Anchor point guide graph clustering network

Abstract

The embodiment of the invention discloses an anchor point guide graph clustering network. The method comprises the following steps: (1) generating a pseudo tag with global distribution characteristics by adopting a characteristic analysis method; (2) learning a mapping relation between sample features and pseudo labels through a model; and (3) in the embedding space, gradually aggregating the samples with the same labels, and dynamically correcting wrong and false labels by the model in the learning process, so that the wrong and false labels are re-aggregated into clusters with similar semantics. According to the method, an anchor guide learning strategy is designed, an anchor point is automatically selected from a high-confidence sample farthest from a distribution boundary, and a k-means + + algorithm is utilized to generate a pseudo tag in a diffusion feature space, so that learning of discriminative characterization is guided. Furthermore, the AGCN integrates an anchor point expansion strategy and a generality enhancement mechanism, and the discrimination of clustering and the precision of a recommendation task are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of multi-view data processing technology, and in particular to an anchor-guided graph clustering network. Background Technology

[0002] In the field of personalized service recommendation, clustering is a key technique. By dividing users with similar behavioral characteristics into different groups, more accurate recommendations can be achieved. Depth graph clustering aims to uncover potential structural patterns in user behavior graphs, thereby effectively grouping users based on behavioral similarity to improve the efficiency of personalized service recommendations.

[0003] Existing graph clustering methods can be broadly categorized into two types: contrastive and generative. Contrastive graph clustering relies on cross-view sample comparison learning to obtain discriminative representations, capturing some clustering information. Generative graph clustering methods typically employ an encoder-decoder structure, learning high-quality user representations within a self-supervised framework. This not only reduces the risk of information loss but also generates more stable recommendation results.

[0004] However, existing comparative graph clustering methods may lose key information during feature dimensionality reduction, leading to representation collapse and thus affecting recommendation performance. Generative graph clustering methods rely heavily on pre-training, and their training process is sensitive to initialization and lacks robustness. Furthermore, they often require repeated high-cost pre-training when adjusting parameters or transferring models. Summary of the Invention

[0005] To address the existing technical problems, this application provides an anchor-guided graph clustering network. The technical solution is as follows: Firstly, an anchor-guided graph clustering network is provided, including: (1) Use feature analysis methods to generate pseudo-labels with global distribution characteristics; (2) The model learns the mapping relationship between sample features and pseudo-labels; (3) In the embedding space, samples with the same label gradually aggregate. The model dynamically corrects erroneous pseudo-labels during the learning process, causing them to re-aggregate into semantically similar clusters.

[0006] Furthermore, the step of generating pseudo-labels with globally distributed features using feature analysis methods includes: The diffusion attribute matrix X of the target undirected graph is processed using k-means++. p Clustering is performed to obtain the pseudo-label one-hot encoding matrix Y. k ; Extracting local important features from the diffusion attribute matrix Xp to construct the enhanced attribute matrix X e X e With symmetric normalized adjacency matrix The data are input together into the model for training, and the one-hot encoding matrix Y is used. k Guided learning of the model; After a quick and coarse training of the model, the easily separable samples that are farthest from the dividing surface and have the largest differences are selected as the anchor samples V for the next stage. a .

[0007] Furthermore, the step of learning the mapping network relationship between sample features and pseudo-labels through the model includes: anchor sample V a One-hot encoding of Y with pseudo-labels of anchor samples a The mapping relationship between them serves as a constraint on the model's results. Conduct training: Generate the target distribution P; Minimize the KL divergence loss between the distributions of Y and P; A threshold update mechanism is used to gradually expand the anchor sample set.

[0008] Furthermore, in the embedding space, samples with the same label are gradually aggregated, and the model dynamically corrects erroneous pseudo-labels during the learning process, causing them to re-aggregate into semantically similar clusters, including: Peel off samples that are closer to the segmentation surface; Using the diffusion attribute matrix X p C cluster centers are generated by combining the anchor samples of the corresponding clusters; Cluster centers C obtained by taking the mean of anchor samples within the same cluster p Calculate Vr and C for each cluster center. p Spatial distance Dis; Find the minimum distance calculated between the sample and the cluster center, and then assign each sample to the cluster with the minimum distance. Complete the extraction of important clustering information from the model.

[0009] Furthermore, this also includes loss functions and complexity analysis.

[0010] Furthermore, the loss function analysis includes: Integrating steps (1), (2), and (3) into the same framework, the total loss function is: in, = U The set of pseudo-labels for all hard samples and anchor samples. = + , This indicates the moment when the anchor sample was selected.

[0011] Further loss complexity analysis includes: The time complexity of the convolutional neural network layer is... ,exist Previous In addition to the model's own operation, the computational complexity of the cross-entropy function is also O(n). ,exist Then, KL divergence was added as the loss function, with a time complexity of O(n log n). The time complexity of anchor sample selection and updating is O(n log n). The time complexity for processing difficult samples is ,because Therefore, the time complexity of the cross-entropy function at this point is still O(n log n). ,because The intervals between training cycles are very small compared to the entire training cycle and can be ignored. Therefore, the final time complexity of the entire model is O(n log n). .

