Driver cargo intelligent pairing method based on intelligent algorithm
By constructing a two-part diagram of driver and cargo and using graph neural network and gated residual strategy, the problem of inefficient matching between driver and cargo in traditional methods is solved, and efficient and accurate logistics resource management is achieved.
Patent Information
- Application Number
- CN202510544464.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-28
- Publication Date
- 2025-08-08
AI Technical Summary
Traditional drivers and goods matching methods are inefficient, prone to inaccurate matching due to human errors in judgment, difficult to cope with the scheduling needs during peak periods, and traditional models are difficult to process complex multimodal data and drivers or goods newly added to the platform, resulting in low quality of recommended results.
By obtaining the historical interactive data of the driver and the goods, a graph neural network is used for message transmission and aggregation, a gated residual strategy training model is introduced, and a multi-layer perceptron is used for accurate matching degree calculation to generate the optimal recommendation.
It significantly improves the accuracy of drivers' matching with goods, reduces resource waste and delivery delays, adapts to order volume fluctuations, optimizes vehicle utilization and transportation routes, and reduces air driving rates and waiting time.
Smart Images
Figure CN120450328A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent cargo pairing, and in particular to a method for intelligent driver-cargo pairing based on an intelligent algorithm. Background Art
[0002] In the modern logistics industry, cargo scheduling and driver matching are critical for improving transportation efficiency and reducing logistics costs. Traditional methods for matching drivers with cargo rely primarily on manual scheduling, simple rules, or traditional recommendation algorithms, which often have significant limitations. Manual rule-based matching requires dispatchers to manually compare cargo and driver information, such as cargo weight, volume, and transportation distance, with the driver's vehicle type, load capacity, and current location. This approach is not only inefficient but also prone to inaccurate matching due to human error. For example, the timeliness of cargo delivery may not match the driver's route, resulting in delayed delivery. Furthermore, with increasing order volume, the complexity of manual scheduling increases exponentially, making it difficult to cope with peak scheduling demands. Traditional recommendation systems often use collaborative filtering or content-based approaches. While these methods have achieved good results in e-commerce and video recommendations, they are often difficult to apply in logistics scenarios. Collaborative filtering methods rely on historical user behavior data, which, in the context of matching drivers with cargo, suffers from a significant cold-start problem. This means that for drivers who have just joined the platform or newly released cargo sources, there is a lack of sufficient historical data support, resulting in low-quality recommendations. At the same time, most traditional models can only process simple features, such as a driver's past order history or basic cargo information, and struggle to integrate complex multimodal data, such as geographic location, cargo type, and driver travel trajectory. In recent years, the rapid development of deep learning, particularly graph neural networks, has brought new opportunities for solving the problem of intelligently matching drivers and cargo. GNNs are capable of processing complex graph-structured data, establishing rich connections between nodes (drivers, cargo) and edges (historical order relationships), and capturing potential matching relationships. However, traditional GNN methods typically only support single-directional or simple rule-based message passing, making them difficult to achieve ideal results with multi-dimensional and multi-relational logistics data. Summary of the Invention
[0003] The purpose of the present invention is to provide a driver-cargo intelligent matching method based on an intelligent algorithm to solve the problems raised in the above background technology.
[0004] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for intelligent matching of drivers and cargo based on an intelligent algorithm, comprising the following steps:
[0005] S1: Obtain driver, cargo, and historical order data, perform data cleaning, feature extraction, and bipartite graph structure construction;
[0006] S2: Based on the historical interaction data between drivers and goods, a bipartite graph about drivers and goods is constructed;
[0007] S3: Aggregate information about node neighborhoods in a bipartite graph through a transfer mechanism to generate a high-dimensional node embedding representation. A gated residual strategy is introduced to train the model to predict the optimal matching of drivers and cargo.
[0008] S4: Calculate similarity based on node embedding, predict the matching degree between the driver and the goods, and recommend goods to the driver according to the matching degree;
[0009] S5: Generate recommendations based on real-time driver and cargo data and provide the best recommendations.
[0010] Furthermore, in step S1, driver, cargo, and historical order data are obtained, and driver characteristics and cargo characteristics are extracted. Driver characteristics include driver ID, geographic location, vehicle type, and load capacity; cargo characteristics include cargo ID, cargo type, weight, volume, origin, and destination.
