A community detection method based on dynamic multi-scale feature fusion network
Through the dynamic multi-scale feature fusion network, the self-coding network and the Transformer network combined with the dynamic attention mechanism are used to solve the problems of low accuracy and limited efficiency improvement in community detection, and efficient detection of complex community structures is achieved.
Patent Information
- Application Number
- CN202211326522.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-27
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-10-27
AI Technical Summary
The existing community detection technology has low accuracy when dealing with complex and dynamically changing community structures, and the shallow depth of deep learning network architecture leads to limited efficiency improvement, and fails to fully integrate flexible network node characteristics and topological structures.
Using a dynamic multi-scale feature fusion network method, a community detection network model is constructed, and a self-encoding network and a Transformer network are used to perform feature learning and information fusion. Combining dynamic attention mechanism and self-supervised clustering modules, a high-precision community is generated. Test results.
It improves the accuracy of community detection, can dynamically capture node information in complex community structures, and improves the efficiency and accuracy of community detection.
Smart Images

Figure CN115964626B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of community detection, and in particular to a community detection method based on a dynamic multi-scale feature fusion network. Background Art
[0002] Community detection is a challenging task in network analysis. It aims to divide the topological graph formed by the network into multiple non-intersecting subgraphs to reveal the hidden relationships within the network. Community detection is widely used in many fields such as interpersonal relationship analysis and character preference analysis in the real world. The current community detection technology is mainly divided into two categories: using probabilistic graph models combined with prior knowledge to infer community structure, and using deep learning technology to convert complex network data into low-dimensional data for representation learning. However, on the one hand, traditional probabilistic graph methods are difficult to apply to complex and dynamically changing community structures, and on the other hand, it is difficult to fuse community feature information, resulting in low accuracy in community detection tasks. Deep learning technology, which has developed rapidly in recent years, converts high-dimensional data into low-dimensional data for representation learning. Therefore, it can quickly adapt to community structures with dynamically changing characteristics, and improve the accuracy of community detection tasks by extracting community node features and structural features between nodes. However, deep learning currently has room for improvement in community detection. On the one hand, since the deep learning network architecture used in community detection is often a shallow network, the efficiency of large-scale community detection has not been greatly improved; on the other hand, community detection technology does not use flexible network node features and network topology structure fusion methods. Therefore, it is very necessary to explore the application of dynamic multi-scale feature fusion networks in community detection. Summary of the invention
[0003] The content of the present invention is used to introduce the concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of the present invention is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.
[0004] In order to solve the technical problem of low accuracy of community detection, the present invention proposes a community detection method based on a dynamic multi-scale feature fusion network.
[0005] The present invention provides a community detection method based on a dynamic multi-scale feature fusion network, the method comprising:
[0006] Get the data set to be detected by the community;
[0007] Preprocessing the data set to be detected in the community to obtain target detection data features;
[0008] According to the target detection data features and the trained community detection network model, the data to be detected in the data set to be detected is clustered to generate clustering result information.
[0009] Furthermore, the preprocessing of the data set to be detected in the community to obtain target detection data features includes:
[0010] According to the data set to be detected as a community, a normalized adjacency matrix is established;
[0011] The target detection data features are generated according to the data set to be detected by the community and the normalized adjacency matrix, wherein the target detection data features are data features that can be further processed.
[0012] Furthermore, the training process of the community detection network model includes:
[0013] Build a community detection network model;
[0014] Obtaining a sample community detection dataset group, wherein the data features corresponding to the sample community detection datasets in the sample community detection dataset group are a sample data class set;
[0015] Preprocessing each sample community detection data set in the sample community detection data set group to obtain sample data features corresponding to the sample community detection data set;
[0016] The community detection network model is trained using the sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group to obtain a trained community detection network model.
[0017] Furthermore, the community detection network model includes multiple hybrid network modules at the same scale, a multi-scale node information fusion and a self-supervised clustering module.
[0018] Furthermore, the method of training a community detection network model by using the sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group to obtain a trained community detection network model includes:
[0019] The sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group are used to learn and fuse the node features of the hybrid network module at the same scale, and multi-scale node information fusion is performed to determine the self-supervised network training loss function, and further obtain the trained community detection network model.
