Supply chain intelligent recommendation system and method based on geographic space constraint and feature matching

Abstract

The invention provides a supply chain intelligent recommendation system and method based on geographic space constraint and feature matching. The method comprises the steps that S1, an enterprise-supplier relation graph is constructed according to supplier feature data and historical relation data, and a standardized matrix is generated; s2, taking the normalized matrix as input, and generating an enterprise feature vector and a supplier feature vector by using a degree normalization-based adaptive feature perception graph convolutional neural network model; s3, based on the geographic coordinates of the suppliers, performing clustering analysis by using a hierarchical clustering algorithm, and identifying a geographic distribution mode of the supply chain; s4, establishing a candidate supplier pool of geographical constraints for each enterprise according to a supply chain geographical distribution mode identification result; and S5, in the candidate pool range, recommending Top-K suppliers to each enterprise based on the matching degree of the enterprise feature vector and the supplier feature vector. According to the method, geographic space constraint and feature matching are effectively fused, and the precision, efficiency and interpretability of supply chain recommendation are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the fields of big data analysis and enterprise digital transformation, and specifically relates to a supply chain intelligent recommendation system and method based on geospatial analysis and feature matching. Background Technology

[0002] As global supply chains become increasingly complex, companies face the challenge of efficiently selecting suitable suppliers. Traditional supplier selection methods rely heavily on human experience and simple scoring systems, resulting in low efficiency, high subjectivity, and difficulty in scaling. In recent years, machine learning-based recommendation systems have achieved significant success in e-commerce and other fields, but their application in supply chain management still faces many challenges: 1) the need to consider the multidimensional characteristics and geographical constraints of suppliers simultaneously; 2) the complex network structure of supply chain relationships; and 3) the need to balance the accuracy and practicality of recommendations.

[0003] In existing technologies, some methods make recommendations based solely on supplier characteristics, neglecting geographical factors; others consider only geographical proximity, ignoring supplier quality matching. How to organically combine geographical constraints with feature matching to construct an intelligent recommendation system that is both accurate and practical has become a pressing technical problem in this field. Summary of the Invention

[0004] The purpose of this invention is to propose an intelligent recommendation system and method for supply chains based on geospatial analysis and feature matching. By creatively introducing a technical framework of "deep fusion of geospatial constraints and network feature learning," a hierarchical processing and modular collaborative intelligent recommendation system is designed. The core of this system is as follows: First, using an adaptive feature perception model based on degree normalization, high-quality potential feature vectors of enterprises and suppliers are distilled from historical transaction networks and supplier attributes, accurately capturing their functional preferences and capability profiles. Second, the geographical distribution patterns of suppliers are analyzed independently and in-depth, and the geographical morphological characteristics of each enterprise's supply chain are quantified through hierarchical clustering and dynamic index calculation. Then, based on the quantified geographical features, an adaptive geographical search radius is dynamically generated for each enterprise, constructing a "balanced" geographical constraint candidate pool that cleverly balances matching accuracy and geographical feasibility. Finally, precise feature similarity matching is performed within this candidate pool to generate the final recommendation list, and the decision-making basis is presented intuitively through visualization. This technical approach achieves a paradigm shift from "global blind matching" to "local precise matching," significantly improving the practicality and interpretability of the results while ensuring recommendation relevance.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A supply chain intelligent recommendation method based on geospatial constraints and feature matching specifically includes the following steps:

[0007] Step 1: Construct an enterprise-supplier relationship graph based on supplier feature data and historical relationship data to generate a normalized matrix for graph convolution operations; the supplier feature data includes multidimensional static attributes of suppliers, and the historical relationship data includes historical transaction records between enterprises and suppliers;

[0008] Step 2: Using the generated normalized matrix as input, a graph convolutional neural network model based on degree normalization adaptive feature perception is used to learn the feature representations of enterprises and suppliers, generating enterprise feature vectors and supplier feature vectors.

[0009] Step 3: Based on the supplier's geographical coordinates, use a hierarchical clustering algorithm to perform cluster analysis and identify the geographical distribution pattern of the supply chain;

[0010] Step 4: Based on the results of supply chain geographical distribution pattern recognition, calculate the adaptive search radius for each enterprise and construct a geographically constrained candidate supplier pool;

[0011] Step 5: Within the candidate pool, recommend Top-K suppliers for each company based on the degree of matching between the company's feature vector and the supplier's feature vector;

[0012] Step 6: Generate visualization reports and statistical analysis results.

[0013] Preferably, the step of constructing an enterprise-supplier relationship graph based on supplier feature data and historical relationship data to generate a normalized matrix for graph convolution operations is as follows:

[0014] Define a graph structure G = (V, E, F), where V represents the set of nodes, and V = V c ∪V s V c V represents the set of enterprise nodes. s Let E represent the set of supplier nodes and E represent the set of edges. V c ×V s This represents the historical procurement transaction relationship between the enterprise and its suppliers; F represents the initial feature matrix of all nodes, initialized as follows:

[0015] For each supplier node v s Assign an initial eigenvector of dimension d. ,vector The data originates from multidimensional static attribute data from suppliers, while each enterprise node v c initial features f c Set to a zero vector or initialize randomly;

[0016] Construct a binary adjacency matrix based on the edge set E. Subsequently, v for each enterprise node is calculated separately. c Each supplier node v s in-degree deg(v) c ), deg(v s ), and for each enterprise node v c Pointing to supplier node v s edge Assign weights For each of the supplier nodes v s Pointing to enterprise node v c edge Assign weights To construct the edge weight matrix W; finally, to generate the normalized matrix used for graph convolution. Where D is the degree matrix, It represents the Hadamardi (or Hadama) stack.

[0017] Preferably, the step of using an adaptive feature-aware graph convolutional neural network model based on degree normalization to learn feature representations of enterprises and suppliers, and generating enterprise feature vectors and supplier feature vectors, is as follows:

[0018] The network structure of the graph convolutional neural network model is composed of two graph convolutional layers stacked sequentially.

[0019] The first layer graph convolution operation is defined as follows: ,in This is the first layer of trainable weight matrix. For graph convolution operations, This is the output matrix of the first layer graph convolution operation;

[0020] The convolution operation of the second layer graph is defined as follows: ,in This is the trainable weight matrix for the second layer. Let be the output matrix of the second layer graph convolution operation; then the graph convolutional neural network model outputs the final d-dimensional feature representation matrix of all nodes. .

[0021] Preferably, the training of the graph convolutional neural network model employs a hybrid loss function L. total :

[0022]

[0023] in, To compare the losses, For the feature reconstruction loss part, and These are preset hyperparameters used to balance the two losses.

