Intelligent industry graph generation and trend prediction method fusing multi-modal heterogeneous data

Abstract

The application discloses a wisdom industry graph generation and trend prediction method fusing multi-modal heterogeneous data, and relates to the technical field of industry graph generation and trend prediction. The method comprises the following steps: acquiring multi-modal heterogeneous data of a plurality of industry subjects in a target region; performing a space-time alignment operation on the multi-modal heterogeneous data of each industry subject to generate a unified space-time data cube taking the industry subject as a unit; extracting a three-dimensional space feature vector of each industry subject within a preset time window; constructing a directed weighted edge to generate an initial industry correlation topology graph with the three-dimensional space feature vectors of all industry subjects as vertices; performing a hierarchical progressive compression operation on the initial industry correlation topology graph, aggregating and replacing the vertex groups meeting the compression determination condition with supernodes to obtain a multi-layer compressed industry graph structure; based on the space-time evolution track of each layer of supernodes in the multi-layer compressed industry graph structure, extracting a cross-layer causal transmission path, and the transmission path generates a wisdom industry graph prediction signal.

Description

Technical Field

[0001] This invention relates to the field of industry map generation and trend prediction technology, specifically to a method for generating and predicting intelligent industry maps by integrating multimodal heterogeneous data. Background Technology

[0002] Industry graph generation relies on the collection and correlation analysis of a large amount of information on industry entities. Existing industry graph construction methods typically cluster and connect entities based on business registration texts or single-dimensional economic statistics, failing to deeply integrate multimodal heterogeneous data sources. This results in the spatial distribution, temporal evolution, and business behavior characteristics of industry entities being fragmented in the correlation modeling process. When it is necessary to simultaneously characterize industry relationships from three dimensions—industry attributes, geographical distribution, and temporal evolution—existing technologies generally suffer from information loss and correlation distortion. For predicting industry trends, existing methods mostly rely on macro-time series extrapolation or expert experience judgment, lacking the ability to dynamically extract causal transmission paths from within the graph structure, making it difficult to capture the causal driving relationships that evolve between industry levels over time. Faced with complex correlation networks composed of massive amounts of industry entities, the graphs generated by conventional graph theory methods are large in scale and have a single hierarchy, failing to provide a multi-granular macro-industry view while preserving fine-grained local correlations, severely restricting the efficiency of extrapolating from micro-events to macro-trends. Therefore, in the generation of industry maps, it is necessary to solve how to uniformly align business registration texts, tax data time series, geospatial vectors, and business event logs into the same spatiotemporal framework, and how to construct a map structure with multi-level compression capabilities from multi-entity relational topology, and automatically extract cross-level causal transmission paths from it, so as to realize the output of predictive signals of industry trends. Summary of the Invention

[0003] This invention provides a method for generating and predicting trends of intelligent industry maps by integrating multimodal heterogeneous data. It aims to perform spatiotemporal alignment and vectorization representation of multidimensional heterogeneous data of industry entities, construct a multi-layer industry map structure through hierarchical progressive compression, and generate prediction signals by utilizing cross-layer causal transmission paths to capture the multi-scale dynamic correlation and vertical causal transmission relationship between industry entities.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for generating and predicting trends of intelligent industry maps by integrating multimodal heterogeneous data, the method comprising the following contents.

[0005] This involves acquiring multimodal heterogeneous data from multiple industry entities within a target region. This multimodal heterogeneous data includes at least business registration documents, time-series tax data, geospatial vector data, and operational event logs. Unified access to this multi-source heterogeneous data allows for the coverage of information on industry entities across multiple dimensions, including registration, operations, geographical location, and time-series activities, providing comprehensive data support for building a unified analytical foundation.

[0006] Spatiotemporal alignment is performed on the multimodal heterogeneous data of each industry entity to generate a unified spatiotemporal data cube with the industry entity as the unit. In this process, business registration text is mapped to geographic grid cells according to the registration timestamp; tax data time series are anchored to the corresponding time slots within the geographic grid cells according to the declaration cycle; geospatial vectors are transformed to the reference projection plane of the geographic grid cells according to the spatial coordinate system; and business event logs are mounted to the time slots and geographic grid cells according to the event occurrence time and location, respectively. For each industry entity, the mapped business registration text, tax data time series, geospatial vectors, and business event logs are associated and concatenated according to a common time slot identifier and geographic grid cell identifier to form a unified spatiotemporal data cube with the industry entity as the unit. By unifying text, time series, spatial, and event data under the same spatiotemporal reference, the heterogeneity gap and spatiotemporal misalignment between multimodal data are effectively eliminated, enabling a unified description of the static attributes and dynamic behaviors of industry entities, and providing a consistent data representation for mining inter-industry relationships.

[0007] Based on the unified spatiotemporal data cube, a three-dimensional spatial feature vector for each industry entity within a preset time window is extracted. This three-dimensional spatial feature vector consists of components along the industry attribute axis, geographical distribution axis, and temporal evolution axis. Specifically, the industry classification code sequence, geographical grid coordinate sequence, and numerical sequence at time slots for each industry entity within the preset time window are read from the unified spatiotemporal data cube. The industry classification code sequence is aggregated using one-hot encoding to obtain the industry attribute component along the industry attribute axis. The centroid offset and coverage radius of the geographical grid coordinate sequence are calculated to obtain the geographical distribution component along the geographical distribution axis. The numerical sequence at the time slots is differentially analyzed to extract the magnitude and direction of numerical changes between adjacent time slots, forming the temporal evolution component along the temporal evolution axis. The industry attribute component, geographical distribution component, and temporal evolution component are then concatenated according to a preset dimensional order to form the three-dimensional spatial feature vector for each industry entity. The feature vector obtained in this way comprehensively characterizes the behavioral patterns of industry entities from three dimensions: industry attributes, spatial distribution, and temporal evolution, enabling the measurement of structural similarity and dynamic synergy among industry entities in a unified vector space.

[0008] As a preferred method, when calculating the geographic distribution component, the geographic grid coordinates where the industry entity first appears within a preset time window are used as the reference coordinates. The average and maximum offset distances of all subsequent geographic grid coordinates relative to the reference coordinates are calculated and used as the centroid offset and the coverage radius, respectively. This calculation method better reflects the actual changes in the location migration and radiation range of the industry entity, enabling the geographic distribution component to accurately reflect the centroid shift and regional expansion trend of the industry entity's spatial activities.

[0009] Using the three-dimensional spatial feature vectors of all industry entities as vertices, directed weighted edges are constructed based on the component differences between any two industry entities in the three-dimensional spatial feature vectors to generate an initial industry association topology. When constructing directed edges, for any two industry entities, the industry attribute component, geographical distribution component, and time evolution component in their respective three-dimensional spatial feature vectors are obtained. The first difference between their industry attribute components, the second difference between their geographical distribution components, and the third difference between their time evolution components are calculated. These differences are then multiplied by preset industry weight coefficients, geographical weight coefficients, and time weight coefficients, respectively, and summed to obtain a comprehensive association distance value. When the comprehensive association distance value is less than a preset association distance threshold, the reciprocal of the comprehensive association distance value is used as the weight of the directed weighted edge, pointing from the industry entity with the smaller comprehensive association distance value to the industry entity with the larger comprehensive association distance value, thus generating the initial industry association topology. This process transforms the comprehensive differences of multidimensional vectors into directional correlation strength, which not only quantifies the closeness between industrial entities, but also expresses the flow trend of industrial influence and resources through directionality, so that the edges in the graph have clear industrial transmission semantics.

