Site selection recommendation method based on custom size block classification
By using custom-sized block classification and machine learning algorithms, a customizable spatial grid system and anchor point distance attenuation model are constructed. This solves the problems of insufficient resolution and inaccurate anchor point impact assessment caused by fixed grid size in existing commercial site selection methods, and realizes scientific and efficient prediction of commercial site selection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-14
AI Technical Summary
Existing commercial site selection methods rely on human experience and fixed-size grid analysis, resulting in insufficient resolution in high-density areas and sparse data in low-density areas. This makes it impossible to accurately assess the impact of anchor facilities and lacks consideration of the differences in anchor facility types and their temporal evolution characteristics, thus affecting the scientific nature of site selection decisions.
By employing a custom-sized block classification method, and through quantitative modeling of POI distribution patterns and historical extrapolation analysis, a customizable spatial grid system and anchor point distance attenuation model are constructed. Combined with machine learning algorithms, commercial site selection prediction and recommendation are performed.
It enables scientific decision-making in business site selection, improves the timeliness and accuracy of site selection, can flexibly adjust grid size, capture the influence of spatial anchor points and temporal evolution trends, and improves the reliability of prediction models and recommendation accuracy.
Smart Images
Figure CN121860693A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of geographic information systems and intelligent spatial data analysis, and in particular to a site selection recommendation method based on custom-sized block classification. Background Technology
[0002] The suitability of a business location directly impacts foot traffic, revenue, and return on investment. Traditional business location selection methods primarily rely on human experience and simple geographic information heatmap analysis. However, business location selection is also influenced by factors such as policy, foot traffic, and demand environment. Existing geographic information system tools typically use fixed-size grids as spatial analysis units. However, fixed sizes can affect the flexibility of judgment: grids that are too large will result in insufficient spatial resolution in high-density areas; grids that are too small will lead to extremely sparse data and unstable statistical characteristics in low-density areas, making it difficult to balance spatial accuracy and data validity.
[0003] According to theories such as central place theory and spatial interaction theory, anchor facilities such as schools, shopping malls, and transportation hubs have a significant attraction effect on surrounding businesses. However, existing methods mostly remain at the level of qualitative analysis or simple distance counting, such as only counting the number of schools within 2km, ignoring the impact of distance differences, lacking differentiated modeling of the influence intensity of different types of anchors (such as universities and primary schools, shopping centers and convenience stores), and failing to fully consider the distance decay effect revealed by the first law of geography. This makes it impossible to accurately assess the actual contribution of anchor facilities, thus affecting the scientific nature of site selection decisions. In addition, the business environment has obvious temporal evolution characteristics (e.g., the continuous increase in the number of restaurants in a certain area over the past five years reflects the trend of the vitality of the catering industry in that area). Existing methods are mainly based on single time-section data analysis, and the use of historical data is limited to simple annual comparisons. Summary of the Invention
[0004] In view of the aforementioned shortcomings of existing technologies, this invention provides a site selection recommendation method based on custom-sized block classification. Through quantitative modeling of POI distribution patterns and historical impact analysis, it predicts and recommends commercial sites for specified building categories. This method constructs a customizable spatial grid system and a distance attenuation impact model for anchor buildings, extracts evolutionary features from historical data, and then combines these with machine learning algorithms for prediction and recommendation, supporting scientific decision-making in commercial site selection.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A site selection recommendation method based on custom-sized block classification includes the following steps: S1. Acquire and preprocess POI data for the target geographic area to generate a structured dataset; S2. Divide the target geographic area into multiple rectangular grid units; S3. Extract anchor buildings and facilities that have an impact on the target business type from the POI data according to preset rules; quantify the spatial influence intensity of anchor buildings and facilities based on the Havelsing distance between rectangular grid cells and anchor buildings and facilities, and construct a distance attenuation model; S4. Based on the spatial influence intensity and distance attenuation model of anchor point buildings and facilities, and combined with the historical inference features extracted from the structured dataset, construct a comprehensive feature set; construct a prediction and recommendation model and train it on the comprehensive feature set to generate predictions of the target business format distribution; S5. Based on the distribution of target business types, construct the final recommendation score of rectangular grid units, output the site selection recommendation list and generate a visual decision support report.
