Staying hotspot sequence inference method based on space-time dual graph representation

Abstract

The invention discloses a space-time dual graph representation-based staying hot spot sequence inference method, which comprises the following steps of: constructing a space-time staying hot spot transfer relation graph based on a waybill track and staying hot spot data, and modeling a relation between a staying hot spot and a driving path; constructing a dual graph coding method based on a graph convolution variational auto-encoder model, and introducing an attention mechanism to jointly learn a space-time dynamic relationship between a staying hot spot and a driving path; performing embedding generation on the query request by using a denoising diffusion model, and capturing correlation between a historical transportation route and a current query; on the basis of a decoder of a recurrent neural network, a mask mechanism is introduced, an unreasonable staying hotspot sequence is effectively pruned, and the inference precision and efficiency are improved; performing model optimization by adopting multi-stage joint loss, and performing training by using an Adam optimizer; and deducing a staying hotspot sequence conforming to the driver behavior mode based on the historical track and the current query request. The method has wide application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent transportation and logistics information technology, and particularly relates to a stay hotspot sequence inference method based on space-time dual graph representation in a bulk freight scene. BACKGROUND

[0002] In the long-distance transportation process of bulk freight, the driver's stay behavior has an important influence on the prediction accuracy of transportation arrival time. Compared with short-distance transportation in urban road networks, bulk freight usually involves longer driving distances, longer transportation times, and wider road network coverage, and a single transportation task may last for ten hours or even several days. The "Regulations on the Implementation of the Road Traffic Safety Law of the People's Republic of China" stipulates that the driver must stop for a rest of 20 minutes after driving a motor vehicle for 4 hours. In addition, in order to meet the needs of rest, dining, refueling, etc., the driver will make multiple voluntary stops during transportation. However, the location, duration and frequency of these stay behaviors are dynamically affected by various factors, including departure time, work period restrictions of transportation start and end points, road traffic rules, weather conditions, etc. Therefore, how to accurately infer the stay hotspot sequence of the driver in a specific transportation task not only plays an important role in improving the rationality of transportation planning, but also effectively reduces the uncertainty of transportation time and improves the overall logistics efficiency.

[0003] In current research and application, by organically integrating the modeling advantages of RNN and graph neural network, the spatio-temporal correlation characteristics contained in the distribution point sequence are deeply mined, and significant results are achieved in the intelligent path planning task of urban distribution logistics. These methods capture the time dependence of the sequence access of distribution points through RNN, and model the spatial constraints of the road network topology using GNN, forming a relatively complete spatio-temporal modeling framework. However, existing methods mainly focus on modeling the correlation patterns between static distribution points, and fail to fully consider the dynamic behavior characteristics unique to the freight scene, leaving the stay hotspot inference facing several challenges. First, the driver's choice of the current stay hotspot is influenced by multiple complex factors, such as the start and end points, the previous stay hotspot, the departure time, the working time limit of the transportation end point, the weather, etc., and the influence weight of these factors is dynamically changing. For example, if the unloading time of the transportation end point is limited, the driver may choose to rest at a location close to the end point to ensure timely delivery. In addition, traditional methods ignore the complex interaction between stay hotspots and driving paths, and how to fully capture the complex factors that change dynamically, especially the complex interaction between stay hotspots and driving paths, is the primary challenge of stay hotspot sequence inference. Second, in the bulk freight scenario, drivers have a large number of potential stay hotspots during driving, but they do not stay at all hotspots, but choose based on individual needs. Existing research has not considered the low search efficiency and local optimal solution problem caused by the huge search space. The limitations of this modeling perspective result in the obvious inadequacy of the reliability and practicality of the prediction results of existing methods in handling the cross-regional, long-time efficient transportation tasks unique to bulk logistics. The direct inference of the complete stay hotspot sequence in existing methods has high computational cost and a huge search space, which may lead to low efficiency of prediction. Therefore, how to effectively reduce the search space and improve the accuracy and efficiency of stay hotspot sequence inference is another key challenge of stay hotspot inference task. Therefore, there is an urgent need for an efficient and accurate stay hotspot inference method to capture the stay patterns of drivers and improve the reliability of transportation scheduling and arrival time prediction. SUMMARY

[0004] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a stay hotspot sequence inference method based on spatio-temporal dual graph representation, which constructs a spatio-temporal stay hotspot transfer relationship graph, learns the dynamic relationship between stay hotspots using graph neural network (GCN-VAE), and captures the correlation between stay hotspots and driving paths by combining multi-head attention mechanism. In addition, the self-encoder is introduced to mine the periodicity of stay hotspots and improve the modeling ability of long-term behavior patterns. Finally, the recurrent neural network is used for stay hotspot sequence inference, and the mask mechanism is used to prune low correlation options, thereby improving the prediction accuracy and computational efficiency. This method can effectively adapt to the bulk freight scenario and improve the reliability of stay behavior prediction, providing data support for transportation path planning, transportation scheduling, etc.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows: A method for inferring dwell hotspot sequences based on spatiotemporal dual graph representation includes the following steps: Step S1: Based on waybill trajectory and stop hotspot data, construct a spatiotemporal stop hotspot transfer relationship graph to model the relationship between stop hotspots and travel routes.

[0006] Step S2: Construct a dual graph encoding method based on the GCN-VAE (Graph Convolutional Variational Autoencoder) model, and introduce an attention mechanism to jointly learn the spatiotemporal dynamic relationship between dwell hotspots and driving paths.

[0007] Step S3: Use the Denoising Diffusion Model (DDPM) to embed the query request and capture the correlation between historical transportation routes and the current query.

[0008] Step S4: A decoder based on a recurrent neural network (LSTM) is used, and a masking mechanism is introduced to effectively prune unreasonable hot spots, thereby improving inference accuracy and efficiency.

[0009] Step S5: Optimize the model using multi-level joint loss and train it using the Adam optimizer.

[0010] Step S6: Based on historical trajectories and current query requests, infer the sequence of stop hotspots that match the driver's behavior pattern.

