A small sample active learning method, system and device for attribute network clustering

Abstract

The present disclosure provides a small-sample active learning method for attribute network clustering, comprising: obtaining an adjacency matrix, an attribute matrix, and an indicator matrix of the attribute network; using the indicator matrix to record the category information of the nodes, and selecting a representative node set of each category in combination with the attribute matrix; obtaining the global clustering results and community cluster prototype properties of the attribute network nodes through each matrix; characterizing the node uncertainty from the adjacency matrix and the global clustering results to obtain the total uncertainty of each node in the attribute network; selecting the node with the highest uncertainty and extracting the label information of the node; selecting all directly connected neighbor nodes from the adjacency relationships of all labeled nodes in the adjacency matrix to construct a subnetwork; obtaining the clustering results of the subnetwork nodes; and using the global clustering results to divide the nodes to obtain the clustering results. The present disclosure solves the problem of small-sample clustering in attribute networks.

Description

Technical Field

[0001] The present disclosure relates to the field of data mining, and in particular to a small sample active learning method, system and device for attribute network clustering. Background Art

[0002] A wide variety of complex networks exist in nature, such as social networks, transportation networks, and power grids. These networks play a vital role in our daily lives. To analyze the properties of complex networks, they are typically modeled as graphs, with entities considered nodes and links as edges. The remaining information is stored elsewhere or converted into edge weights. By leveraging the topological structure of graphs, we can reveal various structures within complex networks and uncover key underlying mechanisms.

[0003] Communities are a crucial network structure, widely present in real-world networks. Community structure is crucial because when networks become too large, their complexity increases dramatically, making them difficult to analyze as a whole. Therefore, it is necessary to group objects into clusters for analysis and to break large networks into smaller subnetworks. This is clustering, also known as community detection. For example, in social networks, communities correspond to groups of people with similar backgrounds or interests. By analyzing the commonalities within these groups, content recommendation algorithms can accurately push appropriate information to users who may be interested. In cancer networks, communities correspond to groups of genes expressing the same or similar functions. By identifying gene communities that express key cancer functions, effective cancer treatments can be discovered.

[0004] Graph clustering (community detection) algorithms aim to discover community structures in complex networks. Traditionally, community detection requires dense connections between nodes in the same group and sparse connections between nodes in different groups. However, with the advancement of computer science, more information has been collected and added to networks, giving rise to attributed networks. Unlike traditional networks, which only contain topological information, attributed networks contain a variety of attribute information, making the previous requirements for community structure no longer applicable.

[0005] Compared to traditional communities, attribute networks require not only close connectivity within each community but also a high degree of attribute similarity among nodes within the same community. Due to the heterogeneity of adjacency structure and attribute information, traditional graph clustering methods are difficult to directly transfer to attribute networks. Therefore, it is necessary to research new methods to eliminate the heterogeneity between the two and achieve better clustering results. Therefore, community detection algorithms on attribute networks require significant research effort, hoping to achieve new breakthroughs.

[0006] In addition to using attribute information, clustering, as a typical unsupervised learning method, has always suffered from lower accuracy compared to supervised classification algorithms. In view of this, it is necessary to provide a small-sample active learning method for attribute network clustering to solve the small-sample clustering problem in attribute networks. Summary of the Invention

[0007] An embodiment of the present invention provides a small-sample active learning method for attribute network clustering, which mainly solves three problems: first, how to use a small amount of supervised information to improve algorithm performance when supervised information is severely insufficient; second, how to use active learning to add labeling information to further stabilize and improve algorithm performance when supervised information is insufficient; third, how to reduce the additional time overhead caused by the active learning algorithm when the network scale is large.

[0008] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0009] First, a small-sample active learning method for attribute network clustering is provided, comprising:

[0010] Processing source data: Obtaining the adjacency matrix, attribute matrix, and indicator matrix of the attribute network. The indicator matrix is used to store the label information of the nodes marked by experts, that is, the category information of the nodes;

[0011] Small sample learning: Using the indicator matrix to record the category information of nodes in the attribute network, and combining the attribute matrix to select a representative node set of each category; obtaining the global clustering results and community cluster prototype attributes of the attribute network nodes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative nodes, and the indicator matrix of the representative nodes;

[0012] Active learning: Characterize node uncertainty from two levels: the adjacency matrix and the global clustering result. Combine the product of the above two types of node uncertainties to obtain the total uncertainty of each node in the attribute network; select the node with the highest uncertainty from the total uncertainty and extract the label information of the node; obtain the adjacency relationship of the marked node set from the adjacency matrix, select all neighbor nodes directly connected to the marked node, and construct a subnetwork; use the adjacency matrix and attribute matrix of the subnetwork to obtain the clustering result of the subnetwork nodes;

[0013] Local update semi-supervised clustering: If the maximum number of uncertain node queries is reached, the latest global clustering result is output; otherwise, active learning is returned to reselect new uncertainty nodes and update the clustering results of the sub-network, while updating the clustering results of the global clustering result sub-network.

