Financial heterogeneous network community discovery method combining comparative learning and generative learning based on matching rate

Abstract

The invention relates to the technical field of heterogeneous networks, in particular to a financial heterogeneous network community discovery method of combined comparative learning and generative learning based on a matching rate, and the method comprises the steps: obtaining a target financial heterogeneous network; the target financial heterogeneous network is input into a preset community discovery model, a community discovery result of the target financial heterogeneous network is output, the community discovery model is used for constructing the target financial heterogeneous network into a high-order fusion graph and obtaining node representation of the high-order fusion graph by constructing the target financial heterogeneous network into a semantic tree, and the node representation of the high-order fusion graph is used for obtaining the community discovery result of the target financial heterogeneous network. And carrying out comparison learning and generation learning to obtain a community discovery result of the target financial heterogeneous network. Compared with the prior art, the community discovery method and model provided by the invention are more effective and reasonable.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous networks, and in particular to a method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning of the same matching rate. Background Art

[0002] As the complexity and diversity of financial markets continue to increase, community discovery technology in heterogeneous financial networks has gradually become an important tool for analyzing and understanding the financial system. This technology reveals the multi-level relationships and interaction patterns between financial entities by identifying the community structure of different types of nodes and edges in the network. In heterogeneous financial networks, the interaction between different entities such as banks, enterprises, and investors makes risk propagation, investment strategy optimization, and market behavior analysis more complicated. Therefore, the use of community discovery methods can effectively monitor financial risks, identify potential fraud, and provide support for investment decisions, thereby improving the stability and transparency of the financial system.

[0003] In the process of community discovery in heterogeneous networks, existing heterogeneous network community discovery methods often extract different meta-paths and treat them as homogeneous networks, and then use common methods such as GNN in homogeneous networks to learn node representations and then perform community discovery. The GNN, HGNN, etc. that the above methods rely on are based on the homogeneity assumption, which assumes that nodes of the same category in the network are more likely to be connected together. Therefore, the above methods often work better on homogeneous heterogeneous networks, but less well on heterogeneous networks with different homogeneity.

[0004] From interpersonal relationships in life to scientific research on protein genes, heterogeneous networks are also ubiquitous in life. Developing powerful models that can solve heterogeneity is a major research hotspot. Currently, researchers have proposed many methods to adjust GNNs so that they can be applied to heterogeneous graphs, promote the learning of graph structures by exploring the high-order structure of heterogeneous networks, or adaptively learn the weights of homogeneous and heterogeneous nodes, and jointly learn node attributes and topological structures. Although the above methods have alleviated the problem of heterogeneity in homogeneous networks to a certain extent, they rely on prior knowledge such as labels for training and cannot be applied to unsupervised community discovery. In addition, they cannot handle heterogeneous networks with multiple types of nodes and edges. Therefore, how to design an unsupervised community discovery algorithm for heterogeneous information networks with heterogeneous information is an important problem that needs to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide a financial heterogeneous network community discovery method based on joint contrastive learning and generative learning of the same matching rate, so as to solve the problems existing in the above-mentioned prior art and break through the limitations of the existing heterogeneous network community discovery algorithm.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio, including:

[0008] Acquire the target financial heterogeneous network;

[0009] The target financial heterogeneous network is input into a preset community discovery model, and a community discovery result of the target financial heterogeneous network is output, wherein the community discovery model is used to construct the target financial heterogeneous network as a high-order fusion graph, and obtain the node representation of the high-order fusion graph by constructing it as a semantic tree, and then perform comparative learning and generative learning to obtain the community discovery result of the target financial heterogeneous network.

[0010] Optionally, the community discovery model includes a neighborhood aggregation module, a contrastive learning module, a semantic tree construction module, and a generative learning module, wherein the neighborhood aggregation module is used to perform neighborhood aggregation on the target financial heterogeneous network to construct a high-order fusion graph; the contrastive learning module is used to perform contrastive learning on the original graph structure of the target financial heterogeneous network and the high-order fusion graph to obtain a community-level contrast loss; the semantic tree construction module is used to construct the high-order fusion graph into a semantic tree, and obtain a node representation of the high-order fusion graph through the semantic tree; the generative learning module is used to perform generative learning on the target financial heterogeneous network through the node representation, and obtain a community discovery result of the target financial heterogeneous network in combination with the community-level contrast loss.

