A cloud platform-based logistics transportation personalized recommended path planning method and system

Abstract

The application discloses a logistics transportation personalized recommendation path planning method and system based on a cloud platform, belongs to the technical field of logistics transportation, and constructs a heterogeneous traffic graph model by fusing road sections, geographic interest points and traffic event information; collects historical transportation task records of multiple users, generates user semantic intention labels based on delivery behaviors, and maps the user semantic intention labels to traffic graph nodes and edge attributes to form a semantic constraint graph structure; uses a heterogeneous graph neural network to jointly model user intentions and path accessibility, generates a path score sequence; collects abnormal events after the completion of a transportation task, dynamically updates a semantic intention label mapping relationship based on a reverse alignment error; and deploys the model and the mapping module on a cloud platform, combines an edge device to realize real-time acquisition of traffic states, realizes online generation and rapid pushing of path recommendations, and effectively improves the personalized matching degree and real-time response capability of path recommendations, and is suitable for intelligent logistics path planning in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of logistics and transportation technology, specifically to a cloud-based personalized route planning method and system for logistics and transportation. Background Technology

[0002] Current mainstream logistics route recommendation methods mostly rely on static road networks and current traffic information for route planning, lacking in-depth modeling of users' transportation intentions and individual preferences. Especially in scenarios involving cross-regional, complex road network structures and abnormal weather, the recommendation results often show delayed response or do not conform to user preferences, leading to reduced transportation efficiency.

[0003] Some existing methods attempt to incorporate machine learning or graph neural networks (GNNs) to process dynamic traffic data, but most are limited to path optimization within a single-region graph structure, making it difficult to effectively extend to cross-regional path scenarios. Furthermore, current models often focus only on temporal and spatial feature changes, ignoring users' varying needs regarding path semantics in different task contexts (such as avoiding specific locations or preferring a particular delivery order), resulting in a lack of "intent-driven" personalization in recommended paths.

[0004] Therefore, there is an urgent need for a personalized path recommendation method that can integrate user semantic intent tags, abnormal event feedback mechanisms, and cross-regional heterogeneous graph modeling capabilities, especially to achieve large-scale heterogeneous data training and real-time path push based on cloud platform architecture, thereby improving the system's intelligence and robustness. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized recommendation route planning method and system for logistics transportation based on a cloud platform, so as to solve the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a cloud-based personalized route planning method for logistics transportation, comprising:

[0007] Construct a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic points of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges.

[0008] Collect historical transportation task records from multiple users, generate user semantic intent tags based on user delivery behavior and historical preferences, and map the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure;

[0009] Based on the constructed semantically constrained traffic map, a heterogeneous graph neural network model is used to jointly model the user intent and path reachability of each transportation task, generating a path score sequence.

[0010] After each transportation task is completed, abnormal transportation events that occurred during its execution are collected, the causal deviation relationship between the recommended path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviation path and the real path.

[0011] The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended route results that meet user preferences online.

[0012] Preferably, the construction of a heterogeneous traffic graph model containing multiple transportation area nodes and path edges includes:

[0013] Acquire vector road data, point of interest distribution data, and historical traffic event data covering multiple transportation areas, and parse their spatial location and temporal attributes respectively to initialize and form corresponding road segment nodes, geographic point of interest nodes, and traffic event nodes;

[0014] Based on road topology and interest point distribution density, geographical adjacency edges between road nodes and interest point nodes are constructed, and edge weights are calculated using a regional proximity function based on multi-scale grid partitioning.

[0015] Based on the time lag relationship between historical traffic events and route execution results, an event influence edge is constructed between event nodes and road nodes, and a decreasing weighting mechanism is used to characterize the degree of influence of event persistence on route accessibility.

[0016] Multidimensional attribute encoding is performed on all nodes and edge structures, including road level, point of interest type, and event urgency, to generate a graph structure feature matrix that supports heterogeneous graph neural network inputs.

