Multimodal travel mode fusion recommendation method based on dynamic transportation network

Abstract

The present invention discloses a multimodal travel mode fusion recommendation method based on a dynamic traffic network. The present invention first constructs the user's spatiotemporal travel trajectory into a heterogeneous information network, and considers the dynamic traffic network that is constantly changing during the travel process; pre-processes historical trajectory data according to the heterogeneous information network; extracts key meta-paths from the network mode corresponding to the heterogeneous information network to enrich the feature embedding of the object; realizes meta-path feature aggregation through aggregation functions; uses a meta-path-guided graph neural network to dynamically model users and travel modes, and obtains feature embedding with rich interactive information; the final feature embedding of users and travel modes is input into the MLP, fully learning the user's preference for travel modes in the spatiotemporal travel trajectory, and recommending travel modes that meet personalized needs to users. The present invention describes the user's personalized preference for travel modes in a fine-grained manner, while alleviating traffic pressure and improving traffic resilience.

Description

Technical Field

[0001] The present invention belongs to the intersection field of transportation travel mode recommendation and heterogeneous graph neural network, and specifically relates to a multi-modal travel mode fusion recommendation method based on a dynamic transportation network. Background Art

[0002] With the continuous development of urbanization and the improvement of economic level, people's travel needs are becoming increasingly diversified. When the demand exceeds the carrying capacity of the transportation system, various traffic problems and even environmental problems will arise, such as traffic jams and automobile exhaust pollution. The complex choice behavior of people in the multi-modal transportation network determines the distribution of travel demand on the transportation network. From the perspective of balancing the distribution of transportation demand, it is of great research value to recommend personalized travel modes.

[0003] Nowadays, in the era of big data, recommendation systems, as information filtering technologies, are particularly important for users to recommend information that meets their personal needs. Heterogeneous information networks take into account different types of objects, different connection relationships between objects, and attribute information, fully reflecting the interaction between different objects in the recommendation system, mining implicit information, learning deeper regularities, greatly improving the accuracy of recommendations, and opening up a new path for personalized recommendation technology. In addition, it can be seen that traditional user-item recommendations are themselves heterogeneous bipartite graphs, so it is inevitable to use HIN (Heterogeneous information networks, HIN) in recommendation technology. Common heterogeneous information networks include social networks, biological neural networks, academic networks, etc., but not only these, the data in the real world is rich and diverse, and information networks are everywhere. For example, the type of movie, director, and user attributes are used to construct a heterogeneous information network to express user preferences in a more fine-grained manner. There are many other studies on recommendation by building heterogeneous information networks, such as malicious account detection and search intent recommendation. This paper finds that users' travel behavior on multimodal transportation networks is also a reflection of information networks. It then proposes to apply representation learning technology in heterogeneous information networks to transportation and recommend multimodal travel methods in combination with dynamic transportation networks.

[0004] Figure 1The interaction process between the user's spatiotemporal travel trajectory and the transportation network is explained. Users use multiple travel modes to continuously switch between complex transportation networks, and finally form a trajectory chain. The complex transportation network is composed of multimodal travel modes. Travel mode is the basic attribute for understanding user mobility behavior, and the transportation mode used by users in the trajectory can reflect the regularity of the user's mobility status. Therefore, most scholars have always paid attention to the research field related to travel mode. Although the field of travel mode has received much attention, there is still little research on personalized recommendation of travel mode. This paper proposes to apply heterogeneous information network representation learning to transportation travel, and completes personalized recommendation of travel mode to fill this gap to a certain extent. Some researchers also use sequence-based methods to recommend personalized travel modes to users, but they do not fully consider the contextual information of the recommended objects and the interactivity between the recommended objects. Our research makes up for this defect well. Summary of the invention

[0005] In response to the above unresolved key issues, the present invention innovatively proposes to construct a heterogeneous information network of users, travel modes, spatiotemporal attributes strongly related to travel modes (such as time, etc.) and start and end locations, and use a graph neural network guided by a meta-path to dynamically model users and travel modes, derive feature embedding with rich interactive information, and fully learn users' travel mode preferences in historical travel trajectories from HIN. The purpose is to balance the distribution of traffic demand in the city and provide residents with personalized travel mode recommendations.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] The multi-modal travel mode fusion recommendation based on dynamic traffic network includes the following steps:

[0008] Step 1: It is proposed to apply heterogeneous information networks to transportation, and to construct a heterogeneous information network with the information of users, travel modes, starting and ending locations, and the spatiotemporal attributes of travel modes;

[0009] Step 2: Based on the heterogeneous information network constructed in step 1), pre-process the spatiotemporal trajectory data, perform validity processing on the user's spatiotemporal travel trajectory data, extract attributes such as users, travel modes, starting and ending locations, and time, construct a mapping relationship, and obtain application data;

[0010] Step 3: Extract key meta-paths from the network patterns corresponding to the heterogeneous information network to enrich the feature embedding of the object;

[0011] Step 4: Aggregate meta-path features through aggregation functions;

[0012] Step 5: Use the graph neural network guided by the meta-path to dynamically model users and travel modes, and obtain feature embedding with rich interactive information. Users are temporal, so LSTM dynamic modeling is used, and travel modes are dynamically modeled using CNN;

[0013] Step 6: The final features of users and travel modes are embedded into the input MLP to fully learn the user's preference for travel mode in the spatiotemporal travel trajectory and recommend travel modes that meet personalized needs to users.

[0014] Preferably, step 1 of building a heterogeneous information network includes the following:

[0015] Step 1-1: Define a set M = {U, V, T, L}, where U represents the set of users; V represents the travel mode used by the user in the travel trajectory, mainly including subway, bus, driving, taxi, self-driving car, and walking; T represents the start time when the user uses a certain travel mode; L represents the start and end locations of the user using a certain travel mode at a certain time; these are all node types in the heterogeneous transportation travel network;

[0016] Step 1-2: In a multimodal transportation network, users take different modes of travel to reach their destinations, and there are certain connections between them. This paper constructs a heterogeneous information network based on users, travel modes, starting and ending locations, and the spatiotemporal attributes of travel modes. By learning the semantic relationships between users and users, users and travel modes, travel modes and their corresponding starting and ending locations, more potential features are obtained (note: the time attribute is closely connected with the user at this time), fully expressing the user's personalized preference for travel mode.

[0017] Preferably, the spatiotemporal travel trajectory preprocessing stage in step 2 includes the following:

[0018] Step 2-1: Data cleaning: Before calculating the features of the data, data filtering is required: cleaning redundant data, deleting records with missing important fields and field values, and data screening;

[0019] Step 2-2: Feature attribute calculation: The four features of speed, acceleration, acceleration rate of change, and steering angle are calculated as trajectory attributes. In CNN, the input shape of all instances must be the same. To meet this challenge, all data segments are restricted to a fixed size of M. Therefore, longer data segments are truncated to M, while shorter data segments are filled with zero values. After generating segments with the same size, two consecutive GPS trajectory points (lat, long, t) are used to calculate the attribute value of each GPS point;

[0020] Step 2-3: Identify the stop points and divide the trip: The trip is divided by identifying the stop points of the user in the trajectory. Each trip contains multiple travel mode segments. This project uses the DBSCAN clustering algorithm to identify the stop points and extracts the corresponding trajectory point semantics in combination with the POI library;

[0021] Step 2-4: Extract travel mode segments by segmenting the trip based on the identification of conversion points: Use the walking segment-based method to segment the single travel mode segment;

[0022] Step 2-5: Identify travel modes based on convolutional neural networks: Integrate the characteristic attributes of each travel mode segment, let the convolutional neural network capture the characteristics of the mode segment, and finally complete the identification of travel modes in GPS trajectories;

[0023] Step 2-6: Establish a mapping relationship between users, travel modes, starting and ending locations, etc. to obtain application data.

[0024] Preferably, step 3 of extracting key meta-paths from the network model corresponding to the heterogeneous information network includes the following:

[0025] Step 3-1: In order to better understand the heterogeneous information network, construct a network model such as Figure 3 As shown, it stipulates the type constraints of the object set and the connection between objects: there is a connection between users and travel modes, indicating the relationship of use and being used; there is a connection between locations and travel modes, indicating the relationship of guidance and being guided, etc.

