A business opportunity identification method and system based on space and semantic cooperation

Abstract

The application discloses a business opportunity identification method and system based on space and semantic cooperation and belongs to the field of data processing. In view of the problems of long time consumption and low accuracy of existing demand identification, the application provides a business opportunity identification method based on space and semantic cooperation: acquiring space data and semantic data; generating a grid level feature vector for the space data; generating a topic level feature vector for the semantic data; simultaneously taking grid nodes and topic nodes as two types of vertices, constructing a demand-supply relationship edge according to the frequency of topics in the grid, and forming a grid-topic bipartite graph; inputting the grid level feature vector and the topic level feature vector into a Relation-GAT model, predicting a grid level opportunity score and outputting and displaying the grid level opportunity score. Through the grid-topic bipartite graph and the Relation-GAT model, bidirectional updating of space constraints and semantic driving is realized, the problem of separation of space features and semantic information in traditional business opportunity identification is solved, and the accuracy and real-time performance of prediction are improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and more specifically, relates to a method and system for identifying business opportunities based on spatial and semantic collaboration. Background Technology

[0002] With the rapid changes in urban commercial landscapes, identifying unmet consumer needs in a timely, accurate, and low-cost manner has become a core pain point for businesses in site selection, store opening, and marketing. Existing technologies can be broadly categorized into three types:

[0003] GIS-based spatial analysis platforms, such as Esri ArcGIS and Baidu Insight, focus on POI, pedestrian flow, and traffic heat maps, but lack a deep understanding of unstructured semantic data such as text and images.

[0004] SaaS products based on user profiles and recommendation algorithms, such as Meituan Cloud Map and Alibaba City Brain, focus on online behavior, but lack sufficient granular characterization of offline spaces and have sparse semantic tags.

[0005] Deep learning-based event extraction or sentiment analysis tools, such as Tencent Public Opinion and TalkingData Semantic Cloud, excel at text insights but cannot be precisely coupled with geographic grids.

[0006] In summary, existing solutions generally suffer from the contradiction of "high spatial dimension but weak semantic dimension, or rich semantic insight but lack of spatial application", resulting in fragmented business opportunity identification results, poor interpretability, and delayed timeliness.

[0007] To address the aforementioned issues, corresponding improvements have been made. For example, Chinese patent application CN202310082744.1, published on August 9, 2024, discloses a method for selecting restaurant locations based on the number of multiple types of POIs in a region. This method includes the following steps: Step S1: Select the administrative region where the store to be opened is located, divide the region into grids, and export the latitude and longitude coordinates of the center point of each grid; Step S2: Input the latitude and longitude coordinates of each center point into the surrounding search API interface of the map software, and search within a predetermined range according to circular areas, crawling POI data of 23 major categories within each circular area. The drawback of this patent is the lack of analysis of dynamic factors, which limits the accuracy of the results.

[0008] For example, Chinese patent application CN202310668730.8, published on August 25, 2023, discloses a method for recommending chain store locations by integrating multi-dimensional features. This method includes: obtaining preliminary store locations through a five-dimensional evaluation model of business recommendation index; acquiring data information about different social circles around the store based on the characteristics of different social circles within the business district, using consumer data analysis, social media analysis, and video surveillance analysis; analyzing and judging the data information to obtain the consumption habits, social methods, brand preferences, store quality requirements, and service staff image requirements of different circles; providing corresponding consumption decisions for consumers in different circles and with different attributes based on the characteristics of the people within each circle; judging the differences in consumer attention and sales volume in different circles based on social circle analysis, through the locked location of the store and the layout and optimization of multiple stores; and determining the degree of influence of each factor on the store using a logistic regression algorithm to identify the key points for optimization of multiple chain stores. The shortcomings of this patent are: video surveillance can only reflect physical space behavior and cannot capture online consumption decisions; the data is not comprehensive enough, and the accuracy is limited. Summary of the Invention

[0009] 1. The problem to be solved

[0010] To address the issues of low accuracy and long processing time in existing business opportunity identification methods, this invention provides a business opportunity identification method and system based on spatial and semantic collaboration. The method utilizes a grid-topic bipartite graph construction and a Relation-GAT model to achieve bidirectional updates driven by spatial constraints and semantics, resolving the problem of the separation between spatial features and semantic information in traditional business opportunity identification, thus improving prediction accuracy and real-time performance. The system employs modular components, resulting in a simple structure and ease of implementation.