[0017] Preferably, the generation of user semantic intent tags based on user delivery behavior and historical preferences includes:

[0018] The user's historical transportation trajectory is sliced, and a multi-dimensional behavior sequence is constructed by combining the task time period, transportation type and delivery scenario context tags;

[0019] A behavior feature encoding network with attention mechanism weighting is used to represent each behavior sequence, generating a high-dimensional embedding vector containing spatial path preference, temporal response characteristics and behavior frequency;

[0020] A contrastive learning mechanism is introduced, using the behavior sequences of the same user in different scenarios as positive sample pairs and the behavior sequences of different users as negative sample pairs.

[0021] Dense clustering is performed in the embedding space to identify multiple cluster centers with stable transport preferences, and each center is assigned an interpretable semantic intent label.

[0022] Preferably, the generated path score sequence includes:

[0023] The traffic map is divided into multiple subgraphs according to the structural roles of various nodes, and an embedding space is constructed for each subgraph. Initial modeling of node feature alignment and semantic association is achieved through meta-path guidance.

[0024] The semantic intent label of the user's current transportation task is used as the graph state initialization factor, and the set of nodes that are strongly related to the intent label are assigned context activation weights to form the starting point of the semantically driven graph structure.

[0025] A multi-view heterogeneous graph attention mechanism is adopted to aggregate different types of neighbor information from the structure view, preference view and event interference view respectively, and dynamically capture the semantic accessibility and behavioral fit of the path.

[0026] The historical score dependency of candidate paths is regularized, and finally a user-customized path score sequence is output for ranking.

[0027] Preferably, the dynamic updating of the semantic intent mapping relationship based on the reverse alignment error between the deviation path and the true path includes:

[0028] The recommended path and the actual execution path are compared at the node level to construct a spatial-temporal alignment matrix, identify path segments with significant deviations, and extract the corresponding path deviation node set.

[0029] For each node in the deviation node set, calculate its cumulative frequency of participation in the deviation path in historical tasks, and establish a node deviation confidence model in combination with the task context to measure the credibility of its semantic mapping.

[0030] For nodes with confidence scores below a threshold, their original positive preference labels are replaced with context-sensitive labels, and the node semantic attribute vector is updated.

[0031] Preferably, the method of combining local traffic conditions obtained in real time with edge computing devices to generate recommended routes that match user preferences online includes:

[0032] The local traffic status reported by edge devices is integrated into the cloud semantic traffic graph in the form of graph patches, and the graph structure is kept in sync with the current traffic environment through a node-level incremental update mechanism.

[0033] Based on user intent labels, we activate associated nodes in the graph and construct a subgraph input for the task, and use a heterogeneous graph neural network model for multi-hop message passing and path candidate generation.

[0034] By introducing dynamic factors based on user context, including task urgency and cargo characteristics, candidate path scores are weighted and ranked to generate real-time path recommendation results.

[0035] This invention also provides a cloud-based personalized route planning system for logistics transportation, comprising:

[0036] Heterogeneous graph construction module: Constructs a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic point of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges.

[0037] Semantic intent generation module: Collects historical transportation task records of multiple users, generates user semantic intent tags based on user delivery behavior and historical preferences, and maps the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure;

[0038] Heterogeneous graph modeling module: Based on the constructed semantically constrained traffic graph, the module uses a heterogeneous graph neural network model to jointly model the user intent and path reachability of each transportation task, generating a path score sequence.

[0039] Path execution feedback module: After each transportation task is completed, abnormal transportation events that occur during its execution are collected, the causal deviation relationship between the recommended path and the actual path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviated path and the actual path.

[0040] Cloud-edge collaborative path recommendation module: The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform, and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended path results that meet user preferences online.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] 1. This invention constructs a heterogeneous traffic graph structure that integrates roads, points of interest, and traffic events, and introduces user semantic intent tags for graph structure constraints. This enables the path recommendation model to not only consider physical accessibility but also deeply reflect users' personalized preferences. Through joint modeling using heterogeneous graph neural networks, multi-dimensional path evaluation from structural, semantic, and behavioral layers is achieved, significantly improving the matching accuracy and user satisfaction of recommended paths.