[0024] The contrast loss portion Lcontrast The aim is to distinguish between positive and negative sample edges: for each positive sample edge that exists in historical transaction relationships Randomly sample a supplier node within the scope of non-historical transaction relationships to generate a corresponding negative sample edge. The number of negative samples and the number of positive samples are kept in a preset ratio, among which Indicates the relationship with enterprise node v c Supplier nodes with historical transactions, Indicates the relationship with enterprise node v c Supplier nodes with no historical transactions, compute node pairs as well as The model outputs cosine similarity of feature vectors and And adopt the loss function: To bring positive sample pairs closer together and push negative sample pairs further apart, For enterprise node v c The model output features, , Supplier nodes Model output features and supplier nodes The model output features, mean(sim(·)) is the average of the similarity of all negative or positive pairs in the current batch (summed and divided by the logarithm).

[0025] The feature reconstruction loss part L MSE The aim is to ensure that the model can effectively reconstruct the original features of suppliers from graph structure information, defined as the model output features of supplier nodes. Initial characteristics of supplier nodes Mean square error between: , where N s The number of suppliers is represented by s, which is the iteration parameter.

[0026] Preferably, the step of using hierarchical clustering algorithms to perform cluster analysis based on supplier geographical coordinates to identify supply chain geographical distribution patterns is as follows:

[0027] Based on the latitude and longitude coordinates of all supplier nodes Calculate the actual distance between each pair of pairs. To construct a symmetric geographic distance matrix ;

[0028] For each enterprise node v c Extract enterprise node v c The set S of all historically associated supplier nodes c and S cThe corresponding coordinate subset; based on the sub-distance matrix corresponding to the coordinate subset, a preset linking method and distance threshold are used. The agglomerative hierarchical clustering algorithm is used to divide S into groups. c Divided into m geographical clusters This ensures that the distance between any two suppliers within the same geographical cluster does not exceed [a certain value]. ;

[0029] For each cluster C k Cluster C k Using the coordinates of all suppliers within the cluster as input, the cluster C is predicted through a GNN model. k center coordinates The training objective of the GNN model is to minimize the regression loss between the prediction center and the geometric center of all supplier coordinates within the cluster.

[0030] For cluster C k Its forecasting center calculates the value of each member supplier's data. i Distance d to the prediction center i And based on the distance d i Calculate dynamic geographic indicators:

[0031] R80 radius: , This indicates the radius that covers 80% of the cluster members;

[0032] Gravity fraction: ,in This is the attenuation coefficient, used to measure the compactness of the cluster around the center;

[0033] Coefficient of variation (CV): cv = std_dist / avg_dist, used to quantify the dispersion of the distance;

[0034] Other statistics include mean distance (avg_dist), standard deviation distance (std_dist), minimum distance (min_dist), maximum distance (max_dist), and median distance (median_dist).

[0035] Preferably, based on the supply chain geographical distribution pattern recognition results, an adaptive search radius is calculated for each enterprise to construct a geographically constrained candidate supplier pool, as follows:

[0036] According to each enterprise v c Each geographic cluster C k Dynamic geographic indicators are used to calculate cluster C according to a preset priority logic. k Corresponding adaptive search radius ;

[0037] For a specific target enterprise, based on each cluster C k Predicted center coordinates and the corresponding adaptive search radius Set circular area filter criteria: for any coordinate in the global supplier database... suppliers j If supplier s j With any cluster center The spherical geographical distance between them satisfies Then supplier s j Preliminarily included in enterprise v c The range of candidates to be considered;

[0038] Collect all suppliers that meet the circular area screening criteria of the target company and merge them to form a preliminary candidate supplier set. For sets Perform deduplication to remove duplicate suppliers, ultimately generating a candidate supplier pool for the target company. .

[0039] Preferably, cluster C is computed according to a preset priority logic. k Corresponding adaptive search radius The specific calculation rules are as follows:

[0040]

[0041] in, For cluster C k The radius of R80, For cluster C k The average distance from each member supplier to the prediction center. The system default radius is used, and `clamp(·)` is a constraint function that ensures the final radius value is limited to the preset minimum value R. min With the maximum value R max The interval formed by [R] min , R max ]Inside.

[0042] Preferably, within the candidate pool, the process of recommending Top-K suppliers for each enterprise based on the matching degree between the enterprise feature vector and the supplier feature vector is as follows:

[0043] For the target company and its pool of candidate suppliers The enterprise node v corresponding to the target enterprise c The model outputs feature vectors With the candidate supplier pool Each candidate supplier j Corresponding supplier nodes The model outputs feature vectors A one-to-one comparison is performed, and the cosine similarity is used to calculate the matching score between the two. c,j After completing the candidate pool After calculating the matching scores of all suppliers, all candidate suppliers are sorted in descending order based on their matching scores, and the top K suppliers are selected to form the final Top-K recommendation set for the target company. .

[0044] A supply chain intelligent recommendation system based on geospatial constraints and feature matching, wherein the system is implemented using any of the above-mentioned supply chain intelligent recommendation methods, including a graph neural network feature learning module, a geospatial clustering and analysis module, a precise matching recommendation module, and a visualization module;

[0045] The graph neural network feature learning module is used to construct an enterprise-supplier relationship graph based on supplier feature data and historical relationship data, generate a normalized matrix for graph convolution operation, and use an adaptive feature-aware graph convolutional neural network model based on degree normalization to learn the feature representations of enterprises and suppliers, and generate enterprise feature vectors and supplier feature vectors.

[0046] The geospatial clustering and analysis module is used to perform clustering analysis based on the supplier's geographical coordinates using a hierarchical clustering algorithm to identify the geographical distribution patterns of the supply chain.

[0047] The precise matching and recommendation module is used to calculate an adaptive search radius for each enterprise based on the results of the supply chain geographical distribution pattern recognition, construct a geographically constrained candidate supplier pool, and recommend Top-K suppliers for each enterprise within the candidate pool based on the degree of matching between the enterprise feature vector and the supplier feature vector.

[0048] The visualization module is used to generate visualization reports and statistical analysis results.

[0049] Preferably, the visualization module generates content including a geographic visualization map displaying the enterprise's location, supplier clusters, adaptive radius, and recommendation results, as well as an analysis report;

[0050] The geographic visualization map is a visual map generated for each target enterprise, covering the historical geographic distribution and clustering results of its suppliers. The visualization map supports zooming and panning operations, and the overlay display content within the same coordinate system includes:

[0051] 1) The geographical location of the target company marked with a specific icon;

[0052] 2) Distribution points of all historical suppliers, displayed using the first category of markers;

[0053] 3) Geographical clusters of different suppliers of the target company, distinguished by different colors;

[0054] 4) The predicted center location of each geographic cluster, marked with a special symbol;

[0055] 5) A circular area is drawn with each prediction center as the center and the corresponding adaptive search radius as the radius, and the geographic coverage of the candidate pool is displayed with semi-transparent filling;

[0056] 6) The locations of the Top-K recommended suppliers selected from the candidate pool, highlighted with the second-category marker;

[0057] The analysis report consists of two parts:

[0058] 1) Enterprise-level detailed report: For a single enterprise, it provides a detailed list of each geographic cluster's ID, prediction center latitude and longitude, adaptive search radius, R80 radius, gravity fraction, coefficient of variation, average distance, number of suppliers within the cluster, and a list of specific supplier IDs.