[0012] The beneficial effects of the technical solution provided in this application are as follows: This application proposes an Anchor-Guided Graph Clustering Network (AGCN) that does not require pre-training. This method automatically selects anchor points from high-confidence samples furthest from the distribution boundary by designing an anchor-guided learning strategy, and generates pseudo-labels in the diffuse feature space using the k-means++ algorithm, thereby guiding the learning of discriminative representations. Furthermore, AGCN integrates an anchor point expansion strategy and a commonality enhancement mechanism, significantly improving the discriminativeness of clustering and the accuracy of recommendation tasks. Theoretical analysis and experimental results on multiple real datasets show that this method outperforms existing state-of-the-art methods in both clustering accuracy and stability. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a core framework diagram of AGCN provided in the embodiments of this application; Figure 2 This is a diagram illustrating the anchor sample selection process provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the selection principle of anchor samples provided in an embodiment of this application; Figure 4This is a graph showing the impact of different pre-training periods on the model, as provided in the embodiments of this application. Figure 5 This is a graph showing the impact of different pre-training strategies provided in the embodiments of this application on the model; Figure 6 This is a visualization of AGCN on Cora provided in the embodiments of this application; Figure 7 This is a visualization of AGCN on Citeseer provided in the embodiments of this application; Figure 8 This is a visualization of AGCN on Amap provided in the embodiments of this application; Figure 9 This is a graph showing the convergence analysis and ablation experiment results of AGCN on Amac and Amap provided in the embodiments of this application; Figure 10 This is a graph showing the convergence analysis and ablation experiment results of AGCN on Citeseer and Cora provided in the embodiments of this application; Figure 11 The figure shows the convergence analysis and ablation experiment results of AGCN on Coauthor_cs and Coauthor_phy provided in the embodiments of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0016] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0017] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0018] In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] Deep attribute graph clustering combines deep learning with attribute graph modeling techniques, comprehensively considering the attribute features and structural relationships of nodes to achieve more accurate node grouping. Typical works, such as SDCN, combine autoencoders with graph neural networks (GNNs) to introduce graph structure information and enhance clustering performance. Subsequently, DFCN employs a dynamic fusion mechanism to improve the information interaction efficiency between autoencoders (AEs) and graph autoencoders (GAEs). In contrast, MvDSCN introduces a multi-view deep subspace joint network to achieve end-to-end fusion of multi-view feature learning and subspace clustering, thereby improving the expressive and fusion capabilities of multi-view clustering. IHGAT constructs a fused heterogeneous graph and introduces an attention mechanism to effectively mine complementary information between different samples and modalities, thus achieving more robust incomplete multimodal clustering under high missing rates.

[0020] While generative graph clustering methods offer advantages in accuracy, they typically rely on complex pre-training processes. This dependence not only increases model training complexity but also limits the model's robustness and practical application capabilities. To address this, this application proposes a novel deep graph clustering method—Anchor-Guided Graph Clustering Network (AGCN). This method replaces the traditional pre-training process by selecting high-confidence anchor samples, guiding the model to directly learn key clustering information from real-world data distributions, thereby significantly improving the model's generalization and adaptability in complex environments.

[0021] Clustering algorithms have wide applications in service computing, improving the classification, retrieval, and recommendation performance of Web services by mining their semantic features and combinatorial information. For example, a clustering framework for macro-flow management has been proposed to cluster network flows based on traffic pattern similarity, achieving an efficient balance between reducing the number of flow table entries and maintaining routing performance. DeepWSC proposes a deep service clustering framework based on heuristic policies, achieving more efficient service clustering results by fusing deep semantic features of service descriptions with composability information in call relationships. Cui et al. proposed a latency-optimized container scheduling algorithm that combines policy gradient reinforcement learning with a self-attention mechanism, significantly reducing task latency by approximately 30% in edge computing clusters. Furthermore, Kim et al. designed an optimal worker node allocation method for multi-host heterogeneous distributed computing environments, transforming the problem into a linear programming solution. Multiple simulation results demonstrate that this method performs excellently in both task allocation efficiency and computational performance.

[0022] In many machine learning tasks, the quality of feature representation directly determines the upper limit of model performance. Positive sample alignment aims to improve the model's discriminative ability by aggregating similar samples in the feature space to make their representations more similar. Huang et al. first systematically analyzed this phenomenon and pointed out that when a sample is clustered into a specific category, its potential similar samples will also be attracted to that cluster, achieving semantic-level aggregation. On the other hand, pseudo-labeling, as a commonly used technique in semi-supervised learning, assigns temporary labels to unlabeled samples using the model's current prediction results to guide subsequent training, thereby enhancing the model's adaptability to data distribution. Iscen et al. (2019) proposed a pseudo-labeling method based on neighborhood graphs; Shi et al. (2018) further combined local neighborhood density to calculate the confidence score of unlabeled samples to improve the reliability of pseudo-labels.

[0023] Based on the two mechanisms mentioned above, this application achieves more accurate sample clustering results than traditional pseudo-labeling mechanisms by learning the sample distribution through the model. Specifically, a simple feature analysis method (such as k-means) is first used to generate pseudo-labels with global distribution features. Then, the model learns the mapping relationship between sample features and pseudo-labels. Here, the model is an AGCN model composed of two layers of convolutional neural networks. The samples can be images or text, such as the Cora and Citeseer datasets for academic papers, and the Amac and Amap datasets for product types. In the embedding space, samples with the same label will gradually aggregate, and the model can also dynamically correct some erroneous pseudo-labels during the learning process, causing them to re-aggregate into semantically similar clusters, as shown in "Model Distribution Learning" in Figure 2.