[0011] To optimize the above technical solution, the specific measures also include: The above step S1 specifically includes: To effectively infer the actual stop sequences of freight drivers and reveal the transfer structure between stop points, a spatiotemporal stop hotspot transfer relationship graph, denoted as G=(V,E), is first constructed based on historical trajectory information, where V represents the set of stop hotspots and E represents the connecting edges between hotspots. Then, a snapshot of the spatiotemporal stop hotspot transfer relationship graph corresponding to the query time is extracted. Next, the node graph snapshot is transformed to obtain the corresponding edge graph. Finally, the node graph and edge graph in the relationship graph are combined to form a dual graph. The dual graph contains information about the stop hotspots and explicitly presents the reachable paths between them.

[0012] Specifically, for a stop hotspot sequence transformed from a waybill trajectory For each consecutive hotspot sequence pair Using consecutive hotspots as nodes, connect nodes are constructed. and edge .

[0013] To further extract the time-dimensional dependencies in the hotspot transfer relationship graph, we first extract the current timestamp. Before A time slice consists of several time slices, i.e. Then extract each time slice. Node features and edge features to generate time slices A snapshot of the hotspot shift relationship diagram. (Each time slice) Snapshot of the image Includes node features .

[0014] For cases with missing data, a nearest neighbor fill strategy is used: (Use data from the previous time slice if features of the current time slice are missing) This strategy ensures the continuity of temporal features.

[0015] Finally, construct the dual graph. To simultaneously model the relationship between dwell hotspots and paths, the original graph... Convert to a graph containing edges Dual graph structure: Node graph The adjacency matrix A satisfies: If two hotspots can be reached directly by a path, then the corresponding element in the adjacency matrix A is 1; otherwise, it is 0. Side graph The adjacency matrix A' satisfies: If two edges have any common vertex, then the corresponding element in the adjacency matrix A' is 1; otherwise, it is 0. The final generated dual graph representation includes: Node feature matrix: ,in This represents the dimension of the feature vector of each node; Edge feature matrix: ,in The dimension of the feature vector of each edge; Dual adjacency: and ; The dual graph is represented as ; This representation method preserves both spatial topology and temporal dynamic features, providing a complete data foundation for subsequent graph neural network processing.

[0016] The above step S2 specifically includes: Given that Graph Convolutional Networks (GCNs) are used to capture neighborhood information of graph nodes to learn the relationships between nodes, while Variational Autoencoders (VAEs) capture the latent structure of data by introducing latent variables and probability distributions, this invention constructs a GCN-VAE model that combines the advantages of both. It aims to obtain the embedded representations of node graphs and edge graphs to learn the rich local feature information and global potential transition relationships between hotspots and edges.

[0017] Specifically, before inputting node and edge features into GCN-VAE, initial representations of nodes and edges are obtained through embedding layers and linear layers, respectively.

[0018] Node characteristics: For the categories of popular stays Discrete features are projected onto an embedding layer. 3D vector space; and for the average dwell time Frequency of passage For continuous features, a linear layer is used to project them onto... A 3D vector space is used to obtain feature embedding representations for discrete and continuous features. The specific definitions of the above feature embedding methods are as follows: ; in, and These are the weights and biases of the linear layer, used for projecting continuous features. Represents continuous node characteristics; This represents the embedding layer, used for projecting discrete features. Represents discrete node characteristics; symbol This represents a splicing operation; Indicates the number of vertices; It is the dimension of the embedding vector, which is equal to the sum of the dimension of the continuous feature projection and the dimension of the discrete feature embedding.

[0019] Edge features: The initialization representation of the edge graph node features is defined as follows: ; in and These are the weights and biases of the linear layer with edge features. Indicates continuous edge characteristics; Indicates the embedding layer. Represents discrete edge features; symbol This represents a splicing operation; It is the dimension of the embedded vector. Let be the number of edges.

[0020] The node graph is used to obtain node embeddings through feature propagation via GCN: ,in, It is a normalized adjacency matrix of a node graph. It is the weight matrix of the l-th layer GCN corresponding to the node graph.

[0021] Similarly, edge features are processed by another GCN to obtain edge embeddings: ,in, It is the normalized adjacency matrix of the edge graph. It is the weight matrix of the l-th layer GCN corresponding to the edge feature map.

[0022] To map node features and edge features to the same latent space, a shared VAE architecture was designed, whose encoder part maps node embeddings to the latent space and generates the distribution parameters of the latent variable z: as well as ;in It is a multilayer perceptron of an encoder.

[0023] Latent variables Sampling was performed using the following methods: ;in, This represents a multivariate standard normal distribution with a mean of 0 and a covariance equal to the identity matrix. This indicates element-wise multiplication; Decoder module reconstructs node features: ; in It is the multilayer perceptron of the decoder. At this point, each time slice is obtained. Feature vector matrix of dwell hotspots and transportation routes in the spatial dimension .

[0024] Furthermore, for the dwell hotspot sequence corresponding to the trajectory This is mapped to an embedded representation of a sequence of nodes and a sequence of edges. The specific processing procedure is as follows: Node sequence embedding construction: Edge sequence embedding construction: ,in, This is the feature vector for the query request.

[0025] Multi-head self-attention mechanisms are used to integrate mutual information between hotspots and / or transportation paths in waybill trajectories. Specifically, multi-head self-attention mechanisms are applied to node sequences and edge sequences respectively: in: , For learnable parameters, ; Finally, the bullish attention output is: ,in: This is for outputting the projection matrix.

[0026] This is the embedding vector of the node sequence; Similarly, the embedding vector of the edge sequence is obtained based on the multi-head self-attention mechanism described above: Let be the embedding vector of the edge sequence, where:

[0027] To integrate the mutual information between dwell hotspots and travel routes in a waybill trajectory, this invention further models the embedded representations of node and edge sequences based on a dual cross-attention mechanism. Specifically, the first cross-attention mechanism in the dual cross-attention mechanism is used to learn the influence of transport routes on dwell hotspots. Specifically, a bidirectional cross-attention mechanism is used to integrate node and edge sequence information: Node → Edge Attention:

[0028]

[0029] Edge → Node Attention:

[0030] Finally, the updated node and edge sequence representations are concatenated to obtain the journey representation: .in, Indicates the length of the node sequence.

