Noise environment-oriented multi-network community discovery method, system and equipment and medium
By constructing a random binary mask matrix and loss function constraints, the problems of noisy edges and redundant edges in multi-networks are solved, achieving stable node representation and community detection, and improving the accuracy and robustness of community detection in multi-networks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 20TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORP
- Filing Date
- 2025-12-24
- Publication Date
- 2026-05-12
Smart Images
Figure CN122020222A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing technology, specifically relating to a method, system, device, and medium for discovering multiple network communities in noisy environments. Background Technology
[0002] Multi-layer networks have been widely used to represent various relationships between objects in the real world. They can be intuitively represented as multi-layered structures where nodes and their attributes are shared across all layers, but each layer exhibits a different topology, reflecting different types of relationships between nodes. For example, in a social network, each node represents a person with unique personality attributes, and the edges in each layer represent different relationships. Clearly, the study of multi-layer networks is a valuable and challenging endeavor.
[0003] Community detection is a crucial technique in multi-view network analysis, aiming to find a set of nodes with dense internal connections and sparse external connections. In image classification, community detection can improve the performance of deep models on noisy data to combat label noise. In recommender systems, it can enhance the accuracy and robustness of user recommendations. Multigraphs, as a special type of heterogeneous graph, contain richer information, undoubtedly posing challenges to graph representation learning. Unsupervised multigraph learning is a powerful tool for multigraph representation, attracting widespread attention due to its ability to utilize information from different views, employing graph neural networks and self-supervised techniques. However, the views in a multigraph network contain complementary but inconsistently distributed information; balancing consistency and effective information utilization during fusion remains a key challenge. Existing methods, by directly concatenating multi-view features, easily introduce redundancy and even noise; others model each view independently, neglecting the inherent relationships between views.
[0004] Current related technologies, such as the Louvain algorithm, Girvan-Newman algorithm, hierarchical clustering, and spectral clustering methods, often neglect the reliability of the graph structure when processing multi-layer networks. However, in real-world graph data, many edges are irrelevant or noisy, which severely impacts the effectiveness of these methods. Noise in multi-layer network data can be categorized into label noise and feature noise: label noise refers to incorrect labeling of entities or edges (e.g., incorrect entity attribute annotations in knowledge graphs). Feature noise arises from biases in the feature vectors of nodes or edges (e.g., abnormal user behavior data in social networks). These issues significantly reduce the accuracy of community discovery results in multi-layer networks. Summary of the Invention
[0005] To overcome the problems of poor accuracy and instability in the detection of multiple network communities in noisy environments due to the presence of noisy edges and redundant edges, this invention provides a method, system, device, and medium for detecting multiple network communities in noisy environments.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] In a first aspect, embodiments of this disclosure provide a method for detecting multiple network communities in noisy environments, comprising the following steps:
[0008] Step S1: Use feature propagation to generate node features of multiple node relationship views in the multi-node network, and process the node features to obtain a refined adjacency matrix;
[0009] Step S2: For the refined adjacency matrix, construct a first random binary mask matrix, and obtain the adjacency mask matrix of the refined adjacency matrix based on the first random binary mask matrix;
[0010] Step S3: For the initial attribute matrix of the node features, construct a second random binary mask matrix, and obtain the attribute mask matrix of the initial attribute matrix based on the second random binary mask matrix;
[0011] Step S4: Input the adjacency mask matrix and attribute mask matrix into the multilayer graph convolutional neural network GCN to obtain the node representation of the relationship view of each node, and concatenate the node representations of the relationship view of each node to obtain the first fusion feature;
[0012] Step S5: Concatenate the initial attribute matrix of the node features into a second fusion matrix, and input the second fusion matrix into a multilayer perceptron (MLP) to obtain the hidden layer representation;
[0013] Step S6: Perform similarity measurement on the same unit hypersphere to obtain the similarity matrix between the node representation and the hidden layer representation. Construct an initial loss function for each node based on the similarity matrix. Aggregate the initial loss functions to obtain an aggregated loss function. Align the node representation and the hidden layer representation in a unified feature space based on the aggregated loss function to obtain aligned node features.