[0059] 2) System-level summary report: Summarizes the overall situation of system processing, including the statistical characteristics of the total number of enterprises processed, the total number of geographical clusters, the size of the candidate pool, and the distribution of enterprises by different supplier number ranges.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. This invention achieves an organic unity between geospatial constraints and deep feature matching, resolving the inherent contradiction between practicality and accuracy in recommendation systems. Traditional methods often use geographic filtering as a rigid pre-processing cutoff or a simple post-processing ranking factor, easily losing high-quality but distant candidates or introducing nearby but low-quality noise. This invention, through its core design of an "adaptive candidate pool," upgrades geographic analysis from simple distance filtering to intelligent constraints based on morphological understanding. The search range (adaptive radius) is dynamically determined based on the unique geographic cluster characteristics (such as compactness and dispersion) of each enterprise's supply chain, making the constraints themselves flexible and customized. Deep feature matching is then performed within this intelligently defined "reasonable activity range," ensuring both the geographic feasibility of the recommended objects and maximizing the retention of high-matching, high-quality candidates, achieving an optimal balance between accuracy and practicality.

[0062] 2. This invention innovatively applies and improves graph neural network models, more accurately characterizing the deep relationships in complex supply chain networks. The proposed adaptive feature-aware graph convolutional network based on degree normalization introduces a degree-aware edge weight allocation mechanism. This more realistically simulates the relationship between the influence of enterprise nodes and their connectivity in real-world scenarios, enhancing the rationality of feature propagation. Simultaneously, it employs a hybrid training objective combining contrastive loss and feature reconstruction loss. This forces the model to capture both network structure information and node attribute information simultaneously, so that the learned feature vectors not only encode the relationship topology of "who cooperates with whom", but also contain the intrinsic attributes of "what kind of supplier is", thereby generating more discriminative and generalizable feature representations, laying a solid foundation for subsequent accurate matching.

[0063] 3. Significantly enhances the interpretability and decision support capabilities of the recommendation system. The multi-layered processing flow and visual output of this invention provide a clear traceability of the decision-making process for each recommendation result. Decision-makers can not only see "who was recommended," but also understand "why it was recommended in this region" through a visual map, understand "the morphological characteristics of the supply chain in this region" (such as very compact or relatively dispersed) through cluster indicator reports, and see "the basis for the search range" through the adaptive radius. This transparent decision-making process greatly enhances the trustworthiness of the system results, transforming the recommendation system from a "black box" tool into a "white box" think tank to assist strategic analysis, helping enterprise managers make more scientific and insightful decisions on strategic issues such as supplier selection, regional warehouse layout, and risk zoning management. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the implementation of an embodiment of the present invention. Detailed Implementation

[0065] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0066] This invention proposes an intelligent supply chain recommendation system based on geospatial constraints and feature matching, implemented using a computer system, comprising the following modules that are sequentially connected and work collaboratively:

[0067] The graph neural network feature learning module is used to construct a bipartite graph of enterprise and supplier. With historical transaction relationships as edges and supplier features as initial node attributes, it learns node representations through an adaptive feature perception model based on degree normalization and outputs the potential feature vectors of enterprises and suppliers.

[0068] The geospatial clustering and analysis module is used to perform hierarchical clustering based on the latitude and longitude coordinates of suppliers, using a preset distance threshold to identify the geographical clusters of suppliers for each enterprise. For each cluster, a dedicated graph neural network is used to predict its geographical center and calculate the radius R80. 80Dynamic geographic metrics including gravity fraction, coefficient of variation (CV), mean distance (avg_dist), standard deviation distance (std_dist), minimum distance (min_dist), maximum distance (max_dist), and median distance (median_dist);

[0069] The adaptive candidate pool construction module is used to calculate the adaptive search radius R for each cluster based on its geographical indicators. adaptive : Prioritize 1.2×R 80 The purpose is to cover the main range of the cluster with a slight extension. If there are no cluster suppliers, then 2.0×avg_dist is used. Otherwise, the default value is used, and the radius is constrained to the range of [50,500]km. Suppliers falling within the circular area of ​​each cluster are selected based on this radius, and after merging and deduplication, a candidate pool of geographical constraints for the enterprise is formed.

[0070] The precise matching recommendation module is used to calculate the cosine similarity sim(f) between the enterprise feature vector and the feature vectors of its suppliers in the candidate pool. c , f s Sort by score in descending order, and select the Top-K suppliers for each company as the final recommendation list;

[0071] The visualization module is used to generate geographic visualization maps and analysis reports that display enterprise locations, supplier clusters, adaptive radii, and recommendation results.

[0072] In this embodiment, the graph neural network feature learning module adopts a graph convolutional neural network model based on degree normalization and adaptive feature perception, specifically including:

[0073] The heterogeneous bipartite graph construction unit is used to transform raw enterprise procurement relationship data and supplier characteristic data into graph structure data that can be processed by graph neural networks; specifically, the graph structure G=(V, E, F) is defined, where the node set V=V c ∪V s Includes enterprise node V c With supplier node V s Edge set E V c ×V s This represents historical procurement transaction relationships; for each supplier node v s V s Assign an initial eigenvector of dimension d. This vector originates from the supplier's static attribute data (such as company size, company type, whether it is listed, science and technology innovation level, etc.), while the initial feature f of the company node... c Set to zero vector or random initialization.

[0074] The normalized adjacency matrix and edge weight calculation unit are used to construct the normalized matrix required for graph convolution operations. First, a binary adjacency matrix is ​​constructed based on the edge set E. Subsequently, v for each enterprise node is calculated separately. c Each supplier node v s in-degree deg(v) c ), deg(v s (i.e., the number of nodes associated with it), and for each enterprise node v c Edge pointing to supplier node vs Assign weights For each edge pointing from supplier node vs to enterprise node vc Assign weights This is used to construct the edge weight matrix W; this reflects that in real-world business, a large enterprise that collaborates with numerous suppliers may have a relatively weak relationship with a single supplier, and vice versa; finally, a normalized matrix is ​​generated for graph convolution. Where D is the degree matrix. This represents the Hadamard product (element-by-element multiplication).

[0075] The degree-normalized graph convolutional network model consists of two stacked graph convolutional layers. It aggregates information within two-hop neighborhoods, enabling enterprise nodes to absorb features from their direct suppliers and neighboring suppliers, and vice versa, thus learning a deeper network context representation. The first-layer graph convolutional operation is defined as follows: Where F is the initial feature matrix of all nodes, This is the first layer of trainable weight matrix. The first layer is a graph convolution operation; the second layer graph convolution operation is defined as: Output the final d-dimensional feature representation matrix of all nodes. .

[0076] The hybrid loss function design and training unit guides the parameter optimization of the graph convolutional network model; this unit defines and minimizes a hybrid loss function L consisting of two parts. total :

[0077] Comparison of loss portion L contrast The aim is to distinguish between positive sample edges (edges present in E) and negative sample edges (randomly sampled edges that do not exist), ensuring that the learned embeddings effectively reflect the network structure. For a batch of positive sample edges... and the corresponding generated negative sample edges ,in This refers to supplier nodes that have historical transactions with the enterprise node (VC). This refers to supplier nodes that have no historical transactions with the enterprise node (VC), and compute node pairs. as well as The model outputs cosine similarity of feature vectors and And adopt the loss function: To bring positive sample pairs closer together and push negative sample pairs further apart, Output features for the enterprise node VC model. , Supplier nodes Model output features and supplier nodes The model output features.