[0011] The data cleaning process involves removing duplicate records, filling in missing values, and correcting erroneous data; feature extraction uses natural language processing technology to parse text information and uses encoding to quantify non-numeric data.
[0012] Furthermore, in step S2, based on the driver and cargo matching in the historical order data as positive samples and the unrelated driver and cargo pairing as negative samples, a bipartite graph of drivers and cargo is constructed, where the edge features in the graph are historical transportation records;
[0013] The steps for constructing a bipartite graph about drivers and goods are:
[0014] Set nodes for drivers, and the driver node set is represented as {d1,d2,…,d m}, where m represents the number of drivers, nodes are set for cargo, and the cargo node set is represented as {g1,g2,…,g n}, where n represents the quantity of goods;
[0015] For the driver node, there is a driver feature set X d ={X d1 ,X d2 ,…,X dm}, the driver feature set represents the set of driver features of m drivers; for the cargo node, there is a cargo feature set X g ={X g1 ,X g2 ,…,X gn}, the cargo feature set represents a set of cargo features of n cargoes;
[0016] For edge feature ei,j , the edge feature e i,j Indicates driver d i With cargo g i Historical transportation records between i With cargo g i If there is a historical transportation record between them, then the edge feature e is recorded. i,j =1, otherwise record edge feature e i,j =0, and then we get the bipartite graph of drivers and goods G = (V, E), where V represents the driver and E represents the goods.
[0017] Furthermore, in step S3, for the Graph Transformer model design, in the graph neural network, the message transmission process of each layer is divided into three steps: message generation, message propagation, and message aggregation. The specific design steps are as follows:
[0018] Message generation: For any edge e in the bipartite graph G of drivers and goods i,j Connected nodes v i and v j , the generated message is expressed as:
[0019]
[0020] e c,ij =W c,e e ij +b c,e ;
[0021] Where c is the number of attention heads, For node v i Features, For node v j Features, providing trainable parameters b c，e Convert to query vector and key vector e c,ij Under the c-th attention head, the edge feature e i,j The encoded edge feature vector;
[0022] Message propagation: The generated message is propagated in the bipartite graph G of drivers and goods, so that each node can receive messages from neighboring nodes. This process is achieved by calculating the attention scores between nodes through the multi-head attention mechanism and adding edge features e i,j As additional information, according to the formula:
[0023]
[0024] Message aggregation: After obtaining the multi-head attention of the bipartite graph G of drivers and goods, perform message aggregation on each edge from node j to node i in the bipartite graph G of drivers and goods according to the following formula:
[0025]
[0026] in and is a trainable parameter, || is the connection operation of c attention heads, is the characteristic of node j, which is obtained through linear change N(i) is the set of all neighbor nodes of all nodes i, is the attention coefficient between node i and its neighbor node j.
[0027] For each attention head C, first traverse all neighbor nodes j of node i. For each neighbor node j, the previously calculated With edge feature e c，ij Add, and then use the attention coefficient The weighted sum is taken to obtain the feature representation of node i under each head after aggregating neighbor information. Finally, the results obtained under C attention heads are concatenated to obtain the updated feature representation of node i at layer l+1.
[0028] To prevent the model from being over-smoothed, a gated residual connection can be used, as follows:
[0029]
[0030] in, and is the weight matrix and bias term, is the input of the current layer, r i (l) is the gating signal of the i-th node in the current layer. is the output of the gating function, which controls information flow. LayerNorm represents a normalization operation. Each training session uses a graph data loader to batch load the features of each node and the relationships between them. The training is optimized by calculating a loss function. Supervised learning is used for training. Based on the labels of historical paired data, the cross-entropy loss is used to calculate the difference between the model's predicted output and the true label. The Adam optimizer is used for gradient descent and parameter updates. By adaptively adjusting the learning rate, the Adam optimizer dynamically adjusts the update step size of each parameter during training, accelerating convergence and avoiding overfitting. After model training is complete, the performance of the model is evaluated by calculating the similarity or accuracy between the predicted results and the actual labels. In practical applications, the trained model will be used to infer new driver-cargo pairings.
[0031] Furthermore, in step S4, the inner product of node embeddings is calculated to obtain a recommended candidate set, and the MLP is used to accurately score the driver and cargo pairs in the candidate set. The best matching solution is recommended based on the ranking of the scores.
