Traffic next position prediction method and device based on multi-adaptive hypergraph cooperation

Abstract

The invention belongs to the field of urban resident traffic trip next position prediction, and discloses a traffic trip next position prediction method and device based on multi-adaptive hypergraph collaboration, and the method comprises the steps: constructing an interactive hypergraph and a directed hypergraph based on an individual traffic trip trajectory sequence, constructing a geographic hypergraph based on geographic feature data of a position, and constructing a traffic trip next position prediction result. The interaction hypergraph and the geographic hypergraph are updated to obtain a self-adaptive interaction hypergraph and a self-adaptive geographic hypergraph, and a position overall embedding matrix and an individual overall embedding matrix are obtained based on the hypergraphs; and outputting the first k positions with the highest probability as a prediction result of the next travel position of the individual through a normalization function. The method has high accuracy and strong robustness for position prediction.

Description

Technical Field

[0001] This invention belongs to the field of next location prediction for urban residents' travel, and specifically relates to a method and apparatus for next location prediction for travel based on multi-adaptive hypergraph collaboration. Background Technology

[0002] Early research on next-location prediction in transportation primarily focused on sequence modeling, with typical methods including Markov chains, recurrent neural networks, and attention mechanisms. In recent years, Graph Neural Networks (GNNs) have demonstrated outstanding performance in modeling complex human travel patterns. They can efficiently model complex graph data by learning high-quality node representations, effectively improving the accuracy of next-location prediction. Existing research has explored various complex graph structures to model the complex dependencies between individual movement and location, including hypergraphs, knowledge graphs, directed graphs, and multi-relation graphs. Simultaneously, several new GNNs have been proposed to extract information from various complex graph data. Despite significant progress in graph-based next-location prediction research, some problems remain. Taking hypergraph structures as an example, the existing problems can be summarized into the following three aspects.

[0003] First, existing research largely relies on static individual travel trajectories and location geographic features to construct hypergraphs, resulting in a rigid graph structure that limits the model's ability to characterize complex individual travel behaviors and location geographic features. This makes it difficult to capture dynamic individual movement patterns and perceived distances between locations. Specifically, perceived distance refers to the effective distance actually perceived by an individual during travel; it depends not only on physical distance in geographic space but is also influenced by factors such as mode of transportation and route selection. Therefore, the models have limitations in terms of the diversity of location predictions and travel convenience. For example... Figure 1 As shown, after visiting the restaurant, the individual's next destination is a store; although store 2 is closer to the residence, the static historical travel trajectory (solid line) may guide the model to predict that the individual's next location is store 1. Similarly, although store 3 is closer in perceived distance (5 minutes by bus is less than 20 minutes on foot), the model may predict store 2, which is closer to the residence, because it only considers the geographical distance between locations.

[0004] Secondly, noise and sparsity in individual historical location check-in data can easily lead to unreliable hyperedges and nodes in the hypergraph, thereby perturbing the hypergraph structure and affecting the accuracy of the model in learning individual travel patterns. For example... Figure 1 As shown, the cinema is a noisy location in a static trajectory, which does not align with individual travel preferences. Nevertheless, the model may still predict the cinema based on historical trajectories, thereby influencing subsequent store predictions and increasing the probability of store 1 (rather than store 2 or store 3) being predicted.

[0005] Finally, individual travel behavior is influenced by a variety of factors, including individual habits, geographical features, and the temporal sequence of visits. However, most graph-based methods use only a single graph to model individual travel patterns, or use multiple graphs for collaborative modeling, without considering the sequential modeling of travel trajectories. For example, a single hypergraph constructed solely from individual travel trajectories, without sequential modeling of these trajectories, can only capture individual travel habits and cannot adequately account for the influence of other factors. Therefore, models that rely on a single graph or neglect sequential modeling struggle to comprehensively understand individual travel patterns. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for predicting the next location in transportation based on multi-adaptive hypergraph collaboration, which has high accuracy and strong robustness in location prediction.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] Firstly, a method for predicting the next location in transportation trips based on multi-adaptive hypergraph collaboration is provided, including:

[0009] Interactive and directed hypergraphs are constructed based on individual travel trajectory sequences, and geographic hypergraphs are constructed based on location-specific geographic feature data.

[0010] The interaction hypergraph and geographic hypergraph are updated based on the feature similarity between each location, and a graph structure enhancement strategy is introduced during the update process. Finally, the updated adaptive interaction hypergraph and adaptive geographic hypergraph are output.

[0011] A trajectory sequence encoder is used to model individual traffic travel trajectory sequences to form an individual sequence embedding matrix;

[0012] Interactive hypergraphs and adaptive interactive hypergraphs are processed by interactive hypergraph convolutional blocks to obtain positional interactive embedding matrices and adaptive positional interactive embedding matrices, and individual interactive embedding matrices are extracted and generated based on the positional interactive embedding matrices.

[0013] Directed hypergraphs are processed using directed hypergraph convolutional blocks to obtain position-directed embedding matrices, and individual directed embedding matrices are extracted and generated based on these position-directed embedding matrices.

[0014] Geographic hypergraphs and adaptive geographic hypergraphs are processed using geographic hypergraph convolutional blocks to obtain location geographic embedding matrices and adaptive location geographic embedding matrices, and individual geographic embedding matrices are generated based on the location geographic embedding matrices.

[0015] The location interaction embedding matrix, the location directed embedding matrix, and the location geographic embedding matrix are aggregated to obtain the overall location embedding matrix. The individual sequence embedding matrix, the individual interaction embedding matrix, the individual directed embedding matrix, and the individual geographic embedding matrix are aggregated to obtain the overall individual embedding matrix.

[0016] Based on the overall location embedding matrix and the overall individual embedding matrix, the top k positions with the highest probabilities are output as the prediction results of the individual's next travel location through a normalization function.

