A Smart Logistics POI Recommendation Method Based on Semantic Knowledge Distillation and Interpretability

Abstract

This application relates to an intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability, comprising: acquiring training data; encoding various types of data according to a feature engineering module, and inputting the obtained user identifier features, POI attribute features, and user behavior features into the deep interaction network of the teacher model; calculating the preliminary interaction score between the user and POI based on the node features of the updated user nodes and POI nodes using a graph neural network; employing a lightweight model as the basic architecture of the student model, inheriting the semantic knowledge of the teacher model through adaptive semantic knowledge distillation, inputting the test data into the student model after inheriting the semantic knowledge, outputting the final interaction score between the user and POI, selecting the POIs with the highest final interaction scores as the recommendation results, and calculating the SHAP value to interpret the recommendation results. This method can simultaneously achieve efficient recommendation, semantic preservation, and decision transparency.

Description

Technical Field

[0001] This application relates to the field of intelligent logistics technology, and in particular to an intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability. Background Technology

[0002] With the development of IoT and big data technologies, intelligent logistics systems have become the core of modern supply chain management. Points of Interest (POI) recommendation, as a key technology in intelligent logistics, is widely used in scenarios such as delivery point selection, route planning, and resource scheduling. However, POI recommendation in the logistics field faces challenges such as data sparsity, complex scenarios, and black-box decision-making. Specifically: 1) User-POI interaction data is sparse and heterogeneous, affected by multiple factors such as traffic, weather, and time. 2) Traditional recommendation models (such as collaborative filtering and neural networks) struggle to balance efficiency and accuracy. 3) Deep learning models lack interpretability, reducing the trust and acceptance of the system among logistics practitioners.

[0003] Knowledge distillation, by transferring knowledge from complex teacher models to lightweight student models, can improve inference efficiency while maintaining performance. Explainable AI technologies enhance system transparency by revealing the rationale behind model decisions. However, existing methods do not fully integrate semantic knowledge distillation with explainable mechanisms, especially in logistics POI recommendation scenarios, where an explainable distillation framework tailored to business needs is lacking. Summary of the Invention

[0004] Therefore, it is necessary to provide an intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability, including:

[0005] S1: Acquire user data, POI data, and user-POI interaction data for training;

[0006] S2: Based on the feature engineering module, each type of data is encoded to obtain user identification features, POI attribute features, and user behavior features.

[0007] S3: Input user identification features, POI attribute features, and user behavior features into the deep interaction network of the teacher model to construct an interaction graph. The graph neural network updates the node features of user nodes and POI nodes in the interaction graph based on each feature, and calculates the preliminary interaction score between the user and the POI based on the updated node features of user nodes and POI nodes.

[0008] S4: A lightweight model is used as the basic architecture of the student model. The semantic knowledge of the teacher model is inherited through adaptive semantic knowledge distillation. The test data is input into the student model after inheriting the semantic knowledge, and the final interaction score between the user and the POI is output. The POIs with the highest final interaction scores are used as the recommendation results, and the SHAP value is calculated to interpret the recommendation results.

[0009] Preferably, the user data includes user ID and user latitude and longitude; the user ID is mapped to user identification features through a hash embedding function, and the user latitude and longitude are converted into Cartesian coordinates and then discretized and embedded into user location features.

[0010] Preferably, the POI data includes POI-ID, POI category, and POI latitude and longitude; the POI attribute features include POI identification features, POI category features, and POI location features.

[0011] Embed the POI-ID as a POI identifier feature;

[0012] Embed POI categories as POI category features;

[0013] After converting the latitude and longitude of the POI to Cartesian coordinates, it is discretized and embedded as the POI location feature.

[0014] Preferably, the user-POI interaction data includes interaction records, interaction timestamps, and task types; the user behavior characteristics include: behavior sequence characteristics, time characteristics, and task type characteristics.