[0010] A hierarchical progressive compression operation is performed on the initial industry association topology graph. Vertex groups that meet the compression criteria are aggregated and replaced with supernodes to obtain a multi-layer compressed industry graph structure. During the hierarchical progressive compression process, the initial industry association topology graph is used as the current layer graph, and all vertices and all directed weighted edges in the current layer graph are obtained. Each vertex in the current layer graph is traversed, and the ratio of the sum of the weights of the incoming edges to the sum of the weights of the outgoing edges of the vertex is calculated as the flow balance coefficient of the vertex. All vertices whose flow balance coefficients are within a preset range are marked as compressible vertices. The weight difference between any two marked compressible vertices on the directed weighted edges is used as the aggregation cost. The compressible vertices with the smallest aggregation cost are combined into a supernode. The supernode is used to replace the merged compressible vertex group, and the weights of the directed weighted edges connected to the supernode are updated to generate the next layer of compressed graph. This process is repeated until there are no compressible vertices in the current layer graph. When generating the next layer of compressed graph, the process further obtains all outgoing edge endpoints and all incoming edge origins for each compressible vertex in the merged compressible vertex group before merging. The supernode is used as the new incoming edge origin for all outgoing edge endpoints, and the supernode itself is used as the new outgoing edge endpoint for all incoming edge origins. The weights of all incoming edges entering the supernode are accumulated to form the merged incoming edge weight of the supernode, and the weights of all outgoing edges leaving the supernode are accumulated to form the merged outgoing edge weight of the supernode. The merged incoming edge weight and the merged outgoing edge weight are used as the new weights of the directed weighted edges between the supernode and external vertices, respectively, to generate the next layer of compressed graph. By employing a flow balance coefficient filtering and aggregation cost minimization merging strategy, tightly connected and mutually balanced industrial groups in the industrial network can be automatically summarized into higher-level industrial structure units. This effectively preserves the macro-level industrial hierarchy during compression, reduces graph complexity, and highlights key transmission links across groups.

[0011] Based on the spatiotemporal evolution trajectories of supernodes in each layer of the multi-layer compressed industry graph structure, cross-layer causal transmission paths are extracted, and intelligent industry graph prediction signals are generated using these paths. During path extraction, for each layer of the multi-layer compressed industry graph structure, the trajectory sequence of the three-dimensional spatial feature vector of each supernode in that layer as a function of time is obtained. This trajectory sequence consists of the three-dimensional spatial feature vectors of the supernode under multiple consecutive time windows. The phase difference between the trajectory sequences of supernodes in adjacent layers is calculated, and pairs of adjacent supernodes with a phase difference less than a preset phase difference threshold are marked as causal transmission node pairs. All consecutive adjacent pairs marked as causal transmission node pairs are connected sequentially from the top layer to the bottom layer to form cross-layer causal transmission paths. During the prediction signal generation process, the terminal values ​​of the three-dimensional spatial feature vectors of all supernodes on each cross-layer causal transmission path are obtained, and these terminal values ​​are weighted and summed according to the path direction to obtain the intelligent industry graph prediction signal. Specifically, for each cross-layer causal transmission path, the three-dimensional spatial feature vector of each supernode on the path is read sequentially along the path direction in the last time window, which is used as the terminal vector of the supernode. From the top layer to the bottom layer along the path direction, the terminal vector of each supernode is multiplied by the preset layer attenuation coefficient of the layer in which the supernode is located to obtain the weighted terminal vector. The weighted terminal vectors of all supernodes on the same cross-layer causal transmission path are accumulated to obtain the path prediction vector of the path. The path prediction vectors of all cross-layer causal transmission paths are normalized according to the path length and then summed to obtain the smart industry graph prediction signal. To accurately label causal transmission node pairs, one technical solution of the present invention calculates the phase difference through frequency domain analysis: A supernode is selected from adjacent layers as an upper-layer supernode and a lower-layer supernode, respectively. The trajectory sequence of the upper-layer supernode is obtained as the upper-layer trajectory sequence, and the trajectory sequence of the lower-layer supernode is obtained as the lower-layer trajectory sequence. A Fourier transform is performed on the upper-layer trajectory sequence to obtain the upper-layer spectrum, and a Fourier transform is performed on the lower-layer trajectory sequence to obtain the lower-layer spectrum. The phase corresponding to the frequency point with the maximum amplitude in the upper-layer spectrum is extracted as the upper-layer principal phase, and the phase corresponding to the frequency point with the maximum amplitude in the lower-layer spectrum is extracted as the lower-layer principal phase. The absolute value of the difference between the upper-layer principal phase and the lower-layer principal phase is calculated as the trajectory sequence phase difference between the upper-layer supernode and the lower-layer supernode. When the trajectory sequence phase difference is less than a preset phase difference threshold, the upper-layer supernode and the lower-layer supernode are labeled as a causal transmission node pair. By exploring the aforementioned multi-layer compression and cross-layer causal transmission paths, this invention can not only characterize the evolution of industrial structures at multiple scales, both macroscopic and microscopic, but also locate the causal transmission chain affecting specific downstream industrial units from changes in top-level industrial groups. The resulting prediction signal integrates the spatiotemporal evolution trends and upstream-downstream transmission effects of industrial entities at different levels, significantly improving the accuracy and robustness of industrial trend prediction. It can provide early and clear quantitative signal guidance for regional industrial planning, risk warning, and policy formulation.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0013] By performing spatiotemporal alignment operations, business registration texts, tax data time series, geospatial vectors, and business event logs are mapped to unified geographic grid cells and time slots, generating a unified spatiotemporal data cube with industry entities as units. This allows multimodal heterogeneous data with different formats, time granularities, and spatial references to be correlated and stitched together under the same benchmark, eliminating data misalignment in the spatiotemporal dimension and providing a self-consistent fused data foundation for subsequent correlation analysis. Based on the unified spatiotemporal data cube, components on the industry attribute axis, geographic distribution axis, and time evolution axis are extracted to form a three-dimensional spatial feature vector. The structured vectors uniformly express the multidimensional characteristics of industry entities, avoiding information fragmentation caused by step-by-step modeling. The differences in industry attribute, geographic distribution, and time evolution between industry entity vectors are calculated and weighted to obtain the comprehensive correlation distance. When the distance is below a threshold, directed weighted edges are established to generate an initial industry correlation topology graph, which can simultaneously quantify the correlation strength of industry entities in the three dimensions of attributes, space, and time with a unified distance metric. The initial industry-related topology graph is subjected to hierarchical progressive compression. The flow balance coefficient is calculated using the ratio of the sum of incoming edge weights to the sum of outgoing edge weights. Vertices with flow balance coefficients within a preset range are marked as compressible vertices. The compressible vertex groups are replaced with supernodes at the aggregation cost of the difference in directed edge weights. The supernodes integrate the incoming and outgoing edge weights of the merged vertices to generate the next layer of compressed graph. This compression is repeated layer by layer until no more vertices can be compressed, forming a multi-layered compressed industry graph structure. This compression mechanism preserves the key flow balance structures in the industry-related network while aggregating micro-level entities into supernodes reflecting meso- and macro-level industry groups. This allows the graph to exhibit multi-granularity association patterns at different compression levels, reducing the analytical complexity of large-scale graphs and showcasing the hierarchical evolution of industry associations from local to global. This paper extracts the 3D spatial feature vector trajectory sequences of supernodes at each layer in a multi-layer compressed industry graph structure under continuous time windows. The principal phase difference of the trajectory sequences of adjacent supernodes after Fourier transform is calculated. Pairs with phase differences less than a threshold are marked as causal transmission node pairs. Cross-layer causal transmission paths are formed by connecting all consecutive adjacent causal transmission node pairs from the top to the bottom layer. Using frequency domain phase difference as the criterion for determining cross-layer causal associations can capture the transmission relationships caused by time lag between industry levels, rather than simple temporal covariance. The terminal vectors of each supernode on each cross-layer causal transmission path are extracted, and weighted summations are performed after applying a hierarchical attenuation coefficient according to the path direction to generate a smart industry graph prediction signal. This transforms the vertical causal transmission inherent in the multi-layer graph structure into a quantitative signal that can be used for industry trend prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a method for generating and predicting trends in a smart industry map by integrating multimodal heterogeneous data;