[0032] The optimal matching recommendation scheme is as follows: the node features updated by the Graph Transformer model will be further input into the recommendation module, and personalized product recommendations will be achieved through similarity calculation of node embeddings. First, the node feature representation vector H = {h1, h2, ..., h N}, where h i Represents node v i Feature representation in high-dimensional embedding space;
[0033] For driver node v d and cargo node v g , the cosine similarity analysis method is used to obtain the driver node v d and cargo node v g The similarity Sim(h d , h g ):
[0034]
[0035] Furthermore, in step S5, after obtaining the similarity scores of all candidate cargo nodes and the target driver node, the similarity scores are arranged in descending order by sorting, and the top N cargo nodes with the highest similarity scores are output to form a recommendation list.
[0036] Compared with the existing technology, the beneficial effects achieved by the present invention are: through systematic data cleaning and feature engineering, multi-source heterogeneous data such as driver's vehicle attributes, historical driving trajectory and cargo type, volume, and timeliness requirements are converted into structured features, and a driver-cargo bipartite graph model is constructed. Historical transportation records are used as edge features to explicitly model the interaction history between the two. At the same time, through the message passing mechanism of the graph neural network, potential associations such as geographical location complementarity and transportation habit fit are implicitly mined, breaking the limitation of traditional methods that only rely on static attribute matching. The system can comprehensively evaluate the matching degree from multiple dimensions such as transportation route planning, vehicle space utilization, and timeliness priority, significantly improving the matching accuracy in complex scenarios and reducing resource waste and delivery delays caused by vehicle type mismatch and route detours; the GraphTransformer model is introduced to dynamically adjust the importance weights of neighbor nodes through the self-attention mechanism, so that the model can adaptively focus on driver groups with high matching degree according to the transportation needs of different goods, avoiding the one-size-fits-all feature aggregation defects of traditional models. Furthermore, the combination of gated residual connections and layer normalization effectively mitigates the vanishing gradient and feature confusion issues in deep network training, ensuring that the model captures global structural features while preserving individual node characteristics when processing large-scale networks of drivers and shipments. This dynamic, hierarchical feature learning capability enables the system to generalize better to newly added drivers or shipments, addressing the issue of unstable recommendation quality in cold-start scenarios. This is particularly applicable to logistics platforms experiencing large order volume fluctuations and frequent new capacity additions. Based on the high-dimensional node embedding vectors generated by the graph neural network, similarity calculations are used to quickly identify potential matching pairs, narrowing the computational scope. A multi-layer perceptron (MLP) is then used to refine the scoring, incorporating real-time dynamic factors. Finally, the optimal recommendation list is output based on the overall matching score. This hierarchical approach preserves the graph model's ability to understand global structure while giving the system the flexibility to respond to dynamic scenarios in real time. This allows logistics platforms to dynamically adjust matching strategies based on supply and demand fluctuations, achieving efficient operations. This reduces driver idle driving rates and cargo waiting times, optimizes vehicle loading rates and transportation route planning, and fundamentally improves the utilization of logistics resources, helping enterprises achieve cost reduction and efficiency gains. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0038] Figure 1 This is a flow chart of a driver-cargo intelligent matching method based on an intelligent algorithm of the present invention. DETAILED DESCRIPTION
[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0040] See also Figure 1 The present invention provides a technical solution: a driver-cargo intelligent matching method based on an intelligent algorithm, comprising the following steps:
[0041] S1: Obtain driver, cargo, and historical order data, perform data cleaning, feature extraction, and bipartite graph structure construction;
[0042] S2: Based on the historical interaction data between drivers and goods, a bipartite graph about drivers and goods is constructed;
[0043] S3: Aggregate information about node neighborhoods in a bipartite graph through a transfer mechanism to generate a high-dimensional node embedding representation. A gated residual strategy is introduced to train the model to predict the optimal matching of drivers and cargo.
[0044] S4: Calculate similarity based on node embedding, predict the matching degree between the driver and the goods, and recommend goods to the driver according to the matching degree;
[0045] S5: Generate recommendations based on real-time driver and cargo data and provide the best recommendations.
