Big data-based real estate enterprise intelligent statistical method, device, equipment and medium

Abstract

The present application belongs to the technical field of city big data governance, and relates to a method, device, equipment and medium for intelligent statistics of people and real estate enterprises based on big data. The method comprises: collecting multi-source heterogeneous government data of public security, housing construction and market supervision, completing cross-department entity alignment through address vectorization and similarity matching, constructing an initial people and real estate enterprise correlation graph, and obtaining an updated target correlation graph by fusing dynamic people flow space-time data. The present application fuses space grid features and electricity and tax business features, quantitatively outputs house vacancy and enterprise false registration risk probability through a random forest model, automatically constructs time sequence training samples based on risk labeling, performs incremental iterative optimization on a long short-term memory network model, and finally realizes time sequence prediction of regional people and real estate enterprise space distribution by using a converged model. The present application remedies the defect that static account books cannot identify false subjects and vacancy status, and effectively improves the accuracy and intelligent level of city real estate and market subject governance.

Description

Technical Field

[0001] This invention belongs to the field of urban big data governance technology, and in particular relates to a smart statistical method, device, equipment and medium for real estate enterprises based on big data. Background Technology

[0002] With the rapid development of smart city governance, government systems such as public security household registration, housing and construction, real estate, market supervision and industry and commerce have accumulated a large amount of static registration data on population, housing, and enterprises. Combined with dynamic time-series data such as urban population flow, energy consumption, and taxation, a rich multi-source spatiotemporal government data resource has been formed. People, housing, and enterprises are the core entities of urban governance and market supervision. Accurately completing their status statistics, risk identification, and spatial distribution prediction is a key support for revitalizing idle urban resources and rectifying entities with false registrations.

[0003] Current traditional methods for statistical analysis of individuals and real estate companies mostly rely on static ledger verification by department, which has significant technical shortcomings. Existing technologies have failed to achieve deep integration of multi-source heterogeneous government data. Data standards, address specifications, and spatial codes of different departments are independent of each other, making it difficult for conventional text matching methods to achieve accurate alignment of cross-source entities. This makes it impossible to build complete and reliable relationships between individuals and real estate companies, resulting in distorted basic statistical data and mismatched data.

[0004] Meanwhile, traditional statistics rely heavily on static registration data, neglecting dynamic spatiotemporal characteristics such as population activity and operational energy consumption. This fails to accurately reflect the actual occupancy status of properties and the current operational status of businesses, making it difficult to effectively identify abnormal risks such as vacant properties and fraudulent business registrations, resulting in insufficient accuracy. Furthermore, existing intelligent assessment models are mostly trained offline with fixed parameters, unable to adaptively iterate and optimize based on dynamic changes in urban business formats, leading to a tendency for model accuracy to decay. Conventional time-series forecasts also lack spatial topological constraints and iterative convergence mechanisms, resulting in poor predictions of the spatial distribution of people, properties, and businesses in a given region. Summary of the Invention

[0005] Based on this, it is necessary to provide a big data-based intelligent statistical method, device, equipment, and medium for real estate enterprises that can achieve accurate alignment and fusion of multi-source government data, joint analysis of dynamic and static features, and prediction of spatiotemporal distribution, in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a big data-based intelligent statistical method for real estate enterprises, including:

[0007] Obtain multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments, and extract unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data.

[0008] A natural language processing model is used to extract semantic features from unstructured address description texts, mapping each address text to a standard address vector of a unified dimension. Based on the similarity calculation results of the standard address vectors, cross-source entity alignment is performed on static registration records to construct an initial real estate enterprise association graph.

[0009] Real-time dynamic pedestrian flow trajectory data of the target area is obtained, and the dynamic pedestrian flow trajectory data is mapped to the geospatial coordinate system corresponding to the initial human-real estate enterprise association map through spatial coordinate mapping rules to obtain the target human-real estate enterprise association map.

[0010] Extract the number of people residing in each spatial grid unit of the target real estate enterprise relationship map, the scale of enterprise registration and employment, and the registered capacity of housing to generate a spatial relationship feature set; simultaneously obtain the electricity frequency data and tax status data of each housing and enterprise entity in the target area to generate a business attribute feature set.

[0011] The spatial association feature set and the business attribute feature set are input into a pre-trained random forest classification model for inference operations, and the vacancy risk probability and false registration risk probability of each house and enterprise entity in the target area are output.

[0012] Risk labeling results for each entity are generated based on the probability of vacancy risk and the probability of false registration risk. Entities marked as warning status are selected based on the risk labeling results. Time series observation data within a preset time window is retrieved. Time series observation data is used as input features and risk labeling results are used as supervision labels to construct time series training samples for incremental training of the time series prediction model.

[0013] Input the time-series training samples into the pre-built long short-term memory network model, perform incremental training and update the model's internal weight parameters to obtain the updated long short-term memory network prediction model.

[0014] Extract the historical time-series state sequence corresponding to the target real estate enterprise association map, input it into the updated long short-term memory network prediction model, and obtain the spatial distribution prediction results of real estate enterprises in the target area within the future preset time period.

[0015] In one embodiment, the initial real estate enterprise association graph is constructed through the following process:

[0016] For unstructured address representation text, a natural language processing model is used to extract semantic features and generate standard address vectors corresponding to each address text.

[0017] The natural language processing model is used to identify multi-level administrative hierarchy elements and spatial location elements in address text, and then the two types of elements are vectorized and encoded separately before being spliced ​​and fused.

[0018] Extract the corresponding spatial coordinate features and address hierarchy features from each standard address vector.

[0019] Based on the spatial coordinate characteristics and address hierarchy characteristics corresponding to different data source addresses, the spatial topology matching degree between addresses is calculated.

[0020] The spatial topology matching degree value is compared with the preset matching threshold.

[0021] If the spatial topology matching degree value is greater than the preset matching threshold, it is determined that the addresses of different data sources point to the same geographic entity, and the corresponding population, housing, and enterprise entities are marked as target entity alignment nodes.

[0022] Establish cross-source entity associations based on the alignment nodes of all target entities, and generate an initial real estate enterprise association graph.

[0023] In one embodiment, the formula for calculating the spatial topology matching degree between addresses is:

[0024]

[0025]

[0026]

[0027] in, This represents the two entity addresses to be matched. With address Spatial topological matching degree, Represents the spatial coordinate similarity components. Address With address The spherical geographic distance between coordinate points is calculated based on latitude and longitude using the Haversine formula. This represents the distance attenuation bandwidth parameter. This indicates the address hierarchy matching degree. Indicates the total number of address levels. Indicates the first Hierarchical matching weights Indicates the first Hierarchical matching identifiers, Represents the spatial topology consistency check factor. This represents the spatial weight balance coefficient.

[0028] Secondly, this application also provides a smart statistical device for real estate enterprises based on big data, the device including:

[0029] The graph construction module is used to acquire multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments. It extracts unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data. It is also used to extract semantic features from the unstructured address description text using a natural language processing model, mapping each address text to a standard address vector of a unified dimension. Based on the similarity calculation results of the standard address vectors, it performs cross-source entity alignment on the static registration records to construct an initial person-real estate-enterprise association graph. It is also used to acquire real-time dynamic pedestrian flow trajectory data of the target area, and map the dynamic pedestrian flow trajectory data to the geospatial coordinate system corresponding to the initial person-real estate-enterprise association graph through spatial coordinate mapping rules to obtain the target person-real estate-enterprise association graph.

[0030] The risk reasoning module is used to extract the number of people residing in each spatial grid unit of the target real estate enterprise entity, the scale of enterprise registration and employment, and the registered capacity of housing from the target real estate enterprise association graph, and generate a spatial association feature set; simultaneously, it acquires the electricity frequency data and tax status data of each housing and enterprise entity in the target area, and generates a business attribute feature set; it is also used to input the spatial association feature set and the business attribute feature set into a pre-trained random forest classification model for reasoning operations, and output the vacancy risk probability and false registration risk probability of each housing and enterprise entity in the target area.

[0031] The incremental training module is used to generate risk labeling results for each entity based on the probability of vacancy risk and the probability of false registration risk; it filters entities marked as warning based on the risk labeling results, retrieves time-series observation data within a preset time window, uses the time-series observation data as input features and the risk labeling results as supervision labels, and constructs time-series training samples for incremental training of the time-series prediction model; it is also used to input the time-series training samples into a pre-constructed long short-term memory network model, perform incremental training and update the internal weight parameters of the model, and obtain the updated long short-term memory network prediction model.

[0032] The distribution prediction module is used to extract the historical time series state sequence corresponding to the target person-real estate enterprise association map, input the updated long short-term memory network prediction model, and obtain the spatial distribution prediction results of people and real estate enterprises in the target area within the future preset time period.