[0026] Step 3-2: Extract four meta-paths used in the present invention: "user-travel mode-starting and ending locations", "user-travel mode-user-travel mode", "travel mode-user-user-starting and ending locations", and "travel mode-user-starting and ending locations-user-travel mode".

[0027] Preferably, step 4 implements meta-path feature aggregation through an aggregation function, including the following:

[0028] Step 4-1: Based on the four meta-paths proposed in step 3-2, count the neighbors guided by the meta-paths: For example, the first-order neighbors of user u2 under the meta-path Φ:UVL are The second-order neighbors of u2 are the first-order neighbors of all nodes in the first-order neighbors, so its second-order neighbors are Therefore, all neighbors of user u2 are expressed as

[0029]

[0030] Step 4-2: Select a suitable aggregation function to aggregate the features of the meta-path and define two meta-paths Φ uul :UVL,Φ UVUV: UVUV, given the initial feature vectors of each node, the two processes of meta-path aggregation are described as follows:

[0031] Meta-path internal aggregation: Select a suitable aggregation function to aggregate the embeddings of the second-order neighbor nodes to obtain the embeddings of the first-order neighbor nodes, and then aggregate the embeddings of the first-order neighbor nodes to obtain the embedding of the target node, as shown below:

[0032]

[0033] f is the aggregation function for aggregating the first-order and second-order neighbor embeddings of u2 on the specified meta-path, is the second-order neighbor node embedding of user u2 on the meta-path, is the first-order neighbor node embedding, which is updated by aggregating the second-order neighbor node embeddings, and then aggregating the first-order neighbor embeddings to get the user u2 in the meta-path Φ uvl The final embedding on .

[0034] Aggregation between meta-paths: Aggregate the embeddings of user u2 obtained from different meta-paths to obtain the final embedding of u2:

[0035]

[0036] Preferably, step 5 uses a graph neural network guided by a meta-path to dynamically model users and travel modes, including the following:

[0037] Step 5-1: Dynamically model the user. First, use the average aggregation function to aggregate the embeddings of the second-order neighbors and the first-order neighbors. The aggregation results are concatenated with the first-order neighbor nodes, and the embeddings of the first-order neighbors are updated. Users have time series, so the LSTM aggregation function is used here: Better aggregation of user embeddings under different meta-paths. i are different meta paths with the user as the target node:

[0038] Step 5-2: Dynamically model the travel mode. First, use the average aggregation function to aggregate the features of neighbor nodes within the meta-path one by one, and then use the CNN aggregation function to complete the feature aggregation between different meta-paths to obtain the final embedding of all travel mode nodes. i These are different meta-paths with travel modes as target nodes:

[0039]

[0040] Preferably, step 6, fusion feature embedding prediction includes the following:

[0041] Step 6-1: Based on the user feature embedding obtained in step 5-1 and the travel mode feature embedding obtained in step 5-2, a fusion prediction is performed, that is,

[0042] Step 6-2: Train the model multiple times, use Precision, Recall, and NDCG as accuracy evaluation indicators, and observe the accuracy of recommendations.

[0043] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0044] This paper proposes to construct the user's spatiotemporal travel trajectory into a heterogeneous information network, consider the dynamic traffic network that is constantly changing during the travel process, and use the graph neural network guided by the meta-path to dynamically model the user and the travel mode, and obtain feature embedding with rich interactive information, so as to fully learn the user's preference for travel mode in the spatiotemporal travel trajectory, recommend travel modes that meet personalized needs to users, alleviate traffic congestion, and improve traffic resilience. Finally, the effectiveness of this method is demonstrated through experimental evaluation on real-world datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of the interaction between the user's spatial and temporal travel trajectory and the transportation network according to the present invention.

[0046] Figure 2 This is a flowchart of the present invention's fusion recommendation of multi-modal travel modes based on a dynamic transportation network.

[0047] Figure 3 It is a schematic diagram of the heterogeneous information network, network model, and meta-path proposed in the present invention.