[0016] Figure 2 This is a flowchart of the cross-layer causal path prediction signal generation process based on trajectory phase difference;

[0017] Figure 3 This is a flowchart for calculating the geographical distribution components of the main industries;

[0018] Figure 4 This is a graph showing the distribution of the time-series changes in tax amount and the proportion of sign bits;

[0019] Figure 5 This is a diagram showing the ranking of industry entities by the absolute value of their comprehensive correlation distance and the correlation distance threshold. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1This invention provides a method for generating and predicting trends in a smart industry map by integrating multimodal heterogeneous data. It acquires multimodal heterogeneous data of multiple industry entities within a target area, including business registration text, tax time series, geospatial vectors, and business event logs. A spatiotemporal alignment operation is performed on the multimodal heterogeneous data of each industry entity to generate a unified spatiotemporal data cube with the industry entity as the unit. Based on the unified spatiotemporal data cube, a three-dimensional spatial feature vector is extracted for each industry entity within a preset time window. This three-dimensional spatial feature vector consists of components along the industry attribute axis, geographical distribution axis, and time evolution axis. Using the three-dimensional spatial feature vectors of all industry entities as vertices, directed weighted edges are constructed based on the component differences between any two industry entities in the three-dimensional spatial feature vectors, generating an initial industry association topology. A hierarchical progressive compression operation is performed on the initial industry association topology, aggregating and replacing vertex groups that meet the compression criteria with supernodes to obtain a multi-layer compressed industry map structure. Based on the spatiotemporal evolution trajectory of each layer supernode in the multi-layer compressed industry graph structure, cross-layer causal transmission paths are extracted, and intelligent industry graph prediction signals are generated using these cross-layer causal transmission paths.

[0022] Example 1:

[0023] In practical implementation, after acquiring multimodal heterogeneous data from multiple industry entities within the target area, a spatiotemporal alignment operation is performed to generate a unified spatiotemporal data cube. For business registration documents, the registration timestamp is parsed and converted into a Coordinated Universal Time (UTC) millisecond-level time value. Based on the registration address field in the business registration document, geocoding is performed, converting the registration address into latitude and longitude coordinates. Then, a pre-defined geographic grid division table is searched for geographic grid cells covering these coordinates. The text identifier, unified social credit code, and business scope text content of the business registration document are mapped to these geographic grid cells. The geographic grid division table divides the entire region into row and column grids with equal latitude and longitude intervals. Each grid cell has a unique geographic grid cell identifier, which is encoded using a combination of row and column numbers.

[0024] For tax data time series, the start and end dates of the reporting period, tax amount, operating income, total assets, and other numerical indicators for each tax record are read. The midpoint of the reporting period is converted into a unified time base. This unified time base is then matched with preset time slot boundaries to determine the corresponding time slot identifier. A time slot is a continuous time slice with a calendar month as the time granularity, and each time slot has a time slot identifier composed of the year and month. Simultaneously, the taxpayer identification number corresponding to the tax record is associated with the industry entity, and the geographic grid unit identifier corresponding to the registered address is obtained from the industry entity's business registration text. The numerical indicators in the tax data time series are stored according to this geographic grid unit identifier and the matched time slot identifier.

[0025] For geospatial vectors, the polygon boundaries or point coordinates within the geospatial vector are transformed from the original coordinate system, such as the WGS84 geographic coordinate system, into a preset reference projection surface using a seven-parameter Helmholtz transformation. The reference projection surface adopts the universal transverse Mercator projection plane coordinate system, with the central meridian selected as the central meridian of the target area, and the coordinate unit is meters. The transformed plane coordinates are then spatially intersected with the grid range of the geographic grid division table to determine the geographic grid cell into which the plane coordinates fall. The geometric objects of the geospatial vector are then serialized and attached to the geographic grid cell identifier.

[0026] For the operational event log, extract the event timestamp and event location description text for each record in the operational event log. Round down the event timestamp according to the time slot granularity, for example, truncate it to the month to determine the time slot identifier corresponding to the event time. After converting the event location description text into latitude and longitude coordinates through the address matching engine, find the geographic grid cell that it falls into to obtain the geographic grid cell identifier. Then, attach the log content such as the operational event type and event attributes to the time slot identifier and geographic grid cell identifier.

[0027] After completing all data mapping, for each industry entity, using its unified social credit code as the primary key, all mapped business registration texts, tax time-series values, geospatial vector objects, and business event log records belonging to the same industry entity are associated and concatenated. This association and concatenation is achieved by constructing an in-memory multidimensional data structure with the industry entity identifier as the dimension and the time slot identifier and geographic grid unit identifier as the joint index. Each cell in this multidimensional data structure stores a set of data entries, which contains all four types of data for that industry entity under that time slot and geographic grid unit, forming a unified spatiotemporal data cube.

[0028] When extracting three-dimensional spatial feature vectors from a unified spatiotemporal data cube, data for each industry entity within a preset time window is read from the unified spatiotemporal data cube. The preset time window is a fixed sliding window that looks back 12 natural months from the current moment. The industry classification code sequence for each industry entity within this preset time window is read, arranged in time slot order. The industry classification code sequence consists of a four-digit code from the National Economic Industry Classification and Code (GB / T4754). A geographic grid coordinate sequence is read, where each coordinate is the plane coordinate of the center point of the geographic grid unit where the industry entity is located at the corresponding time slot. Finally, a numerical sequence is read from the time slot, which represents the specific value of the tax payment indicator in the tax data time series at each time slot.

[0029] When performing one-hot encoding aggregation on the industry classification code sequence, a preset industry classification code dictionary is loaded. This dictionary contains index mappings of all possible four-digit industry classification codes and their corresponding one-hot vector dimensions. The dimension of the one-hot vector is equal to the total number of entries in the dictionary. The industry classification code for each time slot within the preset time window is converted into a one-hot vector based on the dictionary. The index position corresponding to the industry classification code in one one-hot vector is set to 1, and the remaining positions are set to 0. The one-hot vectors for all time slots are summed bitwise to obtain an aggregated vector with the same dimension as the one-hot vectors and all dimensions being non-negative integers. This aggregated vector is then normalized and used as the industry attribute component. The normalization method is to divide the aggregated vector by the total number of time slots within the preset time window.