[0043] 2. This invention introduces a transportation execution feedback mechanism, achieving closed-loop optimization of the path recommendation system's behavior through reverse alignment error analysis and adaptive updating of semantic intent tags. Combining cloud-based heterogeneous modeling capabilities with the real-time perception capabilities of edge devices, the system can achieve low-latency, highly adaptable personalized online path recommendations, significantly improving the intelligence level and responsiveness of logistics transportation in complex traffic environments. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0045] Figure 1 This is a mind map of the method of the present invention.

[0046] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1, please refer to Figure 1 As shown in this embodiment, a personalized route planning method for logistics transportation based on a cloud platform includes:

[0049] Construct a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic points of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges.

[0050] Collect historical transportation task records from multiple users, generate user semantic intent tags based on user delivery behavior and historical preferences, and map the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure;

[0051] Based on the constructed semantically constrained traffic map, a heterogeneous graph neural network model is used to jointly model the user intent and path reachability of each transportation task, generating a path score sequence.

[0052] After each transportation task is completed, abnormal transportation events that occurred during its execution are collected, the causal deviation relationship between the recommended path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviation path and the real path.

[0053] The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended route results that meet user preferences online.

[0054] In this embodiment, to achieve personalized recommendations for logistics transportation routes, the system first constructs a heterogeneous traffic graph model. This model integrates multi-source traffic data and semantic information from multiple transportation areas, serving as the input foundation for subsequent graph neural network models. Specifically, it includes the following steps:

[0055] The system first accesses road vector data in OpenStreetMap format, Point of Interest (POI) data, and structured event data (such as accidents, construction, and traffic control) from a historical traffic event platform. The data is then spatially standardized to unify its projection coordinate system.

[0056] Road vector data is cut into line segments according to its topology and defined as road segment nodes;

[0057] For point-of-interest (POI) data, they are marked as geographic POI nodes according to their type (such as warehouses, gas stations, delivery stations, transportation hubs, etc.);

[0058] For traffic incident data, extract metadata such as the time of occurrence, affected area, and incident type to generate traffic incident nodes, and perform spatial mapping based on the location of the incident center point.

[0059] Each type of node is assigned a set of initial attribute vectors, including spatial coordinates, category code, timestamp, etc.

[0060] When constructing the edge structure between road nodes and points of interest (POIs), a multi-scale regional grid partitioning mechanism is introduced. The system divides the target city or region into multiple grid regions (such as 500m×500m, 1km×1km), calculates the density gradient of POIs and the similarity of road node distribution between grids, and forms the "geographical adjacency".

[0061] Edge weights are defined using the following proximity function: Where d(i,j) is the Euclidean distance between nodes i and j. The difference in interest point density between grids is represented by α and β, which are adjustment parameters. The system connects node pairs with weights greater than a set threshold as geographical adjacency edges.

[0062] To address the correlation between traffic event nodes and road nodes, the system employs a time window-based hysteresis response modeling method. The specific operation is as follows:

[0063] Iterate through each traffic event node and calculate its impact on the traffic speed and actual route execution (such as delays and detours) of surrounding roads within several time windows after its occurrence (such as 0-15 minutes, 15-30 minutes, etc.).

[0064] Add event-affected edges to the road segments affected by the event in each path, and use the following dynamic weighting function: ;in, The type and level are determined, and λ is the attenuation coefficient. This mechanism can be used to simulate the short- and long-term disruptions to traffic conditions caused by the duration and propagation range of events.

[0065] Finally, to support the training of heterogeneous graph neural network models, the system uniformly encodes all nodes and edges into a graph structure feature matrix. Each node embeds the following attributes:

[0066] Road nodes: road class (expressway / national highway / local road), average speed, number of lanes;

[0067] Point of Interest (POI) nodes: POI type, business density, and service time;

[0068] Event Nodes: Event Type (Accident / Construction), Impact Level, Duration;

[0069] Each edge is embedded with a type label (adjacency / influence) and its corresponding weight. All graph structure data is input into the model training module in Tensor format, enabling underlying graph data support for personalized path recommendation.

