CN-121996852-A - Supply chain intelligent recommendation system and method based on geospatial constraint and feature matching

CN121996852ACN 121996852 ACN121996852 ACN 121996852ACN-121996852-A

Abstract

The invention provides a supply chain intelligent recommendation system and method based on geospatial constraint and feature matching, wherein the method comprises the steps of S1, constructing an enterprise-provider relation graph according to provider feature data and historical relation data and generating a normalized matrix, S2, generating enterprise feature vectors and provider feature vectors by using a graph convolution neural network model based on adaptive feature perception of degree normalization by taking the normalized matrix as input, S3, carrying out clustering analysis by using a hierarchical clustering algorithm based on provider geographic coordinates, and identifying a supply chain geographic distribution mode, S4, building a candidate provider pool of geographic constraint for each enterprise according to a supply chain geographic distribution mode identification result, and S5, recommending Top-K providers for each enterprise in the range of the candidate pool based on the matching degree of the enterprise feature vectors and the provider feature vectors. The method effectively fuses the geographic space constraint and the feature matching, and remarkably improves the accuracy, the efficiency and the interpretability of the supply chain recommendation.

Inventors

ZHANG XIAOCHEN
WU QINGLONG
TANG LINGLING
LIN GUIHUI
WANG HAOYANG
ZHENG JUNWEI

Assignees

福建师范大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. The intelligent supply chain recommendation method based on geospatial constraint and feature matching is characterized by comprising the following steps of: Step 1, an enterprise-provider relationship graph is constructed according to provider characteristic data and historical relationship data, so as to generate a normalized matrix for graph convolution operation, wherein the provider characteristic data comprises multidimensional static attributes of a provider, and the historical relationship data comprises enterprise-provider historical transaction records; Step 2, taking the generated normalized matrix as input, and learning the characteristic representation of the enterprise and the supplier by using a graph convolution neural network model based on the self-adaptive characteristic perception of degree normalization to generate an enterprise characteristic vector and a supplier characteristic vector; Step 3, based on the geographic coordinates of the suppliers, carrying out cluster analysis by using a hierarchical clustering algorithm, and identifying a geographic distribution mode of a supply chain; Step 4, calculating an adaptive search radius for each enterprise according to the identification result of the geographic distribution mode of the supply chain, and constructing a candidate supplier pool of geographic constraint; Step 5, recommending Top-K suppliers for each enterprise based on the matching degree of the enterprise feature vector and the supplier feature vector in the candidate pool range; And 6, generating a visual report and a statistical analysis result.
2. The intelligent recommendation method for a supply chain based on geospatial constraint and feature matching according to claim 1, wherein the enterprise-provider relationship graph is constructed according to provider feature data and historical relationship data to generate a normalized matrix for graph convolution operation, specifically as follows: Defining a graph structure g= (V, E, F), where V represents a set of nodes, v=v c ∪V s ,V c represents a set of enterprise nodes, V s represents a set of vendor nodes, E represents a set of edges, E V c ×V s represents the purchasing transaction relation of enterprises and suppliers, F represents the initial feature matrix of all nodes, and the initialization mode is as follows: Each provider node v s is assigned a d-dimensional initial feature vector Vector quantity Multidimensional static attribute data derived from the vendor, and the initial feature f c of each enterprise node v c is set to zero vector or randomly initialized; Construction of binary adjacency matrix from edge set E Then, the degree of ingress deg (v c )、deg(v s ) for each enterprise node v c , each provider node v s , respectively, is calculated and for each edge directed by enterprise node v c to provider node v s Assigning weights For each edge directed by provider node v s to enterprise node v c Assigning weights To construct an edge weight matrix W, and finally generating a normalized matrix for graph convolution Wherein D is a degree matrix, Representing the hadamard product.
3. The intelligent recommendation method for a supply chain based on geospatial constraint and feature matching according to claim 2, wherein the model of a graph roll-up neural network using adaptive feature awareness based on degree normalization learns feature representations of enterprises and suppliers, and generates enterprise feature vectors and supplier feature vectors, specifically as follows: the network structure of the graph roll neural network model is formed by sequentially stacking two layers of graph roll layers; The first layer of graph rolling operations are defined as: Wherein For the first layer of trainable weight matrix, In order to perform the convolution operation of the drawing, An output matrix for a first layer of graph convolution operations; The second layer of graph rolling operations are defined as: Wherein For the second layer of trainable weight matrix, Outputting a final d-dimensional characteristic representation matrix of all nodes for the second-layer graph rolling operation by the graph rolling neural network model 。