[0033] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0034] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0035] The aforementioned big data-based intelligent statistical method, device, computer equipment, and storage media for real estate enterprises first collect multi-source heterogeneous government data from three major government departments: public security, housing and construction, and market supervision. The raw data is then analyzed and classified to separate and extract unstructured address representation data in free text format, as well as structured static registration records corresponding to population registration, property ownership, and enterprise business registration. A natural language processing model is used to extract semantic features and encode elements from the unstructured address text, transforming the inconsistent and disparate address text into standardized address vectors with consistent dimensions and semantic regularity. By calculating the similarity between address vectors, entity matching and alignment across data sources are achieved, eliminating data silos and entity misalignment issues between different government systems. Based on this, an initial real estate enterprise relationship graph containing population, housing, enterprise entities, and their associated relationships is constructed. Based on this, real-time dynamic pedestrian flow trajectory data of the target area is collected. According to preset spatial coordinate mapping rules, the dynamic trajectory data undergoes coordinate system transformation and spatial registration to ensure consistency between the dynamic pedestrian flow data and the geospatial benchmark of the initial person-real estate-enterprise association map. This achieves the fusion and updating of dynamic spatiotemporal data and static entity maps, forming a target person-real estate-enterprise association map that synchronously reflects the static registration attributes and dynamic spatial activity characteristics of entities. Further, using the target person-real estate-enterprise association map as the data foundation, feature statistics are conducted according to preset spatial grid granularity. The number of people residing in each building entity and enterprise entity, the scale of registered enterprise employment, and the registered building capacity are extracted within each spatial grid to construct a spatial association feature set. Simultaneously, electricity frequency data and tax status data corresponding to each building and enterprise entity are collected to construct a business attribute feature set that characterizes the actual operation and usage status of the entities. The spatial association feature set and business attribute feature set are then fused and input into a pre-trained random forest classification model to complete inference calculations, accurately outputting the vacancy risk probability and false registration risk probability for each building and enterprise entity. Risk status labeling of all entities across the entire domain is completed based on two types of risk probabilities. Entities with abnormal risks are screened out for early warning. Continuous time-series observation data of these entities within a fixed time window are retrieved. Using the time-series observation data as model input features and the risk labeling results as supervision labels, a structured time-series training sample for incremental model training is constructed. The constructed time-series training sample is used to incrementally iteratively train the initial Long Short-Term Memory (LSTM) network model. The model weight parameters are continuously optimized through forward propagation, bias calculation, and backpropagation gradient updates. After meeting the convergence condition, an updated and optimized LSM prediction model is obtained. Finally, the historical time-series state sequence corresponding to the target person-real estate enterprise association map is extracted and input into the updated prediction model to complete the time-series extrapolation. The predicted spatial distribution of people and real estate enterprises in the target area within a preset future time period is output, completing the fully automated statistical analysis process from multi-source data fusion, entity association modeling, intelligent risk identification, autonomous sample construction, incremental model optimization, to time-series spatial prediction.

[0036] This invention establishes a fusion link between heterogeneous government data from multiple departments and dynamic spatiotemporal urban data. Through address vectorization and vector similarity matching, it achieves precise cross-source entity alignment, effectively solving the technical problems of data fragmentation and distorted entity matching in traditional statistical methods. By integrating spatial grid residency features, entity labor capacity features, and electricity and tax business features, it achieves objective quantification of the actual operational status of buildings and enterprises, overcoming the shortcomings of traditional static ledger statistical models that rely on registration data and cannot identify false statuses. Simultaneously, based on risk identification results, this invention automatically constructs time-series monitoring samples, enabling continuous incremental iterative updates to the time-series prediction model, solving the problems of weak generalization ability and accuracy decay over time in traditional fixed models. This invention forms an intelligent statistical system integrating data fusion, risk assessment, model self-optimization, and spatiotemporal prediction, capable of stably, accurately, and dynamically completing regional human, housing, and enterprise resource statistics and spatial distribution projections, effectively improving the intelligence and refinement of urban real estate governance and market entity supervision. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A flowchart illustrating the intelligent statistical method for real estate enterprises based on big data provided in this embodiment of the invention;

[0039] Figure 2 The structural block diagram of the intelligent statistical device for real estate enterprises based on big data provided in the embodiments of the present invention is shown. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0041] In one embodiment, such as Figure 1 As shown, this application provides a big data-based intelligent statistical method for real estate enterprises, which may include the following steps:

[0042] Step S101: Obtain multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments, and extract unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data.

[0043] The system uniformly accesses three types of government business databases: public security, housing and construction, and market supervision, acquiring multi-source heterogeneous government data covering population registration, housing property rights, and business registration in batches. This multi-source heterogeneous government data includes structured registration fields and unstructured free text data stored independently by each department. This step first performs structured parsing, field cleaning, and data diversion on the three types of raw government data, removing invalid null values, duplicate and redundant data, and abnormally dirty data. Then, two core data categories are extracted separately from the cleaned government data. The first category consists of unstructured address descriptions freely entered in the address fields of each department, without a unified format, including natural language address content such as house addresses, business registration addresses, and registered residence addresses. The second category consists of standardized structured static registration records, specifically including population registration information, housing property rights filing information, and business registration information.

[0044] Step S102: Use a natural language processing model to extract semantic features from unstructured address description texts and map each address text into a standard address vector of a unified dimension; based on the similarity calculation results of the standard address vectors, perform cross-source entity alignment on static registration records and construct an initial real estate enterprise association graph.

[0045] First, a pre-trained natural language processing model is used to perform semantic element parsing and feature encoding on all unstructured address texts. This automatically identifies multi-level spatial elements in the address text, such as administrative regions, streets / townships, residential buildings, unit floors, and location coordinates. Natural address texts of varying lengths, expressions, and abbreviations are uniformly encoded into fixed-dimensional, semantically normalized standard address vectors, achieving a structured and vectorized representation of unstructured addresses. Subsequently, cross-source similarity metric calculations are performed on all standard address vectors to obtain the matching correlation degree between corresponding addresses in different government systems. Based on the vector similarity matching results, it is determined whether population records, housing records, and enterprise records from different sources point to the same geographic entity, completing cross-departmental and cross-system entity association and alignment binding. After completing the full-domain entity alignment, a three-element topological association structure of people, houses, and enterprises is built based on the attribution, location, and registration relationships between entities. Finally, an initial person-house-enterprise association graph containing population nodes, housing nodes, enterprise nodes, and associated edges is generated, realizing structured association modeling of multi-source static government data.

[0046] Step S103: Obtain real-time dynamic pedestrian trajectory data of the target area, and map the dynamic pedestrian trajectory data to the geospatial coordinate system corresponding to the initial human-real estate-enterprise association map through spatial coordinate mapping rules to obtain the target human-real estate-enterprise association map.

[0047] The dynamic pedestrian trajectory data includes anonymous location sampling data, spatiotemporal dwelling records, area entry and exit identification data, and equipment calibration spatial data, exhibiting characteristics of high dynamism, high temporal sequence, and high spatial dispersion. The initial person-real estate-enterprise association map uses a fixed and unified government geospatial coordinate system, while the coordinate systems of the original pedestrian trajectory data collection equipment are inconsistent. This step, based on preset spatial coordinate mapping and transformation rules, performs coordinate system unification, spatial point correction, and spatiotemporal registration processing on all dynamic pedestrian trajectory data, mapping and aligning all trajectory points to the standard geospatial coordinate system corresponding to the initial person-real estate-enterprise association map, achieving spatial benchmark unification between dynamic data and static maps. Based on complete spatial location matching, dynamic pedestrian dwelling characteristics and spatial aggregation characteristics are bound to the corresponding building entity nodes and enterprise entity nodes in the map, updating the dynamic attribute information of the map nodes, and fusing static registration attributes with dynamic spatiotemporal attributes to finally obtain the target person-real estate-enterprise association map after dynamic and static data fusion and updating.

[0048] Step S104: Extract the number of people residing in each spatial grid unit of the housing entity and enterprise entity, the scale of enterprise registration and employment, and the capacity of housing registration from the target real estate enterprise association map to generate a spatial association feature set; Simultaneously obtain the electricity frequency data and tax status data corresponding to each housing and enterprise entity in the target area to generate a business attribute feature set.

[0049] First, the target area is divided into a global grid according to a preset unified spatial grid granularity, forming standard spatial grid units arranged in a regular pattern. Using each spatial grid unit as a statistical unit, all building and enterprise entity nodes within the grid's coverage area are traversed. Three spatial dimension quantitative indicators are calculated for each entity: the number of valid personnel residing there, the scale of registered and filed employment for enterprises, and the legally registered usable capacity of buildings. The grid statistical results are then correlated and integrated with entity attributes to generate a spatial association feature set covering all entities in the region. Simultaneously, long-term electricity frequency statistics, tax declaration status, and business registration data for each building and enterprise entity within the target area are independently retrieved. Missing values ​​are filled, outliers are removed, and the original business data is standardized and normalized to form a business attribute feature set that accurately reflects the actual usage and operational status of the entities.

[0050] Step S105: Input the spatial association feature set and the business attribute feature set into the pre-trained random forest classification model for inference operation, and output the vacancy risk probability and false registration risk probability of each house and enterprise entity in the target area.

