Financial data multi-source feature selection method and system based on self-paced tensor learning

Abstract

The invention discloses a financial data multi-source feature selection method and system based on self-paced tensor learning, and the method comprises the steps: replacing a pseudo constraint with a pseudo constraint to carry out feature selection, carrying out the optimization of a selection model through L2, p norm regularization, and constructing a co-correlation matrix through integrated member fusion based on the relation of data correlation; information in multiple views can be better explored by utilizing a tensor constraint mechanism; and a global co-correlation matrix is constructed. And self-paced learning is fused into pseudo constraints to guide the model to perform feature selection, so that a feature selection objective function is established, financial data is sorted based on data correlation, and key features are selected. According to the method, a pseudo-constraint special group selection model is combined with a tensor global consensus matrix based on self-paced learning, a pseudo-constraint-guided self-paced tensor constraint feature selection objective function is constructed, and key feature selection of financial multi-source data is optimized.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method and system for multi-source feature selection of financial data based on self-paced tensor learning. Background Art

[0002] Financial data is often scattered across multiple sources (such as finance, legal affairs, public opinion, and behavior), and includes diverse data types (court decisions, financial reports, stock trading behavior, etc.). Furthermore, these data have privacy and security issues, making it difficult to obtain usable labels. Multi-view unsupervised feature selection can extract cross-domain core correlation features, providing data support for enterprise development.

[0003] Furthermore, due to the sensitive nature of financial data, obtaining labeled data can be challenging. Unsupervised learning, particularly multi-view unsupervised feature selection, can uncover underlying patterns in data without explicit labels. This is also true for image recognition, particularly in applications where automatic recognition is required but labeling is difficult, such as customer profiling and segmentation. Using this approach can help reduce the need for manual labeling.

[0004] Multi-view feature selection can effectively extract complementary information from diverse financial data sources. Each source often contains unique behavioral patterns or economic signals, helping to comprehensively identify and deeply analyze the core drivers of financial policy or market changes. However, this approach faces numerous challenges, including achieving a reasonable balance between information weights across data sources, effectively assessing the contribution of each perspective, and addressing the modeling complexity inherent in the high dimensionality and redundancy of financial data. Similarly, in image recognition, multi-view methods can process data from different perspectives or sensors, achieving a more comprehensive understanding of the target object. This concept has also been widely adopted in financial applications. For example, in quantitative analysis and intelligent decision-making systems, technical graphical data such as candlestick charts, volatility heatmaps, or factor network graphs can be extracted through image recognition techniques and then fused with structured data (such as financial statements, macroeconomic indicators, and credit ratings). This cross-modal data integration approach can improve the accuracy of financial market modeling and the depth of policy impact assessment, providing a more comprehensive decision-making basis for investment strategy optimization and risk control.

[0005] Current feature selection technology often has a heavy computational load when processing large-scale data. For example, the Chinese invention patent application with application number CN202510583447.4, "A multi-view feature selection method and system for government data based on consensus clustering," can effectively identify government data. However, when processing financial data, it cannot effectively control the complexity of the model for large-scale financial data, resulting in poor results and limiting the application prospects of multi-view feature selection in financial data analysis. Summary of the Invention

[0006] To address the above problems, the present invention proposes a multi-source feature selection method and system for financial data based on self-paced tensor learning. The method guides feature selection through regularization using pseudo-constraints and the L2,p norm, and introduces a feature selection model that filters features with pseudo-constraints by combining tensor constrained clustering with self-paced learning. This method efficiently integrates multi-view financial data information, improves feature extraction accuracy and efficiency, and effectively reduces data redundancy and computational complexity.

[0007] The specific plan is as follows:

[0008] On the one hand, the multi-source feature selection method for financial data based on self-paced tensor learning includes:

[0009] S1. For each financial multi-source data view, prior knowledge is introduced into the financial data feature selection model through pseudo-constraints to guide feature selection and obtain the key feature set within each financial view. The distance between the financial multi-source data in the key feature set within each financial view is measured using the L2,p norm. The financial data feature selection model is regularized and trained based on the distance between the financial multi-source data to obtain a pseudo-constraint-guided financial data feature selection model.

[0010] S2, fuses the members of each financial data view based on the relationship between samples and clusters, obtains the fusion results within each view, integrates the fusion results within each view into a binary partition matrix, and then integrates the binary partition matrix to generate a sample similarity matrix;

[0011] S3, performing tensor constraints on the sample similarity matrix to construct a tensor-constrained global co-correlation matrix;

[0012] S4, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint guided financial data feature selection model by self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints;

[0013] S5, fuse the data in each financial data view through the feature selection objective function, sort the fused financial multi-source data in descending order, and select the top m features.

