A coal mine multi-relationship big data analysis method with multi-view contrast enhancement

Abstract

The present invention discloses a multi-view contrast-enhanced coal mine multi-relationship big data analysis method, which first obtains data information to be analyzed under complex coal mine scenarios or complex businesses, constructs a heterogeneous graph model for complex coal mine business scenarios, and uses node features in the heterogeneous graph model to calculate the similarity between nodes to generate a similarity graph; the generated similarity graph is sequentially preprocessed by sparseness, non-negativity, and normalization; it is optimized through contrast learning and unsupervised optimization methods; and then a progressive graph structure update strategy is used to update the relationship between the extracted and represented complex data. The present invention fully retains the relationship information of the big data of complex coal mine business scenarios with the help of heterogeneous graph structures, and combines unsupervised learning methods to extract the data structure and semantic features of complex coal mine business scenarios, extract the key relationships of the data of complex coal mine business scenarios, and thus improve the performance on the multi-relationship mining task of complex coal mine data.

Description

Technical Field

[0001] The present invention relates to the technical fields of big data computing and multi-relationship data analysis, and in particular to a coal mine multi-relationship big data analysis method with multi-view contrast enhancement. Background Art

[0002] Multi-relational big data is a widespread phenomenon in real-world production and life. Due to the complexity of scenarios and businesses across various fields, the generated data often contains multiple entity types, with complex and ever-changing relationships between them. It also exhibits characteristics such as high noise, high redundancy, and large scale. For example, in the coal mining sector, workers of different trades may generate different fires (or fire hazards) through their different work behaviors. Traditional coal mine multi-relational analysis methods represent one fire (hazard) entity per record, making it difficult to discover the coupling relationships between multiple fires (hazards). Although some studies have proposed big data analysis methods, these methods have not effectively addressed the problem of extracting multi-relationships from big data. For example, the knowledge graph-based enterprise risk prediction method and system proposed in Patent No. CN201810271489.4 fails to represent the data's characteristic information. Heterogeneous graph models are an effective tool for modeling complex multi-relational data. They are network structures composed of multiple types of entities and their interactions. They are often used to model complex interactions between entities in the real world, such as in knowledge graph entity relationship construction and complex network multi-role modeling. Representation learning for heterogeneous graphs maps multi-type data nodes in the graph into low-dimensional, dense representation vectors to capture the underlying structural characteristics and semantic information between data entities. This underlying structural characteristics and semantic information provide key feature information support for diverse applications, such as enterprise business collaboration analysis, recommendation systems, abnormal activity detection, protein interaction prediction, traffic flow forecasting, and knowledge graph information representation. It serves as a bridge for a deeper understanding of the relationships between business systems and complex, multi-type data. Heterogeneous graph neural networks are an effective representation learning tool with powerful feature capture capabilities, and can be used to capture the underlying features in multi-type complex networks.

[0003] Existing heterogeneous graph neural networks use an attention mechanism to aggregate the features of the target node's neighboring nodes, and use the aggregated features as the updated representation of the target node. This method often relies on a reliable data association structure when learning data representation. However, the complex and multi-type data in real-world scenarios often contain a lot of noise and have a lot of redundancy. At the same time, due to the limitations of the knowledge and business capabilities of data analysts, it is often difficult to model the data to characterize the multiple relationships of the data, resulting in the loss of some relationships or even key relationships. For example, a compact graph structure learning method based on mutual information compression proposed in Patent No. CN202111588758.8 will be affected by relational errors when heterogeneous graph neural networks analyze multi-relational data, resulting in biased analysis results, which seriously affects the accuracy of the analysis task. A method and device for stable graph structure learning proposed in Patent No. CN202010182123.7 also verified that the noisy relationships and incomplete relationships between data produce biased data representations. Existing graph structure learning methods use neural network models to learn feature representations for data nodes while optimizing the graph structure to address node feature aggregation bias caused by noisy or missing relationships between data. These methods all use node domain labels as learning guidance signals, extracting and supplementing key data relationships under this guidance. Therefore, when data lacks domain identifiers or annotations, these methods struggle to extract key associations from complex data and learn accurate data representations.