[0051] First, feature fusion, feature concatenation, and dimensionality normalization are performed on the two feature sets to remove redundant and invalid features, generating unified multi-dimensional feature vectors for each housing entity and enterprise entity. All entity multi-dimensional feature vectors are then batch-input into a random forest classification model that has completed offline pre-training and parameter convergence, and the model's forward inference is performed. The random forest classification model uses parallel branching inference across multiple decision trees and ensemble voting weighted calculation to probabilistically quantify the vacancy status of housing entities and the false registration status of enterprises, ultimately outputting a unified structured risk probability matrix. Each row in this risk probability matrix corresponds to an entity and contains two fixed output indicators: the vacancy risk probability and the false registration risk probability for the current entity.

[0052] Step S106: Generate risk labeling results for each entity based on the probability of vacancy risk and the probability of false registration risk; filter entities labeled as warning status according to the risk labeling results, retrieve time series observation data within a preset time window, and construct time series training samples for incremental training of the time series prediction model using the time series observation data as input features and the risk labeling results as supervision labels.

[0053] Specifically, preset vacancy risk thresholds and false registration risk thresholds are retrieved separately. The probabilities of the two risk categories corresponding to each entity are compared with the corresponding thresholds. Risk binary classification is performed on each entity using threshold determination rules, generating risk labeling results for each entity. These results include labels for normal status and warning abnormal status. After completing the full-domain entity labeling, high-risk entities with warning abnormal status are accurately selected. For the selected high-risk entities, all historical time-series observation data for that entity within a preset continuous time window are retrieved. The time-series observation data includes continuous time-series sequences of the entity's historical spatial characteristics, pedestrian flow characteristics, and electricity consumption and tax business characteristics. In this step, the retrieved time-series observation data is defined as the model training input features, and the risk labeling results corresponding to the entities are defined as model supervision labels. Through a structured combination method that binds features and labels one-to-one, time-series training samples that can be directly used for incremental supervised training of the time-series prediction model are constructed in batches.

[0054] Step S107: Input the time-series training samples into the pre-built long short-term memory network model, perform incremental training and update the internal weight parameters of the model to obtain the updated long short-term memory network prediction model.

[0055] First, pure temporal feature sequences and corresponding supervision labels are extracted from the temporal training samples. These temporal feature sequences are then input into a Long Short-Term Memory (LSTM) network model for forward propagation. Leveraging the model's internal multi-gating mechanism and cell state update logic, temporal deep evolutionary features are extracted, and the final hidden state vector is output. The model's prediction result is obtained through output layer mapping. The difference between the model's prediction result and the corresponding real risk supervision label is calculated to obtain the model training prediction bias. Based on the prediction bias, backpropagation is performed to solve for the incremental gradients of each weight parameter layer by layer. The initial weight matrix is ​​iteratively updated using the incremental gradients to obtain the updated weight matrix for each iteration. This iterative training process of forward propagation, bias calculation, and backpropagation gradient update is repeated, updating the weight matrix in each iteration. The current weight matrix and model loss value are continuously checked to ensure they meet preset convergence conditions. These conditions can be set as the loss value being less than a preset threshold, the change in weight parameters being less than a threshold, or the number of iterations reaching an upper limit. If the convergence conditions are met, incremental training is terminated, and the current weight parameters are fixed, resulting in a LSM prediction model with optimized parameters and updated accuracy.

[0056] Step S108: Extract the historical time series state sequence corresponding to the target real estate enterprise association map, input it into the updated long short-term memory network prediction model, and obtain the prediction result of the spatial distribution of real estate enterprises in the target area within the future preset time period.

[0057] This process continuously extracts historical time-series state sequences of various housing and enterprise entities across multiple time periods and dimensions from the target real estate enterprise relationship graph. The time-series sequences undergo preprocessing such as missing data completion, time-series alignment, and normalization to ensure consistent dimensions and temporal continuity. The standardized historical time-series state sequences are then input into a convergent Long Short-Term Memory (LSTM) network prediction model. The model learns the temporal transition patterns of entity states through forward propagation and outputs the state transition probabilities of each entity node. Combining the pre-defined standard spatial grid cells of the target area, the entity-level state transition probabilities are mapped to the corresponding spatial grid cells. The spatial distribution matrix of real estate enterprises across the entire region is iteratively calculated and updated, with the matrix convergence degree continuously calculated in adjacent iterations. When the matrix convergence degree exceeds a pre-defined convergence threshold, the spatial distribution result is considered stable. The current converged spatial distribution matrix is ​​then determined as the predicted spatial distribution result of real estate enterprises in the target area for the next pre-defined time period, ultimately completing the intelligent time-series prediction output of the spatial distribution of real estate enterprises across the entire region.

[0058] The aforementioned big data-based intelligent statistical method for real estate and enterprise data first collects multi-source heterogeneous government data from public security, housing and construction, and market supervision departments, extracting unstructured address text and static registration records of population, housing, and enterprises. A natural language processing model is used to semantically encode and vectorize the address text, and cross-source entity alignment is achieved based on address vector similarity to construct an initial real estate and enterprise association graph. By performing coordinate mapping and spatial registration on dynamic pedestrian flow trajectory data in the target area, dynamic spatiotemporal features are fused into the initial graph to obtain an updated target real estate and enterprise association graph. This invention extracts spatial association features such as entity occupancy, employment scale, and housing capacity based on grid-based statistics, and constructs a business attribute feature set by combining entity electricity usage frequency and tax status. The fused features are input into a pre-trained random forest classification model, outputting the vacancy risk probability and false registration risk probability of each housing and enterprise entity. This invention completes entity risk labeling based on risk probabilities, filters early warning entities, and extracts time-series observation data within a preset time window. Time-series training samples are constructed using time-series data as input and risk labels as supervisory tags. By incrementally training and iteratively optimizing the weights of a Long Short-Term Memory (LSTM) network model using samples, an updated time-series prediction model is obtained upon convergence. Finally, historical time-series state sequences from the geographic map are extracted and input into the optimized model to predict the spatial distribution of people, real estate, and enterprises in the target area for future periods, achieving fully automated and intelligent statistical analysis. This invention effectively integrates multi-source heterogeneous government data with dynamic spatiotemporal data, improving the accuracy of cross-departmental entity alignment through semantic vectorization matching, thus solving the problems of traditional data fragmentation and matching distortion. By combining spatial residency characteristics and business operation characteristics to quantify the actual operational status of entities, it effectively compensates for the shortcomings of traditional static ledgers in identifying vacancy and false registration. Furthermore, this invention can automatically construct training samples based on risk labeling, enabling incremental iterative updates of the model and avoiding the problem of decaying generalization ability of fixed models, significantly improving the accuracy and intelligence level of urban real estate and market entity governance.

[0059] In one embodiment, the initial real estate enterprise association graph can be constructed through the following process:

[0060] Step S201: For unstructured address representation text, a natural language processing model is used to extract semantic features and generate standard address vectors corresponding to each address text.

[0061] In step S202, the natural language processing model is used to identify multi-level administrative hierarchy elements and spatial location elements in the address text, and then the two types of elements are vectorized and encoded separately before being spliced ​​and fused.

[0062] Optionally, the model first decomposes the input address text into elements, identifying the multi-level administrative hierarchy elements and refined spatial location elements contained within the text. The administrative hierarchy elements include information at the provincial, municipal, district / county, street, and community levels, while the spatial location elements include precise spatial information such as neighborhood, building, unit, address, and geographic location. The model performs independent vectorization encoding operations on both types of elements, obtaining administrative hierarchy vectors and spatial location vectors. Then, the two sets of vectors are dimensionally concatenated and fused to generate a complete standard address vector that simultaneously contains hierarchical semantics and spatial location information, ensuring that the vector features cover both the hierarchical and spatial attributes of the address.

[0063] Step S203: Extract the corresponding spatial coordinate features and address hierarchy structure features from each standard address vector.

[0064] Among them, spatial coordinate features are used to characterize the actual geographical latitude and longitude spatial location corresponding to the address, and address hierarchical structure features are used to characterize the matching structure and hierarchical affiliation of administrative elements at all levels of the address.

[0065] Step S204: Calculate the spatial topology matching degree between addresses based on the spatial coordinate features and address hierarchy features corresponding to different data source addresses.

[0066] Step S205: Compare the spatial topology matching degree value with the preset matching threshold.

[0067] Step S206: If the spatial topology matching degree value is greater than the preset matching threshold, it is determined that the addresses of different data sources point to the same geographic entity, and the corresponding population, housing, and enterprise entities are marked as target entity alignment nodes.

[0068] Based on the threshold comparison results, entity alignment is determined. When the spatial topology matching degree value corresponding to any two cross-data source addresses is greater than the preset matching threshold, it is determined that the two address information with different descriptions and different sources point to the same real geographic entity within the city. Based on the address homogeneity determination result, the population registration entity, housing property entity, and enterprise business entity corresponding to the address are associated and bound, and marked as target entity alignment nodes under the same geographic entity, realizing homogeneous association matching of heterogeneous entity data from multiple departments.

