An intelligent commercial site selection method, device and medium based on big data analysis

Abstract

The application discloses an intelligent commercial site selection method and device based on big data analysis and a medium, relates to the technical field of big data analysis, and comprises the following steps: collecting a commercial site selection coupling data set, performing feature alignment, and forming a standard space-time data stream; adopting an Apache Flink stream processing engine to perform complex event mode recognition on the standard space-time data stream, and forming a reference point address digital portrait; inputting the reference point address digital portrait and the commercial site selection coupling data set into a deep learning site selection model; performing time series modeling and space encoding through a feature coding layer; performing space clustering analysis and similarity calculation through a similarity measurement layer; and outputting a site selection similarity cloud map. Through the Apache Flink stream processing engine and the deep learning site selection model, the application enhances the fusion precision between multi-source data and significantly improves the dynamic prediction capability of commercial site selection.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to an intelligent business site selection method, device, and medium based on big data analytics. Background Technology

[0002] With the rapid pace of urbanization, commercial site selection, as a core element in industries such as retail, catering, and services, directly impacts operational efficiency and market competitiveness through its scientific rigor and accuracy. Traditional methods relied heavily on experience and manual surveys. In recent years, big data analytics, a technology capable of integrating multi-source heterogeneous data and uncovering relationships, has been widely applied to support commercial site selection decisions. Furthermore, existing GIS geographic information analysis methods and machine learning models can also, to some extent, achieve business district identification and potential demand prediction, enhancing the intelligence level of commercial site selection.

[0003] However, existing technologies still face several bottlenecks in practical applications of commercial site selection. First, they fail to adequately address the spatiotemporal heterogeneity of multi-source data and lack effective feature alignment mechanisms, making it difficult to achieve precise coupling between multi-source data and resulting in insufficient representativeness of the final site selection reference system. Second, existing machine learning models, in the process of similarity matching, are limited to static feature comparison and fail to fully explore the deep correlation between spatial clustering and temporal evolution, thus restricting the dynamic predictive capabilities of commercial site selection. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent business site selection method based on big data analysis to solve the problems of insufficient representativeness of the site selection reference system and low predictive ability of business site selection.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent commercial site selection method based on big data analysis, which includes: collecting a commercial site selection coupled dataset and performing feature alignment to form a standard spatiotemporal data stream; and using the Apache Flink stream processing engine to perform complex event pattern recognition on the standard spatiotemporal data stream to form a digital profile of the reference site.

[0008] The reference site digital profile and commercial site selection coupled dataset are input into the deep learning site selection model. The feature encoding layer performs temporal modeling and spatial encoding, and the similarity measurement layer performs spatial clustering analysis and similarity calculation, outputting a site selection similarity cloud map.

[0009] The location is determined by the location similarity cloud map, forming a set of candidate addresses. The TOPSIS algorithm is used to optimize the candidate address set for multiple objectives, and the spatiotemporal decay factor is used to perform collaborative filtering compensation to generate a comprehensive score for the candidate addresses.

[0010] Based on the comprehensive scores of the candidate addresses, the candidate address set is classified, sorted, and one-hot encoded to obtain a priority addressing list. Cross-validation is then performed on the priority addressing list to output the addressing decision package.

[0011] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the business site selection coupled dataset includes a three-dimensional site selection map, historical store operation data, user profile types, and business district activity data.

[0012] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the step of forming a digital profile of the reference site specifically includes the following steps.

[0013] The 3D site selection map, historical store operation data, user profile types and business district activity data are unified with timestamps and spatially registered to form a standard spatiotemporal data stream.

[0014] A real-time data processing pipeline is built based on the Apache Flink stream processing engine. Standard spatiotemporal data streams are input into the real-time data processing pipeline for complex event pattern recognition, and digital profiles of reference points are output.

[0015] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the deep learning site selection model is specifically constructed as follows:

[0016] A feature encoding layer is constructed using a spatiotemporal convolutional network, and a similarity measurement layer is constructed using a graph attention network.

[0017] A deep learning location selection model is constructed by performing gradient backpropagation and cross-layer stacking on the feature encoding layer and the similarity measurement layer.

[0018] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the output of the site selection similarity cloud map specifically includes the following steps.