[0017] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0018] Preferably, the set of hyperedges in the interactive hypergraph consists of individuals, and the set of nodes in the interactive hypergraph consists of locations. Individuals are used as hyperedges to connect all locations visited by the individuals, and connection weights are assigned between individuals and locations in the interactive hypergraph based on the number of times the individuals visit the locations.

[0019] The set of hyperedges in the directed hypergraph consists of source positions, and the set of nodes in the directed hypergraph consists of target positions. The source position is used as a hyperedge to connect all target positions corresponding to the source position, and a connection weight is assigned between the source position and the target position in the directed hypergraph based on the position index interval.

[0020] The set of hyperedges and the set of nodes in the geographic hypermap are both composed of locations. Each location is the center, and adjacent locations are identified within a predefined geographic distance threshold. The center location is used as the hyperedge to connect all adjacent locations.

[0021] Preferably, the update process of the interactive hypergraph is as follows:

[0022] Construct an interaction similarity matrix based on the location interaction embedding matrix;

[0023] For each individual, extract the individuals with the highest similarity based on the interaction similarity matrix that do not belong to the current individual. The extracted location Each location is added to the current individual's individual travel trajectory sequence to obtain the travel trajectory extended sequence;

[0024] Based on the locations involved in the extended travel trajectory sequence, an interaction similarity submatrix is ​​extracted from the interaction similarity matrix, and the submatrix with the lowest similarity is identified. Each location was analyzed, and the least similar sequence in the extended travel trajectory was removed. From each location, a corrected travel trajectory sequence is obtained;

[0025] A graph structure enhancement strategy is used to randomly perturb the travel trajectory correction sequence, and the interaction hypergraph is reconstructed based on the randomly perturbated travel trajectory correction sequence, which is denoted as the adaptive interaction hypergraph.

[0026] Preferably, the update process of the geographic hypermap is as follows:

[0027] Construct a geographic similarity matrix based on the location geographic embedding matrix;

[0028] The adjustment factor is obtained by subtracting the geographical similarity matrix from a matrix in which all elements are 1.

[0029] The adjustment factor is multiplied element by element by the geographic distance matrix to obtain the perceptual distance matrix, which is composed of the geographic distances between each pair of locations.

[0030] The perception distance matrix is ​​randomly perturbed using a graph structure enhancement strategy.

[0031] Based on the perceptual distance matrix after random perturbation, neighboring locations are identified within a predefined perceptual distance threshold, with each location as the center, to obtain the location neighbor set;

[0032] A graph structure enhancement strategy is used to randomly perturb the location neighbor set, and the center position is used as a hyperedge to connect all adjacent positions in the randomly perturbated location neighbor set.

[0033] Preferably, the trajectory sequence encoder includes a time sub-encoder and a sequence sub-encoder. The step of using the trajectory sequence encoder to model individual traffic travel trajectory sequences to form an individual sequence embedding matrix includes:

[0034] The time sub-encoder uses the time2vector method to encode the time index in each individual's traffic trajectory sequence to obtain a time embedding vector. The time embedding vector is then concatenated with the learnable location embedding vector to obtain the trajectory sequence embedding matrix.

[0035] The sequence sub-encoder uses a Transformer encoder to process each trajectory sequence embedding matrix. The output of the Transformer encoder is then averaged to obtain the embeddings of each volume sequence.

[0036] The individual sequence embeddings of all individuals are aggregated to form an individual sequence embedding matrix.

[0037] Preferably, the step of extracting and generating an individual interaction embedding matrix based on the location interaction embedding matrix includes: taking all location interaction embeddings belonging to the same individual in the location interaction embedding matrix, and performing average aggregation on all location interaction embeddings of the same individual to obtain individual interaction embeddings, and aggregating the individual interaction embeddings of all individuals to form an individual interaction embedding matrix.

[0038] The step of extracting and generating an individual directed embedding matrix based on the position directed embedding matrix includes: taking all position directed embeddings belonging to the same individual in the position directed embedding matrix, and performing average aggregation on all position directed embeddings of the same individual to obtain individual directed embeddings, and aggregating the individual directed embeddings of all individuals to form an individual directed embedding matrix.

[0039] The step of extracting and generating an individual geographic embedding matrix based on the location geographic embedding matrix includes: taking all location geographic embeddings belonging to the same individual in the location geographic embedding matrix, and performing average aggregation on all location geographic embeddings of the same individual to obtain individual geographic embeddings, and aggregating the individual geographic embeddings of all individuals to form an individual geographic embedding matrix.

[0040] Preferably, during the training and updating process of the trajectory sequence encoder, interactive hypergraph convolutional block, directed hypergraph convolutional block, and geographic hypergraph convolutional block, the total loss function includes total contrastive learning loss, cross-entropy loss, and KL divergence loss.

[0041] Preferably, the total contrastive learning loss includes: the contrastive learning loss between the location interaction embedding matrix and the adaptive location interaction embedding matrix, the geographic contrastive learning loss between the location geographic embedding matrix and the adaptive location geographic embedding matrix, the contrastive-directed contrastive learning loss between the individual interaction embedding matrix and the individual directed embedding matrix, the contrastive-geographical contrastive learning loss between the individual interaction embedding matrix and the individual geographic embedding matrix, and the contrastive-geographical contrastive learning loss between the individual directed embedding matrix and the individual geographic embedding matrix.

[0042] Preferably, the KL divergence loss includes:

[0043] Construct an interaction similarity matrix based on the location interaction embedding matrix;

[0044] Construct a directed similarity matrix based on the position-directed embedding matrix;

[0045] Construct a geographic similarity matrix based on the location geographic embedding matrix;

[0046] The KL divergence between the interaction similarity matrix and the directed similarity matrix, as well as the KL divergence between the geographic similarity matrix and the directed similarity matrix, are calculated to form the KL divergence loss.

