A deep learning-based enterprise data mining system and method

Abstract

The application discloses a kind of enterprise data mining systems and methods based on deep learning, it is related to big data technical field, comprising the following steps: S1, output is arranged in time stamp index order dynamic snapshot sequence;S2, output is the global importance weight of the influence of representation node in network topology;S3, output is the graph structure guide position coding containing topological physical position and semantic social status;S4, output is local micro-topology enhanced feature by LGNN model;S5, output is the node feature of fusing local mode and global importance by TGN model;S6, output is deep context feature vector by improving Longformer model;S7, output is the target mining result for future time.This application overcomes the limitations of relying on simple rules, insufficient feature extraction capability and early warning lag in traditional enterprise data mining methods, and provides an efficient and accurate solution for enterprise intelligent decision-making and trend analysis.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to an enterprise data mining system and method based on deep learning. Background Technology

[0002] With the accelerating pace of enterprise digital transformation and the deepening application of big data technology in business decision-making, traditional enterprise data mining and business analysis methods are facing increasing challenges. In modern enterprise management, accurately identifying potential business value and future development trends from massive amounts of data is not only a key means to enhance core competitiveness but also an important link in optimizing resource allocation and the accuracy of strategic decision-making. However, most current enterprise data mining methods rely on traditional relational database queries, statistical analysis reports, or expert systems based on simple rules, resulting in complex processing workflows and weak in-depth mining capabilities. While these traditional methods can meet basic business data aggregation needs, their lack of real-time integration and intelligent in-depth analysis capabilities for multi-source heterogeneous data makes them ill-suited to the complex and ever-changing business environment and real-time dynamic decision-making requirements of modern enterprises.

[0003] The main limitations of traditional enterprise data mining methods lie in their insufficient feature extraction capabilities and lagging dynamic evolution analysis. Existing methods typically rely on shallow statistical features, human experience, or static rule models to judge business status, which is time-consuming and struggles to accurately and promptly capture the complex logic behind the data. When enterprise data exhibits high dimensionality, multi-source nature, nonlinearity, and strong time-varying characteristics, the generalization performance and adaptability of traditional mining methods are severely limited. In particular, when faced with diverse data such as transaction records, interaction logs, personnel turnover, and supply chain fluctuations, traditional single-dimensional analysis methods struggle to efficiently and accurately identify potential business patterns or relationships, resulting in the difficulty of timely discovery and early warning of key information, seriously affecting the scientific and timely nature of enterprise decision-making.

[0004] Furthermore, traditional methods in enterprise data mining often overlook the topological relationships and functional dependencies between data entities, making it difficult to comprehensively utilize the potential correlations between network structure information and state data. For example, in enterprise systems with complex organizational structures or business interaction networks, traditional isolated entity analysis methods cannot effectively integrate the correlation information between multi-entity interaction data and operational status, resulting in low prediction accuracy, high false alarm rates, or slow response times. Even when some methods employ general machine learning or deep learning techniques, they fail to fully mine the deep feature relationships between entity topological information and long-term time-series state data, making it difficult to achieve efficient, accurate, and forward-looking enterprise target mining and trend prediction.

[0005] Therefore, how to provide an enterprise data mining system and method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention proposes a deep learning-based enterprise data mining system and method. By combining an improved Longformer model with a graph neural network, it can more accurately mine and predict enterprise goals at future moments. This method can not only automatically extract deep contextual features from multi-source heterogeneous enterprise data, but also combine graph topology information to effectively fuse global importance features with local micro-topological features through an adaptive random walk algorithm and a time graph network, significantly improving the depth of enterprise data mining and the accuracy of time-series prediction. By introducing Riemannian manifold mapping, a two-stream attention deconstruction mechanism, and a global memory compression layer, the traditional Longformer model is improved, enabling the model to adaptively handle high-dimensional nonlinear data and effectively capture long-range dependencies, capturing subtle changes in node states in real time, thereby achieving accurate mining of goals at future moments. This invention overcomes the limitations of traditional enterprise data mining methods, such as reliance on simple rules, insufficient feature extraction capabilities, and delayed early warning, providing an efficient and accurate solution for intelligent enterprise decision-making and trend analysis.

[0007] A deep learning-based enterprise data mining system and method according to an embodiment of the present invention includes the following steps: S1. Collect multi-source heterogeneous data from enterprises, use time discretization operators to perform window segmentation on the continuous time axis, construct topology graph snapshots based on entity instances and interaction relationships within each window, and output a dynamic snapshot sequence arranged in order by timestamp index; S2. Input the dynamic snapshot sequence into the node weight calculation module, iteratively update the random walk damping parameters based on the node degree centrality and edge weights, execute the page ranking algorithm iterative calculation until it converges to a steady-state distribution, and output the global importance weight that represents the influence of the node in the network topology. S3. Input the global importance weight and topological structure index into the location coding generator, extract the node degree, clustering coefficient and multi-hop neighbor number to construct a structured feature vector, perform nonlinear mapping through a multilayer perceptron, and output a graph structure-guided location code containing topological physical location and semantic social status. S4. For the topological graph snapshot, take the high-weight node as the center, sample the local subgraph based on the shortest path distance range, construct the line graph topology and input it into the LGNN model. Through edge-to-node mapping and convolution aggregation, use the self-loop mechanism to reverse map the edge features back to the center node and output the local micro-topological enhancement features. S5. Perform weighted summation and aggregation on the features of neighboring nodes using global importance weights, fuse the results with local micro-topology enhancement features and input them into the TGN model, combine the joint memory module and time-coded attention to extract temporal evolution features, and output node features that fuse local patterns and global importance. S6. Perform additive normalization on node features and graph structure-guided position encoding to generate model input tensors. Input the improved Longformer model to calculate local context using a sliding window expansion attention mechanism, and capture long-range dependencies through global attention tokens to output deep context feature vectors. S7. Input the deep context feature vector into the fully connected layer, perform decision boundary mapping through linear transformation and nonlinear activation function, calculate the posterior probability distribution using a normalized exponential function classifier, and output the target mining results for future time moments.

[0008] Optionally, S1 specifically includes: S11. Parse the enterprise's multi-source heterogeneous data, extract entity identifiers, interaction types and timestamp fields through regular expression matching, and construct the original event stream containing node initial attributes and edge interaction records; S12. Set the time window length and sliding step size, use the time discretization operator to perform discretization mapping on the timestamp values ​​contained in the original event stream to generate discrete indexes, and allocate interactive events to the corresponding time windows according to the discrete indexes to generate a set of discrete time segments. S13. Traverse the set of discrete time segments, extract the unique identifiers of entity instances belonging to each discrete time segment as a set of nodes, extract the entity interaction associations belonging to each discrete time segment as a set of edges, and construct a graph topology skeleton representing the node connection relationship. S14. Calculate the interaction frequency statistics of the node set under the corresponding discrete time segment, map the interaction frequency statistics to edge weights, map the node attributes belonging to the current time segment to node features, and fuse the edge weights and node features into the graph topology skeleton to generate a single-frame topology graph snapshot. S15. Serialize and arrange the single-frame topology snapshots according to the order of timestamp indices, construct data tensors stacked along the time dimension, and output dynamic snapshot sequence.