[0019] The reference site digital profile and the commercial site selection coupled dataset are input into the deep learning site selection model. The feature encoding layer performs temporal modeling and spatial encoding through a spatiotemporal convolutional neural network to generate a dynamic indicator temporal matrix.

[0020] The similarity measurement layer performs spatial clustering analysis using a spectral clustering algorithm and calculates similarity using a cosine similarity formula to generate positional similarity values.

[0021] Spatial interpolation and heatmap rendering are performed on the time series matrix of dynamic indicators and the location similarity values ​​to output a site selection similarity cloud map.

[0022] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the generation of a comprehensive score for candidate addresses specifically includes the following steps:

[0023] Based on the location similarity cloud map, a region growing algorithm is used to detect hotspots and locate regions to form a set of candidate addresses.

[0024] The TOPSIS algorithm is used to perform multi-objective optimization on the candidate address set to generate an initial optimization score. The spatiotemporal decay factor is used to perform decay correction and collaborative filtering compensation on the initial optimization score to generate a comprehensive score for the candidate addresses.

[0025] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the step of obtaining the priority site selection list specifically includes the following steps:

[0026] Based on the comprehensive scores of the candidate addresses, the candidate address set is classified and sorted to generate priority address categories;

[0027] One-hot encoding is performed on priority address categories to obtain a priority address selection list.

[0028] As a preferred embodiment of the intelligent business site selection method based on big data analysis described in this invention, the output site selection decision package specifically refers to performing confidence cross-validation on the priority site selection list using the Bootstrap self-sampling method to output the site selection decision package.

[0029] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the intelligent business site selection method based on big data analysis as described in the first aspect of the present invention.

[0030] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent business site selection method based on big data analysis as described in the first aspect of the present invention.

[0031] The beneficial effects of this invention are as follows: By constructing a standard spatiotemporal data stream and performing feature alignment, the coupling problem caused by the spatiotemporal heterogeneity of multi-source data is effectively solved, significantly improving the fusion accuracy between multi-source data, thereby enhancing the representativeness and reliability of the digital profile of reference locations. Simultaneously, a deep learning-based location selection model is constructed, which not only achieves the collaborative processing of spatial encoding and temporal modeling, but also fully explores the deep correlation between spatial distribution and temporal evolution, significantly improving the dynamic prediction capability of commercial location selection. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of an intelligent business site selection method based on big data analytics.

[0034] Figure 2 A flowchart generated for a standard spatiotemporal data stream.

[0035] Figure 3 The flowchart for generating a digital profile for a reference address.

[0036] Figure 4 A flowchart generated for the location decision package. Detailed Implementation

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0039] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0040] Reference Figures 1-4 This is one embodiment of the present invention, which provides an intelligent business site selection method based on big data analysis, including the following steps:

[0041] S1. Collect commercial site selection coupled datasets and perform feature alignment to form a standard spatiotemporal data stream; use the Apache Flink stream processing engine to perform intelligent matching on the standard spatiotemporal data stream to form a digital profile of the reference site.

[0042] S1.1 Collect commercial site selection coupled dataset, which includes 3D site selection map, historical store operation data, user profile types and business district activity data.

[0043] The 3D site selection map reflects geographical information such as building layout and transportation network, and is collected through remote sensing imagery and laser point cloud; historical store operation data includes operational data such as store sales, customer traffic and operating hours, and is collected using enterprise ERP software (such as SAP or Oracle); user profile types characterize consumer age distribution, spending power and behavioral preferences, and are collected using IoT sensors; business district activity data reflects the dynamic changes in population density, consumption activity and competitive landscape within the area, and is collected using third-party big data software (such as UnionPay Smart Business District).

[0044] S1.2. Unify the timestamps and spatial registration of the 3D site selection map, historical store operation data, user profile types and business district activity data to form a standard spatiotemporal data stream.

[0045] A standard timeline is established using the PTP Precise Time Protocol. Based on this standard timeline, 3D site selection maps, historical store operation data, user profile types, and business district activity data are aligned to a unified time granularity to ensure accurate timestamp synchronization. Data that cannot be directly aligned (such as historical store operation data) is scaled using linear interpolation to eliminate heterogeneity in the time dimension and output a time-aligned commercial site selection coupled dataset.

