A cloud-edge collaborative network representation method
By using cloud-edge collaboration technology, the main network is deployed in the cloud and the sub-networks are deployed at the edge. Combined with the SkipGram model and feature decoupling technology, the problem of network structure changes caused by node failure or addition in multi-view network representation is solved, thereby improving the robustness and accuracy of network representation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING NORMAL UNIV AT ZHUHAI
- Filing Date
- 2024-04-24
- Publication Date
- 2026-07-24
AI Technical Summary
Existing multi-view network representation methods ignore the implicit relationships between nodes and require a lot of computing resources and time to retrain when nodes fail or are added, resulting in low efficiency when the network topology changes.
By employing cloud-edge collaboration technology, the main network is deployed in the cloud and the sub-network is deployed at the edge. The SkipGram model is used to extract initial features, and through feature decoupling and recombination, the efficiency of cloud computing and the low latency of edge computing are combined to achieve real-time response to network structure changes, thereby improving robustness and accuracy.
It enables real-time scheduling when nodes fail or are added, improves the robustness and accuracy of network representation, fully explores the implicit relationships between subnetworks, and reduces computing resources and time costs.
Smart Images
Figure CN118469035B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cloud computing, and more particularly to a network representation method for cloud-edge collaboration. Background Technology
[0002] Network representation is currently the mainstream method for analyzing complex networks. It aims to transform nodes into low-dimensional vectors to obtain the network's fundamental features for downstream tasks such as node clustering, link prediction, and network visualization. Traditional network representation methods primarily focus on single networks and are not suitable for ultra-large-scale, multi-view complex networks. For example, social networking platforms may contain various relationships between users, such as likes, friends, and subscriptions, each constituting a new sub-network perspective. This characteristic prompts us to consider multiple perspectives within the network to learn stable network node representations.
[0003] For multi-view network representation methods, current mainstream approaches mainly employ multi-view clustering techniques or integrate the weights of each view to obtain the representation. However, these methods mostly focus only on the global information of each view, ignoring the implicit relationships between nodes in a multi-view network. Furthermore, current mainstream network representation methods are based on a static network; when a node fails or a new node is introduced, the network topology changes. Retraining a new network would consume significant computational resources and time.
[0004] To address the above issues, this invention designs a network representation method based on cloud-edge collaboration. This method comprehensively considers the computing and storage capabilities of the cloud and the low latency and real-time processing capabilities of the edge. The main network is placed in the cloud for real-time scheduling, while sub-networks from various perspectives are deployed at the edge, utilizing cloud-edge collaboration technology for representation. This approach not only uncovers the implicit relationships between sub-networks but also addresses real-time changes in network structure caused by node failures or the addition of new nodes, further improving the robustness and accuracy of network representation. Summary of the Invention
[0005] The purpose of this invention is to provide a cloud-edge collaborative network representation method that utilizes cloud-edge collaborative technology for representation. This method not only uncovers the implicit relationships between sub-networks but also addresses real-time changes in network structure caused by node failures or the addition of new nodes, further improving the robustness and accuracy of network representation.
[0006] To achieve the above objectives, the technical method of the present invention is as follows:
[0007] A cloud-edge collaborative network representation method is proposed. This method uses the SkipGram model from the text representation domain to extract initial features for each network. Then, a feature decoupling technique is employed to map each network feature to a common space and a difference space, decoupling the features into common and differential features, revealing their inherent attributes and structure. Finally, the features are recombined and concatenated to construct a more comprehensive feature representation. The specific representation algorithm is as follows: Inputting the main network and sub-networks, the method first uses the SkipGram model from text representation to obtain the initial features for each network. Then, the network features are decoupled and mapped to a common space and a difference space, decoupling the features into common and differential features. Finally, the features are recombined and concatenated for training. The details are shown below:
[0008]
[0009]
[0010] The method of extracting initial features for each network using the SkipGram model in the text representation domain is as follows: Initial network node feature acquisition specifically involves using the SkipGram model to obtain the inherent connections within each network. For any node V in the network... i It is necessary to predict the nodes V around this node. j That is, maximizing the average logarithmic probability:
[0011]
[0012] Where V i and V j Let i and j be two distinct nodes, N be the number of nodes, θ be the SkipGram model parameters, and P(V) be the number of nodes. j |V i ;θ) is the softmax function, defined as follows:
[0013]
[0014] Where f i ∈R d ,f' j ∈R d Let d represent the vectors of the center and surrounding nodes respectively, where d is the dimension of the vector, T is the transpose of the vector, and k is the index of the surrounding node.