[0024] The overall architecture of the model in this application is as follows: Figure 1 As shown, the following will be combined with Figure 1 The model in this application is described in detail below. In the diagram, the green line represents the anchor sample selection stage, the blue line corresponds to the anchor-guided feature learning stage, and the purple line highlights the hard sample optimization mechanism in the third stage. The blue and purple paths together depict the iterative clustering and semantic alignment process of the model.

[0025] 1. Selection of distribution edge anchor samples The first stage of AGCN involves using k-means++ to diffuse the attribute matrix of the undirected graph. Clustering is performed to obtain the pseudo-label one-hot encoding matrix. At the same time, extract The local important features constitute the enhanced attribute matrix .Will and They are all input into the model for training, and then used... Guided learning is then applied to the model. After a rapid, coarse training process, easily separable samples that are furthest from and most dissimilar to the partitioned surface are selected as anchor samples for the next stage. .

[0026] 1.1 Obtaining Pseudo-Labels with Better Distribution Typically, the original adjacency matrix It can only represent first-order relationships between samples and is often sparse. However, these limitations can lead to the neglect of higher-order information during data analysis, thus affecting the accuracy and comprehensiveness of the analysis results. The second-order matrix is ​​obtained by multiplying the adjacency matrix by itself. It includes not only first-order relations but also second-order relations. Similarly, Includes samples All relationships within the order.

[0027] Adjacency matrix With attribute matrix Multiplication aggregates the attribute information of a central sample with that of its neighboring samples; this process is also called graph diffusion. This process allows the central sample to acquire the attribute information of its surrounding samples and incorporate this information into its own feature representation. Samples that can be assigned to the same cluster must be similar in attributes or have a strong structural correlation. Therefore, this application proposes the following definition: Within a sub-region centered on a certain central sample, most samples that are directly or indirectly related to that central sample should belong to the same cluster, and samples in the same cluster must have common characteristics in their representation.

[0028] Therefore, according to the above definition, using Gather When considering the properties of in-order neighbors, the resulting diffusion matrix is ​​obtained. It is a weighted attribute matrix, where features common to the central samples have larger weights. From the perspective of the high-dimensional space where the samples reside, through... right After diffusion, samples in the same cluster will cluster together, forming a distribution that is conducive to clustering, as shown in the following process. Figure 1 As shown in steps one and two, from the perspective of the attribute matrix, it forms a representation that is conducive to clustering. The calculation formula is as follows: Based on the above analysis Compare It has better, cluster-friendly representations. Therefore, k-means++ is used to... Clustering can yield a pseudo-label with a better distribution. : here, This indicates the pseudo-labels obtained by applying the k-means++ clustering algorithm to the sample set. Used to guide the next step of model training.

[0029] 1.2 Commonality Enhancement Mechanism To optimize the distribution of samples in their spatial domain, enabling the model to accurately distinguish similar samples and precisely delineate the higher-dimensional space in which the samples reside, a commonality enhancement mechanism was designed. Firstly, from... Extract locally important attributes to construct a locally important feature matrix. ,Right now: here, This indicates that the mean of the variable is being calculated. express medium sample The attributes, for elements, This represents the locally important feature matrix, meaning the features shared within this subspace. Then... and Perform linear superposition: here, It is a hyperparameter. It is incorporated into model training as a common enhancement attribute.

[0030] 1.3 Guided Model Training Through the above steps, a pseudo-label with a good distribution was obtained. Next, use Guided training of the model.

[0031] First, treat the model as a function and enhance common attributes. and As a variable input into the model: in, It is a graph convolution model. It is an activation function, such as ReLU. These are the network's learning parameters. The first layer. This is the clustering result of the model. Pseudo-labels are used. Directly analyze the clustering results of the model Provide guidance, that is: The final result obtained through model training The distribution should be greater than the initial The sample distribution is better. Next, we will base our work on... Select anchor samples based on their distribution .

[0032] 1.4 Anchor Sample Selection Samples that can be quickly identified and assigned to a cluster by the model after a few training cycles are called easily separable samples.

[0033] Using a small number of cycles After rapid training, each cluster in the distribution obtained by the model is sorted, and then the top clusters of each cluster are selected. One sample is used as the anchor sample. Because The clusters are very small, so there is no overlap between them. in, yes elements, It is the result of the model running. It is the first Anchor samples of each cluster, It is a function that sorts variables in descending order. It takes the first one in the list. A function for each sample. This is the anchor sample set. The greater the difference between two samples, the greater the difference should be in the model's output. From a spatial perspective, the model learns to assign different samples to different clusters, essentially dividing different samples into different regions in a high-dimensional space. The model can be viewed as a surface that divides this high-dimensional space. On this surface, the samples furthest apart should show significant differences after model computation. Samples near the surface have high similarity, and simply using well-distributed labels may not be sufficient to distinguish them. Therefore, it's only necessary to roughly divide the different regions using a small number of periods. Within these different regions, the most easily distinguishable samples that are furthest from the dividing surface and have the greatest differences are selected as anchor samples to guide model training in the next stage. Because these anchor samples have high differences, they are unlikely to be assigned to the same cluster. In summary, anchor samples obtained in this way have high reliability. The specific anchor sample selection process is as follows... Figure 2 As shown.