[0069] Step S207: Establish cross-source entity association relationships based on the alignment nodes of all target entities to generate an initial real estate enterprise association graph.

[0070] Specifically, for unstructured address descriptions in multi-source government data, a natural language processing (NLP) model is used to perform refined semantic feature extraction and unified vectorization encoding. The NLP model specifically targets the structural characteristics of address text, automatically identifying and splitting multi-level administrative hierarchy elements and refined spatial location elements within the text. After performing vector encoding on these two independent elements separately, feature concatenation and fusion are performed to generate standard address vectors with unified dimensions, complete semantics, and regular structure, achieving standardized digital representation of various irregular and differentiated unstructured address texts. After address vectorization, quantifiable spatial coordinate features and address hierarchy structure features are individually extracted from each standard address vector, corresponding to the address's geographical location information and administrative hierarchy affiliation information, respectively. For address feature data corresponding to different government data sources, spatial coordinate similarity and hierarchy structure similarity are simultaneously combined to uniformly calculate the spatial topological matching degree between pairs of addresses. This matching degree takes into account both geographic spatial distance constraints and administrative hierarchy matching constraints, objectively reflecting the degree of entity homology between different addresses. The calculated spatial topology matching degree value is compared with the system's preset entity matching threshold. When the spatial topology matching degree value between addresses is greater than the preset matching threshold, it is determined that the address information corresponding to different data sources points to the same real geographic entity. Then, the population entity, housing entity, and enterprise entity corresponding to that address are identified as relatable target entity alignment nodes. After completing the cross-source entity matching verification across the entire domain, cross-departmental and cross-data source entity association correspondences are established based on all target entity alignment nodes. This unifies and connects the three independent entity data of population, housing, and enterprises, ultimately constructing an initial human-housing-enterprise association graph with a complete structure and accurate entity correspondences.

[0071] This embodiment abandons the traditional address fusion method that relies solely on text character matching. Instead, it uses a natural language processing model to decompose unstructured addresses into elements and perform two-dimensional vectorized encoding, solving the matching failure problem caused by address abbreviations, non-standard expressions, and large textual differences. This invention simultaneously integrates spatial coordinate features and address hierarchy structure features to calculate topological matching degree, achieving dual constraints of semantic matching and geographic topological verification. This significantly improves the accuracy and rationality of cross-source entity alignment, avoiding mismatches and omissions caused by single matching rules. By achieving full-domain entity alignment through standardized and quantifiable judgment rules, it effectively breaks down data barriers between multiple departments such as public security, housing and construction, and market supervision, eliminating the problems of fragmented and scattered multi-source heterogeneous data and chaotic entity correspondence.

[0072] In one embodiment, the formula for calculating the spatial topology matching degree between addresses is:

[0073]

[0074]

[0075]

[0076] in, This represents the two entity addresses to be matched. With address Spatial topological matching degree, Represents the spatial coordinate similarity components. Address With address The spherical geographic distance between coordinate points is calculated based on latitude and longitude using the Haversine formula. This represents the distance attenuation bandwidth parameter. This indicates the address hierarchy matching degree. Indicates the total number of address levels. Indicates the first Hierarchical matching weights Indicates the first Hierarchical matching identifiers, Represents the spatial topology consistency check factor. This represents the spatial weight balance coefficient.

[0077] Optionally, , Address Does the coordinate point fall within the address? Within the geographical boundary of the corresponding minimum matching level, the value is 1 if it is true, otherwise the value is the preset attenuation coefficient. ; Address Does the coordinate point fall within the address? Within the geographic boundary of the corresponding minimum matching level, the value is 1 if it is true, and 1 otherwise. When both coordinates fall within the geographic range of the other's address level, The matching degree has no decay; when there is a coordinate out-of-bounds error, the matching degree decays proportionally, and the forced matching result satisfies the spatial topological membership logic.

[0078] This is the spatial weighting balance coefficient, with a value range of [value range missing]. This is used to adjust the fusion ratio of spatial coordinate similarity and address hierarchy structure similarity.

[0079] The larger the value, the more the model focuses on geospatial distance matching; The smaller the value, the more the model focuses on matching the administrative hierarchy structure.

[0080] This is the Gaussian distance decay bandwidth, used to control the rate at which geographical distance similarity decays.

[0081] The smaller the similarity, the faster it decreases as the distance increases, and the higher the requirement for point accuracy.

[0082] The larger the value, the higher the tolerance for address similarity within the neighborhood.

[0083] For the first Matching binary identifiers at the address level:

[0084] If address ,address In the If the content is consistent across administrative levels, then ;

[0085] If the hierarchical content is inconsistent, mismatched, or the hierarchy is missing, then .

[0086] This embodiment constructs a comprehensive spatial topology matching degree calculation method by integrating spatial coordinate similarity, address hierarchy structure similarity, and bidirectional spatial topology consistency verification factors. This overcomes the problems of traditional address matching, which relies solely on text character comparison or single geographical distance calculation, resulting in one-sided matching, poor fault tolerance, and easy generation of cross-level and cross-regional mismatches. By adaptively balancing the matching ratio of spatial and structural dimensions through weight coefficients, and combining Gaussian distance decay mechanism and multi-level weighted matching mechanism, it achieves accurate quantitative matching of non-standard, heterogeneous, and differentiated address texts. At the same time, relying on bidirectional geographical boundary verification to constrain topological membership logic, it effectively improves the accuracy and stability of multi-source cross-department address entity alignment.

[0087] In one embodiment, mapping dynamic pedestrian flow trajectory data to the geospatial coordinate system corresponding to the initial person-real estate company association map using spatial coordinate mapping rules to obtain the target person-real estate company association map may include the following steps:

[0088] Step S301: Perform spatiotemporal preprocessing on the dynamic pedestrian flow trajectory data and extract a continuous trajectory coordinate sequence with timestamps.

[0089] The dynamic pedestrian flow trajectory data includes timestamped sampling sequences of pedestrian locations, entry and exit records of area boundaries, spatial calibration data of sensing devices, and anonymous identification data of pedestrians.

[0090] Step S302: Density clustering algorithm is used to identify dwell points and calculate clustering degree for the trajectory coordinate sequence to obtain the dwell heat value corresponding to each spatial point.

[0091] Optionally, the DBSCAN density clustering algorithm is used to calculate the trajectory coordinate sequence, and the algorithm configures the neighborhood radius. The system uses two parameters, MinPts (Minimum Core Points), to distinguish between long-staying core clusters and passing noise trajectory points, automatically identifying all personnel's spatial locations. The formula for calculating the dwell time intensity is as follows:

[0092]

[0093] In the formula, Spatial point The heat index of the residence; For point Total number of trajectory sampling points within a cluster; This refers to the cumulative time a person spends at that location. This represents the total duration of the statistical period. By traversing all clusters to complete cluster quantification, a unique dwell heat value is obtained for each spatial location, thus realizing the conversion of discrete trajectories into quantifiable indicators of regional human activity.

[0094] Step S303: Based on the dwell heat value and spatial coordinates of each point, generate gridded trajectory dwell feature data with spatial location labels.

[0095] Step S304: Obtain the unified geospatial coordinate system corresponding to the initial real estate enterprise association map, and divide it into several standard spatial grid units according to the preset grid granularity.

[0096] The target area is uniformly divided according to the fixed grid granularity preset by the business, generating standard spatial grid units of consistent size that seamlessly cover the target area. The boundary coordinates and unique grid number of each grid unit are recorded, and a standardized spatial grid index system is constructed to enable dynamic data and static maps to share the same spatial reference.

[0097] Step S305: Spatial matching is performed between the gridded trajectory dwell feature data and the standard spatial grid cells to construct a grid-entity spatial mapping matrix.

[0098] Among them, the grid-entity spatial mapping matrix represents the spatial correspondence between each spatial grid and the building entity nodes and enterprise entity nodes in the map.

[0099] The generated gridded trajectory dwell feature data is spatially overlaid and matched with the divided standard spatial grid cells to establish the spatial mapping relationship between grid cells and map entity nodes, and a two-dimensional grid-entity spatial mapping matrix is ​​constructed. The matrix dimension is the total number of grid cells × the total number of entity nodes, and the matrix elements... A value of 1 represents the first The first standard grid and the first The spatial coverage of each house / enterprise entity node is intersected. A value of 0 indicates that there is no spatial overlap between the two, which fully represents the spatial correspondence between the grid and the graph entities, and builds a mapping bridge between dynamic human flow characteristics and static entity nodes.

[0100] Step S306: Based on the grid-entity spatial mapping matrix, calculate the association edge weights between each house and enterprise entity node and the dynamic pedestrian flow data in the initial human-real estate-enterprise association graph.

[0101] Based on the sum of dwell time of each entity's associated grid according to the grid-entity spatial mapping matrix, the weight of the associated edge between the entity node and the dynamic pedestrian flow data is calculated. The calculation formula is as follows:

[0102]

[0103] In the formula, For entity nodes The corresponding associated edge weights; This represents the total number of standard grid cells across the entire region. For the first The grid-based dwell time of each grid. Edge weights quantify the degree of correlation between the dynamic activities of buildings, business entities, and people; higher values ​​indicate that the entity has high-frequency dwell time.