[0046] The time-aligned commercial site selection coupled dataset is subjected to coordinate system transformation. Furthermore, for the 3D site selection map, affine transformation is applied for coordinate projection to obtain 3D geographic coordinates. For historical store operation data and business district activity data, ArcGIS geocoding service is used for batch address labeling to obtain accurate latitude and longitude coordinates. For user profile types, a spatial autoencoder (directly called from the PyTorch Geometric library) is used to perform user type distribution mapping, converting discrete user profile types into user profile distribution coordinates with spatial continuity.

[0047] It should be noted that the spatial autoencoder achieves user distribution mapping by performing graph convolutional encoding and latent spatial decoding on user profile types.

[0048] Kalman filtering is applied to the 3D geographic coordinates, precise latitude and longitude coordinates, and user profile distribution coordinates to remove outlier noise points. Principal component analysis is then used for feature dimensionality reduction to obtain a standard spatiotemporal data stream. The standard spatiotemporal data stream contains unified time labels, spatial coordinates, and their corresponding multidimensional attribute values, laying the foundation for subsequent real-time stream processing and analysis.

[0049] S1.3. Based on the Apache Flink stream processing engine, a real-time data processing pipeline is built, and standard spatiotemporal data streams are input into the real-time data processing pipeline for complex event pattern recognition, and a digital profile of reference points is output.

[0050] A real-time data processing pipeline is built in the Apache Flink stream processing engine. Further, a Kafka connector is configured in the Flink environment for data subscription and message retrieval, constructing the raw data stream pipeline. In this raw data stream pipeline, out-of-order data is buffered using a Kafka message queue to achieve ordered event processing, resulting in the real-time data processing pipeline. This real-time data processing pipeline not only achieves high-throughput, low-latency data processing but also supports complex event pattern detection, laying the foundation for subsequent complex event pattern recognition.

[0051] It should be noted that the Apache Flink stream processing engine is a distributed computing framework that provides exactly-once state consistency guarantees, high-throughput and low-latency data processing capabilities, and robust fault tolerance mechanisms, making it suitable for commercially viable scenarios.

[0052] The standard spatiotemporal data stream is input into the real-time data processing pipeline for event aggregation. Furthermore, the real-time data processing pipeline divides the standard spatiotemporal data stream into time windows of a fixed size (e.g., 5 minutes, determined based on the specific timeliness requirements of the business site selection scenario) to obtain spatiotemporal event fragments. The spatiotemporal event fragments are then weighted and aggregated to form a candidate event set.

[0053] The relevance of the candidate event set is calculated to generate a business relevance score. The specific mathematical formula is as follows.

[0054] ;

[0055] in, Indicates the score of business relevance. represents the frequency factor, used to characterize the probability of the occurrence frequency of a candidate event; r represents the payoff factor, used to represent the contribution level of the candidate event to the target payoff. This represents the proximity factor, used to measure the decay effect of spatial proximity of candidate events. The network centrality score represents the candidate event, used to quantify the importance of the candidate event in the connection relationships of complex events. This represents a network size normalization constant for candidate events, used to balance the magnitude differences in network centrality scores under complex event relationships of different sizes;

[0056] It should be noted that the frequency factor is obtained by performing sliding window frequency statistics and Min-Max normalization on the candidate events; the payoff factor is obtained by performing payoff aggregation and weight allocation on the candidate events; the proximity factor is obtained by performing coordinate transformation and distance quantization on the candidate events; the network centrality score of the candidate events is obtained by performing graph structure modeling and topological relationship measurement on the candidate events; and the network size normalization constant value of the candidate events is obtained by performing standard deviation normalization on the network centrality score of the candidate events.

[0057] Based on the business relevance score, the candidate event set is categorized and labeled. For example, when the business relevance score is in the first-level threshold range (e.g., 90-100), it indicates that the corresponding candidate event has a high impact on business site selection and is labeled as a core pattern; when the business relevance score is in the second-level threshold range (e.g., 70-90), it indicates that the corresponding candidate event has some reference value for business site selection and is labeled as an auxiliary pattern; when the business relevance score is in the third-level threshold range (e.g., 0-70), it indicates that the corresponding candidate event has low reference value for business site selection and is labeled as a background pattern. The labeled patterns are then integrated to obtain complex event pattern recognition labels.

[0058] It should be noted that the threshold interval is defined based on the cumulative distribution rate of the statistical quantiles of historical business relevance scores.