[0014] In conjunction with the first aspect, in a possible implementation, recording category information of nodes in the attribute network using the indicator matrix includes:

[0015] Determining whether the indicator matrix is sufficient;

[0016] If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix;

[0017] If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

[0018] In conjunction with the first aspect, in a possible implementation, selecting a representative node set of each category in combination with the attribute matrix includes:

[0019] Calculate the Euclidean distance between the attributes of each node in each community and the central attribute, and select several nodes with the smallest Euclidean distance as the representative node set of the community.

[0020] In conjunction with the first aspect, in one possible implementation, obtaining the global clustering results and community cluster prototype attributes of the attribute network nodes by using the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative node, and the indicator matrix of the representative node includes:

[0021] Using the attribute information and category information of the representative node set of each community, the cluster prototype attribute information of each community is obtained;

[0022] By utilizing the adjacency matrix and the attribute matrix of the attribute network and combining the cluster prototype attribute information, a global clustering result of each node in the attribute network is obtained through a given optimization rule.

[0023] In conjunction with the first aspect, in one possible implementation, characterizing node uncertainty from two levels, the adjacency matrix and the global clustering result, and obtaining the total uncertainty of each node in the attribute network by combining the product of the above two types of node uncertainties include:

[0024] Based on the global clustering result and the uncertainty calculation method of the adjacency matrix, the uncertainty M of each node is obtained from the adjacency matrix using a method based on community modularity;

[0025] Calculating the uncertainty η of each node based on the Shannon theorem from the global clustering result;

[0026] The product of each node uncertainty M and each node uncertainty η is calculated as the total uncertainty Ω of the node.

[0027] In conjunction with the first aspect, in a possible implementation, selecting a node with the highest uncertainty from the total uncertainty and extracting label information of the node includes:

[0028] Uncertainty of the node Ω i It is considered to be the product of the uncertainty based on the adjacency matrix and the uncertainty based on the global clustering result, the uncertainty of each node is arranged from large to small, and the node corresponding to the maximum uncertainty is selected as the marked node;

[0029] For the marked nodes, the experts give the correct classification of the marked nodes, so that the number of marked nodes increases.

[0030] Secondly, a small sample active learning system for attribute network clustering is provided, including:

[0031] The source data processing module is used to obtain the adjacency matrix, attribute matrix and indicator matrix of the attribute network. The indicator matrix is used to store the label information of the nodes marked by the experts, that is, the category information of the nodes;

[0032] The small sample learning module uses the indicator matrix to record the category information of the nodes in the attribute network, and selects a representative node set of each category in combination with the attribute matrix; obtains the global clustering results and community cluster prototype attributes of the attribute network nodes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative nodes, and the indicator matrix of the representative nodes;

[0033] An active learning module is used to characterize node uncertainty from two levels: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by combining the product of the above two types of node uncertainties; the node with the highest uncertainty is selected from the total uncertainty, and the label information of the node is extracted; the adjacency relationship of the marked node set is obtained from the adjacency matrix, and all neighbor nodes directly connected to the marked node are selected to construct a subnetwork; the clustering result of the subnetwork nodes is obtained using the adjacency matrix and attribute matrix of the subnetwork;

[0034] The local update semi-supervised clustering module uses the semi-supervised clustering method. If the maximum number of uncertain nodes that can be queried is reached, the latest global clustering result is output; otherwise, the active learning is returned to reselect new uncertainty nodes and update the clustering results of the sub-network, while also updating the global clustering results.

[0035] In conjunction with the second aspect, in a possible implementation, in the small sample learning module, using the indicator matrix to record category information of nodes in the attribute network includes:

[0036] Determining whether the indicator matrix is sufficient;

[0037] If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix;

[0038] If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

[0039] In conjunction with the second aspect, in a possible implementation, in the small sample learning module, selecting a representative node set of each category in combination with the attribute matrix includes:

[0040] Calculate the Euclidean distance between the attributes of each node in each community and the central attribute, and select several nodes with the smallest Euclidean distance as the representative node set of the community.

[0041] In conjunction with the second aspect, in one possible implementation, in the small sample learning module, the global clustering results of the attribute network nodes and the community cluster prototype attributes are obtained through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative node, and the indicator matrix of the representative node, including:

[0042] Using the attribute information and category information of the representative node set of each community, the cluster prototype attribute information of each community is obtained;

[0043] By utilizing the adjacency matrix and the attribute matrix of the attribute network and combining the cluster prototype attribute information, a global clustering result of each node in the attribute network is obtained through a given optimization rule.