[0034] 2. Anchor Sample-Guided Model Training and Update In the anchor sample selection phase described above, all sample distributions are fed into the model for learning. However, the pseudo-labels may contain some erroneous distribution information. To prevent the erroneous information learned by the model from affecting subsequent training, the model parameters are reset before formal training begins.

[0035] 2.1 Anchor Sample-Guided Model Training Selected from the previous stage and The mapping relationship between them serves as a constraint on the model's results. Conduct training: in, This refers to the number of anchor samples. In this case, since the anchor samples represent only a small portion of the sample set, relying solely on this small subset for training could lead to overfitting. To mitigate this issue and provide more space for subsequent anchor sample updates, thus refining feature extraction, a target distribution is generated. in, For soft clustering frequencies, embedding is used. Production target distribution. Generating target distribution. In clustering, pseudo-labels are widely used as soft labels, utilizing partial sample pseudo-labels to guide the model and give it a certain degree of discriminative ability. Therefore, the generated target distribution can, to some extent, closely resemble the true data distribution. To make the distribution generated by the model as close as possible to the target distribution, therefore, minimizing... Distribution and The KL divergence loss between distributions helps the model perform clustering tasks better.

[0036] 2.2 Anchor Sample Update Because the selected anchor samples may contain non-critical clustering information, this information could affect the model's misclassification of samples. As illustrated in the example, an image recognition model might learn background features (such as grass) and fail to correctly identify cows on a beach. To address this issue, a threshold update mechanism is used to gradually expand the anchor sample set. here This represents normalizing the numerical values ​​of variables so that each value is between 0 and 1, and the sum of all values ​​is 1. It can be represented as a probability distribution. That is, the sample Belongs to the The probability of each cluster. For the threshold, It is the updated version The sample set of anchor samples in each cluster. This is achieved by using the updated anchor sample set. Training and guiding the model enhances its ability to extract important clustering information in detail, thereby improving the model's clustering performance.

[0037] 3. Handling of remaining difficult samples After the model converges in the later stages of training, most samples have entered the anchor sample set. However, a small number of samples, due to their similarity to two or more clusters, cannot be precisely assigned to a specific cluster. We can leverage the distribution details of the pseudo-labels of the difficult samples to optimize the clustering details of the remaining local samples, allowing samples with the same label to gradually aggregate and regroup into semantically similar clusters. From a spatial perspective, these samples are closer to the segmentation surface. Therefore, we first isolate these samples: in, To handle this portion of the samples, according to Definition 1, we start with the representations and consider the similarity between the difficult samples and each cluster. It is a weighted attribute matrix, and the local key attribute values ​​are relatively prominent, therefore it adopts... The characterization is more conducive to The samples were correctly assigned to their corresponding clusters. First, select to use... Generate anchor samples with the corresponding clusters respectively Cluster centers: in, For the process The result obtained after calculation Cluster centers of each cluster Indicates the first Anchor sample set The Anchor sample exist The corresponding attributes are as follows. Deep learning clustering uses statistical methods to assign different samples to different clusters. Therefore, samples that can be assigned to the same cluster share common characteristics in their representation. The cluster centers are obtained by taking the average of the anchor samples in the same cluster. The feature weight representing the commonality of samples in the same cluster is larger, while the weights of other features are smaller. Next, we calculate... Spatial distance from each cluster center : in, It is a sample The attributes and the first The distance between the centers of each cluster It is a sample The diffusion properties, It is a difficult sample set. In computation... and When calculating the variance between samples, the objective is to calculate the variance between samples. With cluster center Distance between And its distance is mainly related to The distance is related to the commonalities of the cluster; the greater the commonalities, the smaller the distance, and the smaller the differences. Therefore, we find the minimum distance calculated between the sample and the cluster center, and then assign each sample to the cluster with the smallest distance, i.e.: in, It is a matrix elements, For the sample The independent-hot encoding matrix of the pseudo-labels. Therefore, using Calculation and The cross-entropy function, i.e.: Finally, the model's extraction of important clustering information is completed, resulting in a final improvement.

[0038] 4. Loss Function and Complexity Analysis 4.1 Loss Function The model has three stages. The loss functions for the second and third stages can be written as follows: Integrating the three steps into a single framework, the overall loss function can be written as follows: in, = U The set of pseudo-labels for all hard samples and anchor samples, similarly. = + . This indicates the moment when the anchor sample is selected. The three steps of AGCN are integrated into a unified framework, as follows.

[0039] 4.2 Complexity Analysis The overall time complexity of the model is described as follows: The time complexity of the convolutional neural network layer is... .exist Previous In addition to the model's own operation, the computational complexity of the cross-entropy function is also O(n). .exist Then, KL divergence was added as the loss function, with a time complexity of O(n log n). The time complexity of anchor sample selection and updating is O(n log n). The time complexity for processing difficult samples is O(n log n). .because Therefore, the time complexity of the cross-entropy function at this point is still O(n log n). .because The intervals between training cycles are very small compared to the entire training cycle and can be ignored. Therefore, the final time complexity of the entire model is O(n log n). The symbols and definitions used in this article are shown in Table 1.