[0048] Figure 4 This is a schematic diagram of the overall framework diagram of the fusion recommendation in the present invention.

[0049] Figure 5 It is a schematic diagram of the changes and comparisons of various indicators of the model when meta-paths are gradually added in the present invention.

[0050] Figure 6 This is a schematic diagram of the impact of different aggregation functions on recommendation performance in the present invention. DETAILED DESCRIPTION

[0051] In order to clearly and completely describe the technical solutions in the embodiments of the present invention in combination with the drawings in the embodiments of the present invention, it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0052] like Figure 2As shown, the technical route of the present invention mainly includes 6 steps, namely, proposing to construct a heterogeneous information network of spatiotemporal travel trajectories, preprocessing data, selecting key meta-paths, practicing meta-paths, dynamic modeling of graph neural networks guided by meta-paths, and integrating the features of users and travel modes into prediction results.

[0053] This example uses Geolife spatiotemporal travel data for testing. The following introduces this example from three aspects: building a heterogeneous transportation network, recommending a model for a heterogeneous transportation network, and predicting results.

[0054] 1) Building a heterogeneous transportation network

[0055] In a multimodal transportation network, users take different modes of travel to reach their destinations, and there are certain connections between them. Define a set M = {U, V, T, L}, where U represents the set of users; V represents the travel modes used by users in their travel trajectories, mainly including subway, bus, driving, taxi, self-driving, and walking; T represents the start time when the user uses a certain travel mode; L represents the start and end locations that the user experiences when using a certain travel mode at a certain time; these are all node types in a heterogeneous transportation network. This fusion recommendation analyzes the user's historical travel trajectory and learns the user's personalized preferences for travel modes, combined with the impact of dynamic transportation networks changing over time on travel modes, and ultimately recommends a more time-saving and energy-saving multimodal travel mode that meets personal preferences for users.

[0056] The specific contents are as follows:

[0057] Step 1: Construct a heterogeneous information network by combining users, travel modes, starting and ending locations, and the temporal and spatial attributes of travel modes (e.g. Figure 3 As shown in the left figure), by learning the semantic relationship between users, users and travel modes, travel modes and their corresponding starting and ending locations, more potential features are obtained (note: the time attribute is closely connected with the user at this time), fully expressing the user's personalized preference for travel mode. ;

[0058] Step 2: In order to better understand the heterogeneous information network, construct a network model such as Figure 3 As shown in the middle figure, it stipulates the type constraints of the object set and the connection between objects: there is a connection between users and travel modes, indicating the relationship of use and being used; there is a connection between locations and travel modes, indicating the relationship of guidance and being guided, etc. ;

[0059] Step 3: A meta-path is a path defined on a network model that links two types of objects. It represents a composite relationship between object types and is used to extract interaction information between objects. Figure 3The figure on the right shows the four meta-paths used in this article, in which users and time are closely connected. The meta-path of "user-travel mode-starting and ending locations" indicates that a user departs from a certain place and arrives at a certain place at a certain time; the meta-path of "user-travel mode-user-travel mode" indicates that a user uses a certain travel mode at a certain time, and similar users also use this travel mode and have used other travel modes; "travel mode-user-user-starting and ending locations" indicates that similar users of the user use a certain travel mode when arriving at similar starting and ending locations; "travel mode-user-starting and ending locations-user-travel mode" indicates that similar users of the user use a certain travel mode when arriving at similar locations at similar times.

[0060] Step 4: The present invention mainly integrates rich information by aggregating the neighbors guided by the meta-path. Figure 3 Taking user u2 as an example in the left figure, the neighbors guided by the meta-path are described in detail: the first-order neighbors of user u2 under the meta-path Φ:UVL are The second-order neighbors of u2 are the first-order neighbors of all nodes in the first-order neighbors, so its second-order neighbors are Therefore, all neighbors of user u2 are expressed as

[0061]

