A data intelligent visualization method and system for enterprise management

Abstract

The application discloses a data intelligent visualization method and system for enterprise management, and relates to the field of enterprise management, and comprises the following steps: obtaining historical business data from an enterprise platform; extracting semantic features of the historical business data to obtain a semantic feature vector; constructing a local semantic matrix according to the semantic feature vector; setting a multi-level time window, and performing sequential pattern mining on the historical business data in each time window based on the local semantic matrix to obtain business interaction modes of different time granularities; constructing a business relationship network according to the business interaction modes of different time granularities; performing business trend analysis according to the business relationship network, and generating a business trend analysis visualization result. Since the business relationship of an enterprise is not static, traditional methods often simplify the business relationship into a static network and ignore the evolution process, and the application simultaneously captures the business interaction modes of different time granularities by designing a multi-level time window, so that the dynamic evolution characteristics of the business relationship are analyzed.

Description

Technical Field

[0001] This application relates to the field of enterprise management, and in particular to a data intelligence visualization method and system for enterprise management. Background Technology

[0002] Business relationships within a company are complex, dynamically evolving systems, constantly changing under the influence of various internal and external factors. For example, supply chain relationships adjust to fluctuations in supply and demand and seasonal variations; customer relationships evolve with changes in customer needs, market competition, and adjustments to corporate marketing strategies; and internal business processes are optimized through organizational restructuring, the application of new technologies, and business process reengineering. This dynamism necessitates the incorporation of a time dimension into business relationship analysis to comprehensively and accurately grasp the patterns of business operations.

[0003] Traditional methods for analyzing business relationships have significant shortcomings. First, most methods simplify business relationships into static networks, ignoring their dynamic evolution over time. While this simplification reduces analytical complexity, it fails to reflect the development trends and patterns of change in business relationships, resulting in a lack of time-dimensional insights in the analysis. Second, existing methods typically focus on a single time granularity (such as day, month, or year), making it difficult to simultaneously capture short-term fluctuations and long-term trends, and thus failing to comprehensively reflect the changing characteristics of business behavior across different time scales.

[0004] For example, patent document CN118551040B discloses a method and system for visualizing business data on an enterprise platform. This method includes the following steps: acquiring historical business logs based on the enterprise platform; constructing a global semantic graph from the historical business logs to obtain a semantic feature knowledge graph; performing multi-dimensional anomaly detection on the historical business logs based on the semantic feature knowledge graph to generate outlier difference values; correcting outliers in the historical business logs based on the outlier difference values ​​to obtain outlier-corrected business logs; mining business interactions in the outlier-corrected business logs to obtain business interaction lines; fitting topological associations to the business interaction lines to construct a business topology association network; and obtaining the business attributes of each interaction line based on the outlier-corrected business logs. However, this solution is mainly based on a static semantic knowledge graph and lacks the ability to analyze the dynamic evolution of enterprise business relationships. Summary of the Invention

[0005] Since business relationships are not static but dynamically changing over time, traditional methods often simplify business relationships into static networks, ignoring their evolution process. This application provides a data intelligence visualization method and system for enterprise management. By designing multi-level time windows and capturing business interaction patterns at different time granularities, the dynamic evolution characteristics of business relationships can be analyzed.

[0006] One aspect of this application provides a data intelligence visualization method for enterprise management, comprising: S1, acquiring historical business data from an enterprise platform; S2, extracting semantic features from the historical business data to obtain semantic feature vectors; S3, constructing a local semantic matrix based on the semantic feature vectors; S4, setting multi-level time windows, and within each time window, performing sequential pattern mining on the historical business data based on the local semantic matrix to obtain business interaction patterns at different time granularities; S5, constructing a business relationship network based on the business interaction patterns at different time granularities; and S6, performing business trend analysis based on the business relationship network and generating visualization results of the business trend analysis.

[0007] Furthermore, S2 extracts semantic features from historical business data to obtain semantic feature vectors. This includes: using a self-supervised learning framework to extract semantic features from historical business data, resulting in high-dimensional semantic feature vectors; performing locality-sensitive hashing (LSH) on the high-dimensional semantic feature vectors to obtain local semantic region partitioning results; performing dimensionality reduction on the high-dimensional semantic feature vectors within each local semantic region to generate dimensionality-reduced semantic feature vectors; and merging all dimensionality-reduced semantic feature vectors from all local semantic regions to obtain the final semantic feature vector. The combined application of LSH and dimensionality reduction techniques reduces computational complexity while preserving semantic information, enabling the system to efficiently process large-scale enterprise data and fundamentally solving the performance bottleneck of traditional methods in processing massive amounts of data.

[0008] Furthermore, self-supervised learning frameworks employ BERT, Sim CSE, or Contrastive Loss models.

[0009] Furthermore, in step S3, a local semantic matrix is ​​constructed based on the semantic feature vector, including: constructing a block-based sparse matrix based on the local semantic region segmentation results, where each block represents a local semantic region; calculating the semantic similarity between business data using the dimensionality-reduced semantic feature vector for each local semantic region, and using this as the element value of the corresponding matrix block; setting a similarity threshold and sparsifying the block-based sparse matrix; and performing time-decay weighted processing on the similarity values ​​in the sparsified block-based sparse matrix based on the time attributes of the business data to obtain the final local semantic matrix.

[0010] Furthermore, based on the temporal attributes of the business data, the similarity values ​​in the sparsely segmented block matrix after sparsification are subjected to time-decay weighting processing to obtain the final local semantic matrix. This includes: calculating the time difference between each pair of business data based on the timestamps of the business data; setting a time decay function, which is either an exponential decay function or a half-life function; calculating the time weight between each pair of business data based on the time difference and the time decay function; multiplying the time weight by the corresponding semantic similarity to obtain the time-weighted semantic similarity value; and updating the similarity values ​​in the block sparse matrix to generate a local semantic matrix containing semantic similarity and temporal correlation. By incorporating the time dimension into the semantic similarity calculation through time-decay weighting processing, the system can reflect the real-world characteristic of business relationships naturally decaying over time, solving the technical challenge of traditional static networks being unable to express the impact of time.