[0059] Complex event pattern recognition labels and standard spatiotemporal data streams are dimensionally aligned and subjected to tensor outer product to form a spatiotemporal-semantic coupling matrix. Principal component analysis is used to decompose the spatiotemporal-semantic coupling matrix into matrix elements, and the first u matrix elements are extracted and weighted to generate site selection reference point features. The site selection reference point features are serialized and encoded using the Avro serialization tool. For example, the spatiotemporal coordinates of the site selection reference point features are encoded into binary strings, and the semantic weights of the site selection reference point features are encoded into byte arrays, outputting a digital profile of the reference point address.

[0060] S2. Input the reference site digital profile and commercial site selection coupled dataset into the deep learning site selection model. The feature encoding layer performs spatial encoding and temporal modeling, and the similarity measurement layer performs spatial clustering analysis and similarity calculation to output a site selection similarity cloud map.

[0061] S2.1 Build and train a deep learning location selection model.

[0062] In the PyTorch framework, the spatiotemporal convolutional network is invoked through the nn.Conv3d parameter and initialized. For example, the number of input channels is set to 64, the number of output channels is set to 128, and the kernel size is set to (3,5,5). The GeLU activation function is then applied after the spatiotemporal convolutional network to perform nonlinear transformation, which enhances the feature representation capability. Finally, LayerNorm is used for feature normalization to complete the construction of the feature encoding layer.

[0063] The graph attention network is invoked using the nn.MultiheadAttention parameter and initialized, for example, by setting the embedding dimension to 256, the dropout rate to 0.1, and the number of attention heads to 8. Spatial clustering analysis is performed by spectral clustering after the graph attention network, and the attention weights are normalized by the Softmax function to complete the construction of the similarity measurement layer.

[0064] Residual connections are used to concatenate features in the channel dimension between the feature encoding layer and the similarity measurement layer to obtain multi-scale fused feature representations. The ReLU function is then used for non-linear mapping to generate latent space embeddings. An attention mechanism is used to weight the latent space embeddings to obtain weighted latent space features. The Sigmoid function is used to normalize the probability of the weighted latent space features to generate feature association weights. Based on the feature association weights, the feature encoding layer and the similarity measurement layer are stacked across layers to complete the construction of the deep learning location selection model.

[0065] It should be noted that, compared with existing single convolutional or recurrent neural network models, deep learning site selection models can not only capture the local correlation of spatiotemporal data, but also capture global long-range dependencies. This makes up for the shortcomings of single convolutional or recurrent neural network models in representing the dynamic evolution patterns of complex business districts, and achieves high-precision prediction of business district potential and site selection recommendation.

[0066] Next, the deep learning location selection model is trained. Further, the coupled dataset of historical reference point digitization profiles and commercial location selection is divided into a sample set, a training set, and a validation set. On the sample set, parameters are warmed up using Kaiming initialization, and the learning rate is tuned using the Adam optimizer to form the initial deep learning location selection model parameters. On the training set, backpropagation optimization is performed on the initial deep learning location selection model parameters using cross-entropy loss, and gradient pruning is applied simultaneously for gradient norm constraints to obtain intermediate deep learning location selection model parameters. On the validation set, an early stopping mechanism is used to monitor the loss of the intermediate deep learning location selection model parameters to obtain the validation loss value. When the validation loss value exceeds the convergence threshold for several consecutive rounds (e.g., 5), training terminates, and the trained deep learning location selection model is output simultaneously.

[0067] It should be noted that the convergence threshold is defined based on the weighted average relative rate of change of the validation loss value. For example, when the weighted average relative rate of change of the validation loss value over five consecutive training epochs approaches the lower limit of the stability decision interval, it indicates that the deep learning location model training has entered the stable convergence stage, and the gradient update iteration is terminated. Therefore, the corresponding stability decision interval is used as the range of the convergence threshold, such as 1e-5 to 1e-4.

[0068] S2.2 The feature encoding layer performs temporal modeling and spatial encoding through a spatiotemporal convolutional neural network to generate a dynamic index temporal matrix.