[0030] When calculating the centroid offset and coverage radius of the geographic grid coordinate sequence, the geographic grid coordinates of the first occurrence of the industry entity within the preset time window are used as the reference coordinates, and the reference coordinates are marked as follows: For each geographic grid coordinate in the sequence, from the second coordinate to the last coordinate, they are sequentially numbered as follows: A subsequent coordinate, Values ​​from arrive ,in Subtract the total number of geographic grid coordinate sequences , will the The subsequent coordinates are as follows Calculate the first... The offset distance of each subsequent coordinate relative to the reference coordinate :

[0031]

[0032] in, Indicates the first The offset distance of each subsequent coordinate. The x and y coordinates are the reference coordinates. For the first Calculate the x and y coordinates of each subsequent coordinate. The arithmetic mean of the offset distance values ​​is used as the centroid offset, and all are calculated. The maximum value among the offset distances is used as the coverage radius. The centroid offset and the coverage radius are combined in sequence to form a two-dimensional vector as the geographic distribution component.

[0033] When performing a difference operation on the numerical sequence in the time slot, the numerical sequence is represented as: ,in This represents the total number of time slots within the preset time window. For the first The tax amount for the first time slot. Calculate the tax amount for the first time slot. adjacent difference values , from to Extract each difference value. The sign bit, if If the value is greater than zero, the sign bit is taken. ,like If the value is less than zero, the sign bit is taken. ,like If the value is zero, the sign bit is removed. The sign bits are concatenated to form a sequence of changing directions; each difference value is extracted. The absolute value is used as the change range sequence. The mean, standard deviation, and sign bit of the change range sequence are then used as the data. In the proportion and sign position The proportion is used as a component of four-dimensional time evolution.

[0034] The industrial attribute component, geographical distribution component, and time evolution component are concatenated in a preset dimensional order, with the industrial attribute component first, followed by the centroid offset and coverage radius of the geographical distribution component, and finally the four statistical values ​​of the time evolution component, to form a three-dimensional spatial feature vector for each industrial entity.

[0035] See Figure 4This figure illustrates the scatter plot distribution of two key statistical indicators for the temporal evolution components of multiple industry entities within a preset time window in the unified spatiotemporal data cube described in Example 1: the mean magnitude of change (horizontal axis) and the proportion of "+1" in the sign bit (vertical axis). The horizontal axis, "mean magnitude of change," represents the average of the absolute difference sequence of tax amount indicators in the tax data time series, reflecting the average intensity of changes in the tax amount of industry entities; the vertical axis, "proportion of '+1' in the sign bit," represents the percentage of positive changes in the time slot difference values, reflecting the frequency of the growth trend in the tax amount of industry entities.

[0036] The scatter points of different colors in the graph correspond to different industry classification codes (C2610, C2620, C2630, C2640, C2710, C2720). Each scatter point represents the statistical result of the time evolution component of an industry entity within a preset time window. The overall distribution of the scatter points exhibits the following characteristics:

[0037] The average change range is mainly concentrated in the range of 0 to 50, indicating that the tax amount change of the vast majority of industrial entities remains at a low to medium level. A few scattered points exceed 50, indicating that some industrial entities have a large fluctuation in tax amount.

[0038] The proportion of "+1" in the sign bit is mainly distributed between 0.6 and 0.8, indicating that the tax amount of most industry entities shows a relatively frequent positive growth trend, with no obvious negative growth dominating.

[0039] The scatter plots of the industry classification codes overlap considerably on the graph, failing to demonstrate clear category distinctions. This indicates that the average change in tax amount and the frequency of growth of different industry entities are similar within this preset time window.

[0040] The dense area of ​​the scatter distribution is located in a rectangular region with a mean variation range of 10 to 40 and a sign bit "+1" ratio of 0.65 to 0.75, indicating the typical range of values ​​for the time evolution components of most industry entities.

[0041] Example 2:

[0042] In practical implementation, after obtaining the three-dimensional spatial feature vectors of all industry entities, any two industry entities are denoted as the first industry entity and the second industry entity. The industry attribute component, geographical distribution component, and temporal evolution component of the first industry entity are extracted from its three-dimensional spatial feature vector; similarly, the industry attribute component, geographical distribution component, and temporal evolution component of the second industry entity are extracted from its three-dimensional spatial feature vector.

[0043] Calculate the first difference between the industry attribute components of the primary industry entity and the industry attribute components of the secondary industry entity. The industry attribute components of the primary industry entity are normalized numerical vectors, and the industry attribute components of the secondary industry entity are another normalized numerical vector; both vectors have the same dimension. Subtract the corresponding dimension value from the value in the industry attribute component vector of the secondary industry entity in each dimension of the industry attribute component vector of the primary industry entity to generate an industry attribute difference vector. Calculate the algebraic sum of the values ​​in all dimensions of the industry attribute difference vector, and use this algebraic sum as the first difference. The sign of this algebraic sum indicates the overall strength of the industry attribute of the primary industry entity relative to that of the secondary industry entity.

[0044] Calculate the second difference between the geographical distribution components of the primary industry and the secondary industry. The geographical distribution component of the primary industry contains two elements: a first centroid offset value and a first coverage radius value. The geographical distribution component of the secondary industry contains two elements: a second centroid offset value and a second coverage radius value. Subtract the second centroid offset value from the first centroid offset value to obtain the centroid offset difference; subtract the second coverage radius value from the first coverage radius value to obtain the coverage radius difference. Add the centroid offset difference and the coverage radius difference together, and use the sum as the second difference.

[0045] Calculate the third difference between the time evolution components of the primary industry and the secondary industry. The time evolution component of the primary industry includes four elements: the first mean of the change range, the first standard deviation of the change range, the first positive proportion, and the first negative proportion. The time evolution component of the secondary industry includes four elements: the second mean of the change range, the second standard deviation of the change range, the second positive proportion, and the second negative proportion. Subtract the second mean of the change range from the first mean of the change range to obtain the difference in the mean of the change range; subtract the second standard deviation of the change range from the first standard deviation of the change range to obtain the difference in the standard deviation of the change range; subtract the second positive proportion from the first positive proportion to obtain the difference in the positive proportion; subtract the second negative proportion from the first negative proportion to obtain the difference in the negative proportion. Add the four values: the difference in the mean of the change range, the difference in the standard deviation of the change range, the difference in the positive proportion, and the difference in the negative proportion. Use this sum as the third difference.

[0046] Multiply the first difference, the second difference, and the third difference by the preset industry weight coefficient. Geographic weighting coefficient and time weighting coefficient The summation is then performed to obtain the overall correlation distance value. The calculation formula is expressed as:

[0047]

[0048] in, This is a comprehensive correlation distance value; This is the industry weight coefficient, with a value set to 0.5. The value is determined based on the fact that industry attributes play a dominant role in industry linkages. This is the geographical weighting coefficient, set to 0.3. The value is determined based on the fact that geographical proximity has a secondary impact on industrial linkages. The time weighting coefficient is set to 0.2, based on the fact that the synchronicity of time evolution has a relatively weak impact on industrial linkages. This is the first difference; The second difference; This is the third difference. Industry weight coefficient. Geographic weighting coefficient and time weighting coefficient The value was determined by combining the Delphi method with expert scoring, ensuring that the sum of the three coefficients was 1.