[0011] By establishing a hierarchical time granularity structure with daily, weekly, monthly, quarterly, and annual time windows, comprehensive capture of business relationships across different time scales is achieved. This technically overcomes the limitation of a single time window in simultaneously reflecting short-term fluctuations and long-term trends. The establishment of the time window mapping matrix provides a mathematical foundation for cross-time granularity business relationship analysis, enabling correlation analysis of multi-scale business data and effectively resolving the technical bottleneck of difficulty in integrating analysis results from different time scales.

[0012] Further, in step S4, business interaction patterns at different time granularities are obtained, including: setting multi-level time windows of daily, weekly, monthly, quarterly, and annual periods to form a time granularity hierarchy; establishing a mapping relationship matrix of time windows based on the time attributes of business data and business processes, wherein the mapping relationship matrix reflects the inclusion and overlap relationships between each time window; slicing the local semantic matrix generated in step S3 according to different time windows to obtain sub-matrices within the corresponding time windows; calculating the sequential dependency relationships between business data based on the sub-matrices of each time window to construct a directed weighted graph; wherein nodes are business data and the weights of edges are similarity values ​​in the sub-matrices; performing frequent sequential pattern mining on the directed weighted graphs of different time windows; calculating the time density of frequent sequential patterns; and integrating frequent sequential patterns across time granularities under different time windows based on the mapping relationship matrix of time windows and the time density of frequent sequential patterns to obtain business interaction patterns at different time granularities.

[0013] Furthermore, frequent sequence pattern mining is performed on directed weighted graphs for different time windows, including: normalizing the edge weights of the directed weighted graphs; calculating the weighted average of all edge weights in the frequent sequence pattern as the weighted support of the corresponding pattern based on the normalized edge weights; and setting support thresholds for daily, weekly, monthly, quarterly, and annual time windows respectively. , , , and ;in, Create a node expansion priority queue in each time window; calculate the median of all edge weights in the current time window as a reference threshold; iteratively expand the initial frequent patterns whose weighted support is greater than the corresponding support threshold. When the weight of the candidate expansion edge is greater than the reference threshold, insert the target node connected by the corresponding edge into the head of the node expansion priority queue, otherwise insert it into the tail, thus obtaining the set of frequent order patterns in each time window.

[0014] Among them, the frequent sequence pattern mining algorithm based on directed weighted graphs can capture the temporal dependencies in business processes, mathematically solving the problem of traditional methods ignoring the temporal characteristics of business processes. This is achieved through differentiated support threshold design (…). The system can adaptively adjust the sensitivity of pattern mining in different time windows, technically solving the problem of inconsistent accuracy in long and short-term pattern recognition. The ingenious combination of node expansion priority queues and reference thresholds enables an efficient pattern expansion search strategy, improving the efficiency and accuracy of dynamic pattern mining at the algorithmic level.

[0015] Furthermore, the time density of frequent sequence patterns is calculated, including: extracting the timestamps of business data nodes in each frequent sequence pattern and establishing a mapping table between patterns and times; calculating the time span of the first and last nodes in each frequent sequence pattern. and the variance of timestamps of each node in the frequent sequential pattern. Among them, the time span The variance reflects the overall duration of the pattern, while the time difference reflects the temporal distribution of business interactions within the pattern. The time difference between each pair of adjacent nodes in the frequent sequence pattern is calculated, i.e., the node sequence within the pattern. The corresponding timestamp is Then calculate the time difference sequence. The average of these differences is taken as the average interval time between nodes within the pattern. Based on the time span of frequent sequential patterns And the corresponding time window size T, to calculate the time coverage of the corresponding mode. Based on the variance of each node's timestamp The average interval time between nodes within the pattern calculated above. Time uniformity of the calculation pattern Based on time coverage and time uniformity Calculate the time density of frequent sequential patterns ,in, and The weighting coefficients are and satisfy the following conditions: .

[0016] Based on the calculated time density index D and the time window mapping matrix, the frequent sequence patterns under different time windows are integrated across time granularity, and the patterns with higher time density are given higher retention priority.

[0017] Among them, the designed time density index D, by combining time coverage and time uniformity, mathematically quantifies the time distribution characteristics of the business model, thus solving the technical deficiency of traditional methods in being unable to assess the time stability of the model.

[0018] The average value of the time difference sequence between adjacent nodes is used as a method to calculate the average interval time of the pattern, which can accurately reflect the rhythm characteristics of business interaction and provide technical support for identifying periodic business patterns.

[0019] Weighted combination design of time density indicators This enables adjustable assessment of the time characteristics of business models, technically meeting the differentiated needs for time continuity and uniformity in different enterprise scenarios.

[0020] Another aspect of this application provides a data intelligence visualization system for enterprise management.

[0021] Compared to existing technologies, the advantages of this application are:

[0022] On the one hand, business relationships are not static, but dynamic over time. Traditional methods often simplify business relationships into static networks, ignoring their evolution process. This solution constructs a dynamic business relationship model that integrates the time dimension by using time decay weighting, frequent sequence pattern mining, and business relationship network topology evolution feature calculation. This enables a comprehensive characterization and analysis of the evolution of business relationships over time, solving the fundamental problem that static networks cannot reflect the evolution of business relationships.