[0015] The interaction records are processed by a BERT encoder and a BiLSTM network to generate behavioral sequence features, calculated as follows:

[0016] ;

[0017] in, Represents behavioral sequence features. Indicates weight, Indicates the BERT encoder. Indicates interaction history. Indicates a BiLSTM network;

[0018] The interaction timestamp is decomposed into hours and days of the week, and periodically encoded to map it as a time feature;

[0019] Task types include pickup or delivery, and the task type is embedded as a task type feature.

[0020] Preferably, it also includes filtering users with fewer than 10 interaction records and / or POIs with fewer than 5 visits based on interaction records.

[0021] Preferably, constructing an interaction graph includes:

[0022] Each user remaining after filtering is taken as a user node, and the user identifier feature is taken as the initial node feature of the corresponding user node.

[0023] Each remaining POI after filtering is taken as a POI node, and the POI attribute features are taken as the initial node features of the corresponding POI node.

[0024] The interaction records between the user and the POI are used as edges between the corresponding user node and the POI node;

[0025] Edge weights are calculated based on user location features, POI location features, and time features. The calculation formula is as follows:

[0026] ;

[0027] in, Represents user node With POI nodes Edge weights between them Indicates the spatial attenuation coefficient. Indicates the time decay coefficient. Represents spherical distance. Represents user node User location characteristics, Indicates POI node POI location features, Represents user node The timing characteristics of initiating the interaction Indicates POI node The time characteristics of the accepted interaction It represents the absolute value.

[0028] Preferably, the process of outputting the initial interactive score includes:

[0029] In each layer of the graph neural network, user nodes and POI nodes are updated based on edge weights, using the following formula:

[0030] ;

[0031] ;

[0032] in, Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each user node This represents the activation function. Indicates the first Weights of a layered graph neural network Indicates the first Bias of layered graph neural networks This represents the attention-weighted summation function. Indicates the first In a layered graph neural network, the first Node characteristics of each POI node Indicates the first In a layered graph neural network, the first Node characteristics of each user node Represents user node With POI nodes Edge weights between them Indicates the relationship with the first A set of POI nodes connected to by a user node. Indicates the relationship with the first A collection of user nodes connected to a POI node;

[0033] In the first layer of the graph neural network, the corresponding node features are updated using the initial node features of user nodes or POI nodes;

[0034] The node features of the user nodes and POI nodes are updated using the Lth layer graph neural network. The preliminary interaction score between the user and the POI is calculated using the following formula:

[0035] ;

[0036] in, Represents user node With POI nodes Preliminary interaction scores between them Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each user node This represents the learning matrix.

[0037] Preferred adaptive semantic knowledge distillation includes:

[0038] Based on the interaction scores predicted by the student model and their corresponding actual interaction scores, the cross-entropy loss is calculated using the following formula:

[0039] ;

[0040] in, Let D represent the cross-entropy loss, and let D represent the set of nodes in the interaction graph. Represents user node With POI nodes The actual interaction score between them This indicates that the student model predicts user nodes. With POI nodes Interaction scores between them;

[0041] The knowledge distillation loss is calculated based on the logit values ​​of the teacher and student models, using the following formula:

[0042] ;

[0043] in, This represents the knowledge distillation loss. Denotes KL divergence, This represents the softmax function. This represents the logit value of the teacher model. This represents the logit value of the student model. Indicates temperature parameter;

[0044] Based on the teacher model and student model, the first The feature matching loss is calculated for the updated node features in a layered graph neural network, using the following formula:

[0045] ;

[0046] in, Represents feature matching loss, Represents the cosine distance. In the teacher model, the first The node features of any node updated by a layered graph neural network. In the student model, the first The node features of any node updated by the layered graph neural network;

[0047] The total loss is obtained by weighted fusion of cross-entropy loss, knowledge distillation loss, and feature matching loss, and the formula is as follows:

[0048] ;

[0049] in, Indicates the total loss. Indicates the first loss weight. This indicates the second loss weight. Indicates the third loss weight;

[0050] Adjust temperature parameters and loss weights according to the needs of real-time logistics scenarios to improve the efficiency or effect of semantic knowledge distillation.