[0046] In step S1, driver, cargo, and historical order data are obtained, and driver and cargo characteristics are extracted. Driver characteristics include driver ID, geographic location, vehicle type, and load capacity; cargo characteristics include cargo ID, cargo type, weight, volume, origin, and destination.
[0047] The data cleaning process involves removing duplicate records, filling in missing values, and correcting erroneous data; feature extraction uses natural language processing technology to parse text information and uses encoding to quantify non-numeric data.
[0048] In step S2, based on the driver and cargo matching in the historical order data as positive samples and the unrelated driver and cargo matching as negative samples, a bipartite graph of drivers and cargo is constructed, where the edge features in the graph are historical transportation records;
[0049] The steps for constructing a bipartite graph about drivers and goods are:
[0050] Set nodes for drivers, and the driver node set is represented as {d1,d2,…,d m}, where m represents the number of drivers, nodes are set for cargo, and the cargo node set is represented as {g1,g2,…,gn}, where n represents the quantity of goods;
[0051] For the driver node, there is a driver feature set X d ={X d1 ,X d2 ,…,X dm}, the driver feature set represents the set of driver features of m drivers; for the cargo node, there is a cargo feature set X g ={X g1 ,X g2 ,…,X gn}, the cargo feature set represents a set of cargo features of n cargoes;
[0052] For edge feature e i,j , the edge feature e i,j Indicates driver d i With cargo g i Historical transportation records between i With cargo g i If there is a historical transportation record between them, then the edge feature e is recorded. i,j =1, otherwise record edge feature e i,j =0, and then we get the bipartite graph of drivers and goods G = (V, E), where V represents the driver and E represents the goods.
[0053] In step S3, for the Graph Transformer model design, in the graph neural network, the message transmission process of each layer is divided into three steps: message generation, message propagation, and message aggregation. The specific design steps are as follows:
[0054] Message generation: For any edge e in the bipartite graph G of drivers and goods i,j Connected nodes v i and v j , the generated message is expressed as:
[0055]
[0056] Where c is the number of attention heads, For node v i Features, For node v j Features, providing trainable parameters b c，e Convert to query vector and key vector e c,ij Under the c-th attention head, the edge feature e i,j The encoded edge feature vector;
[0057] Message propagation: The generated message is propagated in the bipartite graph G of drivers and goods, so that each node can receive messages from neighboring nodes. This process is achieved by calculating the attention scores between nodes through the multi-head attention mechanism and adding edge features e i,j As additional information, according to the formula:
[0058]
[0059] Message aggregation: After obtaining the multi-head attention of the bipartite graph G of drivers and goods, perform message aggregation on each edge from node j to node i in the bipartite graph G of drivers and goods according to the following formula:
[0060]
[0061] Where || is the connection operation of c attention heads, and is a trainable parameter;
[0062] To prevent the model from being over-smoothed, a gated residual connection can be used, as follows:
[0063]
[0064] in, and is the weight matrix and bias term, is the input of the current layer, r i (l) is the gating signal of the i-th node in the current layer. is the output of the gating function, which controls information flow. LayerNorm represents a normalization operation. Each training session uses a graph data loader to batch load the features of each node and the relationships between them. The training is optimized by calculating a loss function. Supervised learning is used for training. Based on the labels of historical paired data, the cross-entropy loss is used to calculate the difference between the model's predicted output and the true label. The Adam optimizer is used for gradient descent and parameter updates. By adaptively adjusting the learning rate, the Adam optimizer dynamically adjusts the update step size of each parameter during training, accelerating convergence and avoiding overfitting. After model training is complete, the performance of the model is evaluated by calculating the similarity or accuracy between the predicted results and the actual labels. In practical applications, the trained model will be used to infer new driver-cargo pairings.
[0065] In step S4, the inner product of node embeddings is calculated to obtain a recommended candidate set. The MLP is then used to accurately score the driver and cargo pairs in the candidate set, sort them by score, and recommend the best matching solution.
[0066] The optimal matching recommendation scheme is as follows: the node features updated by the Graph Transformer model will be further input into the recommendation module, and personalized product recommendations will be achieved through similarity calculation of node embeddings. First, the node feature representation vector H = {h1, h2, ..., h N}, where h i Represents node v i Feature representation in high-dimensional embedding space;
[0067] For driver node v d and cargo node v g , the cosine similarity analysis method is used to obtain the driver node v d and cargo node v g The similarity Sim(h d , h g ):
[0068]
[0069] In step S5, after obtaining the similarity scores of all candidate cargo nodes and the target driver node, the similarity scores are sorted in descending order, and the top N cargo nodes with the highest similarity scores are output to form a recommendation list.