[0070] To enhance the personalization of logistics route recommendations, this embodiment proposes a semantic constraint graph structure modeling method that combines user transportation behavior characteristics with historical preferences. This method generates user semantic intent labels and embeds them into the nodes and edges of the traffic graph, thereby explicitly guiding the model to capture user behavior patterns during route planning and achieving personalized route optimization.

[0071] The system collects behavioral records from multiple users in past transportation tasks from the logistics platform's historical order database, including but not limited to:

[0072] Geographic coordinates of the starting point and the destination;

[0073] Transportation task time (time period, weekday / holiday);

[0074] Actual execution path (GPS trajectory data);

[0075] Selection of transfer stations and stops;

[0076] User behavior of actively deviating from the recommended path;

[0077] Task type (express delivery / cold chain / bulk goods, etc.) and cargo attributes.

[0078] The system cleans, denoises, and standardizes the above data to form a set of transportation task sequences that can be used for behavior analysis.

[0079] Each transportation task is represented as a structured sequence containing multiple fields, enhanced by task context (such as weather, region, and road grade). To extract latent semantic features that distinguish user preferences, the following user transportation intent embedding model is constructed:

[0080] Using a multi-layer Transformer structure, contextual feature modeling is performed for each behavior sequence to capture the temporal relationships and preference dependencies between behaviors;

[0081] A trajectory similarity attention mechanism is introduced to assign higher attention weights to key path nodes such as "repeatedly passing through specific road segments" and "preference for non-main roads";

[0082] The sequence embedding results for each user are summarized to obtain their transportation preference vectors in multiple context scenarios.

[0083] An embedding space is constructed using a contrastive learning method, with "the same user's behavior in different scenarios" as positive sample pairs and "similar task behaviors of different users" as negative sample pairs. The model is optimized so that the embedding space has the ability to distinguish users.

[0084] After embedding training is complete, dense clustering is performed in the embedding space. For each cluster center, interpretable semantics are extracted by combining the representative behavioral sequence, such as:

[0085] "Prefers to take detours along main urban roads";

[0086] "Prioritize highway + suburban transit warehouse";

[0087] "Delivery at fixed times every day + specific route dependence".

[0088] Define a semantic intent tag for each type of intent and build a tag index library.

[0089] The generated semantic intent labels are embedded into the heterogeneous traffic map, specifically as follows:

[0090] If a user prefers a certain type of road structure (such as avoiding urban road sections), their intent label is mapped to the "negative preference" attribute of a specific road node;

[0091] If users frequently select specific transit warehouses or points of interest as their stop locations, then add a "high priority" semantic intent tag to the corresponding point of interest node;

[0092] If certain event nodes (such as construction or traffic restrictions) have a significant impact on user path deviation, then a "sensitive label" is embedded on the event edge.

[0093] In this way, the original traffic map is enhanced with semantic intent, forming a semantic constraint graph structure, which can be used as a structured input to a graph neural network model to achieve path optimization reasoning explicitly driven by user behavior.

[0094] To effectively model user semantic intent and path reachability together, this embodiment proposes a path rating generation method based on a multi-view heterogeneous graph neural network. This method not only fully utilizes the structural and semantic information in the traffic map, but also guides graph state initialization through user behavior and introduces a path decoupling mechanism to solve the problem of redundant path inertia, thereby improving the quality of personalized recommendations.

[0095] First, the semantically constrained traffic graph that has been constructed is divided into several subgraph structures, each representing a different structural role relationship:

[0096] Structural subgraph: contains road segment nodes and their geographical adjacency edges, used to represent the physical accessibility structure;

[0097] Preference subgraph: contains interest point nodes and their attribute edges related to user preferences, such as "preferred warehouse" edges and semantic matching edges;

[0098] Disturbance subgraph: contains traffic event nodes and their associated "event impact edges" with path interference.