[0015] The method of decoupling each network feature and mapping it to a common space and a difference space, then recombining and splicing them for training to reveal its inherent properties and structure, is as follows: For each node feature, it is decoupled and mapped to a common space and a difference space respectively. Specifically, a feedforward neural network is used for feature mapping, and then the main network and the sub-networks are spliced pairwise to obtain common features and difference features respectively.
[0016]
[0017]
[0018] Where h b Indicates the characteristics of the main network node. This represents the node characteristics of the i-th subnetwork. Both are parametric matrices, and || denotes vector concatenation.
[0019] Based on the above steps, a cloud-edge collaborative network representation method is proposed. The method includes main network feature reorganization learning and sub-network feature reorganization learning. The main network is concatenated with common features to form a joint feature vector to enhance the feature expression of the main network. The sub-network is concatenated with its common features with the main network and its own differential features to form a new feature vector to highlight the uniqueness of the sub-network. Then, the main network and sub-networks are jointly trained on downstream node classification and link prediction tasks to obtain the final network node representation.
[0020] The aforementioned main network feature recombination learning is as follows: for the feature fusion of the main network, the main network node features h are... b Common features of nodes in all sub-networks spliced together Furthermore, the features of the sub-networks are integrated into the main network, as follows:
[0021]
[0022] in Let be the node characteristics of the i-th subnetwork;
[0023] The sub-network feature recombination learning is as follows: for the feature fusion of each sub-network, the common features of the main network and the current sub-network are combined. Difference characteristics of subnetworks spliced together The basic features of the main network are incorporated into the sub-networks to ensure their uniqueness, as follows:
[0024]
[0025] The collaborative training objective loss is as follows: the main network and the sub-network are jointly trained on link prediction and node classification of downstream tasks to obtain the final network node representation. The main training loss includes downstream task loss and orthogonal loss.
[0026] The downstream task mentioned is link prediction, and its binary classification loss can be expressed as:
[0027]
[0028]
[0029] in and They are nodes V i and V j The representation vector, V is the set of nodes, and |*| is the absolute value function;
[0030] The orthogonal loss ensures that the mapping matrices are mutually orthogonal, so that after the node features are decoupled, the structure and features of the original nodes are preserved in the common space and the difference space, which is expressed as:
[0031]
[0032] in and Let N be the parameter matrices of the common space and difference space of the i-th subnetwork, respectively, and N be the number of subnetworks.
[0033] Therefore, the final loss objective function L total for:
[0034] L total =L cls +λL0,λ∈[0,1], (10)
[0035] Where L cls L0 is the classification loss, L0 is the orthogonal loss, and λ is the weight constant.
[0036] Preferably, the method employs cloud-edge collaborative technology for network representation, deploying the main network in the cloud and sub-networks at the edge, combining the high computing and storage capabilities of cloud computing with the low latency and localized processing capabilities of edge computing. This addresses changes in network topology caused by network node failures or the introduction of new network nodes. Specifically, when sub-network G... j When node A fails, based on the node's representation vector, the similarity with the failed node can be calculated in the cloud. Then, the top k nodes with the highest similarity are selected as the candidate node set C, as shown below:
[0037] C = Top k (cosine(n i,A)),
[0038] stn i ∈(GG∩G j ), i∈[1,N]. (11)
[0039] Where cosine() is the cosine similarity function, G is the main network, and G... j Let N be the number of nodes in the j-th subnetwork, and n be the number of nodes in the subnetwork. i This represents the i-th node.
[0040] Preferably, the method introduces a new network node. Nodes with the same degree have similar topology and attributes, so a node with the same degree as the new node is selected from the main network as a replacement new network node.
[0041] Technical advantages
[0042] This method fully leverages the advantages of cloud computing and storage while also considering the low latency and real-time processing capabilities of the edge. In this architecture, the main network is deployed in the cloud for global real-time scheduling and management. Simultaneously, sub-networks from various perspectives are distributed at the edge to be closer to users and data sources, enabling efficient data processing and real-time response. By employing cloud-edge collaboration technology, this method not only uncovers the implicit relationships between sub-networks but also enables real-time scheduling when node failures or additions cause changes in the network structure, significantly improving the versatility of network representation. Attached Figure Description
[0043] Figure 1 This is a basic flowchart of the method of the present invention.
