A multi-view clustering method based on global and local anchor learning
The multi-view clustering method, which learns from global and local anchor points, solves the problem of balancing local and global structures in existing technologies, achieves efficient multi-view clustering results, and is suitable for large-scale datasets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO UNIV
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-09
AI Technical Summary
Existing multi-view clustering methods cannot effectively balance the complementary information of the local structure of a specific view and the consistency of the global structure across views when utilizing multi-view information, resulting in limited clustering performance.
A multi-view clustering method based on global and local anchor point learning is adopted. By generating an objective function, anchor graph structure learning and global consistency structure learning are performed to construct a two-layer anchor graph optimization objective function. A multi-view clustering model with global and local anchor point learning is constructed using a mapping matrix, and the classification results are obtained by alternately updating the parameters.
It effectively captures local information of specific views and cross-view global consistency of multi-view data, solves the mismatch between local consensus and global structure, improves clustering performance, and demonstrates high efficiency and stability on large-scale datasets.
Smart Images
Figure CN122173959A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing technology, specifically relating to a multi-view clustering method based on global and local anchor point learning. Background Technology
[0002] Clustering is a learning task that divides unlabeled data into clusters based on their underlying structural similarity, resulting in high similarity among samples within the same cluster and significant differences among samples from different clusters. Clustering algorithms have been extensively studied over the past few decades. Traditional clustering algorithms are tailored for data with single-view features. However, in real-world scenarios, data is often collected from multiple different sources, encapsulating different complementary information, forming multi-view data. Traditional single-view clustering methods fail to fully utilize this multi-view information, leading to a waste of valuable information and suboptimal clustering performance. Therefore, it is necessary to develop multi-view clustering (MVC) methods to improve clustering performance and effectively integrate the complementary and consistent information embedded in multi-view features.
[0003] Over the past few decades, MVC has been extensively studied. A typical approach essentially follows the classic subspace clustering framework for single-view data. Generally, subspace clustering assumes that samples are self-representing, and that recovering the underlying cluster structure by seeking a low-dimensional representation of the data is effective. Following this framework, many MVC methods attempt to learn low-dimensional representations of multi-view data within individual views. While these methods have demonstrated good clustering performance in recovering the underlying cluster structure of multi-view data, they face the critical problem of high computational complexity due to their use of singular value decomposition (SVD), which severely limits their applicability in large-scale applications.
[0004] To address this inefficiency, numerous methods have been explored for effectively learning large-scale multi-view data. Among these methods, anchor-based approaches are a typical example. Essentially, they follow the key idea of subspace clustering, using a small number of representative anchors to construct a low-dimensional representation of the data. Notably, compared to subspace clustering, which uses all samples, anchor-based representations significantly reduce dimensionality, greatly reducing the complexity of the learning algorithm and making it suitable for large-scale data scenarios. Existing anchor-based multi-view clustering methods mainly fall into two categories: fixed-anchor methods and adaptive methods. Fixed-anchor methods suffer from the initialization sensitivity and insufficient adaptability to complex data distributions inherent in methods like K-Means. Furthermore, fixed anchors cannot be dynamically adjusted during subsequent learning, negatively impacting the capture of fine-grained data structures and evolving cross-view correlations. The other approach attempts to learn adaptive anchors instead of using fixed anchors throughout the learning process. With adaptive anchors, the MVC method integrates anchor optimization into a unified learning framework for representation and clustering, enabling dynamic adjustment of anchor positions and representativeness to adapt to the underlying data distribution. This not only mitigates the adverse effects of poor initial anchor point selection but also enhances the alignment between anchor points and the global multi-view structure, thereby improving the discriminability of clustering results. However, existing adaptive anchor point-based methods still lack an effective mechanism to adequately balance the complementary information of the local structure of a specific view with the consistency information of the global structure across views, which may limit further improvements in clustering performance.
[0005] Existing techniques, particularly those based on subspace clustering, have been widely applied in multi-view clustering. To improve computational efficiency, many methods select a set of anchors from the data and then use these anchors to learn a low-dimensional representation of the samples through their linear combinations. However, existing techniques, which use K-means to select initial anchors, suffer from the problem of clustering effectiveness being sensitive to anchor points and uneven anchor distribution when selecting complex data. Furthermore, the anchors selected by the k-means method cannot be dynamically adjusted during subsequent learning, which negatively impacts the capture of fine-grained data structures and evolving cross-view correlations.
[0006] Existing technology two is a multi-view clustering method based on adaptive anchor point learning, which allows anchor points to be dynamically refined according to the inherent structure of the data and the clustering objective. However, existing technology two, which uses adaptive anchor points, lacks an effective mechanism to fully balance the complementary information of the local structure of a specific view and the consistency information of the global structure across views, which may limit further improvement in clustering performance. Summary of the Invention
[0007] To address the above problems, this invention proposes a multi-view clustering method based on global and local anchor point learning.