[0009] Optionally, S2 specifically includes: S21. Extract the adjacency matrix and edge weight vector of a single-frame topology snapshot, calculate the sum of the weights of the associated edges of nodes to obtain the node strength, and perform extreme value standardization processing on the node strength using a preset normalization interval to generate a node degree centrality index. S22. Construct a nonlinear mapping function for the damping parameters, input the node degree centrality index into the mapping function to perform numerical transformation, and perform interval constraint operation on the transformed values ​​to generate adaptive damping parameters for balancing random walk and restart probability. S23. Construct a basic transition probability matrix based on the edge weight vector, perform weighted fusion processing on the basic transition probability matrix using adaptive damping parameters to construct a random walk transition matrix, initialize the node importance score vector, iteratively update the node scores through matrix-vector product, and calculate the Euclidean distance between the node score vectors before and after the iterative update. S24. Determine whether the Euclidean distance is less than the preset convergence threshold. If so, stop the iterative calculation and output the current node importance score vector as the global importance weight. Otherwise, continue to update the node importance score vector based on the random walk transition matrix.

[0010] Optionally, S3 specifically includes: S31. Call the position encoding generator to parse the adjacency relationship of a single frame topology snapshot, and calculate the first degree of the node, the local clustering coefficient, and the multi-hop neighbor statistics within the preset number of hops; S32. Map the degree, clustering coefficient and multi-hop neighbor statistics to numerical tensors, perform feature dimension concatenation operation, and generate node local topology vectors. S33. Construct a multilayer perceptron model containing fully connected layers and activation functions. Use the local topological structure vector of the node as the model input for forward propagation. Calculate the product of the input tensor and the weight matrix and add the bias term. Perform nonlinear mapping transformation to generate topological feature embeddings. S34. Obtain the global importance weights, embed the topological features and perform tensor fusion operations with the global importance weights to generate graph structure-guided position encoding.

[0011] Optionally, the LGNN model includes a subgraph sampling module, a line graph construction module, a feature projection module, a line graph convolution module, and a topology mapping reconstruction module. The subgraph sampling module is used to parse the adjacency matrix of the global topology graph, use a preset set of high-weight node indices as anchor points to calculate the shortest path distance matrix between nodes, filter the subset of nodes whose path distance is less than a preset threshold, and construct the adjacency tensor of the local induced subgraph. The line graph construction module is used to perform second-order tensor transformation operations based on the adjacency tensor of the local induced subgraph, mapping the nodes in the local induced subgraph to the edges in the line graph, mapping the connection relationships between nodes to the nodes in the line graph, and generating the adjacency association matrix of the line graph topology. The feature projection module is used to perform a linear transformation on the node feature matrix of the local induced subgraph. By using the index mapping relationship between nodes and edges, the transformed node feature vectors are aligned to the edge dimension of the line graph topology to generate the line graph initialization edge feature matrix. The line graph convolution module is used to construct a message-passing neural network based on the adjacency association matrix of the line graph topology, and to perform neighbor node aggregation and feature update iteration operations on the initial edge feature matrix of the line graph to obtain the hidden state feature matrix of the line graph nodes. The topology mapping and reconstruction module is used to perform inverse dimensional transformation on the hidden state feature matrix of the line graph node, map the features of the line graph node back to the node space of the original topology graph, and update the feature vector of the central node by weighting through the self-loop edge weighting mechanism, and output the local micro-topology enhancement feature matrix.

[0012] Optionally, the TGN model includes a weighted aggregation layer, a feature fusion layer, a joint memory module, a temporal encoding layer, and a temporal attention computation layer: The weighted aggregation layer is used to obtain the neighbor feature matrix and global importance weight vector of a node, perform the Hadamard product operation of the feature matrix and weight vector and the weighted summation operation along the adjacency dimension, and output the aggregated neighbor feature vector. The feature fusion layer is used to perform a tensor concatenation operation along the feature dimension on the aggregated neighbor feature vector and the local micro-topology enhancement feature to generate a node input vector that fuses static features. The joint memory module is used to construct a temporal memory network containing a node memory matrix, parse the timestamp and interactive event stream of the current time step, update the historical memory state through a gated recurrent neural network, and retrieve and output the node memory state vector at the current moment. The time coding layer is used to obtain the original timestamp, map the timestamp to a high-dimensional Riemannian manifold space or perform a nonlinear mapping using a multilayer perceptron to generate a time coding embedding vector. The temporal attention computation layer is used to construct a multi-head attention mechanism. It takes the node input vector as the query vector, the linear combination of the node memory state vector and the temporal encoding embedding vector as the key-value pair, calculates the spatiotemporal correlation score between features, performs weighted aggregation and residual connection based on the score, and outputs node features that fuse local patterns and global importance.

[0013] Optionally, the improved Longformer model includes a Mostinian manifold mapping layer, a reweighted spectral coding layer, a two-stream attention deconstruction layer, and a global memory compression layer: The Mostin manifold mapping layer is used to construct the Poincaré sphere model, obtain the original node feature vectors, project the original node feature vectors from Euclidean space to hyperbolic space through the logarithmic mapping of the tangent space of the Riemann manifold, calculate the gradient update under the manifold curvature parameter, and output the hyperbolic node feature tensor. The reweighted spectral coding layer is used to obtain the graph structure adjacency matrix, construct the normalized graph Laplacian matrix and perform eigenvalue decomposition, reweight the feature components using a preset learnable spectral filter matrix, suppress high-frequency noise components and enhance low-frequency structural components, project the reweighted spectral features back to the hyperbolic space through Riemann exponent mapping, and fuse them with the hyperbolic node feature tensor using Mostine addition to output the manifold structure perception tensor. The dual-stream attention deconstruction layer is used to divide the manifold structure perception tensor into non-overlapping local blocks along the sequence dimension, construct a block diagonal mask matrix, perform multi-head self-attention operation on the node features within the local blocks, and output a local fine-grained feature tensor; at the same time, it extracts the global feature matrix of the manifold structure perception tensor, decomposes it into a tensor ring form of core tensor and factor matrix, calculates global correlation through tensor ring shrinkage operation, and outputs a global low-rank feature tensor. The global memory compression layer is used to initialize the global memory token of the tensor ring structure, perform pooling operation on the global low-rank feature tensor along the node dimension, map the pooling result to each core factor of the tensor ring, perform temporal state update on the core factors using the gated loop unit, and reconstruct the global context vector through tensor ring expansion operation. The local fine-grained feature tensor and the global context vector are concatenated and normalized along the feature dimension. Linear transformation and non-linear activation are then performed through a fully connected layer to output a deep context feature vector.

[0014] Optionally, S7 specifically includes: S71. Obtain the deep context feature vector, construct the multilayer perceptron mapping head, perform linear transformation of the feature space through the weight matrix, and process the output vector layer by layer using the nonlinear activation function to obtain the classification feature vector. S72. Input the classification feature vector into the normalization exponential function layer, calculate the exponential value of the vector element and perform normalization processing to generate the posterior probability distribution belonging to each preset category, and extract the index corresponding to the maximum value in the posterior probability distribution. S73. Map the index to a predefined set of labels, obtain the corresponding category semantics, encapsulate the category semantics with timestamp information, and output the target mining results for future moments.