[0006] Preferably, S3 includes: S31. Identify and extract anchor building facilities that have an impact on the target business from the POI data according to preset rules; S32. Calculate the Havesing distance between each rectangular grid cell and the anchor point building: S33. Construct an anchor point influence model based on Haversin distance to quantify the spatial influence intensity of anchor point buildings; S34. Using nonlinear regression, fit the parameters of the anchor point influence model from the structured dataset, and construct a distance decay model based on the anchor point influence model.
[0007] Preferably, S31 includes further subdividing the anchor point building facilities into different types; for each anchor point building facility in a rectangular grid cell, information including its center coordinates, area, and land use is stored.
[0008] Preferably, S33 includes quantifying the spatial influence intensity of the anchorage facility using a function that includes negative exponential decay and Gaussian decay, based on the first law of geography.
[0009] Preferably, S34 includes: In the distance attenuation model, rectangular mesh cells are defined. In anchor type The overall influence in China is as follows:
[0010] in, Belongs to type The set of all anchor point buildings; sum and iterate through each anchor point building. , For type The initial impact strength of anchorage structures. rectangular grid cells With anchor point building facilities Geographical distance between them For type Distance attenuation coefficient of anchor point building facilities, The fundamental impact that remains even at long distances.
[0011] Preferably, S4 includes: Anchor point influence features are extracted based on the spatial influence intensity and distance attenuation model of anchor point buildings; historical extrapolation features, including historical averages and linear trend coefficients, are extracted from multi-year structured datasets; and a comprehensive feature set including basic statistical features, anchor point influence features, and historical extrapolation features is constructed.
[0012] As a preferred approach, constructing a predictive recommendation model includes: A modal feature is extracted from the comprehensive feature set and a scalar prediction is output. A gradient boosting regression model is used as the predictor, and multiple decision trees are trained in sequence. Each new tree is fitted with the prediction residual of the previous gradient boosting regression model. The prediction of the target business distribution is achieved by combining multiple decision trees additively, including the number or potential value of buildings of a specified category in the rectangular grid cell.
[0013] As a preferred option, S4 also includes: By introducing a temporal smoothing constraint term, a loss function is constructed to train the prediction and recommendation model:
[0014] in, For loss function, For the number of samples, For rectangular grid cell indexing, rectangular grid cells The actual number of buildings in the current year, rectangular grid cells The observation value at the previous moment, It is a smoothing weight.
[0015] As a preferred option, the final recommended score for rectangular grid cells is defined as follows:
[0016] in, This represents the final recommended score for the rectangular mesh cell g. To predict potential weights, , … The weights of various anchor points are affected. For the trend term weight, As a weight for risky projects, For standardized predictions, , , The comprehensive influence of anchor points such as shopping malls, schools, and other commercial buildings on the rectangular grid cell g. As a historical trend characteristic, To standardize risk or cost indicators.
[0017] As a preferred approach, decision support reports are generated using visualization techniques such as functional zoning diagrams, accuracy assessment diagrams, and anchor point relationship diagrams.
[0018] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. Unlike traditional technologies that rely on manual experience, geographic information thermal flow, or single attenuation model assumptions, this invention employs a multi-model adaptive fitting distance attenuation analysis technique. This allows the system to automatically select the optimal extended attenuation model based on actual data characteristics, avoiding the fitting bias problem caused by the mismatch between model assumptions and actual data distribution in traditional methods.
[0019] 2. Unlike traditional technologies that rely on static analysis at a single point in time, short-term dynamic analysis, and prediction based on fixed-size blocks or administrative regions, this invention employs a spatiotemporal feature-based approach that integrates historical data from multiple years within customizable functional blocks. This approach makes restaurant distribution analysis and site selection recommendations more demand-oriented, and the prediction model can simultaneously capture the influence of spatial anchor points and temporal evolution trends, thereby improving the timeliness and reliability of restaurant site selection potential prediction.
