, asset risk aggregation and investment style changes, and provide more reliable technical support.

[0068] See also Figure 4 FIG. 1 is a schematic diagram of the structure of a multi-view financial data clustering and integration device according to an embodiment of the present invention, comprising:

[0069] The label space construction module 401 processes each subspace of the financial multi-view data using a K-Means clustering method to generate an ensemble pool, then processes the ensemble pool using a fuzzy membership function to generate a binary partition matrix; learns the global structure of the label space of the financial multi-view data using the binary partition matrix to obtain an initial co-correlation matrix; and optimizes the initial co-correlation matrix using an LSR model with a Frobenius norm to obtain a label co-correlation matrix.

[0070] A feature space construction module 402 learns a subspace projection matrix, an affinity matrix, and a sample adjacency matrix for each view of the financial multi-view data, adds adaptive weights based on learning errors, and optimizes a co-correlation matrix from the feature space to obtain a feature co-correlation matrix.

[0071] The data clustering integration module 403 combines the label co-correlation matrix and the feature co-correlation matrix to obtain an optimized co-correlation matrix; projects the optimized co-correlation matrix into the constraint space, and constructs an undirected graph based on the projected constraint space; and uses the standardized cut algorithm to partition the undirected graph to obtain the final clustering integration result of the financial multi-view data.

[0072] The functional implementation of the modules in the multi-view financial data clustering and integration device is consistent with that of the multi-view financial data clustering and integration method, and will not be repeated here.

[0073] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-view financial data clustering and integration method, characterized by: include: The K-Means clustering method is used to process each subspace of the financial multi-view data to generate an ensemble pool, and then the fuzzy membership function is used to process the ensemble pool to generate a binary partition matrix; The global structure of the label space of the financial multi-view data is learned by a binary partition matrix to obtain an initial co-correlation matrix; the initial co-correlation matrix is ​​optimized using an LSR model with a Frobenius norm to obtain a label co-correlation matrix; the LSR model represents a least squares model; For each view of the financial multi-view data, the subspace projection matrix, affinity matrix, and sample adjacency matrix are learned, and adaptive weights are added according to the learning error to optimize the co-correlation matrix from the feature space to obtain the feature co-correlation matrix; Combine the label co-correlation matrix and the feature co-correlation matrix to obtain the optimized co-correlation matrix; The optimized co-correlation matrix is ​​projected into the constraint space, and an undirected graph is constructed based on the projected constraint space; the undirected graph is partitioned using a graph cut algorithm to obtain the final clustering integration result of the financial multi-view data; The LSR model with Frobenius norm is used to optimize the initial co-correlation matrix to obtain the label co-correlation matrix, which is expressed as: Among them, B (v) The training data set is a binary partition matrix generated by K-Means clustering and fuzzy membership function, where M represents the initial co-correlation matrix. represents the Frobenius norm, α represents the trade-off parameter of the co-correlation matrix, P represents the clustering ensemble result, L M represents the Laplace matrix constructed based on the co-correlation matrix M, T represents the transpose of the matrix, tr represents the rank of the matrix; v represents the view, and V represents the total number of views; The feature co-correlation matrix is ​​obtained by optimizing the co-correlation matrix from the feature space, which is expressed as: Among them, W (v) represents the subspace projection matrix, X (v) Representing multiple views of financial data,|||| 2,1 represents the row norm and column norm, r (v) represents the adaptive weight; β represents the parameter that controls the weight distribution, A (v) represents the affinity matrix; The optimized co-correlation matrix is ​​obtained by combining the label co-correlation matrix and the feature co-correlation matrix, which is expressed as:

2. The multi-view financial data clustering and integration method according to claim 1, characterized in that: The optimized co-correlation matrix is ​​projected into the constraint space and expressed as: Among them, M ij Represents the elements projected into the constraint space; represents the optimized co-correlation matrix The elements in , i represents the row index, and j represents the column index.

3. The multi-view financial data clustering and integration method according to claim 1, characterized in that: The undirected graph is constructed based on the projection constraint space and is expressed as: in, represents an undirected bipartite graph, represents the edge weight of the undirected bipartite graph node, X represents the sample node set, represents the optimized co-correlation matrix, P v(kl) Represents the fuzzy partition matrix P v The elements in, k represents the row index, l represents the column index; the fuzzy partition matrix P v is the optimized co-correlation matrix The results obtained after spectral clustering and normalization; Represents a node consisting of samples and clusters, Weight(P v(kl) ) represents the fuzzy partition matrix P v , E represents the set of all valid edge connections of the undirected bipartite graph, and (k,l)∈E represents that there is an edge connection between the sample node k and the cluster node l.

4. The multi-view financial data clustering integration method according to claim 3, characterized in that: The graph cutting algorithm is used to divide the undirected graph to obtain the final clustering integration result of the financial multi-view data. Specifically, the standardized cut algorithm is used to cut the undirected bipartite graph into several non-overlapping subgraphs. Each subgraph contains a subset of the sample node set and the nodes corresponding to the co-correlation matrix. The properties of the undirected bipartite graph are maintained. The sample nodes belonging to the same subgraph are divided into the same cluster, thus forming the final consensus partition P. * , where each sample node is uniquely assigned to a cluster, and the clustering integration result of the multi-view financial data is obtained.

5. A multi-view financial data clustering and integration device, characterized in that: The method for implementing the multi-view financial data clustering integration method according to any one of claims 1 to 4 comprises: The label space construction module uses the K-Means clustering method to process each subspace of the financial multi-view data to generate an ensemble pool, and then uses a fuzzy membership function to process the ensemble pool to generate a binary partition matrix. The binary partition matrix is ​​used to learn the global structure of the label space of the financial multi-view data to obtain an initial co-correlation matrix. The initial co-correlation matrix is ​​optimized using an LSR model with a Frobenius norm to obtain a label co-correlation matrix. The LSR model represents a least squares model. The feature space construction module learns the subspace projection matrix, affinity matrix, and sample adjacency matrix for each view of the financial multi-view data, adds adaptive weights based on the learning error, and optimizes the co-correlation matrix from the feature space to obtain the feature co-correlation matrix; The data clustering integration module combines the label co-correlation matrix and the feature co-correlation matrix to obtain an optimized co-correlation matrix; the optimized co-correlation matrix is ​​projected into a constraint space, and an undirected graph is constructed based on the projected constraint space; the undirected graph is partitioned using a graph cut algorithm to obtain the final clustering integration result of the financial multi-view data.

Images