[0023] On the other hand, enterprise business interaction patterns exhibit varying characteristics at different time granularities (short-term, medium-term, and long-term); traditional methods typically focus only on a single time dimension, failing to comprehensively capture business behavior. This solution designs multi-level time windows—daily, weekly, monthly, quarterly, and annual—along with a time window mapping matrix and differentiated support thresholds (…). It enables the simultaneous mining of business models at different time scales; through a two-way integration strategy of bottom-up merging and top-down decomposition, a cross-granularity pattern mapping table is established, which successfully solves the problem of integration and unified analysis of business models at multiple time scales, and technically achieves the collaborative perception and prediction of short-term fluctuations and long-term trends. Attached Figure Description

[0024] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0025] Figure 1 This is an exemplary flowchart illustrating a data intelligence visualization method for enterprise management according to some embodiments of this application;

[0026] Figure 2 This is an exemplary flowchart illustrating the construction of semantic feature vectors according to some embodiments of this application;

[0027] Figure 3 This is an exemplary flowchart illustrating the acquisition of business interaction patterns at different time granularities according to some embodiments of this application;

[0028] Figure 4 This is an exemplary flowchart illustrating the construction of a business relationship network according to some embodiments of this application. Detailed Implementation

[0029] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0030] like Figure 1 As shown, historical business data is obtained from the enterprise platform; semantic features of the historical business data are extracted to obtain semantic feature vectors; a local semantic matrix is ​​constructed based on the semantic feature vectors; multiple time windows are set, and within each time window, sequential pattern mining is performed on the historical business data based on the local semantic matrix to obtain business interaction patterns at different time granularities; a business relationship network is constructed based on the business interaction patterns at different time granularities; business trend analysis is performed based on the business relationship network, and a visualization of the business trend analysis results is generated.

[0031] S1 acquires historical business data from the enterprise platform. First, the system interfaces with multiple internal business systems within the enterprise via APIs, database connections, or ETL tools, including but not limited to Enterprise Resource Planning (ERP) systems, Customer Relationship Management (CRM) systems, Supply Chain Management (SCM) systems, Human Resource Management (HRM) systems, and other business platforms. The system supports the acquisition of various data formats, including structured data (such as database tables), semi-structured data (such as JSON, XML, etc.), and unstructured data (such as business documents, emails, etc.). Second, the system performs data quality assessment on the acquired raw data, including integrity checks, consistency verification, timeliness assessment, and outlier identification. For identified low-quality data, the system uses a rule engine for automatic correction; for data that cannot be automatically corrected, the system marks and records the problem type for subsequent analysis. Then, the system performs time standardization processing on the data, converting timestamps from different source systems into a unified standard time format and setting the time granularity according to business needs, such as accuracy to seconds, minutes, hours, or days. Simultaneously, the system also performs entity standardization on the business data, mapping identifiers representing the same business entities in different systems to a unified entity ID.

[0032] like Figure 2 As shown in Figure S2, semantic features of historical business data are extracted to obtain semantic feature vectors. The system first employs a self-supervised learning framework to extract semantic features from historical business data. Based on the type and characteristics of the business data, the system selects an appropriate pre-trained model: for text-based business data, the system uses the BERT pre-trained model and fine-tunes it using enterprise domain knowledge; for business data requiring similarity comparison, the system uses the Sim CSE model; and for data requiring differentiation of business categories, the system uses a model based on Contrastive Loss.

[0033] During model training, the system employs a domain-adaptive training strategy, using internal enterprise business corpora to perform secondary training on the pre-trained model, enhancing its ability to understand specific business terms and expressions. For each piece of business data, the system calculates its representation vector in a high-dimensional semantic space through forward propagation. The vector dimension is typically between 768 and 1024, enabling it to capture the deep semantic information of the business data.

[0034] For the acquired high-dimensional semantic feature vectors, the system uses the Locality Sensitive Hash (LSH) algorithm to process them, mapping semantically similar business data to the same or similar hash buckets, thereby achieving local region partitioning of the semantic space. In specific implementation, the system uses the multi-level random projection LSH method, generating multiple hash functions through multiple sets of random projection vectors, and generating a hash code for each piece of business data.

[0035] The system clusters the service data based on the hash coding similarity of the service data, and service data with similarity exceeding a preset threshold (usually set to 0.75) is divided into the same local semantic region. This method can effectively solve the "curse of dimensionality" problem in high-dimensional semantic space and improve the efficiency of subsequent analysis. The division result of local semantic regions provides grouping information of service activities, identifies which service activities should be placed in the same local region, and lays a foundation for the subsequent construction of semantic matrix.

[0036] For high-dimensional semantic feature vectors in each local semantic region, the system adopts dimensionality reduction technology for processing, maps the high-dimensional vectors to a low-dimensional space, and preserves the relative relationship between vectors and semantic information at the same time. In specific implementation, the system flexibly selects the dimensionality reduction algorithm according to data characteristics and computing resources: for local regions with a small amount of data, the system adopts the t-SNE algorithm; for local regions with a large amount of data, the system adopts the UMAP algorithm; for scenarios that require retaining the global structure, the system adopts the PCA algorithm.

[0037] In the dimensionality reduction process, the system adaptively adjusts the dimensionality reduction parameters for each local semantic region, such as the target dimension after dimensionality reduction, the number of neighbors and the distance measurement method, etc., to ensure that the dimensionality reduction result can retain the original semantic information to the greatest extent. Generally, the system reduces high-dimensional vectors (768 to 1024 dimensions) to 50 to 100 dimensions, which significantly reduces the complexity of subsequent calculations. The original semantic feature vectors are optimized based on the grouping information of local semantic regions, so that the vectors in the same region have better representation and distinguishability.