[0011] 2. Technical Solution

[0012] To solve the above problems, the present invention adopts the following technical solution.

[0013] A business opportunity identification method based on spatial and semantic collaboration includes the following steps:

[0014] S1: Data Acquisition: Acquiring spatial and semantic data;

[0015] S2: Data preprocessing: Generate grid-level feature vectors for spatial data; generate topic-level feature vectors for semantic data; and simultaneously construct demand-supply relationship edges based on the frequency of topic occurrence within the grid, using grid nodes and topic nodes as two types of vertices, forming a grid-topic bipartite graph.

[0016] S3: Data Prediction: Input grid-level feature vectors and topic-level feature vectors into the Relation-GAT model, update node representations through demand-supply relationship edges; finally, predict grid-level opportunity scores through the demand-supply difference loss function;

[0017] S4: Output Results: Visualize the grid-level chance scores.

[0018] Furthermore, in step S1:

[0019] The acquisition of spatial data includes: acquiring spatial data within the target area divided into grids of a set size through a geographic information platform; the spatial data includes time-segmented pedestrian density, POI distribution density, passenger flow at transportation hubs, housing price gradients, nighttime light intensity, and meteorological data, with time granularity supporting daily updates;

[0020] The acquisition of semantic data includes: acquiring multimodal unstructured data related to business scenarios, including user comments, social media text, short video captions, and investment promotion posters, and extracting their geolocation tags to achieve spatial association.

[0021] Furthermore, after obtaining the target area, the target area is divided into grids, with the set size being 150m×150m~250m×250m.

[0022] Furthermore, the grid-level feature vector in step S2 is generated through a spatiotemporal attention mechanism and includes at least one of the following: pedestrian flow statistics by time period, POI same category density, and complementary category density;

[0023] The topic-level feature vectors are generated through multimodal alignment: Adapter-Transformer encoding is used for text data, and cross-modal adapter encoding is used for image data, followed by clustering to obtain topics.

[0024] Furthermore, step S3, which updates the node representation through the demand-supply relationship, specifically includes:

[0025] Spatial to semantic update: Grid nodes pass spatial features to associated topic nodes, and the regional adaptability representation of topic nodes is updated through weighted aggregation of relation edge weights;

[0026] Semantic-to-spatial update: Topic nodes pass demand features to associated grid nodes, and key signals are filtered through semantic attention to update the opportunity potential representation of grid nodes.

[0027] Furthermore, the Relation-GAT model specifically includes:

[0028] Spatial feature propagation channel: Construct a spatial topology network based on the geographical proximity of grid nodes, and aggregate the spatiotemporal features of adjacent grids through graph convolution operations;

[0029] Semantic association channel: Calculates the semantic similarity weights between topic nodes and realizes topic feature interaction based on the attention mechanism;

[0030] Dynamic relation edge construction: The output signal of the spatial feature propagation channel is concatenated with the output signal of the semantic association channel, and the demand-supply relation edge weights between grid nodes and topic nodes are generated through a multi-layer neural network;

[0031] Bidirectional node update:

[0032] Grid node representation update: It integrates its own spatial features with the semantic features of related topic nodes, and the integration weight is determined by the relation edge weights in the previous step;

[0033] Topic node representation update: It integrates its own semantic features with the spatial features of related grid nodes, and the integration weight is determined by the relation edge weights in the previous step;

[0034] The above steps are executed iteratively until a joint embedding representation that integrates spatial attributes and semantic requirements is generated.

[0035] Furthermore, step S3 also includes: using historical store opening events as intervention variables, calculating the causal effect confidence interval of the opportunity score using a counterfactual diffusion model; and injecting the causal effect value into the demand-supply relationship edge as a weight correction factor.

[0036] Furthermore, step S5 is included: Incremental update process: When new semantic data is input, only the associated topic vectors and relation edge weights of the affected grid are updated; a local graph recalculation strategy is adopted to limit the recalculation range to the 3rd order neighborhood of the affected grid, ensuring that the delay is less than 30 seconds.