[0044] In conjunction with the second aspect, in one possible implementation, in the active learning module, node uncertainty is characterized from two levels: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by multiplying the above two types of node uncertainties, including:

[0045] Based on the global clustering result and the uncertainty calculation method of the adjacency matrix, the uncertainty M of each node is obtained from the adjacency matrix using a method based on community modularity;

[0046] Calculating the uncertainty η of each node based on the Shannon theorem from the global clustering result;

[0047] The product of each node uncertainty M and each node uncertainty η is calculated as the total uncertainty Ω of the node.

[0048] In conjunction with the second aspect, in a possible implementation, the active learning module selects a node with the highest uncertainty from the total uncertainty and extracts label information of the node; including:

[0049] Uncertainty of the node Ω i It is considered to be the product of the uncertainty based on the adjacency matrix and the uncertainty based on the global clustering result, the uncertainty of each node is arranged from large to small, and the node corresponding to the maximum uncertainty is selected as the marked node;

[0050] For the marked nodes, the experts give the correct classification of the marked nodes, so that the number of marked nodes increases.

[0051] In a third aspect, a small sample active learning device for attribute network clustering is provided, characterized by comprising:

[0052] A source data processing unit is used to obtain the adjacency matrix, attribute matrix and indicator matrix of the attribute network, wherein the indicator matrix is used to store the label information of the nodes marked by the experts, that is, the category information of the nodes;

[0053] The small sample learning unit uses the indicator matrix to record the category information of the nodes in the attribute network, and selects a representative node set of each category in combination with the attribute matrix; obtains the global clustering results and community cluster prototype attributes of the attribute network nodes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative nodes, and the indicator matrix of the representative nodes;

[0054] An active learning unit is configured to characterize node uncertainty from two levels: the adjacency matrix and the global clustering result; obtain the total uncertainty of each node in the attribute network by combining the product of the above two types of node uncertainties; select the node with the highest uncertainty from the total uncertainty and extract the label information of the node; obtain the adjacency relationship of the set of marked nodes from the adjacency matrix, select all neighbor nodes directly connected to the marked node, and construct a subnetwork; and obtain the clustering results of the subnetwork nodes using the adjacency matrix and attribute matrix of the subnetwork;

[0055] Locally update the semi-supervised clustering unit and use the semi-supervised clustering method. If the maximum number of uncertain nodes can be queried is reached, the latest global clustering result is output; otherwise, return to active learning to reselect new uncertainty nodes and update the clustering results of the sub-network, and update the global clustering results at the same time. Clustering results of sub-network Clustering results of sub-network

[0056] In conjunction with the third aspect, in a possible implementation, in the small sample learning unit, using the indicator matrix to record category information of nodes in the attribute network includes:

[0057] Determining whether the indicator matrix is sufficient;

[0058] If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix;

[0059] If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

[0060] In conjunction with the third aspect, in a possible implementation, in the small sample learning module, selecting a representative node set of each category in combination with the attribute matrix includes:

[0061] Calculate the Euclidean distance between the attributes of each node in each community and the central attribute, and select several nodes with the smallest Euclidean distance as the representative node set of the community.

[0062] In conjunction with the third aspect, in one possible implementation, in the small sample learning module, the global clustering results of the attribute network nodes and the community cluster prototype attributes are obtained through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative node, and the indicator matrix of the representative node, including:

[0063] Using the attribute information and category information of the representative node set of each community, the cluster prototype attribute information of each community is obtained;

[0064] By utilizing the adjacency matrix and the attribute matrix of the attribute network and combining the cluster prototype attribute information, a global clustering result of each node in the attribute network is obtained through a given optimization rule.

[0065] In conjunction with the third aspect, in one possible implementation, in the active learning module, node uncertainty is characterized from two levels: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by multiplying the above two types of node uncertainties, including:

[0066] Based on the global clustering result and the uncertainty calculation method of the adjacency matrix, the uncertainty M of each node is obtained from the adjacency matrix using a method based on community modularity;

[0067] Calculating the uncertainty η of each node based on the Shannon theorem from the global clustering result;

[0068] The product of each node uncertainty M and each node uncertainty η is calculated as the total uncertainty Ω of the node.

[0069] In conjunction with the third aspect, in a possible implementation, the active learning module selects a node with the highest uncertainty from the total uncertainty and extracts label information of the node; including:

[0070] Uncertainty of the node Ω i It is considered to be the product of the uncertainty based on the adjacency matrix and the uncertainty based on the global clustering result, the uncertainty of each node is arranged from large to small, and the node corresponding to the maximum uncertainty is selected as the marked node;

[0071] For the marked nodes, the experts give the correct classification of the marked nodes, so that the number of marked nodes increases.

[0072] The present disclosure has at least the following technical effects or advantages:

[0073] 1. The embodiments of the present invention adopt prototype network technology and actively learn to effectively solve the small sample learning problem in clustering problems, thereby achieving improvements in semi-supervised clustering results under the premise of small samples.