[0004] Currently, there is a lack of a big data analysis method that can distinguish different types of data entities and the various relationships between them in practical application fields such as coal mines, and conduct qualitative and quantitative evaluations of them; when there are no domain labels of data nodes in the usage scenarios or businesses as guiding information, it can also mine key relational characteristics from the data itself and reduce the errors caused by the initial training of the model. Summary of the Invention

[0005] The technical problem to be solved by the present invention is as follows: When analyzing multi-relationship data with existing heterogeneous graph neural networks, it will be affected by relational errors and obtain biased analysis results, which seriously affects the accuracy of the analysis task. In addition, when domain labels without data nodes in specific scenarios or businesses are used as guiding information, it is difficult to mine key relational characteristics from the data itself. The present invention provides a multi-view contrast-enhanced coal mine multi-relationship big data analysis method, which fully retains the relational information of big data in complex business scenarios of coal mines with the help of heterogeneous graph structures, and combines unsupervised learning methods to extract structural and semantic features in the graph and extract key relationships of the data, thereby improving the performance of multi-relationship mining tasks in complex business scenario data of coal mines.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The present invention proposes a multi-view contrast-enhanced coal mine multi-relationship big data analysis method, which is used to analyze data association relationships in complex coal mine scenarios or complex businesses; the complex scenarios or complex businesses refer to application scenarios or businesses represented by the coal mine field that contain multiple entity types and complex and changeable relationships between data, and the scenarios or businesses do not have domain labels of data nodes as guiding information; the analysis method includes the following steps:

[0008] S1: Obtain data information to be analyzed in complex coal mine scenarios or complex businesses, construct a heterogeneous graph model for complex coal mine business scenarios based on the existing observation relationships of the data information, represent data units and observation relationships, calculate the similarity between nodes using node features in the heterogeneous graph model, and generate a similarity graph to extract key correlation relationships between data;

[0009] S2, according to the characteristics of the coal mine field, the generated similarity graph is preprocessed by sparseness, non-negativity and normalization in sequence;

[0010] S3 optimizes the similarity graph and the original graph structure through contrastive learning and unsupervised optimization methods, and generates a new graph that maintains the original topological information through dynamic fusion, reduces noisy relationships, and completes missing relationships;

[0011] S4, through a progressive graph structure update strategy, reduces the errors introduced by insufficient network fitting capabilities and updates the relationship between the extracted and represented complex data.

[0012] Furthermore, in step S1, the process of generating a similarity graph includes the following sub-steps:

[0013] S11, maps the features of different types of data nodes into the same feature space. The mapping process is as follows:

[0014]

[0015] Where, is a mapping matrix specific to the data node type; d is the feature dimension after mapping all types of data nodes; is the data node feature of type t after mapping;

[0016] S12, use L-layer multilayer perceptron to learn the embedding vector of data nodes, the input of MLP is the mapped data node features The process of learning data node embedding in the l-th layer network is as follows:

[0017]

[0018] Where, is the parameter matrix of the l-th layer network; σ() is the nonlinear activation function;

[0019] S13, for multiple pre-set meta-paths The similarity calculation function is used to learn the similarity graph corresponding to each meta-path, which is recorded as

[0020] Specifically, for a given meta-path P m , m=1,2,…,M, and the similarity graph is obtained by using the similarity calculation function:

[0021]

[0022] Where, φ() is the cosine similarity calculation function; and is the meta-path P m The characteristics corresponding to the data nodes at both ends; is the meta-path P m The type of data node to start with; is the meta-path P m The type of the data node that ends; The generated adjacency matrix of the element path Similarity graph of the same dimension.