[0037] Furthermore, in step S4, when the grid-level chance score is greater than or equal to a set value, the visualization includes:

[0038] Reason: Attributable to the theme with the largest demand-supply gap;

[0039] Action: Generate store opening strategy suggestions based on theme characteristics;

[0040] Results: Predicted increase in customer traffic / sales;

[0041] Confidence: Output the 95% confidence interval for causal effect verification.

[0042] A system using the spatial and semantic co-operational business opportunity identification method as described in any of the above technical solutions includes:

[0043] Data acquisition module: used to acquire spatial and semantic data;

[0044] The data preprocessing module is used to generate grid-level feature vectors for spatial data; generate topic-level feature vectors for semantic data; and construct demand-supply relationship edges based on the frequency of topic occurrence within the grid, using grid nodes and topic nodes as two types of vertices, thus forming a grid-topic bipartite graph.

[0045] Spatial and Semantic Collaborative Model: This model inputs grid-level and topic-level feature vectors into the Relation-GAT model, updates node representations through demand-supply relationship edges, and finally predicts grid-level opportunity scores using the demand-supply difference loss function.

[0046] Visualization module: Used to visualize grid-level chance scores.

[0047] 3. Beneficial effects

[0048] (1) This invention first combines spatial data and semantic data, and realizes spatial association through geographic location tags, which solves the problem of the separation between spatial features and user needs in traditional methods, enabling business opportunity identification to have both macro trend insight and micro demand perception capabilities; secondly, it realizes bidirectional updates of spatial constraints and semantic drive through grid-topic bipartite graph and Relation-GAT model, and directly quantifies the degree of unmet demand as opportunity score by combining demand-supply difference loss function, making the prediction results more in line with actual business logic; and it is visualized to enhance operability and intuitiveness, and improve user experience; the whole method significantly improves the accuracy, real-time performance and interpretability of business opportunity identification, and provides a scientific decision-making tool for retail, real estate, urban planning and other fields.

[0049] (2) When collecting semantic data, this invention maps these data to specific grids through geographic location tags to achieve a direct correlation between “user voices and commercial supply”, avoiding the sample bias problem in traditional surveys; at the same time, the spatial data time granularity supports daily updates, improving the timeliness of prediction; and the target area is divided into grids according to a certain size, taking into account both the consumption of computing resources and the ability to capture spatial details. Too small a grid will lead to an explosive growth in the amount of data and increase the computing delay; too large a grid may blur the community-level consumption hotspots, while this size range can cover a single commercial entity and avoid excessive subdivision, ensuring the statistical significance of dynamic indicators.

[0050] (3) This invention automatically weights the density of people at different times through an attention mechanism to capture the demand patterns of time and space, avoiding the loss of time period information caused by static statistics; it reflects the intensity of regional competition through the density of the same category and reveals the commercial synergy effect through the density of complementary categories. The attention mechanism dynamically allocates feature weights, avoiding the limitations of manually setting thresholds and improving the adaptability to complex commercial scenarios; it clusters multimodal features into commercial themes, making semantic information structured from fragmented data, which is convenient for subsequent models to capture the supply and demand relationship between themes and regions; at the same time, it establishes a relationship between grid-level features and theme-level features in the preprocessing stage, providing high-quality input for the subsequent Relation-GAT model to model the "demand-supply" relationship and reducing noise interference in model training.

[0051] (4) This invention achieves mutual enhancement of spatial constraints and semantic drive through a bidirectional update mechanism. The weighted aggregation of relation edges and attention screening ensure the retention of key signals and improve the representation ability of node representation. The parallel design of spatial feature propagation channels and semantic association channels, combined with dynamic relation edge generation, enables the model to adapt to the demand-supply correlation strength under different scenarios. The iterative fusion mechanism gradually optimizes the joint embedding representation and finally generates high-order features that simultaneously contain spatial dynamism and semantic demand, making the prediction results more in line with actual business logic.