[0051] Preferably, the calculation of the SHAP value to interpret the recommendation results includes:

[0052] The SHAP value is calculated based on the final interaction score, using the following formula:

[0053] ;

[0054] in, Representing target features SHAP value, Represents the complete set of features. express Not included Any feature subset, express Size, express Size, Indicates that in the known and In this case, the student model outputs the final interaction score. Indicates only known The final interaction score output by the student model in this case;

[0055] Target features include current logistics network load matching degree, regional empty mileage rate, delivery distance, delivery efficiency, road conditions, and distance from the main route;

[0056] Based on the SHAP value of the target features, the reasons for selecting each recommendation result are quantified, and the recommendation results are explained.

[0057] Beneficial effects: This method acquires training data; encodes various data types according to the feature engineering module, and inputs the obtained user identifier features, POI attribute features, and user behavior features into the deep interaction network of the teacher model to construct an interaction graph. Based on the node features of user nodes and POI nodes updated by the graph neural network, the initial interaction score between users and POIs is calculated; a lightweight model is used as the basic architecture of the student model, inheriting the semantic knowledge of the teacher model through adaptive semantic knowledge distillation. The test data is input into the student model after inheriting semantic knowledge, and the final interaction score between users and POIs is output. The POIs with the highest final interaction scores are used as recommendation results, and the SHAP value is calculated to interpret the recommendation results. This method can simultaneously achieve efficient recommendation, semantic preservation, and decision transparency, and improve recommendation accuracy, thus supporting the actual deployment and application of intelligent logistics systems. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability in the embodiments of this application. Detailed Implementation

[0060] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0061] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0062] like Figure 1 As shown, this embodiment provides an intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability, including:

[0063] S1: Obtain user data, POI data, and user-POI interaction data for training.

[0064] In this embodiment, the ratio of training data: validation set: test data is 7:1:2.

[0065] S2: Based on the feature engineering module, each type of data is encoded to obtain user identification features, POI attribute features, and user behavior features.

[0066] Specifically:

[0067] The user data includes user ID, user latitude and longitude, user role (dispatcher, deliveryman, warehouse manager), and delivery area; the user ID is mapped to user identification features through a hash embedding function, and the user latitude and longitude are converted into Cartesian coordinates and then discretized and embedded into user location features.

[0068] The POI data includes POI-ID, POI category, POI latitude and longitude, and service area; the POI attribute features include POI identification features, POI category features, and POI location features.

[0069] Embed the POI-ID as a POI identifier feature;

[0070] Embed POI categories as POI category features;

[0071] After converting the latitude and longitude of the POI to Cartesian coordinates, it is discretized and embedded as the POI location feature.

[0072] The user-POI interaction data includes interaction records, interaction timestamps, and task types; the user behavior characteristics include: behavior sequence characteristics, time characteristics, and task type characteristics.

[0073] The interaction records are processed by a BERT encoder and a BiLSTM network to generate behavioral sequence features, calculated as follows:

[0074] ;

[0075] in, Represents behavioral sequence features. Indicates weight, Indicates the BERT encoder. This represents an interaction record (such as a sequence of Points of Interest (POIs) of a user's most recent interactions). Indicates a BiLSTM network;

[0076] The interaction timestamp is decomposed into hours and days of the week, and periodically encoded to map it as a time feature;

[0077] Task types include pickup or delivery, and the task type is embedded as a task type feature.

[0078] In this embodiment, it also includes filtering users with fewer than 10 interaction records and / or POIs with fewer than 5 visits based on the interaction records.

[0079] In this embodiment, the initialization parameters of the feature engineering module are: the feature corresponding to the user and the feature embedding dimension corresponding to the POI are set to 64, and the number of BERT layers is set to 4.