[0031] The above step S3 specifically includes: Since all the stopping locations for each transportation task can be known in advance during training, the actual transportation routes (i.e., the stopping hotspot sequence) cannot be obtained during the actual prediction phase. Therefore, the goal of this step is to address the issue given a query request. and its corresponding historical transportation routes (represented as) This involves learning the correlation between historical transportation routes and query requests to generate transportation route embeddings. Most methods for POI (point of interest) recommendation tasks are based on Generative Adversarial Networks (GANs) and contrastive learning. However, GANs rely on large amounts of labeled data and are difficult to train, while contrastive learning, although it can learn in an unsupervised or self-supervised manner and reduce its dependence on labeled data, has a poor ability to capture the underlying distribution of the data. Inspired by the successful application of denoising diffusion models in trajectory generation and other fields, this step utilizes a denoising diffusion probabilistic model (DDPM) to model the data distribution, which mainly consists of a forward diffusion process and a backward denoising process. The forward diffusion process gradually introduces noise into the signal, eventually transforming it into a simple prior distribution (e.g., a Gaussian distribution). To determine the underlying distribution based on the query request... The process generates an embedded representation of the line, and the inverse denoising process progressively transforms the noisy vectorized line embedded representation into a noise-free representation. In other words, the inverse denoising process gradually removes noise, ultimately generating a representation that is consistent with the original signal. The corresponding noiseless line embedding representation.

[0032] In the denoising diffusion model, the key to generating the transportation route embedding representation lies in learning a neural network-based denoiser. Used to predict time steps The amount of noise is used to generate a representation for n-1 time steps. Inspired by the successful application of Unet in various diffusion models, this step introduces it into the line embedding representation generation task. This invention designs, as follows: Figure 1 The denoising unit consists of M downsampling blocks and upsampling blocks, each containing components such as a residual block and a Transformer. Downsampling is performed using max pooling, while upsampling is achieved through linear interpolation.

[0033] To effectively learn the noise at each time step and guide the generation of the line embedding representation, the denoiser first performs an embedding query request through the coding layer. and time step Specifically, time step The sampled blocks are encoded using positional encoding techniques and concatenated with the feature vector of the query request before being input into each residual block. Subsequently, each sampled block undergoes an attention mechanism (i.e., ...). Transformer The component learns the correlation between query requests and routes to obtain output. .at last, Further processing is performed using standard residual network blocks in Unet.

[0034] The above step S4 specifically includes: This step aims to be based on the transportation route embedding representation. Embedded representation of hotspots and transportation routes The study uses an LSTM equipped with a masked attention mechanism to predict the hotspot sequence at the starting point.

[0035] The decoding process in this step is performed iteratively. Its core idea is that in each decoding step, the output node from the previous step (represented as...) is used... The most likely neighbor node is selected as the next visited node (represented as...). The selected node is then concatenated with the embedded representation of the edge connecting it, and input into a single-layer LSTM to record the current decoding state, thereby guiding the next step of the selection. Specifically, in the first... During the decoding process, The attention query vector is generated by concatenation, and the neighbor node embedding representation of the output node in the previous step is used as the key vector to generate the output probability distribution of the neighbor node.

[0036] To ensure the rationality of the predicted route, this step introduces a masking mechanism to filter the candidate transfer node set. The masking rules include the following three aspects: (1) Masking already visited nodes: avoiding repeated visits to already visited nodes, i.e., masking the set of nodes. (2) Masking based on historical data: Based on historical access frequency and distance information, the masking transition probability is lower than the threshold. The nodes, i.e., the mask set (3) Masking based on driver's historical trajectory: According to the driver's historical transition probability distribution, the nodes with a transition probability of zero are masked, i.e., the mask set. .

[0037] The masked node will not be selected as a candidate node. The masking mechanism in this step is implemented using the following formula:

[0038] in, For learnable parameters, The query vector is tanh, and the activation function is tanh. Mask node set. Defined as: .

[0039] Finally, this step uses the softmax layer to calculate the... The probability of each output node in each step:

[0040] Based on the above formula, this step selects the node with the highest probability as the current step output, and iterates until the endpoint is reached or the preset iteration limit is reached.

[0041] The above step S5 specifically includes: The model training process of this invention is mainly divided into two stages. The first stage is reconstruction training based on a decoder architecture that combines a waybill trajectory encoder and sequence inference prediction. This stage aims to extract effective graph embedding representations from the waybill trajectory using the GCN-VAE model, and then reconstruct the dwell hotspot sequence from the decoding module using the waybill trajectory encoding results. The loss function of GCN-VAE includes reconstruction error, which measures the difference between the original features and the reconstructed features, and KL divergence, which measures the difference between the latent variable distribution and the standard normal distribution. The specific form of the loss function is as follows:

[0042]

[0043] The first term is the reconstruction error, which is usually measured using mean squared error or cross-entropy loss; the second term is the KL divergence, which is used to regularize the distribution of latent variables to make them closer to the standard normal distribution.

[0044] Furthermore, this invention treats route prediction as a multi-class classification problem and uses the cross-entropy function as the loss function, which is formally defined as follows:

[0045] in It is the number of waybill trajectories in the training dataset. Indicates the first The number of hotspots for each waybill trajectory sample.

[0046] In summary, the total loss function of the first stage... The definition is as follows: ,in These are the weight coefficients for each loss term, used to balance the contribution of each component to the total loss. This training process yields the trained model. ; After completing the waybill trajectory encoding and sequence decoding steps, the second stage uses the waybill trajectory embedding representation generated in the waybill trajectory encoding step to train the denoiser used in the transportation route embedding representation generation step, thus obtaining the model. Specifically, this stage utilizes the waybill trajectory embedding representation output by the trajectory encoding module to specifically train the denoiser in the transportation route embedding generation module. The denoiser learns the spatial distribution characteristics of the transportation route embedding through a step-by-step denoising process.