[0078] Feature reconstruction loss part L MSE The aim is to ensure that the model can effectively reconstruct the original features of suppliers from graph structure information, and to ensure that the embedded vectors retain the inherent attribute information of suppliers. It is defined as the mean squared error between the model output features of supplier nodes and the initial features. , where N s This represents the number of suppliers.

[0079] The total loss function is a weighted sum of the two: in and The hyperparameters are preset to balance the objectives of the two losses. By minimizing this loss through backpropagation, the model ultimately outputs a low-dimensional, dense, and semantically rich latent feature vector of all enterprise nodes and supplier nodes.

[0080] In this embodiment, the geospatial clustering and analysis module is independent of feature learning and specifically processes the supplier's geographic coordinate data. Its aim is to reveal the geographic distribution patterns and cluster characteristics of each enterprise's supply chain network. Specifically, it includes the following processing units executed sequentially:

[0081] The geographic distance matrix calculation unit is used to quantify the spatial proximity between suppliers; this unit is based on the latitude and longitude coordinates of all supplier nodes. Use spherical geographic distance formulas (such as the Haversine formula) to calculate the actual distance between any two pairs of objects. ,in, Construct a symmetric geographic distance matrix .

[0082] Enterprise supplier subgraph partitioning and hierarchical clustering units are used to identify the geographical distribution patterns of suppliers within each enterprise; for each enterprise node v c First, extract the set S of all historically associated supplier nodes. cand its corresponding coordinate subset; subsequently, based on the sub-distance matrix corresponding to this subset, a preset linking method (such as full linking) and a distance threshold are used. (For example, 800 km) Agglomerative hierarchical clustering algorithm is used to divide S c Divided into m geographical clusters This ensures that the distance between any two suppliers within the same cluster does not exceed [a certain value]. This allows the identification of multiple geographical "locations" or "regional centers" within the company's supply chain.

[0083] Cluster center prediction unit, used to infer the potential center of each geographic cluster using a graph neural network model; for each cluster C k Using the coordinates of all suppliers within the cluster as input, a dedicated optimized locator GNN model (training a dedicated optimized locator graph neural network model for each identified cluster) predicts the center coordinates of the cluster. This GNN model takes a complete graph composed of supplier coordinates within the cluster as input, and aggregates information through multi-layer graph convolution and global attention pooling layers to predict a coordinate point that better represents the "center of gravity" or "optimal center" of the cluster. k lon k Its training objective is to minimize the regression loss between the prediction center and the geometric center of all supplier coordinates within the cluster.

[0084] A dynamic geographic indicator calculation unit is used to evaluate the geographic characteristics of each cluster based on the prediction center; for cluster C k Its forecasting center calculates the value of each member supplier's data. i Distance d to the center i And based on this, calculate:

[0085] R80 radius: That is, the 80th quantile of the distance from the set, which represents the radius covering 80% of the members of the cluster;

[0086] Gravity fraction: ,in This is the attenuation coefficient, used to measure the compactness of the cluster around the center;

[0087] Coefficient of variation (CV): cv = std_dist / avg_dist, used to quantify the dispersion of the distance;

[0088] Other statistics include mean distance (avg_dist), standard deviation distance (std_dist), minimum distance (min_dist), maximum distance (max_dist), and median distance (median_dist).

[0089] These metrics serve as key inputs to the adaptive candidate pool building module.

[0090] In this embodiment, the adaptive candidate pool construction module is responsible for the first and most crucial fusion of feature matching requirements and geospatial constraints. Its core innovation lies in dynamically calculating the search radius of each cluster based on the geographic analysis results, rather than using a fixed value. The calculation rules have priority and robustness. Specifically, it implements the following functions:

[0091] Adaptive search radius calculation unit: Based on the geospatial clustering and analysis module, for each enterprise v c Each geographic cluster C k The output dynamic geographic metrics are used to calculate the adaptive search radius for the cluster according to a preset priority logic. The specific calculation rules are as follows:

[0092]

[0093] in, For cluster C k The radius of R80, The system uses a default radius (e.g., 100 kilometers) as a preset value. `clamp(x,a,b)` is a constraint function that ensures the final radius value is limited to the preset minimum value `R`. min (e.g., 50 km) and the maximum value R max The interval [R] consisting of (e.g., 500 km) min , R max [Within], to prevent extreme values ​​from occurring.

[0094] Geospatial candidate region filtering unit: based on the predicted cluster C k center coordinates and its corresponding adaptive search radius Define several circular filtering areas in geographic space. For any supplier s in the global supplier database... j Its coordinates are If it is related to any cluster center The spherical geographical distance between them satisfies Then the supplier s j Preliminarily included in enterprise v c The range of candidates to be considered.

[0095] Candidate pool aggregation and deduplication unit: targeting enterprise v c Collect all suppliers that meet any of the above circular area screening criteria and merge them to form a preliminary candidate supplier set. Subsequently, a deduplication operation is performed on the set to remove duplicate suppliers, ultimately generating a geographically constrained pool of potential candidate suppliers specific to the enterprise. Candidate pool This will be used as input to the precise matching recommendation module; this step transforms the global search problem into a local search problem, significantly narrowing the computational scope of feature matching.

[0096] In this embodiment, the precise matching recommendation module specifically implements the following functions:

[0097] Feature matching calculation unit within the candidate pool: for target enterprise v c and its dedicated pool of geographically constrained candidate suppliers The feature vector f of the enterprise node c (Output from the feature learning module) and each candidate supplier node in the pool eigenvector f sj Perform a one-to-one comparison; calculate the matching score between the two using cosine similarity. c,j The calculation formula is as follows: The score ranges from [-1, 1] and is used to quantify the directional consistency between enterprise feature preferences and supplier features. A higher score indicates a better match.

[0098] Matching result sorting and filtering unit: After completing the candidate pool After calculating the matching scores of all suppliers, this unit sorts all candidate suppliers in descending order based on their scores, generating an ordered list. Where s (1) It has the highest matching score. Then, based on the system's preset recommendation count K, it is selected from the sorted list. The top K suppliers are selected to form the company's final Top-K recommendation set. ,Right now Since the candidate pool has already undergone geographical screening, this recommendation list naturally possesses the dual attributes of "excellent feature matching" and "strong geographical feasibility".

[0099] Full-system batch recommendation generation unit: This unit generates recommendations for all enterprise nodes v in the system. c Repeat the above two units to generate a Top-K recommendation set for each enterprise. This process fully utilizes the results of the adaptive candidate pool construction module, decomposing the global, computationally intensive problem of matching all suppliers into candidate pools that are independent for each enterprise and significantly smaller in scale. This addresses the local matching problem within the system, thereby significantly improving the overall computational efficiency and scalability of the recommendation system.