[0014] Furthermore, in S1, the financial data feature selection model is regularized and trained based on the distance between financial multi-source data. The calculation formula is as follows:

[0015]

[0016] in, X v represents the data of each financial view, n represents the number of financial view samples, d vRepresents the sample dimension of each financial view; W v represents the feature selection matrix, Used to learn the relationship between data point features and clusters; Q represents the pseudo constraint matrix, Q∈R n×n , k is the projection dimension; α is a hyperparameter used to control the trade-off of the L2,p regularization term; ‖W v ‖ 2,p Represents the feature selection matrix W v Carry out L 2,p norm; Represents the feature selection matrix W v The minimum value of .

[0017] Furthermore, the S2 specifically includes:

[0018] In each financial view, the relationship between samples and clusters is applied to generate a binary partition matrix S v ∈{0,1} n×c ; Where n represents the number of financial view samples, and c represents the generated binary partition cluster;

[0019] Perform b binary partitions on the binary partition matrix and integrate them to obtain the integrated binary partition matrix Among them, c i The clusters generated for each binary partition;

[0020] The integrated binary partition matrix is ​​fused into a similarity matrix. The calculation formula is as follows:

[0021]

[0022] Among them, Q v Represents the obtained similarity matrix.

[0023] Furthermore, in S3, the sample similarity matrix is ​​tensor-constrained, and the calculation formula is as follows:

[0024]

[0025] stF={Q 1 ,Q 2 ,...,Q v}

[0026] Where F represents the tensor matrix; V represents the number of financial views; || || * Represents a tensor constraint; represents the minimum value of the tensor matrix F; st represents the constraint condition;

[0027] The tensor constraint global co-correlation matrix is ​​constructed through tensor constraints. The calculation formula is as follows:

[0028]

[0029] Where β represents the learning rate for hyperparameter adjustment Q.

[0030] Furthermore, in S4, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints. The calculation formula is as follows:

[0031]

[0032] Among them, S v represents the self-paced learning coefficient, S v ∈R n×n , used to control the amount of financial data in the next iteration; represents the selection vector, z v The coefficient used to control the quantity of financial data for the next iteration.

[0033] On the other hand, the multi-source feature selection system for financial data based on self-paced tensor learning includes:

[0034] A multi-feature selection model construction module is used to introduce prior knowledge into the financial data feature selection model through pseudo-constraints for each financial multi-source data view to guide feature selection and obtain a key feature set within each financial view. The module also measures the distance between the financial multi-source data in the key feature set within each financial view using the L2,p norm, and performs regularized training on the financial data feature selection model based on the distance between the financial multi-source data to obtain a pseudo-constraint-guided financial data feature selection model.

[0035] The similarity matrix construction module is used to fuse the members of each financial data view based on the relationship between samples and clusters, obtain the fusion results within each view, integrate the fusion results within each view into a binary partition matrix, and then integrate the binary partition matrices to generate a sample similarity matrix;

[0036] A tensor constraint module is used to construct a tensor-constrained global co-correlation matrix by performing tensor constraints on the sample similarity matrix;

[0037] A feature selection objective function construction module is used to integrate the tensor-constrained global co-correlation matrix into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints;

[0038] The feature selection module is used to fuse the data in each financial data view through the feature selection objective function, sort the fused financial multi-source data in descending order, and select the top m features.

[0039] The present invention adopts the above technical solution and has the following beneficial effects:

[0040] (1) The present invention performs subspace feature selection through pseudo-constraints and adopts L2,p norm for regularization constraints, which can prevent the model from blindly learning from massive data;

[0041] (2) The present invention generates multiple binary partition matrices through binary partitioning, then obtains a similarity matrix based on member correlation, and uses tensor constraints to obtain key information of each view, thereby significantly improving the accuracy and efficiency of feature extraction;

[0042] (3) The present invention optimizes the key feature selection of financial view data by combining the consensus co-correlation matrix after tensor constraint with the multi-view feature selection model, and adopts self-paced learning for feature screening to construct a self-paced tensor-guided feature selection objective function. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flowchart of a method for multi-source feature selection of financial data based on self-paced tensor learning according to an embodiment of the present invention;

[0044] Figure 2 This is a structural block diagram of a multi-source feature selection system for financial data based on self-paced tensor learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0046] like Figure 1 As shown, the multi-source feature selection method for financial data based on self-paced tensor learning of the present invention includes:

[0047] S1. For each financial multi-source data view, prior knowledge is introduced into the financial data feature selection model through pseudo-constraints to guide feature selection and obtain the key feature set within each financial view. The distance between the financial multi-source data in the key feature set within each financial view is measured by the L2,p norm. The financial data feature selection model is regularized and trained based on the distance between the financial multi-source data to obtain a financial data feature selection model guided by pseudo-constraints.