[0040] 5. Analysis 5.1 Anchor samples exhibit strong inter-class differences. Assume there are two inter-cluster anchor samples and Their differences are relatively small. The mapping relationship of the model can be represented as... Treat the model as a function ,have: Since vector calculations are involved, the sample is calculated first. unit vector Let the sample The angle between the coordinate system and the coordinate system is ,sample The included angle is Next, construct the adjustment matrix: It is the identity matrix, through and adjust the matrix to transform the vector Convert to vector : because and The difference is small, then when , hour, It can be simplified to: Will Substitute into function calculate The results showed that the sample and The anchor samples should be located near the segmentation surface, not furthest from it. This contradicts the assumption, hence the high dissimilarity of the anchor samples selected in section 1.4.

[0041] 5.2 By learning the distribution of samples through a model, a more accurate sample distribution can be obtained than that obtained through pseudo-labels.

[0042] First, the samples are labeled using data analysis methods (such as k-means++). Assume there are three samples: two with similar attributes and pseudo-labels... samples and and a pseudo tag with similar attributes. samples The model learns through the cross-entropy function. , and Its relationship with pseudo-tags: The loss function can be expanded to the following form: Due to positive sample alignment, samples with the same label and similar attributes... and Approaching naturally within the embedded space: Indicates proximity. Indicating reasoning, This indicates that they are close to each other. It is logical that the results of samples a and b are similar: Based on inference 1, samples with similar attributes , and The result should be similar: But from the loss function From the perspective of the sample Should be closer : At this time, the sample was discovered. Cluster affiliation conflicts occurred for the samples. It should be close to We should also get closer Their similarity contradicts that of pseudo-labels. According to formula (24), if near ,So It will become smaller, and This will increase, leading to an overall increase in the loss function, which violates the principle of minimizing the loss function. Therefore, during the overall training process, the samples... It will tend to get closer This meets the requirements of clustering—grouping similar samples into the same cluster.

[0043] When obtaining pseudo-labels, due to the limitations of data analysis, , and Such samples are prevalent in the sample set. Therefore, during model learning, the model tends to minimize the loss as much as possible, attracting similarity samples to the correct clusters. This is also the reason for selecting anchor samples in section 1.4. Based on the above proof, the specific principle of anchor sample selection can be vividly demonstrated, which can be used... Figure 3 express.

[0044] 5.3 Model Convergence Analysis The initially selected anchor sample set suffers from significant inter-cluster differences, which is detrimental to the model's ability to extract important clustering information, limiting its generalization capability and leading to a suboptimal solution. Before analyzing whether anchor sample updates can extract important clustering information, we propose a definition: in the ideal case, only a small amount of inter-cluster differences and intra-cluster similarities are needed when identifying samples.

[0045] For example, there are many characteristics in life that can be used to distinguish between men and women, but even with so many characteristics, it is still not completely accurate. However, the key characteristic that can accurately distinguish between men and women is to look at whether a person's 23rd pair of chromosomes is XX or XY; this characteristic is called key clustering information. Therefore, in the process of model learning, it is necessary to gradually reduce the differences between samples in different clusters in order to find the key clustering information that can distinguish samples.

[0046] Suppose there is a sample set The sample size is Therefore, the attributes between samples can be divided into different categories. and the same attributes The model is represented as The learning parameters are and , It is a row vector, which is added to the matrix during computation through vector diffusion.

[0047] (31) Will Rewrite it in matrix form and observe the characteristics of each row: (32) Through observation, it was found that Each row is identical; this special matrix can be labeled as follows: (33) here, and Both indicate that each row of the matrix has the same size, only the way they are expressed is different. Then formula (31) can be rewritten as: (34) here At this point, observation reveals that distinguishing different samples depends on attributes that are not entirely identical between the samples. Now, let's shift our focus from all the data to the attributes within a specific cluster. Suppose that the location of a cluster in the embedding space is... The sample size is Then you can Divided into samples within the same cluster with not all identical attributes and the same attribute The equation can then be written as: (35) because Since the values ​​in each row are the same, the formula can be rewritten as follows: (36) because and Since all rows of these matrices are identical, when the model finally converges... It should also be the same for every row.

[0048] Then from the entire sample set From this perspective, assuming a simple solution for the model, when the model converges... It should be equal to zero, but The result only followed In It changes with the changes. But if it occurs... When other than samples, The results will exhibit a degree of randomness, which is related to the model training strategy and the assumption that partial sample dependency has already formed. Therefore, the sample similarity principle in the theorem can be used to increase the number of anchor samples through a threshold update mechanism, thereby increasing the richness of samples learned through the above formula and continuously reducing the number of samples within the same cluster. The number of attributes reduces the dependence on samples and improves the model's ability to extract important clustering information.