[0020] Furthermore, the learning and fusing of node features of the hybrid network modules at the same scale includes:
[0021] The sample data features are input into two deep learning networks respectively, and the features learned by the two deep learning networks are fused using a dynamic attention mechanism, and finally further feature extraction is performed with the structural information of the data.
[0022] Furthermore, the sample data features are input into two deep learning networks respectively, the features learned by the two deep learning networks are integrated by a dynamic attention mechanism, and finally further feature extraction is performed with the structural information of the data, including:
[0023] Using autoencoder networks for feature learning;
[0024] The Transformer network is used to learn the data feature representation. The Transformer network consists of three modules: a data input module, an encoding module, and a task-based multi-layer perceptron module. The data input module contains a linear layer, the encoding module contains a self-attention layer, a feedforward network layer, and a residual connection layer, and the task-based multi-layer perceptron module contains multiple linear layers and multiple nonlinear layers.
[0025] Autoencoder networks and Transformer networks are used for dynamic hybrid learning to represent data information, and the weights are solved through the self-attention mechanism. The weights are added to the corresponding network, and then the structural characteristics of the data are integrated to finally learn the data information.
[0026] Furthermore, the formula for determining the self-supervised network training loss function is:
[0027] ζ=αζ R +βζ C
[0028] Among them, ζ is the self-supervised network training loss function, α and β are two coefficients, ζ C is the clustering loss, ζ R is the reconstruction loss.
[0029] Furthermore, the formula corresponding to establishing the normalized adjacency matrix is:
[0030]
[0031] in, is the normalized adjacency matrix, D is the degree matrix, A is the adjacency matrix of the graph, where the graph can be the graph constructed by the community detection dataset, and I is the identity matrix.
[0032] Furthermore, the formula for generating the target detection data feature is expressed as:
[0033]
[0034] Among them, X′ is the target detection data feature, is the normalized adjacency matrix, and X is the data feature corresponding to the data set to be detected by the community.
[0035] The present invention has the following beneficial effects:
[0036] The present invention adopts a dynamic multi-scale feature fusion network and uses two modules TDCN-M and TDCNS to dynamically capture node information, which can be applied to the dynamic and complex community detection field and improve the accuracy of community detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0038] Figure 1 A flowchart of a community detection method based on a dynamic multi-scale feature fusion network according to the present invention;
[0039] Figure 2 A schematic diagram of a single Transformer model framework according to the present invention;
[0040] Figure 3 A schematic diagram of learning data feature representation using a Transformer network according to the present invention;
[0041] Figure 4 A schematic diagram of dynamically hybrid learning to represent data information using an autoencoder network and a Transformer network according to the present invention. DETAILED DESCRIPTION
[0042] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.
[0043] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0044] The present invention provides a community detection method based on a dynamic multi-scale feature fusion network, the method comprising the following steps:
[0045] Get the data set to be detected by the community;
[0046] Preprocess the community detection dataset to obtain target detection data features;
[0047] According to the characteristics of the target detection data and the trained community detection network model, the data to be detected in the community detection data set are clustered to generate clustering result information.
[0048] The following is a detailed explanation of each of the above steps:
[0049] refer to Figure 1 , shows the process of some embodiments of a community detection method based on a dynamic multi-scale feature fusion network according to the present invention. The community detection method based on a dynamic multi-scale feature fusion network includes the following steps:
[0050] Step S1, obtaining a data set to be detected by the community.
[0051] In some embodiments, a dataset to be detected as a community may be obtained.
[0052] The data to be detected in the data set to be detected may be data to be detected. For example, the data set to be detected may include but is not limited to: a set of purchased goods of a target user, a set of urban population, and a set of shelf items. The target user may be a user who purchases goods.
[0053] As an example, when the data set to be detected as a community is a set of purchased items of a target user, community detection is subsequently performed on the set of purchased items of the target user to obtain the purchase preferences of the target user.
[0054] As another example, the Pytorch framework can be used to read in the dataset to be detected by the community.
[0055] Step S2, preprocessing the community detection data set to obtain target detection data features.
[0056] In some embodiments, the above-mentioned community detection data set can be preprocessed to obtain target detection data features.
[0057] The target detection data feature may be a pre-processed data set to be detected in the community. The target detection data feature may also be a data feature that can be further processed.
[0058] As an example, this step may include the following steps:
[0059] The first step is to establish a normalized adjacency matrix based on the above-mentioned community detection data set.