[0011] Optionally, the neighborhood aggregation module performs neighborhood aggregation on the target financial heterogeneous network to construct a high-order fusion graph, including:

[0012] Decomposing the target financial heterogeneous network through meta-paths to obtain several homogeneous sub-networks;

[0013] Using neighborhood aggregation to extract information from the homogeneous sub-network to obtain common information;

[0014] The high-order fusion graph is constructed based on the common information.

[0015] Optionally, the high-order fusion graph is:

[0016]

[0017] Among them, S is a high-order fusion graph, φ is a meta-path, φ∈Φ, Φ is all meta-paths, is the graph structure corresponding to the meta-path φ, θ φ is the attenuation factor, A φ Represents the high-order adjacency matrix corresponding to the meta-path φ, od φrepresents the optimal order of high-order aggregation of the graph structure corresponding to the meta-path φ, α and β are adjustable parameters, and Ω is a matrix representing the first-order structural similarity of nodes.

[0018] Optionally, the community-level contrast loss is:

[0019]

[0020] in, represents the attribute matrix of the first view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the community prototype matrix of the second view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the attribute matrix of the second view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the community prototype matrix of the first view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion.

[0021] Optionally, the semantic tree construction module constructs the high-order fusion graph into a semantic tree, including:

[0022] Define the structural entropy of the semantic tree;

[0023] The high-order fusion graph is constructed into an optimal quality semantic tree by minimizing the structural entropy, wherein the optimal quality semantic tree is:

[0024]

[0025] Among them, T * represents the best quality semantic tree, height(T) represents the height of the semantic tree T, K represents the depth of the semantic tree T, H T (S) represents the structural entropy of the semantic tree T corresponding to the graph S.

[0026] Optionally, the generation learning module performs generation learning on the target financial heterogeneous network through the node representation, including:

[0027] A masked autoencoder is used to mask and reconstruct the target financial heterogeneous network in terms of structure and attributes to obtain the structure reconstruction loss and the attribute reconstruction loss.

[0028] Optionally, the structure reconstruction loss is:

[0029] L Edge =∑ φ∈{Φ∪S} α φ ·l φ ;

[0030] Among them, L Edge represents the structural reconstruction loss, α φrepresents the weight of the meta-path φ, l φ Represents the edge reconstruction loss of the meta-path φ.

[0031] Optionally, the attribute reconstruction loss is:

[0032]

[0033] Among them, L Feat represents the attribute reconstruction loss, Represents the attribute matrix after masking, Z v Represents the node attributes reconstructed using the decoder. represents the edge matrix after masking, and γ2 represents the corresponding scaling factor.

[0034] The beneficial effects of the present invention are:

[0035] The present invention proposes a neighborhood aggregation module based on the same matching rate, which expresses the network in a finer granularity and constructs a more refined high-order fusion graph by extracting the common information of homogeneous sub-networks with different same matching rates; proposes a contrastive learning module based on heterogeneous structures, which helps to promote the model to generate more discriminative node representations; proposes a semantic tree construction module, which helps to mine the hierarchical structure of the tree between meta-paths and learn stable node representations; proposes a generative learning module, which uses a heterogeneous network masked autoencoder to achieve the purpose of dynamically enhancing the learning of target type node attributes and heterogeneous network semantics. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be 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.