[0052] (5) This invention introduces historical store opening events as intervention variables, calculates the confidence interval of causal effect through counterfactual diffusion model, effectively eliminates the interference of confounding variables, and makes the chance score causal explanatory; the weight correction factor dynamically adjusts the relation edges, improving the model's ability to simulate real business decisions; at the same time, the local graph recalculation strategy greatly reduces the consumption of computing resources, and the addition of semantic data triggers only updates the affected grid, ensuring that the system response delay is <30 seconds, meeting the real-time incremental update requirements, and reducing operation and maintenance costs;

[0053] (6) The system of the present invention presents a modular design, from data acquisition to preprocessing to collaborative model and finally to visualization display, ensuring independent optimization of each link and strong system scalability; the spatial and semantic collaborative model, as the core module, can be embedded into existing business analysis platforms to improve technical adaptability and reduce integration difficulty; the whole system achieves the simultaneous performance of spatial features and semantic information, as well as the accurate spatial positioning of demand gaps, improving the accuracy, timeliness and completeness of business opportunity identification. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the process of the present invention;

[0055] Figure 2 This is a schematic diagram of the processing flow of the spatial and semantic collaborative model;

[0056] Figure 3 This is a schematic diagram of the incremental update process;

[0057] Figure 4 This is a flowchart of Example 1. Detailed Implementation

[0058] The present invention will be further described below with reference to specific embodiments and accompanying drawings. It should be noted that a business opportunity refers to an unmet need or resource optimization space that exists in the market and can be utilized by enterprises to achieve profit goals; specifically in this application, a business opportunity is essentially a high-potential, low-competition, and feasible supply and demand gap identified from massive data through spatial and semantic collaborative analysis, such as opening a store.

[0059] like Figure 1 As shown, a business opportunity identification method based on spatial and semantic collaboration includes the following steps:

[0060] S1: Data Acquisition: Acquiring spatial and semantic data;

[0061] Specifically, in step S1, spatial data such as pedestrian density and POI distribution reflect the supply and demand in the physical world (e.g., high pedestrian grids have a customer base), but cannot capture consumer intentions; semantic data such as user reviews and social media text imply unmet needs (e.g., "cannot find a low-calorie restaurant"), but lacks spatial feasibility; therefore, this step simultaneously acquires spatial and semantic data to solve the problem of blind spots in a single data dimension.

[0062] S2: Data preprocessing: Generate grid-level feature vectors for spatial data; generate topic-level feature vectors for semantic data; and simultaneously construct demand-supply relationship edges based on the frequency of topic occurrence within the grid, using grid nodes and topic nodes as two types of vertices, forming a grid-topic bipartite graph.

[0063] Specifically, the essence of step S2 preprocessing is to establish a heterogeneous connection bridge between space and semantics; grid-level feature vectors (spatial vectors): spatiotemporal attention aggregation captures dynamic patterns (e.g., morning peak traffic variance <15% → stable consumption scenario); topic-level feature vectors (topic vectors): Adapter-Transformer achieves cross-modal alignment (text "salad" and salad image are mapped to the same vector); because early splicing of spatial and semantic features in traditional processing leads to information confusion, this step establishes a grid-topic bipartite graph, where grid nodes only retain spatial features and topic nodes only retain semantic features, thereby ensuring the purity of features;

[0064] S3: Data Prediction: Input grid-level feature vectors and topic-level feature vectors into the Relation-GAT model, update node representations through demand-supply relationship edges; finally, predict grid-level opportunity scores through the demand-supply difference loss function;

[0065] Specifically, the Relation-GAT model in step S3 is designed specifically for bipartite graphs. This model employs a dual-channel parallel strategy, performing graph convolution between grid nodes through the spatial channel to capture the geographic radiation effect; calculating attention between topic nodes through the semantic channel to identify demand symbiotic relationships; then dynamically updating the relationship edges; and finally combining the demand-supply difference loss function to directly quantify the degree of unmet demand as an opportunity score, making the prediction results more accurate and logical. Meanwhile, the demand-supply difference loss function can be defined as loss function = MSE(opportunity score, logarithm of sales in the 30 days after actual store opening) + 0.1 × causal regularization term.

[0066] S4: Output Results: Visualizes the grid-level chance scores. Visualization improves both presentation and understandability.