[0060] For example, the formula for establishing a normalized adjacency matrix can be:
[0061]
[0062] in, is the normalized adjacency matrix. D is the degree matrix. A is the adjacency matrix of the graph, where the graph can be the graph constructed from the community detection dataset. I is the identity matrix.
[0063] The second step is to generate target detection data features based on the above-mentioned community detection data set and normalized adjacency matrix.
[0064] The formula for generating target detection data features is expressed as:
[0065]
[0066] Among them, X′ is the target detection data feature (processed data feature). is the normalized adjacency matrix. X is the data feature corresponding to the data set to be detected.
[0067] Step S3: clustering the data to be detected in the community detection data set according to the target detection data characteristics and the trained community detection network model to generate clustering result information.
[0068] In some embodiments, the data to be detected in the data set to be detected may be clustered according to the target detection data features and the trained community detection network model to generate clustering result information.
[0069] The community detection network model can be used to cluster the data to be detected in the community detection data set. The clustering result information can represent the result of community detection on the community detection data set.
[0070] As an example, clustering index calculation and community detection analysis. Among them, clustering indicators may include: accuracy (ACC), normalized mutual information (NMI), average Rand index (ARI) and F1 recall. The higher these four indicators are, the better the clustering effect is, and thus the better the community detection effect is.
[0071] Optionally, the training process of the community detection network model may include the following steps:
[0072] The first step is to build a community detection network model.
[0073] The community detection network model includes multiple hybrid network modules (TDCN-M) at the same scale, a multi-scale node information fusion (TDCN-S), and a self-supervised clustering module. Both TDCN-M and TDCN-S are composed of an autoencoder architecture (AE) and a Transformer structure.
[0074] For example, the parameters in the community detection network model can be initialized. For example, the number of hybrid network modules at the same scale can be set to a first preset number, the number of data set clusters can be set to a second preset number, the number of encoding blocks in the Transformer network can be set to a third preset number, and the number of multilayer perceptron modules in the Transformer network can be set to a fourth preset number. The framework diagram of a single Transformer model can be as follows: Figure 2 As shown. Among them, the first preset number, the second preset number, the third preset number and the fourth preset number can be preset numbers. The first preset number, the second preset number, the third preset number and the fourth preset number may not be equal. The first preset number can be recorded as M. The second preset number can be recorded as k. The third preset number can be recorded as K. The fourth preset number can be recorded as N. The number of clusters of the data set can be the number of categories obtained by clustering through the community detection network model.
[0075] The second step is to obtain a sample community detection dataset group.
[0076] Among them, the data type of the sample community detection data in the sample community detection data set in the above sample community detection data set group may be the same as the data type of the data to be community detected. The sample data class set corresponding to the sample community detection data set may be known. The sample data class set corresponding to the sample community detection data set may be a category set obtained by clustering the sample community detection data in the sample community detection data set in advance. The sample data class set may characterize the division of the sample community detection data in the sample community detection data set. The data features corresponding to the sample community detection data set in the above sample community detection data set group are the sample data class sets.
[0077] In actual situations, the more sample community detection datasets there are in a sample community detection dataset group, the better the effect of training the community detection network model is.
[0078] The third step is to preprocess each sample community detection dataset in the sample community detection dataset group to obtain sample data features corresponding to the sample community detection dataset.
[0079] The specific implementation method of this step can refer to step S2. The sample community detection dataset can be used as the dataset to be community detected. Step S2 is executed, and the target detection data features obtained are the sample data features.
[0080] In the fourth step, the community detection network model is trained using the sample data features and data features corresponding to each sample community detection data set in the above sample community detection data set group to obtain a trained community detection network model.
[0081] For example, the sample data features and data features corresponding to each sample community detection data set in the above sample community detection data set group can be used to learn and fuse the node features of the hybrid network module at the same scale, and perform multi-scale node information fusion to determine the self-supervised network training loss function, and further obtain the trained community detection network model. This step may include the following sub-steps:
[0082] In the first sub-step, the node features of TDCN-M (Multi-Network Feature Fusion Module) at the same scale are learned and fused.