[0080] S3: Input user identification features, POI attribute features, and user behavior features into the deep interaction network of the teacher model to construct an interaction graph. The graph neural network updates the node features of user nodes and POI nodes in the interaction graph based on each feature, and calculates the preliminary interaction score between the user and the POI based on the updated node features of user nodes and POI nodes.

[0081] Specifically, constructing an interaction diagram includes:

[0082] Each user remaining after filtering is taken as a user node, and the user identifier feature is taken as the initial node feature of the corresponding user node.

[0083] Each remaining POI after filtering is taken as a POI node, and the POI attribute features are taken as the initial node features of the corresponding POI node.

[0084] The interaction records between the user and the POI are used as edges between the corresponding user node and the POI node;

[0085] Edge weights are calculated based on user location features, POI location features, and time features. The calculation formula is as follows:

[0086] ;

[0087] in, Represents user node With POI nodes Edge weights between them Indicates the spatial attenuation coefficient. Indicates the time decay coefficient. Represents spherical distance. Represents user node User location characteristics, Indicates POI node POI location features, Represents user node The timing characteristics of initiating the interaction Indicates POI node The time characteristics of the accepted interaction It represents the absolute value.

[0088] Furthermore, the process of outputting preliminary interaction scores includes:

[0089] In each layer of the graph neural network, user nodes and POI nodes are updated based on edge weights, using the following formula:

[0090] ;

[0091] ;

[0092] in, Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each user node This represents the activation function. Indicates the first Weights of a layered graph neural network Indicates the first Bias of layered graph neural networks This represents the attention-weighted summation function. Indicates the first In a layered graph neural network, the first Node characteristics of each POI node Indicates the first In a layered graph neural network, the first Node characteristics of each user node Represents user node With POI nodes Edge weights between them Indicates the relationship with the first A set of POI nodes connected to by a user node. Indicates the relationship with the first A collection of user nodes connected to a POI node;

[0093] In the first layer of the graph neural network, the corresponding node features are updated using the initial node features of user nodes or POI nodes;

[0094] The node features of the user nodes and POI nodes are updated using the Lth layer graph neural network. The preliminary interaction score between the user and the POI is calculated using the following formula:

[0095] ;

[0096] in, Represents user node With POI nodes Preliminary interaction scores between them Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each user node This represents the learning matrix.

[0097] In this embodiment, the number of attention heads in the graph neural network is set to 8, and the number of layers in the graph neural network (GNN) is set to 3. The Adam optimizer is used, with the interaction score prediction loss (MSE loss) as the objective, and training stops when there is no improvement for 3 consecutive epochs on the validation set NDCG@10, thus completing the training of the teacher model.

[0098] S4: A lightweight model is used as the basic architecture of the student model. The semantic knowledge of the teacher model is inherited through adaptive semantic knowledge distillation. The test data is input into the student model after inheriting the semantic knowledge, and the final interaction score between the user and the POI is output. The POIs with the highest final interaction scores are used as the recommendation results, and the SHAP value is calculated to interpret the recommendation results.

[0099] In this embodiment, the lightweight model includes regularized matrix factorization. The student model specifically includes:

[0100] Embedded layer: Inherits the feature engineering module.

[0101] Infrastructure: Learning through low-rank approximation: ,in, This represents the node features input to the user nodes in the student model. Indicates transpose. This represents the node features input to the POI nodes in the student model; the number of parameters (node ​​features) in the student model is reduced by more than 70% compared to the teacher model, and the computational complexity is reduced from O(n^2). 2 The value decreases to O(n).

[0102] Feature enhancement layer: Introducing a shallow MLP to fuse static features on top of the ID embedding: , , , These are the static characteristics of users and POIs (such as user delivery area and POI service range). This represents a multilayer perceptron.