[0067] This embodiment first combines spatial and semantic data, and achieves spatial association through geographic location tags, solving the problem of the disconnect between spatial features and user needs in traditional methods. This enables business opportunity identification to simultaneously possess macro-trend insights and micro-demand perception capabilities. Secondly, it achieves bidirectional updates of spatial constraints and semantic drives through grid-topic bipartite graphs and the Relation-GAT model. Combined with the demand-supply difference loss function, it directly quantifies unmet demand as an opportunity score, making the prediction results more aligned with actual business logic. Furthermore, it provides visualization, enhancing operability and intuitiveness, and improving user experience. The entire method significantly improves the accuracy, real-time performance, and interpretability of business opportunity identification, providing a scientific decision-making tool for retail, real estate, urban planning, and other fields.

[0068] In one specific implementation, in step S1:

[0069] The acquisition of spatial data includes: acquiring spatial data within the target area divided into grids of a set size through a geographic information platform; the spatial data includes time-segmented pedestrian density, POI distribution density, passenger flow at transportation hubs, housing price gradients, nighttime light intensity, and meteorological data, with time granularity supporting daily updates to improve the timeliness of predictions;

[0070] The acquisition of semantic data includes: acquiring multimodal unstructured data related to business scenarios, including user comments, social media text, short video captions, and investment promotion posters, and extracting their geographic location tags to achieve spatial association; mapping these data to specific grids through geographic location tags to achieve a direct association between "user voices" and "business supply", avoiding the sample bias problem in traditional surveys.

[0071] More specifically, after acquiring the target area, the target area is divided into grids, with the set size being 150m×150m to 250m×250m. This balances computational resource consumption with spatial detail capture capabilities. A grid that is too small would lead to an explosive growth in data volume and increase computational latency; a grid that is too large might obscure community-level consumer hotspots. This size range covers a single commercial entity while avoiding over-segmentation, ensuring the statistical significance of dynamic indicators.

[0072] In one specific implementation, the grid-level feature vector in step S2 is generated through a spatiotemporal attention mechanism, outputting a 128-dimensional spatial vector; it includes at least one of the following: pedestrian flow time-segmentation statistics, POI same category density, and complementary category density;

[0073] The topic-level feature vectors are generated through multimodal alignment: text data is encoded using Adapter-Transformer, and image data is encoded using cross-modal adapter, and then clustered to obtain topics; topics may include consumer demand, supply gaps, and price sensitivity, etc.; that is, Adapter-Transformer is used to encode text and images separately, and then the alignment layer maps them to the same 256-dimensional semantic space to achieve topic consistency and location accuracy.

[0074] In one specific implementation, step S3, represented by the demand-supply relationship edge update node, specifically includes:

[0075] Spatial to semantic update: Grid nodes pass spatial features to associated topic nodes, and the regional adaptability representation of topic nodes is updated through weighted aggregation of relation edge weights;

[0076] Semantic-to-spatial update: Topic nodes pass demand features to associated grid nodes, and key signals are filtered through semantic attention to update the opportunity potential representation of grid nodes.

[0077] like Figure 2 As shown, specifically, the Relation-GAT model described in this embodiment includes:

[0078] Spatial feature propagation channel: Construct a spatial topology network based on the geographical proximity of grid nodes, and aggregate the spatiotemporal features of adjacent grids through graph convolution operations;

[0079] Semantic association channel: Calculates the semantic similarity weights between topic nodes and realizes topic feature interaction based on the attention mechanism;

[0080] Dynamic relation edge construction: The output signal of the spatial feature propagation channel is concatenated with the output signal of the semantic association channel, and the demand-supply relation edge weights between grid nodes and topic nodes are generated through a multi-layer neural network;

[0081] Bidirectional node update:

[0082] Grid node representation update: It integrates its own spatial features with the semantic features of related topic nodes, and the integration weight is determined by the relation edge weights in the previous step;

[0083] Topic node representation update: It integrates its own semantic features with the spatial features of related grid nodes, and the integration weight is determined by the relation edge weights in the previous step;

[0084] The above steps are executed iteratively until a joint embedding representation that integrates spatial attributes and semantic requirements is generated.

[0085] In other words, the Relation-GAT model in this embodiment has a three-layer architecture.

[0086] Spatial layer: Using a preset size (250m×250m) grid as the basic unit, it incorporates spatiotemporal data such as pedestrian flow, POI, traffic, housing prices, nighttime light, and weather.