[0049] Set a preset association distance threshold Association distance threshold The determination method is as follows: For all pairs of industry entities, calculate the comprehensive correlation distance value for each pair of industry entities. absolute value , all absolute values Sort the data in ascending order, and take the absolute value of the 10th percentile after sorting as the association distance threshold. .

[0050] When the comprehensive correlation distance value of a pair of industrial entities absolute value Less than the association distance threshold At this point, construct a directed weighted edge. The weight value of the directed weighted edge is set to... The direction of the directed weighted edge is determined according to the comprehensive correlation distance value. The sign is determined: if If the calculation result is negative, it indicates that the comprehensive association distance value of the primary industry entity is less than that of the secondary industry entity, then the directed weighted edge points from the primary industry entity to the secondary industry entity; if If the calculation result is positive, it indicates that the comprehensive association distance value of the primary industry entity is greater than that of the secondary industry entity. Therefore, the directed weighted edge points from the secondary industry entity to the primary industry entity. This process of judgment and edge construction is performed on all industry entity pairs in the above manner to generate an initial industry association topology graph.

[0051] See Figure 5 The figure depicts the relationship based on the absolute value of the comprehensive correlation distance. The distribution curve of the absolute value of the comprehensive correlation distance of all industry entity pairs arranged in ascending order (blue solid line) and the preset correlation distance threshold. (Red dashed line). The horizontal axis represents the sequence number of the industry entity pair, which is based on the absolute value of the comprehensive correlation distance. The data is sorted from smallest to largest, with the sequence number covering all industry entities; the vertical axis represents the absolute value of the comprehensive correlation distance. The values ​​are displayed using a linear scale.

[0052] The blue curve in the graph shows a clear characteristic of phased changes: in approximately 90% of the interval before the serial number, The values ​​remained low and stable, concentrated between 0 and 0.01, with a gentle curve and small fluctuations; subsequently, in the last 10% of the interval, The value shows a rapid upward trend, with the curve rising steeply and the maximum value reaching over 300. This indicates that the comprehensive correlation distance between most industrial entities is relatively small, while the differences between a small number of industrial entities are significant, reflecting obvious long-tail distribution characteristics.

[0053] The red dashed line represents the association distance threshold. The value is set as the absolute value corresponding to the 10th percentile after sorting, located at approximately 0.004. This threshold divides industry entities into two categories: absolute value... Industry entity pairs below the threshold are located in the left region of the blue curve, accounting for 10% of all industry entity pairs, thus meeting the criteria for constructing a directed weighted edge; absolute value Industry entities with values ​​greater than the threshold are located in the right-hand region, which does not meet the edge construction conditions, and therefore will not form edge connections in the initial industry association topology graph.

[0054] Example 3:

[0055] In practical implementation, the initial industry association topology is used as the current layer graph, and all vertices and all directed weighted edges in the current layer graph are obtained. Each vertex in the current layer graph corresponds to an industry entity, and each directed weighted edge in the current layer graph corresponds to a directed connection relationship consisting of a pair of start and end vertices and a weight value. All vertices in the current layer graph are then labeled with vertex identifiers. This indicates that the set of all directed weighted edges is identified by its edge identifier. This means that each edge stores the identifier of the starting vertex, the identifier of the ending vertex, and the weight value of the edge.

[0056] Traverse the set of vertex identifiers in the current layer of the graph. Each vertex identifier in the array is used to record the currently traversed vertex as the current vertex. Get the current vertex. For all incoming edges to the endpoint, the set of incoming edges is denoted as the incoming edge set. Get the current vertex Let the set of all outgoing edges originating from a given point be denoted as the outgoing edge set. From the set of incoming edges Extract the weight value of each incoming edge, and sum all the incoming edge weight values ​​to obtain the sum of the incoming edge weights. From the outgoing edge set Extract the weight value of each outgoing edge, and sum all the outgoing edge weight values ​​to obtain the sum of the outgoing edge weights. Calculate the current vertex. Flow balance coefficient Flow balance coefficient The calculation formula is:

[0057]

[0058] in, Represents the current vertex The flow balance coefficient, Represents the current vertex The sum of the weights of the incoming edges, where the sum of the weights of the incoming edges means that the weights of the incoming edges are the weights of the edges that enter the current vertex. The sum of the weights of all directed weighted edges. Represents the current vertex The sum of the outgoing edge weights, where the sum of the outgoing edge weights represents the weights of the edges leaving the current vertex. The sum of the weights of all directed weighted edges. If the sum of the weights of the outgoing edges... If the value is zero, then the flow balance coefficient will be... Set as a maximum constant This is to avoid division by zero exceptions.

[0059] Obtain the preset lower bound threshold of the flow balance coefficient. Upper threshold of flow balance coefficient Lower threshold of flow balance coefficient Set to 0.8, upper limit threshold of flow balance coefficient The thresholds are set to 1.2. The rationale for these thresholds is that when the flow balance coefficient is within the range [0.8, 1.2], the inflow and outflow intensities of the industrial entity are relatively balanced, indicating that the industrial entity has the structural conditions to be compressed to a higher level. The flow balance coefficient is set to... With flow balance coefficient lower threshold Upper threshold of flow balance coefficient If a comparison is made, and Then the current vertex Marked as compressible vertices.

[0060] Repeat the above process of calculating and comparing flow balance coefficients until the entire set of vertex identifiers has been traversed. By identifying all the vertices in the dataset, a compressible vertex set can be obtained. .

[0061] In a compressible vertex set In this context, for any two vertices marked as compressible vertices, we denote them as the first compressible vertex. Second compressible vertex Obtain the first compressible vertex. Second compressible vertex The relationship between two points on all directed weighted edges is given by the set of directed weighted edges. There exists a first compressible vertex. Pointing to the second compressible vertex The edge may have a second compressible vertex. Point to the first compressible vertex If an edge is found to be compressible, its weight value is read, and the reciprocal of the absolute value of this weight value is taken as the first compressible vertex. Second compressible vertex The aggregation cost between them. If there is a set of directed weighted edges. There is no first compressible vertex in it. Second compressible vertex For direct edges between them, the aggregation cost is set to a maximum constant. This indicates that there is no direct connection between the two vertices, making aggregation extremely difficult. For compressible vertex sets... Calculate the aggregation cost for all pairwise vertex pairs, select the two compressible vertices with the smallest aggregation cost, and combine the two selected compressible vertices into a supernode.

[0062] When generating the compressed graph of the next layer, obtain the first compressible vertex from the two compressible vertices that are being merged. The set of all outgoing edge endpoints before merging and the set of all incoming edge origins Second compressible vertex The set of all outgoing edge endpoints before merging and the set of all incoming edge origins The merged supernodes As the first compressible vertex The set of outgoing edge endpoints The new starting point of the incoming edges for all vertices in the supernode will be... As the second compressible vertex The set of outgoing edge endpoints Find the new starting point of incoming edges for all vertices in the array, and also add the supernode. As the first compressible vertex The set of incoming edge starting points The new outgoing edge endpoints of all vertices in the supernode will be... As the second compressible vertex The set of incoming edge starting points The new outgoing edge endpoints for all vertices in the middle.