[0015] The beneficial effects of this invention are: (1) This invention significantly improves the ability to extract features of deep temporal dependencies and complex structures in enterprise data mining by introducing an improved Longformer model. Although the traditional Transformer model has advantages in processing long sequence data, it often suffers from high computational complexity, difficulty in effectively capturing hierarchical structures, and easy neglect of local fine-grained features when facing high-dimensional heterogeneous enterprise data. To solve these problems, this invention innovatively introduces a Mostinian manifold mapping layer and a reweighted spectral coding layer. By using Riemannian manifold and graph Laplacian spectral decomposition techniques, node features are projected onto hyperbolic space and low-frequency structural components are enhanced, effectively suppressing high-frequency noise interference. At the same time, a dual-stream attention deconstruction layer and a global memory compression layer are adopted. Through block diagonal masking and tensor ring memory mechanisms, the computational complexity is reduced while achieving efficient capture of local context and long-range dependencies, thereby significantly improving the model's ability to represent complex business logic and its prediction accuracy.

[0016] (2) This invention effectively solves the technical challenge of integrating local micro-topological features and global dynamic evolution patterns in multi-source heterogeneous enterprise data by employing a combination of TGN and LGNN technologies. Traditional data mining methods often neglect the complex interactions and dynamic changes between entity nodes, making it difficult to accurately capture deep patterns and potential risks in business interactions. This invention automatically identifies key nodes by constructing dynamic snapshot sequences and using an adaptive random walk strategy to calculate global importance weights; it accurately mines the micro-relationships between edges and nodes by combining the LGNN model with line graph convolution and feature reconstruction of the local subgraphs of high-weight nodes; and it dynamically updates node features by integrating node memory states and time encoding through the joint memory module and temporal attention mechanism of the TGN model. This method breaks through the limitations of traditional methods in handling complex network topologies and temporal dynamics, significantly improving the timeliness and accuracy of target mining, and providing efficient and accurate data support for intelligent enterprise decision-making. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based enterprise data mining system and method proposed in this invention. Figure 2 This is a flowchart illustrating the working principle of the improved Longformer model of an enterprise data mining system and method based on deep learning proposed in this invention. Detailed Implementation

[0018] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A deep learning-based enterprise data mining system and method, comprising the following steps: S1. Collect multi-source heterogeneous data from enterprises, use time discretization operators to perform window segmentation on the continuous time axis, construct topology graph snapshots based on entity instances and interaction relationships within each window, and output a dynamic snapshot sequence arranged in order by timestamp index; S2. Input the dynamic snapshot sequence into the node weight calculation module, iteratively update the random walk damping parameters based on the node degree centrality and edge weights, execute the page ranking algorithm iterative calculation until it converges to a steady-state distribution, and output the global importance weight that represents the influence of the node in the network topology. S3. Input the global importance weight and topological structure index into the location coding generator, extract the node degree, clustering coefficient and multi-hop neighbor number to construct a structured feature vector, perform nonlinear mapping through a multilayer perceptron, and output a graph structure-guided location code containing topological physical location and semantic social status. S4. For the topological graph snapshot, take the high-weight node as the center, sample the local subgraph based on the shortest path distance range, construct the line graph topology and input it into the LGNN model. Through edge-to-node mapping and convolution aggregation, use the self-loop mechanism to reverse map the edge features back to the center node and output the local micro-topological enhancement features. S5. Perform weighted summation and aggregation on the features of neighboring nodes using global importance weights, fuse the results with local micro-topology enhancement features and input them into the TGN model, combine the joint memory module and time-coded attention to extract temporal evolution features, and output node features that fuse local patterns and global importance. S6. Perform additive normalization on node features and graph structure-guided position encoding to generate model input tensors. Input the improved Longformer model to calculate local context using a sliding window expansion attention mechanism, and capture long-range dependencies through global attention tokens to output deep context feature vectors. S7. Input the deep context feature vector into the fully connected layer, perform decision boundary mapping through linear transformation and nonlinear activation function, calculate the posterior probability distribution using a normalized exponential function classifier, and output the target mining results for future time moments.

[0020] In this embodiment, S1 specifically includes: S11. Parse the enterprise's multi-source heterogeneous data, extract entity identifiers, interaction types, and timestamp fields through regular expression matching, and construct an original event stream containing node initial attributes and edge interaction records; the regular expression matching includes setting regular expression rules, scanning the original data text to match user account strings, transaction type codes, and timestamp values ​​accurate to milliseconds in a preset format; extract the successfully matched field data to a memory buffer, and build corresponding index files according to the extracted data types; if a data record is missing any one of the entity identifier, interaction type, or timestamp fields, the record is marked as abnormal data and removed; arrange the cleaned complete data records in chronological order to form an original event stream containing node initial attributes and edge interaction records.

[0021] S12. Set the time window length and sliding step size, and use the time discretization operator to perform discretization mapping on the timestamp values ​​contained in the original event stream to generate discrete indexes. Based on the discrete indexes, allocate the interactive events to the corresponding time windows to generate a set of discrete time segments. The time discretization operator includes setting the time window length to 24 hours and the sliding step size to 1 hour; reading the timestamp value of each record in the original event stream, calculating the difference between the timestamp value and the starting reference time, dividing the difference by the sliding step size and rounding down to obtain the corresponding discrete index; grouping interactive events with the same discrete index value into the same group, with each group representing a discrete time segment; traversing all interactive events to complete the grouping operation, generating a set of multiple discrete time segments arranged in chronological order.

[0022] S13. Traverse the set of discrete time segments, extract the unique identifiers of entity instances belonging to each discrete time segment as a node set, extract the entity interaction associations belonging to each discrete time segment as an edge set, and construct a graph topology skeleton representing the node connection relationship; the traversal includes reading each group in the discrete time segment set, scanning the interaction type field of all interaction events in the current group; if the interaction type indicates that there is an interaction behavior, extract the main entity identifier and the guest entity identifier involved in the event, and add these two identifiers to the node set; at the same time, create an edge object connecting the main entity and the guest entity, and add the edge object to the edge set; after the traversal is completed, construct an adjacency list based on the node set and the edge set, and generate a graph topology skeleton representing the node connection relationship.

[0023] S14. Calculate the interaction frequency statistics of the node set under the corresponding discrete time segment, map the interaction frequency statistics to edge weights, and map the node attributes belonging to the current time segment to node features. Merge the edge weights and node features into the graph topology skeleton to generate a single-frame topology snapshot. The calculation includes, for each edge in the graph topology skeleton, counting the total number of interaction behaviors between corresponding entities in the current discrete time segment as the interaction frequency statistics; directly assigning the interaction frequency statistics to the corresponding edge as the edge weight; dividing the interaction frequency statistics by a preset value of 50 and normalizing it to a value between 0 and 1 as the normalized edge weight; reading the node attribute data when the last interaction occurred in the current discrete time segment and converting it into a 128-dimensional numerical vector as the node feature; writing the calculated edge weights and node features into the data structure of the graph topology skeleton to generate a single-frame topology snapshot.

[0024] S15. Serialize and arrange the single-frame topology map snapshots according to their timestamp indices to construct a data tensor stacked along the time dimension, and output a dynamic snapshot sequence. The serialization and arrangement includes reading all generated single-frame topology map snapshots and obtaining the timestamp index corresponding to each snapshot; sorting the single-frame topology map snapshots according to their timestamp indices from smallest to largest; stacking the sorted single-frame topology map snapshots sequentially to construct a three-dimensional data tensor containing time dimension, node dimension, and feature dimension; storing the data tensor in a preset storage path as input data for subsequent deep learning models, and outputting a dynamic snapshot sequence.