[0087] Semantic layer: Multimodal encoding is performed on text and image tags such as reviews, microblogs, short video subtitles, and investment posters to extract topic vectors such as consumer demand, supply gaps, and price sensitivity;

[0088] Collaboration layer: Construct a grid-topic bipartite graph, use a dual-channel GNN (spatial convolution + semantic attention) to learn the joint embedding of grids and topics, and output an opportunity score; the opportunity score can be the predicted value of the logarithm of sales within a specific number of days after a new store is opened in a specific grid or other values ​​that can characterize the unmet business demand potential of the grid.

[0089] In one specific implementation, step S3 further includes: using historical store opening events as intervention variables, calculating the causal effect confidence interval of the opportunity score using a counterfactual diffusion model; and injecting the causal effect value into the demand-supply relationship edge as a weight correction factor. In other words, this embodiment uses real store opening events as intervention variables, utilizes Do-Calculus and a diffusion-based counterfactual generation model to calculate the causal effect confidence interval of the "opportunity score"; and uses the causal weight as an attribute of the graph relationship edge to achieve causally interpretable recommendations, thereby improving the model's ability to simulate real business decisions.

[0090] like Figure 3 As shown, in one specific implementation, step S5 is also included: Incremental update process: When new semantic data is input, only the associated topic vectors and relation edge weights of the affected grids are updated; a local graph recalculation strategy is adopted to limit the recalculation range to the third-order neighborhood of the affected grids, ensuring that the latency is <30 seconds. The local graph recalculation strategy significantly reduces the consumption of computing resources. The addition of semantic data triggers only the update of the affected grids, ensuring that the system response latency is <30 seconds, meeting the requirements of real-time incremental updates, and reducing operation and maintenance costs.

[0091] In one specific implementation, in step S4, when the grid-level chance score is greater than or equal to a set value, a quadruple visualization is performed, and a text and graphic report is automatically generated using a natural language template; the quadruple includes:

[0092] Reason: It is attributed to the theme with the greatest demand-supply gap; it directly targets unmet demand themes (such as "community fresh food demand gap"), avoiding users blindly guessing the source of opportunities and improving the targeting of decision-making;

[0093] Actions: Generate store opening strategy suggestions based on theme characteristics; automatically match business format suggestions (such as "small-area high-turnover restaurant" and "24-hour convenience store") according to theme type (such as "trendy restaurant" and "community service"), reducing the threshold for users to formulate strategies;

[0094] Results: Predicts the increase in customer traffic / sales; facilitates user evaluation of ROI; the effect prediction is linked with grid-level feature vectors (such as time-segmented statistics of pedestrian traffic) to ensure that the prediction results are consistent with the upper limit of regional consumption capacity and avoid overly optimistic expectations;

[0095] Confidence: Outputs the 95% confidence interval for causal effect verification; by eliminating confounding factors such as seasonality and policy through counterfactual diffusion model, it outputs the statistical correlation strength of "opportunity score - real effect" (such as "confidence interval [20%, 40%]"), avoiding misjudging accidental phenomena as opportunities.

[0096] In other words, this embodiment constructs a complete decision-making chain, avoiding information gaps and optimizing the entire process from opportunity identification to decision implementation. This not only improves decision-making efficiency but also ensures the scientific nature of the decisions through quantification and verification mechanisms. Furthermore, only grid-level opportunity scores greater than or equal to a set value are displayed in a four-element format, avoiding the waste of resources from uniform displays and preventing information overload, ensuring that users can focus their resources on the opportunities most likely to generate benefits.

[0097] In one specific embodiment, a system using a spatial and semantic co-operation-based business opportunity identification method as described in any of the above embodiments includes:

[0098] Data acquisition module: used to acquire spatial and semantic data;

[0099] The data preprocessing module is used to generate grid-level feature vectors for spatial data; generate topic-level feature vectors for semantic data; and construct demand-supply relationship edges based on the frequency of topic occurrence within the grid, using grid nodes and topic nodes as two types of vertices, thus forming a grid-topic bipartite graph.

[0100] Spatial and Semantic Collaborative Model: This model inputs grid-level and topic-level feature vectors into the Relation-GAT model, updates node representations through demand-supply relationship edges, and finally predicts grid-level opportunity scores using the demand-supply difference loss function.