[0044] Figure 2 This is a schematic diagram of cloud-edge collaborative scheduling of the present invention.
[0045] Figure 3 This is a multi-perspective network representation framework diagram based on cloud-edge collaboration in this invention.
[0046] Figure 4 This is a schematic diagram of network representation training according to the present invention.
[0047] Figure 5 This is a comparison chart of collaborative training results on link prediction in the implementation example.
[0048] Figure 6 This is a diagram showing the result of changing the network topology in the example.
[0049] Figure 7 This is a clustering graph representing the vectors of the implementation example. Detailed Implementation
[0050] The present invention will now be described in detail with reference to the accompanying drawings.
[0051] like Figure 1-4 As shown, a cloud-edge collaborative network representation method is described in the following flowchart. Figure 1 As shown, the specific representation algorithm is as follows:
[0052]
[0053] Among them, such as Figure 3 The diagram shown is a multi-view network representation framework based on cloud-edge collaboration. The main network is placed in the cloud, and each sub-network is arranged at the edge. The features of each sub-network are obtained through cloud-edge collaboration.
[0054] Among them, such as Figure 4 The diagram shown illustrates network representation training. The initial node representations are decoupled into a common feature space and a difference feature space. Then, the features of the main network and sub-networks are reorganized to perform representation learning.
[0055] Specifically, this method first uses the SkipGram model from the field of text representation to extract initial features for each network. Then, a feature decoupling technique is employed to map the features of each network to a common space and a difference space, respectively, to reveal their inherent properties and structure. Based on this, the features are reorganized and concatenated to construct a more comprehensive feature representation.
[0056] For the main network, it is concatenated with common features to form a joint feature vector, thereby strengthening the feature representation of the main network. Simultaneously, for each sub-network, its common features with the main network and its own unique features are concatenated to form a new feature vector, highlighting the uniqueness of the sub-network. Finally, the main network and sub-networks are jointly trained on downstream node classification and link prediction tasks to obtain the final network node representation. Specific details are as follows:
[0057] 1) Initial Network Node Feature Acquisition: For each network, the SkipGram model is used to obtain the inherent connections within the network. For any node V in the network... i It is necessary to predict the nodes V around this node. j That is, maximizing the average logarithmic probability:
[0058]
[0059] Where N is the number of nodes, P(V) j |V i ;θ) is the softmax function, defined as follows:
[0060]
[0061] Where f i ∈R d,f' j ∈R d Let d represent the vectors of the center and surrounding nodes, respectively, where d is the dimension of the vector, T is the transpose of the vector, and k is the index of the surrounding node.
[0062] 2) Acquisition of common and differential features: For each node feature, it is mapped to the common space and the differential space respectively through decoupling. This method uses a feedforward neural network for feature mapping, and then concatenates the main network and the sub-networks pairwise to obtain the common features and differential features respectively.
[0063]
[0064]
[0065] Where h b Indicates the characteristics of the main network node. This represents the node characteristics of the i-th subnetwork. Both are parametric matrices, and || denotes vector concatenation.
[0066] 3) Main network feature learning: For the feature fusion of the main network, the main network node features h are... b Common features of nodes in all sub-networks spliced together Furthermore, the features of the sub-networks are integrated into the main network. Specifically, as shown below:
[0067]
[0068] in Let be the node characteristics of the i-th subnetwork.
[0069] 4) Sub-network feature learning: For feature fusion of each sub-network, this method combines the common features of the main network and the current sub-network. Difference characteristics of subnetworks The purpose of splicing is to integrate the basic features of the main network into the sub-network while ensuring the uniqueness of the sub-network.
[0070]
[0071] 5) Co-training objective loss: This method jointly trains the main network and sub-networks on downstream task link prediction and node classification to obtain the final network node representation. Specifically, the main training loss includes two items: downstream task loss and orthogonality loss.