[0014] Step S7: Based on the aligned node feature representation, the K-means algorithm is used to obtain the community partitioning results.
[0015] Furthermore, obtaining the refined adjacency matrix in step S1 includes the following steps:
[0016] Node features are generated for each node's relationship view using feature propagation. :
[0017] ;
[0018] In the formula, , It is the identity matrix; The initial adjacency matrix, Indicates the first One view; The initial attribute matrix for the nodes;
[0019] The similarity matrix between nodes is obtained based on node features Xl. :
[0020] ;
[0021] In the formula, and Both represent nodes; Indicates similarity to cosine;
[0022] After symmetrizing the similarity matrix Hl, the kNN algorithm is used to retain the first few nodes. Each neighbor is used to generate a refined adjacency matrix after standardization. :
[0023] ;
[0024] In the formula, This indicates standardized processing.
[0025] Furthermore, in step S2, the adjacency mask matrix include:
[0026] ;
[0027] In the formula, Represents element-wise multiplication, refining the adjacency matrix , The number of nodes, the first random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
[0028] Furthermore, in step S3, the attribute mask matrix include:
[0029] ;
[0030] In the formula, This indicates element-wise multiplication. This is the initial attribute matrix for node features. , For the feature dimension, the second random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
[0031] Furthermore, the node representation Zl of each node relationship view in step S4 includes:
[0032] ;
[0033] in, , It is the identity matrix. For the first Layer output, where , This is the weight matrix. It is the ReLU activation function;
[0034] Representing each node in the node relationship view Vertical splicing is the first fusion feature , .
[0035] Furthermore, in step S5, the initial attribute matrix of the node features is vertically or longitudinally concatenated to obtain the second fusion matrix. The second fusion matrix Inputting the data into a multilayer perceptron (MLP) yields the hidden layer representations. include:
[0036] .
[0037] Furthermore, in step S6, the similarity matrix include:
[0038] ;
[0039] In the formula, L2 normalization is performed on the node representation and the hidden layer representation respectively to obtain... and ;
[0040] The infoNCE loss for each node is constructed based on the similarity matrix, serving as the initial loss function. :
[0041] ;
[0042] In the formula, and All are similarity matrices elements, For temperature reference; is a natural constant, serving as the base of the natural exponential function;
[0043] The aggregation loss function include:
[0044] ;
[0045] In the formula, The total number of network nodes;
[0046] In a unified feature space, the node representations and hidden layer representations are aligned based on an aggregation loss function, and a contrastive learning framework is used to obtain aligned node features. :
[0047] ;
[0048] In the formula, argmin represents the optimization achieved within the contrastive learning framework. ,make In the embedded space and Maintain maximum consistency.
[0049] In a second aspect, embodiments of this disclosure provide a system for multiple network community detection in noisy environments, comprising:
[0050] The initialization unit is configured to: generate node features of multiple node relationship views in the multi-node network using feature propagation, and process the node features to obtain a refined adjacency matrix;
[0051] The first masking unit is configured to: construct a first random binary masking matrix for the refined adjacency matrix, and obtain the adjacency masking matrix of the refined adjacency matrix based on the first random binary masking matrix;
[0052] The second masking unit is configured to: construct a second random binary mask matrix for the initial attribute matrix of the node features, and obtain the attribute mask matrix of the initial attribute matrix based on the second random binary mask matrix;
[0053] The first fusion unit is configured to: input the adjacency mask matrix and the attribute mask matrix into the multilayer graph convolutional neural network GCN to obtain the node representation of each node relationship view, and concatenate the node representations of each node relationship view to obtain the first fusion feature;
[0054] The second fusion unit is configured to: concatenate the initial attribute matrix of the node features into a second fusion matrix, and input the second fusion matrix into a multilayer perceptron (MLP) to obtain the hidden layer representation;
[0055] The alignment unit is configured to: perform similarity measurement on the same unit hypersphere to obtain a similarity matrix between the node representation and the hidden layer representation; construct an initial loss function for each node based on the similarity matrix; aggregate the initial loss functions to obtain an aggregated loss function; and align the node representation and the hidden layer representation in a unified feature space based on the aggregated loss function to obtain aligned node features.