[0069] The reference site digitization profile and the commercial site selection coupled dataset are input into the deep learning site selection model through a data loading interface (such as DataLoader). The feature encoding layer uses a two-layer spatiotemporal convolutional neural network for temporal modeling. Further, the first layer performs temporal convolution and padding operations on the reference site digitization profile and the commercial site selection coupled dataset to ensure the integrity of temporal information and obtain primary temporal features. The second layer applies dilated convolution to expand the receptive field and fuse multi-scale features of the primary temporal features to achieve cross-temporal dependency modeling and form advanced temporal features. The GeLU activation function is applied to perform nonlinear transformation on the primary and advanced temporal features, and the LayerNorm layer is applied for normalization to output enhanced temporal features.

[0070] It should be noted that receptive field expansion refers to the process of increasing the perceptual range of primary temporal features by increasing the dilation coefficient of dilated convolution.

[0071] Pyramid convolution is applied to spatially encode the enhanced temporal features. Further, the lower-level convolutions in the pyramid convolution (such as 3x3 standard convolution) are used to perform convolutional kernel filtering and downsampling on the enhanced temporal features to capture local spatial features (such as the distribution of commercial facilities within a 50-meter range). The enhanced temporal features are then subjected to dilated convolution and average pooling through higher-level convolutions (such as 5x5 dilated convolution) to gradually expand the receptive field and capture macro-regional features (such as the morphology of a business district within a 500-meter range). In the feature channels, local spatial features and macro-regional features are weighted and fused, and the Swish activation function is applied for nonlinear projection to generate a dynamic index temporal matrix.

[0072] S2.3 The similarity measurement layer performs spatial clustering analysis on the time series matrix of dynamic indicators and uses the cosine similarity formula to perform similarity calculation to form positional similarity values.

[0073] Graph convolution is used to aggregate neighborhood information on the time series matrix of dynamic indicators to obtain node embedding representations. Graph autoencoder is used to perform feature space projection on the node embedding representations to reduce feature dimensionality and preserve topological structure, resulting in low-dimensional dense feature representations. Spectral clustering is used to perform spatial clustering on the low-dimensional dense feature representations to obtain feature clusters. Vector normalization is then performed on the feature clusters to obtain spatial cluster vectors.

[0074] It should be noted that spectral clustering achieves spatial clustering by constructing a similarity graph and decomposing eigenvectors from low-dimensional dense feature representations.

[0075] The cosine similarity formula is used to calculate the similarity of spatial clustering vectors, forming a positional similarity value. The specific mathematical formula is as follows.

[0076] ;

[0077] in, Indicates the positional similarity value. Indicates the candidate position index. Indicates the first Spatial clustering vectors of candidate locations Indicates the reference address index. Indicates the first Spatial clustering vectors of reference points, Indicates the first The magnitude of the spatial clustering vector for each candidate position. Indicates the first The magnitude of the spatial clustering vector of each reference point address;

[0078] It should be noted that the candidate locations are potential store sites to be evaluated, and the reference locations refer to the site selection samples of historically successful stores; the magnitude of the spatial clustering vector of the candidate location and the magnitude of the spatial clustering vector of the reference location are obtained by performing the sum of squares and the square root operation on the spatial clustering vector of the candidate location and the spatial clustering vector of the reference location, respectively.

[0079] S2.4 Perform spatial interpolation and heatmap rendering on the time series matrix of dynamic indicators and location similarity values, and output the site selection similarity cloud map.

[0080] Importance sampling is performed on the location similarity values ​​to obtain a spatial interpolation point set. Based on the spatial interpolation point set, Kriging spatial interpolation is used to perform grid filling and edge smoothing on the location similarity values ​​to generate a location similarity distribution. The location similarity distribution is fitted with a nonlinear surface using the least squares method to generate a continuous similarity surface. The continuous similarity surface is then weighted and superimposed with key spatiotemporal features in the dynamic index time series matrix (such as the time series changes in pedestrian flow and consumption trends) to obtain the index-similarity coupling feature.

[0081] It should be noted that the least squares method achieves nonlinear surface fitting by performing a polynomial expansion of the position similarity distribution and minimizing the residuals.

[0082] Heatmap rendering is performed on the indicator-similarity coupling features. Furthermore, in the Mapbox GL visualization framework, the indicator-similarity coupling features are subjected to color band mapping and contrast adjustment to generate optimized rendering data. The WebGL rendering engine is then used to perform layer compositing and resolution adaptation on the optimized rendering data to form a high-precision heatmap. Pixel fusion is performed on the high-precision heatmap through bilinear interpolation to smooth pixel boundaries and output a location similarity cloud map.