[0099] The subgraphs mentioned above are input into graph embedding encoding modules of different dimensions. Using a meta-path-guided method, such as "point of interest-road-point of interest" or "event-road-warehouse" path, a multi-semantic cross space is constructed, and the node features are semantically aligned to generate an initial embedding vector shared across subgraphs.

[0100] Traditional route models use "origin" or "location pair" as input, while this method uses the semantic intent tag of the user's current transportation task as input guidance. The system first activates tags according to task requirements, such as "avoid urban areas," "prioritize highways + fixed cargo space," and "time-sensitive priority."

[0101] For nodes that highly match the labels (such as road segments and points of interest that frequently appear under this preference category), activation weights are assigned to the graph, and semantic biases are superimposed on the initial embedding vector to form a task-specific semantic activation graph state. This skips the traditional "location-space" graph construction and directly forms the initial state of the user preference-driven graph structure.

[0102] After constructing a unified graph structure, the system introduces a three-view heterogeneous attention aggregation mechanism:

[0103] Structural view: Encodes physical adjacency relationships and calculates the structural consistency of road topological accessibility;

[0104] Preference view: Encodes semantic edges of interest point nodes and extracts the influence of interest regions consistent with user preferences;

[0105] Disturbance view: Encodes the dynamic relationship between traffic event nodes and paths to identify risk areas that a task may encounter.

[0106] Each node calculates type attention weights based on the different type characteristics of its neighbors in the three views and performs weighted fusion to obtain a node representation with comprehensive "structure-semantics-perturbation" features.

[0107] To avoid the scoring model being affected by historical path inertia, the system is designed with a path dependency decoupling module:

[0108] The performance of candidate paths in historical tasks is compared to identify potential redundancy biases (such as "a certain path is always recommended first").

[0109] A regularization penalty term is introduced to reduce the weight of high-frequency paths and low-quality paths in the path score calculation;

[0110] Execute a cumulative scoring function for each candidate path: ;in Let be the final representation of the i-th node in the path. These are the weighting coefficients. The path historical deviation score is given, where λ is the adjustment parameter and n is the total number of nodes.

[0111] Finally, a personalized path rating sequence for the user's current task is generated, and the paths with the highest ratings are output as recommendations.

[0112] To enhance the self-correction capability of path recommendation systems within the user feedback loop, this embodiment proposes a dynamic semantic intent optimization method based on a reverse alignment error feedback mechanism. This method identifies semantic mapping deviation nodes by analyzing the behavioral differences between the recommended path and the actual path, and adaptively updates their corresponding labels and graph structures, thereby strengthening the recommendation system's ability to model users' true preferences.

[0113] After each transport mission is completed, the system records two path sequences:

[0114] The system recommends the following path: obtained by ranking the scores from a heterogeneous graph neural network.

[0115] Actual execution path: obtained by reconstructing vehicle trajectory data (such as GPS).

[0116] The system uses a node-level dynamic time warping algorithm to compare paths, construct a space-time alignment matrix between the recommended path and the actual path, and identify locations where there are significant deviations between path segments.

[0117] For areas exhibiting persistent deviations (such as users circumventing the same route or point of interest in multiple tasks), extract the corresponding set of nodes, denoted as the path deviation node set. And record relevant contextual information (time, region, event impact, etc.).

[0118] right For each node n, count the frequency with which it participated in the deviation path in the past N tasks. Furthermore, a node bias confidence model is constructed by combining consistency measures of the task context (such as whether they are from the same time period, the same type of goods, or the same type of intent labels): ;in, Here, σ is the context consistency coefficient, and σ is the Sigmoid normalization function. The output result is... This indicates the semantic credibility of a node. When... If the value is below a set threshold (e.g., 0.3), it is considered that the semantic intent currently bound to it is biased.