[0104] Step S307: If the weight of the associated edge is greater than the preset associated threshold, the residence heat feature of the corresponding spatial grid is associated with the corresponding entity node of the initial real estate enterprise association graph and the node attributes are updated to obtain the target real estate enterprise association graph.

[0105] Specifically, the first step is to acquire dynamic pedestrian flow trajectory data for the target area. This data includes timestamped sampling sequences of pedestrian locations, entry and exit records at area boundaries, spatial calibration data from sensing devices, and anonymous identification data of pedestrians. Spatiotemporal preprocessing is performed on the raw dynamic pedestrian flow trajectory data to remove abnormal trajectory points and invalid noise data, resulting in a continuous, time-aligned trajectory coordinate sequence with precise timestamps. Density clustering algorithms are then used to perform cluster analysis on the preprocessed trajectory coordinate sequence, automatically identifying spatial points where people have long resided and calculating the population density at each spatial point. The residency heat value corresponding to each spatial point is quantified and output, achieving a structured transformation from discrete trajectory points to regional residency intensity characteristics. Combining the residency heat value of each spatial point with its corresponding spatial coordinate information, gridded regularization processing of the global trajectory features is completed, generating gridded trajectory residency feature data with precise spatial location labels. This solution further matches the unified geospatial coordinate system used in the initial human-real estate-enterprise association map, dividing the target area into a whole according to a preset fixed grid granularity, generating standard spatial grid units with uniform size and spatial continuity. The regularized gridded trajectory dwell feature data is overlaid and matched with standard spatial grid cells across the entire domain to establish the positional correspondence between grid space and entity nodes, constructing a grid-entity spatial mapping matrix. This matrix is ​​used to quantitatively represent the one-to-one or one-to-many spatial association between each standard spatial grid and the building entity nodes and enterprise entity nodes within the map. Based on the constructed grid-entity spatial mapping matrix, the association edge weights of each building entity node and enterprise entity node in the initial person-real estate-enterprise association map and the dynamic pedestrian trajectory data are statistically calculated. These edge weights represent the degree of association between entities and dynamic pedestrian activities. The calculated association edge weights are compared with a preset association threshold. For entity nodes with association edge weights greater than the preset threshold, the dwell heat feature of the corresponding spatial grid is bound and updated to the entity node attributes of the initial person-real estate-enterprise association map, completing the fusion update of static map attributes and dynamic pedestrian features, and finally obtaining the target person-real estate-enterprise association map with fused static and dynamic data.

[0106] This embodiment achieves the structured and feature-based transformation of discrete pedestrian trajectory data through spatiotemporal preprocessing and density clustering algorithms. This enables accurate extraction of real-time hotspots of pedestrian activity, avoiding the problems of discrete, noisy, and unquantifiable raw trajectory data. By unifying the geographic coordinate system and standardizing grid partitioning, the spatial benchmark between dynamic spatiotemporal data and static correlation maps is unified. Based on the grid-entity spatial mapping matrix, quantifiable and reproducible spatial correlation rules are established, solving the technical pain point of accurately matching and integrating dynamic pedestrian data with building and enterprise entities. By setting correlation edge weights and threshold filtering mechanisms, only highly correlated and effective dynamic features are updated to map nodes, filtering out invalid and weakly correlated interference data, ensuring the accuracy and effectiveness of map updates. This solution achieves deep integration of static registered entity attributes and dynamic spatiotemporal features of pedestrian flow, enabling the final constructed target person-building-enterprise correlation map to simultaneously reflect the inherent registration attributes and real-time dynamic usage status of entities.

[0107] In one embodiment, inputting the spatial association feature set and the business attribute feature set into a pre-trained random forest classification model for inference operations, and outputting the vacancy risk probability and false registration risk probability of each house and enterprise entity within the target area, may include the following steps:

[0108] Step S401: Based on the entity nodes and relationships between nodes in the target real estate enterprise association graph, a multi-dimensional attribute entity association network is constructed by integrating spatial association feature set and business attribute feature set.

[0109] Step S402: Extract the multidimensional feature vectors corresponding to each house and enterprise entity from the multidimensional attribute entity association network.

[0110] Traverse all building and enterprise entity nodes within the multidimensional attribute entity association network, read all standardized feature indicators attached to a single entity node, and concatenate and combine them in a fixed order to form a multidimensional feature vector with unified dimensions. subscript Representing the One entity, To ensure a unified total dimension for fixed features, the feature vectors output by different entities are made with identical lengths and sorting rules. Empty feature fields with no numerical meaning are removed, and all vectors are guaranteed to have standardized formats and aligned dimensions. This results in a standardized vector sample set that can be directly input into classification models in batches, completing the conversion of graph-structured data into numerical vectors for model input.

[0111] Step S403: Input the multidimensional feature vector into the pre-trained random forest classification model for inference operation and output the risk probability matrix.

[0112] The risk probability matrix includes the vacancy risk probability and the false registration risk probability for each property and business entity.

[0113] The output multi-dimensional feature vectors of all entities are batch-fed into a random forest classification model that has been trained offline and whose loss has converged, to perform forward inference computation. The random forest uses multiple independent decision trees to make judgments in parallel, outputting predicted probabilities for the two types of risks respectively. A stable output result is obtained through ensemble voting weighting, and finally a two-dimensional risk probability matrix is ​​generated. Each row in the matrix corresponds to a single property or business entity. The matrix contains two columns of fixed indicators; the first column represents the probability of the entity being vacant. The second column represents the probability of false registration of an entity. It fully records the dual-dimensional risk quantification results of all entities across the entire domain, and achieves unified quantitative output of the risk status of all entities.

[0114] Step S404: Compare the vacancy risk probability and false registration risk probability of each entity in the risk probability matrix with the corresponding preset risk thresholds.

[0115] The system pre-stores two sets of independent critical thresholds and a vacancy risk threshold. False registration risk threshold The two thresholds can be adjusted according to governance needs. Traverse each row of entity data in the risk probability matrix and perform two independent numerical comparison operations: and contrast, and The comparison fully records the relationship between the two probabilities of each entity and the corresponding threshold, providing a quantitative basis for screening high-risk entities. By comparing the two types of risks separately, single-type anomalies such as vacant properties and false business registrations can be independently identified, avoiding blind spots caused by a single threshold-based judgment.

[0116] Step S405: If any risk probability corresponding to any entity is greater than the corresponding preset risk threshold, the corresponding entity is identified as a high-risk entity, and the vacancy risk probability and false registration risk probability corresponding to the high-risk entity are output.

[0117] like or Then mark the entity. Entities identified as high-risk are classified as such. For entities deemed high-risk, the complete values ​​of the entity's vacancy risk probability and false registration risk probability are simultaneously retained and output uniformly. Entities with both probabilities below their corresponding thresholds are marked as normal entities and are not included in the high-risk output set.

[0118] Specifically, using the inherent spatial and attribution topology of the target real estate enterprise association graph as its underlying framework, the spatial association feature set and business attribute feature set generated in the previous steps are concatenated and normalized according to feature dimensions and then bound to the corresponding entity nodes to build a multi-dimensional attribute entity association network that carries multiple indicators such as static registration, spatial flow, and business operations. All housing entities and enterprise entities within the multi-dimensional attribute entity association network are traversed, and all standardized feature indicators bound to each entity node are extracted and combined to generate independent and dimensionally unified multi-dimensional entity feature vectors. The multi-dimensional feature vectors corresponding to all entities are batch-input into a random forest classification model that has completed offline pre-training and parameter convergence, and forward inference is performed. The model integrates the discrimination results of multiple decision trees to output a structured risk probability matrix. This matrix, on an entity-by-entity basis, synchronously stores two types of quantitative indicators for each record: the vacancy risk probability and the false registration risk probability of the corresponding housing or enterprise entity. The system retrieves the pre-set thresholds for determining vacancy risk and false registration risk, and compares the two risk probability values ​​of each entity in the risk probability matrix with the matched thresholds. When the risk probability of any one type of an entity exceeds the corresponding threshold, the entity is determined to be a high-risk entity, and the two risk probability values ​​corresponding to the entity are completely stored and output, thus completing the quantitative identification of risk classification for entities across the entire domain.

[0119] This embodiment integrates two types of feature datasets based on the original topological structure of the association graph, achieving unified modeling of spatial pedestrian flow indicators and business operation indicators. This avoids the data fragmentation problem caused by independent analysis of the two types of features and enriches the basis for entity status discrimination. By leveraging a random forest ensemble learning mechanism to synthesize multi-dimensional features and complete probabilistic risk output, the identification results are more objective compared to single-indicator judgment methods. Through independent comparison rules with categorized thresholds, it distinguishes between vacancy and false registration risks, accurately screening high-risk entities with abnormal states while fully retaining the risk quantification score. This achieves automated batch risk screening and provides accurate annotation basis for subsequent time-series training sample construction, improving the completeness and reliability of identifying risks such as vacant properties and false enterprise registration.