[0048] Specifically, the financial data feature selection model is regularized and trained based on the distance between financial multi-source data. The calculation formula is as follows:

[0049]

[0050] in, X vrepresents the data of each financial view, n represents the number of financial view samples, d v Represents the sample dimension of each financial view; W v represents the feature selection matrix, Used to learn the relationship between data point features and clusters; Q represents the pseudo constraint matrix, Q∈R n×n , k is the projection dimension; α is a hyperparameter used to control the trade-off of the L2,p regularization term; ‖W v ‖ 2,p Represents the feature selection matrix W v Carry out L 2,p norm; Represents the feature selection matrix W v The minimum value of .

[0051] This model automatically selects the most important features that best represent the original information from multiple perspectives of financial data. For example, consider financial images from multiple sources (such as stock price charts or trading volume graphs), each providing information from a different perspective. This feature selection method automatically finds features that are representative across multiple perspectives, thereby preserving the essence of the data while removing redundant or noisy information.

[0052] The data processed in this way can not only improve the accuracy of subsequent classification, but also significantly reduce the computational complexity.

[0053] S2, through the relationship between samples and clusters, members of each financial data view are fused to obtain the fusion results within each view, the fusion results within each view are integrated into a binary partition matrix, and then the binary partition matrix is ​​integrated to generate a sample similarity matrix.

[0054] Specifically, the S2 includes:

[0055] In each financial view, the relationship between samples and clusters is applied to generate a binary partition matrix S v ∈{0,1} n×c ; Where n represents the number of financial view samples, and c represents the generated binary partition cluster;

[0056] Perform b binary partitions on the binary partition matrix and integrate them to obtain the integrated binary partition matrix Among them, c i The clusters generated for each binary partition;

[0057] The integrated binary partition matrix is ​​fused into a similarity matrix. The calculation formula is as follows:

[0058]

[0059] Among them, Q vRepresents the obtained similarity matrix.

[0060] S3, performs tensor constraints on the sample similarity matrix to construct a tensor-constrained global co-correlation matrix.

[0061] Specifically, the sample similarity matrix is ​​tensor-constrained, and the calculation formula is as follows:

[0062]

[0063] stF={Q 1 ,Q 2 ,...,Q v}

[0064] Where F represents the tensor matrix; V represents the number of financial views; || || * Represents a tensor constraint; represents the minimum value of the tensor matrix F; st represents the constraint condition;

[0065] The tensor constraint global co-correlation matrix is ​​constructed through tensor constraints. The calculation formula is as follows:

[0066]

[0067] Where β represents the learning rate for hyperparameter adjustment Q.

[0068] S4, through self-paced learning, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint-guided financial data feature selection model, thereby constructing a feature selection objective function based on self-paced tensor constraints.

[0069] Specifically, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints. The calculation formula is as follows:

[0070]

[0071] Among them, S v represents the self-paced learning coefficient, S v ∈R n×n , used to control the amount of financial data in the next iteration; represents the selection vector, z v The coefficient used to control the quantity of financial data for the next iteration.

[0072] Through the self-learning coefficient S v and the quantitative coefficient z of financial data vDynamically adjust the number of samples and features selected for each perspective to ensure that the key data points that best represent the original message are retained in the data fusion. v =0.5, then S v Can automatically adjust according to z v During training, we continuously evaluate which data samples are most representative, and give these samples a weight of 1, thereby determining the information retained by the current perspective.

[0073] S5, combining the data in each financial data view through the feature selection objective function, sorting the combined financial multi-source data in descending order, and selecting the top m features.