[0119] For nodes with low confidence, the system introduces an adaptive semantic intent label transfer mechanism, which specifically includes:

[0120] Label weight reduction: The weight of the original bound label (such as "preferred route") in the graph embedding is reduced, and its impact on the recommendation model is adjusted by updating the semantic attribute vector of the node;

[0121] Semantic hierarchy transfer: If the original label of a node is "high priority preference", it is downgraded to "conditional preference" or "context sensitive" label under the premise of repeated avoidance behavior, reflecting its non-general adaptability;

[0122] Category reclassification: By combining the semantic embedding space and changes in cluster centers, the category to which a node belongs in the semantic space is reassessed, and its classification label is updated synchronously in the label index.

[0123] These update operations are automatically completed using the K-nearest neighbor relocation algorithm in the embedding space, ensuring that label transfer takes into account both semantic consistency and behavioral matching.

[0124] In addition to updating node attributes, the system will also dynamically adjust the structural relationships of the off-target nodes in the semantic graph:

[0125] If a node frequently misleads the propagation of path scores, then an edge weight reduction operation is performed to reduce the intensity of information flow between it and other preferred nodes.

[0126] If a node is circumvented for a long time and has lost its semantic representativeness, a local node blocking mechanism can be triggered to temporarily remove it from the candidate path score in future tasks.

[0127] The updated node and edge structure are synchronously written into the heterogeneous graph database and used as input for the next round of task path planning model, thereby realizing a semantic graph closed-loop self-optimization mechanism based on reverse alignment error.

[0128] This embodiment proposes a cloud-based path scoring and recommendation method driven by real-time graph structure updates and user preference tags. By combining a heterogeneous graph neural network model with a cloud-edge collaborative communication mechanism, it achieves efficient and personalized path recommendation capabilities, which is particularly suitable for complex scenarios with drastic traffic dynamics or significant user preferences.

[0129] To ensure the input graph structure of the recommendation model is synchronized with the actual traffic conditions, the system employs an edge graph patching mechanism. Traffic data collected by edge devices (such as vehicle density, road congestion, and event broadcasts) is structured into local graph patches, with the following format:

[0130] New or modified nodes (such as new accident nodes, speed reduction nodes);

[0131] Changes in edge weights (such as an increase in road travel time);

[0132] Node status changes (e.g., changes in the business status of a point of interest).

[0133] The cloud-based graph database uses a node-level hash index matching mechanism to compare the differences between the patch and the original graph structure and perform incremental node and edge updates, avoiding the overhead of full graph reconstruction and thus achieving millisecond-level graph synchronization.

[0134] When a user initiates a route request, the system reads their current transportation task parameters and preference intent tags (such as "Avoid City Center", "Prioritize Expressway", "Fixed Delivery Station"), and selects nodes and edges that meet the following conditions from the real-time graph structure to construct a task subgraph:

[0135] Points of interest or road segments that are highly correlated with preference tags;

[0136] The starting and ending points and the main relay nodes on the reachable paths;

[0137] Traffic anomaly nodes that are currently affected by edge graph patches.

[0138] This subgraph serves as the input to the heterogeneous graph neural network model, ensuring that the model's computational scope is limited to regions strongly related to the task, thereby improving real-time performance and computational efficiency.

[0139] To achieve more refined path ranking, the system introduces a user task context factor weighting mechanism during the path scoring phase. This factor includes, but is not limited to:

[0140] Task urgency (e.g., whether it is a time-limited delivery);

[0141] Type of goods (e.g., perishable, heavy-duty);

[0142] User's current behavioral deviation record (such as recent continuous changes to recommended paths).

[0143] The system dynamically adds adjustment factors to the scoring function of candidate paths: ;in For the model's basic score, Let i be the function of the influence of the i-th context factor on the path. The weighting coefficients are assigned. After ranking the scores, the Top-N paths are selected as the recommended output.

[0144] To quickly transmit recommendation results to user terminals, the system is designed with an adaptive communication channel selection mechanism:

[0145] If the current network conditions are stable and the latency is low, the system will prioritize pushing the path data to the vehicle terminal via a 5G direct link.