[0049] 6. Experimental Design 6.1 Limitations of Current Pre-training Mechanisms Pre-training has become an indispensable step in the model clustering process. However, according to this study, the pre-training constraints of the encoder-decoder only involve the reconstruction of attributes or structures. This approach is insufficient to effectively constrain the model's pre-training and may lead to model collapse during the pre-training stage, thus affecting the final training results. To explore the impact of pre-training on model training, the training results of SDCN, DFCN, and R2FGC models on the Cora dataset are used as examples. The pre-training of existing models is usually completed within 100 epochs. The model was pre-trained for 50 epochs, and the pre-training parameters for each epoch were recorded to simulate the impact of pre-training from insufficient to sufficient. DFCN requires pre-training AE and GAE separately before pre-training the entire DFCN model. To observe the impact of pre-training on the model, the parameters of different AE and GAE pre-training were paired to pre-train the entire DFCN model. Specifically, the AE pre-training parameters of each epoch were trained together with the corresponding GAE pre-training parameters for 50 epochs to generate the DFCN pre-training parameters for that epoch. Since R2FGC does not have publicly available pre-training code, we used DFCN's three-stage pre-training mode and a unified pre-training mode to pre-train it, in order to observe the impact of different pre-training strategies on the performance. (Reference) Figure 4 and Figure 5 Observe the pre-training status of SDCN, DFCN, and R2FGC: SDCN consists of AE and GAE. According to the SDCN documentation, only the AE needs to be pre-trained. Observation Figure 4 The relationship between AE pre-training and formal model training was investigated, revealing that improvements in AE accuracy generally lead to positive changes in model accuracy. Sufficient pre-training helps improve model performance, while insufficient pre-training can lead to performance degradation or even model crashes. For example, using pre-training parameters from the first 15 epochs resulted in model crashes during formal training.

[0050] DFCN requires pre-training AE and GAE separately, followed by a final pre-training of the entire DFCN model. During this process, DFCN uses principal component analysis (PCA) to process the data, which causes the model to crash during AE pre-training. In contrast, GAE pre-training performs better. When the pre-training parameters of AE and GAE are applied together to DFCN, pre-training performance improves, and the changes in the pre-training curves of GAE and DFCN are generally consistent. However, due to the mismatch between the pre-training strategy and the actual training of DFCN, the final training effect is worse than the pre-training effect. Although the trends are consistent, the study found that DFCN pre-training has poor robustness, with the best results appearing in the 6th epoch. In actual model training, pre-training cannot adapt well to the real environment and usually requires more iterations. Therefore, this paper recognizes that the pre-training strategy must match the model training; otherwise, it may lead to poor model performance or even model crash (e.g., Figure 5 (Pre-training status of R2FGC in China).

[0051] contrast Figure 5 The comparison between three pre-training iterations and one pre-training iteration for R2FGC revealed significantly different impacts of different pre-training strategies on the model. Using the three-stage pre-training strategy with DFCN, the R2FGC model collapsed immediately, while the R2FGC model with one-stage pre-training performed slightly better than the DFCN strategy. Furthermore, the final training results were consistent with the pre-training trends. Table 3 shows that the accuracy of R2FGC reached 58.08%. This is because the pre-training epochs were set to 500, while the actual pre-training epochs are typically 50 to 200. This also reveals the vulnerability of encoder-decoder pre-training.

[0052] In summary, the research indicates that models based on encoder-decoder pre-training have the following problems: Simple encoder-decoder pre-training lacks effective constraints, which may lead to problems in the pre-training stage or even cause the model to crash.

[0053] To ensure model performance, pre-training must be consistent with model training. Loss of consistency can severely impact model performance and even lead to model crashes.

[0054] These issues make the model ill-equipped to handle the challenges of complex real-world environments. Furthermore, adjusting parameters or choosing different pre-training strategies necessitates retraining the model, increasing resource consumption.

[0055] 6.2 Experimental Setup and Evaluation Indicators Datasets: Six datasets of different sizes were used, including purchase relationship graphs (Amac, Amap), coauthor graphs (Coauthor_CS, Coauthor_Phy), and citation network graphs (Citeseer, Cora). Statistics are shown in Table 2.

[0056] Evaluation metrics: Accuracy (ACC), Normalized Mutual Information (NMI), Average RAND Index (ARI), and Macro F1 Score (F1) were used. Higher scores indicate better clustering results.

[0057] Parameter Setting: AGCN is implemented on the PyTorch platform and runs on an NVIDIA 3090 GPU. AGCN Application Global Structure To propagate attributes, and then after diffusion Perform k-means++ clustering. Diffusion parameters Choose from {2, 4, 5, 7}. The hidden size of the latent features varies from {800, 1000}. The learning rate of the Adam optimizer is set to... In the anchor sample selection phase, the number of rounds was chosen from {40, 60, 80}. For each dataset, the top 0.1%, 1%, or 2% of samples in each cluster were selected as anchor samples. The update frequency was chosen from {10, 20, 80, 180}, and the maximum number of iterations was set to 1000.