[0074] Specifically, the multi-source feature selection method and system for financial data based on self-paced tensor learning of the present invention are operated based on the Python version of Windows11 3.0.13. By introducing prior knowledge into the financial data feature selection model through pseudo-constraints, we guide feature selection and obtain a set of key features for each financial view. We then perform membership fusion on each financial data view based on the relationship between samples and clusters, thereby obtaining a sample similarity matrix and applying tensor constraints to it, effectively reducing model complexity. Furthermore, the experimental datasets used in this example consist of an image database (Caltech101-7) widely used for object recognition tasks and a sports news dataset (BBC Sport). Both datasets are similar to financial data and can well reflect the performance of this method. For comparison, we selected LS (Laplacian score for feature selection), ASVW (adaptive similarity and view weights for multi-view unsupervised feature selection), CRV-DGL (cross-view local structure-preserving diversity and consistency learning for multi-view unsupervised feature selection), NSGL (multi-view feature selection based on non-negative structure graph learning), TLR (multi-view unsupervised feature selection based on tensor low-rank minimization), CCSFS (consensus clustering structure-guided multi-view unsupervised feature selection), and PTFS (multi-view unsupervised feature selection based on partitioned hierarchical tensor learning) as comparison methods. In the performance evaluation, accuracy (ACC) is used as the metric, accuracy∈[0,1], and the higher its value, the higher the accuracy.

[0075] Table 1 Comparative analysis of accuracy measurements;

[0076]

[0077] According to the results in Table 1, the method of the present invention ranks first in accuracy in both image and news datasets, indicating that it has excellent performance and stability in the multi-view feature selection task.

[0078] In general, the core of this application is to achieve effective analysis and feature extraction of financial data through a series of steps to improve the quality of decision support. For each financial multi-source data view, prior knowledge is introduced into the model through pseudo-constraints to guide feature selection; then the distance between financial multi-source data is measured by L2,p norm, so as to perform regularized training, and the partitioning results of each financial data view are obtained by binary partitioning, and the binary partitioning results in each view are integrated into a binary partitioning matrix; thus, a fusion is performed to generate a sample similarity matrix; then, based on the sample similarity matrix, we impose tensor constraints on it, so as to better explore the information within each view, so as to better construct a tensor-constrained global co-correlation matrix; then, based on the tensor-constrained global co-correlation matrix, we integrate the pseudo-constraint-guided feature selection model through self-paced learning, so as to guide the model to screen features. This constructs a feature selection objective function based on self-paced tensor constraints. Finally, the government view data is sorted and selected according to the constructed objective function to ensure that the most representative features are ultimately selected. Pseudo-constraints are used to guide feature learning to avoid interference from noise. The tensor constraint mechanism is used to obtain key information of the similarity matrix of each view to optimize the construction of the global consensus matrix to accurately capture key financial information, and pseudo-constraints are introduced into the tensor-constrained global consensus matrix to guide feature selection. Finally, features are screened through the self-paced learning mechanism to guide the feature selection process, thereby ensuring the robustness and efficiency of the model when processing multi-view government data. This forms a feature selection objective function guided by the self-paced tensor matrix and achieves efficient and accurate selection of financial view data. Overall, these technical effects work together to improve the accuracy, efficiency and generalization ability of feature selection in financial data analysis.

[0079] like Figure 2 As shown, this embodiment also discloses a multi-source feature selection system for financial data based on self-paced tensor learning, including:

[0080] A multi-feature selection model construction module 21 is configured to, for each financial multi-source data view, introduce prior knowledge into the financial data feature selection model through pseudo-constraints to guide feature selection, thereby obtaining a key feature set within each financial view; measure the distance between the financial multi-source data in the key feature set within each financial view through the L2,p norm, and perform regularized training on the financial data feature selection model based on the distance between the financial multi-source data to obtain a pseudo-constraint-guided financial data feature selection model;

[0081] A similarity matrix construction module 22 is used to perform member fusion on each financial data view based on the relationship between samples and clusters, obtain the fusion results within each view, integrate the fusion results within each view into a binary partition matrix, and then integrate the binary partition matrices to generate a sample similarity matrix;

[0082] A tensor constraint module 23 is used to construct a tensor constraint global co-correlation matrix by performing tensor constraints on the sample similarity matrix;

[0083] a feature selection objective function construction module 24 for integrating the tensor-constrained global co-correlation matrix into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing a feature selection objective function based on the self-paced tensor constraint;

[0084] The feature selection module 25 is used to fuse the data in each financial data view through the feature selection objective function, sort the fused financial multi-source data in descending order, and select the top m features.