[0146] If the network fluctuates greatly or the edge device is offline, the system will cache the path results at the local edge relay node and automatically send them back when the most recent connection is made, or wait for the device to come online before pushing them.

[0147] The recommended data format includes node sequences, turning instructions, expected times, and semantic annotations to ensure that the navigation system can directly parse and execute it.

[0148] Ultimately, after receiving the path recommendation, end users can obtain context-sensitive, real-time, and efficient personalized path guidance without waiting for large-scale graph updates and complex model inference.

[0149] Example 2, please refer to Figure 2 As shown in this embodiment, a cloud-based personalized route planning system for logistics transportation includes:

[0150] Heterogeneous graph construction module: Constructs a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic point of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges.

[0151] Semantic intent generation module: Collects historical transportation task records of multiple users, generates user semantic intent tags based on user delivery behavior and historical preferences, and maps the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure;

[0152] Heterogeneous graph modeling module: Based on the constructed semantically constrained traffic graph, the module uses a heterogeneous graph neural network model to jointly model the user intent and path reachability of each transportation task, generating a path score sequence.

[0153] Path execution feedback module: After each transportation task is completed, abnormal transportation events that occur during its execution are collected, the causal deviation relationship between the recommended path and the actual path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviated path and the actual path.

[0154] Cloud-edge collaborative path recommendation module: The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform, and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended path results that meet user preferences online.

[0155] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A personalized route planning method for logistics transportation based on a cloud platform, characterized in that: include: Construct a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic points of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges. The construction of a heterogeneous traffic graph model, which includes multiple transportation area nodes and path edges, includes: Acquire vector road data, point of interest distribution data, and historical traffic event data covering multiple transportation areas, and parse their spatial location and temporal attributes respectively to initialize and form corresponding road segment nodes, geographic point of interest nodes, and traffic event nodes; Based on road topology and interest point distribution density, geographical adjacency edges between road nodes and interest point nodes are constructed, and edge weights are calculated using a regional proximity function based on multi-scale grid partitioning. Edge weights are defined using a region proximity function: Where d(i,j) is the Euclidean distance between nodes i and j. The difference in interest point density between grids is represented by α and β, which are adjustment parameters. Node pairs with edge weights greater than a set threshold are connected as geographic adjacency edges. Based on the time lag relationship between historical traffic events and route execution results, an event influence edge is constructed between event nodes and road nodes, and a decreasing weighting mechanism is used to characterize the degree of influence of event persistence on route accessibility. Iterate through each traffic event node and calculate its impact on the traffic speed and actual route execution of surrounding roads within several time windows after its occurrence; Add event-affected edges to the road segments affected by the event in each path, and use a dynamic weighting function: Where Δt is determined by the type and level of the event, and λ is the attenuation coefficient; Multidimensional attribute encoding is performed on all nodes and edge structures, including road level, point of interest type and event urgency, to generate a graph structure feature matrix that supports heterogeneous graph neural network input; Collect historical transportation task records from multiple users, generate user semantic intent tags based on user delivery behavior and historical preferences, and map the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure; Specifically: if a user prefers a certain type of road structure, their intent label is mapped to the "negative preference" attribute of the road node; if a user selects a specific transit warehouse or point of interest as a stop location, a "high priority" semantic intent label is added to the corresponding point of interest node; if an event node has a significant impact on the user's path deviation, a "sensitive label" is embedded on the event edge. Based on the constructed semantically constrained traffic map, a heterogeneous graph neural network model is used to jointly model the user intent and path reachability of each transportation task, generating a path score sequence. Specifically, this includes: dividing the traffic map into multiple subgraphs based on the structural roles of various nodes, constructing embedding spaces for each, and using meta-path guidance to achieve initial modeling of node feature alignment and semantic association; using the semantic intent label of the user's current transportation task as the graph state initialization factor, assigning context activation weights to the node set strongly correlated with the intent label, forming the starting point of the semantically driven graph structure; employing a multi-view heterogeneous graph attention mechanism to aggregate different types of neighbor information from the structural view, preference view, and event interference view, dynamically capturing the semantic accessibility and behavioral fit of the path; regularizing the historical score dependency of candidate paths, and finally outputting a user-customized path score sequence for ranking; After each transportation task is completed, abnormal transportation events that occur during its execution are collected, the causal deviation relationship between the recommended path and the actual path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviation path and the actual path. The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended route results that meet user preferences online.