[0025] In this embodiment, S2 specifically includes: S21. Extract the adjacency matrix and edge weight vector of a single-frame topology snapshot, calculate the sum of the weights of the edges associated with each node to obtain the node strength, and perform extreme value standardization processing on the node strength using a preset normalization interval to generate a node degree centrality index. The extraction includes reading the data structure of the single-frame topology snapshot to obtain the adjacency matrix containing node connection relationships and the edge weight vector recording the connection strength. The calculation of node strength includes finding all edges directly connected to each node in the adjacency matrix, summing the weight values ​​corresponding to these edges, and using the sum as the node strength of that node. The extreme value standardization processing includes setting a preset normalization interval of 0 to 1, finding the maximum and minimum values ​​of node strength among all nodes, subtracting the minimum value from the node strength and dividing by the difference between the maximum and minimum values ​​to calculate the normalized value, and using this value as the node degree centrality index of that node.

[0026] S22. Construct a nonlinear mapping function for the damping parameters. Input the node degree centrality index into the mapping function to perform numerical transformation, and perform interval constraint operations on the transformed values ​​to generate adaptive damping parameters for balancing random walks and restart probabilities. The construction of the nonlinear mapping function for the damping parameters includes setting a sigmoid function as the mapping function, setting the slope parameter of the function to 2, and setting the offset to 0. The numerical transformation includes inputting the node degree centrality index calculated in step S21 as the independent variable into the sigmoid function to calculate a transformed value between 0 and 1. The interval constraint operation includes setting the lower limit of the damping parameter to 0.15 and the upper limit to 0.85, comparing the calculated transformed value with the upper and lower limits, and forcibly correcting the transformed value to 0.15 if it is less than 0.15, and forcibly correcting it to 0.85 if it is greater than 0.85, thus obtaining the adaptive damping parameters for the node.

[0027] S23. Construct a basic transition probability matrix based on the edge weight vector. Perform weighted fusion processing on the basic transition probability matrix using adaptive damping parameters to construct a random walk transition matrix. Initialize the node importance score vector. Iteratively update the node scores through matrix-vector multiplication and calculate the Euclidean distance between the node score vectors before and after the iterative update. The construction of the basic transition probability matrix includes dividing the elements of each row in the adjacency matrix by the sum of all elements in that row, so that the sum of the values ​​in each row equals 1. The weighted fusion processing includes multiplying the basic transition probability matrix by the formula (1 minus the adaptive damping parameter), and then adding the adaptive damping parameter multiplied by the teleportation matrix to construct the random walk transition matrix. The initialization of the node importance score vector includes creating a vector with a length equal to the total number of nodes and initializing the values ​​of all elements to 1 divided by the total number of nodes. The iterative update includes performing matrix multiplication on the random walk transition matrix and the node importance score vector to obtain the updated score vector. Calculate the sum of squares of the differences between corresponding elements in the score vectors before and after the update, and take the square root of the sum of squares to obtain the Euclidean distance.

[0028] S24. Determine whether the Euclidean distance is less than a preset convergence threshold. If so, stop the iterative calculation and output the current node importance score vector as the global importance weight. Otherwise, continue to update the node importance score vector based on the random walk transition matrix. The determination includes setting the preset convergence threshold to 0.0001. Stopping the iterative calculation includes determining that the score vector has converged when the calculated Euclidean distance is less than 0.0001, recording the current iteration number, and terminating the calculation loop. The output includes reading the node importance score vector at the time of termination and using its value as the global importance weight to measure the node's influence in the network topology. The continued updating includes using the updated score vector as the input for the next iteration when the calculated Euclidean distance is greater than or equal to 0.0001, repeating the matrix multiplication operation and Euclidean distance calculation process until the convergence condition is met or the number of iterations reaches the maximum value of 1000.

[0029] In this embodiment, S3 specifically includes: S31. Call the position encoding generator to parse the adjacency relationship of the single-frame topology snapshot, calculate the first degree of the node, the local clustering coefficient, and the multi-hop neighbor statistics within a preset number of hops; the parsing includes reading the adjacency matrix data of the single-frame topology snapshot, traversing each node in the matrix; counting the number of edges directly connected to the current node, and using the counted number as the first degree of the node; querying all neighbor nodes of the current node, calculating the ratio of the actual number of edges between these neighbor nodes to the theoretical maximum number of edges, and using the ratio as the local clustering coefficient; setting the preset number of hops to 3 hops, using the breadth-first search algorithm to find all neighbor nodes of the current node within the 3-hop range, and using the total number of neighbor nodes obtained as the multi-hop neighbor statistics.

[0030] S32. Map the degree, clustering coefficient, and multi-hop neighbor statistics to numerical tensors, and perform a feature dimension concatenation operation to generate a local topological structure vector of the node. The mapping includes converting the calculated first-order degree, local clustering coefficient, and multi-hop neighbor statistics into 32-bit floating-point scalar values ​​respectively; filling these three scalar values ​​into a preset one-dimensional numerical tensor of shape 1; performing a feature dimension concatenation operation to connect the three one-dimensional numerical tensors containing degree, clustering coefficient, and multi-hop neighbor statistics end to end in the feature dimension to generate a one-dimensional array of shape 3, and using this array as the local topological structure vector of the node.

[0031] S33. Construct a multilayer perceptron model containing fully connected layers and activation functions. Use the node's local topological structure vector as model input for forward propagation. Calculate the product of the input tensor and the weight matrix, and add a bias term. Perform a nonlinear mapping transformation to generate a topological feature embedding. The construction includes setting the number of neurons in the input layer of the multilayer perceptron model to 3, corresponding to the dimension of the local topological structure vector, and setting the number of neurons in the output layer to 64. Initialize the weight matrix and bias term. The values ​​of the weight matrix are randomly generated using a normal distribution, and the initial value of the bias term is set to 0. The forward propagation includes performing matrix multiplication on the node's local topological structure vector and the weight matrix, adding the result to the bias term, and inputting the result into the activation function for nonlinear calculation. The output is a topological feature embedding with a dimension of 64.

[0032] S34. Obtain the global importance weights, perform tensor fusion operation on the topological feature embeddings and the global importance weights, and generate the graph structure guidance position code; the obtaining includes reading the global importance weight vector calculated and output in step S2, and extracting the importance value corresponding to the current node; the tensor fusion operation includes performing a Hadamard product operation on each element of the topological feature embedding vector and the global importance weight value, or performing a scalar multiplication operation on the topological feature embedding vector as a coefficient; normalize the fused feature vector to keep the output value within a preset range, and generate the final graph structure guidance position code.

[0033] In this embodiment, the LGNN model includes a subgraph sampling module, a line graph construction module, a feature projection module, a line graph convolution module, and a topology mapping reconstruction module. The subgraph sampling module is used to parse the adjacency matrix of the global topology graph. Using a preset set of high-weight node indices as anchors, it calculates the shortest path distance matrix between nodes, filters a subset of nodes whose path distance is less than a preset threshold, and constructs an adjacency tensor for a local induced subgraph. The parsing process includes reading the adjacency matrix data of the global topology graph, selecting the top 10 most important nodes from all nodes as a preset set of high-weight node indices, using each node in the high-weight node indices as a starting anchor, traversing the global topology graph using a breadth-first search algorithm, calculating the shortest path length from the anchor to all other nodes in the graph, and constructing a symmetric matrix containing distance information for all nodes. Setting the preset threshold to 2, it filters nodes with shortest path lengths less than 2 from the symmetric matrix, extracts these filtered nodes along with their connecting edges, combines them to construct a local induced subgraph, and converts the connection relationships of this subgraph into an adjacency tensor of a preset dimension.