[0058] 6.3 Comparison of Experimental Results The clustering results of AGCN on seven datasets are shown in Tables 3, 4, 5, and 6. These datasets are the original datasets. In the context of graph attribute clustering, contrastive graph clustering methods generally outperform simple single-view generative clustering methods. For example, methods like CCGC and HSAN utilize the properties of samples in the view... Figure 1 Heshi Figure 2 The differences between clustering methods are represented, and important clustering information is extracted through contrastive learning to obtain better clustering results. On the other hand, single generative clustering methods such as SDCN, while effective in themselves, rely to some extent on the accuracy of model pre-training. This leads to unsatisfactory model performance or even model collapse when the pre-training of the encoder-decoder structure cannot adapt to the dataset. Furthermore, the encoder-decoder pre-training strategy cannot adapt to large datasets, and may even lead to pre-training memory overflow (OOM). Next, this paper will analyze and compare AGCN with previous graph clustering methods from different perspectives and dataset sizes, comprehensively demonstrating the superiority of AGCN. AGCN's clustering in Figure 6 , Figure 7 and Figure 8 The visualization vividly demonstrates its overwhelming advantage over current single-view clustering methods, while also showcasing its superior performance compared to multi-view clustering methods.

[0059] On the lightweight datasets Citeseer and Cora, pre-trained generative clustering methods and comparative methods perform similarly. AGCN is a single-view clustering method. Compared to generative clustering methods, AGCN does not require pre-training, thus avoiding the problem of poor model training performance caused by pre-training with encoder-decoder structures being unable to adapt to datasets, such as SDCN and DFCN on the Cora dataset. Compared to generative clustering methods, AGCN can adapt to more complex scenarios, thus outperforming the best method FT-VGAE by 0.87% and 1.74%, respectively. Compared to multi-view clustering, by progressively expanding anchor samples and handling difficult samples, AGCN's ability to extract important clustering information is more refined, improving performance by 0.57% and 1.11% compared to the best-performing comparative clustering method HSAN.

[0060] On the medium-sized datasets Amac and Amap, the performance gap between generative single-view methods and contrastive multi-view methods becomes more pronounced. As the dataset size increases, the pre-training cost of generative single-view clustering methods rises, making it difficult to obtain satisfactory pre-training results for model training. In contrast, contrastive graph clustering methods can maintain good performance by leveraging the contrast between views. The AGCN method avoids the negative impact of pre-training by using anchor samples to guide model training. Therefore, compared to generative clustering methods, AGCN improves performance by 18.08% and 3.4%, respectively. Due to the model's more granular feature extraction, AGCN also outperforms the best-performing contrastive graph clustering method on both datasets, by 2.36% and 2.19%, respectively.

[0061] On large-scale graph datasets such as Coauthor_cs, Coauthor_phy, and ogbn-arXiv, some generated single-view clustering methods cannot be pre-trained due to excessive pre-training costs. Pre-training processes exceed the available memory on 3090 GPUs, leading to unsuccessful pre-training, such as DFCN. This limitation severely impacts the performance of these methods on large-scale datasets. For complex multi-view structures, sampling-based training is attempted, but most multi-view methods require comparisons between global and local data, making model training infeasible. However, on even larger datasets, such as Coauthor_cs and Coauthor_phy, AGCN outperforms previous methods by 11.49% and 24.75%, respectively. This highlights the effectiveness of AGCN in handling large-scale and structurally complex datasets, enabling it to handle more complex real-world scenarios. On the ogbn-arXiv dataset, AGCN achieves a significant performance improvement over previous methods.

[0062] 6.4 Effectiveness Analysis of the Selection Strategy for Distribution Margin Anchor Samples This section analyzes the effectiveness of the distribution edge anchor sample selection strategy, and describes the dataset accuracy at each stage and the final selected anchor samples step by step. The accuracy rate was high. Experimental results demonstrate that the selected anchor samples have a high confidence level.

[0063] First, the accuracy of the original dataset was analyzed, and the results are shown in the "Original Data" column. Due to insufficient information redundancy in the original data, its accuracy was low. The diffusion matrix generated after the diffusion operation is then analyzed. The accuracy of the diffusion method was improved compared to the "Original Data," as shown in the "Original Diffusion" column. This improvement may be related to the sparsity of the dataset; for datasets with high sparsity (such as Cora and Citeseer), the accuracy improvement after diffusion is more significant. Next, the clustering results after model training were compared with the diffusion results, and the "After Model Training" column in Table 7 shows this difference. It was found that the clustering results after model training were better than the diffusion results: the model learning process can obtain a more accurate sample distribution than the pseudo-labels. Finally, the accuracy of the selected anchor samples was evaluated. The "Selected Anchor Samples" column in Table 7 shows the accuracy of anchor samples in different datasets. In most datasets, AGCN achieved high accuracy, indicating that the anchor sample selection strategy at the distribution edges has high confidence and reliability. The results further show that anchor samples play an important guiding role in clustering and can effectively improve the overall clustering performance of the model.

[0064] 6.5 Ablation and Convergence Analysis After the first phase, a subset of samples is selected as anchor samples. The proportion of anchor samples is relatively small, for example, 1% in the Cora and Citeseer datasets, and 2% in the Coauthor_cs and Coauthor_phy datasets. If the anchor samples are not updated, the limited training data may lead to inaccurate feature extraction, resulting in overfitting due to excessive sensitivity to abnormal feature changes.

[0065] from Figure 9 , Figure 10 and Figure 11 Analysis shows that without updating anchor samples, the model accuracy oscillates within a certain range and gradually decreases. This indicates that the model becomes dependent on specific samples, causing random fluctuations in the results for other samples. Updating anchor samples not only increases the number of training samples but also helps the model better extract key clustering information, stabilizing model performance. Ablation studies compared the results with and without anchor sample updates, confirming the advantages of AGCN in extracting key clustering information.