[0047] The second aspect: provides a traffic next location prediction device based on multi-adaptive hypergraph collaboration, including a processor and a memory storing a number of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the traffic next location prediction method based on multi-adaptive hypergraph collaboration.

[0048] The present invention proposes a method and apparatus for predicting next location of travel based on multi-adaptive hypergraph collaboration (MAHCN), which has the following advantages compared with the prior art:

[0049] (1) An Adaptive Hypergraphs Update Module (AHUM) is proposed to dynamically update the graph structures of the interaction hypergraph and the geographic hypergraph, thereby enhancing the representation of the complex dependencies between individual travel preferences and location. In this process, the graph structure solidification problem caused by static individual travel trajectories and location geographic features is gradually alleviated, enabling the model to output more diverse and convenient locations.

[0050] (2) The Perturbation-Driven Augmentation Strategy (PDAS) is integrated into AHUM. This strategy uses a hyperparameter to control the degree of perturbation, and implements the random removal and addition of hyperedges and nodes to simulate perturbed hypergraph structures, thereby enhancing the model's ability to model imperfect graph structures.

[0051] (3) A Multiple Hypergraphs Collaboration Module (MHCM) is proposed to model the impact of multidimensional factors on individual travel behavior. This module not only considers the temporal patterns in trajectory sequences, but also uses contrastive learning to achieve collaborative modeling of multiple hypergraphs, thereby providing a more comprehensive understanding of individual travel behavior patterns. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the interference of fixed graph structures and data noise on store forecasts.

[0053] Figure 2 This is a flowchart of a traffic travel next location prediction method based on multi-adaptive hypergraph collaboration according to the present invention;

[0054] Figure 3 This is a flowchart of the adaptive interactive hypergraph update submodule of the present invention;

[0055] Figure 4 The figure shows the experimental results of the hypergraph construction method of the present invention.

[0056] Figure 5 This is a visualization of the t-SNE results of the location embedding in the NYC dataset of this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0059] Example 1:

[0060] Accurate prediction of human travel patterns is crucial for many downstream applications, such as traffic management, pandemic control, and personalized service recommendations. The core task is to predict an individual's next location based on historical travel data. Existing research widely employs graph neural networks and their derived complex graph structures to characterize individual travel behavior patterns. However, these methods still face numerous challenges, such as graph structure rigidity, noise and sparsity in check-in data, and insufficient consideration of multidimensional factors influencing individual travel preferences, leading to limited predictive model performance. To address these issues, this invention proposes a method for predicting the next location of travel based on multi-adaptive hypergraph collaboration. First, a multi-type hypergraph based on an adaptive hypergraph update mechanism is constructed. By dynamically optimizing the structures of the interaction hypergraph and the geographic hypergraph, the impact of graph structure rigidity is mitigated, thereby enhancing its ability to represent the complex dependencies between individual travel preferences and location. Second, a perturbation-based graph structure enhancement strategy is introduced to learn individual travel preferences within perturbed hypergraph structures, improving the model's ability to model imperfect graph structures. Furthermore, contrastive learning is used to strengthen the collaborative relationships between multiple hypergraphs, thereby enhancing the model's multi-factor modeling capabilities.

[0061] like Figure 2The present embodiment of the traffic travel next location prediction method based on multi-adaptive hypergraph collaboration includes the following three key modules: 1) Initial Hypergraph Construction Module (IHCM): This module constructs three initial hypergraphs (interaction hypergraph, directed hypergraph, and geographic hypergraph) based on individual traffic travel trajectory and location geographic feature data to capture complex individual travel behavior patterns. 2) Adaptive Hypergraph Update Module (AHUM): This module dynamically adjusts the structure of the interaction hypergraph and geographic hypergraph, and introduces the PDAS strategy to simulate the perturbed hypergraph structure, thereby improving the prediction model's ability to model imperfect graph structures. 3) Multi-Hypergraph Collaboration Module (MHCM): This module extracts and integrates information from trajectory sequences and multiple hypergraphs, and uses a collaborative contrastive learning mechanism to enhance the model's multi-factor modeling ability. Specifically, the method of this invention includes the following steps:

[0062] Step 1: Construct an interactive hypergraph and a directed hypergraph based on individual travel trajectory sequences, and construct a geographic hypergraph based on location geographic feature data.

[0063] Individual travel trajectory sequences are spatiotemporal sequence data, mainly composed of an ordered combination of timestamps and geographic coordinates. Attributes such as travel mode and stops can also be added, representing the locations an individual has traversed from a preset time period to the current moment. Geographic feature data describes the attributes, environment, and related information of a geographic location, and can be used to integrate and generate a set of locations.

[0064] Pick and Let them represent the set of individuals and the set of locations, respectively. The total number of individuals, This represents the total number of positions. Each position... Corresponding to a spatial location in the real world, using latitude and longitude coordinates express, Represents latitude and longitude coordinates. Represents latitude and longitude coordinates. Individual. Individual travel trajectories in sequence It means that among them Represents an individual exist Location accessed at any time , It is the length of an individual's travel trajectory sequence.

[0065] The next position prediction task aims to predict individual The k locations for the next trip can be represented by a nonlinear prediction function. As shown in formula (1).

[0066] (1)

[0067] in, It is a probability vector that records an individual The probability of visiting each location next; It is a function All parameters are included. Based on the predicted access probability, this embodiment sorts all positions in descending order and selects the top k positions to form a prediction result list.