2. The personalized recommendation route planning method for logistics transportation based on a cloud platform according to claim 1, characterized in that: The user semantic intent tags generated based on user delivery behavior and historical preferences include: The user's historical transportation trajectory is sliced, and a multi-dimensional behavior sequence is constructed by combining the task time period, transportation type and delivery scenario context tags; A behavior feature encoding network with attention mechanism weighting is used to represent each behavior sequence, generating a high-dimensional embedding vector containing spatial path preference, temporal response characteristics and behavior frequency; A contrastive learning mechanism is introduced, using the behavior sequences of the same user in different scenarios as positive sample pairs and the behavior sequences of different users as negative sample pairs. Dense clustering is performed in the embedding space to identify multiple cluster centers with stable transport preferences, and each center is assigned an interpretable semantic intent label.

3. The personalized recommendation route planning method for logistics transportation based on a cloud platform according to claim 1, characterized in that: The dynamic updating of semantic intent mapping based on the reverse alignment error between the deviation path and the true path includes: The recommended path and the actual execution path are compared at the node level to construct a spatial-temporal alignment matrix, identify path segments with significant deviations, and extract the corresponding path deviation node set. For each node in the deviation node set, calculate its cumulative frequency of participating in the deviation path in historical tasks, and establish a node deviation confidence model in combination with the task context to measure the credibility of its semantic mapping. For nodes with confidence scores below a threshold, their original positive preference labels are replaced with context-sensitive labels, and the node semantic attribute vector is updated.

4. The personalized recommendation route planning method for logistics transportation based on a cloud platform according to claim 1, characterized in that: This includes combining real-time local traffic data acquired through edge computing devices to generate recommended routes that match user preferences online, including: The local traffic status reported by edge devices is integrated into the cloud semantic traffic graph in the form of graph patches, and the graph structure is kept in sync with the current traffic environment through a node-level incremental update mechanism. Based on user intent labels, we activate associated nodes in the graph and construct a subgraph input for the task, and use a heterogeneous graph neural network model for multi-hop message passing and path candidate generation. By introducing dynamic factors based on user context, including task urgency and cargo characteristics, candidate path scores are weighted and ranked to generate real-time path recommendation results.

5. A cloud-based personalized route planning system for logistics transportation, used to implement the cloud-based personalized route planning method for logistics transportation as described in any one of claims 1-4, characterized in that: include: Heterogeneous graph construction module: Constructs a heterogeneous traffic graph model containing multiple transportation area nodes and path edges. The traffic graph nodes include road segment nodes, geographic point of interest nodes, and traffic event nodes. The path edges include time-related edges, geographic adjacency edges, and event impact edges. Semantic intent generation module: Collects historical transportation task records of multiple users, generates user semantic intent tags based on user delivery behavior and historical preferences, and maps the semantic intent tags to the node and edge attributes of the traffic graph to form a semantic constraint graph structure; Heterogeneous graph modeling module: Based on the constructed semantically constrained traffic graph, the module uses a heterogeneous graph neural network model to jointly model the user intent and path reachability of each transportation task, and generates a path score sequence. Path execution feedback module: After each transportation task is completed, abnormal transportation events that occur during its execution are collected, the causal deviation relationship between the recommended path and the actual path is analyzed, and the semantic intent mapping relationship is dynamically updated based on the reverse alignment error between the deviated path and the actual path. Cloud-edge collaborative path recommendation module: The heterogeneous graph neural network model and semantic intent mapping module are deployed on the core computing node of the cloud platform, and combined with edge computing devices to obtain local traffic status in real time, so as to generate recommended path results that meet user preferences online.

Images