[0038] Finally, the system merges the semantic feature vectors after dimensionality reduction of all local semantic regions to obtain the final set of semantic feature vectors. In the merging process, the system adopts a region adaptive normalization strategy to ensure that vectors from different local regions are comparable after merging. At the same time, the system adds metadata tags to each final semantic feature vector, including information such as the local semantic region it belongs to, original service data ID, and time stamp, which facilitates subsequent cross-region analysis and temporal pattern mining.

[0039] Through the combined application of locality-sensitive hashing and dimensionality reduction processing technology, the system significantly reduces the computational complexity while retaining key semantic information, the spatial complexity changes from to (where n is the number of service data, d is the dimension of the original high-dimensional vector, k is the dimension after dimensionality reduction, and k<<d), the time complexity changes from to . This processing method enables the system to efficiently process large-scale enterprise data containing millions of records, and fundamentally solves the performance bottleneck of traditional methods when processing massive data.

[0040] S3. Construct a local semantic matrix based on the semantic feature vectors. Based on the local semantic region segmentation results obtained in step S2, the system constructs a block-based sparse matrix structure. This matrix adopts a block-based diagonal form, with an overall dimension of N×N, where N is the total amount of business data. The matrix is ​​divided into several non-overlapping matrix blocks, each corresponding to a local semantic region. Elements within a block represent the relationships between business data within the same semantic region, while elements outside the block are initially set to zero or a minimum value.

[0041] In practice, the system first determines the block structure of the matrix based on the number K of local semantic regions, generating K main diagonal matrix blocks. The dimension of each matrix block is determined by the amount of business data within the corresponding local semantic region. The system uses a Block Compressed Sparse Row (BCSR) storage format to store this sparse matrix, significantly reducing storage space requirements and optimizing subsequent computational efficiency.

[0042] For each local semantic region, the system calculates the semantic similarity between business data within the region using the dimensionality-reduced semantic feature vectors obtained in step S2. In specific implementation, the system selects an appropriate similarity measurement method based on the characteristics of the business data: for numerical feature vectors, cosine similarity is used; for feature vectors containing category information, adjusted Jaccard similarity is used; and for business data with sequence characteristics, edit distance transformed similarity is used.

[0043] The system employs an optimized matrix calculation method to batch calculate similarity. For a local semantic region containing m business data points, the time complexity is reduced from that of the naive method. (k is the dimension of the feature vector) optimized to (b is the batch size). The calculated similarity values ​​range from [0, 1], with larger values ​​indicating stronger semantic relevance between business data. These similarity values ​​are filled into the element positions of the corresponding matrix blocks to form the initial block semantic similarity matrix.

[0044] To further improve computational efficiency and highlight important business relationships, the system performs sparsification on the block-based sparse matrix. The system employs an adaptive threshold strategy, automatically determining the similarity threshold based on the similarity distribution characteristics within each local semantic region. Specifically, it calculates the distribution statistics of similarity values ​​within each region (such as the mean μ and standard deviation σ), and sets the threshold accordingly. , where α is an adjustable parameter, typically ranging from [0.5, 1.5].

[0045] When the similarity value in the matrix is ​​lower than the threshold of the corresponding region, the system sets the value to zero, indicating that there is no significant semantic relationship between the corresponding business data. Furthermore, the system retains the connection between each business data point and its top-N most similar data (typically N is between 5 and 10), ensuring that important business relationships are not lost due to sparsity. Through this sparsity processing, the proportion of non-zero elements in the matrix can typically be reduced from 40% to 60% to 5% to 15%, significantly reducing the subsequent computational burden while preserving key business relationships.

[0046] To reflect the natural decay of business relationships over time, the system applies time-decay weighted processing to the similarity values ​​in the sparsely segmented block matrix after sparsification. The system first calculates the time difference between each pair of business data based on their timestamps. The time difference calculation employs standardization, converting the original timestamps into a uniform time unit (such as day, hour, or minute) to facilitate subsequent weighted calculations. For data with missing timestamps, the system uses either data association inference or a strategy of defaulting to the oldest timestamp.

[0047] The system supports two typical time decay functions: exponential decay function: Where λ is the decay coefficient, usually set based on business domain knowledge. For example, sales might take a larger value (0.01 to 0.05), indicating faster relationship decay; while cooperation might take a smaller value (0.001 to 0.005), indicating slower relationship decay. Half-life function: Where T is the half-life, representing the time interval required for the similarity to halve. The half-life can be set according to different business scenarios; for example, it might be set to 30 days for order relationships and 180 days for supplier relationships. The system also supports business personnel to customize the decay function according to actual needs, such as a piecewise function or a seasonally adjusted decay function, to adapt to the time characteristics of different business scenarios.

[0048] The system calculates the time difference. And with the selected time decay function f, calculate the time weight between each pair of business data. The time weight has a range of (0, 1]. When the timestamps of two business data are the same, the time weight is 1; as the time difference increases, the time weight gradually approaches 0.

[0049] The system also introduces a time-effect adjustment parameter β, and the final time weight calculation formula is as follows: Where β∈[0,1] controls the strength of the influence of the time factor. The closer β is to 1, the stronger the influence of the time factor; the closer β is to 0, the weaker the influence of the time factor. The system supports dynamic adjustment of the β value under different business scenarios.

[0050] The system multiplies the time weight by the corresponding semantic similarity to obtain a time-weighted semantic similarity value: This weighting mechanism ensures that relationships between business data that are close in time are given higher importance, while relationships between business data that are far apart in time are given lower importance, thus reflecting the reality that business relationships naturally decay over time.

[0051] The system updates the corresponding elements in the block sparse matrix using the calculated time-weighted semantic similarity values, generating the final local semantic matrix. This matrix simultaneously incorporates the semantic similarity and temporal correlation between business data. By organically integrating the time dimension into the semantic similarity calculation, the system successfully constructs a local semantic matrix that dynamically reflects the evolution of business relationships. This matrix overcomes the limitations of traditional static networks, accurately capturing the natural decay characteristics of business relationships over time, and fundamentally solving the technical challenge of static networks' inability to express the impact of time.