[0034] The line graph construction module is used to perform second-order tensor transformation operations based on the adjacency tensor of the local induced subgraph, mapping the nodes in the local induced subgraph to edges in the line graph, and mapping the connection relationships between nodes to nodes in the line graph, generating an adjacency correlation matrix of the line graph topology. The second-order tensor transformation operation includes traversing the adjacency tensor of the local induced subgraph, identifying each directly connected original edge; assigning a unique identifier to each original edge, adding the identifier as a new node to the line graph topology, thereby transforming the original edge into a line graph node; if two original edges in the local induced subgraph share the same original node, establishing a connection relationship between the corresponding two line graph nodes; counting all line graph nodes and their connection relationships, constructing a sparse matrix with all zero values ​​as the adjacency correlation matrix of the line graph topology, and setting the values ​​at the corresponding connection positions to 1.

[0035] The feature projection module performs a linear transformation on the node feature matrix of the local induced subgraph. Utilizing the index mapping relationship between nodes and edges, it aligns the transformed node feature vectors to the edge dimension of the line graph topology, generating an initial line graph edge feature matrix. The linear transformation includes constructing a neural network layer with a preset number of layers (1), setting the input dimension to 128 and the output dimension to 64. The feature vector of each node in the local induced subgraph is input into this neural network layer for calculation, compressing the feature dimension from 128 to 64. The index mapping relationship includes establishing a lookup table between the original node index and the original edge index, finding the two original nodes connected by the original edge according to this lookup table, and performing element-wise addition on the 64-dimensional feature vectors of these two original nodes after the linear transformation. The resulting 64-dimensional vector is used as the initial feature of the line graph node corresponding to the edge, ultimately generating an initial line graph edge feature matrix containing the features of all line graph nodes.

[0036] The line graph convolution module is used to construct a message-passing neural network based on the adjacency correlation matrix of the line graph topology, and to perform neighbor node aggregation and feature update iterative operations on the initial edge feature matrix of the line graph to obtain the hidden state feature matrix of the line graph nodes. The construction includes setting the number of layers of the message-passing neural network to 2 and defining the message-passing function as the average aggregation function. The neighbor node aggregation includes finding the directly connected neighbor line graph nodes of each line graph node according to the adjacency correlation matrix, summing the feature vectors of the neighbor line graph nodes and taking the average value. The feature update includes inputting the aggregated average vector into the activation function for nonlinear transformation, and using the transformation result as the new feature vector of the current line graph node. The above aggregation and update process is repeated twice. After the iteration is completed, the hidden state feature matrix of the line graph nodes containing the final feature states of all line graph nodes is output.

[0037] The topology mapping and reconstruction module performs an inverse dimensional transformation on the hidden state feature matrix of the line graph nodes, mapping the features of the line graph nodes back to the node space of the original topology graph. It then updates the feature vectors of the central nodes using a self-loop edge weighting mechanism, outputting a local micro-topology enhancement feature matrix. The inverse dimensional transformation includes reassigning the hidden state vectors of the line graph nodes to the corresponding edges in the original topology graph based on previous mapping records. The self-loop edge weighting mechanism involves traversing each central node in the original topology graph and performing a weighted summation of the hidden state vectors of all edges with that central node as an endpoint. The weight coefficient for self-loop edges is preset to 2, and the weight coefficient for non-self-loop edges is preset to 1. The weighted update involves concatenating the summation result with the original feature vector of the central node, mapping the dimension of the concatenated vector to a preset output dimension, and generating the final local micro-topology enhancement feature matrix.

[0038] In this embodiment, the TGN model includes a weighted aggregation layer, a feature fusion layer, a joint memory module, a temporal encoding layer, and a temporal attention computation layer: The weighted aggregation layer is used to obtain the neighbor feature matrix and global importance weight vector of a node, perform the Hadamard product operation of the feature matrix and weight vector and the weighted summation operation along the adjacency dimension, and output the aggregated neighbor feature vector. The obtaining includes reading the feature matrix of a preset number of neighbor nodes and the global importance weight vector corresponding to each neighbor node, setting the dimension of the neighbor feature matrix to 128, and setting the length of the weight vector to the number of neighbors. The Hadamard product operation includes performing element-wise multiplication of each value in the global importance weight vector with the feature vector of the corresponding row in the neighbor feature matrix to obtain the weighted neighbor feature matrix. The weighted summation operation includes summing the weighted neighbor feature matrix along the row dimension, using the summation result as the aggregated feature representation, and outputting the aggregated neighbor feature vector with a dimension of 128.

[0039] The feature fusion layer is used to perform a tensor concatenation operation along the feature dimension on the aggregated neighbor feature vector and the local micro-topology enhancement feature to generate a node input vector that fuses static features. The tensor concatenation operation includes concatenating the aggregated neighbor feature vector and the local micro-topology enhancement feature vector end-to-end along the feature dimension, where the aggregated neighbor feature vector has a dimension of 128 and the local micro-topology enhancement feature has a dimension of 64. After concatenation, a one-dimensional vector with a dimension of 192 is generated. This one-dimensional vector is used as the node input vector for fusing static features and is used as the input for subsequent temporal feature extraction steps.

[0040] The joint memory module is used to construct a temporal memory network containing a node memory matrix, parse the timestamp and interaction event stream of the current time step, update the historical memory state through a gated recurrent neural network, and retrieve and output the node memory state vector at the current moment. The construction includes initializing a node memory matrix with a dimension of 256 to store the state information of nodes at historical time steps. The update includes inputting the timestamp and interaction event stream of the current time step into the gated recurrent neural network unit, calculating the activation state of the forget gate, input gate, and output gate, and using the activation state to update the historical memory state numerically. The retrieval includes extracting the row vector corresponding to the current node from the updated node memory matrix, using it as the node memory state vector at the current moment, and outputting a vector with a dimension of 256.

[0041] The time encoding layer is used to obtain the original timestamp, map the timestamp to a high-dimensional Riemannian manifold space, or perform a nonlinear mapping using a multilayer perceptron to generate a time encoding embedding vector. The obtaining includes reading the timestamp value of the current interaction event and converting the timestamp value into the number of hours relative to the reference time. The nonlinear mapping includes constructing a multilayer perceptron with three hidden layers, with an input layer dimension of 1 and an output layer dimension of 64. The relative hours are input into the multilayer perceptron, and after linear transformation and calculation using the Sigmoid nonlinear activation function, a time encoding embedding vector containing time periodic features is generated, with an output time encoding embedding vector of dimension 64.

[0042] The temporal attention computation layer is used to construct a multi-head attention mechanism. It uses the node input vector as the query vector and the linear combination of the node memory state vector and the temporal encoding embedding vector as the key-value pair. It calculates the spatiotemporal correlation score between features, performs weighted aggregation and residual connection based on the score, and outputs node features that fuse local patterns and global importance. The construction includes setting the number of attention heads to 8, with each attention head having a dimension of 32. The linear combination includes adding the node memory state vector (dimension 256) to the temporal encoding embedding vector (dimension 64) to generate a key vector and a value vector. The computation includes calculating the dot product of the query vector and the key vector, dividing the dot product by a preset scaling factor of 8, and then calculating the correlation score using a normalized exponential function. The weighted aggregation includes weighted summation of the value vector based on the correlation score, performing residual addition with the summation result and the node input vector, and finally outputting node features with a dimension of 192.