[0083] For example, the sample data features can be input into two deep learning networks respectively, and the features learned by the two deep learning networks can be fused with the dynamic attention mechanism, and finally further feature extraction can be performed with the structural information of the data. Specifically, the following steps can be included:
[0084] First, feature learning is performed using an autoencoder network, that is, the sample data features are input into a first preset number of layers of autoencoder networks to obtain the features of each layer of the autoencoder network.
[0085] For example, the formula corresponding to the characteristics of each layer of the autoencoder network can be:
[0086] H (l) =σ(W (l) H (l-1) +b (l) )
[0087] Among them, H (l) is the characteristic of the l-th layer autoencoder network. (l-1) is the characteristic of the l-1th layer autoencoder network. l represents the number of layers. The value range of l is {1, 2, 3, ..., M}. M is the first preset number. When l = 1, H (0) is the sample data feature. σ is a nonlinear activation function, such as the Relu function. W (l) and b (l) are the learning parameters of the lth layer in the autoencoder model.
[0088] Next, the Transformer network is used to learn the data feature representation. The Transformer network mainly consists of three modules: data input module, encoding module, and task-based multi-layer perceptron module. The data input module mainly contains a linear layer. The encoding module contains a self-attention layer, a feedforward network layer, and a residual connection layer. The task-based multi-layer perceptron module (MLP) contains multiple linear layers and multiple nonlinear layers. For specific implementation methods, refer to Figure 3 The following steps may be included:
[0089] Step 301, the data input module learns the input data representation.
[0090] For example, the node features of the lth Transformer network input can be recorded as Z (l) , where l also represents the serial number of the Transformer network. The node feature Z of the first Transformer network input (0) is the sample data feature. (l) Passed into the linear layer of the neural network, the calculation formula is as follows:
[0091] Z (l) =W (l) Z (l) +b (l)
[0092] Step 302: the encoding module further processes the data processed by the data input module.
[0093] For example, the incoming data feature represents Z (l) After K encoding modules, a new data representation Z is generated (l) In order to input into the multi-layer perceptron module. Wherein, K is a third preset number. Further, the formula of a single encoding module is as follows:
[0094] Z′ (p) =LN(MHA(Z′ (p) )+Z′ (p) )
[0095] Z (p) =LN(FFN(Z′ (p) )+Z′ (p) )
[0096] Among them, the value range of p is {1, 2, 3, …, K}.
[0097] Step 303: MLP further processes the data.
[0098] For example, the data feature Z processed by the encoding module (l)Input into MLP to obtain a new data representation for data information fusion. Further, the formula of a single MLP is as follows:
[0099] Z′ (p) =W (p) Z (p) +b (p)
[0100] Z (p) =ELU(Z′ (p) )
[0101] The value range of p can also be {1, 2, 3, ..., N}. MLP is a multi-layer perceptron.
[0102] Then, the autoencoder network and the Transformer network are used to perform dynamic hybrid learning to represent the data information. The weights are solved through the self-attention mechanism. That is, the autoencoder network and the Transformer network can solve the weights through the self-attention mechanism. The weights are added to the corresponding network, and then the structural characteristics of the data are integrated to finally learn the information of the data. For the specific implementation method, please refer to Figure 4 The following steps may be included:
[0103] Step 401, calculate the weights of the two networks respectively.
[0104] The calculation formula for this step is as follows:
[0105] Ψ l =l2(softmax(a T (LeakyReLU([Z (l) ||H (l) ]W (l) ))))
[0106] Where a is a learnable parameter during training. l2 is L2-Norm. L2 is normalization. Z (l) is the feature learned by the Transformer module. (l) is the output of the autoencoder module AE. (l) is a learnable parameter of the neural network. l It is of 2d×2 dimension, where d is the dimension of the output of the hybrid network.
[0107] Step 402, applying the calculated weights to the output of the corresponding network.
[0108] For example, the calculated weight Ψ l , applied to the output of the corresponding network, the calculation method is as follows:
[0109] Z (l) =(ψ l,1)⊙Z (l) +(ψ l,2 )⊙H (l)
[0110] Where ⊙ is the Hadamard product operation. (ψ l,1 ) is the matrix Ψ l The first column of l,2 ) is the matrix Ψ l The second column.
[0111] Step 403, final calculation of node features of TDCN-M.
[0112] For example, the adjacency matrix of the data is integrated into the learned features to obtain new data feature information as the input of the next TDCN-M. The calculation formula is as follows:
[0113]
[0114] Among them, Z (l) It is the feature learned in the first layer. is the normalized adjacency matrix. Transformer() means that it is input into the Transformer module in the form of brackets.