[0008] The technical solution of this invention is: a multi-view clustering method based on global and local anchor point learning, comprising the following steps:
[0009] S1. Generate objective functions for all views;
[0010] S2. Based on the objective function, perform anchor graph structure learning and global consistency structure learning to construct a two-layer anchor graph optimization objective function;
[0011] S3. Optimize the objective function based on a two-layer anchor graph and construct a multi-view clustering model for global and local anchor point learning using a mapping matrix;
[0012] S4. Solve the multi-view clustering model for global and local anchor point learning to obtain the classification results.
[0013] Furthermore, in S1, the expression for the objective function is:
[0014] ;
[0015] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. Represents the transpose of a matrix. Indicates the first equilibrium parameter. Represents an identity matrix of appropriate size.
[0016] Furthermore, in S2, the expression for the objective function of the two-layer anchor graph optimization is:
[0017] ;
[0018] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. Represents an identity matrix of appropriate size.
[0019] Furthermore, in S3, the expression for the multi-view clustering model of global and local anchor point learning is:
[0020] ;
[0021] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. This represents the third equilibrium parameter. Represents the orthogonal projection matrix. Represents an identity matrix of appropriate size.
[0022] Furthermore, S4 includes the following sub-steps:
[0023] S41. Initialize parameters;
[0024] S42. Alternately update the initialized parameters;
[0025] S43. Repeat S42, and when the convergence condition is met or the maximum number of iterations is reached, determine the cluster labels and output the classification results.
[0026] Furthermore, in S41, the initialization parameters include the anchor point matrix of each view, the local cross-view consensus affinity matrix, the global cross-view anchor graph matrix, and the orthogonal projection matrix.
[0027] Furthermore, in S43, singular value decomposition is performed on the local cross-view consensus affinity matrix to obtain the spectral embedding. The K-means algorithm is then used to process the spectral embedding to determine the clustering labels and output the classification results.
[0028] The beneficial effects of this invention are:
[0029] (1) This invention learns the local anchor points of a specific view under the guidance of cross-view global anchor point selection, thereby effectively capturing the local information of a specific view and cross-view global consistency of multi-view data.
[0030] (2) This invention encodes a global shared structure for multi-view data and provides cross-view... Figure 1 Consistency provides a unified reference, fundamentally solving the mismatch between local consensus and global structure.
[0031] (3) This invention supports specific structures in local views and global cross-views. Figure 1A coherent two-layer joint optimization is employed. Through an orthogonal mapping matrix, global structure guidance is precisely transmitted to the learning of local consensus anchor graphs. This strategy avoids the conflict between local structure and global multi-view. Figure 1 This approach addresses structural deviations while preserving complementary local information captured by specific view anchors for each view, breaking the long-standing trade-off between local fitting and global alignment in existing methods.
[0032] (4) This invention provides a unified global criterion for multi-view clustering. Theoretical complexity analysis shows that GL4-MVC is scalable in terms of sample size, which makes it highly efficient and suitable for large-scale datasets. Numerous experimental results confirm the effectiveness and efficiency of GL4-MVC and verify its superiority in multi-view clustering. Attached Figure Description
[0033] Figure 1 A flowchart of a multi-view clustering method for learning global and local anchor points;
[0034] Figure 2 A t-SNE visualization of GL4-MVC on the NGs dataset;
[0035] Figure 3 A comparison chart showing the clustering performance of GL4-MVC with eight state-of-the-art baseline methods on six datasets;
[0036] Figure 4 A comparison graph showing the runtime of GL4-MVC across all datasets and different anchoring methods;
[0037] Figure 5 This is a graph showing the parameter sensitivity of GL4-MVC on Mfeat;
[0038] Figure 6 A comparison chart of the performance of GL4-MVC and ablation variants across all datasets;
[0039] Figure 7 The graph shows the convergence curves of the objective function of GL4-MVC on NGs and Mfeat. Detailed Implementation
[0040] The embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0041] like Figure 1 As shown, this invention provides a multi-view clustering method based on global and local anchor point learning, including the following steps:
[0042] S1. Generate objective functions for all views;
[0043] S2. Based on the objective function, perform anchor graph structure learning and global consistency structure learning to construct a two-layer anchor graph optimization objective function;
[0044] S3. Optimize the objective function based on a two-layer anchor graph and construct a multi-view clustering model for global and local anchor point learning using a mapping matrix;
[0045] S4. Solve the multi-view clustering model for global and local anchor point learning to obtain the classification results.