[0100] Recommendation list output interface: generates a Top-K recommendation set for each enterprise. The data is structured and integrated, and output in a predetermined format as a direct result for presentation to end users or for use by downstream systems. This output typically includes company identification, a list of recommended supplier identifiers, and corresponding matching scores.

[0101] In this embodiment, the key results of each processing stage of the system are visualized in the form of charts and reports, specifically including:

[0102] Enterprise Supplier Geographic Distribution Visualization Unit: For each target enterprise, a visualization map of its historical supplier geographic distribution and clustering results is generated. This map is overlaid on the same coordinate system, displaying: 1) the target enterprise's geographic location (marked with a specific icon); 2) the distribution points of all its historical suppliers (displayed with first-type markers); 3) the geographic clusters of suppliers divided by the geospatial clustering and analysis module, distinguished by different colors; 4) the predicted center location of each geographic cluster (marked with a special symbol); 5) a circular area drawn with each predicted center as the center and the corresponding adaptive search radius as the radius, with semi-transparent filling to show the geographic coverage of the candidate pool; 6) the locations of the Top-K recommended suppliers selected from the candidate pool (highlighted with second-type markers). This visualization map supports zooming and panning.

[0103] Structured Analysis Report Generation Unit: Automatically generates analysis reports containing text and key data. The report content is divided into two parts: 1) Company-level detailed report: For a single company, it lists in detail the ID of each geographic cluster, the latitude and longitude of the prediction center, the adaptive search radius, the R80 radius, the gravity fraction, the coefficient of variation, the average distance, the number of suppliers in the cluster, and a list of specific supplier IDs; 2) System-level summary report: Summarizes the overall situation of the system processing, including the total number of companies processed, the total number of geographic clusters, the statistical characteristics of the candidate pool size (mean, maximum, minimum, median), and the distribution of companies according to different supplier number ranges.

[0104] Visualization output interface: Save the generated map images, statistical chart images, and analysis report text files to the specified file system path, and record the generation log for users to view and access.

[0105] In this embodiment, the system adopts a four-level processing architecture, with each level connected via data files or structured data, specifically including:

[0106] Level 1: Feature representation learning, the input of which is the original relational data D. relation (Including historical transaction records between enterprises and suppliers) and supplier characteristic data D feature(Including the supplier's multidimensional static attributes), processed by the graph neural network feature learning module; the output of this stage is two feature matrices: the enterprise feature matrix. Supplier Feature Matrix Where N c N represents the total number of companies. s d represents the total number of suppliers and d represents the feature dimension; the feature matrix output at this level is stored in a structured file for subsequent levels to access.

[0107] Level 2: Geographic pattern analysis, which takes geographic coordinate data as input. This involves analyzing the latitude and longitude coordinates of all suppliers, which are then processed by the geospatial clustering and analysis module. The output of this stage is a set of supplier clustering results for each enterprise. Each cluster C k It includes a group of supplier members and comes with dynamic geographic metrics such as the predicted center coordinates, R80 radius, gravity fraction, coefficient of variation, and average distance for the cluster. This level of output is typically saved as an Excel file containing multiple worksheets, which record detailed information for each company and each cluster.

[0108] Level 3: Candidate space construction, whose input is the clustering results output from Level 2. The corresponding indicators are processed through the adaptive candidate pool construction module; this level outputs a geographically constrained candidate supplier pool S for each enterprise. c Its specific content is as follows: based on each cluster C k The prediction center and adaptive search radius are used to filter suppliers that fall within all clustered search areas from the global supplier database, and the supplier IDs are merged and deduplicated to form a set. The output of this level is usually saved as an Excel file, where each row corresponds to a company and lists all supplier IDs in its candidate pool.

[0109] Level 4: Precise recommendation generation, its input is the enterprise feature matrix F output from Level 1. c and supplier feature matrix F s And the candidate supplier pool S for each enterprise in the third-level output. c The process is handled by the precise matching and recommendation module; the output of this stage is the final recommendation list. TopK(c) represents the Top-K supplier IDs recommended by enterprise c and their matching scores; the output of this level can be integrated into the output file of the third level or saved separately, and used as input to the visualization module to generate the final report and charts.

[0110] In this embodiment, the specific data flow and module execution order of the four-level processing architecture are as follows:

[0111] The first level (feature representation learning level) of data input and output: This level accepts the original relational data file and the supplier feature data file as input; wherein, the original relational data file contains at least user_id (enterprise identifier) ​​and item_id (supplier identifier); the supplier feature data file contains d-dimensional static attribute vectors of all suppliers; this level calls the graph neural network feature learning module and performs the following steps to generate a feature vector file: 1) Construct an enterprise-supplier bipartite graph G based on the original relational data file, wherein user_id is mapped to enterprise nodes and item_id is mapped to supplier nodes to construct edges; 2) Load the supplier feature data as the initial features of supplier nodes, and set the initial features of enterprise nodes to zero vectors; 3) Train G using a degree-normalized graph convolutional network model, with the training objective being to minimize the mixed loss function. in and 4) After training, save the final output features (enterprise node features and supplier node features) of all nodes in the model as a feature matrix file.

[0112] Data input and output at the second level (geographic pattern analysis level): This level accepts enterprise coordinate files and supplier coordinate files as input. This level calls the geospatial clustering and analysis module and performs the following steps to generate a clustering result file: 1) Load enterprise coordinates and supplier coordinates, and establish a mapping from the enterprise to its historical supplier set based on the original relational data file; 2) For each enterprise, calculate the geographic distance matrix between all its historical suppliers, and perform agglomerative hierarchical clustering using a preset threshold (e.g., 800km) to form geographic clusters; 3) For each geographic cluster, use an optimized locator GNN model to predict its center coordinates and calculate dynamic indicators including R80 radius, gravity fraction, and coefficient of variation; 4) Summarize all cluster information for each enterprise (including cluster ID, predicted center coordinates, various dynamic indicators, and a list of supplier IDs within the cluster) and save it as a timestamped Excel file.

[0113] The third level (candidate space construction level) input and output: The input of this level is the clustering result Excel file generated in the second level; this level calls the adaptive candidate pool construction module and performs the following steps to generate the candidate pool file: 1) Parse the clustering result file and read the predicted center coordinates and R80 radius of each geographical cluster of each enterprise; 2) Determine the search radius for each cluster according to the adaptive radius calculation formula and filter out all suppliers falling within the radius; 3) Merge the filtering results of all clusters under the same enterprise, remove duplicate suppliers, and form the final candidate supplier pool of the enterprise; 4) Save the candidate pool information of each enterprise (including company ID, candidate supplier ID list, candidate pool size, etc.) as another Excel file with a timestamp.