[0023] Furthermore, in step S2, the k-nearest neighbor classification algorithm is used to perform sparse processing on the generated similarity graph, and the specific method is as follows:

[0024] For each data node, only the k neighbors that are most similar to the data node are retained, and the remaining interaction weights are set to 0. For each similarity graph under each meta-path The process of sparsification is as follows:

[0025]

[0026] Where top-k() is the k-value selection function, which is used to select the k maximum values ​​in the row vector; is the i-th row of the similarity graph adjacency matrix; f sp () represents the sparsification function;

[0027] After the sparse operation, the similarity graph structure that only retains the k most similar neighbors is obtained.

[0028] Furthermore, in step S2, the edge weights of the similarity graph are non-negative through the ReLU activation function. For each similarity graph under a meta-path, The process of non-negativity is as follows:

[0029]

[0030] Where, f act () represents the non-negative function; σ() is the nonlinear activation function;

[0031] After the non-negative operation, the sparse and non-negative similarity graph is obtained.

[0032] Furthermore, in step S2, the similarity graph under each meta-path is Perform normalization operation to place the edge weight values ​​in the similarity graph in the range of [0, 1]. The normalization calculation process is as follows:

[0033]

[0034] Where, Similarity graph The node degree matrix of is a diagonal matrix, and the diagonal elements are the degrees of the corresponding nodes; the output is the normalized Laplacian matrix of the similarity graph.

[0035] Furthermore, in step S3, the similarity graph and the original graph structure are optimized by contrastive learning and unsupervised optimization methods, and a new graph that can maintain the original business topology relationship is generated by dynamic fusion, reducing noise relationships, and completing missing relationships. The process includes the following steps:

[0036] S31, setting a comparative learning view;

[0037] The comparative learning view includes the learning view and the fixed view; the learning view is the processed similarity graph Data node characteristics For each similarity graph under a meta-path, its corresponding learning view is constructed and recorded as Fixed view is composed of metapath graph and node features It is used to provide high-order semantic information of heterogeneous graphs for the similarity graph in contrastive learning; for each meta-path graph under a meta-path, its corresponding fixed view is constructed, which is recorded as

[0038] S32, adopts the data augmentation method of feature masking and edge dropping to generate different augmented data from the perspective of features and structure respectively;

[0039] S33, based on multi-view comparative learning, optimizes the similarity graph and the original graph structure; specifically includes:

[0040] S331, use the graph neural network model to learn the representation of nodes on the two augmented views; the learning process is as follows:

[0041]

[0042] Where, is the learning meta-path P m Model of lower node representation; It is the node representation learned in the learning view; It is the node representation learned under a fixed view; is the feature matrix of the t-th type of data node learned on the learning view; is the feature matrix of the t-th type of data node learned on a fixed view;

[0043] S332 uses MLP to map the learned data node representations under the two views into the same feature space for calculating the node-level contrast loss. The mapping process is as follows:

[0044]

[0045] Where, is the metapath P m The MLP mapping function of the node representation under; After MLP mapping After MLP mapping

[0046] S333, construct the contrast loss function as follows:

[0047]

[0048]

[0049] Where φ() is the cosine similarity calculation function; t is the control parameter used to scale the similarity value; is the representation of data node i on the learning view; It is the representation of data node i in a fixed view.

[0050] Furthermore, in step S32, a data augmentation method of feature masking and edge discarding is adopted to generate different augmented data from the perspectives of features and structures, including the following sub-steps:

[0051] Data augmentation for feature masking of learning views, given node features For each class of feature X t All with p (f)The probability of masking its feature dimension and generating its corresponding feature mask vector The feature masking process is as follows:

[0052]

[0053]

[0054] Where x ti , is the feature of the ith node of type t; ⊙ represents the Hadamard product; T fm () is the feature masking conversion function; is the feature matrix after data augmentation; It is on the metapath P m , m = 1, 2, ..., M under the learning view after data augmentation;

[0055] Augment the data by discarding edges from the fixed view to generate a graph structure mask matrix of the same dimension as the meta-path graph and similarity graph With p (a) The probability of discarding the interaction between the two views is ; for each view under a meta-path, the edge discarding process is as follows:

[0056]

[0057]

[0058] Where, is the adjacency matrix that needs to be discarded; ⊙ represents the Hadamard product; T ed () is the edge discard function; is the adjacency matrix after data augmentation; It is on the metapath P m , m=1, 2,…,M fixed views after data augmentation.