4. A geospatial constraint and feature matching based supply chain intelligent recommendation method in accordance with claim 3 wherein said graph rolling neural network model is trained using a hybrid loss function L total : Wherein, the In order to contrast the lost portion, The lost portion is reconstructed for the feature, And Is a preset super parameter and is used for balancing the two targets of loss; The contrast loss portion L contrast is intended to distinguish positive sample edges from negative sample edges for each positive sample edge present in the historical trade relationship Randomly sampling a provider node in a range of non-historical transaction relationships to generate a corresponding negative sample edge The number of negative samples is kept in a preset ratio to the number of positive samples, wherein Representing a provider node having historical transactions with enterprise node v c , Provider node representing no history transactions with enterprise node v c , compute node pairs And Cosine similarity of model output feature vector of (a) And And adopts a loss function: to pull the positive pair of samples closer and push the negative pair of samples farther, wherein, The features are output for the model of enterprise node v c , 、 Respectively, provider nodes Model output features and provider nodes of (a) Model output characteristics of (2); The feature reconstruction loss portion L MSE is intended to ensure that the model can effectively reconstruct the original features of the provider from the graph structure information, defined as the model output features of the provider nodes Initial features with vendor node Mean square error between: where N s represents the number of suppliers and s is the traversal parameter.
5. The intelligent supply chain recommendation method based on geospatial constraint and feature matching according to claim 1, wherein the clustering analysis is performed by using a hierarchical clustering algorithm based on geographic coordinates of suppliers to identify a geographic distribution pattern of a supply chain, specifically comprising the following steps: longitude and latitude coordinates based on all provider nodes Calculating the field distance between every two To construct a symmetrical geographical distance matrix ; For each enterprise node v c , extracting coordinate subsets corresponding to all the historically associated provider node sets S c and S c of the enterprise node v c , and adopting a preset link mode and a distance threshold value based on a sub-distance matrix corresponding to the coordinate subsets Dividing S c into m geographic clusters by a condensation hierarchical clustering algorithm So that the distance between any suppliers in the same geographic cluster is not more than ; For each cluster C k , the center coordinates of cluster C k are predicted by the GNN model with the coordinates of all suppliers within cluster C k as input The training objective of the GNN model is to minimize regression loss between the prediction center and the geometric center of all vendor coordinates within the cluster; For cluster C k and its prediction center, calculate the distance d i of each member provider s i to the prediction center, and calculate the dynamic geographic index from the distance d i : Radius R80: , representing the radius covering 80% of the members of the cluster; gravity score: Wherein As an attenuation coefficient, the method is used for measuring the compactness of the cluster around the center; cv=std_dist/avg_dist, for quantifying the degree of dispersion of the distance; Other statistics include average distance avg_dist, standard deviation distance std_dist, minimum distance min_dist, maximum distance max_dist, median distance media_dist.
6. The intelligent supply chain recommendation method based on geospatial constraint and feature matching according to claim 5, wherein the method calculates an adaptive search radius for each enterprise according to the result of supply chain geographic distribution pattern recognition, and constructs a candidate provider pool of geographic constraint, specifically as follows: According to the dynamic geographic index of each geographic cluster C k of each enterprise v c , calculating the self-adaptive search radius corresponding to cluster C k according to preset priority logic ; For a certain target enterprise, the predicted center coordinates based on each cluster C k Corresponding adaptive search radius Setting round area screening conditions, namely, setting any coordinate in a global supplier base as If the supplier s j , the supplier s j and any cluster center The spherical geographic distance between them satisfies Then provider s j is initially taken into consideration as a candidate for enterprise v c ; Collecting all suppliers meeting the circular area screening condition of the target enterprise, and combining to form a preliminary candidate supplier set Pair aggregation Performing a deduplication operation, removing duplicate suppliers therein, and finally generating a candidate supplier pool of the target enterprise 。