[0047] First, considering that the core task of the reverse denoising stage is to predict the amount of noise to be removed at the current time step, a noise matching loss function is defined based on the mean squared error (MSE):

[0048] in, This represents the true noise sampled from a standard normal distribution. The noise predictor generates an estimate under current conditions, with input consisting of the diffused embedding representation and the time step. It consists of the query request and the predicted noise. This loss metric measures the error between the predicted noise and the actual noise, directly reflecting the model's ability to restore noise in denoising tasks.

[0049] However, relying solely on noise errors can lead to a lack of global structural consistency in the generated paths, especially when errors accumulate during multi-step denoising, potentially resulting in offsets or distortions. Therefore, this invention further introduces a structural alignment loss to constrain the final generated path representation. Embedded representation with the output of the trajectory coding module Similarity between them:

[0050] in, It is the final representation of the waybill trajectory sequence by the trajectory encoder. The noise-free path embedding reconstructed from the reverse process is calculated as follows:

[0051] This structural alignment term not only ensures controllable accuracy of the model's generated results at local points but also improves semantic consistency across the entire trajectory. The combined effect of these two losses allows the model to achieve a more stable optimization direction during training, while avoiding the limitation of optimizing only low-level noise targets and neglecting structural representation.

[0052] Finally, the overall loss function is defined as follows:

[0053] This objective function jointly considers both the accuracy of local noise prediction and the quality of global structure reconstruction, providing a dual guarantee for the generation of path embedding representations. This training process yields the trained model. .

[0054] The above step S6 specifically includes: First, based on the query request Matching the initial and lasting hotspots And calculate departure time update Next, construct the timeline from departure time. Start of time slice sequence And generate a snapshot sequence of hotspot transfer relationship diagrams based on the time slice sequence. Then, the trained model is used. Generate and embed a dwell hotspot map and edge embedding and using the model Generate the embedded representation of the transportation route corresponding to the query request. Finally, the model Used to infer transportation route sequences, ultimately returning the inferred stop hotspot sequences. .at last, Together They are sent together to the decoding module to infer the hot spot sequence.

[0055] The beneficial effects of this invention include: Addressing the problem that driver stop selection is dynamically influenced by multiple factors, this invention proposes a method combining a spatiotemporal stop hotspot transfer relationship graph and GCN-VAE dual graph encoding to capture the dynamic interaction between stop hotspots and driving paths. Furthermore, a denoising diffusion model is introduced to generate a high-dimensional transportation route embedding representation, and a recurrent neural network is used for sequence decoding. Specifically, to address the issue of a large stop hotspot search space, a masking mechanism is introduced during the decoding process to prune unreasonable stop hotspot selections, thereby improving the accuracy of sequence inference. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a framework diagram of a method for inferring dwell hotspot sequences based on spatiotemporal dual graph representation.

[0058] Figure 2 This is an example diagram illustrating the spatiotemporal hotspot transfer relationship.

[0059] Figure 3 This is an example diagram of a dual graph.

[0060] Figure 4 This is a framework diagram of a denoiser module based on Unet. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the present invention are all common knowledge and general knowledge in the art, and the present invention does not have any particular limitations.

[0062] This invention addresses the problem of driver stop selection being dynamically coupled by multiple factors and the large search space for stop hotspots, providing a stop hotspot sequence inference method based on spatiotemporal dual graph representation. The method first constructs a stop hotspot transition graph to model the complex dynamic relationship between stop behavior and driving patterns, using GCN-VAE for graph structure encoding and incorporating a multi-head attention mechanism to capture the interaction relationships between stop hotspots and their interaction with the driving path. Furthermore, this invention introduces an autoencoder to model the periodic dynamic changes of stop hotspots, thereby enhancing the model's ability to learn long-term behavioral patterns. To reduce the search space for stop hotspot sequence inference, this invention designs a transportation route embedding representation generation method based on a denoising diffusion model (DDPM), dynamically generating high-dimensional representations of stop hotspot sequences based on query requests and combining this with LSTM for sequence decoding. To further improve prediction accuracy and computational efficiency, this invention introduces a masking mechanism during the decoding process to prune unreasonable stop hotspot selections, thereby optimizing the accuracy and computational efficiency of stop hotspot sequence inference.

[0063] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0064] This invention discloses a framework for inferring dwell hotspot sequences based on spatiotemporal dual graph representation, such as... Figure 1 As shown, the overall framework for hotspot sequence inference is an encoder-decoder architecture, mainly divided into three modules: dual graph encoding based on GCN-VAE, query request embedding generation based on denoising diffusion model, and sequence decoding based on masking mechanism.

[0065] like Figure 1 As shown, the present invention employs a six-step method for inferring dwell hotspot sequences, including the following steps: S1: Based on waybill trajectory and stop hotspot data, construct a spatiotemporal stop hotspot transfer relationship graph to model the relationship between stop hotspots and travel routes.

[0066] In this embodiment, step S1 specifically includes: To facilitate the subsequent modeling and representation of hotspots and their reachable paths, a hotspot transfer graph is first constructed based on historical waybill trajectories, with hotspots as nodes and reachability relationships between hotspots as edges. Then, a snapshot of the hotspot transfer graph at the corresponding query time is extracted. Next, the node graph snapshot is transformed to obtain the corresponding edge graph. Finally, the node graph and edge graph are combined to form a dual graph. The dual graph contains information about hotspots and explicitly presents the reachable paths between hotspots.

[0067] Specifically, for a stop hotspot sequence transformed from a waybill trajectory For each consecutive hotspot sequence pair Build connection nodes and edge .

[0068] To further extract the time-dimensional dependencies in the hotspot transfer relationship graph, we first extract the current timestamp. Before A time slice consists of several time slices, i.e. Then extract each time slice. Node features and edge features to generate time slices A snapshot of the hotspot shift relationship diagram. (Each time slice) Snapshot of the image Includes the following features: Node features Includes both static and dynamic features:

[0069] in: Embed for hotspot types (such as service areas, gas stations, etc.) latitude and longitude coordinates Number of visits within a time slice These represent the average length of stay and the median, respectively. Edge features Reflecting path attributes:

[0070] in: Euclidean distance Number of passes These represent the average driving time and the median, respectively. For cases with missing data, a nearest neighbor fill strategy is used: (Use data from the previous time slice if features of the current time slice are missing) This strategy ensures the continuity of temporal features.