[0059] S35, calculate the metapath P m , m=1, 2, …, M, the contrast loss is calculated as follows:

[0060]

[0061]

[0062] Where φ() is the cosine similarity calculation function; t is the temperature control parameter used to scale the similarity value; is the representation of data node i on the learning view; It is the representation of data node i in a fixed view.

[0063] Furthermore, in step S4, the process of updating the relationship between the extracted and represented complex data by using a progressive graph structure update strategy to reduce the error introduced by insufficient network fitting ability includes the following steps:

[0064] For each meta-path, the meta-path graph is optimized using progressive relation fusion.

[0065]

[0066] Where τ∈[0,1] is the fusion rate; step is the number of times the set fusion rate is updated step by step; q=1,2,…,Q is the number of iterations of the current model training; Q is the total number of training iterations of the model; represents the fusion rate controlled by the number of model training times;

[0067] Output optimized meta-path graph

[0068] The present invention adopts the above technical solution, and compared with the prior art, its significant technical effects are as follows:

[0069] 1. The multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention fully considers the characteristics of big data in complex business scenarios of coal mining enterprises, such as different data types, different data semantics, and multiple relationships between data. It uses heterogeneous graphs to capture such complex information and maintains the heterogeneous characteristics in complex coal mine scenarios and business big data.

[0070] 2. The multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention fully considers the existence of data redundancy, data noise, and lack of relational semantics in the big data of complex business scenarios of coal mining enterprises. The present invention uses a heterogeneous graph neural network model to capture key data relationships in complex coal mine data.

[0071] 3. The multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention fully considers the situation of extracting data relationships in complex enterprise scenarios and complex businesses without domain annotations. When domain labels without data nodes in the use scenario or business are used as guiding information, the comparative learning method is used to mine key relationship characteristics from the data itself to form guiding information to guide the learning of data node representation and the extraction of key data relationships.

[0072] 4. The multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention fully considers the data relationship error caused by the introduction of biased relationships due to the weak learning ability of the neural network in the initial stage of training when the learned captured relationships are merged with the original relationship network to form a new data relationship network in the data relationship fusion stage. A progressive relationship fusion method is designed, and the learned relationships and the original relationships are fused in a variable proportion as the model is trained, thereby reducing the error caused by the initial training of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a structural diagram of the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention.

[0074] Figure 2 This is a heterogeneous structure diagram of the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention.

[0075] Figure 3 This is an example diagram of commonly used methods for heterogeneity processing in the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention.

[0076] Figure 4 This is a flow chart of the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention. DETAILED DESCRIPTION

[0077] The present invention will be described in further detail below with reference to the accompanying drawings and specific implementations.

[0078] Figure 1 This is a structural diagram of the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention. Figure 4 This is a flowchart of the multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention. The embodiment of the present invention discloses a multi-view contrast-enhanced coal mine multi-relationship big data analysis method, which is used to analyze data association relationships in complex coal mine scenarios or complex businesses. Complex scenarios or complex businesses refer to application scenarios or businesses represented by the coal mine field that contain multiple entity types, have complex and changeable relationships between data, and lack domain labels for data nodes as guiding information. The analysis method includes the following steps:

[0079] S1, obtains the data information to be analyzed in complex coal mine scenarios or complex businesses, builds a heterogeneous graph model based on the existing observation relationship of the data information, represents the data units and observation relationships, uses the node features in the heterogeneous graph model to calculate the similarity between nodes, and generates a similarity graph for extracting key correlation relationships between data.

[0080] S2, according to the characteristics of the coal mine field, the generated similarity graph is preprocessed by sparseness, non-negativity and normalization in sequence.