[0070] Example 1: First, data preprocessing is performed. The platform collects driver information, including their location, vehicle type and load capacity, such as the upper limit of a van's load capacity and the urban areas where it is usually active; cargo information covers type, weight, transportation starting point and destination, such as a batch of home appliances needing to be transported from a warehouse in place A to a shopping mall in place B. The data cleaning process will eliminate duplicate records, such as multiple identical information for the same driver; missing vehicle load data will be supplemented by the average load of vehicles of the same type; obvious errors in cargo weight will be corrected to ensure the accuracy of basic data. For address text information, natural language processing technology is used to parse out specific cities and regions, and non-numerical information such as vehicle type is converted into computable codes, such as assigning different feature labels to "refrigerated trucks" and "flatbed trucks".
[0071] Next, a network of connections between drivers and shipments is constructed. Based on historical transport records, if a driver has previously transported a certain shipment, a connection is established between the two, forming a bipartite graph reflecting this historical partnership. To improve model training, negative examples are randomly generated for driver and shipment combinations that have never collaborated before. For example, pairing a driver who has never been to a certain location with shipments from that location enriches the training data.
[0072] During the model training phase, a graph neural network incorporating a self-attention mechanism is employed. Within each network layer, driver and cargo nodes "pay attention" to each other's neighboring nodes, dynamically adjusting attention weights based on historical collaboration frequency, geographic proximity, and other factors. For example, drivers and cargo that frequently transport similar goods in the same area are assigned higher association weights. Position encoding is also introduced to capture the relative position of nodes in the network, allowing the model to learn the interaction patterns of nodes in different regions. During training, historically collaborative pairs serve as positive samples, while random pairs serve as negative samples. An optimization algorithm is used to adjust model parameters, enabling the model to accurately distinguish between valid and invalid matches. To prevent information from becoming blurred during multi-layer transmission, a gating mechanism is implemented to control the flow of information within each layer, ensuring that key features are not diluted.
[0073] Finally, matching recommendations are made. After model training is complete, a vector representation containing multi-dimensional features is generated for each driver and cargo. By calculating the similarity between the two vectors, for example, comparing the overlap between the driver's frequently used routes and the cargo transportation routes, and the matching degree between the vehicle load and the cargo weight, the scores of the candidate matching pairs are obtained. All candidate cargoes are sorted from high to low by score, and the cargoes with the highest scores are selected and recommended to the driver to form the final matching solution. For example, if a driver is located in Location C and has no load, the system will prioritize recommending cargoes with a starting point near Location C and a load requirement that matches the vehicle's capacity, thus achieving efficient allocation of transportation resources. The entire process is completed automatically, significantly improving the efficiency and accuracy of logistics scheduling.
[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.
Claims
1. A method for intelligent matching of drivers and cargo based on an intelligent algorithm, characterized by: The method comprises the following steps: S1: Obtain driver, cargo, and historical order data, perform data cleaning, feature extraction, and bipartite graph structure construction; S2: Based on the historical interaction data between drivers and goods, a bipartite graph about drivers and goods is constructed; S3: Aggregate information about node neighborhoods in a bipartite graph through a transfer mechanism to generate a high-dimensional node embedding representation. A gated residual strategy is introduced to train the model to predict the optimal matching of drivers and cargo. S4: Calculate similarity based on node embedding, predict the matching degree between the driver and the goods, and recommend goods to the driver according to the matching degree; S5: Generate recommendations based on real-time driver and cargo data and provide the best recommendations.
2. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 1, characterized in that: In step S1, driver, cargo, and historical order data are obtained, and driver and cargo characteristics are extracted. Driver characteristics include driver ID, geographic location, vehicle type, and load capacity; cargo characteristics include cargo ID, cargo type, weight, volume, origin, and destination. The data cleaning process involves removing duplicate records, filling missing values, and correcting erroneous data; Feature extraction uses natural language processing techniques to parse text information and uses encoding to quantify non-numeric data.
3. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 2, characterized in that: In step S2, based on the driver and cargo matching in the historical order data as positive samples and the unrelated driver and cargo pairing as negative samples, a bipartite graph about drivers and cargo is constructed, and the edge features in the graph are historical transportation records.
4. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 3, characterized in that: The steps for constructing a bipartite graph about drivers and goods are: Set nodes for drivers, and the driver node set is represented as {d1,d2,…,d m }, where m represents the number of drivers, nodes are set for cargo, and the cargo node set is represented as {g1,g2,…,g n }, where n represents the quantity of goods; For the driver node, there is a driver feature set X d ={X d1 ,X d2 ,…,X dm }, the driver feature set represents the set of driver features of m drivers; for the cargo node, there is a cargo feature set X g ={X g1 ,X g2 ,…,X gn }, the cargo feature set represents a set of cargo features of n cargoes; For edge feature e i,j , the edge feature e i,j Indicates driver d i With cargo g i Historical transportation records between i With cargo g i If there is a historical transportation record between them, then the edge feature e is recorded. i,j =1, otherwise record edge feature e i,j =0, and then we get the bipartite graph of drivers and goods G = (V, E), where V represents the driver and E represents the goods.
5. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 4, characterized in that: In step S3, for the Graph Transformer model design, in the graph neural network, the message transmission process of each layer is divided into three steps: message generation, message propagation, and message aggregation. The specific design steps are as follows: Message generation: For any edge e in the bipartite graph G of drivers and goods i,j Connected nodes v i and v j , the generated message is expressed as: have been c,ij =W c，e have been ij +b c，e ; Where c is the number of attention heads, For node v i Features, For node v j Features, providing trainable parameters W c，e , b c,e Convert to query vector and key vector e c,ij Under the c-th attention head, the edge feature e i,j The encoded edge feature vector; Message propagation: The generated message is propagated in the bipartite graph G of drivers and goods, so that each node can receive messages from neighboring nodes. The attention scores between nodes are calculated through the multi-head attention mechanism, and the edge feature e is added. i,j As additional information, calculate the attention score between nodes in is the query vector, is the key vector, N(i) is the set of neighbor nodes of node i, so j∈N(i). c,ij is the feature vector of edge (i, j) under the c-th attention head, and the attention coefficient is calculated Its value is between 0 and 1, so that when performing subsequent message aggregation, the neighbor node information can be reasonably aggregated according to the weight of the attention coefficient. Message aggregation: After obtaining multi-head attention of the bipartite graph G of drivers and goods, message aggregation is performed on each edge from node j to node i in the bipartite graph G of drivers and goods.
6. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 5, characterized in that: Use the gated residual strategy, which includes: in, and are the weight matrices and bias terms, is the feature input of node i in layer l, generating the residual term r of node i in layer l i (l) , r i (l) The information of the current node features after a specific linear transformation is retained. The updated features of the message aggregation Residual term r i (l) And the difference between the two is spliced, and then the weight matrix is used The matrix performs linear transformation on the splicing vector and then activates it through the sigmoid function to obtain the gating coefficient according to right and r i (l) Perform weighted fusion, then normalize with LayerNorm and activate with ReLU to obtain the updated features of node i in layer l+1 7. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 6, characterized in that: In step S4, the recommended candidate set is obtained by calculating the inner product of the node embedding, and the driver and cargo pairs in the candidate set are accurately scored. They are sorted according to the scores and the best matching solution is recommended.
8. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 7, characterized in that: The optimal matching recommendation scheme is as follows: the node features updated by the Graph Transformer model will be further input into the recommendation module, and personalized product recommendations will be achieved through similarity calculation of node embeddings. First, the node feature representation vector H = {h1, h2, ..., h N }, where h i Represents node v i Feature representation in high-dimensional embedding space; For driver node v d and cargo node v g , the cosine similarity analysis method is used to obtain the driver node v d and cargo node v g The similarity Sim(h d ,h g ):
9. The method for intelligent driver-cargo pairing based on an intelligent algorithm according to claim 8, characterized in that: In step S5, after obtaining the similarity scores of all candidate cargo nodes and the target driver node, the similarity scores are sorted in descending order, and the top N cargo nodes with the highest similarity scores are output to form a recommendation list.