[0020] 3. This invention enables parameterized configuration of the grid size, allowing for flexible adjustment based on different city sizes and analytical accuracy requirements, overcoming the limitations of traditional fixed grids. Simultaneously, it establishes a quantitative model of the influence of anchorage facilities based on the distance decay law, with a model fit goodness R... 2 It outperforms traditional linear models and offers strong interpretability. Furthermore, it proposes "historical inference features" to characterize their contribution to prediction, improving the accuracy of prediction and recommendation. In addition, by integrating geographical theory with machine learning methods, the model demonstrates high recommendation accuracy in real-world scenarios, consistent with factual logic. Attached Figure Description
[0021] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a diagram of the rectangular grid cell processing framework of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the distance attenuation model fitting in Embodiment 2 of the present invention; Figure 4 This is a graph showing the trend of the number of restaurants around the anchor point from 2018 to 2023 in Embodiment 2 of the present invention. Figure 5 This is a graph showing the trend of the average number of restaurants around the anchor point from 2018 to 2023 in Embodiment 2 of the present invention. Figure 6 This is a graph showing the prediction of 2024 data based on data from 2018 to 2023 in Embodiment 2 of the present invention. Figure 7 Based on data from 2018 to 2023 in Embodiment 2 of the present invention, the top 5 recommended restaurant locations for 2024 are predicted. Figure 8 This is a schematic diagram of the top 5 predicted blocks and the top 10 actual growth blocks in Embodiment 2 of the present invention. Detailed Implementation
[0022] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.
[0023] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.
[0024] Example 1: like Figure 1 The site selection recommendation method based on custom-sized block classification, as shown, includes the following steps: S1. Acquire and preprocess POI data for the target geographic area to generate a structured dataset; The process involves acquiring Point of Interest (POI) data for a target geographic area over several consecutive years, then performing encoding and identification, outlier coordinate removal, coordinate transformation, and category standardization operations on the raw POI files to generate a structured dataset that can be directly used in subsequent processes. The POI data includes information such as location, category, attribute, and time.
[0025] like Figure 2 The rectangular mesh element processing content in steps S2-S4 shown below: S2. Divide the target geographic area into multiple rectangular grid units; Based on the user-specified rectangular grid cell size parameters, the geographical area is divided into m×n rectangular grid cells. The latitude and longitude mapping formula for each grid cell is as follows:
[0026] in and The calculated row and column indices of the rectangular grid cells. , Represented as the longitude and latitude of the POI to be mapped. , The step size of the rectangular grid cells in the longitude and latitude directions. and These represent the minimum longitude and minimum latitude within the mapped region.
[0027] S3. Extract anchor buildings and facilities that have an impact on the target business type from the POI data according to preset rules; quantify the spatial influence intensity of anchor buildings and facilities based on the Havelsing distance between rectangular grid cells and anchor buildings and facilities, and construct a distance attenuation model; S31. From massive POI data, identify and extract anchor buildings and facilities that significantly impact the target business type, such as schools, shopping malls, transportation hubs, and hospitals, according to preset rules, and further subdivide their types. For example, school anchors can be subdivided into universities, colleges, high schools, junior high schools, primary schools, and kindergartens; education anchors can be subdivided into higher education, basic education, and early childhood education institutions; commercial anchors can be divided into large complexes, community businesses, and convenience stores; and transportation anchors can be divided into rail hubs and ground bus stations. Based on this subdivision, the filtering rules can be quickly adjusted for different city or industry scenarios. For each rectangular grid cell, store its center coordinates, area, and land use information, and ensure that data from different years can be aligned on the same spatial benchmark, laying the foundation for subsequent quantitative analysis.
[0028] S32. Calculate the Havesing distance between each rectangular grid cell and the anchor point building:
[0029] in, rectangular grid cells and anchorage building facilities The distance between them, with Earth as the reference point, Represented as latitude, It is expressed as longitude.
[0030] S33. Modeling the impact of anchor points based on Haversin distance: According to the first law of geography, functions such as negative exponential decay and Gaussian decay are used to quantify the spatial impact intensity of anchor point facilities, and an anchor point impact model is constructed: ;
[0031] in, The distance is The influence value of rectangular grid cells, The initial gain intensity provided when the distance approaches 0. This is the distance attenuation coefficient (the larger the value, the faster the effect attenuates with distance). It is the distance between the anchor point and the rectangular grid cell to be evaluated. For distant background values (specified as type building density or potential without anchor point influence), The feature distance is denoted as .
[0032] S34. By using nonlinear regression, model parameters are fitted from structured datasets accumulated over many years, thereby accurately depicting the law of influence decay with distance.
[0033] For regions where anchor points change significantly, the correlation between the anchor point increment and the target POI increment is evaluated, and a feature vector is generated for each rectangular grid cell:
[0034] in, It is a rectangular grid unit. In anchor type China's overall influence Belongs to type The set of all anchor points, summed and iterated through each anchor point. , For type The initial influence strength of the anchor point rectangular grid cells With anchor point Geographical distance between them For type Anchor point distance attenuation coefficient, The fundamental impact that remains even at long distances.