[0046] Fixed anchor point methods lack adaptability and global representativeness, while adaptive anchor point methods often fail to effectively balance the local complementarity of specific views with global consistency across views. To bridge this gap, this invention proposes a two-layer anchor graph optimization framework that learns view-specific local anchor points under the guidance of cross-view global anchor points within a unified model. Specifically, the proposed framework combines two complementary anchor point learning mechanisms: it first selects global anchor points across views and then guides the learning of view-specific local anchor points. Therefore, cross-view anchor points can well preserve the global structure of the data, while view-specific anchor points capture the local structure of the data.
[0047] Anchor-based MVC methods, such as LMVSC and FPMVS, offer high computational efficiency and clustering accuracy, but they struggle to balance global structural consistency with view-specific complementary information. Specifically, static anchors in methods like LMVSC cannot adapt to global data distribution, while FPMVS overemphasizes cross-view anchor consistency while neglecting view-specific local features. To address this trade-off, this invention proposes a two-layer anchor guidance framework that integrates cross-view global consistency with view-specific complementary information, achieving synergistic optimization of local adaptation and global alignment.
[0048] In this embodiment of the invention, in S1, an adaptive optimization strategy is employed to learn view-specific anchor points to ensure sufficient capture of view-specific local complementary information. This strategy is used to generate view-specific anchor points and further enhance the consistency among the learned anchor point graphs, thereby obtaining the objective function.
[0049] The expression for the objective function is:
[0050] ;
[0051] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. Represents the transpose of a matrix. Indicates the first equilibrium parameter. Represents an identity matrix of appropriate size.
[0052] In this embodiment of the invention, in S2, in order to solve the problem that locally consistent anchor graphs may deviate from the global data structure, resulting in suboptimal clustering performance, the present invention proposes a two-layer anchor graph optimization framework, which jointly realizes the learning of local anchor graph structure and global consistency structure in a unified model.
[0053] The expression for the objective function of the two-layer anchor graph optimization is:
[0054] ;
[0055] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data. Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. Represents an identity matrix of appropriate size.
[0056] Indicates a series of multi-view features. This represents the sum of the dimensions of all views. This indicates that the k-means method is used in... The globally consistent anchor matrix selected above. This represents the number of global anchor points, where This represents the global cross-view anchor graph matrix.
[0057] In this embodiment of the invention, in S3, the global and local terms in the preceding paradigm are separate and lack the necessary connection. A mapping matrix is further introduced into the equation. This bridges the gap between local learning and global guidance, thus forming the multi-view clustering model of global and local anchor learning of the present invention.
[0058] The expression for the multi-view clustering model that learns global and local anchor points is:
[0059] ;
[0060] in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. This represents the third equilibrium parameter. Represents the orthogonal projection matrix. Represents an identity matrix of appropriate size.
[0061] It is worth noting that cross-view global anchor point selection plays a crucial role in the model from several perspectives. First, it overcomes the inherent shortcomings of existing anchor-based MVC methods in global structure modeling. Unlike specific view-local anchor points that only capture a single view's local portion, this invention constructs a globally shared structure for multi-view data and provides local view anchor points for global structure modeling. Figure 1 Consistency provides a unified reference, fundamentally resolving the mismatch between local consensus and global structure. Secondly, it supports specific structures in local views and global cross-view functionality. Figure 1 A coherent two-layer joint optimization is employed. Through an orthogonal mapping matrix, global structure guidance is precisely transmitted to the learning of local consensus anchor graphs. This strategy avoids the conflict between local structure and global multi-view. Figure 1 This approach avoids structural deviations while preserving complementary local information captured by specific view anchors for each view, breaking the long-standing trade-off between local fitting and global alignment in existing methods. Third, it provides a unified global criterion for multi-view clustering. It ensures that all views converge to the same intrinsic structure during optimization, rather than generating independent or potentially conflicting clustering criteria based on their respective local anchors. This not only enhances the cross-view consistency of clustering results... Figure 1 This improves consistency and enhances the model's robustness to noise and specific view variations, laying a solid foundation for high-quality multi-view clustering.
[0062] In this embodiment of the invention, S4 includes the following sub-steps:
[0063] S41. Initialize parameters;
[0064] S42. Alternately update the initialized parameters;
[0065] S43. Repeat S42, and when the convergence condition is met or the maximum number of iterations is reached, determine the cluster labels and output the classification results.
[0066] In this embodiment of the invention, in S41, the initialization parameters include the anchor point matrix of each view, the local cross-view consensus affinity matrix, the global cross-view anchor map matrix, and the orthogonal projection matrix.