[0120] In one embodiment, inputting time-series training samples into a pre-built long short-term memory network model, performing incremental training, and updating the model's internal weight parameters to obtain an updated long short-term memory network prediction model may include the following steps:

[0121] Step S501: Extract the temporal feature sequence and corresponding supervision label from the temporal training sample, input the temporal feature sequence into the pre-constructed long short-term memory network model, and output the corresponding hidden state vector.

[0122] Optionally, standardized temporal feature sequences and their unique corresponding supervision labels are extracted from the constructed temporal training samples. The temporal feature sequence of a single sample is defined as... ,in The time series length is defined. The time-series feature sequence is input into a pre-built and initially pre-trained Long Short-Term Memory (LSTM) network model. The model collaboratively filters and updates time-series information through forgetting, input, and output gates, deeply encoding the feature change patterns across continuous time dimensions, and outputting the final hidden state vector corresponding to the current time series. This completes the transformation from raw time-series data to high-dimensional time-series features.

[0123] Step S502: Map the hidden state vector to generate the prediction result, and compare the prediction result with the corresponding supervision label to calculate the prediction bias.

[0124] The hidden state vector output by the Long Short-Term Memory network By connecting to a fully connected mapping layer, the model's predicted output value is obtained through linear mapping operations. The predicted value With sample real supervision label Error calculation is performed, and the prediction deviation is quantified using the mean squared error loss function. The calculation formula is as follows: In the formula The larger the loss function value, the greater the fitting error between the model's current prediction and the true label.

[0125] Step S503: Perform backpropagation of error based on prediction bias to generate weight increment gradient, and adjust the initial weight matrix of the model according to the weight increment gradient to obtain the updated weight matrix.

[0126] The gradients of the weight parameters at each layer of the model are calculated using the chain rule, yielding the weight increment gradients. Combined with the preset learning rate For the initial weight matrix of the model The weight update is performed iteratively, and the formula is as follows: The model parameters are corrected using gradient descent to reduce the overall prediction loss. An updated weight matrix is ​​generated after each iteration. .

[0127] Step S504: Iterate through the forward and backward propagation processes to determine whether the current iteration meets the preset convergence conditions.

[0128] The entire training process—forward propagation feature encoding, prediction result generation, loss bias calculation, and backpropagation gradient update—is repeatedly executed to achieve continuous iterative optimization of model parameters. After each iteration, the current model loss value and parameter update magnitude are recorded and compared with preset convergence conditions. Convergence conditions include three criteria: the loss value is below a preset threshold, the loss no longer decreases after multiple consecutive iterations, and the number of iterations reaches a preset maximum value. These conditions are used to quantitatively control when model training terminates.

[0129] Step S505: If the current iteration satisfies the preset convergence condition, then the updated long short-term memory network prediction model is generated using the current updated weight matrix.

[0130] When any of the preset convergence conditions is met during the iteration process, the model training and fitting are considered complete, and the incremental iterative training process is terminated. The latest weight matrix obtained in the current iteration is then fixed. The original initialization parameters of the model were replaced, and the overall parameters of the model were updated to obtain the converged updated Long Short-Term Memory Network prediction model. This model was incrementally optimized using real business time-series samples and can accurately fit the time-series evolution pattern of real estate enterprise risks.

[0131] Specifically, the process begins by separating and extracting standardized temporal feature sequences and their corresponding supervised label data from the constructed temporal training samples. These temporal feature sequences are then input into a pre-built and pre-trained Long Short-Term Memory (LSTM) network model. The model uses an internal control mechanism to perform temporal dependency learning and feature compression on the temporal feature sequences, outputting a hidden state vector that characterizes the temporal evolution of entity risk. Subsequently, the model performs dimensionality transformation on the hidden state vector through a fully connected mapping layer, generating a prediction result for the entity risk state. The difference between the model's output prediction result and the corresponding real supervised label is calculated to obtain the prediction bias of a single inference. Based on this prediction bias, backpropagation is performed to calculate the incremental gradient of the weights for each network parameter layer by layer. This incremental gradient is used to iteratively correct the original initial weight matrix, generating an updated weight matrix. The iterative training process of forward feature inference, prediction bias calculation, and backward weight update is continuously executed. After each iteration, the convergence status of the current model parameters and loss value is verified to determine whether the preset convergence conditions are met. When the iteration process meets the preset convergence condition, the incremental training process is terminated, the updated weight matrix obtained in the current iteration is solidified, and finally an updated long short-term memory network prediction model with optimized parameters and adapted to real-time scene data is generated.

[0132] This embodiment employs an incremental iterative training mode based on real business risk samples. While preserving the model's original basic learning capabilities, it continuously adapts to the dynamic changes in the status of urban real estate entities, addressing the problems of fixed parameters, long-term accuracy decay, and insufficient generalization ability inherent in traditional static models. Through forward propagation and temporal hidden feature extraction mechanisms, it fully mines the temporal evolutionary correlation features of risk data, enhancing the model's ability to learn from changes in entity status. Relying on the error backpropagation gradient update mechanism, it achieves adaptive parameter fine-tuning, ensuring that model iteration updates are based on rigorous data-driven principles. By setting standardized convergence criteria, it ensures stable and controllable incremental training effects, avoiding overfitting or underfitting issues. Ultimately, the updated prediction model accurately adapts to the spatiotemporal evolution characteristics of regional real estate entities, effectively improving the accuracy and reliability of subsequent spatial distribution temporal predictions.

[0133] In one embodiment, extracting the historical time-series state sequence corresponding to the target person-real estate enterprise association map, inputting it into the updated long short-term memory network prediction model, and obtaining the spatial distribution prediction result of people-real estate enterprises in the target area within a preset future time period may include the following steps:

[0134] Step S601: Obtain the multi-dimensional historical time sequence of the corresponding housing entity nodes and enterprise entity nodes in the target real estate enterprise association graph.

[0135] Based on the target real estate enterprise association graph with completed dynamic feature fusion, all housing entity nodes and enterprise entity nodes within the graph are traversed. All attribute data for each entity node within a continuous fixed time window are read, including spatial residency characteristics, personnel activity characteristics, electricity consumption characteristics, tax operation characteristics, and historical risk status data. Historical data with different dimensions and sampling frequencies undergoes temporal alignment, missing data completion, and normalization preprocessing, resulting in a unified, consistent, and temporally continuous multi-dimensional historical time-series state sequence that can be directly used for model inference.

[0136] Step S602: The updated Long Short-Term Memory network model is used to perform forward propagation operation on the temporal state sequence, and the state transition probability of each entity node is output.

[0137] The model uses forward propagation to read the dependencies between time-series data, learns the long-term evolution patterns of entity states, and outputs the state transition probabilities of each house and enterprise entity node in future time periods. State transition probabilities are used to quantify the likelihood of an entity transitioning from its current normal, vacant, or falsely registered state to various target states.

[0138] Step S603: Combining the pre-divided standard spatial grid units of the target area, the state transition probability of each entity node is mapped to the corresponding spatial grid to construct the spatial distribution matrix of people, real estate and enterprises.

[0139] The state transition probabilities of each entity node are matched and mapped to the standard grid cell to which it belongs. Using the spatial grid as the basic statistical unit, the state transition probabilities of all building entities and enterprise entities within the grid coverage area are aggregated to construct a globally regular spatial distribution matrix of people, real estate, and enterprises. The matrix expression is: ,in Representing the The distribution characteristic values ​​of each grid, entities within the mesh The state transition probability matrix fully records the future distribution of entities in each grid across the entire domain.

[0140] Step S604: Iteratively update the spatial distribution matrix and calculate the matrix convergence degree corresponding to adjacent iteration rounds.

[0141] The spatial distribution matrix of real estate companies is iteratively updated multiple times. In each iteration, the grid probability aggregation values ​​are adjusted according to the temporal evolution of entities, gradually optimizing the spatial distribution results. After each iteration, the matrix convergence of the current iteration and the previous iteration is calculated. The convergence calculation formula is as follows: In the formula, For the convergence of the matrix, For the first The spatial distribution matrix of the round iteration, The convergence factor is the matrix L2 norm, used to quantify the overall difference between two adjacent round matrices. The higher the convergence value, the smaller the matrix iteration change and the more stable the result.

[0142] Step S605: If the matrix convergence is greater than the preset convergence threshold, then the current spatial distribution matrix is ​​determined as the spatial distribution prediction result of the target area.