[0052] Furthermore, the system's multi-level time windows—daily, weekly, monthly, quarterly, and annual—form a complete hierarchical structure of time granularity, enabling comprehensive capture of business relationships across different time scales. This multi-level time window design fundamentally solves the limitation of a single time window in simultaneously reflecting short-term fluctuations and long-term trends, allowing the system to capture short-term changes in business relationships at a fine-grained level while simultaneously grasping long-term development trends at a coarse-grained level.

[0053] The system establishes a time window mapping matrix that clearly records the inclusion and overlap relationships between different time windows, providing a solid mathematical foundation for cross-time-granularity business relationship analysis. Through this matrix, the system can effectively integrate analysis results from different time scales, achieving correlation analysis of multi-scale business data and successfully solving the technical bottleneck of traditional methods that struggle to integrate analysis results from different time scales.

[0054] like Figure 3As shown in Figure S4, multiple time windows are set up. Within each time window, sequential pattern mining is performed on historical business data based on the local semantic matrix to obtain business interaction patterns at different time granularities. Daily time window: 24 hours in length, can be divided by natural days or implemented using a sliding window method; Weekly time window: 7 days in length, can be divided by natural weeks or implemented using a continuous 7-day sliding window; Monthly time window: divided by natural months, with a length of 28-31 days; Quarterly time window: each quarter contains 3 natural months; Annual time window: divided by natural year, with a length of 365 / 366 days. The system uses a tree structure to organize these time windows, with the annual window as the root node, the quarterly window as its child node, the monthly window as the child node of the quarterly window, and so on. The system supports flexible adjustment of the specific parameters of the time windows, such as adding hourly windows or adjusting the sliding method of the windows, to adapt to the business characteristics and analysis needs of different enterprises.

[0055] The system establishes a time window mapping matrix M based on the time attributes of business data and the characteristics of business processes. This matrix is ​​a five-dimensional matrix with dimensions D×W×M×Q×Y (corresponding to five levels: day, week, month, quarter, and year). The matrix element M(d, w, m, q, y) represents the mapping relationship between the daily window d, the weekly window w, the monthly window m, the quarterly window q, and the annual window y.

[0056] For each piece of business data, the system determines which time windows it belongs to based on its timestamp and records the business data ID in the corresponding mapping matrix element. The system uses multi-granularity indexing technology to accelerate mapping queries and achieve... The time complexity is as follows. The mapping relationship matrix not only records the inclusion relationship between each time window (such as which week-level window a certain day-level window is included in), but also records the overlap relationship (such as which month-level windows a certain week-level window overlaps with).

[0057] The system slices the local semantic matrix generated in step S3 according to different time windows. During the slicing process, the system traverses each time window and extracts the corresponding rows and columns from the local semantic matrix based on the set of business data IDs contained in the window, forming a submatrix.

[0058] To improve computational efficiency, the system employs sparse matrix slicing technology, extracting only the rows and columns containing non-zero elements while maintaining the sparse storage format of the original matrix. For large-scale data, the system uses a distributed computing framework to process the slicing operations of multiple time windows in parallel, significantly improving processing speed. Each submatrix inherits the semantic similarity and temporal weight information of the original local semantic matrix, while also attaching metadata for the time window.

[0059] The system constructs a corresponding directed weighted graph based on the submatrix of each time window. In the diagram, node V represents business data, edge E represents the relationship between business data, and the weight W of the edge is the similarity value in the corresponding submatrix.

[0060] During the construction process, the system first determines the direction of the edges based on the temporal sequence of the business data to ensure that the directed graph accurately reflects the temporal characteristics of the business process. If the timestamp of business data A is earlier than that of business data B, a directed edge is constructed from A to B. If the timestamps of the two business data are the same, a directed edge is constructed after determining the priority based on the business logic (such as process step ID, operation type, etc.).

[0061] To ensure a reasonable graph structure, the system controls the out-degree and in-degree of each node, typically retaining the out- and in-edges with the highest weights from the TOP-K (K value is set according to business complexity, generally 3 to 7) to avoid overly complex graph structures. The system uses an adjacency list structure to store the directed weighted graph, optimizing memory usage and query efficiency.

[0062] For frequent sequential pattern mining, the system normalizes the edge weights in the directed weighted graph to ensure the comparability of weight values ​​between different subgraphs. The normalization method uses Min-Max standardization, mapping the weight values ​​to the [0, 1] interval. Where W is the original weight, and These are the minimum and maximum values ​​of all edge weights within the current time window, respectively. For weight values ​​with extremely uneven distribution, the system uses a quantile normalization method to remap them according to the percentile of the weight distribution, reducing the influence of extreme values ​​and making the normalized weight distribution more balanced.

[0063] The system defines frequent sequence patterns as path segments in a directed weighted graph. For each candidate pattern (path), the system calculates the weighted average of the weights of all edges within it as the weighted support of the corresponding pattern. in, Let be the normalized weight of the i-th edge in the path. These are the corresponding weighting coefficients.

[0064] The system supports multiple weighting coefficient setting strategies: uniform weighting (all) Equal), position-weighted (path front edge) Larger), time-weighted (edges with smaller time intervals) (Larger). Different strategies are suitable for different business scenarios, and the system supports dynamically selecting the appropriate strategy based on enterprise needs.

[0065] The system sets support thresholds for five time windows: daily, weekly, monthly, quarterly, and annual. , , , and ,satisfy The decreasing relationship. This embodiment is configured as follows: , , , , The specific values ​​can be adjusted according to business characteristics and data distribution.