[0062] 2) Recommendation model for heterogeneous transportation networks

[0063] The idea of ​​this fusion recommendation is to construct the user's spatiotemporal travel trajectory into a heterogeneous information network, select different meta-paths, and use the different semantic information expressed by them to construct a graph neural network. By aggregating the neighbor node information guided by the meta-path, the node embedding of users and travel modes is enriched, that is, to better learn the representation of users and travel modes, and recommend travel modes that meet personalized needs for users. Figure 4 The overall framework of this fusion recommendation model is presented. The heterogeneous information network is used as input, and the embedding layer preprocesses the model data with content features to generate initial node embeddings. Then, in the meta-path aggregation layer, according to the selected meta-path, the neighboring nodes guided by it are aggregated for dynamic modeling, that is, the semantic information between the target node and its neighbors is aggregated to obtain rich embeddings of users and travel modes. Finally, the feature embeddings of users and travel modes are fused to recommend multimodal travel modes.

[0064] First, different types of node features in the heterogeneous graph are mapped to the same vector space, and a specific type of linear transformation is applied to each type of node, where x v ∈R d is the original eigenvector, E v ∈R d′ is the projected latent vector of node v. A ∈Rd′×d is the parameter weight matrix of type A nodes. After applying this operation, the projected features of all nodes have the same dimension, which facilitates the aggregation process of the next model component:

[0065]

[0066] like Figure 4 The meta-path aggregation layer indicates that the user is first taken as the target node (here it can be a multi-hop meta-path or a single-hop meta-path), and then neighbor node features are aggregated one by one according to the selected meta-path. Taking the meta-path UVL as an example, the initial embedding of the travel mode and the start and end locations is known, and the aggregation process is as follows:

[0067] Aggregation between nodes within a single meta-path

[0068] Use the average aggregation function to aggregate the embeddings of the second-order neighbors, concatenate the aggregation results with the first-order neighbor nodes, and update the embeddings of the first-order neighbors:

[0069]

[0070] Similarly, the average aggregation function is used to aggregate the first-order neighbor node embeddings and update the feature embedding of the target node:

[0071]

[0072] Aggregation between different meta-paths

[0073] In this recommendation model, time plays a key role in recommending travel modes that meet personalized needs to users. Therefore, when dynamically modeling users, users have temporal characteristics. Here, the LSTM aggregation function is used to better aggregate the embedding of users under different meta-paths. i are different meta paths with the user as the target node:

[0074]

[0075] All users are dynamically modeled in the above way to obtain the final embedding of all user nodes:

[0076]

[0077] The dynamic modeling of travel modes is similar to that of the original model. According to different meta-paths with travel modes as target nodes, the average aggregation function is first used to aggregate the features of neighbor nodes within the meta-path one by one, and then the CNN aggregation function is used to complete the feature aggregation between different meta-paths to obtain the final embedding of all travel mode nodes. i These are different meta-paths with travel modes as target nodes:

[0078]

[0079] Finally, the user and travel mode nodes are embedded and connected and fused, and finally input into the MLP to predict the score.

[0080]

[0081] f is an MLP layer with only one output, sigmoid is the activation layer, is the embedding chaining operator.

[0082] The loss function in this model uses a point-by-point loss function. The model is adjusted through the loss function to obtain the optimal result. They are respectively the instance sets of positive samples and negative samples:

[0083]

[0084] 3) Prediction results

[0085] In this paper, the entire user travel trajectory data of the dataset is randomly divided into a training set (80%) and a test set (20%), and the K-th precision (Prec@K), K-th recall (Recall@K) and K-th normalized discounted cumulative gain (NDCG@K) are used as evaluation indicators. The larger the NDCG value, the better the performance. For stability, we use different random split training / test sets for multiple runs and take the average results.

[0086] The present invention considers three different representative recommendation methods for effect comparison. The first is the collaborative filtering-based method: itemKNN, MF, SVD; then the more advanced neural network-based model: GMF, MLP, NeuMF; and finally the HIN-based model: metapath2vec; as well as several variants of the proposed model.