[0037] Figure 1 This is a flow chart of a method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning of the same matching rate according to an embodiment of the present invention;

[0038] Figure 2 This is a graph showing the results of the method on the real dataset DBLP according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] This embodiment provides a method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning of the same matching rate, including:

[0042] Acquire the target financial heterogeneous network;

[0043] The target financial heterogeneous network is input into a preset community discovery model, and the community discovery result of the target financial heterogeneous network is output, wherein the community discovery model is used to construct the target financial heterogeneous network into a high-order fusion graph, and obtain the node representation of the high-order fusion graph by constructing it into a semantic tree, and then perform contrastive learning and generative learning to obtain the community discovery result of the target financial heterogeneous network. The community discovery model specifically includes a neighborhood aggregation module, a contrastive learning module, a semantic tree construction module, and a generative learning module, wherein the neighborhood aggregation module is used to perform neighborhood aggregation on the target financial heterogeneous network to construct a high-order fusion graph; the contrastive learning module is used to perform contrastive learning on the original graph structure of the target financial heterogeneous network and the high-order fusion graph to obtain the community-level contrast loss; the semantic tree construction module is used to construct the high-order fusion graph into a semantic tree, and obtain the node representation of the high-order fusion graph through the semantic tree; the generative learning module is used to perform generative learning on the target financial heterogeneous network through the node representation, and obtain the community discovery result of the target financial heterogeneous network in combination with the community-level contrast loss.

[0044] Specifically, this implementation proposes a neighborhood aggregation module based on the same matching rate, which expresses the network in a more fine-grained manner and constructs a more refined high-order fusion graph by extracting the common information of homogeneous sub-networks with different same matching rates; proposes a contrastive learning module based on heterogeneous structures, which helps to promote the model to generate more discriminative node representations; proposes a semantic tree construction module, which helps to mine the hierarchical structure of the tree between meta-paths and learn stable node representations; proposes a generative learning module, which uses a heterogeneous network masked autoencoder to achieve the purpose of dynamically enhancing the learning of target type node attributes and heterogeneous network semantics.

[0045] Combine the following Figure 1 , Figure 2A financial heterogeneous network community discovery method and a community discovery model based on joint contrastive learning and generative learning of the same matching rate proposed in this embodiment are described in detail. The community discovery model structure includes a neighborhood aggregation module, a contrastive learning module, a semantic tree construction module, and a generative learning module. The community discovery model workflow includes:

[0046] S1. The neighborhood aggregation module uses different neighborhood aggregation methods to extract the common information of multiple subgraphs from the homogeneous sub-networks obtained by meta-path decomposition of the target financial heterogeneous network, and constructs a refined high-order fusion graph;

[0047] S2. The contrastive learning module masks the original graph structure and calculates the community-level contrast loss of graph structures generated by different masks;

[0048] S3. The semantic tree construction module constructs the high-order fusion graph into a semantic tree with a stable structure, encodes the fusion graph from the perspective of the tree, and obtains a new node representation;

[0049] S4. The generative learning module uses a heterogeneous network masked autoencoder to mask and reconstruct the network from both structural and attribute perspectives, and discovers the target financial network community from a generative self-supervised perspective.

[0050] Specifically, the neighborhood aggregation module in S1 uses different neighborhood aggregation methods to obtain the homogeneous sub-networks obtained after the target financial heterogeneous network is decomposed by the meta-path, extracts the common information of multiple sub-graphs, and constructs a refined high-order fusion graph, including: decomposing the target financial heterogeneous network by the meta-path to obtain several homogeneous sub-networks; using neighborhood aggregation to extract information from the homogeneous sub-networks to obtain common information; and constructing the high-order fusion graph based on the common information.

[0051] In this embodiment, pseudo labels are used to calculate the matching rate of each view, and the matching rate is used to calculate the optimal order of high-order aggregation of each view. f , for Z f Clustering extracts community information as pseudo labels for model initialization. Then use the pseudo labels Z f To guide the learning of more sophisticated graph structures:

[0052]

[0053] Among them, the Ω matrix can be used to represent the first-order structural similarity of nodes.

[0054] Since different orders of the adjacency matrix A contain different structural information, the high-order graph structure can capture more relationships. The high-order neighborhood relationship is defined as:

[0055]

[0056] Among them, k represents the aggregation order of neighbor relationships, the meta-path φ∈Φ, A φ Represents the high-order adjacency matrix corresponding to the meta-path φ.