[0035] S4. Based on the spatial influence intensity and distance attenuation model of anchor point buildings and facilities, and combined with the historical extrapolation features extracted from structured data, construct a comprehensive feature set; construct a prediction and recommendation model and train it on the comprehensive feature set to generate predictions of the target business format distribution; S41. Construct a comprehensive feature set; based on spatial relationships, anchor point influence patterns, and historical evolution trends, construct a multi-dimensional feature vector for each rectangular grid cell, including spatial statistics, anchor point influence, and temporal evolution features; on the basis of basic statistical features and anchor point influence features, systematically extract historical extrapolation features from multi-year structured data, including historical averages, linear trend coefficients, etc., and combine the above spatial features, anchor point influence features, and historical extrapolation features to form a comprehensive feature set. This organically integrates the evolutionary patterns in the time dimension with the anchor point influence patterns in the spatial dimension; S42. Construct a predictive recommendation model: Based on this, gradient boosting is used to train on a comprehensive feature set, and the distribution of the target business type is predicted by additive combination of multiple decision trees:
[0036] in, It is responsible for extracting features from a specific modality (anchor point, history, regional attributes, risk, etc.) and outputting scalar predictions; The feature vectors of all time steps and all rectangular grid cells were collected; In order to be in Time The feature vector of a rectangular grid cell For historical time steps, For single-time-step feature dimensions, This represents the total number of rectangular grid cells.
[0037] A gradient boosting regression model is used as the predictor, and multiple decision trees are trained sequentially. Each new tree fits the prediction residual of the previous model, gradually reducing the overall prediction error.
[0038] in, For model predictions about rectangular grid cells The number or potential value of buildings of a specified category; The number of decision trees participating in the integration; For input grid features The Sub-model, This is the learning rate.
[0039] S43. Model training aims to minimize prediction error. The basic loss function uses mean squared error to measure the deviation between predicted and true values. To further enhance the continuity and stability of prediction results over time, a temporal smoothing constraint term is introduced, constructing a composite loss function:
[0040] in, For the overall loss function, For the number of samples, For rectangular grid cell indexing, rectangular grid cells The actual number of buildings in the current year, rectangular grid cells The observation value at the previous moment, It is a smoothing weight that controls the strength of the smoothing term and prevents unreasonable and drastic fluctuations.
[0041] S5. Based on the distribution of target business types, construct the final recommendation score of rectangular grid units, output the site selection recommendation list and generate a visual decision support report.
[0042] The system forecasts business potential for the target year, outputs a list of recommended locations, and performs cross-validation. Finally, it generates an intuitive and easy-to-understand decision support report using various visualization techniques such as functional zoning maps, accuracy assessment maps, and anchor point relationship diagrams.
[0043] in, This represents the final recommended score for the rectangular mesh cell g. To predict potential weights, , … The weights of various anchor points are affected. For the trend term weight, As a weight for risky projects, For standardized predictions, , , The standardization influence of anchor points such as shopping malls, schools, and other commercial buildings on the rectangular grid cell g. As a historical trend characteristic, To standardize risk or cost indicators.
[0044] Example 2: Application scenario: A chain restaurant company plans to open 3-5 new stores in Qiantang District, Hangzhou City, and needs to select locations scientifically to reduce investment risks.
[0045] This embodiment uses open data POIs to obtain POI data for Qiantang District, Hangzhou City from 2018 to 2024. The range selected is the coordinates (120.290699, 30.227846) to (120.706377, 30.396451). The data includes name, coordinates, address information, category, etc., with 30,000 to 36,000 data entries per year.
[0046] Phase 1: Pattern Extraction 1. Construct a 250m grid system: Based on the data coverage area, without distinguishing the size of blocks under a specified area, a uniform 250-meter grid unit is established, generating a total of 2,862 grid cells. The entire grid is created using coordinate boundaries and unit steps. Each POI undergoes data cleaning and string matching processing, and its grid index is calculated. POI data is read year by year, and each year's POI is mapped to a unified grid. After grid partitioning, it can be attached and displayed on the map.
[0047] 2. Anchor Point POI Extraction and Segmentation: Extract 23 school-related keywords, filter address information, and segment school-related POIs into types such as universities, colleges, high schools, junior high schools, primary schools, and kindergartens. Shopping mall POIs are categorized into types such as large complexes, community businesses, and convenience stores.