[0066] This differentiated threshold design reflects the difference in pattern saliency across different time granularities: business activities within short time windows (such as daily levels) are highly concentrated and have clear patterns, thus requiring a higher support threshold; while business activities within long time windows (such as annual levels) are dispersed and contain multiple patterns, thus requiring a lower support threshold. The system supports automatically recommending threshold configurations based on historical data distribution, and also allows business personnel to manually adjust them according to actual needs.

[0067] The system creates a node expansion priority queue in each time window to guide the path expansion order during pattern mining. The priority queue is implemented using a heap structure, supporting maximum value retrieval in O(1) time complexity. Insertion operations with time complexity.

[0068] The elements in the queue are nodes to be expanded. The priority of a node is determined by a combination of factors: the weight of the connecting edges, the node's centrality index (such as PageRank value or betweenness centrality), and the node's business importance (predefined based on business domain knowledge). The system uses a weighted summation method to calculate the final priority score.

[0069] The system calculates the median of all edge weights within the current time window as a reference threshold. The use of the median makes the reference threshold insensitive to outliers, and can stably reflect the typical level of edge weights within the current window. For time windows with particularly large amounts of data, the system uses a random sampling method to estimate the median, improving computational efficiency.

[0070] The reference threshold is used to determine the importance of candidate expansion edges during pattern expansion and is a determining factor in the node insertion priority queue position. The system supports methods for dynamically adjusting the calculation of the reference threshold based on window characteristics, such as using the 75th percentile instead of the median in highly sparse graphs.

[0071] The system employs an iterative approach to expand frequent order patterns. First, the system identifies all initial frequent patterns (i.e., one-sided paths) with a weighted support greater than the corresponding support threshold. Then, the system iteratively expands these initial patterns as follows: Starting from the end node of the current pattern (path), all possible expansion edges are examined; for each candidate expansion edge, the system evaluates the weighted support of the new pattern formed after adding the edge; if the weighted support of the new pattern is still greater than the support threshold for the corresponding time window, the expansion edge is valid. For valid expansion edges, the system compares their weights with a reference threshold: if the edge weight is greater than the reference threshold, the corresponding target node is inserted at the head of the priority queue; if the edge weight is less than or equal to the reference threshold, the corresponding target node is inserted at the tail of the priority queue; the system retrieves the highest-priority node from the priority queue for the next round of expansion until the queue is empty or the preset maximum pattern length is reached. Through this priority-based expansion strategy, the system can prioritize exploring high-weight paths, improving the efficiency and quality of pattern mining. For each time window, the system ultimately obtains a set of frequent order patterns, containing patterns of different lengths and support levels.

[0072] The system establishes a time mapping table for each frequently detected sequential pattern, extracts the timestamp of each business data node in the pattern, and forms a pattern-timestamp mapping relationship. The mapping table adopts a key-value pair structure, where the key is the pattern ID and the value is an ordered list containing the timestamps of all nodes for that pattern.

[0073] The system assigns a unique identifier to each pattern and records metadata such as the time window to which the pattern belongs, the pattern's structural description (node ​​sequence and edge relationships), and weighted support, facilitating subsequent analysis and retrieval. For large-scale pattern sets, the system employs distributed storage and indexing technologies to improve data management efficiency.

[0074] The system calculates the time characteristic indicators for each frequent sequential pattern: time span. The difference between the timestamp of the last node and the timestamp of the first node in the pattern reflects the overall duration of the pattern; timestamp variance. The statistical variance of timestamps across all nodes in the model reflects the uniformity of the time distribution of business interactions within the model. The system supports calculations for multiple time units (such as hours, days, and weeks) and automatically selects the appropriate unit based on analysis requirements. For business data spanning different time zones or daylight saving time, the system performs standardization processing to ensure consistency in time calculations.

[0075] The system calculates the time difference sequence between adjacent nodes in a frequent sequential pattern. For a pattern containing k nodes, its node sequence... The corresponding timestamp is The system calculates the time difference sequence. The average of these differences is taken as the average interval of the pattern. The average interval reflects the rhythmic characteristics of business interactions and is an important indicator for identifying periodic business patterns. The system supports removing abnormal interval values ​​(such as extreme values ​​exceeding 3 standard deviations) before calculating the average, improving the stability and representativeness of the indicator.

[0076] Based on the above calculation results, the system further calculates two important time characteristic indicators: time coverage. : Where T is the size of the corresponding time window. Time coverage reflects the coverage range of the model within the time window, with a value range of [0, 1]. The closer the value is to 1, the wider the time range covered by the model. Time uniformity : ,in The variance of the node timestamps. This represents the average interval time. Time uniformity reflects the evenness of the time distribution of business interactions within a pattern; a smaller value indicates a more uniform time distribution, while a larger value indicates a more uneven time distribution. The system supports adjusting the calculation methods of these indicators according to different business scenarios. For example, when dealing with highly seasonal business, a seasonal adjustment factor can be introduced to optimize the uniformity calculation.

[0077] The system is based on time coverage. and time uniformity Calculate the time density index D of frequent sequential patterns: ,in, and The weighting coefficients are and satisfy the following conditions: ; This represents the maximum temporal uniformity value for all patterns within the current time window. The temporal density index D ranges from [0, 1], with values ​​closer to 1 indicating better temporal characteristics (wider coverage and more uniform distribution). The system supports adjustments based on enterprise needs. and The value can be increased for scenarios that prioritize pattern persistence. The value (e.g.) , For scenarios that emphasize pattern regularity, the [data / capacity] can be increased. The value (e.g.) , The system also supports adaptive weight allocation, automatically adjusting weight coefficients based on the time distribution characteristics of business data.

[0078] Based on the time window mapping matrix established earlier, the system constructs a pattern index structure across time granularities. This structure is implemented using a multi-level hash table. The first-level key is the time granularity (day, week, month, quarter, year), the second-level key is the time window ID, and the value is the set of frequent sequential patterns within that time window.