[0087] like Figure 5 The performance changes of various indicators of the model in the process of adding meta-paths are shown. It can be observed that the recommendation performance of the model is significantly improved by adding more meta-paths. At the same time, different meta-paths have different effects on the recommendation performance. In particular, we can find that when UVUV is added, the performance is significantly improved. Here, similar users of the user are added, which is no longer just the preferences reflected by the user's personal data. For example, their nearby neighbors or friends may have similar preferences for travel modes, which is also the role of the semantic information extracted by the meta-path. The results show that adding meta-paths plays an important role in learning task-related embeddings, and also illustrate that the heterogeneous graph neural network method guided by meta-paths proposed in this paper is very effective for recommending multimodal travel modes.

[0088] This article explains this from two perspectives. Under the same conditions for the training set and the test set, different aggregation function models are used to compare the precision and recall performance indicators (such as Figure 6 As shown in the left figure). We divide the data into five equal parts, one of which is used as the test set, and the remaining data is stacked at 20%, 40%, 60%, and 80% as the training set, and compared with different aggregators (such as Figure 6 As shown in the right figure, we can find that as the training set data increases, the improvement of recommendation performance gradually decreases. The results show that using heterogeneous information networks to obtain richer semantic information is very effective in improving recommendation performance. The proposed recommendation model uses meta-paths to give full play to its advantages on heterogeneous information networks.

[0089] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the art should understand that certain modifications and changes can be made to the present invention without exceeding the spirit and essence of the present invention, but they should be included in the protection scope of the present invention.