7. The intelligent recommendation method for the supply chain based on geospatial constraint and feature matching according to claim 6, wherein the adaptive search radius corresponding to cluster C k is calculated according to a preset priority logic The specific calculation rules are as follows: Wherein, the For the R80 radius of cluster C k , For each member provider in cluster C k to the average value of the distance to the prediction center, The default radius preset for the system, clip (·) is a constraint function, ensuring that the final radius value is limited within the interval [ R min , R max ] formed by the preset minimum value R min and maximum value R max .
8. The intelligent supply chain recommendation method based on geospatial constraint and feature matching according to claim 1, wherein in the candidate pool range, top-K suppliers are recommended for each enterprise based on the matching degree of the enterprise feature vector and the supplier feature vector, specifically as follows: For a target enterprise and candidate provider pool thereof Outputting the model of the enterprise node v c corresponding to the target enterprise to the feature vector With a pool of candidate suppliers The provider node corresponding to each candidate provider s j in the network Model output feature vector of (a) Performing a successive comparison, calculating a matching score c,j between the two by using cosine similarity, and completing the candidate pool After calculating the matching scores of all suppliers in the enterprise, arranging all candidate suppliers in a descending order according to the matching scores, and selecting the first K suppliers to form a final Top-K recommendation set of the target enterprise 。
9. A supply chain intelligent recommendation system based on geospatial constraint and feature matching, which is characterized in that the system is realized by adopting the supply chain intelligent recommendation method according to any one of claims 1-8, and comprises a graphic neural network feature learning module, a geospatial clustering and analyzing module, an accurate matching recommendation module and a visualization module; the graph neural network feature learning module is used for constructing an enterprise-provider relationship graph according to provider feature data and historical relationship data, generating a normalized matrix for graph convolution operation, learning feature representations of enterprises and providers by using a graph convolution neural network model based on adaptive feature perception of degree normalization, and generating enterprise feature vectors and provider feature vectors; The geospatial clustering and analyzing module is used for carrying out clustering analysis by using a hierarchical clustering algorithm based on geographic coordinates of suppliers and identifying a geographic distribution mode of a supply chain; the accurate matching recommendation module is used for calculating the self-adaptive search radius for each enterprise according to the identification result of the geographic distribution mode of the supply chain, constructing a candidate provider pool of geographic constraint, and recommending Top-K providers for each enterprise based on the matching degree of the enterprise feature vector and the provider feature vector in the range of the candidate pool; The visualization module is used for generating a visualization report and a statistical analysis result.
10. The geospatial constraint feature matching based supply chain intelligent recommendation system of claim 9 wherein said visualization module generates content including a geographic visualization map showing business locations, vendor clusters, adaptive radii and recommendation results and an analysis report; The geographic visual map is a visual map which is generated for each target enterprise and covers the geographic distribution and clustering result of the historical suppliers, the visual map supports scaling and translation operations, and the superposition display content of the map under the same coordinate system comprises: 1) The geographic location of the target business marked with a particular icon; 2) Distribution points of all historical suppliers displayed with a first type of mark; 3) Different provider geographic clusters of target businesses differentiated by different colors; 4) A predicted central location of each geographic cluster marked with a special symbol; 5) Taking each prediction center as a circle center, and corresponding self-adaptive searching radius as a radius to draw a circular area, and filling and displaying the geographic coverage of the candidate pool in a semitransparent manner; 6) Top-K recommended supplier locations screened from the candidate pool highlighted with a second type of label; the analysis report content includes two parts of content: 1) Enterprise-level detail report, namely, for a single enterprise, detailing the ID of each geographic cluster, the longitude and latitude of a prediction center, the self-adaptive search radius, the radius R80, the gravity fraction, the variation coefficient, the average distance, the number of suppliers in the cluster and a specific supplier ID list; 2) Summarizing the overall situation processed by the system, wherein the overall situation comprises the number of processed overall enterprises, the number of overall geographic clusters, the statistical characteristics of the size of the candidate pool and the enterprise distribution situation among different provider number intervals.