[0083] It should be noted that the Mapbox GL visualization framework is an open-source JavaScript library with hardware-accelerated rendering and dynamic style editing capabilities. It achieves color banding and contrast adjustment through declarative style specifications (JSON format) and color interpolation algorithms. The WebGL rendering engine is a browser-based low-level graphics interface based on the OpenGL ES standard. It achieves layer compositing and resolution adaptation for optimized rendering data through parallel vertex shader and fragment shader programming.

[0084] S3. Based on the location similarity cloud map, perform regional positioning to form a candidate address set. Use the TOPSIS algorithm to perform multi-objective optimization on the candidate address set and use the spatiotemporal decay factor to perform collaborative filtering compensation to generate a comprehensive score for the candidate addresses.

[0085] S3.1. Based on the location similarity cloud map, use the region growing algorithm to perform hotspot detection and region localization to form a set of candidate addresses.

[0086] High-value pixels (e.g., pixels with a location similarity value greater than 0.8) are extracted from the location similarity cloud map as initial seed points. The initial seed points are then expanded into neighborhoods to obtain initial growth regions. Contour parameters of the initial growth regions are extracted, and the Douglas-Peucker algorithm is applied to perform polygon approximation to generate region contour polygons. Gaussian convolution is used to filter and denoise the region contour polygons to eliminate the noise effects of spatial jitter, resulting in smooth contour polygons. The smooth contour polygons are then overlapped and merged to output candidate hotspot regions.

[0087] It should be noted that the Douglas-Peucker algorithm is a curve simplification method based on a vertical distance threshold. It approximates polygons by extracting feature points and fitting line segments to contour parameters.

[0088] Candidate hotspot areas are located. Further, the area function of the Shapely library is used to identify the area of ​​each candidate hotspot. If the area is less than the minimum commercial area (e.g., 500 square meters), the corresponding candidate hotspot is defined as a discrete region and eliminated. Morphological closing operations are then used to optimize the edges and fill holes in the remaining candidate hotspots to restore regional integrity, resulting in optimized candidate regions. GeoPandas spatial analysis tools are then used to construct a spatial index for the optimized candidate regions, generating a set of candidate addresses. Each candidate address includes attributes such as center latitude and longitude, boundary polygon vertex sequence, and regional commercial potential score.

[0089] It should be noted that the GeoPandas spatial analysis tool constructs spatial indexes by creating R-tree indexes and using spatial reference systems on optimized candidate regions.

[0090] S3.2. Perform multi-objective optimization on the candidate address set using the TOPSIS algorithm to generate an initial optimization score. Use the spatiotemporal decay factor to perform collaborative filtering compensation on the initial optimization score to generate a comprehensive score for the candidate addresses.

[0091] The TOPSIS algorithm is used to perform multi-objective optimization on the candidate address set to generate an initial optimization score. The specific mathematical formula is as follows.

[0092] ;

[0093] in, Indicates the alternative address index. Indicates the first The initial optimization score of the candidate addresses, This indicates that the target index should be optimized. This represents the total number of optimization objectives. Indicates the first The weighting coefficients of each optimization objective. Indicates the first The first alternative address is in the Standardized values ​​over an optimization objective express The negative ideal solution of an optimization objective This represents the positive ideal solution to the optimization objective.

[0094] It should be noted that the optimization objective is a key evaluation dimension of the candidate address set, including commercial potential, operating costs, and risk level; the weight coefficient of the optimization objective is defined based on the variance contribution rate of the optimization objective, with an exemplary value range of 0.1 to 0.5; the standardized value is obtained by performing range processing and dimensionless transformation on the candidate addresses and optimization objectives; the process of obtaining the negative ideal solution and the positive ideal solution is as follows: according to the indicator type classification rules (based on the benefit attribute definition of the optimization objective), the optimization objectives are sorted in descending order to obtain an ordered indicator sequence; the last value in the ordered indicator sequence is extracted as the negative ideal solution, and the first value in the obtained ordered indicator sequence is extracted as the positive ideal solution.

[0095] The initial optimization score is attenuated and compensated by a spatiotemporal decay factor, and the comprehensive score of the candidate addresses is output. The specific mathematical formula is as follows.