[0057] Since unsupervised community discovery cannot use the true labels of nodes, we consider using pseudo labels R * Calculate the HR for each view and use HR to calculate the optimal order of high-order aggregation for each view:

[0058]

[0059] in, is the floor operation, and δ is the coefficient for filtering low homogeneous graphs. A smaller order is assigned to low homogeneous graphs, and a higher order is assigned to high homogeneous graphs. Under the guidance of pseudo labels, each view is refined to obtain the refined graph structure:

[0060]

[0061] Among them, θ φ is the attenuation factor.

[0062] Finally, the refined graph structures obtained from all meta-paths are fused to generate a new high-order fused graph S, where α and β are adjustable parameters:

[0063]

[0064] Specifically, the contrastive learning module in S2 masks the original graph structure and calculates the community-level contrast loss of graph structures generated by different masks, including:

[0065] For a given heterogeneous graph G = (V, A, T V ,T E ,X,Φ), construct the meta-path-based adjacency matrix A for each meta-path φ∈Φ φ A cross-view and cross-community comparison mechanism is used to compare the node representation of one view with the community prototype of another view. Formally, the community comparison is performed by:

[0066]

[0067] Where m represents the current training epoch, ∈ is a hyperparameter, and the weight of the contrast loss is gradually increased during model training. i Represents the community to which the i-th node belongs. w(i,k) is the weight function of RBF. This maximizes the similarity of nodes in the same community and minimizes the similarity of nodes in different communities. The final community-level contrast loss of the meta-path view and the high-order fusion graph is as follows:

[0068]

[0069] in, represents the attribute matrix of the first view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Community prototype matrix representing the second view after graph augmentation.

[0070] Specifically, the semantic tree construction module in S3 constructs the high-order fusion graph into a semantic tree with a stable structure, encodes the fusion graph from the perspective of the tree, and obtains new node representations including:

[0071] The generated high-order fusion graph is constructed into a stable semantic tree. The number of nodes in the fusion graph is the same as the number of nodes in the semantic tree. By minimizing the instability of the graph structure, the optimal heterogeneous semantic tree is constructed, and the optimal representation of the nodes on the fusion graph can be obtained.

[0072] First, we define the structural entropy of the semantic tree: For a semantic tree T with a depth of K, the structural entropy H of the tree is defined as follows:

[0073]

[0074] Among them, H T (S) is the structural entropy of the semantic tree T corresponding to the graph S. The edge set is divided according to the non-leaf nodes. For the non-leaf node non-root node α, assuming that the number of its child nodes is N, α (i) is the i-th child node of node α. The root node is denoted as root, T root =T, after segmentation, multiple subsets are obtained [T α ,T α / T root ], the total weight of all edges in the set is w(α), and vol(S) is the sum of the degrees of all edges in the graph. - Represents the parent node of α. vol(T α ) represents T α The sum of the degrees of the nodes inside.

[0075] The semantic tree with the smallest structural entropy is defined as:

[0076]

[0077] Among them, height(T) represents the height of the semantic tree T. Using MLP as the encoder of the semantic tree f M (·), using the hierarchical structure of the semantic tree, the features of leaf nodes are initialized, and the features of non-leaf nodes are obtained by aggregating the features of their child nodes, that is, each non-leaf node is an abstraction of its child nodes.

[0078] Specifically, the generative learning module in S4 uses a heterogeneous network masked autoencoder to mask and reconstruct the network from both structural and attribute perspectives, and performs target financial network community discovery from the perspective of generative self-supervision, including: using a masked autoencoder to mask and reconstruct the target financial heterogeneous network from both structural and attribute perspectives, and obtaining structural reconstruction loss and attribute reconstruction loss.

[0079] After applying the mask, replace the attribute X with the masked adjacency matrix Input encoder f E (·) Get the learned node attributes.

[0080]

[0081] Then send it to the decoder f D (·) Generate node embedding after decoding Use the sigmoid activation function to get the reconstructed adjacency matrix A φ′ , calculate the reconstructed adjacency matrix and the original adjacency matrix A φ The reconstruction loss between them.