[0074] 2. The small sample learning method used in the embodiments of the present invention is a learning method that can still achieve good results even when supervised information is severely insufficient. The embodiments of the present invention adopt a small sample learning method based on the prototype network strategy, using prior information or the KMEANS algorithm to select prototypes of different clusters, and guiding clustering through the relationship between each node and the cluster prototype, thereby obtaining more accurate clustering results.

[0075] 3. The present invention uses active learning (a strategy that uses an algorithm to select labeled nodes) to proactively select samples to be labeled and obtain supervisory information for further learning. The present invention utilizes community modularity and Shannon's theorem to select nodes to be labeled, characterizing node uncertainty from both structural and node membership perspectives. This derives the total uncertainty for each node and labels the node with the greatest uncertainty.

[0076] 4. Because the active learning of the embodiment of the present invention has a limit on the number of queries, and the model needs to be further learned before each query, it will take more time. In order to reduce time overhead, the embodiment of the present invention adopts a local update semi-supervised clustering algorithm. More specifically, the embodiment of the present invention adopts a method of locally updating the information of nodes around the marked node to update the clustering results after each query. The algorithm uses cluster prototypes to obtain the attribute preferences of each cluster, and combines the marked nodes obtained in active learning to update the clustering results of the nodes adjacent to the marked nodes. Except for overly dense networks, this algorithm can greatly reduce the time complexity of the clustering algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments of the present invention or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0078] Figure 1 A flow chart of a small sample active learning method for attribute network clustering provided by an embodiment of the present disclosure;

[0079] Figure 2 An organizational chart of a small sample active learning method for attribute network clustering provided by an embodiment of the present disclosure;

[0080] Figure 3 This is a principle block diagram of the active learning semi-supervised clustering algorithm based on a small sample attribute network provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0081] The present disclosure is described in detail below with reference to the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present disclosure, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in this field based on these embodiments are all within the scope of protection of the present disclosure.

[0082] See also Figure 1 The present disclosure provides a small sample active learning method for attribute network clustering, including:

[0083] Process source data: obtain the adjacency matrix, attribute matrix and indicator matrix of the attribute network, where the indicator matrix is used to store the label information of the nodes marked by experts.

[0084] In practical applications, the attribute network mentioned above is recorded as G = (V, E, F, L), where V = {v1, ..., v n} is the set of nodes, n represents the number of nodes, E={(v i , v j )} is the set of edges, The attribute set corresponding to the node set V, L = {y1, ..., y o} is the existing supervision information.

[0085] The adjacency matrix described above can be the topological structure of the attribute network, which can reflect the adjacency relationships of the attribute network. The attribute matrix described above can be the attribute information set of all nodes in the attribute network. The indicator matrix described above is obtained using the label information in the attribute network, which indicates which class the node belongs to. Obtaining the adjacency matrix, attribute matrix, and indicator matrix of the attribute network can also be referred to as obtaining the adjacency relationships, node attributes, and label information of the network.

[0086] See also Figure 3 In the embodiment of the present disclosure, a network is constructed from source data and communities are obtained through network analysis. Communities are obtained through network analysis based on matrix decomposition methods, such as prototype networks, active learning, and joint matrix decomposition.

[0087] Please continue reading Figure 1 and Figure 2 In the small sample learning method of the disclosed embodiment, the indicator matrix is used to record the category information of nodes in the attribute network. A representative set of nodes for each category is selected in combination with the attribute matrix. The global clustering results and community cluster prototype properties of the attribute network nodes are obtained by combining the adjacency matrix, the attribute matrix, the indicator matrix, and the attribute matrix of the representative nodes.

[0088] See also Figure 2 , the indicator matrix is used to record the category information of nodes in the attribute network, including:

[0089] Determine whether the indicator matrix is sufficient;

[0090] If the judgment result is sufficient, the average value of the attributes of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the indicator matrix;

[0091] If the judgment result is insufficient, the KMeans algorithm is used to cluster the network attribute matrix to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the indicator matrix is used to record the category information of the b nodes.

[0092] The above-mentioned global clustering results of attribute network nodes and community cluster prototype attributes are obtained through the adjacency matrix, attribute matrix, indicator matrix, attribute matrix of representative nodes, and indicator matrix of representative nodes, including:

[0093] Using the attribute information and category information of the representative node set of each community, the cluster prototype attribute information of each community is obtained;

[0094] By using the adjacency matrix and attribute matrix of the attribute network and combining the cluster prototype attribute information, the global clustering results of each node in the attribute network are obtained through given optimization rules.

[0095] The global clustering results from the small-sample learning described above also become a membership matrix. Using existing label information or label information preprocessed using KMeans, the attribute averages of the labeled non-edge nodes of each community are calculated. The Euclidean distance between each node in the community and the central attribute is calculated, and the nodes with the smallest Euclidean distance are selected as the representative node set of the community. Using the attribute and category information of the representative node set of each community, the prototype attribute information of the community cluster is obtained. Using the adjacency and attribute information of the attribute network, combined with the cluster prototype attribute information obtained in the previous step, the node membership matrix is obtained using a given optimization rule.