[0079] The system adds granularity markers and parent-child relationship pointers to each pattern, recording the pattern's position and associated patterns at different time granularity levels. The index structure supports fast cross-granularity pattern queries and association analysis, such as "finding all weekly patterns related to monthly pattern M" or "finding all quarterly high-density patterns within the annual window Y".

[0080] The system employs a bottom-up merging strategy, combining high-time-density patterns within small time windows (such as daily or weekly) to form candidate composite patterns within larger time windows (such as monthly, quarterly, or annual). The specific steps are as follows: The system sorts the patterns within the small time windows in descending order based on their time density (D) value; a time density threshold is set. (Typically 0.7 to 0.8), select high-density patterns with D values ​​greater than the threshold; based on the time window mapping relationship matrix, identify the set of high-density patterns in small windows belonging to the same large time window; analyze the temporal distribution and structural characteristics of these patterns, and identify possible combination patterns; for each potential combination pattern, the system calculates its merging support and consistency index; if the merging support and consistency index meet the preset conditions, then these small window patterns are merged into candidate composite patterns of the large window.

[0081] The system also employs a top-down decomposition strategy, breaking down patterns within a large time window into smaller time windows to pinpoint their precise components. The specific steps are as follows: For each pattern within a large time window, the system identifies all smaller time windows covered by its time span; within these smaller time windows, it searches for sub-patterns with structures similar to the large window pattern; it calculates the matching degree between the large window pattern and each smaller window sub-pattern, including node coverage, edge coverage, and temporal consistency; based on the matching degree, it constructs a mapping graph between the large window pattern and the smaller window patterns; and it applies a graph segmentation algorithm to decompose the large window pattern into the optimal combination of smaller window patterns. This decomposition strategy reveals the fine structure of large-scale business patterns at smaller time scales, helping enterprises understand the composition and evolution of complex business patterns.

[0082] The system calculates structural and semantic similarity between related patterns at different time granularities: Structural similarity: Calculated based on the topological structure of the pattern, primarily examining the overlap of node sets and the consistency of edge connection patterns. The system employs multiple graph similarity algorithms, such as graph edit distance, maximum common subgraph ratio, and graph kernel method, selecting the algorithm most suitable for the current business scenario. Semantic similarity: Calculated based on the local semantic matrix, examining the semantic relevance of corresponding business data within the pattern. The system extracts semantic feature vectors corresponding to nodes in the pattern and calculates the aggregated similarity between vector sets, such as mean cosine similarity and the distance the vector sets have traveled on Earth. The system combines structural and semantic similarity into a comprehensive similarity through a weighted average, serving as the basis for judging the relevance of patterns at different time granularities. The weight configuration can be flexibly adjusted according to business needs, such as a configuration emphasizing structure (0.7:0.3) or a configuration emphasizing semantics (0.3:0.7).

[0083] The system establishes a cross-granularity pattern mapping table based on structural similarity, semantic similarity, and temporal density D. The mapping table adopts a graph structure design, where nodes represent business patterns at different temporal granularities, edges represent the mapping relationship between patterns, and the weight of the edges is the comprehensive similarity.

[0084] Each record in the mapping table contains information such as source pattern ID, target pattern ID, time granularity mapping relationship (e.g., day → week, week → month), comprehensive similarity value, and time overlap ratio. The system supports many-to-many mapping relationships, meaning that a large granularity pattern may map to multiple small granularity patterns, and a small granularity pattern may also participate in the formation of multiple large granularity patterns.

[0085] Finally, the system merges cross-granularity patterns with consistent mapping relationships to construct a hierarchical business interaction pattern library: the system identifies pattern sets with consistent mapping relationships and high similarity from the mapping table; applies a pattern fusion algorithm to each set to generate a unified pattern representation across granularities; preserves the unique attributes of patterns at each time granularity, such as support, time density, and typical instances; constructs hierarchical relationships between patterns to form a complete business pattern genealogy from daily to annual levels; the final generated business interaction pattern library contains business interaction patterns at different time granularities and their relationships, providing enterprises with a multi-scale business process view, supporting comprehensive business analysis from macro to micro levels, and providing a solid foundation for business process improvement, anomaly detection, and trend prediction.

[0086] like Figure 4As shown in Figure S5, a business relationship network is constructed based on business interaction patterns at different time granularities. This includes: extracting business entities and interaction relationships from the obtained business interaction patterns at different time granularities as the basic elements of the business relationship network; designing a multi-level business relationship network structure, with different levels corresponding to different time granularities, and connecting the levels through an established cross-granularity pattern mapping table; performing community discovery on the business relationship network at each level to identify closely interacting business groups; calculating the centrality index of nodes in the business relationship network, including degree centrality, betweenness centrality, and eigenvector centrality, to identify key business nodes in the network; assigning weights to the edges in the network based on the calculated time density index D to reflect the temporal stability of business interactions; and calculating the importance of edges in the network based on the obtained weighted support to identify core business process paths.

[0087] S6. Based on the business relationship network, perform business trend analysis and generate visualized results, including: calculating the topological evolution characteristics of the business relationship network within different time windows based on the constructed multi-level business relationship network, including the changing trends of network density, clustering coefficient, and average path length; combining the time density index D, performing time-series analysis on patterns in the business relationship network to identify periodic, growing, and decaying business interaction patterns; comparing and analyzing core business processes within different time windows based on weighted support to identify optimization directions for business processes; constructing an evolutionary map of business patterns using cross-time granularity pattern integration results to demonstrate the changes and development trends of business patterns; generating interactive visualization interfaces containing network structure views, time evolution views, and pattern comparison views for different levels of business relationship networks; designing highlighting and interactive exploration functions in the visualization interface based on identified key business nodes and core business process paths; providing business anomaly detection and early warning functions, triggering an early warning mechanism and marking anomalies in the visualization interface when real-time business data deviates significantly from historical business interaction patterns.