[0114] The fourth level (precise recommendation generation level) data input and output: The input of this level includes the feature matrix file generated in the first level and the candidate pool file generated in the third level; this level calls the precise matching recommendation module and performs the following steps to generate the final recommendation: 1) Load the enterprise feature matrix F c and supplier feature matrix F s ;2) Load the candidate supplier pool S for each enterprise c 3) For each enterprise c, calculate its feature vector f. c With S c 4) Calculate the cosine similarity of each supplier's feature vector and select the top K suppliers with the highest similarity as the recommendation results; 5) Integrate the recommendation results with the candidate pool information and update them to the candidate pool Excel file, or save them as a new recommendation result file as the final output of the system.

[0115] This invention also proposes a geospatial-constrained intelligent recommendation method for supply chains, comprising the following steps:

[0116] Step 1: Load supplier characteristic data and historical relationship data to construct an enterprise-supplier relationship graph;

[0117] Step 2: Use a degree-normalized adaptive feature-aware graph convolutional neural network model to learn feature representations of enterprises and suppliers, and generate feature preference vectors;

[0118] Step 3: Based on the supplier's geographical coordinates, use a hierarchical clustering algorithm to perform cluster analysis and identify the geographical distribution pattern of the supply chain;

[0119] Step 4: Calculate the adaptive search radius for each company and build a geographically constrained pool of candidate suppliers;

[0120] Step 5: Within the candidate pool, recommend Top-K suppliers for each company based on feature cosine similarity;

[0121] Step 6: Generate visualization reports and statistical analysis results.

[0122] In this embodiment, the graph neural network training in step 2 further includes the following optimization steps: 1) Random negative sample generation strategy: During model training, a dynamic negative sample generation method based on the number of positive samples is adopted; for each positive sample edge existing in historical transaction relationships, a supplier node is randomly sampled within the range of non-historical transaction relationships to generate the corresponding negative sample edge; the number of negative samples and the number of positive samples are kept in a preset ratio to ensure the balance of training data; 2) Hybrid loss function training: When training the graph convolutional network model based on degree normalization, the contrastive loss and feature reconstruction loss are optimized simultaneously; the contrastive loss is used to distinguish between positive and negative sample edges; the feature reconstruction loss ensures the reconstruction capability of supplier features; 3) Model selection and saving mechanism: The model performance is evaluated periodically during training, and when the recommendation recall rate on the validation set reaches the historical best, the current model parameters are saved; after training, the saved best model parameters are loaded to generate the final feature representation.

[0123] like Figure 1 As shown in the figure, this embodiment provides a detailed introduction to the design, construction, and operation of the proposed intelligent recommendation system for the supply chain based on geospatial constraints and feature matching. Figure 1 The system's overall architecture and data flow are demonstrated, with its processing workflow following the logic of "parallel feature learning and spatial analysis → spatial constraints → precise matching → visualization." The specific implementation includes the following steps:

[0124] 1) Raw data collection and preprocessing

[0125] This step prepares the basic data for system operation. Three types of raw data need to be collected:

[0126] Historical transaction relationship data: at least includes the fields company_id (unique enterprise identifier) ​​and supplier_id (supplier identifier), used to build an enterprise-supplier relationship network.

[0127] Supplier feature data: This includes static attributes of all suppliers, such as supplier_id, firm_type (company type), firm_size (company size), and listed (whether publicly listed). This data needs to be cleaned and standardized (e.g., Min-Max normalization) to form a d-dimensional initial feature vector for each supplier.

[0128] Geographic coordinate data: Contains the precise geographical location of all businesses and suppliers, including at least the latitude and longitude fields.

[0129] 2) Specific implementation of the graph neural network feature learning module, the goal of which is to learn the deep feature representations of enterprises and suppliers in the transaction network.

[0130] Constructing a firm-supplier bipartite graph: Define the graph structure G=(V, E, F). Where the node set V=V c ∪V s V c V is a set of enterprise nodes. s Let V be the set of supplier nodes. Based on historical transaction data, if enterprise c and supplier s have a transaction record, then establish an edge e=(c,s), and all edges constitute the edge set E. For each supplier node v... s V s Assign an initial eigenvector of dimension d. The initial features of enterprise nodes are uniformly set to d-dimensional zero vectors.

[0131] Constructing a degree-normalized adjacency matrix: First, construct a binary adjacency matrix. If (c,s)∈E, then A cs =1, otherwise 0. Calculate v for each enterprise node. c degree Assign weights to each edge (c,s). Finally, the weight matrix W is formed, and the normalized matrix used for graph convolution is calculated: Where D is the degree matrix. This represents the Hadamard product (element-by-element multiplication).

[0132] Construct and train an adaptive feature-aware model: Build a two-layer graph convolutional network (GCN).

[0133] First convolution operation:

[0134] Second convolution operation:

[0135] Where F is the initial feature matrix of all nodes, , H is a trainable weight matrix. (2) That is, the final output feature matrix F of all nodes. out The loss function used in this network is defined as follows: in and These are preset hyperparameters used to balance the two losses. The contrast loss L... contrast The aim is to bring together node pairs with transactional relationships and push away unrelated node pairs. Feature reconstruction loss L MSE : Ensure that the output features of the supplier nodes retain their original semantics.

[0136] In this embodiment, the following is taken =0.2, =0.8. Train using the Adam optimizer until convergence. After training, save the enterprise feature matrix F. C and supplier feature matrix F S .

[0137] 3) Specific implementation of the geospatial clustering and analysis module, which independently analyzes the geographical distribution pattern of each enterprise's supplier network.

[0138] Extracting a subset of enterprise suppliers: For each enterprise c, based on historical transaction data, obtain the set S of all its cooperating suppliers. c and the corresponding set of coordinates.

[0139] Hierarchical clustering for geographic cluster identification: For the set of supplier coordinates of company c, calculate the spherical geographic distance between each pair of nodes (e.g., using the Haversine formula). An agglomerative hierarchical clustering algorithm is employed, with a preset distance threshold. Clustering was performed using distances of 800km and the "fully connected" method. Distances less than [a certain value] were [clustered]. Furthermore, suppliers connected through linking criteria are grouped into the same geographical cluster, ultimately leading to S c Divided into m geographical clusters {C1, C2, ..., C m}

[0140] Predict cluster centers and calculate dynamic indicators for each geographical cluster C. k :

[0141] Center prediction: Using a dedicated optimized locator GNN model, taking the coordinates of all suppliers within the cluster as input, predicts the center coordinates of the cluster. The network is trained using the geometric center of the coordinates within the cluster as the supervision signal.

[0142] Calculate dynamic geographic metrics: Based on the prediction center, calculate s for each supplier within the cluster. i Distance d to the center i .

[0143] R80 radius : Distance set {d i The 80th percentile of} represents the main range of the cluster.

[0144] Gravity fraction (k) : The attenuation coefficient =0.005, the larger the value, the more compact the cluster.

[0145] Coefficient of variation (CV) （k） :cv （k） =std_dist （k） / avg_dist （k） This is used to quantify the degree of dispersion of the distance.

[0146] Other statistics: including mean distance avg_dist （k） Standard deviation distance std_dist （k） minimum distance min_dist （k） Maximum distance max_dist （k） median distance （k） .