[0056] The output unit is configured to obtain the community partitioning result based on the alignment node features using the K-means algorithm.
[0057] In a third aspect, embodiments of this disclosure provide an electronic device, characterized in that the electronic device comprises:
[0058] At least one processor; and,
[0059] The memory is communicatively connected to the at least one processor; wherein,
[0060] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for discovering multiple network communities in noisy environments.
[0061] In a fourth aspect, embodiments of this disclosure provide a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method for multiple network community discovery in noisy environments.
[0062] Compared with the prior art, the present invention has the following beneficial technical effects:
[0063] This application provides a method, system, device, and medium for community detection in multi-network environments with noisy conditions. The method constructs a first random binary mask matrix and a second random binary mask matrix, randomly masking some edges and attribute features to generate adjacency mask matrices and attribute mask matrices. This forces the model to learn stable node representations under conditions of missing, incomplete, or even noisy data, effectively avoiding the over-reliance on specific structures or features of a particular view in existing related technologies for GCNs. It also addresses the problem of poor structural reliability caused by noisy and redundant edges in real-world multi-networks. Subsequently, this application constrains the node representations generated by different views in a unified representation space through a loss function, ensuring consistency and comparability between structural and attribute features. This fully integrates complementary information between multiple views, avoiding feature conflicts or pattern shifts, and providing more stable and discriminative node representations for community partitioning, significantly improving the quality of feature representations and the accuracy of community detection. Attached Figure Description
[0064] Figure 1 A flowchart illustrating a method for multiple network community detection in a noisy environment according to an embodiment of this disclosure is shown.
[0065] Figure 2 A block diagram illustrating the method principle of an embodiment of this disclosure is shown;
[0066] Figure 3 A diagram of an apparatus for a method of discovering multiple network communities in a noisy environment, according to an embodiment of this disclosure, is shown. Detailed Implementation
[0067] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0068] To address the challenge of accurately detecting multiple network communities in noisy environments in social platform task analysis, content recommendation, and knowledge graph clustering technologies, this disclosure provides a method flow 100 for detecting multiple network communities in noisy environments, such as... Figure 1 and Figure 2 As shown, it includes the following steps:
[0069] In step S101, feature propagation is used to generate node features of multiple node relationship views in the multi-node network, and the node features are processed to obtain a refined adjacency matrix.
[0070] Specifically, obtaining the refined adjacency matrix in step S101 includes the following steps:
[0071] To simplify computation and preserve low-frequency information, feature propagation is used to generate node features for each node's relationship view. :
[0072] ;
[0073] In the formula, , It is the identity matrix; The initial adjacency matrix, Indicates the first One view; The initial attribute matrix for the nodes;
[0074] The similarity matrix between nodes is obtained based on node features Xl. :
[0075] ;
[0076] In the formula, and Both represent nodes; Indicates similarity to cosine;
[0077] To ensure the symmetry and non-negativity of the adjacency matrix, the similarity matrix Hl is symmetricized, and the kNN algorithm is used to retain the first few nodes of each node. The adjacency matrix is calculated by standardizing the neighboring elements to ensure numerical stability. This standardization process generates a refined adjacency matrix. :
[0078] ;
[0079] In the formula, This indicates standardized processing.
[0080] It should be noted that the adjacency matrix of each node relationship view reflects the node connections under a specific relationship. However, the node relationship view often contains many noisy or redundant edges, which can lead to poor representations learned by the graph neural network and is detrimental to downstream tasks. The refined adjacency matrix is obtained through the graph structure optimization process in step S101. By removing redundant connections, the adjacency matrix of each node's relational view can be refined, enhancing the expressive power of the graph structure and providing high-quality input data for subsequent feature learning.
[0081] Next, proceed to step S102.
[0082] In step S102, for the refined adjacency matrix, a first random binary mask matrix is constructed, and the adjacency mask matrix of the refined adjacency matrix is obtained based on the first random binary mask matrix.
[0083] Specifically, in step S102, the adjacency mask matrix include:
[0084] ;
[0085] In the formula, Represents element-wise multiplication, refining the adjacency matrix , The number of nodes, the first random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
[0086] Next, proceed to step S103.