[0071] Finally, construct the dual graph. To simultaneously model the relationship between dwell hotspots and paths, the original graph... Convert to a graph containing edges Dual graph structure: Node graph The adjacency matrix A satisfies:

[0072] Side graph The adjacency matrix A' satisfies:

[0073] The final generated dual graph representation includes: Node feature matrix:

[0074] Edge feature matrix:

[0075] Dual adjacency: and

[0076] S2: A dual graph encoding method based on GCN-VAE (Graph Convolutional Variational Autoencoder) is proposed, and an attention mechanism is introduced to jointly learn the spatiotemporal dynamic relationship between dwell hotspots and driving paths.

[0077] In this embodiment, step S2 specifically includes: Given that Graph Convolutional Networks (GCNs) are used to capture neighborhood information of graph nodes to learn the relationships between nodes, while Variational Autoencoders (VAEs) capture the latent structure of data by introducing latent variables and probability distributions, this invention designs a GCN-VAE model that combines the advantages of both. The aim is to obtain the embedded representations of the node graph and the edge graph to learn the rich local feature information and global potential transition relationships between hotspots and edges.

[0078] Specifically, before inputting node and edge features into GCN-VAE, initial representations of nodes and edges are obtained through embedding layers and linear layers, respectively. For the hotspot category... Discrete features are projected onto an embedding layer. A dimensional vector space. And for the average dwell time... Frequency of passage For continuous features, a linear layer is used to project them onto... A dimensional vector space. The specific definitions of the above feature embedding methods are as follows:

[0079] in, and These are the weights and biases of the linear layer, used for projecting continuous features. Represents continuous node characteristics; This represents the embedding layer, used for projecting discrete features. Represents discrete node characteristics; symbol This represents a splicing operation; Indicates the number of vertices; It is the dimension of the embedding vector, which is equal to the sum of the dimension of the continuous feature projection and the dimension of the discrete feature embedding.

[0080] The initialization representation of the edge graph node features is defined as follows:

[0081] in and These are the weights and biases of the linear layer with edge features. Indicates continuous edge characteristics; Indicates the embedding layer. Represents discrete edge features; symbol This represents a splicing operation; It is the dimension of the embedded vector. Let be the number of edges.

[0082] Node graphs propagate features via GCN:

[0083] in, It is a normalized adjacency matrix. This is the weight matrix of the l-th layer GCN. Similarly, the edge features are processed by another GCN to obtain the edge embeddings: ,in, It is the normalized adjacency matrix of the edge graph.

[0084] To map node features and edge features to the same latent space, a shared VAE architecture was designed, whose encoder part maps node embeddings to the latent space and generates the distribution parameters of the latent variable z: as well as , It is a multilayer perceptron of an encoder.

[0085] Latent variables Sampling was performed using the following methods: ;in, This represents a multivariate standard normal distribution with a mean of 0 and a covariance equal to the identity matrix. This indicates element-wise multiplication; Decoder partially reconstructs node features: ; in The decoder's multilayer perceptron. At this point, each time slice is obtained. Feature vector matrix of dwell hotspots and transportation routes in the spatial dimension .

[0086] Furthermore, for the dwell hotspot sequence corresponding to the trajectory This is mapped to an embedded representation of a sequence of nodes and a sequence of edges. The specific processing procedure is as follows: Node sequence embedding construction: Edge sequence embedding construction: ,in This is the feature vector for the query request.

[0087] Multi-head self-attention mechanisms are used to integrate mutual information between hotspots and / or transportation paths in waybill trajectories. Specifically, Apply multi-head self-attention mechanism to the node sequence and edge sequence respectively: in: , For learnable parameters,

[0088] Finally, the bullish attention output is:

[0089] in: This is the output projection matrix. Similarly, the embedding vector of the edge sequence is obtained based on the multi-head self-attention mechanism described above. .

[0090] To integrate the mutual information between dwell hotspots and travel routes in a waybill trajectory, this invention further models the embedded representations of node and edge sequences based on a dual cross-attention mechanism. Specifically, the first cross-attention mechanism in the dual cross-attention mechanism is used to learn the influence of transport routes on dwell hotspots. Specifically, a bidirectional cross-attention mechanism is used to integrate node and edge sequence information: Node → Edge Attention:

[0091]

[0092] Edge → Node Attention:

[0093] Finally, the updated node and edge sequence representations are concatenated to obtain the journey representation: .

[0094] S3: Use the Denoising Diffusion Model (DDPM) to embed the query request and capture the correlation between historical transportation routes and the current query.

[0095] In this embodiment, step S3 specifically includes: Since all the stopping locations for each transportation task can be known in advance during training, the actual transportation routes (i.e., the stopping hotspot sequence) cannot be obtained during the actual prediction phase. Therefore, the goal of this step is to address the issue given a query request. and its corresponding historical transportation routes (represented as) This involves learning the correlation between historical transportation routes and query requests to generate transportation route embeddings. Most POI recommendation tasks are based on Generative Adversarial Networks (GANs) and contrastive learning. However, GANs rely on large amounts of labeled data and are difficult to train, while contrastive learning, although it can learn in an unsupervised or self-supervised manner and reduce its dependence on labeled data, has a poor ability to capture the underlying distribution of the data. Inspired by the successful application of denoising diffusion models in trajectory generation and other fields, this step utilizes a denoising diffusion probabilistic model (DDPM) to model the data distribution, which mainly consists of a forward diffusion process and a backward denoising process. The forward diffusion process gradually introduces noise into the signal, eventually transforming it into a simple prior distribution (e.g., a Gaussian distribution). To determine the underlying distribution based on the query request... The process generates an embedded representation of the line, and the inverse denoising process progressively transforms the noisy vectorized line embedded representation into a noise-free representation. In other words, the inverse denoising process gradually removes noise, ultimately generating a representation that is consistent with the original signal. The corresponding noiseless line embedding representation.