[0068] Individual travel behavior patterns are highly complex and difficult to fully represent using a single graph structure. Therefore, this embodiment constructs an interaction hypergraph, a directed hypergraph, and a geographic hypergraph to capture the interaction relationships between individuals and locations, the directed transfer relationships between locations, and the geographic feature associations between locations, respectively.

[0069] (1) Interactive Hypergraph: Interactive Hypergraph It is constructed based on individual travel trajectory sequences, where the super-edge set Composed of individuals, a set of nodes Composed of locations, each individual is used as a hyperedge to connect all locations visited by that individual, and connection weights are assigned between individuals and locations in the interaction hypergraph based on the number of times each individual visits a location.

[0070] like Figure 2 As shown in the IHCM module, individual Location visited , and Therefore, in In the middle, super-border Connected nodes , and Furthermore, since individuals have different preferences for different locations, the degree of preference for a particular location increases with the number of times it is visited. Therefore, in the process of constructing the interactive hypergraph, this embodiment assigns different connection weights to individuals and locations, as shown in formula (2).

[0071] (2)

[0072] in, express medium-sized individuals and location Connection weights; Indicates individual travel trajectory Middle position The number of times it was accessed; Indicates individual travel trajectory Middle position The number of times it was accessed.

[0073] (2) Directed Hypergraph: Directed hypergraph based on individual travel trajectory sequences Construction. Specifically, two locations visited sequentially in an individual's travel trajectory sequence are represented as directed tuples (source location, target location). Used to characterize the directed transfer relationship between each source location and its corresponding target location. In the superedge set Composed of source locations, a set of nodes It consists of target locations, with the source location as the hyperedge connecting all target locations corresponding to the source location, and assigns connection weights between the source location and the target location in the directed hypergraph based on the location index interval.

[0074] like Figure 2 As shown in the IHCM module, individual The locations visited were... , and And individuals The locations visited were... , and Therefore, the source location Corresponding to three directed tuples: , and Therefore, in In the middle, super-border Connected nodes , and .

[0075] Considering that there is usually a stronger correlation between positions with close index intervals in the trajectory, this embodiment proposes a connection weight calculation method based on the position index interval to quantify the tightness of the directed transfer relationship, as shown in formula (3).

[0076] (3)

[0077] in, Indicates source location and target location exist Connection weights in the data; and They represent and The index value in an individual's travel trajectory sequence. If multiple identical directed tuples exist, then... The weight values ​​of the corresponding connections are accumulated.

[0078] (3) Geographic Hypermap: Geographic Hypermap Location-based geographic feature data construction, where hyperedge sets and node set All locations are determined by their positions. First, the geographic distances between all locations are calculated using the Haversine formula based on latitude and longitude, forming a geographic distance matrix. Then, centered on each location, within a predefined geographic distance threshold... The inner part identifies its adjacent positions, and uses the center position as a hyperedge to connect all adjacent positions. For example... Figure 2 As shown in the IHCM module, in In the equation (4), the center position is used as a hyperedge, connecting all its adjacent positions.

[0079] (4)

[0080] in, Indicates position and Geographical distance between them; yes The set of neighbors.

[0081] Step 2: Update the interaction hypergraph and geographic hypergraph based on the similarity between locations, and introduce graph structure enhancement strategies during the update process to finally output the updated adaptive interaction hypergraph and adaptive geographic hypergraph.

[0082] The initial interaction hypergraph and geographic hypergraph are constructed based on static individual travel trajectories and location geographic features, resulting in a rigid graph structure that struggles to capture dynamic individual travel preferences and perceived distances between locations. To address this, the AHUM module dynamically adjusts the structure of the two hypergraphs using an adaptive interaction hypergraph update submodule (AHUM-I) and an adaptive geographic hypergraph update submodule (AHUM-G), thereby enhancing the representation of the complex dependencies between individual travel preferences and locations. This allows the model to predict more diverse and convenient locations. Simultaneously, the PDAS strategy is integrated with both submodules to promote hypergraph structure optimization, resulting in an enhanced hypergraph. Figure 3 The specific process of the AHUM-I module integrating the PDAS strategy is demonstrated.

[0083] (1) Adaptive interactive hypergraph update submodule:

[0084] Existing research typically constructs a graph adjacency matrix using the similarity of embedding vectors between locations. In contrast, the AHUM-I module proposed in this embodiment directly adjusts the structure of the interactive hypergraph based on a similarity matrix formed from the interactive embeddings of locations, such as... Figure 3As shown. Specifically, the AHUM-I module first generates a location interaction embedding matrix using the interaction hypergraph convolutional blocks from the MHCM module, and then constructs an interaction similarity matrix based on the location interaction embedding matrix. The AHUM-I module uses a weighted cosine similarity function to calculate the similarity between the embedding vectors at each position, as shown in formula (5).

[0085] (5)

[0086] in, Represents the positional interaction embedding matrix; This represents a learnable weight matrix; Represents the Hadamard product. This indicates the transpose operation. Describes the Euclidean norm. This indicates that the transpose operation is performed first, and then the Euclidean norm is calculated.

[0087] Next, this module according to The values ​​are sorted. To reveal an individual's potential travel preferences, this module, based on the sorting results, extracts the most similar individuals that do not belong to the current individual from the interaction similarity matrix for each individual. The extracted location Each location is added to the current individual's travel trajectory sequence to obtain the travel trajectory extension sequence, thus achieving trajectory extension. Value and length of travel trajectory sequence Relevant, and proportional Reduce. Furthermore, since the structure of the interaction hypergraph is independent of the access time of each location, this process does not consider... The access times and their order of locations within the travel trajectory. The generation process is shown in formula (6).