[0096] The small sample active learning method of the embodiment of the present disclosure mainly includes semi-supervised clustering and active learning methods. The basic framework is to use the semi-supervised clustering method to obtain preliminary clustering results, that is, the membership matrix U. After the semi-supervised learning method obtains preliminary results, the membership value U of the critical subgraph of the marked node is updated by the active learning method. s , and finally added to the membership matrix U to obtain the final node clustering result.

[0097] Given the attribute network G with minimal supervision information in L, V is divided into k groups {C1, ..., C k The embodiment of the present invention is a hard-partitioned clustering problem, which requires that the intersection of different groups is an empty set, that is, In addition, based on the requirements for cluster homogeneity and compactness, nodes in the same cluster also require higher homogeneity and a tighter topological structure.

[0098] In the disclosed embodiments, the indicators of homogeneity and topological structure tightness are not directly defined. Instead, the algorithm uses a small amount of supervised information to extract the cluster prototype of each community, determine the attribute preferences of the community, and use this as the center to determine the classification results of each result.

[0099] Due to the high cost of labeling nodes, for the small sample attribute network clustering problem, only a few or no nodes in each category can be labeled by experts. (1) When the number of labeled nodes meets the requirements and the experts believe that the node set is not an edge node set, the set can be directly used to select the most representative node subset for each category. Assume that the attribute network has k communities, each community has at least z non-edge nodes labeled, and its attribute average value is h. Calculate the Euclidean distance of each node in the community to h, and select the b nodes closest to h as the most representative node set of the cluster. Based on this set, the cluster prototype information of each category can be learned by non-negative matrix decomposition. (2) When experts are absent or labeling information is insufficient, use the KMeans algorithm to pre-determine k communities and select the center point attribute of each community as the average attribute h. On this basis, calculate the Euclidean distance of each node to h and select the b nodes closest to h. Then, based on this node set, learn the cluster prototype of each category. In this case, this algorithm can be regarded as zero-shot learning in small sample learning.

[0100] After performing the above operations on each community and extracting kb nodes, the attribute matrix is constructed using the attribute information and label information of the node set. and And use the non-negative matrix decomposition method to obtain the attribute information C∈R of each cluster prototype k×m The objective function is as follows:

[0101]

[0102] The small sample learning part uses the cluster prototype attribute information C selected in the first step and uses the joint non-negative matrix factorization method to obtain the node membership matrix for each cluster. In order to reduce the impact of sampling errors on node attributes, the method makes the learning process of cluster prototype attribute information also affected by the rest of the nodes. That is, let the network's cluster prototype attribute matrix U∈R n×k The node membership matrix C is affected by both the labeled node information and the global network information, which improves the clustering results:

[0103]

[0104] Among them, the adjacency matrix of the attribute network Property Matrix In order to obtain U and C that minimize the loss function, the following update methods U and C are designed based on the non-negative matrix factorization method:

[0105]

[0106]

[0107] Please continue reading Figure 1 and Figure 2 Active learning in the disclosed embodiment characterizes node uncertainty from two perspectives: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by multiplying the two types of node uncertainty. The node with the highest uncertainty is selected from the total uncertainty, and the label information of the node is extracted. A subnetwork is constructed by selecting all directly connected neighbor nodes from the adjacency relationships of all labeled nodes in the adjacency matrix. The clustering results of the subnetwork nodes are obtained using the adjacency matrix and attribute matrix of the subnetwork.

[0108] The above describes the node uncertainty from two levels: the adjacency matrix and the global clustering result. The product of the above two types of node uncertainties is combined to obtain the total uncertainty of each node in the attribute network, including:

[0109] Based on the uncertainty calculation method of global clustering results and adjacency matrix, the uncertainty M of each node is obtained from the adjacency matrix using the community modularity method;

[0110] Calculate the uncertainty η of each node based on the Shannon theorem from the global clustering results;

[0111] The product of each node uncertainty M and each node uncertainty η is calculated as the total uncertainty Ω of the node.

[0112] The above process selects the node with the highest uncertainty from the total uncertainty and extracts the label information of the node; including:

[0113] Uncertainty of the node Ω i It is considered to be the product of the uncertainty based on the adjacency matrix and the uncertainty based on the global clustering result. The uncertainty of each node is arranged from large to small, and the node corresponding to the maximum uncertainty is selected as the marked node;

[0114] For the marked nodes, the experts give the correct classification of the marked nodes, so that the number of marked nodes increases.

[0115] The above active learning combines the two node uncertainties obtained from the topological structure and the membership matrix to obtain the total uncertainty of the node; using the uncertainty of the node, the node with the highest uncertainty is selected and the label of the node is obtained; using the adjacency relationship of all labeled nodes, all directly connected neighbor nodes are selected to construct a sub-network; using the adjacency matrix and attribute matrix of the sub-network, the membership matrix of the sub-network nodes is obtained, requiring that the membership vector of the labeled node is as consistent as possible with the correct label.