[0085] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. A multi-source feature selection method for financial data based on self-paced tensor learning, characterized by: include: S1, for each financial multi-source data view, introduce prior knowledge into the financial data feature selection model through pseudo constraints to guide feature selection and obtain the key feature set within each financial view; The distance between the financial multi-source data in the key feature set of each financial view is measured by the L2,p norm. The financial data feature selection model is regularized and trained based on the distance between the financial multi-source data to obtain a pseudo-constraint guided financial data feature selection model. S2, fuses the members of each financial data view based on the relationship between samples and clusters, obtains the fusion results within each view, integrates the fusion results within each view into a binary partition matrix, and then integrates the binary partition matrix to generate a sample similarity matrix; S3, performing tensor constraints on the sample similarity matrix to construct a tensor-constrained global co-correlation matrix; S4, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint guided financial data feature selection model by self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints; S5, fuse the data in each financial data view through the feature selection objective function, sort the fused financial multi-source data in descending order, and select the top m features.

2. The method for multi-source feature selection of financial data based on self-paced tensor learning according to claim 1, characterized in that: In S1, the financial data feature selection model is regularized and trained based on the distance between financial multi-source data. The calculation formula is as follows: in, X v represents the data of each financial view, n represents the number of financial view samples, d v Represents the sample dimension of each financial view; W v represents the feature selection matrix, Used to learn the relationship between data point features and clusters; Q represents the pseudo constraint matrix, Q∈R n×n , k is the projection dimension; α is a hyperparameter used to control the trade-off of the L2,p regularization term; ‖W v ‖ 2,p Represents the feature selection matrix W v Carry out L 2,p norm; Represents the feature selection matrix W v The minimum value of .

3. The method for multi-source feature selection of financial data based on self-paced tensor learning according to claim 1, characterized in that: Said S2 specifically includes: In each financial view, the relationship between samples and clusters is applied to generate a binary partition matrix S v ∈{0,1} n×c ; Where n represents the number of financial view samples, and c represents the generated binary partition cluster; Perform b binary partitions on the binary partition matrix and integrate them to obtain the integrated binary partition matrix Among them, c i The clusters generated for each binary partition; The integrated binary partition matrix is ​​fused into a similarity matrix. The calculation formula is as follows: Among them, Q v Represents the obtained similarity matrix.

4. The method for multi-source feature selection of financial data based on self-paced tensor learning according to claim 3, characterized in that: In S3, the sample similarity matrix is ​​tensor-constrained and the calculation formula is as follows: s.t.F={Q 1 ,Q 2 ,...,Q v } Where F represents the tensor matrix; V represents the number of financial views; |||| * Represents a tensor constraint; represents the minimum value of the tensor matrix F; st represents the constraint condition; The tensor constraint global co-correlation matrix is ​​constructed through tensor constraints. The calculation formula is as follows: Where β represents the learning rate for hyperparameter adjustment Q.

5. The method for multi-source feature selection of financial data based on self-paced tensor learning according to claim 4, characterized in that: In S4, the tensor-constrained global co-correlation matrix is ​​integrated into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing the feature selection objective function based on self-paced tensor constraints. The calculation formula is as follows: Among them, S v represents the self-paced learning coefficient, S v ∈R n×n , used to control the amount of financial data in the next iteration; represents the selection vector, z v The coefficient used to control the quantity of financial data for the next iteration.

6. A multi-source feature selection system for financial data based on self-paced tensor learning, characterized by: include: A multi-feature selection model building module is used to introduce prior knowledge into the financial data feature selection model through pseudo-constraints for each financial multi-source data view to guide feature selection and obtain the key feature set within each financial view; The distance between the financial multi-source data in the key feature set of each financial view is measured by the L2,p norm. The financial data feature selection model is regularized and trained based on the distance between the financial multi-source data to obtain a pseudo-constraint guided financial data feature selection model. The similarity matrix construction module is used to fuse the members of each financial data view based on the relationship between samples and clusters, obtain the fusion results within each view, integrate the fusion results within each view into a binary partition matrix, and then integrate the binary partition matrices to generate a sample similarity matrix; A tensor constraint module is used to construct a tensor-constrained global co-correlation matrix by performing tensor constraints on the sample similarity matrix; A feature selection objective function construction module is used to integrate the tensor-constrained global co-correlation matrix into the pseudo-constraint-guided financial data feature selection model through self-paced learning, thereby constructing a feature selection objective function based on self-paced tensor constraints; The feature selection module is used to fuse the data in each financial data view through the feature selection objective function, sort the fused financial multi-source data in descending order, and select the top m features.

Images