[0103] Adaptive semantic knowledge distillation includes:

[0104] Based on the interaction scores predicted by the student model and their corresponding actual interaction scores, the cross-entropy loss is calculated using the following formula:

[0105] ;

[0106] in, Let D represent the cross-entropy loss, and let D represent the set of nodes in the interaction graph. Represents user node With POI nodes The actual interaction score between them This indicates that the student model predicts user nodes. With POI nodes Interaction scores between them;

[0107] The knowledge distillation loss is calculated based on the logit values ​​of the teacher and student models, using the following formula:

[0108] ;

[0109] in, This represents the knowledge distillation loss. Denotes KL divergence, This represents the softmax function. This represents the logit value of the teacher model. This represents the logit value of the student model. Indicates temperature parameter;

[0110] Based on the teacher model and student model, the first The feature matching loss is calculated for the updated node features in a layered graph neural network, using the following formula:

[0111] ;

[0112] in, Represents feature matching loss, Represents the cosine distance. In the teacher model, the first The node features of any node updated by a layered graph neural network. In the student model, the first The node features of any node updated by the layered graph neural network;

[0113] The total loss is obtained by weighted fusion of cross-entropy loss, knowledge distillation loss, and feature matching loss, and the formula is as follows:

[0114] ;

[0115] in, Indicates the total loss. Indicates the first loss weight. This indicates the second loss weight. Indicates the third loss weight;

[0116] Adjust temperature parameters and loss weights according to the needs of real-time logistics scenarios to improve the efficiency or effect of semantic knowledge distillation.

[0117] This student model categorizes real-time data features in logistics scenarios into peak and off-peak periods to automatically adjust loss weights and temperature parameters. For example, during peak periods, the weight of "feature matching loss" is increased to ensure the accuracy of POI recommendations, while during off-peak periods, the weight is reduced to accelerate model inference. Simultaneously, "distillation difficulty awareness" is introduced to individually enhance KL divergence constraints for users / POIs with sparse interaction data, reducing knowledge transfer bias.

[0118] In this embodiment, the training of the student model includes:

[0119] Initialize the student model: set the embedding dimension to 32, and the shallow MLP to 2 layers (128 hidden units).

[0120] Configure distillation parameters: Temperature parameters =3.0, loss weight , , .

[0121] Progressive distillation training: First, freeze the embedding layer of the student model and train only the MLP layer (50 epochs); then unfreeze the embedding layer and jointly optimize the total loss (50 epochs). During training, monitor the validation set NDCG@10 and inference time in real time.

[0122] Lightweight model deployment: Convert the trained student model into ONNX format and deploy it to the logistics scheduling terminal or cloud server.

[0123] Furthermore, the calculation of the SHAP value to interpret the recommendation results includes:

[0124] The SHAP value is calculated based on the final interaction score, using the following formula:

[0125] ;

[0126] in, Representing target features SHAP value, Represents the complete set of features. express Not included Any feature subset, express Size, express Size, Indicates that in the known and In this case, the student model outputs the final interaction score. Indicates only known The final interaction score output by the student model in this case;

[0127] Target features include current logistics network load matching degree, regional empty mileage rate, delivery distance, delivery efficiency, road conditions, and distance from the main route;

[0128] Based on the SHAP value of the target features, the reasons for selecting each recommendation result are quantified, and the recommendation results are explained.

[0129] Scene-aware interpretation generation: Extracting Top-K positive and negative SHAP value features to generate natural language interpretation templates that adapt to the needs of different user roles. An example is given below:

[0130] For dispatchers: "POI [p] is recommended because it has a high degree of matching with the current logistics network load (SHAP value: 0.82) and can reduce the regional empty-run rate by 15% (SHAP value: 0.65)".

[0131] For delivery drivers: "POI [p] is recommended because it is only 3.2 kilometers away from the previous delivery point (SHAP value: 0.91), and the delivery efficiency is optimal in the current time period (afternoon) (SHAP value: 0.73)".

[0132] Counterfactual explanation supplement: Compare the characteristic differences between the target POI and the alternative POI to generate an explanation for "why not recommended": "POI is not recommended [p'] because the POI has a congestion index of 0.8 (SHAP value: -0.52) and is 2.1 km off the main path (SHAP value: -0.38)."