[0081] S3 optimizes the similarity graph and the original graph structure through contrastive learning and unsupervised optimization methods, and generates a new graph that can maintain the original topological information through dynamic fusion, reduces noise relationships, and completes missing relationships.

[0082] S4, through a progressive graph structure update strategy, reduces the errors introduced by insufficient network fitting capabilities and updates the relationship between the extracted and represented complex data.

[0083] For example, in a coal mine fire scenario, many factors influence coal mine fires, including not only the workers and coal mine operations but also other factors. If the workers' situation changes, the coal mine operations also change. It's unclear whether the coal mine fire was caused by the workers, the coal mine operations, or some other factor. Therefore, in this example, we define the relationships between workers in coal mine work as two different meta-paths. The first meta-path represents the relationship between workers and their collaborative work; the second meta-path represents the relationship between workers performing work in the same fire incident.

[0084] 1. Generate Similarity Graph

[0085] Figure 2 This is a heterogeneous structure diagram of the coal mine multi-relationship big data analysis method with multi-view contrast enhancement of the present invention. Figure 3 This diagram illustrates a commonly used method for heterogeneity processing in the multi-view contrast-enhanced coal mine multi-relationship big data analysis method of the present invention. A heterogeneous graph model is constructed based on the observable relationships of complex coal mine big data. Data units and observation relationships are represented. The similarity between nodes in the graph is calculated using the features of the nodes in the graph. A similarity graph is constructed based on the feature similarity to extract key correlation relationships between data. The specific steps are as follows:

[0086] In order to calculate the feature similarity between data nodes, it is necessary to map the features of different types of data nodes into the same feature space. The calculation method is as follows:

[0087]

[0088] Where, is a mapping matrix specific to the data node type; d is the feature dimension after mapping all types of data nodes; is the data node feature of type t after mapping.

[0089] This method uses L-layer MLP to learn the representation vector of the node. The input of MLP is the mapped node features. The process of learning node embedding in the l-th layer network is as follows:

[0090]

[0091] Where, is the parameter matrix of the l-th layer network. Since different types of data node features have been mapped to the same feature space, the same parameter matrix is ​​used for each type of data node feature; σ() is the nonlinear activation function; this paper uses the ReLU function as the nonlinear activation function in MLP.

[0092] This embodiment uses cosine similarity to calculate the feature similarity between data nodes and generate a similarity graph. m , m=1,2,…,M, the similarity graph calculation process is as follows:

[0093]

[0094] Where, φ() is the cosine similarity calculation function; and is the meta-path P m The characteristics corresponding to the data nodes at both ends; is the meta-path P m The type of data node to start with; is the meta-path P m The type of the data node that ends; The generated adjacency matrix of the element path Similarity graph of the same dimension.

[0095] For multiple pre-set meta-paths The similarity calculation function is used to learn the similarity graph corresponding to each meta-path and recorded as

[0096] 2. Preprocessing the generated similarity graph

[0097] According to the domain characteristics, the generated similarity graph is subjected to preprocessing operations such as symmetry, sparseness, and normalization in sequence, so that it can more realistically express the coal mining domain nature of the data relationship. The specific steps are as follows:

[0098] Sparsification: This embodiment uses the K-nearest neighbor (KNN) graph to achieve sparsification of the similarity graph. For each data node, only the k most similar neighbors are retained, and the interaction weights with the rest are set to 0. For each similarity graph under a meta-path The process of sparsification is as follows:

[0099]

[0100] Where top-k() is the k-value selection function, which is used to select the k maximum values ​​in the row vector; is the i-th row of the similarity graph adjacency matrix; f sp () represents the sparsification function; after the sparsification operation, this method obtains a similarity graph structure that only retains the k most similar neighbors

[0101] Non-negative: This embodiment uses an activation function to non-negatively convert the edge weights of the similarity graph to ensure that the edge weight values ​​are within the non-negative value range. Its non-negativity and symmetry process are as follows:

[0102]

[0103] Where, f act () represents the non-negative function; σ() is a non-linear activation function; this embodiment uses the ReLU activation function to non-negatively transform the edge weights in the similarity graph to obtain a similarity graph that has been sparsely and non-negatively transformed.