Description

Supply chain intelligent recommendation system and method based on geospatial constraint and feature matching Technical Field The invention belongs to the fields of big data analysis, enterprise digital transformation and the like, and particularly relates to an intelligent supply chain recommendation system and method based on geospatial analysis and feature matching. Background With the complexity of the globalization supply chain, enterprises face challenges in how to efficiently select suitable suppliers. The traditional supplier selection method mainly depends on manual experience and a simple scoring system, and has the problems of low efficiency, strong subjectivity, difficulty in large scale and the like. In recent years, recommendation systems based on machine learning have been remarkably successful in the fields of electronic commerce and the like, but the application in supply chain management still faces a plurality of challenges, namely 1) the multi-dimensional characteristics and geographic constraints of suppliers need to be considered simultaneously, 2) the supply chain relation has a complex network structure, and 3) the accuracy and the practicability of recommendation need to be balanced. In the prior art, some methods only recommend based on the characteristics of suppliers, neglecting geographic factors, and some methods only consider geographic proximity, neglecting quality matching of suppliers. How to organically combine geographic constraint and feature matching and construct an accurate and practical intelligent recommendation system becomes a technical problem to be solved in the field. Disclosure of Invention The invention aims to provide a supply chain intelligent recommendation system and method based on geospatial analysis and feature matching, and designs an intelligent recommendation system with layered processing and module cooperation by creatively introducing a technical framework of 'deep fusion of geospatial constraint and network feature learning'. The system is characterized in that a self-adaptive feature perception model based on degree normalization is utilized to distill high-quality potential feature vectors of enterprises and suppliers from historical transaction network and supplier attributes, functional preference and capability images of the enterprises and the suppliers are accurately captured, a supplier geographic distribution mode is independently and deeply analyzed, geographic morphological features of each enterprise supply chain are quantized through hierarchical clustering and dynamic index calculation, further, a self-adaptive geographic search radius is dynamically generated for each enterprise based on the quantized geographic features, a geographic constraint candidate pool with degree of tightness is built, matching precision and geographic feasibility are ingeniously balanced, finally, accurate feature similarity matching is conducted in the candidate pool, a final recommendation list is generated, and decision basis is visually presented through a visual means. The technical path realizes the normal form transition from global blind matching to local accurate butt joint, and the practicability and the interpretability of the result are obviously improved while the recommendation correlation is ensured. In order to achieve the above purpose, the technical scheme of the invention is as follows: A supply chain intelligent recommendation method based on geospatial constraint and feature matching specifically comprises the following steps: Step 1, an enterprise-provider relationship graph is constructed according to provider characteristic data and historical relationship data, so as to generate a normalized matrix for graph convolution operation, wherein the provider characteristic data comprises multidimensional static attributes of a provider, and the historical relationship data comprises enterprise-provider historical transaction records; Step 2, taking the generated normalized matrix as input, and learning the characteristic representation of the enterprise and the supplier by using a graph convolution neural network model based on the self-adaptive characteristic perception of degree normalization to generate an enterprise characteristic vector and a supplier characteristic vector; Step 3, based on the geographic coordinates of the suppliers, carrying out cluster analysis by using a hierarchical clustering algorithm, and identifying a geographic distribution mode of a supply chain; Step 4, calculating an adaptive search radius for each enterprise according to the identification result of the geographic distribution mode of the supply chain, and constructing a candidate supplier pool of geographic constraint; Step 5, recommending Top-K suppliers for each enterprise based on the matching degree of the enterprise feature vector and the supplier feature vector in the candidate pool range; And 6, generating a visual report and a statistical ana