[0096] ;

[0097] in, For the first The overall score of the candidate addresses, For the first The initial optimization score of the candidate addresses, Represents the spatiotemporal decay factor. Indicates the compensation coefficient. This represents the collaborative filtering compensation value;

[0098] It should be noted that the spatiotemporal decay factor is defined based on the spatiotemporal decay law of the historical initial optimization score, and the exemplary value range is [0, 0.3]. The compensation coefficient is obtained by performing negative correlation mapping and exponential smoothing on the spatiotemporal decay factor. The process of obtaining the collaborative filtering compensation value is as follows: the initial optimization score is collaboratively weighted by the compensation coefficient to obtain the basic compensation amount, and the basic compensation amount is numerically normalized to obtain the collaborative filtering compensation value.

[0099] S4. Based on the comprehensive score of the candidate addresses, classify, sort and perform one-hot encoding on the candidate address set to obtain a priority addressing list, perform cross-validation on the priority addressing list, and output the addressing decision package.

[0100] S4.1 Based on the comprehensive score of the candidate addresses, classify and sort the candidate address set to generate priority address categories, perform one-hot encoding on the priority address categories, and obtain the priority address list.

[0101] When the overall score of the candidate addresses is in a high range (e.g., 0.8-1.0), it indicates excellent business potential. The corresponding candidate address set is classified as a priority location and ranked as first priority. When the overall score of the candidate addresses is in a medium range (e.g., 0.6-0.8), it indicates good business potential. The corresponding candidate address set is classified as a secondary location and ranked as second priority. When the overall score of the candidate addresses is in a low range (e.g., 0.4-0.6), it indicates average business potential. The corresponding candidate address set is classified as a reserve and ranked as third priority. The classified candidate addresses and priority sequences are then integrated to output the priority address classification.

[0102] It should be noted that the score range is based on the cumulative distribution rate of the overall scores of historical candidate addresses.

[0103] One-hot encoding of priority address categories is performed using the OneHotEncoder function of the Scikit-learn library. Further, binary encoding is applied to these priority address categories; for example, "preferred address" is encoded as [1,0,0], "secondary address" as [0,1,0], and "alternate reserve" as [0,0,1], resulting in binary encoded vectors. These vectors are then hashed using the MD5 hash algorithm to map them to hash fingerprint strings. Finally, key-value pairs are concatenated between the binary encoded vectors and the hash fingerprint strings to form a priority address list, enabling structured storage and fast retrieval.

[0104] S4.2. Use the Bootstrap self-sampling method to perform confidence cross-validation on the priority site selection list and output the site selection decision package.

[0105] Random sampling with replacement is performed on the priority location list to generate multiple Bootstrap samples (e.g., 1000). The capacity of each Bootstrap sample is the same as that of the priority location list but includes duplicate samples. The confidence of each Bootstrap sample is measured using the percentile method to obtain the initial confidence. The initial confidence is then aggregated by quantiles to obtain the sample confidence vector.

[0106] It should be noted that the percentile method achieves confidence measurement by sorting and truncating the Bootstrap samples.

[0107] When the sample confidence vector exceeds the confidence threshold, it indicates that the site selection result is statistically significant, and the corresponding priority site selection list is deemed to have passed the confidence verification. When the sample confidence vector is below the confidence threshold, it indicates that the site selection result is not stable enough, and the corresponding priority site selection list is deemed to have failed the confidence verification, requiring resampling and evaluation. The priority site selection list that has passed the confidence verification is then discretized to obtain optimized site selection parameters. The optimized site selection parameters are then formatted and encapsulated using the Apache Avro tool to output a site selection decision package, providing statistical reliability assurance for commercial site selection.

[0108] It should be noted that the confidence threshold is defined based on the statistical characteristics of the quantiles of the historical sample confidence vectors. For example, the sample confidence vectors of Bootstrap samples from historical successful site selection cases are collected, and the 95th percentile of the sample confidence vectors is taken as the benchmark reference value. The benchmark reference value is then smoothed by a sliding window to obtain the dynamic confidence threshold. An exemplary value range is [0.85, 0.95].

[0109] This embodiment also provides a computer device applicable to the intelligent business site selection method based on big data analysis, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent business site selection method based on big data analysis as proposed in the above embodiment.