[0104] Normalization: In order to place the edge weight values ​​in the similarity graph in the range [0, 1], this embodiment normalizes the similarity graph under each meta-path. Perform the normalization operation, the calculation process is as follows:

[0105]

[0106] Where, Similarity graph The node degree matrix of is a diagonal matrix, and the diagonal elements are the degrees of the corresponding nodes. is the normalized Laplacian matrix of the similarity graph.

[0107] 3. Optimize

[0108] First, set up the comparative study view:

[0109] Learning view: The learning view consists of the processed similarity graph Data node characteristics For each similarity graph under a meta-path, its corresponding learning view is constructed and recorded as

[0110] Fixed view: Fixed view is composed of meta path diagram Data node characteristics It is used to provide high-level semantic information of heterogeneous graphs for the similarity graph in contrastive learning. For each meta-path graph under a meta-path, its corresponding fixed view is constructed and recorded as

[0111] Then perform data augmentation operation:

[0112] Data augmentation for feature masking of learning views, given node features For each class of feature X t All with p (f) The probability of masking its feature dimension and generating its corresponding feature mask vector The feature masking process is as follows:

[0113]

[0114]

[0115] Where x ti , is the feature of the ith node of type t; ⊙ represents the Hadamard product; T fm () is the feature masking conversion function; is the feature matrix after data augmentation; It is on the metapath P m , m = 1, 2, …, M under the learning view after data augmentation.

[0116] Augment the data by discarding edges from the fixed view to generate a graph structure mask matrix of the same dimension as the meta-path graph and similarity graph With p (a) The probability of discarding the interaction between the two views is ; for each view under a meta-path, the edge discarding process is as follows:

[0117]

[0118]

[0119] Where, is the adjacency matrix that needs to be discarded; ⊙ represents the Hadamard product; T ed () is the edge discard function; is the adjacency matrix after data augmentation; It is on the metapath P m , m=1, 2,…,M fixed views after data augmentation.

[0120] This embodiment uses a graph neural network model to learn the representation of nodes on the two augmented views. m , two views under m=1, 2, ..., M and Its node representation learning process is as follows:

[0121]

[0122] Where, is the learning meta-path P m Model of the next node representation; set and Respectively shown in Figure and The node representation learned under and Represent the feature matrices of the t-th type of data node learned on the two views respectively.

[0123] After obtaining the representation of the data nodes through the graph convolutional neural network, this embodiment uses MLP to map the learned data node representations under the two views into the same feature space for calculating the node-level contrast loss. m , m = 1, 2, ..., M, the two node representations learned, the MLP mapping process is as follows:

[0124]

[0125] Where, is the metapath P m The MLP mapping function of the node representation under and Respectively represent the MLP mapping and

[0126] Finally, in the metapath P m , m=1, 2, …, M, the contrast loss is calculated as follows:

[0127]

[0128]

[0129] Where φ() represents the cosine similarity calculation function; t is the temperature control parameter used to scale the similarity value; and They represent the representation of data node i in the learning view and the fixed view respectively.

[0130] 4. Update

[0131] Through a progressive graph structure update strategy, we reduce the errors introduced by insufficient network fitting capabilities and update the relationship between the extracted and represented complex data. The specific solution is as follows:

[0132] This embodiment designs a progressive relational fusion method, a relational fusion method with a variable fusion ratio depending on the number of training times. For each meta-path, the calculation process is as follows:

[0133]

[0134] Where τ∈[0,1] is the fusion rate; step is the number of times the set fusion rate is updated step by step; q=1,2,…,Q is the number of iterations of the current model training; Q is the total number of training iterations of the model; Represents the fusion rate controlled by the number of model training times.

[0135] When the preset number of iterations is reached, the multi-relational data representation and the key relationships between the data are output. If the preset number of iterations is not reached, the similarity graph is regenerated for optimization iteration.