[0116] The small-sample active learning method of an embodiment of the present invention includes semi-supervised clustering and active learning methods. The basic framework is to use the semi-supervised clustering method to obtain the global clustering result, namely the membership matrix U. After the semi-supervised learning method obtains preliminary results, the membership value U3 of the adjacent subgraph (adjacency matrix) of the marked node is updated through the active learning method (i.e., the clustering result of the subnetwork nodes). Finally, it is added to the membership matrix U to obtain the final node clustering result.

[0117] The embodiment of the present invention obtains the node membership matrix through a small sample, but due to the defects of the clustering algorithm itself and the characteristics of each data set, the clustering results of some nodes may show a high degree of uncertainty and are easily affected by noise and slight deviations in the initialization nodes. At the same time, the surrounding nodes of this type of node often have similar tendencies. To this end, the embodiment of the present invention proposes a corresponding algorithm to reduce the negative impact of this type of node on the algorithm, mainly through active learning methods to extract nodes with higher uncertainty and obtain their category information, and further update the neighboring nodes around this type of node.

[0118] The uncertainty of the above measurement nodes is as follows:

[0119] To discover nodes with high uncertainty, this embodiment of the present invention designs an uncertainty calculation method based on upstream clustering results and topological structure. Based on the topological structure, a method based on community modularity is used to obtain the node uncertainty M. Based on the upstream clustering results, the node uncertainty η is calculated based on Shannon's theorem. Finally, the product of the two is calculated as the total node uncertainty Ω.

[0120] (1) Uncertainty of nodes in topological structure

[0121] Given the adjacency matrix W, the node pair <V i , V j > Community Modularity Q ij The calculation formula is as follows:

[0122]

[0123] Among them, d i =∑ h W ih , represents the number of edges connecting a node in the network, also known as its degree. e represents the total number of edges in G.

[0124] Q ij Corresponding to the uncertainty of the edge, in order to calculate the uncertainty of each node, the uncertainty of the node in the topological structure is defined as follows:

[0125]

[0126] (2) Uncertainty of nodes in clustering results

[0127] Shannon's theorem obtains the uncertainty of each node by weighted mutual information. Given the membership matrix U ij , then node v i The uncertainty η i The calculation formula is as follows:

[0128]

[0129] (3) Total uncertainty of nodes

[0130] Due to the uncertainty Ω of the node in the embodiment of the present invention i It is considered to be the product of the uncertainty based on the topology structure and the uncertainty based on the membership matrix, and is calculated as follows:

[0131] Ω i =M i ×η i

[0132] When it is obtained After the value is obtained, the algorithm arranges the uncertainty of each node from large to small, and selects the node corresponding to the maximum uncertainty as the node that needs to be marked.

[0133] Please continue reading Figure 1 and Figure 2 In the locally updated semi-supervised clustering of the disclosed embodiment, if the maximum number of uncertain node queries is reached, the clustering results of the sub-network are used to update the global clustering results of the global network. Otherwise, active learning is returned to reselect new uncertainty nodes and update the clustering results of the sub-network, while also updating the global clustering results.

[0134] In other words, in the above-mentioned local update semi-supervised clustering, if the maximum number of uncertain nodes can be queried is reached, the membership matrix of the sub-network is used to update the membership matrix of the global network. Conversely, a new maximum uncertain node is selected and the sub-network membership matrix is updated. The disclosed embodiment also uses the global membership matrix to classify nodes into corresponding categories.

[0135] The above semi-supervised clustering includes:

[0136] After selecting the node with the greatest uncertainty, the expert gives the correct classification for that node, increasing the number of labeled nodes. Because the active learning algorithm's model needs to update the clustering results after each query before the next query can be performed, the algorithm's time complexity will increase significantly when the time overhead of a single query is high.

[0137] Assume that the number of queries that can be made in active learning is γ. There are currently two common solutions to this problem. The first strategy is to select γ nodes to be marked at one time based on the existing uncertainty, and then re-update the clustering results of the entire network. The advantage of this method is that it can reduce the large amount of time overhead caused by multiple queries, and the time overhead of active learning for γ times is reduced to the time overhead of a single active learning. The disadvantage is that the results of each active learning cannot be used to select the node with the maximum uncertainty for the next time. It may happen that a group of nodes with multiple high uncertainties are selected at one time to complete the active learning process. The second method is to locally update the clustering results of the nodes surrounding the marked node. Unlike the former, this strategy does not correct the clustering results of the entire network, but updates the neighboring nodes around the marked node, thereby reducing the time spent on each query.

[0138] Since the second method can use the results of the previous query to improve the next query, the embodiment of the present invention selects the method of locally updating the clustering results of the nodes around the marked node, and the target loss function used is as follows:

[0139]

[0140] in, and The adjacency matrix, membership matrix and attribute matrix of the surrounding nodes corresponding to the marked node, ψ represents the total number of marked nodes and their neighboring nodes; R is used to select the rows corresponding to the marked nodes in Uh, and the rows of the marked nodes can be used to compare with the indicator matrix of the marked nodes For comparison, 1 represents the total number of nodes with a marker.