[0096] In the denoising diffusion model, the key to generating the transportation route embedding representation lies in learning a neural network-based denoiser. Used to predict each time step The amount of noise, thus generating Inspired by the successful application of Unet in various diffusion models, this step introduces it into the line embedding representation generation task. This invention designs, as follows: Figure 1 The denoising unit consists of M downsampling blocks and upsampling blocks. Each sampling block contains components such as a residual block and a Transformer. Downsampling is performed using max pooling, while upsampling is achieved through linear interpolation.

[0097] To effectively learn the noise at each time step and guide the generation of the line embedding representation, the denoiser first performs an embedding query request through the coding layer. and time step Specifically, time step The sampled blocks are encoded using positional encoding techniques and concatenated with the feature vector of the query request before being input into each residual block. Subsequently, each sampled block undergoes an attention mechanism (i.e., ...). Transformer The component learns the correlation between query requests and routes to obtain output. .at last, Further processing is performed using standard residual network blocks in Unet.

[0098] S4: Design a decoder based on a recurrent neural network (LSTM) and introduce a masking mechanism to effectively prune unreasonable hotspot sequences, thereby improving inference accuracy and efficiency.

[0099] In this embodiment, step S4 specifically includes: This step aims to be based on the transportation route embedding representation. Embedded representation of hotspots and transportation routes The method uses an LSTM equipped with a masked attention mechanism to predict the hotspot sequence at the starting point.

[0100] The decoding process in this step is performed iteratively. Its core idea is that in each decoding step, the output node from the previous step (represented as...) is used... The most likely neighbor node is selected as the next visited node (represented as...). The selected node is then concatenated with the embedded representation of the edge connecting it, and input into the LSTM to record the current decoding state, thus guiding the next step of the selection. Specifically, in the first... During the decoding process, The attention query vector is generated by concatenation, and the neighbor node embedding representation of the output node in the previous step is used as the key vector to generate the output probability distribution of the neighbor node.

[0101] To ensure the rationality of the predicted route, this step introduces a masking mechanism to filter the candidate transfer node set. The masking rules include the following three aspects: (1) Masking already visited nodes: avoiding repeated visits to already visited nodes, i.e., masking the set of nodes. (2) Masking based on historical data: Based on historical access frequency and distance information, the masking transition probability is lower than the threshold. The nodes, i.e., the mask set (3) Masking based on driver's historical trajectory: According to the driver's historical transition probability distribution, the nodes with a transition probability of zero are masked, i.e., the mask set. .

[0102] The masking mechanism in this step is implemented using the following formula:

[0103] in, For learnable parameters, For query vectors. Mask node set. Defined as: .

[0104] Finally, this step uses the softmax layer to calculate the... The probability of each output node in each step:

[0105] Based on the above formula, this step selects the node with the highest probability as the current step output, and iterates until the endpoint is reached or the preset iteration limit is reached.

[0106] S5: The model is optimized using a multi-level joint loss and trained using the Adam optimizer.

[0107] In this embodiment, step S5 specifically includes: This invention is mainly trained in two stages. The first stage is reconstruction training based on a decoder architecture that combines a waybill trajectory encoder and sequence inference prediction. This stage aims to extract effective graph embedding representations from waybill trajectories using the GCN-VAE model, and then reconstruct the dwell hotspot sequence from the decoding module using the waybill trajectory encoding results. The GCN-VAE loss function includes reconstruction error, which measures the difference between the original features and the reconstructed features, and KL divergence, which measures the difference between the latent variable distribution and the standard normal distribution. The specific form of the loss function is as follows:

[0108]

[0109] The first term is the reconstruction error, which is usually measured using mean squared error or cross-entropy loss; the second term is the KL divergence, which is used to regularize the distribution of latent variables to make them closer to the standard normal distribution.

[0110] Furthermore, this invention treats route prediction as a multi-class classification problem and uses the cross-entropy function as the loss function, which is formally defined as follows:

[0111] in It is the number of waybill trajectories in the training dataset. Indicates the first The number of hotspots for each waybill trajectory sample.

[0112] In summary, the total loss function of the first stage... The definition is as follows: in It is the weighting coefficient of each loss term, used to balance the contribution of each part to the total loss.

[0113] After completing the waybill trajectory encoding and sequence decoding steps, the waybill trajectory embedding representation generated by the waybill trajectory encoding step is used to train the denoiser used for the transportation route embedding representation generation step.

[0114] S6: Based on historical trajectories and current query requests, infer the sequence of stop hotspots that match the driver's behavior patterns.

[0115] In this embodiment, step S6 specifically includes: First, based on the query request Matching the initial and lasting hotspots And calculate departure time update Next, construct the timeline from departure time. Start of time slice sequence And generate a snapshot sequence of hotspot transfer relationship diagrams based on the time slice sequence. Then, the trained model is used. Generate and embed a dwell hotspot map and edge embedding and using the model Generate the embedded representation of the transportation route corresponding to the query request. Finally, the model Used to infer transportation route sequences, ultimately returning the inferred stop hotspot sequences. .at last, , Together They are sent together to the decoding module to infer the hot spot sequence.

[0116] Furthermore, the specific implementation process also explains the settings of the main optimal hyperparameters selected for the model. In the autoencoder-based dual graph representation learning module, the time slice sequence length of the dwell hotspot transition graph is set to 12, and the time slice size is set to 2 hours. The embedding representation dimension d of the nodes and edges of the dwell hotspot transition graph is set to 128, the number of layers L of the GCN is set to 3, and the number of attention heads P is set to 4. The number of self-attention and cross-attention layers in the dual cross-attention network is set to 4. In the denoising diffusion encoder-based transportation route embedding generation module, the denoising steps N are set to 800, the number of sampling blocks M of the route embedding denoiser is set to 4, and β1 and βn, used to control the noise level, are set to 0.0001 and 0.015, respectively. Finally, the model is built using PyTorch, trained using the Adam optimizer, with a learning rate of 0.005 and an upper limit of 1000 training iterations.