[0087] In step S103, for the initial attribute matrix of the node features, a second random binary mask matrix is constructed, and the attribute mask matrix of the initial attribute matrix is obtained based on the second random binary mask matrix.
[0088] Specifically, the attribute mask matrix in step S103 include:
[0089] ;
[0090] In the formula, This indicates element-wise multiplication. This is the initial attribute matrix for node features. , For the feature dimension, the second random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
[0091] It should be noted that steps S102 and S103 form random perturbations through the first random binary mask matrix and the second random binary mask matrix, so that in step S104, the model can learn robust node representations from incomplete information and avoid overfitting to noise or random features of a specific view.
[0092] Next, proceed to step S104.
[0093] In step S104, the adjacency mask matrix and attribute mask matrix are input into the multilayer graph convolutional neural network GCN to obtain the node representation of each node relationship view. The node representations of each node relationship view are then concatenated to obtain the first fusion feature.
[0094] Specifically, the node representation Zl of each node relationship view in step S104 includes:
[0095] ;
[0096] in, , It is the identity matrix. For the first Layer output, where , This is the weight matrix. It is the ReLU activation function;
[0097] Representing each node in the node relationship view Vertical splicing is the first fusion feature , .
[0098] Next, proceed to step S105.
[0099] In step S105, the initial attribute matrix of the node features is concatenated into a second fusion matrix, and the second fusion matrix is input into a multilayer perceptron (MLP) to obtain the hidden layer representation.
[0100] Specifically, in step S105, the initial attribute matrix of the node features is vertically or longitudinally concatenated to obtain the second fusion matrix. The second fusion matrix Inputting the data into a multilayer perceptron (MLP) yields the hidden layer representations. include:
[0101] .
[0102] 7. The method for detecting multiple network communities in noisy environments according to claim 1, characterized in that the similarity matrix in step S6... include:
[0103] ;
[0104] In the formula, L2 normalization is performed on the node representation and the hidden layer representation respectively to obtain... and ;
[0105] The infoNCE loss for each node is constructed based on the similarity matrix, serving as the initial loss function. :
[0106] ;
[0107] In the formula, and All are similarity matrices elements, For temperature reference; is a natural constant, serving as the base of the natural exponential function;
[0108] The aggregation loss function include:
[0109] ;
[0110] In the formula, The total number of network nodes;
[0111] In a unified feature space, the node representations and hidden layer representations are aligned based on an aggregation loss function, and a contrastive learning framework is used to obtain aligned node features. :
[0112] ;
[0113] In the formula, argmin represents the optimization achieved within the contrastive learning framework. ,make In the embedded space and Maintain maximum consistency.
[0114] It should be noted that, in this embodiment, the similarity matrix is used. This approach maximizes the similarity of the same node across two views while suppressing the similarity between different nodes, ensuring consistency between the two features at both the semantic and structural levels. An initial loss function is introduced. It can align node representations and hidden layer representations in a unified feature space.
[0115] As the overall training objective, the aggregation loss function It effectively avoids feature bias caused by a single view, suppresses feature drift caused by noisy attributes or edges, and makes the node representation and hidden layer representation complementary and constrained. The node representation mainly describes the topological dependencies of nodes, while the hidden layer representation Zfusion reflects their global attribute associations; in the aggregation loss function... Under the influence of this mechanism, these two types of information are forcibly aligned, resulting in a robust node representation that possesses both structural optimization and attribute advantages. This is achieved through continuous optimization of the aggregation loss function. The model can extract cross-view Figure 1 The latent representation, which is unaffected by weak graph noise, provides more accurate and stable node embeddings for subsequent clustering tasks.
[0116] Next, proceed to step S107.
[0117] In step S107, the community partitioning result is obtained by using the K-means algorithm based on the alignment node features.
[0118] Specifically, align node features Input K-means algorithm, the aligned node features Under feature alignment constraints, it possesses the characteristics of consistent scale, compact intra-class structure, and separation between classes, thus enabling it to stably adapt to the distance calculation assumption of the K-means clustering algorithm and output community partitioning results.