[0141] To obtain the local optimal solution of the algorithm, update U as follows: h

[0142]

[0143] When the number of active queries γ is used up, the active learning algorithm obtains the membership matrix U3. The algorithm uses U3 to update the membership matrix U learned from the small sample, and obtains the final clustering result through the updated U.

[0144] Please continue reading Figure 1 and Figure 2 , the final clustering of the embodiment of the present disclosure: using the global clustering results of the attribute network and the clustering results of the sub-network to divide the nodes and obtain the clustering results.

[0145] The embodiments of the present invention have the following advantages:

[0146] Cluster prototype selection: This embodiment of the present invention screens the prototype nodes that may become each community in the attribute network through analysis of topological structure and attribute information, and designs two strategies for the presence or absence of expert participation as the basis for small sample learning.

[0147] Small sample clustering: Based on the cluster prototype idea of the prototype network, combined with the topological structure and node attribute information in the attribute network, it guides the node clustering process in the attribute network and obtains the membership matrix of the entire network nodes to each community, thereby obtaining the global clustering result of the network.

[0148] Uncertainty Assessment: Based on the results of the upstream clustering algorithm, the uncertainty of the clustering results is calculated using Shannon's theorem. In addition, the model also extracts the uncertainty of each node's topological structure from the community modularity, and uses the product of the two as the comprehensive node uncertainty.

[0149] Active learning clustering method with local update: Based on the existing cluster prototype attributes and prior supervision information, only the membership matrix of the neighboring nodes of the nodes with reliable results (existing supervision information) is updated, thereby improving the accuracy of the model without increasing the time complexity of the algorithm.

[0150] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present disclosure. They are not intended to limit the scope of protection of the present disclosure. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present disclosure should be included in the scope of protection of the present disclosure.

[0151] It will be apparent to those skilled in the art that the present disclosure is not limited to the details of the exemplary embodiments described above and that the present disclosure can be implemented in other specific forms without departing from the spirit or essential characteristics of the present disclosure. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present disclosure is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed within the present disclosure. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0152] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A small sample active learning method for attribute network clustering, characterized by: Perform community detection on attribute networks, including: Processing source data: Obtaining the adjacency matrix, attribute matrix, and indicator matrix of the attribute network. In social networks, communities correspond to groups of people with similar backgrounds or interests. By analyzing the commonalities of these groups, content recommendation algorithms can accurately push appropriate information to users who may be interested. In cancer networks, communities correspond to a class of genes that express the same or similar functions. The indicator matrix is used to store the label information of nodes marked by experts, that is, the node category information. Small sample learning: Using the indicator matrix to record the category information of nodes in the attribute network, and combining the attribute matrix to select a representative node set of each category; obtaining the global clustering results and community cluster prototype attributes of the attribute network nodes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative nodes, and the indicator matrix of the representative nodes; Active learning: Characterize node uncertainty from two levels: the adjacency matrix and the global clustering result. Combine the product of the above two types of node uncertainties to obtain the total uncertainty of each node in the attribute network; select the node with the highest uncertainty from the total uncertainty and extract the label information of the node; obtain the adjacency relationship of the marked node set from the adjacency matrix, select all neighbor nodes directly connected to the marked node, and construct a subnetwork; use the adjacency matrix and attribute matrix of the subnetwork to obtain the clustering result of the subnetwork nodes; Local update semi-supervised clustering: If the maximum number of uncertain node queries is reached, the latest global clustering result is output; otherwise, active learning is returned to reselect new uncertainty nodes and update the clustering results of the sub-network, while also updating the global clustering results.

2. The small sample active learning method according to claim 1, characterized in that The step of recording the category information of nodes in the attribute network by using the indicator matrix includes: Determining whether the indicator matrix is sufficient; If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix; If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

3. The small sample active learning method according to claim 1, characterized in that The selecting of representative node sets of each category in combination with the attribute matrix includes: Calculate the Euclidean distance between the attributes of each node in each community and the central attribute, and select several nodes with the smallest Euclidean distance as the representative node set of the community.

4. The small sample active learning method according to claim 1, characterized in that The method of obtaining the global clustering results of attribute network nodes and community cluster prototype attributes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative node, and the indicator matrix of the representative node includes: Using the attribute information and category information of the representative node set of each community, the cluster prototype attribute information of each community is obtained; By utilizing the adjacency matrix and the attribute matrix of the attribute network and combining the cluster prototype attribute information, a global clustering result of each node in the attribute network is obtained through a given optimization rule.