[0088] (6)

[0089] in, yes The element in the array represents the position. and Interaction similarity between them; This is the lower bound of the average interaction similarity. Among them, the ones with the highest similarity but not belonging to the current individual... The location is determined by extracting individual travel trajectory sequences from the interaction similarity matrix. Each position With a sequence of travel routes that do not belong to an individual's travel trajectory Location The interaction similarity is calculated, and the average of all extracted interaction similarities is taken. As a position Extended similarity is used to calculate all sequences of traffic trajectories that do not belong to individual travel trajectories. The expanded similarity of the positions is taken, and the expanded similarity is greater than 1. And the largest One position.

[0090] Subsequently, the module from Extract each Corresponding interaction similarity submatrix This submatrix only describes middle The interaction similarity between each location. Based on the similarity submatrix, determine the location with the lowest similarity in the interaction similarity submatrix. Each location, and remove the extended sequence of travel trajectories. The one with the lowest similarity Based on the given location, a travel trajectory correction sequence is obtained, enabling the correction of individual travel trajectories and maintaining their length within a specified range. Obtain the travel trajectory correction sequence. The generation process is shown in formula (7).

[0091] (7)

[0092] in, This is the upper bound of the maximum interaction similarity. Among them, the one with the lowest similarity in the interaction similarity submatrix is ​​determined. The method for determining the location is as follows: for the extended sequence of travel trajectories Each position in Take the position in the interaction similarity submatrix. Related interaction similarity The maximum value in the sequence is used as the similarity to remove the line trajectory extension sequence. Remove those with similarity less than And the lowest One position.

[0093] Finally, based on the obtained travel trajectory correction sequence Refactoring To achieve interactive hypergraph The update is denoted as the adaptive interactive hypergraph. .

[0094] (2) Adaptive hypermap update submodule:

[0095] First, this module constructs a geographic similarity matrix based on the location geographic embedding matrix. Its construction method and Similarly. Next, the module utilizes... Calculate the perceived distance between locations instead of the original geographic distance calculated by the Haversine formula.

[0096] Specifically, this module first utilizes a matrix where all elements are 1. minus This yields the inverse of the geographical similarity, with the corresponding numerical range being... Convert to This is to facilitate subsequent sensing distance calculation. Then, the module will adjust the factor. Geographic distance matrix Element-wise multiplication yields the perceptual distance matrix. Next, the module reconstructs the location based on the sensing distance matrix. Neighbor set To achieve a hypergraph Adaptive updates are performed. During this process, the adjustment factor is controlled. The impact on sensing distance, the sensing distance threshold is set to The above process is shown in formulas (8) and (9).

[0097] (8)

[0098] (9)

[0099] in, yes The element in the array represents the position. and The perceived distance between them.

[0100] (3) Perturbation-based graph structure enhancement strategy:

[0101] In real-world check-in data, noisy locations and sparse trajectories are common. To enhance the robustness of the prediction model, the PDAS strategy simulates the perturbed hypergraph structure by adding or removing hyperedges and nodes during AHUM-I and AHUM-G processes, thus constructing an enhanced adaptive hypergraph.

[0102] Specifically, in the AHUM-I module, the PDAS policy uses perturbation... Implement the addition and removal of hyperedges and nodes to enhance adaptive hypergraphs. The process is as follows Figure 3 As shown, the PDAS strategy is used to randomly perturb the travel trajectory correction sequence, and the interactive hypergraph is reconstructed based on the randomly perturbed travel trajectory correction sequence. In the AHUM-G module, the PDAS strategy is implemented by perturbing... and To enhance the adaptive hypergraph, dynamic adjustments to hyperedges and nodes are implemented. The process involves: First, randomly perturbing the sensing distance matrix using the PDAS strategy. Then, based on this perturbed matrix, identifying neighboring locations within a predefined sensing distance threshold, centered at each location, to obtain a neighborhood set. Finally, randomly perturbing the neighborhood set using the graph PDAS strategy, using the center location as a hyperedge to connect all neighboring locations within the perturbed neighborhood set. In the MHCM module, two enhanced adaptive hypergraphs are used. and As hypergraphs and By comparing the views, and then using comparative learning, the model's ability to model imperfect graph structures can be enhanced.

[0103] Individual travel behavior patterns are influenced by a variety of factors, including individual habits, geographical features, and the order of visits. In the MHCM module, a trajectory sequence encoder and a hypergraph convolutional network were designed to extract information from trajectory sequences and multiple hypergraphs, respectively. A contrastive learning framework was also constructed to enhance the model's multi-factor modeling capabilities, thereby improving the MHCM module's ability to capture the diversity and dynamic characteristics of individual travel preferences.

[0104] Step 3: Use a trajectory sequence encoder to model individual traffic travel trajectory sequences to form an individual sequence embedding matrix.

[0105] While multiple hypergraphs can effectively capture individual travel behavior patterns and geographical relationships, they fail to capture temporal patterns within trajectory sequences. To address this deficiency, this embodiment further designs a trajectory sequence encoder for modeling individual travel trajectory sequences. The time pattern is described. The encoder consists of a time sub-encoder and a sequence sub-encoder.

[0106] The time sub-encoder uses the time2vector method to... The time index is encoded. Specifically, the time... Transform into a dimension Embedded vector Embedded vectors The Middle The calculation method for each element is shown in formula (10).

[0107] (10)

[0108] in, This refers to the time2vector function; Represents the sinusoidal activation function; and These are learnable parameters. Encoded temporal embeddings. With learnable position embedding matrix Embedded in By piecing them together one by one, a trajectory sequence embedding matrix is ​​finally formed. .

[0109] The sequence sub-encoder is implemented based on the Transformer framework to model... The time patterns in the formulas are shown in formulas (11), (12) and (13).