[0110] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0111] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the intelligent business location selection method based on big data analysis as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0112] In summary, this invention effectively solves the coupling problem caused by the spatiotemporal heterogeneity of multi-source data by constructing a standard spatiotemporal data stream and performing feature alignment, significantly improving the fusion accuracy between multi-source data and thus enhancing the representativeness and reliability of the digital profile of reference locations. Simultaneously, a deep learning-based location selection model is constructed, which not only achieves collaborative processing of spatial encoding and temporal modeling but also fully explores the deep correlation between spatial distribution and temporal evolution, significantly improving the dynamic prediction capability of commercial location selection.

[0113] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart business site selection method based on big data analysis, characterized in that: include, Collect a commercial site selection coupled dataset and perform feature alignment to form a standard spatiotemporal data stream; The Apache Flink stream processing engine is used to perform complex event pattern recognition on standard spatiotemporal data streams to form digital profiles of reference points. The reference site digital profile and commercial site selection coupled dataset are input into the deep learning site selection model. The feature encoding layer performs temporal modeling and spatial encoding, and the similarity measurement layer performs spatial clustering analysis and similarity calculation, outputting a site selection similarity cloud map. The specific construction process of the deep learning location selection model is as follows. A feature encoding layer is constructed using a spatiotemporal convolutional network, and a similarity measurement layer is constructed using a graph attention network. A deep learning location selection model is constructed by performing gradient backpropagation and cross-layer stacking on the feature encoding layer and the similarity measurement layer; The location is determined by the location similarity cloud map, forming a set of candidate addresses. The TOPSIS algorithm is used to optimize the candidate address set for multiple objectives, and the spatiotemporal decay factor is used to perform collaborative filtering compensation to generate a comprehensive score for the candidate addresses. Based on the comprehensive scores of the candidate addresses, the candidate address set is classified, sorted, and one-hot encoded to obtain a priority addressing list. Cross-validation is then performed on the priority addressing list to output the addressing decision package.

2. The intelligent business site selection method based on big data analysis as described in claim 1, characterized in that: The commercial site selection coupled dataset includes a 3D site selection map, historical store operation data, user profile types, and business district activity data.

3. The intelligent business site selection method based on big data analysis as described in claim 2, characterized in that: The process of creating a digital profile of the reference address specifically includes the following steps. The 3D site selection map, historical store operation data, user profile types and business district activity data are unified with timestamps and spatially registered to form a standard spatiotemporal data stream. A real-time data processing pipeline is built based on the Apache Flink stream processing engine. Standard spatiotemporal data streams are input into the real-time data processing pipeline for complex event pattern recognition, and digital profiles of reference points are output.

4. The intelligent business site selection method based on big data analysis as described in claim 3, characterized in that: The output location similarity cloud map specifically includes the following steps. The reference site digital profile and the commercial site selection coupled dataset are input into the deep learning site selection model. The feature encoding layer performs temporal modeling and spatial encoding through a spatiotemporal convolutional neural network to generate a dynamic indicator temporal matrix. The similarity measurement layer performs spatial clustering analysis using a spectral clustering algorithm and calculates similarity using a cosine similarity formula to generate positional similarity values. Spatial interpolation and heatmap rendering are performed on the time series matrix of dynamic indicators and the location similarity values ​​to output a site selection similarity cloud map.

5. The intelligent business site selection method based on big data analysis as described in claim 4, characterized in that: The process of generating a comprehensive score for candidate addresses specifically includes the following steps: Based on the location similarity cloud map, a region growing algorithm is used to detect hotspots and locate regions to form a set of candidate addresses. The TOPSIS algorithm is used to perform multi-objective optimization on the candidate address set to generate an initial optimization score. The initial optimization score is attenuated and compensated by a spatiotemporal decay factor to generate a comprehensive score for the candidate addresses.

6. The intelligent business site selection method based on big data analysis as described in claim 5, characterized in that: The process of obtaining the priority addressing list specifically includes the following steps. Based on the comprehensive scores of the candidate addresses, the candidate address set is classified and sorted to generate priority address categories; One-hot encoding is performed on priority address categories to obtain a priority address selection list.

7. The intelligent business site selection method based on big data analysis as described in claim 6, characterized in that: The output location decision package specifically refers to the output location decision package generated by performing confidence cross-validation on the priority location list using the Bootstrap self-sampling method.

8. 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 steps of the intelligent business site selection method based on big data analysis as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent business site selection method based on big data analysis as described in any one of claims 1 to 7.

Images