[0143] Specifically, the historical operational data of all housing entity nodes and enterprise entity nodes in the target real estate enterprise relationship graph are first retrieved. This data is then organized, aligned, and normalized to obtain a multi-dimensional historical time-series state sequence for each entity. The sequence covers the entity's historical spatial residency characteristics, business operation characteristics, and risk status changes, providing complete historical input data for time-series projection. The normalized time-series state sequence is then input into an incrementally trained Long Short-Term Memory (LSTM) network model. Through forward propagation, the model mines the temporal evolution patterns and state transition characteristics of entity states, outputting the state transition probability of each housing entity and enterprise entity in the future. Based on pre-divided standard spatial grid units in the target area, and according to the inherent spatial mapping relationship between entity nodes and spatial grids, the state transition probabilities output by each entity node are mapped one by one to the corresponding grid location, initially constructing a spatial distribution matrix of people, real estate enterprises, and enterprises covering the entire region. The matrix values ​​are used to characterize the future distribution probability level of people, housing, and enterprise entities within each spatial grid. Through multiple rounds of iterative computation, the distribution probability values ​​of each grid within the matrix are continuously corrected, and the matrix convergence between adjacent iterations is calculated synchronously to quantify the stability of the overall spatial distribution results. When the convergence of the matrix is ​​greater than the preset convergence threshold, it is determined that the spatial distribution probability of the entire domain tends to be stable, the iterative update is stopped, and the current converged spatial distribution matrix is ​​determined as the prediction result of the spatial distribution of people and enterprises in the target area for the future preset time period.

[0144] This embodiment relies on real historical time-series data from the geographic map to drive predictive extrapolation, abandoning traditional static statistics and empirical estimation methods. It objectively reflects the temporal evolution of the status of entities such as people, real estate companies, and enterprises. By mapping and matching the entity status transition probability with a standard spatial grid, it achieves accurate conversion of entity-level prediction results to regional grid-level distribution results, ensuring that the prediction results have fine spatial granularity. Through matrix iterative updates and convergence verification mechanisms, the stability of the prediction results is constrained, effectively avoiding the problems of high randomness and low reliability in single extrapolation results. This solution can accurately output the future spatial distribution of regional resources related to people, real estate companies, and enterprises, providing reliable data support for revitalizing idle urban resources, predicting market entity risks, and refining spatial governance, thereby improving the temporal foresight and spatial accuracy of intelligent statistics on people, real estate companies, and enterprises.

[0145] In one embodiment, such as Figure 2 As shown, this application also provides a smart statistical device for real estate enterprises based on big data. The device may include:

[0146] The graph construction module 701 is used to acquire multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments. It extracts unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data. It is also used to extract semantic features from the unstructured address description text using a natural language processing model, mapping each address text to a standard address vector of a unified dimension. Based on the similarity calculation results of the standard address vectors, it performs cross-source entity alignment on the static registration records to construct an initial person-real estate-enterprise association graph. It is also used to acquire real-time dynamic pedestrian flow trajectory data of the target area, and map the dynamic pedestrian flow trajectory data to the geospatial coordinate system corresponding to the initial person-real estate-enterprise association graph through spatial coordinate mapping rules to obtain the target person-real estate-enterprise association graph.

[0147] The risk reasoning module 702 is used to extract the number of people residing in each spatial grid unit of the housing entities and enterprise entities, the scale of enterprise registration and employment, and the capacity of housing registration from the target real estate enterprise association graph, and generate a spatial association feature set; simultaneously acquire the electricity frequency data and tax status data of each housing and enterprise entity in the target area, and generate a business attribute feature set; it is also used to input the spatial association feature set and the business attribute feature set into a pre-trained random forest classification model for reasoning operation, and output the vacancy risk probability and false registration risk probability of each housing and enterprise entity in the target area.

[0148] The incremental training module 703 is used to generate risk labeling results for each entity based on the probability of vacancy risk and the probability of false registration risk; to filter entities labeled as warning status according to the risk labeling results; to retrieve time series observation data within a preset time window; to construct time series training samples for incremental training of the time series prediction model using the time series observation data as input features and the risk labeling results as supervision labels; and to input the time series training samples into a pre-constructed long short-term memory network model, to perform incremental training and update the internal weight parameters of the model, thereby obtaining the updated long short-term memory network prediction model.

[0149] The distribution prediction module 704 is used to extract the historical time series state sequence corresponding to the target person-real estate enterprise association map, input the updated long short-term memory network prediction model, and obtain the spatial distribution prediction results of people and real estate enterprises in the target area within the future preset time period.

[0150] The aforementioned intelligent statistical device for people, real estate companies, and enterprises based on big data, through the collaborative operation of a graph construction module, a risk reasoning module, an incremental training module, and a distribution prediction module, completes the entire intelligent processing of multi-source data fusion, entity risk assessment, model adaptive optimization, and temporal and spatial prediction. Specifically, the graph construction module is responsible for the underlying data collection and graph construction. This module acquires multi-source heterogeneous government data from public security, housing and construction, and market supervision departments, separating unstructured address description text and static registration records of population household registration, property ownership, and enterprise business registration. It then uses a natural language processing model to extract address semantic features and perform standardized vectorization conversion, achieving entity alignment of cross-departmental static registration records based on address vector similarity, and constructing an initial people, real estate companies, and enterprise association graph. Simultaneously, this module accesses real-time dynamic pedestrian flow trajectory data of the target area, completing spatial registration and fusion of dynamic trajectory data and static graphs through unified spatial coordinate mapping, generating a target people, real estate companies, and enterprise association graph that combines static registration attributes with dynamic spatiotemporal attributes. The risk inference module relies on the target entity and real estate enterprise association graph to complete multi-dimensional feature construction and intelligent risk identification. This module uses spatial grids as statistical units to extract spatial indicators such as the number of people residing in houses and enterprises, the scale of registered employees, and the capacity of the buildings to construct a spatial association feature set. Simultaneously, it collects operational data such as the frequency of electricity use and tax status to construct a business attribute feature set. The two feature sets are then merged and input into a pre-trained random forest classification model to complete inference calculations, accurately outputting the vacancy risk probability and false registration risk probability for each house and enterprise entity. The incremental training module enables adaptive iterative optimization of the prediction model. This module completes full-domain entity risk labeling based on entity risk probabilities, filters entities in warning status, and retrieves time-series observation data within a preset time window. Using the time-series observation data as input features and the risk labeling results as supervision labels, it constructs time-series training samples. These samples are used to perform incremental training on the long short-term memory network model, continuously updating the model's internal weight parameters to obtain an updated long short-term memory network prediction model adapted to real-time scenario features. The distribution prediction module is responsible for completing the final spatiotemporal extrapolation and prediction. This module extracts the historical time-series state sequence of the target person-real estate enterprise association map, inputs it into the updated and optimized long short-term memory network prediction model, and outputs the spatial distribution prediction results of people and real estate enterprises in the target area within the future preset time period through temporal pattern extrapolation.

[0151] The modules in this embodiment are seamlessly integrated, and the overall technical solution possesses significant technical advantages and practical value. This device breaks down heterogeneous data barriers between multiple departments through a map construction module, and combines address vectorization matching with a dynamic spatiotemporal data fusion mechanism to solve the problems of data fragmentation, entity matching distortion, and inability to simultaneously consider static registration and dynamic operational status in traditional statistical methods. Through a risk reasoning module, spatial grid features and entity business operation features are integrated, and a dual-dimensional risk probability quantification output is achieved based on an ensemble learning model. This overcomes the limitations of traditional manual judgment and single-indicator assessment, improving the objectivity and accuracy of identifying risks such as vacant properties and false enterprise registration. An incremental training module enables dynamic self-optimization of the model, iteratively updating model parameters based on real-time business risk samples, effectively addressing the shortcomings of traditional fixed models, such as weak generalization ability and long-term accuracy decay. Finally, a distribution prediction module mines the temporal evolution patterns of entities, outputting refined spatial distribution prediction results. This achieves a full-process intelligent upgrade from static statistics and risk assessment to dynamic prediction, effectively improving the intelligent level of urban real estate governance and refined market entity supervision.

[0152] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0153] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the big data-based intelligent statistical method for real estate enterprises as described above.

[0154] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0155] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0156] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A big data-based intelligent statistical method for real estate enterprises, characterized in that: The method includes: Obtain multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments, and extract unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data; A natural language processing model is used to extract semantic features from the unstructured address text, mapping each address text to a standard address vector of a uniform dimension; based on the similarity calculation results of the standard address vectors, cross-source entity alignment is performed on the static registration records to construct an initial person-real estate enterprise association graph; Real-time dynamic pedestrian flow trajectory data of the target area is obtained, and the dynamic pedestrian flow trajectory data is mapped to the geospatial coordinate system corresponding to the initial human-real estate enterprise association map through spatial coordinate mapping rules to obtain the target human-real estate enterprise association map. Extract the number of people residing in each spatial grid unit of the housing entity and enterprise entity, the scale of enterprise registration and employment, and the capacity of housing registration from the target real estate enterprise association map to generate a spatial association feature set; simultaneously obtain the electricity frequency data and tax status data of each housing and enterprise entity in the target area to generate a business attribute feature set. The spatial association feature set and the business attribute feature set are input into a pre-trained random forest classification model for inference operations, and the vacancy risk probability and false registration risk probability of each house and enterprise entity in the target area are output. Based on the vacancy risk probability and the false registration risk probability, risk labeling results are generated for each entity; entities labeled as warning status are selected according to the risk labeling results, time series observation data within a preset time window is retrieved, and time series training samples for incremental training of the time series prediction model are constructed using the time series observation data as input features and the risk labeling results as supervision labels. The time-series training samples are input into a pre-built long short-term memory network model, incremental training is performed, and the internal weight parameters of the model are updated to obtain an updated long short-term memory network prediction model. Extract the historical time-series state sequence corresponding to the target real estate enterprise association map, input it into the updated long short-term memory network prediction model, and obtain the spatial distribution prediction result of real estate enterprises in the target area within a future preset time period.