[0067] In this embodiment of the invention, in S43, singular value decomposition is performed on the local cross-view consensus affinity matrix to obtain the spectral embedding. The K-means algorithm is then used to process the spectral embedding to determine the clustering label and output the classification result.
[0068] In an embodiment of the present invention, Figure 2 This is a t-SNE visualization of GL4-MVC on the NGs dataset. The figure shows the first and second views of the NGs dataset, as well as the spectral embedding matrix. The t-SNE visualization clearly shows that before the clustering of this invention, the samples in NGs are highly overlapping in the original space, while after the clustering of this invention, the samples belonging to different clusters can be clearly separated in the low-dimensional space.
[0069] Figure 3 This figure compares the clustering performance of GL4-MVC with eight state-of-the-art baseline methods on six datasets. The figure primarily compares four standard clustering metrics: accuracy (ACC), normalized mutual information (NMI), purity, and F-score. As can be seen from the figure, the proposed method achieves clustering performance comparable to or better than other methods across all datasets, demonstrating its superiority.
[0070] Figure 4 This figure compares the runtime of GL4-MVC with different anchoring methods across all datasets. As shown, this invention demonstrates efficiency advantages on both large-scale and small-to-medium-scale datasets.
[0071] Figure 5 This is a demonstration of the parameter sensitivity of GL4-MVC on Mfeat. As shown in the figure, it illustrates the impact of different values for the three balancing parameters on the clustering results. It can be seen that GL4-MVC can achieve good clustering results over a wide range of parameters, demonstrating the stability and potential applicability of GL4-MVC in practical applications.
[0072] Figure 6 This is a performance comparison of GL4-MVC and its ablation variant across all datasets. The ablation variant ignores the global anchor graph guidance component. As shown in the figure, GL4-MVC achieves better clustering results than the ablation variant on all datasets. GL4-MVC learns the underlying structure of the data more effectively under the guidance of global anchor points, confirming the importance of global anchor guidance in the GL4-MVC model.
[0073] Figure 7 This figure shows the convergence curves of the objective function of GL4-MVC on NGs and Mfeat. As can be clearly seen from the figure, GL4-MVC converges to the objective function value within approximately 20 iterations. These observations empirically demonstrate the efficient convergence behavior of GL4-MVC and its practical applicability in real-world scenarios.
[0074] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A multi-view clustering method based on global and local anchor point learning, characterized in that, Includes the following steps: S1. Generate objective functions for all views; S2. Based on the objective function, perform anchor graph structure learning and global consistency structure learning to construct a two-layer anchor graph optimization objective function; S3. Optimize the objective function based on a two-layer anchor graph and construct a multi-view clustering model for global and local anchor point learning using a mapping matrix; S4. Solve the multi-view clustering model for global and local anchor point learning to obtain the classification results.
2. The multi-view clustering method based on global and local anchor point learning according to claim 1, characterized in that, In S1, the expression for the objective function is: ； in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. Represents the transpose of a matrix. Indicates the first equilibrium parameter. Represents an identity matrix of appropriate size.
3. The multi-view clustering method based on global and local anchor point learning according to claim 1, characterized in that, In S2, the expression for the objective function of the two-layer anchor graph optimization is: ； in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data. Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. Represents an identity matrix of appropriate size.
4. The multi-view clustering method based on global and local anchor point learning according to claim 1, characterized in that, In S3, the expression for the multi-view clustering model of global and local anchor point learning is: ； in, This represents the anchor point matrix of the v-th learned view, where V represents the number of views. This represents the feature data of the v-th view. Represents the local cross-view consensus affinity matrix. This represents the global cross-view anchor graph matrix. Represents the concatenated feature data of multiple views. Represents global anchor data. Represents the transpose of a matrix. Indicates the first equilibrium parameter. This represents the second equilibrium parameter. This represents the third equilibrium parameter. Represents the orthogonal projection matrix. Represents an identity matrix of appropriate size.
5. The multi-view clustering method based on global and local anchor point learning according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Initialize parameters; S42. Alternately update the initialized parameters; S43. Repeat S42, and when the convergence condition is met or the maximum number of iterations is reached, determine the cluster labels and output the classification results.
6. The multi-view clustering method based on global and local anchor point learning according to claim 5, characterized in that, In step S41, the initialization parameters include the anchor point matrix of each view, the local cross-view consensus affinity matrix, the global cross-view anchor map matrix, and the orthogonal projection matrix.
7. The multi-view clustering method based on global and local anchor point learning according to claim 5, characterized in that, In step S43, singular value decomposition is performed on the local cross-view consensus affinity matrix to obtain the spectral embedding. The K-means algorithm is then used to process the spectral embedding to determine the clustering labels and output the classification results.