[0119] This embodiment tests the proposed method on three real-world multi-reusable network datasets: two article citation networks, "ACM" and "DBLP," and a social review network, "Yelp." Testing metrics include four commonly used evaluation metrics: NMI (Normalized Mutual Information), ARI (Adjusted Rand Index), ACC (Accuracy), and F1-score. For all four metrics, higher values indicate better algorithm performance. The heterogeneous graph is transformed into a structured graph with multiple structures by extracting meta-paths, and experiments are then conducted.
[0120] The hyperparameter settings and experimental environment settings for the experiment include:
[0121] The parameter settings are as follows: k-means iteration count is 10, KNN neighbor count is 5, contrastive loss function temperature coefficient is 0.2, GCN hidden layer feature dimension is 128, and edge mask code rate is... 0.7 and feature mask rate The sensitivity of these two important parameters was 0.6, and sensitivity analysis was performed in the parameter sensitivity experiment. The MLP had 3 layers, with hidden layer dimensions of 256 and 128 respectively. The Adam optimizer was used, with a learning rate of 0.001 and a maximum number of iterations of 200. This experiment used publicly available Python libraries such as NumPy, PyTorch, Scikit-learn, and DGL to implement the specific code. The processor was an Intel(R) Xeon(R) W-2245 CPU@3.9GHz, the graphics card was an NVIDIA RTX3090 GPU (24GB), and the memory was 64GB. The results were obtained by running the experiment on a server.
[0122] Table 1 presents the experimental results of this method on three datasets: ACM, DBLP, and Yelp. All metrics are expressed in units of %.
[0123] Table 1. Experimental results of this method on the ACM, DBLP, and Yelp datasets.
[0124] Indicator Dataset NMI ARI ACC F1 ACM 74.34±0.68 78.49±0.65 92.21±0.26 92.21±0.28 DBLP 74.19±0.22 79.11±0.28 91.16±0.13 90.61±0.13 Yelp 73.06±0.38 75.38±0.87 92.07±0.30 92.96±0.25
[0125] As can be seen from Table 1, the proposed method in this embodiment has achieved excellent performance in NMI, ARI, ACC and F1 metrics on the three datasets ACM, DBLP and Yelp, proving the effectiveness of the graph representation learning strategy based on dynamic mask and feature alignment proposed in this embodiment.
[0126] Furthermore, the table shows that the variance of the experimental results of the proposed method in this embodiment is less than 1%, with relatively small overall fluctuations, indicating high performance stability across multiple training iterations. In summary, the proposed method demonstrates good generalization ability and robustness on different types of graph data, providing reliable support for subsequent graph community discovery tasks.
[0127] A second embodiment of the present invention also provides a system for multiple network community detection in noisy environments, comprising:
[0128] The initialization unit is configured to: generate node features of multiple node relationship views in the multi-node network using feature propagation, and process the node features to obtain a refined adjacency matrix;
[0129] The first masking unit is configured to: construct a first random binary masking matrix for the refined adjacency matrix, and obtain the adjacency masking matrix of the refined adjacency matrix based on the first random binary masking matrix;
[0130] The second masking unit is configured to: construct a second random binary mask matrix for the initial attribute matrix of the node features, and obtain the attribute mask matrix of the initial attribute matrix based on the second random binary mask matrix;
[0131] The first fusion unit is configured to: input the adjacency mask matrix and the attribute mask matrix into the multilayer graph convolutional neural network GCN to obtain the node representation of each node relationship view, and concatenate the node representations of each node relationship view to obtain the first fusion feature;
[0132] The second fusion unit is configured to: concatenate the initial attribute matrix of the node features into a second fusion matrix, and input the second fusion matrix into a multilayer perceptron (MLP) to obtain the hidden layer representation;
[0133] The alignment unit is configured to: perform similarity measurement on the same unit hypersphere to obtain a similarity matrix between the node representation and the hidden layer representation; construct an initial loss function for each node based on the similarity matrix; aggregate the initial loss functions to obtain an aggregated loss function; and align the node representation and the hidden layer representation in a unified feature space based on the aggregated loss function to obtain aligned node features.
[0134] The output unit is configured to obtain the community partitioning result based on the alignment node features using the K-means algorithm.