[0115] The second sub-step is multi-scale node information fusion (TDCN-S). TDCN-S (Multi-Scale Heterogeneous Fusion Module) generates the output Z of L TDCN-Ms. (l) The output of the Mth layer of the autoencoder network is weighted and fused to further explore the characteristics of node features at multiple scales. Specifically, the following steps may be included:
[0116] First, connect the output of each module of TDCN-M and the last layer output of the autoencoder network to obtain the hidden state Z′. The calculation formula is as follows:
[0117] Z′=[Z (1) ||Z (2) ||…||Z (l) ||H (M) ]
[0118] Among them, the value range of l is {1, 2, 3, …, L}.
[0119] Finally, the final representation Z is calculated. The attention calculation method can refer to step 401 to obtain the weight Ψ. The weight Ψ is also calculated in the form of Hadamard product to calculate the hidden state Z′. The calculation of Z refers to step 403.
[0120] The third sub-step is to determine the loss function of the self-supervised network training. The loss function mainly consists of two parts: clustering loss ζ C and reconstruction loss ζ R The formula for determining the loss function of the self-supervised network training is as follows:
[0121] ζ=αζ R +βζ C
[0122] Where ζ is the self-supervised network training loss function. α and β are two coefficients. C is the clustering loss. R is the reconstruction loss.
[0123] Furthermore, R The calculation method is obtained from the decoding part of the autoencoder network. The decoder is the feature H output by the encoder (m) Reconstruct and get the predicted value of node features The original node features and the predicted features are normalized. The specific formula is as follows:
[0124]
[0125] Furthermore, the clustering loss ζ C It is composed of the final representation Z and the self-encoding output H (m) The KL divergence calculated for each.
[0126] The present invention adopts a dynamic multi-scale feature fusion network and uses two modules TDCN-M and TDCNS to dynamically capture node information, which can be applied to the dynamic and complex community detection field and improve the accuracy of community detection.
[0127] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
Claims
1. A community detection method based on a dynamic multi-scale feature fusion network, characterized in that: The following steps are involved: Get the data set to be detected by the community; Preprocessing the data set to be detected in the community to obtain target detection data features; Clustering the data to be detected in the data set to be detected according to the target detection data features and the trained community detection network model to generate clustering result information; The training process of the community detection network model includes: Build a community detection network model; Obtaining a sample community detection dataset group, wherein the data features corresponding to the sample community detection datasets in the sample community detection dataset group are a sample data class set; Preprocessing each sample community detection data set in the sample community detection data set group to obtain sample data features corresponding to the sample community detection data set; Using the sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group, the community detection network model is trained to obtain a trained community detection network model; The community detection network model includes multiple hybrid network modules at the same scale, a multi-scale node information fusion and a self-supervisory network; The method of training the community detection network model by using the sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group to obtain a trained community detection network model includes: Using the sample data features and data features corresponding to each sample community detection data set in the sample community detection data set group, the node features of the hybrid network module at the same scale are learned and integrated, and multi-scale node information is integrated to determine the self-supervised network training loss function, and further obtain a trained community detection network model; The learning and fusing of node features of hybrid network modules at the same scale includes: The sample data features are input into two deep learning networks respectively, and the features learned by the two deep learning networks are fused with the dynamic attention mechanism, and finally further feature extraction is performed with the structural information of the data; The sample data features are input into two deep learning networks respectively, the features learned by the two deep learning networks are integrated with the dynamic attention mechanism, and finally further feature extraction is performed with the structural information of the data, including: Using autoencoder networks for feature learning; The Transformer network is used to learn the data feature representation. The Transformer network consists of three modules: a data input module, an encoding module, and a task-based multi-layer perceptron module. The data input module contains a linear layer, the encoding module contains a self-attention layer, a feedforward network layer, and a residual connection layer, and the task-based multi-layer perceptron module contains multiple linear layers and multiple nonlinear layers. Use autoencoder networks and Transformer networks to perform dynamic hybrid learning to represent data information, and solve weights through the self-attention mechanism, add weights to the corresponding network, and then integrate the structural characteristics of the data to finally learn the data information; The sample data features are input into the first preset number