[0082]

[0083] Among them, l φ represents the edge reconstruction loss of the meta-path φ, and γ1 is the corresponding scaling factor. Since the importance of each meta-path is different, the semantic level attention mechanism is introduced here to calculate the importance score c of different meta-paths. φ , and then use the softmax function to φ Normalization is performed to obtain the weights of different meta-paths.

[0084]

[0085] Where W represents the weight matrix and b is the bias vector. The final structure reconstruction loss L for all meta-paths is Edge The calculation is as follows:

[0086] L Edge =∑ φ∈{Φ∪S} α φ ·l φ (17)

[0087] Among them, L Edge represents the structural reconstruction loss, α φ represents the weight of the meta-path φ, l φ Represents the edge reconstruction loss of the meta-path φ.

[0088] The attribute matrix after masking And the original adjacency matrix A are sent to the encoder to calculate the learned node attribute H3:

[0089]

[0090] Z=f D (A,H3) (19)

[0091] Use the decoder to obtain the reconstructed node attribute Z, and calculate the attribute reconstruction loss before and after the reconstruction of the masked node attributes.

[0092]

[0093] Among them, L Feat represents the attribute reconstruction loss, Represents the attribute matrix after masking, Z v Represents the node attributes reconstructed using the decoder. Represents the edge matrix after masking.

[0094] Furthermore, the model is optimized:

[0095] Final objective function It is defined as the multi-view contrast loss L cos , edge reconstruction loss l based on meta-path Edge , attribute reconstruction loss l Feat Weighted combination formation.

[0096] L = λ × L cos +μ×L Edge +η×L Feat (twenty one)

[0097] Among them, λ, μ, η are community-level contrast loss L cos , structure reconstruction loss L Edge , attribute reconstruction loss L Feat The weight coefficient of .

[0098] Experimental verification:

[0099] (1) Dataset:

[0100] In order to verify the effectiveness of the model, this embodiment uses four real-world heterogeneous datasets for experiments, namely ACM, DBLP, FIN, and IMDB datasets.

[0101] ACM. Common citation network dataset, including three types of nodes: author, paper, and topic.

[0102] DBLP. A common citation network dataset, including four types of nodes: authors, papers, terms, and locations.

[0103] FIN. Financial network dataset, including three types of nodes: investment entities, listed securities companies, and unlisted securities companies. The network is constructed based on stock time series information, and the stock market entity network is constructed using similarity information such as network attribute structure.

[0104] IMDB. A dataset of online movie networks, including nodes such as actors, directors, and plots.

[0105] (2) Comparison method:

[0106] Four models are selected for comparison with the HoCMAE model proposed in this embodiment. These comparison models are briefly introduced below.

[0107] 1) DFCN: This model uses a dynamic fusion module to fuse embeddings learned from both local and global perspectives, and combines it with a triple self-supervision mechanism for consensus representation learning.

[0108] 2) HGMAE: This model uses a dynamic masked autoencoder to learn node representations by encoding attributes, topology, and location information, and then performs community discovery.

[0109] 3)DDGAE: This model builds views based on high-order modularity and attribute information, and uses a deep multi-attention mechanism to perform community discovery by reconstructing topological structures and attributes.

[0110] 4) HAESF: This model uses graph autoencoders to aggregate heterogeneous network structure information, uses non-negative matrix factorization to learn network community semantic information, and jointly optimizes the two.

[0111] Compared with the performance of other models on four data sets, this model can achieve higher results in heterogeneous networks with heterogeneity and homogeneous heterogeneous networks. Table 1-4 shows the experimental results of this model and other comparison models. It can be seen that this model performs well overall and is better than other models overall.

[0112] Table 1

[0113]

[0114] Table 2

[0115]

[0116] Table 3

[0117]

[0118] Table 4

[0119]

[0120] This experiment is based on the real data DBLP, and the community discovery results of this method on the data set DBLP are obtained, as shown in Figure 2 As shown, it can be seen that the HoCMAE model proposed in this embodiment can clearly identify nodes in different communities, the nodes in the same community are condensed into a relatively dense cluster, and the boundaries between different communities are relatively clear.