[0048] 3. Establish a distance attenuation model: Summarize samples from all years to form school anchor point influence data. Simultaneously, statistically analyze the mean, median, standard deviation, and sample size for each item based on arithmetic distance. Using the POI information G(x) and the preset anchor point information A(A_type) as input, iterate through grid combinations containing restaurants, and calculate the attenuation using Havelsing distance aggregation. Figure 3 As shown, the school anchor point is obtained. (school) The optimal parameters returned are (2.966, 1.131, 8.458), which corresponds to the school impact model. Goodness of fit R 2 =0.8201; Similarly, the shopping mall impact model is obtained: Goodness of fit R 2 =0.7823, the two models are used as anchor point influence quantification data, and the comprehensive influence of the two types of anchor points on each grid is aggregated.
[0049] 4. Analysis of temporal evolution patterns: Extracting temporal features from multi-year historical data. For example... Figure 4 As shown, the annual trend of the number of restaurants around each anchor point (taking schools as an example) is illustrated; Figure 5 As shown, the average trend of the number of restaurants around all anchor points is displayed. Further comparative analysis is performed by dividing the grid into two categories based on whether it includes schools. Figure 6 As shown, the grids are divided into two categories: "grids with schools" and "grids without schools". The food service density in grids with schools is 3.7 times that in grids without schools, and both types of areas show an increasing trend: between 2018 and 2023, the food service density in grids with schools increased from 6.07 to 8.77 (an increase of 44.5%), while that in grids without schools increased from 0.30 to 0.46 (an increase of 53.8%). These statistics, along with the anchor point strength, serve as the historical extrapolation input for Phase Two.
[0050] Phase Two: Site Selection Forecast and Recommendations for 2024 1. Model Training: Spatial statistical features, anchor point influence features, and historical extrapolation features are standardized and then fused to form a comprehensive feature vector. A total of 16,536 grid samples from 2018-2023 are used, each containing spatial features of the location, anchor point strength, and temporal evolution trend. Gradient boosting ensemble training is employed, with temporal smoothing constraints introduced. The training and validation sets are divided into 80%:20% sets. The model achieves R on the validation set. 2It reached 0.7014, indicating good generalization ability.
[0051] 2. Training Strategy: The comprehensive location score for each grid cell is calculated based on the final recommendation score formula. The configurations are: η = 0.15, learning rate = 1e-3, batch size = 512, and convergence after 60 epochs. In this experiment... , , , , The values are 0.42, 0.22, 0.15, 0.11, and 0.10, respectively. , This corresponds to historical trends and risks.
[0052] 3. Verification results: such as Figure 7 and Figure 8 As shown, among the top 5 recommendations, the actual number of restaurants in 4 blocks is relatively high (90-222), and the recommendation accuracy rate reaches 80%. The characteristics of the top 5 blocks are an average of 7.2 schools, an average of 31.6 shops, and a historical average of 85.0 restaurants. The recommendation accuracy rate of 80% is within the effective influence radius of the university, which is consistent with the anchor point influence theory and historical extrapolation patterns.
[0053] The method patent of this invention has the following contributions: This patent addresses the issues of insufficient quantification of the influence mechanism of fixed spatial analysis units and anchor points. To resolve the problems of unstable statistical characteristics and difficulty in balancing spatial accuracy and data validity caused by using fixed-size grids as spatial analysis units, this patent partitions fixed blocks into different sizes (e.g., 250*250m and 1000*1000m grids) during pre-input, and separately statistically analyzes dense and sparse areas. Taking restaurant site selection as an example, for high-influence anchor points such as schools and shopping malls, further subdivision is performed, classifying schools into universities, middle schools, and primary schools, and shopping malls into large shopping malls and convenience stores. Based on the first law of geography and models such as negative exponential decay, power-law decay, and Gaussian decay, a nonlinear regression method is used to select an applicable extended decay function to quantify the spatial influence intensity of anchor point buildings. Model parameters are fitted from historical data to accurately characterize the distance-based influence decay law.
[0054] This invention addresses the problem of insufficient utilization and integration of historical time-series information. Data analysis based on a single time segment limits the use of historical data to simple year-to-year comparisons, resulting in insufficient information utilization. This patent, based on the distribution of similar blocks with custom dimensions, specifies the coordinates of similar buildings and their historical change trends, integrating the historical evolution patterns of anchor buildings as predictive features into the model.