[0117] To verify the effectiveness of the proposed hotspot inference framework (hereinafter referred to as ST-GVDM), this invention selected a real dataset from a logistics company platform in Shandong Province. The experimental data is shown in Table 1. This dataset includes historical hotspot sequence data of bulk freight drivers in two directions. Five of the latest comparative methods in the trip recommendation field were used for experimental verification, including POI-Popularity, Query2Trip, BERTTrip, SHSI, Graph2Route, and CP-Route. This invention utilizes... , and The experiment was evaluated using three evaluation indicators, and the calculation formula is as follows:

[0118]

[0119]

[0120] Given two sequences of dwelling hotspots, This represents the actual sequence of hotspots for stays. This represents the predicted sequence of hotspots. Represents the first part of the true sequence A popular spot for staying Indicates the first part of the predicted sequence A popular spot for staying. Used to measure the top of the prediction results The hit rate of each trending topic. Indicates the length of the transportation route. and These represent the actual and predicted locations of popular tourist attractions, respectively. Represents geographical coordinates. Used to measure the degree of deviation between the predicted result and the true label. , , This indicates the accuracy of predicting hotspot pairs in the predicted route. Recall is represented by Pairs-F1, which is used to evaluate the sequential relationship of hotspot sequences in the predicted route.

[0121] The comparative evaluation results are shown in Table 2. In the data from the two cities of LinYi and QingDao, the hotspot sequence inference method of this invention performs well. , and It outperformed other comparison methods in all three evaluation metrics.

[0122] Table 1 Dataset Description Table

[0123] Table 2. Results of the comparative experiment on hotspot sequence inference.

[0124]

[0125] Experimental results show that BERT-Trip achieves better results than the greedy strategy-based POI-Popularity method. This may be attributed to BERT's masked language model, which can perform bidirectional context encoding, thus capturing driver dwell patterns more accurately. Based on this, the improved SHSI and the Seq2Seq-related Query2Trip method also show significant performance improvements compared to POI-Popularity. On the other hand, graph-based models outperform sequence-based methods in most cases. This is because the main limitation of sequence models is their difficulty in integrating global spatial information when handling long-distance dwell hotspot transitions, which can easily lead to prediction biases, such as unreasonable predictions due to excessively large spatial distances between adjacent dwell hotspots. Although graph model solutions (such as Graph2Route and CP-Route) introduce graph encoders to improve this problem, they still do not adequately model the complex relationship between the spatiotemporal interaction characteristics of dwell hotspots and driving paths and driver decision-making behavior, thus their prediction performance is not yet optimal. In summary, the ST-GVDM method proposed in this invention successfully solves the task of inferring dwell hotspot sequences in complex spatiotemporal environments by combining a dynamic spatiotemporal dual graph neural network model with a mask-based LSTM decoder, and demonstrates stronger predictive capabilities. Furthermore, it can be observed that most methods perform better on LinYi than on QingDao, which may be because the dwell hotspot sequences on LinYi are shorter, and the search space for the dwell hotspot sequence inference task is smaller than that on the QingDao dataset.

[0126] In summary, to address the issue of insufficient modeling of the complex interactions between hotspots and travel routes, this invention proposes a GCN-VAE dual graph encoding method incorporating an attention mechanism. This method captures the interactions between hotspots, between travel routes, and between hotspots and travel routes, fully integrating the features of both. Furthermore, a denoising diffusion model is introduced to generate a high-dimensional transportation route embedding representation, which is then combined with a recurrent neural network for sequence decoding. To address the problem of a large search space for hotspots, a masking mechanism is introduced during decoding to prune inappropriate hotspot selections, improving the accuracy and efficiency of sequence inference.

[0127] The scope of protection of this invention is not limited to the above embodiments. Any variations and advantages that can be conceived by those skilled in the art without departing from the spirit and scope of this invention are included in this invention and are protected by the appended claims.

Claims

1. A method for inferring dwell hotspot sequences based on spatiotemporal dual graph representation, characterized in that, Includes the following steps: Step S1: Based on waybill trajectory and stop hotspot data, construct a spatiotemporal stop hotspot transfer relationship diagram and model the relationship between stop hotspots and travel routes; Step S2: Construct a dual graph encoding method based on a graph convolutional variational autoencoder model, and introduce an attention mechanism to jointly learn the spatiotemporal dynamic relationship between dwell hotspots and travel paths; Step S3: Use a denoising diffusion model to embed and generate the query request, capturing the correlation between historical transportation routes and the current query; Step S4: A decoder based on a recurrent neural network is used, and a masking mechanism is introduced to effectively prune unreasonable hot spots, thereby improving inference accuracy and efficiency; Step S5: Optimize the model using multi-level joint loss and train it using the Adam optimizer; Step S6: Based on historical trajectories and current query requests, infer the sequence of stop hotspots that match the driver's behavior pattern.

2. The method according to claim 1, characterized in that, In step S1, the nodes in the spatiotemporal dwelling hotspot transfer relationship graph represent dwelling hotspots, and the edges represent the reachability relationships between different dwelling hotspots; The spatiotemporal hotspot transfer relationship graph is represented as G=(V,E), where V represents the set of hotspots and E represents the connecting edge between hotspots; Extract a snapshot of the spatiotemporal hotspot transfer relationship graph corresponding to the query time, convert the node graph into an edge graph, and combine the node graph and edge graph in the relationship graph to form a dual graph; And / or, The nearest neighbor filling strategy is used to complete the missing node features in the time slice; And / or, Original image Transform into dual graph Simultaneously modeling dwell hotspots and path relationships; among them, Indicates a set of hotspots where people stay. Indicates the connecting edges between hotspots. Represents the node feature matrix, Represents the adjacency matrix of a node graph. Represents the adjacency matrix of the edge graph.

3. The method according to claim 2, characterized in that, Extract current timestamp Before A time slice consists of several time slices, i.e. Then extract each time slice Node features and edge features to generate time slices A snapshot of the dwell hotspot transfer relationship graph is used to extract the spatiotemporal dwell hotspot transfer relationship graph's dependence in the time dimension; Each time slice Snapshot of the image Includes node features ; And / or, When using the nearest neighbor filling strategy, the node feature data in the previous time slice is used to replace the missing node feature data in the current time slice.