[0043] In this embodiment, the improved Longformer model includes a Mostin manifold mapping layer, a reweighted spectral coding layer, a two-stream attention deconstruction layer, and a global memory compression layer: The Mostin manifold mapping layer is used to construct a Poincaré sphere model, obtain the original node feature vectors, project the original node feature vectors from Euclidean space to hyperbolic space through the logarithmic mapping of the tangent space of the Riemannian manifold, calculate the gradient update under the manifold curvature parameter, and output the hyperbolic node feature tensor. The construction of the Poincaré sphere model includes setting the curvature parameter of the hyperbolic space to 1.0 and defining the dimension of the model as 64. The logarithmic mapping of the tangent space includes constructing a tangent plane at the position of the original node feature vector, calculating the mapping relationship between the Euclidean space vector and the origin of the hyperbolic space, and projecting the original node feature vectors to the hyperbolic space using the logarithmic mapping formula to obtain the preliminary hyperbolic features. The gradient update includes calculating the gradient value of the manifold curvature parameter according to the preset learning rate of 0.001 during backpropagation, dynamically adjusting the curvature parameter to optimize the feature distribution, and finally outputting a hyperbolic node feature tensor with a dimension of 64.

[0044] The reweighted spectral encoding layer is used to obtain the graph structure adjacency matrix, construct a normalized graph Laplacian matrix and perform eigenvalue decomposition. It reweights the feature components using a pre-defined learnable spectral filter matrix, suppressing high-frequency noise components and enhancing low-frequency structural components. The reweighted spectral features are projected back into hyperbolic space via Riemann exponent mapping and fused with the hyperbolic node feature tensor using Mostine addition to output a manifold structure-aware tensor. Constructing the normalized graph Laplacian matrix includes calculating the degree matrix of the graph structure adjacency matrix, calculating the Laplacian matrix using the degree matrix and adjacency matrix, and performing normalization. The eigenvalue decomposition includes... The eigenvalue decomposition of the uniform graph Laplacian matrix is ​​performed to extract the top 50 eigenvalues ​​and their corresponding eigenvectors. The reweighting includes initializing a learnable spectral filter vector of length 50, multiplying the eigenvectors by the filter vector, with the filter values ​​corresponding to high frequencies preset close to 0 and the filter values ​​corresponding to low frequencies preset close to 2, thereby suppressing the noise enhancement structure. The Mostin addition fusion includes projecting the reweighted spectral features back to the hyperbolic space through the Riemann exponent mapping, and using the Mostin addition formula to perform manifold addition operation on the projected spectral features and the hyperbolic node feature tensor, outputting a manifold structure perception tensor with a dimension of 64.

[0045] The dual-stream attention deconstruction layer is used to divide the manifold structure perception tensor into non-overlapping local blocks along the sequence dimension, construct a block diagonal mask matrix, perform multi-head self-attention operation on the node features within the local blocks, and output a local fine-grained feature tensor; simultaneously, it extracts the global feature matrix of the manifold structure perception tensor, decomposes it into a tensor ring form of core tensor and factor matrix, calculates global correlation through tensor ring shrinkage operation, and outputs a global low-rank feature tensor; the partitioning includes dividing the manifold structure perception tensor with a sequence length of 512 into 4 local blocks with a length of 128; The block diagonal mask matrix involves constructing a 512x512 matrix, setting valid values ​​only at the corresponding local block positions, and setting negative infinity at the remaining positions; the multi-head self-attention operation involves setting the number of attention heads to 4, calculating attention scores using the block diagonal mask matrix, extracting local features, and outputting a local fine-grained feature tensor with a dimension of 64; the tensor ring decomposition involves decomposing the global feature matrix into 3 core tensors, each with a dimension of 4, simulating global contextual relevance through tensor ring contraction operations, and outputting a global low-rank feature tensor with a dimension of 64.

[0046] The global memory compression layer is used to initialize the global memory tokens of the tensor ring structure. It performs pooling operations along the node dimension of the global low-rank feature tensor, maps the pooling results to the core factors of the tensor ring, and uses a gated recurrent unit to perform temporal state updates on the core factors. The global context vector is reconstructed through tensor ring expansion. The initialization includes constructing three 4-dimensional tensors as global memory tokens. The pooling operation includes calculating the average value of the global low-rank feature tensor along the node dimension to obtain a pooling vector. The mapping includes linearly transforming the pooling vector and adding it to the three tensor core factors. The temporal state update includes inputting the updated core factors into the gated recurrent unit to calculate the new memory state. The reconstruction includes expanding the core factors containing the new memory state using the tensor ring, recombine them into a complete vector, and output a 64-dimensional global context vector.

[0047] The process involves concatenating and normalizing the local fine-grained feature tensor and the global context vector along the feature dimension, followed by linear transformation and GELU nonlinear activation through a fully connected layer to output a deep context feature vector. The concatenation process involves linking the 64-dimensional local fine-grained feature tensor and the 64-dimensional global context vector end-to-end along the feature dimension to generate a combined feature vector of dimension 128. The layer normalization process involves calculating the mean and variance of the combined feature vector along the feature dimension and standardizing the vector using the mean and variance. The fully connected layer includes constructing a linear transformation layer with an input dimension of 128 and an output dimension of 64. The GELU nonlinear activation process involves inputting the result of the linear transformation into the GELU activation function for computation, outputting a final deep context feature vector of dimension 64.

[0048] In this embodiment, S7 specifically includes: S71. Obtain the deep context feature vector, construct a multilayer perceptron mapping head, perform a linear transformation of the feature space through the weight matrix, and process the output vector layer by layer using the GELU nonlinear activation function to obtain a classification feature vector; the construction includes setting the multilayer perceptron mapping head to contain two fully connected layers, the first layer has an input dimension of 64 and an output dimension of 32, and the second layer has an input dimension of 32 and an output dimension of 16; the linear transformation includes initializing the weight matrices of the first and second layers, and performing matrix multiplication operations on the deep context feature vector and the weight matrix of the first layer; the layer-by-layer processing includes inputting the operation result of the first layer into the Tanh nonlinear activation function for calculation, using the calculated result as the input of the second layer for matrix multiplication operations again, and finally outputting a classification feature vector with a dimension of 16.

[0049] S72. Input the classification feature vector into the normalized exponential function layer, calculate the exponential values ​​of the vector elements and perform normalization processing to generate posterior probability distributions belonging to each preset category, and extract the index corresponding to the maximum value in the posterior probability distribution; the calculation includes reading the 16 element values ​​in the classification feature vector, and calculating the exponential operation result of each element value using the natural constant as the base; the normalization processing includes summing all exponential operation results, calculating the proportion of each exponential operation result to the sum, and combining all the calculated proportion values ​​into a one-dimensional vector, which is the posterior probability distribution; the extraction includes traversing all proportion values ​​in the posterior probability distribution vector, finding the largest proportion value, and recording the position index of the maximum value in the vector as an index.

[0050] S73. Map the index to a predefined tag set, obtain the corresponding category semantics, encapsulate the category semantics and timestamp information, and output the target mining result for a future time. The mapping includes establishing a predefined tag set containing 16 element category names, searching for the category name corresponding to the position number extracted in the previous step in the tag set, and using the category name as the category semantics. The encapsulation includes obtaining the timestamp value of the current data processing, and combining and packaging the category semantics and timestamp value according to a preset key-value pair format. The output includes outputting the packaged data as the final result through a preset interface or file to generate a target mining result for a future time containing the predicted category and occurrence time.