[0072] 5.1) Downstream Task Loss: Taking link prediction as an example, the binary classification loss of the downstream task can be expressed as:
[0073]
[0074]
[0075] in and These are the vector representations of the two nodes, and They are nodes V i and V j The representation vector, V is the set of nodes, and |*| is the absolute value function;
[0076] 5.2) Orthogonal Loss: Orthogonal loss aims to ensure that the mapping matrices are mutually orthogonal, so that after the node features are decoupled, the structure and features of the original nodes are preserved in the common space and the difference space. It is expressed as:
[0077]
[0078] Therefore, the final loss objective function is:
[0079] L total =L cls +λL0,λ∈[0,1], (10)
[0080] Where L cls L0 is the classification loss, L0 is the orthogonal loss, and λ is the weight constant.
[0081] Furthermore, based on the core ideas described above, to address network topology changes caused by node failures or the introduction of new nodes, this method employs cloud-edge collaboration technology for network representation. The concept of cloud-edge collaboration originates from the field of edge computing, combining the high computing and storage capabilities of cloud computing with the low latency and localized processing capabilities of edge computing, representing an innovative solution. Specifically, such as... Figure 2 The diagram shown illustrates cloud-edge collaborative scheduling, where the main network is deployed in the cloud and the sub-networks are deployed at the edge, using cloud-edge collaborative technology to address changes in network topology.
[0082] Among them, when subnetwork G j When node A fails, this method calculates the similarity with the failed node in the cloud based on the node's representation vector, and then selects the top k nodes with high similarity as the candidate node set C, as shown below:
[0083] C = Top k (cosine(n i ,A)),
[0084] stn i ∈(GG∩G j ), i∈[1,N]. (11)
[0085] Where cosine() is the cosine similarity function, G is the main network, and G... j Let N be the number of nodes in the j-th subnetwork, and n be the number of nodes in the subnetwork. i This represents the i-th node.
[0086] When introducing a new node, considering that nodes with the same degree have similar topologies and attributes, this method selects a node with the same degree as the new node from the main network as a replacement node. For example... Figure 2 The cloud-edge collaborative scheduling diagram shown can be used to obtain similar nodes by querying the adjacency matrix.
[0087] Comparative Example
[0088] like Figure 5 The graph shown compares the collaborative training results for link prediction. Figure 5 In the link prediction task, the results of collaborative training are significantly better than those of independent training, especially in the performance of sub-networks. This is because sub-networks can obtain more information about the relationships between nodes from shared features, while also providing additional information to the main network.
[0089] Furthermore, such as Figure 6 The diagram shows the network topology changes, where CC-Net represents the proposed method, Random represents the random assignment method, and Retraining represents the ideal retraining method. Figure 6 In this study, when the network topology changes, and the missing or new rate is less than 20%, the method proposed in this invention is closer to the ideal retraining method and is significantly better than the random assignment representation method.
[0090] Finally, as Figure 7 The clustering graph of the representation vectors shown is in Figure 7 In terms of the results of representation vector clustering, the vectors of the collaborative training method adopted in this invention are more concentrated and compact, which further illustrates that the method proposed in this invention can learn more accurate and robust node representations.
[0091] In summary, the present invention has many advantages and beneficial effects compared with the prior art, specifically reflected in the following aspects:
[0092] 1) Compared with existing multi-view network representation algorithms, this algorithm considers the intrinsic structure of the network and further explores the implicit relationships in the sub-networks.
[0093] 2) A cloud-edge collaboration framework is introduced, placing the main network in the cloud while deploying sub-networks from various perspectives at the edge, which can respond in real time to situations where node failure or the addition of new nodes causes changes in the network structure.
[0094] 3) Compared with existing representation methods, a collaborative training framework is adopted, which further improves the robustness and accuracy of network representation.
[0095] The above is a detailed description of the present invention in conjunction with specific embodiments and comparative examples. It should not be construed that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, any equivalent substitutions or obvious modifications made without departing from the concept of the present invention, and which have the same performance or use, should be considered to fall within the patent protection scope defined by the submitted claims.