[0101] Visualization module: Used to visualize grid-level chance scores.

[0102] The system in this embodiment presents a modular design, from data acquisition to preprocessing to collaborative modeling and finally to visualization, ensuring independent optimization of each link and strong system scalability. The spatial and semantic collaborative model, as the core module, can be embedded into existing business analysis platforms to improve technical adaptability and reduce integration difficulty. The entire system achieves the simultaneous performance of both spatial features and semantic information, as well as the accurate spatial positioning of demand gaps, improving the accuracy, timeliness, and completeness of business opportunity identification.

[0103] It is also worth noting that in an experiment with 1.2 million real data points in 8 cities, the system in this embodiment achieved a 42% higher Top-100 hit rate for opportunity identification compared to traditional GIS solutions and a 36% higher rate compared to pure semantic solutions. The interpretable path resulted in a user comprehension score (1-5) of 4.7. It supports three deployment models: SaaS, private deployment, and edge boxes, and has been implemented in 6 industries, including chain restaurants, convenience stores, and shared offices, demonstrating a very broad prospect for application.

[0104] To further understand the technical solution of this application, the following examples are provided:

[0105] Example 1

[0106] like Figure 4 As shown, taking "a light food chain seeking new store opportunities in Nanjing" as an example, the steps are as follows:

[0107] Data Access: Spatial Data: Using the Gaode API, we obtained three months of daily data on pedestrian flow, POI, and subway exit passenger flow in a 7×7km area of ​​Nanjing with a 250m×250m grid.

[0108] Semantic data: We crawled 320,000 reviews from Dianping containing keywords such as "light meals", "salads" and "low-calorie" and 58,000 food photos with location tags.

[0109] Preprocessing: Grid-level features: Take the mean and variance of pedestrian flow over different time periods, and calculate the density of similar and complementary light meals for POIs; Semantic features: Use Adapter-BERT to extract 128-dimensional topic vectors, and then use K-Means to obtain 12 categories of demand topics (such as "weight loss lunch" and "fitness meal replacement").

[0110] Model training: Construct a bipartite graph with 14,520 grids × 12 topics, totaling 174,240 nodes;

[0111] Trained for 50 epochs using Relation-GAT (2 layers, 256 hidden layers), with a learning rate of 1e-3 and a batch size of 2048.

[0112] Loss function = MSE (opportunity score, logarithm of sales in the 30 days after the store opens) + 0.1 × causal regularization term.

[0113] Online reasoning:

[0114] The user entered "new opportunities for light meals within 3km of Nanjing Hexi";

[0115] The system retrieves the corresponding grid and outputs the Top-5 chance grids and their scores:

[0116] ① 250m north of Yuantong Metro Station, opportunity score 0.92 (high demand, few competitors, dense night runners);

[0117] ②…

[0118] Each result is accompanied by a causal explanation: If a store were to open here, the average daily customer traffic is expected to increase by 18% (CI: 12%-24%).

[0119] Deployment and Expansion: SaaS: Alibaba Cloud ACK + Pulsar streaming link, supporting tens of thousands of concurrent connections; Edge: NVIDIA Jetson Xavier runs a pruned 30M parameter model, with a single card covering 50km. 2 API: RESTful / gRPC dual protocol, latency <200ms.

[0120] Through the above embodiments, enterprises can complete the process from "submitting a request" to "obtaining a list of high-confidence opportunities with explanations" within 2 hours, significantly reducing the cost of trial and error in site selection.