4. The method according to claim 1, characterized in that, In step S2, a GCN-VAE model is constructed to obtain the embedding representations of the node graph and edge graph, and local and global learning is performed. And / or, Discrete features are projected using an embedding layer, and continuous features are projected using a linear layer to obtain feature embedding representations of discrete and continuous features, which are then initialized. And / or, The node and edge graph are propagated through GCN embedding representations, which are then used by the VAE module to generate latent variables; And / or, The multi-head self-attention mechanism integrates information from the node sequence and the edge sequence respectively; And / or, A dual cross-attention mechanism is used to fuse information bidirectionally between nodes and edges to obtain a complete journey representation.

5. The method according to claim 4, characterized in that, For node features, the feature embedding initialization of the discrete and continuous features is expressed as follows: ;in, and These are the weights and biases of the linear layer, respectively. Indicates continuous node characteristics. Represents discrete node characteristics. Indicates the number of vertices; Indicates the embedding layer; symbol This represents a splicing operation; It is the dimension of the embedded vector; For edge features, the initial representation is defined as follows: Among them, and These are the weights and biases of the linear layer with edge features. Indicates continuous edge characteristics; Indicates the embedding layer. Represents discrete edge features; symbol This represents a splicing operation; It is the dimension of the embedded vector. The number of edges; And / or, Node embedding is represented as: ,in, It is the normalized adjacency matrix of the node graph. It is the weight matrix of the l-th layer GCN corresponding to the node graph; Edge embedding is represented as: ,in, It is the normalized adjacency matrix of the edge graph. It is the weight matrix of the l-th layer GCN corresponding to the edge feature map; And / or, By sharing the VAE architecture, node features and edge features are mapped to the same latent space, generating the distribution parameters of the latent variable z: , ;in, This refers to a multilayer perceptron for the encoder in a VAE. The latent variable z is sampled using the following formula: ;in, This represents a multivariate standard normal distribution with a mean of 0 and a covariance equal to the identity matrix. This indicates element-wise multiplication; Multilayer perceptron using decoder The reconstructed node features are represented as follows: ; Get each time slice The feature vector matrices of the hotspots and transportation routes in the spatial dimension ; And / or, The hotspot sequence corresponding to the trajectory is mapped to an embedded representation of a node sequence and an edge sequence. The embedded representation of the node sequence is as follows: The embedding representation of the edge sequence is as follows: ;in, The query request feature vector; And / or, By integrating mutual information between hotspots and / or transportation paths in the waybill trajectory using a multi-head attention mechanism, the output is: , in, To output the projection matrix; And / or, The dual cross-attention mechanism is used to integrate and update the node sequence and edge sequence information. After concatenation, the complete journey is obtained, represented as: ;in, An embedding vector representing a sequence of nodes. The embedding vector representing the edge sequence. Indicates the length of the node sequence.

6. The method according to claim 1, characterized in that, In step S3, given a query request and the corresponding historical transportation route, the correlation between the historical transportation route and the query request is learned to generate an embedded representation of the transportation route. The embedded representation is modeled using a denoising diffusion probability model, which includes a forward diffusion process and a reverse denoising process; The forward diffusion process gradually introduces noise into the signal, eventually transforming the signal into a priori distribution, while the reverse denoising process gradually transforms the noisy vectorized line embedding representation into a noise-free representation. And / or, The denoiser adopts the Unet architecture and consists of multiple downsampling blocks and upsampling blocks. Each block contains a residual block and a Transformer. And / or, The query request features and time steps are concatenated using location encoding and input into the residual network to guide the denoising process.

7. The method according to claim 1, characterized in that, In step S4, the decoding process is carried out iteratively. At each decoding step, the neighboring node with the highest probability is selected as the next access node, and the embedded representation of the selected node and the edge connecting the node are concatenated and input into a single-layer LSTM to record the current decoding state and guide the selection of the next step. And / or, The candidate transfer node set is filtered using a masking mechanism. The masking rules include: 1) Block already visited nodes; 2) Nodes with a transition probability below a threshold are blocked; 3) Nodes with a historical trajectory transfer probability of zero are blocked; Nodes that satisfy any one or more of the above mask rules will not be selected as candidate nodes; And / or, The probability of the current selectable node is normalized using softmax, and the process is repeated until the end.

8. The method according to claim 1, characterized in that, In step S5, model training is performed in two stages; The first phase of training involves encoding the trajectory map using GCN-VAE and reconstructing the dwell hotspot sequence using a decoder; The loss function of the GCN-VAE includes the reconstruction error, which measures the difference between the original features and the reconstructed features, and the KL divergence, which measures the difference between the distribution of the latent variables and the standard normal distribution, as follows: ， ， The loss for the multi-class classification problem in route prediction is calculated using the cross-entropy function, as shown below: ， in, It is the number of waybill trajectories in the training dataset. Indicates the first Number of hotspots for each waybill trajectory sample; The total loss function for the first stage is expressed as: ， in, These are the weighting coefficients for each loss term, used to balance the contributions of each part to the total loss; In the second stage, the embedding representation of the waybill trajectory generated in the waybill trajectory encoding step is used to train the denoiser used in the transportation route embedding representation generation step, resulting in the model. Learn the spatial distribution of embedded transportation routes; Based on the mean squared error, the noise matching loss function is defined as follows: ， in, This represents the true noise sampled from a standard normal distribution. The noise predictor generates an estimate under current conditions, with input consisting of the diffused embedding representation and the time step. And the query request consists of; And / or, Introducing structural alignment loss to constrain the final generated path representation Embedded representation with the output of the trajectory coding module The similarity between them is expressed as follows: ， in, It is the final representation of the waybill trajectory sequence by the trajectory encoder. The noise-free path embedding reconstructed from the reverse process is calculated as follows: ， The total loss is expressed as: .

9. The method according to claim 1, characterized in that, In step S6, the starting hotspot is matched according to the query request and the departure time is updated. The time slice sequence after the departure time is constructed and a snapshot sequence of the hotspot transfer relationship graph is generated. The graph embedding and edge embedding of the hotspot, as well as the transportation route embedding representation corresponding to the query request, are generated and input into the decoder to complete the hotspot sequence inference.

Images