[0051] Example 1: To verify the effectiveness of this invention in data mining and decision support in complex business environments, the method was applied to the intelligent business intelligence analysis platform of a large multinational enterprise group (hereinafter referred to as "Group Q"). Traditional enterprise operational data analysis systems typically employ rule-based database queries or simple time-series statistical models. These methods not only struggle to automatically extract deep semantic relationships from multi-source heterogeneous data, but also fail to effectively capture the complex dynamic evolution patterns and long-range dependencies between enterprise entities, easily leading to misjudgments of future market trends or the omission of key business opportunities. To address these issues, Group Q decided to adopt the deep learning-based enterprise data mining system and method proposed in this invention.

[0052] During implementation, Group Q first utilized its internal ERP system, CRM platform, and external social media interfaces to collect multi-source heterogeneous data, including transaction records, customer interaction logs, and market sentiment texts. It then used a time discretization operator to segment the continuous time axis into windows, setting the time window length to 24 hours and the sliding step to 1 hour, constructing a dynamic snapshot sequence stacked along the time dimension. Simultaneously, Group Q's data scientists precisely categorized and labeled key historical events with target values, serving as a benchmark for model training and performance evaluation.

[0053] Group Q, through its node weight calculation module, iteratively updates the random walk damping parameters based on node degree centrality and edge weights, executes the page ranking algorithm until convergence, and outputs the global importance weights representing the influence of nodes in the network topology. Next, using a Line Graph Neural Network (LGNN), it samples local subgraphs centered on high-weight nodes, within a range limited by the shortest path distance. Through edge-to-node mapping and convolution aggregation, it uses a self-loop mechanism to back-map edge features back to the central node, outputting local micro-topological enhancement features, effectively eliminating the interference of redundant information. Subsequently, the nodes incorporating these local micro-topological enhancement features are input into the TGN model. Combined with a joint memory module and temporal encoding attention, it extracts temporal evolution features, outputting node features that integrate local patterns and global importance.

[0054] In the core mining and prediction stage, this invention introduces an improved Longformer model, utilizing a Mostinian manifold mapping layer to project the original node feature vectors onto a hyperbolic space, and combining a reweighted spectral coding layer to suppress high-frequency noise and enhance low-frequency structural components. Simultaneously, a dual-stream attention deconstruction layer and a global memory compression layer are employed, using a sliding window expansion attention mechanism to calculate local context and capturing long-range dependencies through global attention tokens, outputting a deep context feature vector. Finally, the system calculates the posterior probability distribution through a fully connected layer and a normalized exponential function classifier, accurately outputting the enterprise target mining results for future timeframes.

[0055] During implementation, the technical team at Group Q discovered that, compared to traditional data analysis methods and conventional data mining algorithms, the method of this invention significantly improves the accuracy and timeliness of enterprise target prediction. Traditional methods cannot automatically integrate global topological structure and local micro-features, and have a weak ability to capture long-term nonlinear dependencies. In contrast, the method of this invention, through an adaptive random walk strategy, Riemannian manifold mapping, and a two-stream attention deconstruction mechanism, effectively achieves deep characterization of complex business data and accurate prediction of future trends.

[0056] To further verify the actual performance of the method of the present invention, Group Q conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Comparison of Data Mining and Target Prediction Performance of Group Q Enterprise

[0057] As shown in Table 1, the performance of the enterprise data mining and target prediction system was comprehensively improved after applying the method of this invention. The target mining accuracy increased from 78.4% with traditional methods to 94.2%, and the long-term trend prediction error rate decreased from 25.6% to 6.3%, significantly improving the reliability of decision-making. The key entity identification recall rate increased from 81.5% to 96.5%, effectively avoiding the omission of important business information. The time for multi-source data fusion was significantly reduced from 450 seconds to 120 seconds, significantly enhancing system processing efficiency. Furthermore, the early warning lead time increased from 3 days to 15 days, and the cost of manual analysis decreased from 2 million yuan / year to 900,000 yuan / year, significantly reducing operating costs. Business department satisfaction also significantly improved, increasing from 88.0% to 98.5%.

[0058] Through the method of this invention, Group Q has successfully achieved in-depth value mining of multi-source heterogeneous data and accurate prediction of future goals, effectively improving market response speed and strategic decision-making level, significantly enhancing the intelligence and foresight of enterprise operations, significantly reducing the workload of business analysts, enhancing the robustness and interpretability of the data mining system, and providing strong technical support for enterprise digital transformation and intelligent upgrading.

[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based enterprise data mining system and method, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data from enterprises, use time discretization operators to perform window segmentation on the continuous time axis, construct topology graph snapshots based on entity instances and interaction relationships within each window, and output a dynamic snapshot sequence arranged in order by timestamp index; S2. Input the dynamic snapshot sequence into the node weight calculation module, iteratively update the random walk damping parameters based on the node degree centrality and edge weights, execute the page ranking algorithm iterative calculation until it converges to a steady-state distribution, and output the global importance weight that represents the influence of the node in the network topology. S3. Input the global importance weight and topological structure index into the location coding generator, extract the node degree, clustering coefficient and multi-hop neighbor number to construct a structured feature vector, perform nonlinear mapping through a multilayer perceptron, and output a graph structure-guided location code containing topological physical location and semantic social status. S4. For the topological graph snapshot, take the high-weight node as the center, sample the local subgraph based on the shortest path distance range, construct the line graph topology and input it into the LGNN model. Through edge-to-node mapping and convolution aggregation, use the self-loop mechanism to reverse map the edge features back to the center node and output the local micro-topological enhancement features. S5. Perform weighted summation and aggregation on the features of neighboring nodes using global importance weights, fuse the results with local micro-topology enhancement features and input them into the TGN model, combine the joint memory module and time-coded attention to extract temporal evolution features, and output node features that fuse local patterns and global importance. S6. Perform additive normalization on node features and graph structure-guided position encoding to generate model input tensors. Input the improved Longformer model to calculate local context using a sliding window expansion attention mechanism, and capture long-range dependencies through global attention tokens to output deep context feature vectors. S7. Input the deep context feature vector into the fully connected layer, perform decision boundary mapping through linear transformation and nonlinear activation function, calculate the posterior probability distribution using a normalized exponential function classifier, and output the target mining results for future time moments.

2. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Parse the enterprise's multi-source heterogeneous data, extract entity identifiers, interaction types and timestamp fields through regular expression matching, and construct the original event stream containing node initial attributes and edge interaction records; S12. Set the time window length and sliding step size, use the time discretization operator to perform discretization mapping on the timestamp values ​​contained in the original event stream to generate discrete indexes, and allocate interactive events to the corresponding time windows according to the discrete indexes to generate a set of discrete time segments. S13. Traverse the set of discrete time segments, extract the unique identifiers of entity instances belonging to each discrete time segment as a set of nodes, extract the entity interaction associations belonging to each discrete time segment as a set of edges, and construct a graph topology skeleton representing the node connection relationship. S14. Calculate the interaction frequency statistics of the node set under the corresponding discrete time segment, map the interaction frequency statistics to edge weights, map the node attributes belonging to the current time segment to node features, and fuse the edge weights and node features into the graph topology skeleton to generate a single-frame topology graph snapshot. S15. Serialize and arrange the single-frame topology snapshots according to the order of timestamp indices, construct data tensors stacked along the time dimension, and output dynamic snapshot sequence.

3. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Extract the adjacency matrix and edge weight vector of a single-frame topology snapshot, calculate the sum of the weights of the associated edges of nodes to obtain the node strength, and perform extreme value standardization processing on the node strength using a preset normalization interval to generate a node degree centrality index. S22. Construct a nonlinear mapping function for the damping parameters, input the node degree centrality index into the mapping function to perform numerical transformation, and perform interval constraint operation on the transformed values ​​to generate adaptive damping parameters for balancing random walk and restart probability. S23. Construct a basic transition probability matrix based on the edge weight vector, perform weighted fusion processing on the basic transition probability matrix using adaptive damping parameters to construct a random walk transition matrix, initialize the node importance score vector, iteratively update the node scores through matrix-vector product, and calculate the Euclidean distance between the node score vectors before and after the iterative update. S24. Determine whether the Euclidean distance is less than the preset convergence threshold. If so, stop the iterative calculation and output the current node importance score vector as the global importance weight. Otherwise, continue to update the node importance score vector based on the random walk transition matrix.

4. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, Specifically, S3 includes: S31. Call the position encoding generator to parse the adjacency relationship of a single frame topology snapshot, and calculate the first degree of the node, the local clustering coefficient, and the multi-hop neighbor statistics within the preset number of hops; S32. Map the degree, clustering coefficient and multi-hop neighbor statistics to numerical tensors, perform feature dimension concatenation operation, and generate node local topology vectors. S33. Construct a multilayer perceptron model containing fully connected layers and activation functions. Use the local topological structure vector of the node as the model input for forward propagation. Calculate the product of the input tensor and the weight matrix and add the bias term. Perform nonlinear mapping transformation to generate topological feature embeddings. S34. Obtain the global importance weights, embed the topological features and perform tensor fusion operations with the global importance weights to generate graph structure-guided position encoding.

5. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, The LGNN model includes a subgraph sampling module, a line graph construction module, a feature projection module, a line graph convolution module, and a topology mapping reconstruction module. The subgraph sampling module is used to parse the adjacency matrix of the global topology graph, use a preset set of high-weight node indices as anchor points to calculate the shortest path distance matrix between nodes, filter the subset of nodes whose path distance is less than a preset threshold, and construct the adjacency tensor of the local induced subgraph. The line graph construction module is used to perform second-order tensor transformation operations based on the adjacency tensor of the local induced subgraph, mapping the nodes in the local induced subgraph to the edges in the line graph, mapping the connection relationships between nodes to the nodes in the line graph, and generating the adjacency association matrix of the line graph topology. The feature projection module is used to perform a linear transformation on the node feature matrix of the local induced subgraph. By using the index mapping relationship between nodes and edges, the transformed node feature vectors are aligned to the edge dimension of the line graph topology to generate the line graph initialization edge feature matrix. The line graph convolution module is used to construct a message-passing neural network based on the adjacency association matrix of the line graph topology, and to perform neighbor node aggregation and feature update iteration operations on the initial edge feature matrix of the line graph to obtain the hidden state feature matrix of the line graph nodes. The topology mapping and reconstruction module is used to perform inverse dimensional transformation on the hidden state feature matrix of the line graph node, map the features of the line graph node back to the node space of the original topology graph, and update the feature vector of the central node by weighting through the self-loop edge weighting mechanism, and output the local micro-topology enhancement feature matrix.

6. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, The TGN model includes a weighted aggregation layer, a feature fusion layer, a joint memory module, a temporal encoding layer, and a temporal attention computation layer. The weighted aggregation layer is used to obtain the neighbor feature matrix and global importance weight vector of a node, perform the Hadamard product operation of the feature matrix and weight vector and the weighted summation operation along the adjacency dimension, and output the aggregated neighbor feature vector. The feature fusion layer is used to perform a tensor concatenation operation along the feature dimension on the aggregated neighbor feature vector and the local micro-topology enhancement feature to generate a node input vector that fuses static features. The joint memory module is used to construct a temporal memory network containing a node memory matrix, parse the timestamp and interactive event stream of the current time step, update the historical memory state through a gated recurrent neural network, and retrieve and output the node memory state vector at the current moment. The time coding layer is used to obtain the original timestamp, map the timestamp to a high-dimensional Riemannian manifold space or perform a nonlinear mapping using a multilayer perceptron to generate a time coding embedding vector. The temporal attention computation layer is used to construct a multi-head attention mechanism. It takes the node input vector as the query vector, the linear combination of the node memory state vector and the temporal encoding embedding vector as the key-value pair, calculates the spatiotemporal correlation score between features, performs weighted aggregation and residual connection based on the score, and outputs node features that fuse local patterns and global importance.

7. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, The improved Longformer model includes a Mostin manifold mapping layer, a reweighted spectral coding layer, a two-stream attention deconstruction layer, and a global memory compression layer: The Mostin manifold mapping layer is used to construct the Poincaré sphere model, obtain the original node feature vectors, project the original node feature vectors from Euclidean space to hyperbolic space through the logarithmic mapping of the tangent space of the Riemann manifold, calculate the gradient update under the manifold curvature parameter, and output the hyperbolic node feature tensor. The reweighted spectral coding layer is used to obtain the graph structure adjacency matrix, construct the normalized graph Laplacian matrix and perform eigenvalue decomposition, reweight the feature components using a preset learnable spectral filter matrix, suppress high-frequency noise components and enhance low-frequency structural components, project the reweighted spectral features back to the hyperbolic space through Riemann exponent mapping, and fuse them with the hyperbolic node feature tensor using Mostine addition to output the manifold structure perception tensor. The dual-stream attention deconstruction layer is used to divide the manifold structure perception tensor into non-overlapping local blocks along the sequence dimension, construct a block diagonal mask matrix, perform multi-head self-attention operation on the node features within the local blocks, and output a local fine-grained feature tensor; at the same time, it extracts the global feature matrix of the manifold structure perception tensor, decomposes it into a tensor ring form of core tensor and factor matrix, calculates global correlation through tensor ring shrinkage operation, and outputs a global low-rank feature tensor. The global memory compression layer is used to initialize the global memory token of the tensor ring structure, perform pooling operation on the global low-rank feature tensor along the node dimension, map the pooling result to each core factor of the tensor ring, perform temporal state update on the core factors using the gated loop unit, and reconstruct the global context vector through tensor ring expansion operation. The local fine-grained feature tensor and the global context vector are concatenated and normalized along the feature dimension. Linear transformation and non-linear activation are then performed through a fully connected layer to output a deep context feature vector.

8. The enterprise data mining system and method based on deep learning according to claim 1, characterized in that, S7 includes the following steps: S71. Obtain the deep context feature vector, construct the multilayer perceptron mapping head, perform linear transformation of the feature space through the weight matrix, and process the output vector layer by layer using the nonlinear activation function to obtain the classification feature vector. S72. Input the classification feature vector into the normalization exponential function layer, calculate the exponential value of the vector element and perform normalization processing to generate the posterior probability distribution belonging to each preset category, and extract the index corresponding to the maximum value in the posterior probability distribution. S73. Map the index to a predefined set of labels, obtain the corresponding category semantics, encapsulate the category semantics with timestamp information, and output the target mining results for future moments.

Images