Claims

1. A data intelligence visualization method for enterprise management, characterized in that, include: S1, retrieves historical business data from the enterprise platform; S2, extract the semantic features of historical business data to obtain semantic feature vectors; S3, Construct a local semantic matrix based on the semantic feature vector; S4. Set up multi-level time windows. Within each time window, perform sequential pattern mining on historical business data based on the local semantic matrix to obtain business interaction patterns at different time granularities. S5 constructs a business relationship network based on business interaction patterns at different time granularities; S6 performs business trend analysis based on the business relationship network and generates visualized results of the business trend analysis. S3, based on the semantic feature vectors, constructs a local semantic matrix, including: Based on the local semantic region segmentation results, a block sparse matrix is ​​constructed, where each block represents a local semantic region; For each local semantic region, the semantic similarity between business data is calculated using the dimensionality-reduced semantic feature vector, and used as the element value of the corresponding matrix block; Set a similarity threshold to sparsify the block sparse matrix; Based on the time attributes of the business data, the similarity values ​​in the sparsely segmented block matrix after sparsification are subjected to time decay weighting to obtain the final local semantic matrix. S4, in particular, acquires business interaction patterns at different time granularities, including: Set up multi-level time windows for days, weeks, months, quarters, and years to form a hierarchical structure of time granularity; A time window mapping matrix is ​​established based on the time attributes of business data and business processes. The mapping matrix reflects the inclusion and overlap relationships between time windows. The local semantic matrix generated in step S3 is sliced ​​according to different time windows to obtain the sub-matrix within the corresponding time window. Based on the submatrix of each time window, calculate the sequential dependencies between business data and construct a directed weighted graph; where nodes are business data and the weights of edges are the similarity values ​​in the submatrix. Frequent sequential pattern mining is performed on directed weighted graphs with different time windows; Calculate the temporal density of frequent sequential patterns; Based on the mapping matrix of time windows and the time density of frequent sequence patterns, frequent sequence patterns under different time windows are integrated across time granularities to obtain business interaction patterns at different time granularities. This includes frequent sequence pattern mining of directed weighted graphs at different time windows, including: Normalize the edge weights of the directed weighted graph; Based on the normalized edge weights, the weighted average of all edge weights in the frequent sequence pattern is calculated as the weighted support of the corresponding pattern. Set support thresholds for daily, weekly, monthly, quarterly, and annual time windows respectively. , , , and ;in, ; Create a node expansion priority queue in each time window; Calculate the median of all edge weights within the current time window as a reference threshold; The initial frequent patterns with weighted support greater than the corresponding support threshold are iteratively expanded. When the weight of the candidate expansion edge is greater than the reference threshold, the target node connected by the corresponding edge is inserted into the head of the point expansion priority queue, and otherwise inserted into the tail, thus obtaining the set of frequent order patterns in each time window. The calculation of the time density of frequent sequential patterns includes: Extract the timestamps of business data nodes in each frequent sequential pattern and establish a mapping table between the pattern and time; Calculate the time span of the first and last nodes in each frequent sequence pattern. and the variance of timestamps of each node in the frequent sequential pattern. Among them, the time span The variance reflects the overall duration of the pattern and the time distribution of business interactions within the pattern. Calculate the difference sequence of timestamps between adjacent nodes in a frequent sequential pattern, and calculate the average of the difference sequence as the average interval time of the corresponding pattern. ; Based on time span Calculate the time coverage of the corresponding mode based on the size T of the corresponding time window. ; Based on the variance of each node's timestamp and the average interval time of the corresponding mode Calculate the time uniformity of the corresponding pattern ; Based on time coverage and time uniformity Calculate the time density D of frequent sequential patterns.

2. The data intelligent visualization method for enterprise management according to claim 1, characterized in that: S2, extract the semantic features of historical business data to obtain a semantic feature vector, including: A self-supervised learning framework is used to extract semantic features from historical business data, resulting in high-dimensional semantic feature vectors. Locality-sensitive hashing is applied to high-dimensional semantic feature vectors to obtain local semantic region partitioning results; Dimensionality reduction is performed on the high-dimensional semantic feature vectors of each local semantic region to generate dimensionality-reduced semantic feature vectors. The final semantic feature vector is obtained by merging the dimensionality-reduced semantic feature vectors of all local semantic regions.

3. The data intelligent visualization method for enterprise management according to claim 2, characterized in that: Self-supervised learning frameworks employ BERT, Sim CSE, or Contrastive Loss models.

4. The data intelligent visualization method for enterprise management according to claim 1, characterized in that: Based on the temporal attributes of the business data, the similarity values ​​in the sparsely segmented block matrix after sparsification are subjected to time-decay weighting to obtain the final local semantic matrix, including: Calculate the time difference between each pair of business data based on the timestamps of the business data; Set a time decay function, wherein the time decay function is an exponential decay function or a half-life function; Calculate the time weight between each pair of business data based on the time difference and time decay function; Multiply the time weight by the corresponding semantic similarity to obtain the time-weighted semantic similarity value; Update the similarity values ​​in the block sparse matrix to generate a local semantic matrix that includes semantic similarity and temporal correlation.

5. The data intelligent visualization method for enterprise management according to claim 4, characterized in that: The time density D of frequent sequential patterns is calculated using the following formula: ;in, and The weighting coefficients are and satisfy the following conditions: , This represents the maximum time uniformity value for all patterns within the current time window.

6. A system based on the data intelligence visualization method for enterprise management as described in any one of claims 1 to 5.

Images