[0136] The above embodiments are only for illustrating the technical ideas of the present invention and cannot be used to limit the protection scope of the present invention and the application field of the present invention. Any changes made on the basis of the technical solutions in accordance with the technical ideas proposed by the present invention fall within the protection scope of the present invention.

Claims

1. A coal mine multi-relationship big data analysis method with multi-view contrast enhancement, characterized in that: The analysis method is used to analyze data association relationships in complex coal mine scenarios or complex businesses; the complex scenarios or complex businesses refer to application scenarios or businesses represented by the coal mine field that contain multiple entity types and complex and changeable relationships between data, and the scenarios or businesses do not have domain labels of data nodes as guiding information; the analysis method includes the following steps: S1: Obtain data information to be analyzed in complex scenarios or complex businesses, build a heterogeneous graph model based on the existing observation relationships of the data information, represent data units and observation relationships, calculate the similarity between nodes using node features in the heterogeneous graph model, and generate a similarity graph to extract key correlation relationships between data; S2, according to the domain characteristics, the generated similarity graph is preprocessed by sparseness, non-negativity and normalization in sequence; S3 optimizes the similarity graph and the original graph structure through contrastive learning and unsupervised optimization methods, and generates a new graph that maintains the original topological information through dynamic fusion, reduces noisy relationships, and completes missing relationships; S4, through a progressive graph structure update strategy, reduces the errors introduced by insufficient network fitting capabilities and updates the relationship between the extracted and represented complex data; In step S3, the similarity graph and the original graph structure are optimized through contrastive learning and unsupervised optimization methods. A new graph that can maintain the original topological information is generated through dynamic fusion, reducing noise relationships and completing missing relationships. The process includes the following steps: S31, setting a comparative learning view; The comparative learning view includes the learning view and the fixed view; the learning view is the processed similarity graph Data node characteristics For each similarity graph under a meta-path, its corresponding learning view is constructed and recorded as Fixed view is composed of metapath graph and node features It is used to provide high-order semantic information of heterogeneous graphs for the similarity graph in contrastive learning; for each meta-path graph under a meta-path, its corresponding fixed view is constructed, which is recorded as S32, adopts the data augmentation method of feature masking and edge dropping to generate different augmented data from the perspective of features and structure respectively; S33, based on multi-view comparative learning, optimizes the similarity graph and the original graph structure; specifically includes: S331, use the graph neural network model to learn the representation of nodes on the two augmented views; the learning process is as follows: Where, is the learning meta-path P m Model of lower node representation; It is the node representation learned in the learning view; It is the node representation learned under a fixed view; is the feature matrix of the t-th type of data node learned on the learning view; is the feature matrix of the t-th type of data node learned on a fixed view; S332 uses MLP to map the learned data node representations under the two views into the same feature space for calculating the node-level contrast loss. The mapping process is as follows: Where, is the metapath P m The MLP mapping function of the node representation under; After MLP mapping After MLP mapping S333, construct the contrast loss function as follows: Where φ() is the cosine similarity calculation function; t is the temperature control parameter used to scale the similarity value; is the representation of data node i on the learning view; It is the representation of data node i in a fixed view.

2. The multi-view contrast-enhanced coal mine multi-relationship big data analysis method according to claim 1 is characterized in that: In step S1, the process of generating a similarity graph includes the following sub-steps: S11, maps the features of different types of data nodes into the same feature space. The mapping process is as follows: Where, is a mapping matrix specific to the data node type; d is the feature dimension after mapping all types of data nodes; is the data node feature of type t after mapping; S12, use L-layer multilayer perceptron to learn the embedding vector of data nodes, the input of MLP is the mapped data node features The process of learning data node embedding in the l-th layer network is as follows: Where, is the parameter matrix of the l-th layer network; σ() is the nonlinear activation function; S13, for multiple pre-set meta-paths The similarity calculation function is used to learn the similarity graph corresponding to each meta-path, which is recorded as Specifically, for a given meta-path P m , m=1,2,…,M, and the similarity graph is obtained by using the similarity calculation function: Where, φ() is the cosine similarity calculation function; and is the meta-path P m The characteristics corresponding to the data nodes at both ends; is the meta-path P m The type of data node to start with; is the meta-path P m The type of the data node that ends; The generated adjacency matrix of the element path Similarity graph of the same dimension.