[0110] (11)

[0111] (12)

[0112] (13)

[0113] in, This represents the Transformer encoder, which includes a multi-head self-attention mechanism and a point-by-point feedforward network; and This represents the input and output embedding matrices of the Transformer encoder; express The Each input feature embedding; express The A sequence embedding; express The An index code is used to inject index order information from individual traffic trajectory sequences into the Transformer encoder; express The Each output feature embedding; This indicates element-wise addition; express Individuals obtained after average pooling Sequence embedding.

[0114] Finally, the sequence embeddings of all individuals are aggregated to form an individual sequence embedding matrix. .

[0115] Step 4: Process the interactive hypergraph and the adaptive interactive hypergraph using an interactive hypergraph convolutional network to obtain the location interactive embedding matrix and the adaptive location interactive embedding matrix, and extract the individual interactive embedding matrix based on the location interactive embedding matrix; process the directed hypergraph using a directed hypergraph convolutional network to obtain the location directed embedding matrix, and extract the individual directed embedding matrix based on the location directed embedding matrix; process the geographic hypergraph and the adaptive geographic hypergraph using a geographic hypergraph convolutional network to obtain the location geographic embedding matrix and the adaptive location geographic embedding matrix, and extract the individual geographic embedding matrix based on the location geographic embedding matrix.

[0116] To extract higher-order information and obtain high-quality node representations from multiple constructed hypergraphs, this embodiment employs an association matrix. Indicates containing Each node and A hypergraph with a superedge This embodiment has , , They are respectively , and The corresponding correlation matrix is ​​used to construct a hypergraph convolutional network. This network consists of interaction hypergraph convolutional blocks, directed hypergraph convolutional blocks, and geographic hypergraph convolutional blocks, with each convolutional block stacked. A hypergraph propagation layer. The definition of a hypergraph propagation layer is shown in formula (14).

[0117] (14)

[0118] in, and They represent the first The node feature matrices of the layer input and output; Represents a non-linear activation function; A diagonal matrix representing the degree of nodes; A diagonal matrix representing the weights of the hyperedges; A diagonal matrix representing the cardinality of hyperedges; Indicates the first The learnable convolutional parameter matrix of the layer; Indicates the first The residual connectivity parameter matrix of the layer; It is the first The dimension of the feature vectors of the output nodes of the layer. For adaptive hypergraphs... and Its correlation matrix has elements that are either 0 or 1, while for adaptive hypergraphs... and as well as The elements of the correlation matrix are the connection weights of nodes and hyperedges, and their specific calculation is shown in formulas (2) and (3).

[0119] Interactive Hypergraph Convolutional Blocks for Adaptive Hypergraph and The operations are shown in formulas (15), (16) and (17).

[0120] (15)

[0121] (16)

[0122] (17)

[0123] in, Represents the stacked elements in the hypergraph convolutional blocks. One hypergraph propagation layer; and These represent adaptive hypergraphs. and The corresponding position embedding matrix serves as the initial node feature matrix. and Input to the first hypergraph propagation layer; and This represents the node feature matrix output by the last hypergraph propagation layer, which is used to update the location embedding matrix. and .

[0124] Directed Hypergraph Convolutional Block Processing Adaptive Hypergraph The method, and the adaptive hypergraph processing of geographic hypergraph convolutional blocks. , The methods are all related to the adaptive hypergraph processing of convolutional blocks in interactive hypergraphs. , The method is similar.

[0125] Updated position embedding matrix , and They are used to construct the interaction similarity matrix respectively. Directed similarity matrix Geographic similarity matrix .in, and The calculation method is similar to As shown in formula (5). And according to the position embedding matrix , and Generate individual interaction embedding matrix Individual Directed Embedding Matrix and individual geographic embedding matrix The generation process is similar: take all position embeddings belonging to the same individual in the position embedding matrix, and perform average aggregation (add and average) on all position embeddings of the same individual to obtain individual embeddings, and aggregate all individual embeddings to form an individual embedding matrix.

[0126] Step 5: Aggregate the location interaction embedding matrix, the location directed embedding matrix, and the location geographic embedding matrix to obtain the overall location embedding matrix. The overall individual embedding matrix is ​​obtained by aggregating the individual sequence embedding matrix, the individual interaction embedding matrix, the individual directed embedding matrix, and the individual geographic embedding matrix. .

[0127] Step 6: Based on the overall location embedding matrix and the overall individual embedding matrix, output the top k positions with the highest probabilities using a normalization function as the prediction results for the individual's next travel location.

[0128] This embodiment utilizes multi-hypergraph collaboration based on contrastive learning to train and update trajectory sequence encoders, interactive hypergraph convolutional networks, directed hypergraph convolutional networks, and geographic hypergraph convolutional networks. The multi-hypergraph collaborative contrastive learning framework simultaneously includes comparisons between a hypergraph and its augmenting hypergraph, as well as comparisons between different hypergraphs.

[0129] The comparison with augmented hypergraphs is based on the position embedding matrix, which aims to learn more robust hypergraph node representations. Adaptive hypergraphs... and Comparative learning loss The definition is shown in formula (18).

[0130] (18)

[0131] in, Represents an adaptive hypergraph Positional interaction embedding matrix; and These represent the positional interaction embedding matrices respectively. and adaptive positional interaction embedding matrix Middle position The embeddings form positive sample pairs; express Middle position Embedded, and Form negative sample pairs; Represents the cosine similarity function; It's a temperature hyperparameter. Adaptive hypergraph. and The contrastive learning loss between them is defined as Based on location geographic embedding matrix and adaptive location geographic embedding matrix Calculation, its calculation method and similar.

[0132] SuperMap ( , and The comparison and optimization between these components improved the model's multi-factor modeling capability. This process is based on the individual embedding matrix. , and These matrices are generated from the embeddings of corresponding positions in individual travel trajectory sequences. The loss is learned pairwise through comparative calculation of the three hypergraphs. , and Adding the three together yields... As shown in formulas (19) and (20).