[0133] This embodiment also includes verifying the consistency of interpretation on the test set; if it is less than 80%, the loss weight and template content are adjusted.

[0134] Application and monitoring of the method:

[0135] Application scenario integration: The method is integrated with the logistics scheduling platform to support functions such as route planning (input deliveryman ID and origin, output Top-5 delivery points) and pre-position warehouse location selection (input regional requirements, output Top-3 warehouse locations).

[0136] Real-time monitoring: Monitor system inference time, recommendation accuracy, and user feedback.

[0137] Model iteration: Each quarter, based on newly collected logistics data, the teacher and student models are updated, and temperature parameters and interpretation templates are optimized.

[0138] The intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretability provided in this embodiment has the following beneficial effects:

[0139] To address the issues of data sparsity, complex multi-objective decision-making, and model "black box" problems in existing intelligent logistics POI recommendation systems, this embodiment provides an intelligent logistics POI recommendation method based on semantic knowledge distillation and interpretable AI, achieving synergistic optimization of improved recommendation accuracy, lightweight model deployment, and transparent decision-making. The student model module inherits the semantic knowledge of the teacher model through adaptive semantic knowledge distillation, achieving lightweight deployment. Recommendation explanations are generated based on SHAP values. These three components are linked through data interfaces and parameter passing mechanisms, forming a complete recommendation process of "semantic learning - knowledge transfer - explanation generation." This embodiment provides a POI recommendation method that can simultaneously achieve efficient recommendation, semantic preservation, and transparent decision-making, supporting the practical deployment and application of intelligent logistics systems.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A smart logistics POI recommendation method based on semantic knowledge distillation and interpretability, characterized in that, include: S1: Acquire user data, POI data, and user-POI interaction data for training; the user-POI interaction data includes interaction records, interaction timestamps, and task types; S2: Based on the feature engineering module, each type of data is encoded to obtain user identification features, POI attribute features, and user behavior features. The user behavior characteristics include: behavior sequence characteristics, time characteristics, and task type characteristics; The interaction records are processed by a BERT encoder and a BiLSTM network to generate behavioral sequence features, calculated as follows: ； in, Represents behavioral sequence features. Indicates weight, Indicates the BERT encoder. Indicates interaction history. Indicates a BiLSTM network; The interaction timestamp is decomposed into hours and days of the week, and periodically encoded to map it as a time feature; Task types include pickup or delivery, and task types are embedded as task type features; Filter users with fewer than 10 interaction records and / or POIs with fewer than 5 visits based on interaction records; S3: Input user identification features, POI attribute features, and user behavior features into the deep interaction network of the teacher model to construct an interaction graph. The graph neural network updates the node features of user nodes and POI nodes in the interaction graph based on each feature, and calculates the preliminary interaction score between the user and the POI based on the updated node features of user nodes and POI nodes. Building an interaction diagram includes: Each user remaining after filtering is taken as a user node, and the user identifier feature is taken as the initial node feature of the corresponding user node. Each remaining POI after filtering is taken as a POI node, and the POI attribute features are taken as the initial node features of the corresponding POI node. The interaction records between the user and the POI are used as edges between the corresponding user node and the POI node; Edge weights are calculated based on user location features, POI location features, and time features. The calculation formula is as follows: ； in, Represents user node With POI nodes Edge weights between them Indicates the spatial attenuation coefficient. Indicates the time decay coefficient. Represents spherical distance. Represents user node User location characteristics, Indicates POI node POI location features, Represents user node The timing characteristics of initiating the interaction Indicates POI node The time characteristics of the accepted interaction Represents absolute value; The process of generating the initial interaction score includes: In each layer of the graph neural network, user nodes and POI nodes are updated based on edge weights, using the following formula: ； ； in, Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each POI node This represents the activation function. Indicates the first Weights of a layered graph neural network Indicates the first Bias of layered graph neural networks This represents the attention-weighted summation function. Indicates the first In a layered graph neural network, the first Node characteristics of each POI node Indicates the first In a layered graph neural network, the first Node characteristics of each user node Represents user node With POI nodes Edge weights between them Indicates the relationship with the first A set of POI nodes connected to by a user node. Indicates the relationship with the first A collection of user nodes connected to a POI node; In the first layer of the graph neural network, the corresponding node features are updated using the initial node features of user nodes or POI nodes; The node features of the user nodes and POI nodes are updated using the Lth layer graph neural network. The preliminary interaction score between the user and the POI is calculated using the following formula: ； in, Represents user node With POI nodes Preliminary interaction scores between them Indicates the first The updated first layer in a graph neural network Node characteristics of each user node Indicates the first The updated first layer in a graph neural network Node characteristics of each POI node Represents the learning matrix; S4: A lightweight model is used as the basic architecture of the student model. The semantic knowledge of the teacher model is inherited through adaptive semantic knowledge distillation. The test data is input into the student model after inheriting the semantic knowledge, and the final interaction score between the user and the POI is output. The POIs with the highest final interaction scores are used as the recommendation results, and the SHAP value is calculated to interpret the recommendation results.