2. The method according to claim 1, characterized in that, The initial real estate enterprise association graph was constructed through the following process: For the unstructured address representation text, a natural language processing model is used to extract semantic features and generate standard address vectors corresponding to each address text. The natural language processing model is used to identify multi-level administrative hierarchy elements and spatial location elements in address text, and then concatenates and fuses the two types of elements after vectorizing and encoding them respectively. Extract the corresponding spatial coordinate features and address hierarchy structure features from each of the standard address vectors; Based on the spatial coordinate features and address hierarchy features corresponding to different data source addresses, calculate the spatial topology matching degree between addresses; The spatial topology matching degree value is compared with a preset matching threshold; If the spatial topology matching degree value is greater than the preset matching threshold, it is determined that the addresses corresponding to different data sources point to the same geographic entity, and the corresponding population, housing, and enterprise entities are marked as target entity alignment nodes. Based on the alignment nodes of all the target entities, establish cross-source entity association relationships and generate an initial real estate enterprise association graph.

3. The method according to claim 2, characterized in that, The formula for calculating the spatial topology matching degree between addresses is: in, This represents the two entity addresses to be matched. With address Spatial topological matching degree, Represents the spatial coordinate similarity components. Address With address The spherical geographic distance between coordinate points is calculated based on latitude and longitude using the Haversine formula. This represents the distance attenuation bandwidth parameter. This indicates the address hierarchy matching degree. Indicates the total number of address levels. Indicates the first Hierarchical matching weights Indicates the first Hierarchical matching identifiers, Represents the spatial topology consistency check factor. This represents the spatial weight balance coefficient.

4. The method according to claim 1, characterized in that, The process of mapping the dynamic pedestrian flow trajectory data to the geospatial coordinate system corresponding to the initial person-real estate company association map using spatial coordinate mapping rules to obtain the target person-real estate company association map includes: The dynamic pedestrian flow trajectory data is preprocessed in time and space to extract a continuous trajectory coordinate sequence with timestamps; The dynamic pedestrian trajectory data includes timestamped sampling sequences of pedestrian locations, entry and exit records of area boundaries, spatial calibration data of sensing devices, and anonymous identification data of pedestrians. Density clustering algorithm is used to identify dwell points and calculate clustering degree of the trajectory coordinate sequence to obtain the dwell heat value corresponding to each spatial point. Based on the dwell heat value and spatial coordinates of each point, gridded trajectory dwell feature data with spatial location labels is generated; Obtain the unified geospatial coordinate system corresponding to the initial real estate enterprise association map, and divide it into several standard spatial grid units according to the preset grid granularity; The gridded trajectory dwell feature data is spatially matched with the standard spatial grid cell to construct a grid-entity spatial mapping matrix; The grid-entity spatial mapping matrix represents the spatial correspondence between each spatial grid and the building entity nodes and enterprise entity nodes in the map; Based on the grid-entity spatial mapping matrix, calculate the association edge weights between each house and enterprise entity node and the dynamic pedestrian flow data in the initial human-real estate-enterprise association graph; If the weight of the associated edge is greater than the preset associated threshold, the residence heat feature of the corresponding spatial grid is associated with the corresponding entity node of the initial real estate enterprise association graph and the node attributes are updated, and the target real estate enterprise association graph is obtained by fusion.

5. The method according to claim 1, characterized in that, The step of inputting the spatial association feature set and the business attribute feature set into a pre-trained random forest classification model for inference operations, and outputting the vacancy risk probability and false registration risk probability of each house and enterprise entity within the target area, includes: Based on the entity nodes and inter-node relationships in the target real estate enterprise association graph, a multi-dimensional attribute entity association network is constructed by integrating the spatial association feature set and the business attribute feature set. Extract the multidimensional feature vectors corresponding to each house and enterprise entity from the multidimensional attribute entity association network; The multidimensional feature vector is input into a pre-trained random forest classification model for inference operations, and a risk probability matrix is ​​output. The risk probability matrix includes the vacancy risk probability and the false registration risk probability for each house and business entity. The vacancy risk probability and the false registration risk probability of each entity in the risk probability matrix are compared with the corresponding preset risk thresholds. If any risk probability corresponding to any entity is greater than the corresponding preset risk threshold, the corresponding entity will be identified as a high-risk entity, and the vacancy risk probability and false registration risk probability corresponding to the high-risk entity will be output.

6. The method according to claim 1, characterized in that, The step of inputting the time-series training samples into a pre-constructed long short-term memory network model, performing incremental training and updating the model's internal weight parameters to obtain an updated long short-term memory network prediction model includes: Temporal feature sequences and corresponding supervision labels are extracted from the temporal training samples. The temporal feature sequences are input into a pre-constructed long short-term memory network model, and the corresponding hidden state vectors are output. The hidden state vector is mapped to generate a prediction result, and the prediction result is compared with the corresponding supervision label to calculate the prediction deviation; Based on the prediction deviation, backpropagation of the error is performed to generate the weight increment gradient. The initial weight matrix of the model is then adjusted according to the weight increment gradient to obtain the updated weight matrix. The forward and backward propagation processes are executed iteratively, and it is determined whether the current iteration meets the preset convergence condition. If the current iteration satisfies the preset convergence condition, then the updated Long Short-Term Memory network prediction model is generated using the currently updated weight matrix.

7. The method according to claim 1, characterized in that, The step of extracting the historical time-series state sequence corresponding to the target real estate enterprise association map, inputting it into the updated Long Short-Term Memory network prediction model, and obtaining the spatial distribution prediction results of real estate enterprises in the target area within a preset future time period includes: Obtain the multi-dimensional historical time-series state sequence corresponding to the housing entity nodes and enterprise entity nodes in the target real estate enterprise association graph; The updated long short-term memory network model is used to perform forward propagation on the temporal state sequence, and the state transition probability of each entity node is output. By combining the pre-divided standard spatial grid units of the target area, the state transition probabilities of each entity node are mapped to the corresponding spatial grid to construct a spatial distribution matrix of people, real estate and enterprises; Iteratively update the spatial distribution matrix and calculate the matrix convergence degree corresponding to adjacent iteration rounds; If the convergence of the matrix is ​​greater than the preset convergence threshold, then the current spatial distribution matrix is ​​determined to be the spatial distribution prediction result of the real estate enterprises in the target area.

8. A smart statistical device for real estate enterprises based on big data, characterized in that: The device includes: The graph construction module is used to acquire multi-source heterogeneous government data from public security departments, housing and construction departments, and market supervision departments; extract unstructured address description text and static registration records corresponding to population household registration, housing property rights, and enterprise business registration from the multi-source heterogeneous government data; it is also used to extract semantic features from the unstructured address description text using a natural language processing model, mapping each address text to a standard address vector of a unified dimension; based on the similarity calculation results of the standard address vectors, cross-source entity alignment is performed on the static registration records to construct an initial person-real estate-enterprise association graph; it is also used to acquire real-time dynamic pedestrian flow trajectory data of the target area, and map the dynamic pedestrian flow trajectory data to the geospatial coordinate system corresponding to the initial person-real estate-enterprise association graph through spatial coordinate mapping rules to obtain the target person-real estate-enterprise association graph; The risk reasoning module is used to extract the number of people residing in each spatial grid unit of the housing entities and enterprise entities, the scale of enterprise registration and employment, and the capacity of housing registration from the target real estate enterprise association graph, and generate a spatial association feature set; simultaneously acquire the electricity frequency data and tax status data corresponding to each housing and enterprise entity in the target area, and generate a business attribute feature set; it is also used to input the spatial association feature set and the business attribute feature set into a pre-trained random forest classification model for reasoning operations, and output the vacancy risk probability and false registration risk probability corresponding to each housing and enterprise entity in the target area; The incremental training module is used to generate risk labeling results for each entity based on the vacancy risk probability and the false registration risk probability; to filter entities labeled as warning status according to the risk labeling results; to retrieve time-series observation data within a preset time window; to construct time-series training samples for incremental training of the time-series prediction model using the time-series observation data as input features and the risk labeling results as supervision labels; and to input the time-series training samples into a pre-constructed long short-term memory network model, perform incremental training, and update the internal weight parameters of the model to obtain an updated long short-term memory network prediction model. The distribution prediction module is used to extract the historical time-series state sequence corresponding to the target person-real estate enterprise association map, input it into the updated long short-term memory network prediction model, and obtain the spatial distribution prediction result of people and real estate enterprises in the target area within a preset future time period.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images