[0133] (19)

[0134] (20)

[0135] in, , and These represent adaptive hypergraphs. and , and , and Comparative learning loss between them; and They represent and The embedding of individual u in the middle, the two constitute a positive sample pair; express Embedding of individual v in the middle, and Forming negative sample pairs; and The calculation method and similar.

[0136] Therefore, the total contrastive learning loss is obtained. As shown in formula (21).

[0137] (twenty one)

[0138] Based on position embedding matrix , and and individual embedding matrix , , and In this embodiment, the average aggregation method is used to obtain the overall position embedding matrix. and individual-to-whole embedding matrix As shown in formulas (22) and (23).

[0139] (twenty two)

[0140] (twenty three)

[0141] Next, calculate the individual Next position prediction probability distribution As shown in formula (24).

[0142] (twenty four)

[0143] in, express medium-sized individuals The overall embedding. Predicting loss. The cross-entropy function is used to measure the predicted value. and real labels The difference between them is shown in formula (25).

[0144] (25)

[0145] Where, if individual u's next visit is the th If there are several positions, then It equals 1, otherwise it equals 0.

[0146] Furthermore, the update of the adaptive hypergraph depends on the embedding quality of the locations, which may lead to suboptimal solutions and training difficulties in the early stages. To alleviate this problem, this embodiment introduces KL divergence loss. For the similarity matrix in the AHUM module and Regularization is performed to optimize the update process of the hypergraph. Specifically, the KL divergence between the interaction similarity matrix and the directed similarity matrix, as well as the KL divergence between the geographic similarity matrix and the directed similarity matrix, are calculated to form the KL divergence loss, as shown in formula (26).

[0147] (26)

[0148] in, Calculate the KL divergence between two probability distributions; Indicates the first Distribution of locations with directed transition relationships; This represents the directed similarity matrix.

[0149] In summary, by combining predicted losses Total comparative learning loss and KL divergence loss Total loss The definition is shown in formula (27).

[0150] (27)

[0151] in, , and These are weight hyperparameters used to balance multiple losses; Represents all parameters Regularization.

[0152] To visually demonstrate the advantages of the method of this invention, the following experiments were conducted:

[0153] (1) Experimental setup:

[0154] (1-1) Dataset:

[0155] This experiment selected three publicly available real-world check-in datasets to evaluate the proposed MAHCN model: Foursquare-NYC, Foursquare-TKY, and Gowalla-CA. These datasets come from two location-based social media platforms: Foursquare and Gowalla. Compared to Foursquare, Gowalla data has a wider geographical coverage and a longer time span.

[0156] The three datasets underwent the following three preprocessing steps: 1) Sort the check-ins for each individual in chronological order; 2) Divide the individual check-ins into trajectories at 24-hour intervals, filtering out locations visited less than 5 times, trajectories with fewer than 3 check-ins, and inactive individuals with fewer than 3 trajectories; 3) Divide the individual travel trajectories in chronological order, using the first 80% of the check-in data as the training set, the middle 10% as the validation set, and the last 10% as the test set. The statistical information of the datasets after preprocessing is shown in Table 1.

[0157] Table 1 Statistical information of the dataset

[0158]

[0159] (1-2) Baseline method:

[0160] The performance of the MAHCN proposed in this invention is compared with the following baseline methods:

[0161] FPMC (Factorized Personalized Markov Chain): A hybrid approach that combines matrix factorization with Markov chains to decompose individual travel preferences and model sequential behavior.

[0162] PRME (Personalized Ranking with Metric Embedding): A personalized ranking method that uses metric embedding to predict the next position.

[0163] STGN (Spatial-Temporal Gated Network): An enhanced LSTM based on time-gated and distance-gated methods for capturing an individual's order preference.

[0164] LSTPM (Long-Short Term Preference Modeling with Contextual Non-Local Networks and Geographically Expanded RNNs): A framework that improves the performance of next location prediction by modeling an individual's long- and short-term travel preferences using contextual non-local networks and geographically expanded recurrent neural networks respectively.

[0165] STAN (Spatial-Temporal Attention Network): Employs a two-layer attention mechanism to capture point-to-point spatiotemporal interactions and aggregate related locations, thereby explicitly modeling the correlation between non-adjacent locations and non-continuous visits.

[0166] SGRec (Sequence and Graph-based Recommendation): A GNN-based approach that extracts collaborative information from individual travel trajectories to model multi-level sequence preference patterns.

[0167] GETNext (Graph-Enhanced Transformer for Next POI Recommendation): A graph-enhanced Transformer model that uses cooperative signals in a global trajectory flow graph to predict the future movement of individuals.

[0168] HCCF (Hypergraph-based Collaborative Filtering with Contrastive Learning): A self-supervised collaborative filtering method that effectively combines hypergraph structure with contrastive learning to capture both local and global collaborative relationships.

[0169] DisenPOI (Disentangled Dual-Graph Framework for POI Recommendation): A decoupled dual-graph framework that combines contrastive learning to learn interpretable representations from sequence graphs and geographic graphs, respectively.

[0170] AGRAN (Adaptive Graph Representation Learning for Dynamic Preference Modeling) learns dynamic similarity features between embedding vectors at different locations to adaptively adjust the graph structure, thereby accurately capturing dynamic individual travel preferences.

[0171] STHGCN (Spatial-Temporal Hypergraph Convolutional Network): A Transformer model based on hypergraphs that can capture high-order collaborative relationships between the travel trajectories of a target individual and the travel trajectories of other individuals to alleviate the cold start problem.

[0172] DCHL (Decoupled Contrastive Learning on Hypergraph for POI Recommendation) is a contrastive learning method based on hypergraph decoupling that effectively alleviates the problem of entanglement in individual representations and improves the ability to model complementary recommendation effects.