[0066] 6.6 Validity Analysis of the Anchor Sample Guidance Model To verify the universality of the anchor sample-guided model, this paper ported it to the SDCN and DFCN models, trained it on the Cora, Citeseer, and Amap datasets respectively, and compared it with pre-trained methods to demonstrate the effectiveness and advantages of anchor sample-guided training: Referring to Table 8, on the Cora dataset, SDCN and DFCN performed poorly using pre-training, while anchor sample-guided training improved their accuracy by 20.44% and 17.26%, respectively. The encoder-decoder pre-training structure is unsuitable for the Cora dataset, causing the model to fail to achieve the expected results. This demonstrates that the anchor sample method proposed in this paper has stronger adaptability and can adapt to more complex real-world scenarios.

[0067] On the Citeseer dataset, the anchor sample guidance method has a slight advantage over the pre-training method, improving the accuracy of SDCN and DFCN by 0.44% and 1.31%, respectively.

[0068] In the Amap dataset, SDCN failed to effectively explore the data distribution, while the anchor sample-guided method improved its accuracy by 10.55%, and DFCN by 0.04%.

[0069] In summary, compared with pre-trained generative clustering methods, anchor sample-guided methods are applicable to more complex real-world scenarios and exhibit stronger adaptability and better performance.

[0070] This application proposes a novel generative clustering method, AGCN. This method guides model training through anchor samples, effectively replacing the traditional pre-training process, and enhances the model's ability to extract key clustering information through iterative updates of anchor samples. Inspired by positive sample alignment and pseudo-labeling techniques, this application designs a Distribution-aware Anchor Selection Mechanism to select samples with significant discriminative power to guide the training process. Experimental results verify the important role of anchor samples in improving the model's clustering performance, especially demonstrating excellent performance in complex and varied real-world scenarios.

[0071] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An anchor-guided graph clustering network, characterized in that, include: (1) Use feature analysis methods to generate pseudo-labels with global distribution characteristics; (2) The model learns the mapping relationship between sample features and pseudo-labels; (3) In the embedding space, samples with the same label gradually aggregate. The model dynamically corrects erroneous pseudo-labels during the learning process, causing them to re-aggregate into semantically similar clusters.

2. The anchor-guided graph clustering network according to claim 1, characterized in that, The generation of pseudo-labels with global distribution features using feature analysis methods includes: The diffusion attribute matrix X of the target undirected graph is processed using k-means++. p Clustering is performed to obtain the pseudo-label one-hot encoding matrix Y. k ; Extracting local important features from the diffusion attribute matrix Xp to construct the enhanced attribute matrix X e , will X e With symmetric normalized adjacency matrix The data are input together into the model for training, and the one-hot encoding matrix Y is used. k Guided learning of the model; After a quick and coarse training of the model, the easily separable samples that are farthest from the dividing surface and have the largest differences are selected as the anchor samples V for the next stage. a .

3. The anchor-guided graph clustering network according to claim 1, characterized in that, The method of learning the mapping network relationship between sample features and pseudo-labels through the model includes: anchor sample V a One-hot encoding of Y with pseudo-labels of anchor samples a The mapping relationship between them serves as a constraint on the model's results. Conduct training: Generate the target distribution P; Minimize the KL divergence loss between the distributions of Y and P; A threshold update mechanism is used to gradually expand the anchor sample set.

4. The anchor-guided graph clustering network according to claim 1, characterized in that, In the embedding space, samples with the same label are gradually aggregated. During the learning process, the model dynamically corrects erroneous pseudo-labels, causing them to re-aggregate into semantically similar clusters, including: Peel off samples that are closer to the segmentation surface; Using the diffusion attribute matrix X p C cluster centers are generated by combining the anchor samples of the corresponding clusters; Cluster centers C obtained by taking the mean of anchor samples within the same cluster p Calculate Vr and C for each cluster center. p Spatial distance Dis; Find the minimum distance calculated between the sample and the cluster center, and then assign each sample to the cluster with the minimum distance. Complete the extraction of important clustering information from the model.

5. The anchor-guided graph clustering network according to claim 1, characterized in that, Also includes: Loss function and complexity analysis.

6. The anchor-guided graph clustering network according to claim 5, characterized in that, The loss function analysis includes: Integrating steps (1), (2), and (3) into the same framework, the total loss function is: in, = U The set of pseudo-labels for all hard samples and anchor samples. = + , This indicates the moment when the anchor sample was selected.

7. The anchor-guided graph clustering network according to claim 5, characterized in that, Loss complexity analysis includes: The time complexity of the convolutional neural network layer is... ,exist Previous In addition to the model's own operation, the computational complexity of the cross-entropy function is also O(n). ,exist Then, KL divergence was added as the loss function, with a time complexity of O(n log n). The time complexity of anchor sample selection and updating is O(n log n). The time complexity for processing difficult samples is ,because Therefore, the time complexity of the cross-entropy function at this point is still O(n log n). ,because The intervals between training cycles are very small compared to the entire training cycle and can be ignored. Therefore, the final time complexity of the entire model is O(n log n). .

Images