[0147] 4) Adaptive candidate pool construction, such as Figure 1 As shown, this module receives the results from modules 2 and 3 and performs spatial constraint fusion.

[0148] Calculate the adaptive search radius: for each geographic cluster C of enterprise c k Based on its dynamic indicators, the adaptive search radius is calculated according to the following priority rules. :

[0149]

[0150] Where clamp(x,a,b) is the constraint function, and in this embodiment, R is set to... min = 50km, R max = 500km, =100km.

[0151] Geospatial filtering constructs a candidate pool: for each supplier s in the global supplier database j (coordinates (lat) j ,lon j This involves determining whether a given cluster falls within the search range of any geographical cluster of enterprise C. In other words, if a cluster C exists... k This makes the spherical distance Then s j Incorporate into the candidate pool. Collect all qualified suppliers, merge and remove duplicates to form a geographically constrained candidate supplier pool specific to enterprise C. .

[0152] 5) Precise matching recommendation: This module performs efficient feature matching within the candidate pool.

[0153] Feature similarity calculation: For target company c, from feature matrix F C Obtain its feature vector f c Traverse its candidate pool Each supplier in j From the characteristic matrix F S Obtain its feature vector f sj Calculate the cosine similarity:

[0154]

[0155] Sorting and generating Top-K recommendations: [This involves] sorting the candidate pool. All suppliers are ranked according to their similarity score. c,j Sort the lists in descending order. Based on the system's preset recommendation quantity K (e.g., K=10), select the top K suppliers from the sorted list to form the final recommendation list TopK(c)={s (1) ,s (2) ,...,s (K)}

[0156] 6) Visual output: This module generates intuitive visualizations to support decision-making.

[0157] (a) Generate a geographic map and display the following features in layers overlaid in the same view:

[0158] Target company location (icon marker).

[0159] Its historical supplier distribution (scatter plot, can be colored by cluster).

[0160] Predicted center locations for each geographic cluster (specially marked).

[0161] A semi-transparent circular area (representing the search range) is drawn with each prediction center as the center and the adaptive radius as the radius.

[0162] Top-K recommended supplier positions (highlighted, such as a red star).

[0163] (b) Generate structured analysis reports: Automatically generate two reports:

[0164] Detailed enterprise-level report: Lists all dynamic metrics for each geographic cluster in the enterprise in tabular form, including ID, prediction center, adaptive radius, R80 radius, gravity fraction, cluster members, and Top-K recommendation list.

[0165] System-level summary report: Summarizes the overall situation of this operation, including the total number of enterprises processed, the total number of clusters identified, and the statistical characteristics of the candidate pool size (mean, median, quantiles, etc.).

[0166] Through the above design, a supply chain intelligent recommendation system based on geospatial constraints and feature matching, as described in this invention, is constructed, enabling accurate, practical, and interpretable intelligent supplier recommendations. The system effectively integrates the deep mining capabilities of graph neural networks for complex relationships with the quantification capabilities of geospatial analysis for real-world constraints, achieving an optimal balance between "precise matching" and "geographical feasibility" through an adaptive candidate pool mechanism. The visualization module greatly enhances the transparency and decision support value of the recommendation results. This system helps enterprises quickly and scientifically locate high-quality potential partners in the global supply chain network, optimize supply chain layout, and improve resilience and competitiveness.

[0167] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0172] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0173] This invention is not limited to the preferred embodiments described above. Anyone inspired by this invention can derive other forms of intelligent supply chain recommendation systems and methods based on geospatial constraints and feature matching. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A supply chain intelligent recommendation method based on geospatial constraints and feature matching, characterized in that, Specifically, the following steps are included: Step 1: Construct an enterprise-supplier relationship graph based on supplier feature data and historical relationship data to generate a normalized matrix for graph convolution operations; the supplier feature data includes multidimensional static attributes of suppliers, and the historical relationship data includes historical transaction records between enterprises and suppliers; Step 2: Using the generated normalized matrix as input, a graph convolutional neural network model based on degree normalization adaptive feature perception is used to learn the feature representations of enterprises and suppliers, generating enterprise feature vectors and supplier feature vectors. Step 3: Based on the supplier's geographical coordinates, use a hierarchical clustering algorithm to perform cluster analysis and identify the geographical distribution pattern of the supply chain; Step 4: Based on the results of supply chain geographical distribution pattern recognition, calculate the adaptive search radius for each enterprise and construct a geographically constrained candidate supplier pool; Step 5: Within the candidate pool, recommend Top-K suppliers for each company based on the matching degree between the company's feature vector and the supplier's feature vector; Step 6: Generate visualization reports and statistical analysis results.

2. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 1, characterized in that, The process of constructing an enterprise-supplier relationship graph based on supplier feature data and historical relationship data to generate a normalized matrix for graph convolution operations is as follows: Define a graph structure G = (V, E, F), where V represents the set of nodes, and V = V c ∪V s V c V represents the set of enterprise nodes. s Let E represent the set of supplier nodes and E represent the set of edges. V c ×V s This represents the historical procurement transaction relationship between the enterprise and its suppliers; F represents the initial feature matrix of all nodes, initialized as follows: For each supplier node v s Assign an initial eigenvector of dimension d. ,vector The data originates from multidimensional static attribute data from suppliers, while each enterprise node v c initial features f c Set to a zero vector or initialize randomly; Construct a binary adjacency matrix based on the edge set E. Subsequently, v for each enterprise node is calculated separately. c Each supplier node v s in-degree deg(v) c ), deg(v s ), and for each enterprise node v c Pointing to supplier node v s edge Assign weights For each of the supplier nodes v s Pointing to enterprise node v c edge Assign weights To construct the edge weight matrix W; finally, to generate the normalized matrix used for graph convolution. Where D is the degree matrix, It represents the Hadamardi (or Hadama) stack.

3. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 2, characterized in that, The graph convolutional neural network model based on degree normalization and adaptive feature perception learns feature representations of enterprises and suppliers, generating enterprise feature vectors and supplier feature vectors, as detailed below: The network structure of the graph convolutional neural network model is composed of two graph convolutional layers stacked sequentially. The first layer graph convolution operation is defined as follows: ,in This is the first layer of trainable weight matrix. For graph convolution operations, This is the output matrix of the first layer graph convolution operation; The convolution operation of the second layer graph is defined as follows: ,in This is the trainable weight matrix for the second layer. Let be the output matrix of the second layer graph convolution operation; then the graph convolutional neural network model outputs the final d-dimensional feature representation matrix of all nodes. .

4. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 3, characterized in that, The training of the graph convolutional neural network model uses a hybrid loss function L. total : in, To compare the losses, For the feature reconstruction loss part, and These are preset hyperparameters used to balance the two losses. The contrast loss portion L contrast The aim is to distinguish between positive and negative sample edges: for each positive sample edge that exists in historical transaction relationships Randomly sample a supplier node from outside the scope of historical transaction relationships to generate a corresponding negative sample edge. The number of negative samples and the number of positive samples are kept in a preset ratio, among which Indicates the relationship with enterprise node v c Supplier nodes with historical transactions, Indicates the relationship with enterprise node v c Supplier nodes with no historical transactions, compute node pairs as well as The model outputs cosine similarity of feature vectors and And adopt the loss function: To bring positive sample pairs closer together and push negative sample pairs further apart, For enterprise node v c The model output features, , Supplier nodes Model output features and supplier nodes The model output features; The feature reconstruction loss part L MSE The aim is to ensure that the model can effectively reconstruct the original features of suppliers from graph structure information, defined as the model output features of supplier nodes. Initial characteristics of supplier nodes Mean square error between: , where N s The number of suppliers is represented by s, which is the iteration parameter.

5. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 1, characterized in that, The method involves using hierarchical clustering algorithms to perform cluster analysis based on supplier geographic coordinates to identify supply chain geographic distribution patterns, as detailed below: Based on the latitude and longitude coordinates of all supplier nodes Calculate the actual distance between each pair of pairs. To construct a symmetric geographic distance matrix ; For each enterprise node v c Extract enterprise node v c The set S of all historically associated supplier nodes c and S c The corresponding coordinate subset; based on the sub-distance matrix corresponding to the coordinate subset, a preset linking method and distance threshold are used. The agglomerative hierarchical clustering algorithm is used to divide S into groups. c Divided into m geographical clusters This ensures that the distance between any two suppliers within the same geographical cluster does not exceed [a certain value]. ; For each cluster C k Cluster C k Using the coordinates of all suppliers within the cluster as input, the cluster C is predicted through a GNN model. k center coordinates The training objective of the GNN model is to minimize the regression loss between the prediction center and the geometric center of all supplier coordinates within the cluster. For cluster C k Its forecasting center calculates the value of each member supplier's data. i Distance d to the prediction center i And based on the distance d i Calculate dynamic geographic indicators: R80 radius: , This indicates the radius that covers 80% of the cluster members; Gravity fraction: ,in This is the attenuation coefficient, used to measure the compactness of the cluster around the center; Coefficient of variation (CV): cv = std_dist / avg_dist, used to quantify the dispersion of the distance; Other statistics include mean distance (avg_dist), standard deviation distance (std_dist), minimum distance (min_dist), maximum distance (max_dist), and median distance (median_dist).

6. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 5, characterized in that, Based on the results of supply chain geographical distribution pattern recognition, an adaptive search radius is calculated for each enterprise to construct a geographically constrained candidate supplier pool, as detailed below: According to each enterprise v c Each geographic cluster C k Dynamic geographic indicators are used to calculate cluster C according to a preset priority logic. k Corresponding adaptive search radius ; For a specific target enterprise, based on each cluster C k Predicted center coordinates and the corresponding adaptive search radius Set circular area filter criteria: for any coordinate in the global supplier database... suppliers j If supplier s j With any cluster center The spherical geographical distance between them satisfies Then supplier s j Preliminarily included in enterprise v c The range of candidates to be considered; Collect all suppliers that meet the circular area screening criteria of the target company and merge them to form a preliminary candidate supplier set. For sets Perform deduplication to remove duplicate suppliers, ultimately generating a candidate supplier pool for the target company. .

7. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 6, characterized in that, Cluster C is computed according to a preset priority logic. k Corresponding adaptive search radius The specific calculation rules are as follows: in, For cluster C k The radius of R80, For cluster C k The average distance from each member supplier to the prediction center. The system default radius is used, and `clamp(·)` is a constraint function that ensures the final radius value is limited to the preset minimum value R. min With the maximum value R max The interval formed by [R] min , R max ]Inside.

8. The intelligent supply chain recommendation method based on geospatial constraints and feature matching according to claim 1, characterized in that, Within the candidate pool, Top-K suppliers are recommended for each enterprise based on the degree of matching between the enterprise's feature vector and the supplier's feature vector, as detailed below: For the target company and its pool of candidate suppliers The enterprise node v corresponding to the target enterprise c The model outputs feature vectors With the candidate supplier pool Each candidate supplier j Corresponding supplier nodes The model outputs feature vectors A one-to-one comparison is performed, and the cosine similarity is used to calculate the matching score between the two. c,j After completing the candidate pool After calculating the matching scores of all suppliers, all candidate suppliers are sorted in descending order based on their matching scores, and the top K suppliers are selected to form the final Top-K recommendation set for the target company. .

9. A supply chain intelligent recommendation system based on geospatial constraints and feature matching, characterized in that, The system is implemented using the intelligent supply chain recommendation method as described in any one of claims 1-8, including a graph neural network feature learning module, a geospatial clustering and analysis module, a precise matching recommendation module, and a visualization module; The graph neural network feature learning module is used to construct an enterprise-supplier relationship graph based on supplier feature data and historical relationship data, generate a normalized matrix for graph convolution operation, and use an adaptive feature-aware graph convolutional neural network model based on degree normalization to learn the feature representations of enterprises and suppliers, and generate enterprise feature vectors and supplier feature vectors. The geospatial clustering and analysis module is used to perform clustering analysis based on the supplier's geographical coordinates using a hierarchical clustering algorithm to identify the geographical distribution patterns of the supply chain. The precise matching and recommendation module is used to calculate an adaptive search radius for each enterprise based on the results of the supply chain geographical distribution pattern recognition, construct a geographically constrained candidate supplier pool, and recommend Top-K suppliers for each enterprise within the candidate pool based on the degree of matching between the enterprise feature vector and the supplier feature vector. The visualization module is used to generate visualization reports and statistical analysis results.

10. A supply chain intelligent recommendation system based on geospatial constraints and feature matching according to claim 9, characterized in that, The visualization module generates content including geographic visualization maps and analysis reports that display enterprise location, supplier clusters, adaptive radius, and recommendation results; The geographic visualization map is a visual map generated for each target enterprise, covering the historical geographic distribution and clustering results of its suppliers. The visualization map supports zooming and panning operations, and the overlay display content within the same coordinate system includes: 1) The geographical location of the target company marked with a specific icon; 2) Distribution points of all historical suppliers, displayed using the first category of markers; 3) Geographical clusters of different suppliers of the target company, distinguished by different colors; 4) The predicted center location of each geographic cluster, marked with a special symbol; 5) A circular area is drawn with each prediction center as the center and the corresponding adaptive search radius as the radius, and the geographic coverage of the candidate pool is displayed with semi-transparent filling; 6) The locations of the Top-K recommended suppliers selected from the candidate pool, highlighted with the second-category marker; The analysis report consists of two parts: 1) Enterprise-level detailed report: For a single enterprise, it provides a detailed list of each geographic cluster's ID, prediction center latitude and longitude, adaptive search radius, R80 radius, gravity fraction, coefficient of variation, average distance, number of suppliers within the cluster, and a list of specific supplier IDs. 2) System-level summary report: Summarizes the overall situation of system processing, including the statistical characteristics of the total number of enterprises processed, the total number of geographical clusters, the size of the candidate pool, and the distribution of enterprises by different supplier number ranges.

Images