Claims
1. A network representation method for cloud-edge collaboration, characterized in that: The method described uses the SkipGram model from the text representation domain to extract initial features for each network. Then, a feature decoupling technique is employed to map each network feature to a common space and a difference space, respectively, to reveal its inherent attributes and structure. Finally, the features are recombined and concatenated to construct a more comprehensive feature representation. The specific representation algorithm is as follows: Inputting the main network and sub-networks, the SkipGram model from text representation is first used to obtain the initial features for each network. Then, the network features are decoupled and mapped to a common space and a difference space, respectively. The decoupled features are then recombined and concatenated for training. A cloud-edge collaborative network representation algorithm is shown below: Input: Main network Sub-network Output: Network node representation The main network is deployed in the cloud, and the sub-networks are deployed at the edge. The SkipGram model is used to obtain the initial representation features of each network. ; ， Feature decoupling yields common features and differences , ， Main network feature reorganization ， Sub-network feature recombination, ， Cooperative training joint loss, ， Optimize the loss to obtain the final network representation. ; The method of extracting initial features for each network using the SkipGram model in the text representation domain is as follows: Initial network node feature acquisition specifically involves using the SkipGram model to obtain the inherent connections within each network for any node in the network. It is necessary to predict the nodes around this node. That is, maximizing the average logarithmic probability: (1) in and These are two different nodes. and , The number of nodes For SkipGram model parameters, The softmax function is defined as follows: (2) in Let these be vectors representing the center and surrounding nodes, respectively. Let the dimension be the vector. This is the transpose of the vector. Use the subscript index of surrounding nodes; The method of decoupling each network feature and mapping it to a common space and a difference space, then recombining and splicing them for training to reveal its inherent properties and structure, is as follows: For each node feature, it is decoupled and mapped to a common space and a difference space respectively. Specifically, a feedforward neural network is used for feature mapping, and then the main network and the sub-networks are spliced pairwise to obtain common features and difference features respectively. (3) (4) in Indicates the characteristics of the main network node. Indicates the first Node characteristics of a subnetwork All are parameter matrices. This indicates vector concatenation; The method described uses cloud-edge collaboration technology for network representation, deploying the main network in the cloud and sub-networks at the edge. This combines the high computing and storage capabilities of cloud computing with the low latency and localized processing capabilities of edge computing to address network topology changes caused by network node failures or the introduction of new network nodes. Specifically, when a sub-network... Middle node When a node fails, based on its representation vector, the similarity to the failed node can be calculated in the cloud, and then the following steps can be taken: Nodes with high similarity are selected as candidate nodes. , means as follows: , (11) in The cosine similarity function is used. Main network, For the first Individual networks, For the number of nodes, Indicates the first Each node.
2. The cloud-edge collaborative network representation method as described in claim 1, characterized in that: The method includes main network feature reorganization learning and sub-network feature reorganization learning. The main network is concatenated with common features to form a joint feature vector to strengthen the feature representation of the main network. The sub-network is concatenated with its common features with the main network and its own differential features to form a new feature vector to highlight the uniqueness of the sub-network. Then, the main network and sub-network are jointly trained on downstream node classification and link prediction tasks to obtain the final network node representation. The aforementioned main network feature recombination learning is as follows: for the feature fusion of the main network, the main network node features are... Common features of nodes in all sub-networks spliced together Furthermore, the features of the sub-networks are integrated into the main network, as follows: (5) in For the first Node characteristics of a subnetwork; The sub-network feature recombination learning is as follows: for the feature fusion of each sub-network, the common features of the main network and the current sub-network are combined. Difference characteristics of subnetworks spliced together The basic features of the main network are integrated into the sub-networks to ensure the uniqueness of the sub-networks, as follows: (6)。 3. The cloud-edge collaborative network representation method as described in claim 1, characterized in that: The collaborative training objective loss is as follows: the main network and the sub-network are jointly trained on link prediction and node classification of downstream tasks to obtain the final network node representation. The main training loss includes downstream task loss and orthogonal loss. The downstream task mentioned is link prediction, and its binary classification loss can be expressed as: (7) (8) in and They are nodes and The representation vector, For a set of nodes, It is an absolute value function; The orthogonal loss ensures that the mapping matrices are mutually orthogonal, so that after the node features are decoupled, the structure and features of the original nodes are preserved in the common space and the difference space, which is expressed as: (9) in and The first The parameter matrices of the common space and difference space of each subnetwork. The number of subnetworks; Therefore, the final loss objective function for: (10) in For classifying losses, For orthogonal loss, This is a weight constant.
4. The cloud-edge collaborative network representation method as described in claim 1, characterized in that: The introduction of new network nodes involves nodes with the same degree having similar topology and attributes. Therefore, a node with the same degree as the new node is selected from the main network as the replacement new network node.