[0121] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning of the same matching ratio, characterized in that: include: Acquire the target financial heterogeneous network; The target financial heterogeneous network is input into a preset community discovery model, and a community discovery result of the target financial heterogeneous network is output, wherein the community discovery model is used to construct the target financial heterogeneous network as a high-order fusion graph, and obtain the node representation of the high-order fusion graph by constructing it as a semantic tree, and then perform comparative learning and generative learning to obtain the community discovery result of the target financial heterogeneous network.

2. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 1 is characterized in that: The community discovery model includes a neighborhood aggregation module, a contrastive learning module, a semantic tree construction module, and a generative learning module, wherein the neighborhood aggregation module is used to perform neighborhood aggregation on the target financial heterogeneous network to construct a high-order fusion graph; the contrastive learning module is used to perform contrastive learning on the original graph structure of the target financial heterogeneous network and the high-order fusion graph to obtain a community-level contrast loss; the semantic tree construction module is used to construct the high-order fusion graph into a semantic tree, and obtain a node representation of the high-order fusion graph through the semantic tree; the generative learning module is used to perform generative learning on the target financial heterogeneous network through the node representation, and obtain a community discovery result of the target financial heterogeneous network in combination with the community-level contrast loss.

3. The method for discovering financial heterogeneous network communities based on joint contrastive learning and generative learning based on the same matching rate according to claim 2 is characterized in that: The neighborhood aggregation module performs neighborhood aggregation on the target financial heterogeneous network to construct a high-order fusion graph including: Decomposing the target financial heterogeneous network through meta-paths to obtain several homogeneous sub-networks; Using neighborhood aggregation to extract information from the homogeneous sub-network to obtain common information; The high-order fusion graph is constructed based on the common information.

4. The method for discovering financial heterogeneous network communities based on joint contrastive learning and generative learning based on the same matching rate according to claim 3 is characterized in that: The high-order fusion graph is: Among them, S is a high-order fusion graph, φ is a meta-path, φ∈Φ, Φ is all meta-paths, is the graph structure corresponding to the meta-path φ, θ φ is the attenuation factor, A φ Represents the high-order adjacency matrix corresponding to the meta-path φ, od φ represents the optimal order of high-order aggregation of the graph structure corresponding to the meta-path φ, α and β are adjustable parameters, and Ω is a matrix representing the first-order structural similarity of nodes.

5. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 2 is characterized in that: The community-level contrast loss is: in, represents the attribute matrix of the first view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the community prototype matrix of the second view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the attribute matrix of the second view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion, Represents the community prototype matrix of the first view of the homogeneous subnetwork corresponding to the meta-path φ after graph expansion.

6. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 2 is characterized in that: The semantic tree construction module constructs the high-order fusion graph into a semantic tree, including: Define the structural entropy of the semantic tree; The high-order fusion graph is constructed into an optimal quality semantic tree by minimizing the structural entropy, wherein the optimal quality semantic tree is: Among them, T * represents the best quality semantic tree, height(T) represents the height of the semantic tree T, K represents the depth of the semantic tree T, H T (S) represents the structural entropy of the semantic tree T corresponding to the graph S.

7. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 2 is characterized in that: The generation learning module generates and learns the target financial heterogeneous network through the node representation, including: A masked autoencoder is used to mask and reconstruct the target financial heterogeneous network in terms of structure and attributes to obtain the structure reconstruction loss and the attribute reconstruction loss.

8. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 7 is characterized in that: The structural reconstruction loss is: L Edge ＝∑ φ∈{Φ∪S} α φ ·l φ ； Among them, L Edge represents the structural reconstruction loss, α φ represents the weight of the meta-path φ, l φ Represents the edge reconstruction loss of the meta-path φ.

9. The method for discovering communities in financial heterogeneous networks based on joint contrastive learning and generative learning based on the same matching ratio according to claim 7 is characterized in that: The attribute reconstruction loss is: Among them, L Feat represents the attribute reconstruction loss, Represents the attribute matrix after masking, Z v Represents the node attributes reconstructed using the decoder. represents the edge matrix after masking, and γ2 represents the corresponding scaling factor.

Images