5. The small sample active learning method according to claim 1, characterized in that The node uncertainty is characterized from two levels: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by combining the product of the above two types of node uncertainties, including: Based on the global clustering result and the uncertainty calculation method of the adjacency matrix, the uncertainty M of each node is obtained from the adjacency matrix using a method based on community modularity; Calculating the uncertainty η of each node based on the Shannon theorem from the global clustering result; The product of each node uncertainty M and each node uncertainty η is calculated as the total uncertainty Ω of the node.

6. The small sample active learning method according to claim 5, characterized in that The step of selecting a node with the highest uncertainty from the total uncertainty and extracting label information of the node comprises: Uncertainty of the node Ω i It is considered to be the product of the uncertainty based on the adjacency matrix and the uncertainty based on the global clustering result, the uncertainty of each node is arranged from large to small, and the node corresponding to the maximum uncertainty is selected as the marked node; For the marked nodes, the experts give the correct classification of the marked nodes, so that the number of marked nodes increases.

7. A small sample active learning system for attribute network clustering, characterized by: Perform community detection on attribute networks, including: The source data processing module is used to obtain the adjacency matrix, attribute matrix, and indicator matrix of the attribute network. In social networks, communities correspond to groups of people with similar backgrounds or hobbies. By analyzing the commonalities of these groups, the content recommendation algorithm can accurately push appropriate information to users who may be interested. In cancer networks, communities correspond to a class of genes that express the same or similar functions. The indicator matrix is used to store the label information of nodes marked by experts, that is, the node category information. A small sample learning module selects a representative node set of each category using the indicator matrix combined with the attribute matrix; obtains the global clustering results of attribute network nodes and community cluster prototype attributes through the adjacency matrix, the attribute matrix, the indicator matrix, the attribute matrix of the representative nodes, and the indicator matrix of the representative nodes; An active learning module is used to characterize node uncertainty from two levels: the adjacency matrix and the global clustering result. The total uncertainty of each node in the attribute network is obtained by combining the product of the above two types of node uncertainties; the node with the highest uncertainty is selected from the total uncertainty, and the label information of the node is extracted; the adjacency relationship of the marked node set is obtained from the adjacency matrix, and all neighbor nodes directly connected to the marked node are selected to construct a subnetwork; the clustering result of the subnetwork nodes is obtained using the adjacency matrix and attribute matrix of the subnetwork; The local update semi-supervised clustering module uses the semi-supervised clustering method. If the maximum number of uncertain node queries is reached, the clustering results of the sub-network are used to update the global clustering results of the global network; otherwise, active learning is returned to re-select new uncertainty nodes and update the clustering results of the sub-network, while also updating the global clustering results.

8. The small sample active learning system according to claim 7, characterized in that: In the small sample learning module, the method for recording the category information of nodes in the attribute network using the indicator matrix includes: Determine whether the category information is sufficient; If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix; If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

9. A small sample active learning device for attribute network clustering, characterized in that: Perform community detection on attribute networks, including: A source data processing unit is used to obtain the adjacency matrix, attribute matrix, and indicator matrix of the attribute network. In social networks, communities correspond to groups of people with similar backgrounds or hobbies. By analyzing the commonalities of these groups, the content recommendation algorithm can accurately push appropriate information to users who may be interested. In cancer networks, communities correspond to a class of genes that express the same or similar functions. The indicator matrix is used to store the label information of nodes marked by experts, that is, the node category information. The small sample learning unit selects a representative node set of each category using the indicator matrix combined with the attribute matrix; obtains the global clustering results of the attribute network nodes and the community cluster prototype attributes through the adjacency matrix, the attribute matrix, the indicator matrix and the attribute matrix of the representative nodes and the indicator matrix of the representative nodes; An active learning unit is configured to characterize node uncertainty from two levels: the adjacency matrix and the global clustering result; obtain the total uncertainty of each node in the attribute network by combining the product of the above two types of node uncertainties; select the node with the highest uncertainty from the total uncertainty and extract the label information of the node; obtain the adjacency relationship of the set of marked nodes from the adjacency matrix, select all neighbor nodes directly connected to the marked node, and construct a subnetwork; and obtain the clustering results of the subnetwork nodes using the adjacency matrix and attribute matrix of the subnetwork; The semi-supervised clustering unit is locally updated. If the maximum number of uncertain node queries is reached, the latest global clustering result is output; otherwise, active learning is returned to reselect new uncertainty nodes and update the clustering results of the sub-network, while also updating the clustering results of the global clustering result sub-network.

10. The small sample active learning device according to claim 9, characterized in that: In the small sample learning unit, the indicator matrix is used to record the category information of the nodes in the attribute network, including: Determining whether the indicator matrix is sufficient; If the judgment result is sufficient, the attribute average of each community is calculated, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix; If the judgment result is insufficient, the network attribute matrix is clustered using the KMeans algorithm to obtain the center point of each community, and the b nodes closest to the center of each community are extracted, and the category information of the b nodes is recorded using the representative node indicator matrix.

Images