[0121] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A business opportunity identification method based on spatial and semantic collaboration, characterized in that: Includes the following steps: S1: Data Acquisition: Acquiring spatial and semantic data; S2: Data preprocessing: Generate grid-level feature vectors for spatial data; generate topic-level feature vectors for semantic data; and simultaneously construct demand-supply relationship edges based on the frequency of topic occurrence within the grid, using grid nodes and topic nodes as two types of vertices, forming a grid-topic bipartite graph. S3: Data Prediction: Input grid-level feature vectors and topic-level feature vectors into the Relation-GAT model, update node representations through demand-supply relationship edges; finally, predict grid-level opportunity scores through the demand-supply difference loss function; S4: Output Results: Visualize the grid-level chance scores; The Relation-GAT model specifically includes: Spatial feature propagation channel: Construct a spatial topology network based on the geographical proximity of grid nodes, and aggregate the spatiotemporal features of adjacent grids through graph convolution operations; Semantic association channel: Calculates the semantic similarity weights between topic nodes and realizes topic feature interaction based on the attention mechanism; Dynamic relation edge construction: The output signal of the spatial feature propagation channel is concatenated with the output signal of the semantic association channel, and the demand-supply relation edge weights between grid nodes and topic nodes are generated through a multi-layer neural network; Bidirectional node update: Grid node representation update: It integrates its own spatial features with the semantic features of related topic nodes, and the integration weight is determined by the relation edge weights in the previous step; Topic node representation update: It integrates its own semantic features with the spatial features of related grid nodes, and the integration weight is determined by the relation edge weights in the previous step; The above steps are executed iteratively until a joint embedding representation that integrates spatial attributes and semantic requirements is generated.

2. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: In step S1: The acquisition of spatial data includes: acquiring spatial data within the target area divided into grids of a set size through a geographic information platform; the spatial data includes time-segmented pedestrian density, POI distribution density, passenger flow at transportation hubs, housing price gradients, nighttime light intensity, and meteorological data, with time granularity supporting daily updates; The acquisition of semantic data includes: acquiring multimodal unstructured data related to business scenarios, including user comments, social media text, short video captions, and investment promotion posters, and extracting their geolocation tags to achieve spatial association.

3. A business opportunity identification method based on spatial and semantic collaboration according to claim 2, characterized in that: After obtaining the target area, the target area is divided into grids, with the set size being 150m×150m~250m×250m.

4. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: The grid-level feature vector in step S2 is generated through a spatiotemporal attention mechanism and includes at least one of the following: pedestrian flow time-segmentation statistics, POI same category density, and complementary category density. The topic-level feature vectors are generated through multimodal alignment: Adapter-Transformer encoding is used for text data, and cross-modal adapter encoding is used for image data, followed by clustering to obtain topics.

5. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: Step S3, specifically represented by updating nodes along the demand-supply relationship, includes: Spatial to semantic update: Grid nodes pass spatial features to associated topic nodes, and the regional adaptability representation of topic nodes is updated through weighted aggregation of relation edge weights; Semantic-to-spatial update: Topic nodes pass demand features to associated grid nodes, and key signals are filtered through semantic attention to update the opportunity potential representation of grid nodes.

6. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: Step S3 further includes: using historical store opening events as intervention variables, calculating the causal effect confidence interval of the opportunity score using a counterfactual diffusion model; and injecting the causal effect value into the demand-supply relationship edge as a weight correction factor.

7. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: It also includes step S5: Incremental update process: When new semantic data is input, only the associated topic vectors and relation edge weights of the affected grid are updated; a local graph recalculation strategy is adopted to limit the recalculation range to the 3rd order neighborhood of the affected grid to ensure that the delay is <30 seconds.

8. The business opportunity identification method based on spatial and semantic collaboration according to claim 1, characterized in that: In step S4, when the grid-level chance score is greater than or equal to a set value, the visualization includes: Reason: Attributable to the theme with the largest demand-supply gap; Action: Generate store opening strategy suggestions based on theme characteristics; Results: Predicted increase in customer traffic / sales; Confidence: Output the 95% confidence interval for verifying the causal effect.

9. A system using the business opportunity identification method based on spatial and semantic collaboration as described in any one of claims 1-8, characterized in that: include: Data acquisition module: used to acquire spatial and semantic data; Data preprocessing module: used to generate grid-level feature vectors from spatial data; The semantic data is used to generate topic-level feature vectors; at the same time, grid nodes and topic nodes are used as two types of vertices, and demand-supply relationship edges are constructed according to the frequency of topic occurrence in the grid to form a grid-topic bipartite graph. Spatial and Semantic Collaborative Model: This model inputs grid-level and topic-level feature vectors into the Relation-GAT model, updates node representations through demand-supply relationship edges, and finally predicts grid-level opportunity scores using the demand-supply difference loss function. Visualization module: Used to visualize grid-level chance scores.

Images