Claims
1. A site selection recommendation method based on custom-sized block classification, characterized in that, Includes the following steps: S1. Acquire and preprocess POI data for the target geographic area to generate a structured dataset; S2. Divide the target geographic area into multiple rectangular grid units; S3. Extract anchor buildings and facilities that have an impact on the target business type from the POI data according to preset rules; quantify the spatial influence intensity of anchor buildings and facilities based on the Havelsing distance between rectangular grid cells and anchor buildings and facilities, and construct a distance attenuation model; S4. Based on the spatial influence intensity and distance attenuation model of anchor point buildings and facilities, and combined with the historical inference features extracted from the structured dataset, construct a comprehensive feature set; construct a prediction and recommendation model and train it on the comprehensive feature set to generate predictions of the target business format distribution; S5. Based on the distribution of target business types, construct the final recommendation score of rectangular grid units, output the site selection recommendation list and generate a visual decision support report.
2. The site selection recommendation method based on custom-sized block classification according to claim 1, characterized in that, S3 include: S31. Identify and extract anchor building facilities that have an impact on the target business from the POI data according to preset rules; S32. Calculate the Havesing distance between each rectangular grid cell and the anchor point building: S33. Construct an anchor point influence model based on Haversin distance to quantify the spatial influence intensity of anchor point buildings; S34. Using nonlinear regression, fit the parameters of the anchor point influence model from the structured dataset, and construct a distance decay model based on the anchor point influence model.
3. The site selection recommendation method based on custom-sized block classification according to claim 2, characterized in that, S31 includes further subdivision of anchor point building facilities; for each anchor point building facility in a rectangular grid cell, information including its center coordinates, area, and land use is stored.
4. The site selection recommendation method based on custom-sized block classification according to claim 2, characterized in that, S33 includes quantifying the spatial influence intensity of anchorage facilities using functions including negative exponential decay and Gaussian decay, based on the first law of geography.
5. The site selection recommendation method based on custom-sized block classification according to claim 2, characterized in that, S34 includes: In the distance attenuation model, rectangular mesh cells are defined. In anchor type The overall influence in China is as follows: ,in, Belongs to type The set of all anchor point buildings; sum and iterate through each anchor point building. , For type The initial impact strength of anchorage structures. rectangular grid cells With anchor point building facilities Geographical distance between them For type Distance attenuation coefficient of anchor point building facilities, The fundamental impact that remains even at long distances.
6. The site selection recommendation method based on custom-sized block classification according to claim 1, characterized in that, S4 include: The influence characteristics of anchor points are extracted based on the spatial influence intensity and distance attenuation model of anchor point buildings; Historical extrapolation features, including historical averages and linear trend coefficients, are extracted from multi-year structured datasets; a comprehensive feature set, including basic statistical features, anchor point influence features, and historical extrapolation features, is constructed.
7. The site selection recommendation method based on custom-sized block classification according to claim 1, characterized in that, Building a predictive recommendation model includes: A modal feature is extracted from the comprehensive feature set and a scalar prediction is output. A gradient boosting regression model is used as the predictor, and multiple decision trees are trained in sequence. Each new tree is fitted with the prediction residual of the previous gradient boosting regression model. The prediction of the target business distribution is achieved by combining multiple decision trees additively, including the number or potential value of buildings of a specified category in the rectangular grid cell.
8. The site selection recommendation method based on custom-sized block classification according to claim 7, characterized in that, S4 also includes: By introducing a temporal smoothing constraint term, a loss function is constructed to train the prediction and recommendation model: ， in, For loss function, For the number of samples, For rectangular grid cell indexing, rectangular grid cells The actual number of buildings in the current year, rectangular grid cells The observation value at the previous moment, It is a smoothing weight.
9. The site selection recommendation method based on custom-sized block classification according to claim 1, characterized in that, The final recommended score for defining rectangular mesh cells: ,in, This represents the final recommended score for the rectangular mesh cell g. To predict potential weights, , … The weights of various anchor points are affected. For the trend term weight, As a weight for risky projects, For standardized predictions, , , The comprehensive influence of anchor points such as shopping malls, schools, and other commercial buildings on the rectangular grid cell g. As a historical trend characteristic, To standardize risk or cost indicators.
10. The site selection recommendation method based on custom-sized block classification according to claim 1, characterized in that, Decision support reports are generated using visualization techniques such as functional zoning diagrams, accuracy assessment diagrams, and anchor point relationship diagrams.