2. The intelligent logistics POI recommendation method according to claim 1, characterized in that, The user data includes user ID and user latitude and longitude; the user ID is mapped to user identification features through a hash embedding function, and the user latitude and longitude are converted into Cartesian coordinates and then discretized and embedded into user location features.

3. The intelligent logistics POI recommendation method according to claim 2, characterized in that, The POI data includes POI-ID, POI category, and POI latitude and longitude; the POI attribute features include POI identification features, POI category features, and POI location features. Embed the POI-ID as a POI identifier feature; Embed POI categories as POI category features; After converting the latitude and longitude of the POI to Cartesian coordinates, it is discretized and embedded as the POI location feature.

4. The intelligent logistics POI recommendation method according to claim 1, characterized in that, Adaptive semantic knowledge distillation includes: Based on the interaction scores predicted by the student model and their corresponding actual interaction scores, the cross-entropy loss is calculated using the following formula: ； in, Let D represent the cross-entropy loss, and let D represent the set of nodes in the interaction graph. Represents user node With POI nodes The actual interaction score between them This indicates that the student model predicts user nodes. With POI nodes Interaction scores between them; The knowledge distillation loss is calculated based on the logit values ​​of the teacher and student models, using the following formula: ； in, This represents the knowledge distillation loss. Denotes KL divergence, This represents the softmax function. This represents the logit value of the teacher model. This represents the logit value of the student model. Indicates temperature parameter; Based on the teacher model and student model, the first The feature matching loss is calculated for the updated node features in a layered graph neural network, using the following formula: ； in, Represents feature matching loss, Represents the cosine distance. In the teacher model, the first The node features of any node updated by a layered graph neural network. In the student model, the first The node features of any node updated by the layered graph neural network; The total loss is obtained by weighted fusion of cross-entropy loss, knowledge distillation loss, and feature matching loss, and the formula is as follows: ； in, Indicates the total loss. Indicates the first loss weight. Indicates the second loss weight. Indicates the third loss weight; Adjust temperature parameters and loss weights according to the needs of real-time logistics scenarios to improve the efficiency or effect of semantic knowledge distillation.

5. The intelligent logistics POI recommendation method according to claim 4, characterized in that, The calculation of the SHAP value to interpret the recommendation results includes: The SHAP value is calculated based on the final interaction score, using the following formula: ； in, Representing target features SHAP value, Represents the complete set of features. express Not included Any feature subset, express Size, express Size, Indicates that in the known and In this case, the student model outputs the final interaction score. Indicates only known The final interaction score output by the student model in this case; Target features include current logistics network load matching degree, regional empty mileage rate, delivery distance, delivery efficiency, road conditions, and distance from the main route; Based on the SHAP value of the target features, the reasons for selecting each recommendation result are quantified, and the recommendation results are explained.

Images