[0063] Will enter the first compressible vertex The weights of all incoming edges and the weights of edges entering the second compressible vertex The weights of all incoming edges are summed. If there are duplicate external starting vertices, they are all connected to the first compressible vertex. Second compressible vertex Then, the weights of the corresponding two edges are combined and added together, and the sum is used as the supernode. The weight of the merged edge. It will leave the first compressible vertex. The weights of all outgoing edges and the weights of edges leaving the second compressible vertex. The weights of all outgoing edges are summed. If there are duplicate external endpoint vertices that are also the first compressible vertex, then... Second compressible vertex For a connection, the weights of the corresponding two edges are combined and added together, and the sum is used as the supernode. The weights of the outgoing edges are merged. The weights of the incoming edges are used as the weights of the incoming supernodes. The new weights of the directed weighted edges are used to determine the weights of the edges leaving the supernode. The new weights of the directed weighted edges will be applied to the supernodes. Replace the original first compressible vertex Second compressible vertex This generates the compressed map for the next layer.

[0064] In the previous layer of the compressed graph, the process of traversing each vertex in the current layer of the graph to calculate the flow balance coefficient and mark compressible vertices is repeated. The compressible vertices are then combined and merged until there are no more compressible vertices in the current layer of the graph. At this point, the compression process is terminated, and a multi-layer compressed industry graph structure is obtained.

[0065] Example 4:

[0066] In specific implementation, please refer to Figure 2 After obtaining the multi-layer compressed industry map structure, for each layer in the multi-layer compressed industry map structure, this layer is denoted as the th layer. Layer, in which The numbers are sequentially increased from the top to the bottom, with the top level being the [number]. The first layer, the second layer is the third layer. layer, This represents the total number of layers in the multi-layer compressed industry map structure. (The remaining text appears to be incomplete and requires further context.) The three-dimensional spatial feature vector of each supernode in the layer, the first The three-dimensional spatial feature vector of each supernode in the layer is obtained by taking a weighted average of the three-dimensional spatial feature vectors of all original industry entities contained in that supernode. For the first... For each supernode in the layer, the 3D spatial feature vector of the current supernode is extracted under multiple consecutive time windows, with the number of consecutive time windows being a preset number. There are several time windows of equal length, each spanning a period of 1 / 3. The sliding step size of adjacent time windows is [number] calendar months. One natural month. (The sentence is incomplete and lacks context.) The 3D spatial feature vectors of the current supernode under consecutive time windows are arranged in chronological order to form a trajectory sequence, which contains... The three-dimensional spatial feature vectors arranged chronologically. For the... Each supernode in the layer performs a trajectory sequence generation operation to obtain the first... The set of trajectory sequences of all supernodes in the layer.

[0067] Select a supernode from each of the two adjacent layers, and denote the upper layer of the two adjacent layers as the first layer. The lower layer is denoted as the th. Layer. From the first Select a supernode from the set of layer trajectory sequences, and denote it as the upper-layer supernode. From the first Select a supernode from the set of layer trajectory sequences and denote it as the lower-level supernode. Get the parent supernode. The trajectory sequence is used as the upper-level trajectory sequence to obtain the lower-level supernode. The trajectory sequence is used as the lower-level trajectory sequence. The upper-level trajectory sequence is processed as follows: the sum of squares and the square root of the square root of all components of the three-dimensional spatial feature vector of each time window in the upper-level trajectory sequence are performed to obtain an upper-level scalar trajectory value, which is then continuously processed. The upper-level scalar trajectory values ​​of each time window are arranged in chronological order to form an upper-level scalar trajectory sequence. A Fast Fourier Transform (FFT) is then performed on the upper-level scalar trajectory sequence, with the number of points set to [value missing]. ,like no If the value is an integer power, then the end of the upper-level scalar trajectory sequence will be padded with zeros until the nearest integer power. The integer power of . After Fast Fourier Transform, we obtain . A sequence of upper-level complex spectral values ​​is formed, with each complex number corresponding to a frequency point. The magnitude of each complex number in the upper-level complex spectral sequence is calculated, and this magnitude is used as the amplitude of the corresponding frequency point. All frequency points are iterated through, and the frequency point with the largest amplitude is found and designated as the upper-level principal frequency point. The argument of the complex number corresponding to the upper-level principal frequency point is extracted, and this argument is used as the upper-level principal phase. .

[0068] For lower-level supernodes The same processing is applied to the lower-level trajectory sequence: the sum of squares and the square root of the square of all components of the three-dimensional spatial feature vector of each time window in the lower-level trajectory sequence are performed to obtain a lower-level scalar trajectory value, forming the lower-level scalar trajectory sequence. A Fast Fourier Transform is then performed on the lower-level scalar trajectory sequence, with the number of points also set to be padded to... integer powers This yields the lower-level complex spectrum sequence. The modulus of each complex number in the lower-level complex spectrum sequence is calculated, and the frequency point with the largest amplitude is identified as the lower-level principal frequency point. The argument of the complex number corresponding to the lower-level principal frequency point is extracted as the lower-level principal phase. .

[0069] Calculate the upper-level principal phase With lower-level main phase The absolute value of the difference between them is used as the upper-level supernode. With lower-level supernodes Phase difference between trajectory sequences :

[0070]

[0071] Obtain the preset phase difference threshold Preset phase difference threshold Set as The basis for this setting is that when the principal phase deviation between the two sequences is less than... At a given time, the periodic fluctuations of the two sequences exhibit high temporal synchronization, satisfying the temporal conditions for causal transmission. The phase difference between the trajectory sequences... With respect to the preset phase difference threshold Compare, if Less than Then the upper-level supernode With lower-level supernodes This is marked as a causal transit node pair. For the first... Layer and first All possible combinations of upper and lower supernodes in the layer are subjected to the above phase difference calculation and comparison, and all combinations that meet the conditions are marked as causal transit node pairs.

[0072] After marking causal transit node pairs between all adjacent layers, connect all marked causal transit node pairs between consecutive adjacent layers sequentially from top to bottom to form a cross-layer causal transit path. The connection method is as follows: starting from the top layer... Starting from the marked causal transit node pairs in the layer, find the upper-level supernode of the pair as the starting point of the path, and use the lower-level supernode of the pair as the transition point, in the ... Layer and first In the layer's labeled causal transit node pairs, find new pairs where the upper-level supernode and the transition point are the same, and continue connecting their lower-level supernodes downwards, iterating in this way until the bottom layer is reached. Each layer forms a complete sequence of supernodes from the top to the bottom, and this complete sequence of supernodes is denoted as a cross-layer causal transmission path. This connection process is repeated for all possible starting points to generate a set of cross-layer causal transmission paths, which includes... A cross-layer causal transmission path.

[0073] For each cross-level causal transit path in the set of cross-level causal transit paths, denoted as the , A cross-layer causal transmission path, reading the first causal link sequentially from the top layer to the bottom layer along the path direction. The three-dimensional spatial feature vector of each supernode on the cross-layer causal propagation path in the last time window is used as the first... The first cross-layer causal transmission path The terminal vectors of each supernode ,in This represents the index of the supernode within the path. from to , For the first The total number of supernodes contained in a cross-layer causal transmission path.