of layers of autoencoder network, and the formula corresponding to the features of each layer of autoencoder network is obtained as follows: Among them, H (l) is the characteristic of the l-th layer autoencoder network; H (l-1) is the characteristic of the l-1th layer autoencoder network; the value range of l is {1, 2, 3, ..., M}; M is the first preset number; when l = 1, H (0) is the sample data feature; σ is the nonlinear activation function; W1 (l) and is the learning parameter of the lth layer in the autoencoder model; The Transformer network is used to learn the data feature representation. The Transformer network mainly includes three modules: data input module, encoding module, and task-based multi-layer perceptron module. The data input module mainly includes a linear layer. The encoding module includes a self-attention layer, a feedforward network layer, and a residual connection layer. The task-based multi-layer perceptron module includes multiple linear layers and multiple nonlinear layers. The specific implementation method includes the following steps: The data input module learns the input data representation by recording the node features of the l-1th Transformer network input as Among them, the node features of the first Transformer network input is the sample data feature; Passed into the linear layer of the neural network, the calculation formula is: in, and is the learning parameter of the lth Transformer network; The encoding module further processes the data processed by the data input module, including: the incoming data feature representation After K encoding modules, a new data representation is generated In order to input into the multi-layer perceptron module; wherein K is the third preset number; the formula of a single encoding module is: Among them, the value range of p is {1, 2, 3, ..., K}; MLP further processes the data including: the data features after being processed by the encoding module Input into MLP to obtain a new data representation; the formula for a single MLP is: Among them, MLP is a multi-layer perceptron; and is the learning parameter of the pth MLP; The autoencoder network and the Transformer network solve the weights through the self-attention mechanism, add the weights to the corresponding network, and then integrate the structural characteristics of the data to finally learn the information of the data. The specific implementation method includes the following steps: The formula for calculating the weights of the two networks is: Among them, a is a learnable parameter during training; l2 is L2-Norm; L2 is normalization; is the feature learned by the lth Transformer network; H (l) is the output of the lth autoencoder network AE; is a learnable parameter of the neural network; l It is 2d×2 dimensional, where d is the dimension of the hybrid network output; The calculated weight Ψ l , applied to the output of the corresponding network, the calculation method is: Where, ⊙ is the Hadamard product operation; (ψ l,1 ) is the matrix Ψ l The first column of l,2 ) is the matrix Ψ l The second column of The final calculation of the node features of the hybrid network module includes: integrating the adjacency matrix of the data into the learned features to obtain new data feature information as the input of the next hybrid network module. The calculation formula is: in, is the feature learned by the lth layer; is the normalized adjacency matrix; Transformer() means it is input into the Transformer network in the form of brackets.
2. According to claim 1, a community detection method based on a dynamic multi-scale feature fusion network is characterized in that: The preprocessing of the community detection data set to obtain target detection data features includes: According to the data set to be detected as a community, a normalized adjacency matrix is established; The target detection data features are generated according to the data set to be detected by the community and the normalized adjacency matrix, wherein the target detection data features are data features that can be further processed.
3. According to claim 1, a community detection method based on a dynamic multi-scale feature fusion network is characterized in that: The formula for determining the self-supervised network training loss function is: z=az R +bz C Among them, ζ is the self-supervised network training loss function, α and β are two coefficients, ζ C is the clustering loss, ζ R is the reconstruction loss.
4. According to claim 2, a community detection method based on a dynamic multi-scale feature fusion network is characterized in that: The formula corresponding to establishing the normalized adjacency matrix is: in, is the normalized adjacency matrix, D is the degree matrix, A is the adjacency matrix of the graph, where the graph is the graph constructed by the community detection dataset, and I is the identity matrix.
5. The community detection method based on a dynamic multi-scale feature fusion network according to claim 2, characterized in that: The formula for generating the target detection data features is expressed as: Among them, X′ is the target detection data feature, is the normalized adjacency matrix, and X is the data feature corresponding to the data set to be detected by the community.
Citation Information
Patent Citations
Multi-scale community discovery method for online social network
CN106789285A
Method and system for detecting overlapping communities in social network based on inter-node similarity
CN107153713A
Brain network classification method based on community detection and double-channel self-coding
CN113642596A
Community structure detection method and system for movie social network
CN114020999A
Large-scale network community detection algorithm based on multi-feature fusion
CN114202035A