[0135] The third embodiment of the present invention also provides an electronic device, the electronic device comprising:
[0136] At least one processor; and,
[0137] The memory is communicatively connected to the at least one processor; wherein,
[0138] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for discovering multiple network communities in a noisy environment according to any of the foregoing embodiments.
[0139] The fourth embodiment of the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method for multiple network community discovery in noisy environments as described in any of the foregoing embodiments.
[0140] The fifth embodiment of the present invention also provides a computer program product, which includes a computing program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions that, when executed by a computer, cause the computer to perform the method for multiple network community discovery in noisy environments according to any of the foregoing embodiments.
[0141] Figure 3 The illustration shows a method or device 1000 implementing an embodiment of the present invention. In some embodiments, more or fewer devices may be included than illustrated. In some embodiments, it may be implemented using a single or multiple devices. In some embodiments, it may be implemented using cloud-based or distributed devices.
[0142] like Figure 3As shown, device 1000 includes a processor 1001, which can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) 1002 or programs and / or data loaded from storage portion 1008 into random access memory (RAM) 1003. Processor 1001 may be a multi-core processor or may contain multiple processors. In some embodiments, processor 1001 may include a general-purpose main processor and one or more special coprocessors, such as a central processing unit (CPU), graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for the operation of device 1000 are also stored in RAM 1003. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Input / output (I / O) interface 1005 is also connected to bus 1004.
[0143] The processor and memory described above are used together to execute programs stored in the memory. When the program is executed by a computer, it can implement the methods, steps, or functions described in the above embodiments.
[0144] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, touchscreen, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1010 as needed so that computer programs read from it can be installed into storage section 1008 as needed. Figure 3 The diagram only shows a portion of the components and does not imply that the device 1000 only includes... Figure 3 The components shown.
[0145] The systems, devices, modules, or units described in the above embodiments can be implemented by a computer or its associated components. The computer may be, for example, a mobile terminal, smartphone, personal computer, laptop computer, in-vehicle human-machine interface device, personal digital assistant, media player, navigation device, game console, tablet computer, wearable device, smart TV, Internet of Things system, smart home, industrial computer, server, or a combination thereof.
[0146] Although not shown, in this embodiment of the invention, a computer-readable storage medium is provided having a computer program / instructions stored thereon that, when executed by a processor, implements the method for multiple network community discovery in noisy environments as described in the embodiments.
[0147] Storage media in embodiments of the present invention include articles that are permanent and non-permanent, removable and non-removable, capable of storing information by any method or technology. Examples of storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.
[0148] Although not shown, embodiments of the present invention also provide a computer program product, including: a computer program / instructions that, when executed by a processor, implement the method for multiple network community discovery in noisy environments as described in the embodiments.
[0149] The methods, programs, systems, apparatuses, etc., in embodiments of the present invention can be executed or implemented in one or more networked computers, or practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks can be performed by remote processing devices connected via a communication network.
Claims
1. A method for detecting multiple network communities in noisy environments, characterized in that, Includes the following steps: Step S1: Use feature propagation to generate node features of multiple node relationship views in the multi-node network, and process the node features to obtain a refined adjacency matrix; Step S2: For the refined adjacency matrix, construct a first random binary mask matrix, and obtain the adjacency mask matrix of the refined adjacency matrix based on the first random binary mask matrix; Step S3: For the initial attribute matrix of the node features, construct a second random binary mask matrix, and obtain the attribute mask matrix of the initial attribute matrix based on the second random binary mask matrix; Step S4: Input the adjacency mask matrix and attribute mask matrix into the multilayer graph convolutional neural network GCN to obtain the node representation of the relationship view of each node, and concatenate the node representations of the relationship view of each node to obtain the first fusion feature; Step S5: Concatenate the initial attribute matrix of the node features into a second fusion matrix, and input the second fusion matrix into a multilayer perceptron (MLP) to obtain the hidden layer representation; Step S6: Perform similarity measurement on the same unit hypersphere to obtain the similarity matrix between the node representation and the hidden layer representation. Construct an initial loss function for each node based on the similarity matrix. Aggregate the initial loss functions to obtain an aggregated loss function. Align the node representation and the hidden layer representation in a unified feature space based on the aggregated loss function to obtain aligned node features. Step S7: Based on the aligned node feature representation, the K-means algorithm is used to obtain the community partitioning results.
2. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, When obtaining the refined adjacency matrix in step S1, the following steps are included: Node features are generated for each node's relationship view using feature propagation. : ； In the formula, , It is the identity matrix; The initial adjacency matrix, Indicates the first One view; The initial attribute matrix for the nodes; The similarity matrix between nodes is obtained based on node features Xl. : ； In the formula, and Both represent nodes; Indicates similarity to cosine; After symmetrizing the similarity matrix Hl, the kNN algorithm is used to retain the first few nodes. Each neighbor is used to generate a refined adjacency matrix after standardization. : ； In the formula, This indicates standardized processing.
3. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, The adjacency mask matrix in step S2 include: ； In the formula, Represents element-wise multiplication, refining the adjacency matrix , The number of nodes, the first random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
4. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, The attribute mask matrix in step S3 include: ； In the formula, This indicates element-wise multiplication. This is the initial attribute matrix for node features. , For the feature dimension, the second random binary mask matrix and with probability Will mask elements in Set to 0, with probability ( Will mask elements in Set it to 1.
5. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, The node representation Zl of each node relationship view in step S4 includes: ； in, , It is the identity matrix. For the first Layer output, where , This is the weight matrix. It is the ReLU activation function; Representing each node in the node relationship view Vertical splicing is the first fusion feature , .
6. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, In step S5, the initial attribute matrix of the node features is vertically or longitudinally concatenated to obtain the second fusion matrix. The second fusion matrix Inputting the data into a multilayer perceptron (MLP) yields the hidden layer representations. include: 。 7. The method for multiple network community detection in noisy environments according to claim 1, characterized in that, The similarity matrix in step S6 include: ； In the formula, L2 normalization is performed on the node representation and the hidden layer representation respectively to obtain... and ; The infoNCE loss for each node is constructed based on the similarity matrix, serving as the initial loss function. : ； In the formula, and All are similarity matrices elements, For temperature reference; is a natural constant, serving as the base of the natural exponential function; The aggregation loss function include: ； In the formula, The total number of network nodes; In a unified feature space, the node representations and hidden layer representations are aligned based on an aggregation loss function, and a contrastive learning framework is used to obtain aligned node features. : ； In the formula, argmin represents the optimization achieved within the contrastive learning framework. ,make In the embedded space and Maintain maximum consistency.
8. A system for detecting multiple network communities in noisy environments, characterized in that, The method for multiple network community detection in noisy environments based on any one of claims 1-7 includes: The initialization unit is configured to: generate node features of multiple node relationship views in the multi-node network using feature propagation, and process the node features to obtain a refined adjacency matrix; The first masking unit is configured to: construct a first random binary masking matrix for the refined adjacency matrix, and obtain the adjacency masking matrix of the refined adjacency matrix based on the first random binary masking matrix; The second masking unit is configured to: construct a second random binary mask matrix for the initial attribute matrix of the node features, and obtain the attribute mask matrix of the initial attribute matrix based on the second random binary mask matrix; The first fusion unit is configured to: input the adjacency mask matrix and the attribute mask matrix into the multilayer graph convolutional neural network GCN to obtain the node representation of each node relationship view, and concatenate the node representations of each node relationship view to obtain the first fusion feature; The second fusion unit is configured to: concatenate the initial attribute matrix of the node features into a second fusion matrix, and input the second fusion matrix into a multilayer perceptron (MLP) to obtain the hidden layer representation; The alignment unit is configured to: perform similarity measurement on the same unit hypersphere to obtain a similarity matrix between the node representation and the hidden layer representation; construct an initial loss function for each node based on the similarity matrix; aggregate the initial loss functions to obtain an aggregated loss function; and align the node representation and the hidden layer representation in a unified feature space based on the aggregated loss function to obtain aligned node features. The output unit is configured to obtain the community partitioning result based on the alignment node features using the K-means algorithm.
9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for multi-network community discovery in noisy environments as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method for multi-network community discovery in noisy environments as described in any one of claims 1 to 7.