Claims

1. A multi-modal travel mode fusion recommendation method based on a dynamic traffic network, characterized in that: The following steps are involved: Step 1: It is proposed to apply heterogeneous information networks to transportation, and to construct a heterogeneous information network with the information of users, travel modes, starting and ending locations, and the spatiotemporal attributes of travel modes; Step 2: Based on the heterogeneous information network constructed in step 1, pre-process the spatiotemporal trajectory data, perform validity processing on the user's spatiotemporal travel trajectory data, extract the user, travel mode, starting and ending locations, and time, construct a mapping relationship, and obtain application data; Step 3: Extract key meta-paths from the network patterns corresponding to the heterogeneous information network to enrich the feature embedding of the object; Step 4: Aggregate meta-path features through aggregation functions; Step 5: Use the graph neural network guided by the meta-path to dynamically model users and travel modes, and obtain feature embedding with rich interactive information. Users are temporal, so LSTM dynamic modeling is used, and travel modes are dynamically modeled using CNN; Step 6: The final features of users and travel modes are embedded into the input MLP to fully learn the user's preference for travel mode in the spatiotemporal travel trajectory and recommend travel modes that meet personalized needs to users; Step 1 builds a heterogeneous information network as follows: Step 1-1: Define a set M = {U, V, T, L}, where U represents the set of users; V represents the travel mode used by the user in the travel trajectory, including subway, bus, driving, taxi, self-driving car, and walking; T represents the start time when the user uses a certain travel mode; L represents the start and end locations of the user using a certain travel mode at a certain time; It belongs to the node type in the heterogeneous transportation travel network; Step 1-2: In a multimodal transportation network, users take different modes of travel to reach their destinations, and there are certain connections between them. A heterogeneous information network is constructed by combining users, travel modes, starting and ending locations, and the spatiotemporal attributes of travel modes. More potential features are obtained by learning the semantic relationships between users, users and travel modes, and travel modes and their corresponding starting and ending location objects, fully expressing users' personalized preferences for travel modes. The spatiotemporal travel trajectory preprocessing stage described in step 2 includes the following: Step 2-1: Data cleaning: Before calculating the features of the data, data filtering is required: cleaning redundant data, deleting records with missing important fields and field values, and data screening; Step 2-2: Feature attribute calculation: The four features of speed, acceleration, acceleration change rate, and steering angle are calculated as trajectory attributes; in CNN, all data segments are limited to a fixed size of M; longer data segments are truncated to M, and shorter data segments are filled with zero values; after generating segments with the same size, two consecutive GPS trajectory points are used to calculate the attribute value of each GPS point; Step 2-3: Identify the stop points and divide the trip: The trip is divided by identifying the stop points of the user in the trajectory. Each trip contains multiple travel mode segments. The DBSCAN clustering algorithm is used to identify the stop points, and the corresponding trajectory point semantics are extracted in combination with the POI library. Step 2-4: Extract travel mode segments by segmenting the trip based on the identification of conversion points: Use the walking segment-based method to segment the single travel mode segment; Step 2-5: Identify travel modes based on convolutional neural networks: Integrate the characteristic attributes of each travel mode segment, let the convolutional neural network capture the characteristics of the mode segment, and finally complete the identification of travel modes in GPS trajectories; Step 2-6: Establish a mapping relationship between users, travel modes, and starting and ending locations to obtain application data; Step 3 of extracting key meta-paths from the network model corresponding to the heterogeneous information network includes the following: Step 3-1: In order to better understand the heterogeneous information network, a network model is constructed, which stipulates the type constraints of the object set and the connection between objects: there is a connection between users and travel modes, indicating the relationship of use and being used; there is a connection between locations and travel modes, indicating the relationship of guidance and being guided; Step 3-2: Extract four meta-paths: "user U—travel mode V—starting and ending locations L", "user—travel mode—user—travel mode", "travel mode—user—user—starting and ending locations", and "travel mode—user—starting and ending locations—user—travel mode"; The implementation of meta-path feature aggregation through the aggregation function in step 4 includes the following: Step 4-1: Based on the four meta-paths proposed in step 3-2, count the neighbors guided by the meta-path: The first-order neighbors of user u2 under the meta-path Φ:UVL are The second-order neighbors of u2 are the first-order neighbors of all nodes in the first-order neighbors, so its second-order neighbors are All neighbors of user u2 are denoted as N u ={u2,v1,v2,v3,v4,l1,l2,l3,l4,l5,l6}; Step 4-2: Select a suitable aggregation function to aggregate the features of the meta-path and define two meta-paths Φ uvl :UVL,Φ UVUV :UVUV, given the initial feature vectors of each node, the two processes of meta-path aggregation are described as follows: Meta-path internal aggregation: Select a suitable aggregation function to aggregate the embeddings of the second-order neighbor nodes to obtain the embeddings of the first-order neighbor nodes, and then aggregate the embeddings of the first-order neighbor nodes to obtain the embedding of the target node, as shown below: f is the aggregation function for aggregating the first-order and second-order neighbor embeddings of u2 on the specified meta-path, is the second-order neighbor node embedding of user u2 on the meta-path, is the first-order neighbor node embedding, which is updated by aggregating the second-order neighbor node embeddings, and then aggregating the first-order neighbor embeddings to get the user u2 in the meta-path Φ uvl The final embedding on ; Aggregation between meta-paths: Aggregate the embeddings of user u2 obtained from different meta-paths to obtain the final embedding of u2: The process of using the graph neural network guided by the meta-path to dynamically model the user and travel mode and obtain feature embedding with rich interactive information in step 5 includes the following: Step 5-1: Dynamically model the user. First, use the average aggregation function to aggregate the embeddings of the second-order neighbors and the first-order neighbors. The aggregation result is spliced ​​with the first-order neighbor nodes, and the embedding of the first-order neighbors is updated. Users have time sequence, and the LSTM aggregation function is used here to aggregate the embeddings of users under different meta-paths. Step 5-2: Dynamically model the travel mode. First, use the average aggregation function to aggregate the features of neighbor nodes within the meta-path one by one. Then use the CNN aggregation function to complete the feature aggregation between different meta-paths to obtain the final embedding of all travel mode nodes: In step 6, the final features of the user and the travel mode are embedded into the input MLP to fully learn the user's preference for travel mode in the spatiotemporal travel trajectory, and the process of recommending travel modes that meet the personalized needs of the user includes the following: Step 6-1: Based on the user feature embedding obtained in step 5-1 and the travel mode feature embedding obtained in step 5-2, a fusion prediction is performed, that is, Among them, f is an MLP layer with only one output, sigmoid is the activation layer, is the embedding chaining operator; Step 6-2: Perform multiple trainings, use Precision, Recall, and NDCG as accuracy evaluation indicators, and observe the accuracy of recommendations.

Images