[0074] Following the path direction from top to bottom, sequentially... The terminal vectors of each supernode Multiply by the first Preset level attenuation coefficient of the layer where each supernode is located ,in Indicates the first The layer number of each supernode in the multi-layer compressed industry graph structure. Preset layer attenuation coefficient. The rule for determining the value is: layer number for hour, Set as Layer number Greater than hour, Set as That is, for each layer descending from the top to the bottom, the attenuation coefficient is multiplied by . The reason why the attenuation coefficient decreases layer by layer is that the influence of industrial linkages attenuates at each layer as it is transmitted along the hierarchy. The attenuation rate is determined by empirical parameters of industrial network diffusion. The terminal vector... Multiply by the corresponding preset level attenuation coefficient to obtain the first level. The weighted terminal vector of the nth supernode. The nth... All cross-layer causal transmission paths The weighted terminal vectors of the nth supernodes are accumulated element by element to obtain the nth... Path prediction vector of each path The accumulation process is represented as:

[0075]

[0076] in, Indicates the first The path prediction vector corresponding to each cross-layer causal transmission path. It is a vector with the same dimension as the feature vector in three-dimensional space; Indicates the first The total number of supernodes contained in a cross-layer causal transmission path; Indicates the first The first path The preset level attenuation coefficient of the layer where each supernode is located; Indicates the first The first path The three-dimensional spatial feature vectors of each supernode in the last time window.

[0077] For each path in the set of cross-layer causal transmission paths, the corresponding path prediction vector is calculated using the above process. Then, the path prediction vectors of all cross-layer causal transmission paths are normalized according to their path lengths and summed. The specific method for normalization and summation is as follows: for the ... Path prediction vector Divide each element value in by the first element. Path length of the path The normalized path prediction vector is obtained. All The normalized path prediction vectors of each path are summed element-wise to obtain the smart industry map prediction signal vector. Each component of the smart industry map prediction signal vector represents the comprehensive prediction strength of future trends on the industry attribute axis, geographical distribution axis, and time evolution axis.

[0078] Example 5:

[0079] In specific implementation, please refer to Figure 3After reading the geographic grid coordinate sequence of each industry entity within a preset time window from the unified spatiotemporal data cube, the centroid offset and coverage radius are calculated on the geographic grid coordinate sequence to obtain the geographic distribution components.

[0080] For a geographic grid coordinate sequence, all coordinates in the sequence are arranged chronologically according to their time slots, with the earliest coordinate at the beginning and the latest coordinate at the end. The geographic grid coordinate corresponding to the earliest coordinate in the sequence is designated as the reference coordinate, and the abscissa of the reference coordinate is... The ordinate of the reference coordinate is marked as The reference coordinates are the geographic grid coordinates of the industry entity when it first appears within the preset time window. They are the plane coordinates of the center point of the geographic grid unit bound to the industry entity in the first time slot within the preset time window. The plane coordinates of the center point are directly read from the grid geometry attributes of the geographic grid division table. The coordinate values ​​are in the universal transverse Mercator projection plane coordinate system, and the unit is meters.

[0081] After the baseline coordinates, the remaining coordinates in the geographic grid coordinate sequence are obtained in chronological order, and the remaining coordinates are arranged in the order of their appearance from... Begin by numbering sequentially, with the serial number denoted as . , The range of values ​​is from arrive ,in Subtract the total number of coordinates in the geographic grid coordinate sequence within the preset time window. The serial number is Subsequent coordinates are marked as subsequent coordinates Subsequent coordinates The horizontal coordinate mark Subsequent coordinates The longitudinal coordinates are marked as For each subsequent coordinate Calculate subsequent coordinates Offset distance relative to the reference coordinates offset distance The calculation method is as follows:

[0082]

[0083] in, Indicates the sequence number is The offset distance of the subsequent coordinates relative to the reference coordinates, the offset distance The unit is meters. The x-coordinate value represents the baseline coordinate. The vertical coordinate value represents the reference coordinate. Indicates the sequence number is The x-coordinate value of the subsequent coordinates, Indicates the sequence number is The subsequent coordinates are the ordinate values. Both the x-coordinate and ordinate values ​​are real numbers in the general transverse Mercator projection plane coordinate system.

[0084] right Perform the above offset distance calculation operation on all subsequent coordinates to obtain the result from The offset distance set consisting of offset distance values Calculate the offset distance for all values ​​in the set. The arithmetic mean of the offset distance values ​​is calculated by taking... Add up the offset distance values ​​and divide by . The calculated arithmetic mean is used as the centroid offset. The centroid offset represents the average degree of displacement of the overall geographical distribution of the industry entities relative to their initial positions within a preset time window, and is measured in meters.

[0085] Extract the largest offset distance value from the offset distance set by iterating through all values ​​in the offset distance set. The maximum offset distance is determined by comparing each pair of offset distance values ​​and retaining the larger value. This maximum offset distance is used as the coverage radius. The coverage radius represents the maximum geographical range of an industry entity within a preset time window, expressed in meters.

[0086] The calculated centroid offset and coverage radius are combined sequentially into a two-dimensional numerical vector. The first element of this vector is the centroid offset value, and the second element is the coverage radius value. This two-dimensional numerical vector is used as the geographic distribution component. The geographic distribution component, together with the industry attribute component and the time evolution component, constitutes the three-dimensional spatial feature vector of the industry entity.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for generating and predicting trends of intelligent industry maps by integrating multimodal heterogeneous data, characterized in that: include: Acquire multimodal heterogeneous data of multiple industry entities within the target area. The multimodal heterogeneous data includes at least business registration text, time series of tax data, geospatial vector data, and business event logs. Perform spatiotemporal alignment operations on the multimodal heterogeneous data of each industry entity to generate a unified spatiotemporal data cube with the industry entity as the unit; Based on the unified spatiotemporal data cube, a three-dimensional spatial feature vector of each industry entity is extracted within a preset time window. The three-dimensional spatial feature vector is composed of components on the industry attribute axis, the geographical distribution axis, and the time evolution axis. Using the three-dimensional spatial feature vectors of all industry entities as vertices, and based on the component differences between any two industry entities in the three-dimensional spatial feature vectors, directed weighted edges are constructed to generate an initial industry association topology graph. A hierarchical compression operation is performed on the initial industry association topology graph. Vertex groups that meet the compression criteria are aggregated and replaced with supernodes to obtain a multi-layer compressed industry graph structure. Based on the spatiotemporal evolution trajectory of each layer supernode in the multi-layer compressed industry graph structure, cross-layer causal transmission paths are extracted, and intelligent industry graph prediction signals are generated using the cross-layer causal transmission paths.

2. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 1, characterized in that, The step of performing spatiotemporal alignment operations on the multimodal heterogeneous data of each industry entity to generate a unified spatiotemporal data cube with industry entities as units specifically includes: The business registration text is mapped to a geographic grid cell according to the registration timestamp, the tax data time series is anchored to the corresponding time slot in the geographic grid cell according to the declaration cycle, the geographic spatial vector is converted to the reference projection plane of the geographic grid cell according to the spatial coordinate system, and the business event log is mounted to the time slot and the geographic grid cell according to the event occurrence time and event occurrence location respectively. For each industry entity, the mapped business registration text, the time series of tax data, the geospatial vector, and the business event log are associated and spliced ​​according to a common time slot identifier and geographic grid unit identifier to form a unified spatiotemporal data cube with the industry entity as the unit.

3. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 2, characterized in that, The step of extracting a three-dimensional spatial feature vector for each industry entity within a preset time window based on the unified spatiotemporal data cube, wherein the three-dimensional spatial feature vector is composed of components on the industry attribute axis, geographical distribution axis, and time evolution axis, specifically includes: Read the industry classification code sequence, geographic grid coordinate sequence, and numerical sequence of time slots for each industry entity within a preset time window from the unified spatiotemporal data cube; One-hot encoding aggregation is performed on the industry classification code sequence to obtain the industry attribute components on the industry attribute axis; The centroid offset and coverage radius of the geographic grid coordinate sequence are calculated to obtain the geographic distribution components on the geographic distribution axis; Perform a difference operation on the numerical sequence in the time slot to extract the magnitude and direction of numerical change between adjacent time slots, and form the time evolution component on the time evolution axis. The industry attribute component, the geographical distribution component, and the time evolution component are concatenated in a preset dimensional order to form a three-dimensional spatial feature vector for each industry entity.

4. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 3, characterized in that, The step of generating an initial industry association topology graph by constructing directed weighted edges based on the component differences between any two industry entities in the three-dimensional spatial feature vectors, using the feature vectors of all industry entities as vertices, specifically includes: For any two industry entities, obtain the industry attribute component, geographical distribution component, and time evolution component from the three-dimensional spatial feature vectors of each entity. Calculate the first difference between the industry attribute components, the second difference between the geographical distribution components, and the third difference between the time evolution components; The first difference, the second difference, and the third difference are multiplied by preset industry weight coefficients, geographical weight coefficients, and time weight coefficients, respectively, and then summed to obtain the comprehensive correlation distance value. When the comprehensive association distance value is less than the preset association distance threshold, the reciprocal of the comprehensive association distance value is used as the weight of the directed weighted edge, and the industry entity with the smaller comprehensive association distance value points to the industry entity with the larger comprehensive association distance value, thus generating an initial industry association topology graph.

5. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 4, characterized in that, The step of performing a hierarchical compression operation on the initial industry association topology graph, aggregating and replacing vertex groups that meet the compression criteria with supernodes to obtain a multi-layer compressed industry graph structure, specifically includes: Using the initial industry association topology as the current layer graph, obtain all vertices and all directed weighted edges in the current layer graph; Traverse each vertex in the current layer graph, calculate the ratio of the sum of the weights of the incoming edges to the sum of the weights of the outgoing edges of the vertex as the flow balance coefficient of the vertex, and mark all vertices whose flow balance coefficients are within a preset range as compressible vertices. Using the weight difference on the directed weighted edge between any two vertices marked as compressible as the aggregation cost, the compressible vertices with the smallest aggregation cost are combined into a supernode. Replace the merged compressible vertex group with the supernode, update the weights of the directed weighted edges connected to the supernode, generate the next layer of compressed graph, and repeat until there are no compressible vertices in the current layer of graph.

6. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 5, characterized in that, The specific steps for generating the upper-level compressed map include: Get all outgoing edge endpoints and all incoming edge starting points of each compressible vertex in the merged compressible vertex group before merging; The super node is used as the new starting point of all outgoing edges, and the super node is used as the new ending point of all incoming edges. The weights of all incoming edges entering the supernode are summed to form the merged incoming edge weight of the supernode, and the weights of all outgoing edges leaving the supernode are summed to form the merged outgoing edge weight of the supernode. The weights of the merged inbound edges and the merged outbound edges are used as the new weights of the directed weighted edges between the supernode and the external vertices to generate the next layer of compressed graph.

7. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 6, characterized in that, The steps of extracting cross-layer causal transmission paths based on the spatiotemporal evolution trajectories of supernodes in the multi-layer compressed industry graph structure, and generating intelligent industry graph prediction signals using the cross-layer causal transmission paths specifically include: For each layer in the multi-layer compressed industry map structure, obtain the trajectory sequence of the three-dimensional spatial feature vector of each supernode in that layer as a function of time. The trajectory sequence is composed of the three-dimensional spatial feature vector of the supernode under multiple consecutive time windows. Calculate the phase difference of the trajectory sequences of supernodes between adjacent layers, and mark the pairs of supernodes in adjacent layers with a phase difference less than a preset phase difference threshold as causal transmission node pairs; All consecutive adjacent layers marked as causal transit nodes are connected sequentially from the top layer to the bottom layer to form a cross-layer causal transit path; Obtain the terminal values ​​of the three-dimensional spatial feature vectors of all supernodes on each cross-layer causal transmission path, and sum the terminal values ​​according to the path direction to obtain the smart industry map prediction signal.

8. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 7, characterized in that, The step of calculating the phase difference of the trajectory sequences of supernodes between adjacent layers and marking adjacent layer supernode pairs with a phase difference less than a preset phase difference threshold as causal transmission node pairs specifically includes: Select one supernode from each of the adjacent layers as the upper-level supernode and the lower-level supernode respectively, obtain the trajectory sequence of the upper-level supernode as the upper-level trajectory sequence, and obtain the trajectory sequence of the lower-level supernode as the lower-level trajectory sequence. The upper-layer spectrum is obtained by performing a Fourier transform on the upper-layer trajectory sequence, and the lower-layer spectrum is obtained by performing a Fourier transform on the lower-layer trajectory sequence. The phase corresponding to the frequency point with the maximum amplitude in the upper-layer spectrum is extracted as the upper-layer main phase, and the phase corresponding to the frequency point with the maximum amplitude in the lower-layer spectrum is extracted as the lower-layer main phase. The absolute value of the difference between the upper-level main phase and the lower-level main phase is calculated as the trajectory sequence phase difference between the upper-level supernode and the lower-level supernode. When the trajectory sequence phase difference is less than a preset phase difference threshold, the upper-level supernode and the lower-level supernode are marked as a causal transmission node pair.

9. The method for generating and predicting trends of intelligent industry maps by integrating multimodal heterogeneous data according to claim 7, characterized in that, The steps for obtaining the smart industry map prediction signal specifically include: For each cross-layer causal transmission path, read the three-dimensional spatial feature vector of each supernode on the path in the last time window in sequence along the path direction, and use it as the end vector of the supernode. Following the path direction from top to bottom, the terminal vector of each supernode is multiplied by the preset level attenuation coefficient of the layer in which the supernode is located to obtain the weighted terminal vector. The path prediction vector of the path is obtained by summing the weighted terminal vectors of all supernodes on the same cross-layer causal transmission path; the path prediction vectors of all cross-layer causal transmission paths are normalized according to the path length and then summed to obtain the smart industry graph prediction signal.

10. The method for generating and predicting smart industry maps by integrating multimodal heterogeneous data according to claim 3, characterized in that, When calculating the centroid offset and coverage radius of the geographic grid coordinate sequence, the geographic grid coordinates in which the industry entity first appears within a preset time window are used as the reference coordinates. The average and maximum offset distances of all subsequent geographic grid coordinates relative to the reference coordinates are calculated as the centroid offset and coverage radius, respectively.

Images