3. The coal mine multi-relationship big data analysis method with multi-view contrast enhancement according to claim 1 is characterized in that: In step S2, the k-nearest neighbor classification algorithm is used to perform sparse processing on the generated similarity graph. The specific method is as follows: For each data node, only the k neighbors that are most similar to the data node are retained, and the remaining interaction weights are set to 0. For each similarity graph under each meta-path The process of sparsification is as follows: Where top-k() is the k-value selection function, which is used to select the k maximum values ​​in the row vector; is the i-th row of the similarity graph adjacency matrix; f sp () represents the sparsification function; After the sparse operation, the similarity graph structure that only retains the k most similar neighbors is obtained.

4. The method for analyzing coal mine multi-relationship big data with multi-view contrast enhancement according to claim 1, characterized in that: In step S2, the edge weights of the similarity graph are non-negative through the ReLU activation function. For each similarity graph under a meta-path, The process of non-negativity is as follows: Where, f act () represents the non-negative function; σ() is the nonlinear activation function; After the non-negative operation, the sparse and non-negative similarity graph is obtained.

5. The coal mine multi-relationship big data analysis method with multi-view contrast enhancement according to claim 1 is characterized in that: In step S2, the similarity graph under each meta-path is Perform normalization operation to place the edge weight values ​​in the similarity graph in the range of [0, 1]. The normalization calculation process is as follows: Where, Similarity graph The node degree matrix of is a diagonal matrix, and the diagonal elements are the degrees of the corresponding nodes; the output is the normalized Laplacian matrix of the similarity graph.

6. The method for analyzing coal mine multi-relationship big data with multi-view contrast enhancement according to claim 1, characterized in that: In step S32, a data augmentation method using feature masking and edge discarding is used to generate different augmented data from the perspectives of features and structure, including the following sub-steps: Data augmentation for feature masking of learning views, given node features For each class of feature X t All with p (f) The probability of masking its characteristic dimension, Generate its corresponding feature mask vector The feature masking process is as follows: Where x ti , i=1,2,…, is the feature of the ith node of type t; ⊙ represents Hadamard product; T fm () is the feature masking conversion function; is the feature matrix after data augmentation; It is on the metapath P m , m = 1, 2, ..., M under the learning view after data augmentation; Augment the data by discarding edges from the fixed view to generate a graph structure mask matrix of the same dimension as the meta-path graph and similarity graph With p (a) The probability of discarding the interaction between the two views is ; for each view under a meta-path, the edge discarding process is as follows: Where, is the adjacency matrix that needs to be discarded; ⊙ represents the Hadamard product; T ed () is the edge discard function; is the adjacency matrix after data augmentation; It is on the metapath P m , m = 1, 2, ..., M fixed view after data augmentation; S35, calculate the metapath P m , m=1, 2, …, M, the contrast loss is calculated as follows: Where φ() is the cosine similarity calculation function; t is the temperature control parameter used to scale the similarity value; is the representation of data node i on the learning view; It is the representation of data node i in a fixed view.

7. The coal mine multi-relationship big data analysis method with multi-view contrast enhancement according to claim 1 is characterized in that: In step S4, the process of updating the relationship between the extracted and represented complex data by using a progressive graph structure update strategy to reduce the error introduced by insufficient network fitting ability includes the following steps: For each meta-path, the meta-path graph is optimized using progressive relation fusion. Where τ∈[0,1] is the fusion rate; step is the number of times the set fusion rate is updated step by step; q=1,2,…,Q is the number of iterations of the current model training; Q is the total number of training iterations of the model; represents the fusion rate controlled by the number of model training times; Output optimized meta-path graph

Images