[0173] To enable a fair comparison with methods that do not utilize category information, this embodiment removes the modules involving category information from SGRec, GETNext, and STHGCN.

[0174] (1-3) Evaluation indicators:

[0175] This embodiment uses two evaluation metrics: Recall@k and NDCG@k (normalized depreciated cumulative gain) to assess the predictive performance of the proposed MAHCN compared to baseline methods. Recall@k measures whether the first k positions of the predicted results list contain the individual's true label; NDCG@k evaluates the ranking accuracy of the predicted results, with k taking values ​​of 5 and 10. Both metrics are positive indicators; higher values ​​indicate better model performance.

[0176] (1-4) Experimental setup:

[0177] The MAHCN proposed in this embodiment is implemented based on the PyTorch framework. All experiments were conducted on a server equipped with an NVIDIA Tesla A800 GPU and an Intel Xeon Platinum 8358 CPU @ 2.60GHz. The main hyperparameters and training configuration are as follows:

[0178] Dimensions of location and individual embedding The number of dimensions for time embedding is 128. The number is 64. This refers to the number of hypergraph propagation layers in a hypergraph convolutional block. The value is 3, indicating a temperature over-parameter. The value is 0.3. The Transformer encoder is configured with 2 layers and 4 attention heads. A predefined threshold is used. , and The values ​​are 2.5, 0.1, and 0.0, respectively. The model uses the Adam optimizer for gradient optimization, with the learning rate and learning rate decay rate set to 2.5, 0.1, and 0.0, respectively. and The batch size is 128, the maximum number of training rounds is 100, the number of early stop and patience rounds is 10, and the minimum number of rounds is 50.

[0179] (2) Comparative experiment:

[0180] Table 2 presents the comparative experimental results of MAHCN and the baseline method. The indices (Recall and NDCG) are indicated by their first letters.

[0181] Table 2 shows the results of the comparative experiments on the three datasets.

[0182]

[0183] Based on the experimental results in Table 2, the following conclusions can be drawn:

[0184] (1) Among all baseline methods, graph-based methods (SGRec, GETNext, HCCF, DisenPOI, AGRAN, STHGCN, DCHL) outperform sequence-based methods (FPMC, PRME, STGN, LSTPM, STAN). This is because graph-based methods can fully utilize the correlation between access locations of different individuals, thereby capturing the collaborative relationships between individuals from a global perspective.

[0185] (2) On all three datasets, the proposed MAHCN outperforms all baseline methods, demonstrating its outstanding overall prediction performance and good effectiveness and robustness.

[0186] (3) STAN with self-attention mechanism is significantly better than other sequence-based methods, indicating that self-attention mechanism can more effectively model individual traffic travel trajectory sequences compared with sequence models such as RNN.

[0187] (4) Compared with other graph-based methods, the hypergraph-based methods (HCCF, STHGCN, DCHL) have better overall prediction performance, which indicates that the hypergraph structure is more helpful in capturing high-order collaborative information between individual travel trajectories.

[0188] (5) Compared with other graph-based methods (excluding hypergraph-based methods), AGRAN using only a single graph structure has better prediction performance, indicating that the graph topology is optimized after GSL, thereby enhancing its node representation capability.

[0189] (6) Compared to the NYC dataset, MAHCN shows a slightly lower improvement in prediction performance on the TKY and CA datasets. This indicates that on smaller datasets like NYC, the adaptive hypergraph constructed in this invention has a better graph structure and can better model individual travel behavior patterns.

[0190] (3) Ablation experiment:

[0191] To evaluate the contribution of each module of the model to the overall prediction performance, ablation experiments were conducted on several variants of MAHCN on different datasets. The results are shown in Table 3. The specific variant model designs are as follows:

[0192] w / o AHUM-I: This variant model removes the AHUM-I module.

[0193] w / o AHUM-G: This variant model removes the AHUM-G module.

[0194] w / o PDAS: This variant model removes the PDAS policy.

[0195] w / o MHCM-CL: This variant model removes the contrastive learning framework from the AHUM module.

[0196] Table 3 Ablation experimental results of different MAHCN variant models

[0197]

[0198] Based on the experimental results in Table 3, the following conclusions can be drawn from this experiment:

[0199] (1) The prediction performance of the variant models w / o AHUM-I and w / o AHUM-G is weaker than that of MAHCN, indicating that the adaptive update of the hypergraph optimizes the prediction performance. and The graph structure was modified to enhance its node representation capabilities. Furthermore, on the TKY and CA datasets, the performance degradation of the variant models w / o AHUM-I and w / o AHUM-G was more significant, indicating that as the graph size increases, the weakening effect of graph structure rigidity on node representation capabilities becomes more pronounced.

[0200] (2) The prediction performance of the variant model w / o PDAS is weaker than that of MAHCN, indicating that the PDAS strategy can effectively simulate hypergraph structure perturbation and significantly improve the robustness of the model in scenarios where check-in data is generally noisy and sparse.

[0201] (3) The predictive performance of the variant w / o MHCM-CL is weaker than that of MAHCN, indicating that by fusing the interaction between individuals and locations, the directed transfer relationships between locations, and the geographic feature associations of locations, the contrastive learning framework can effectively promote collaboration between hypergraphs, thereby enhancing the model's multi-factor modeling capabilities. Furthermore, compared to the other two datasets, this variant exhibits a greater performance drop on the TKY dataset, because the average number of visits to each location is higher in the TKY dataset, and the relationship between locations and individuals is more complex. In this case, hypergraph collaboration is particularly important for a comprehensive understanding of individual travel behavior patterns.

[0202] (4) Analytical experiments on hypergraph construction and updating

[0203] To evaluate the effectiveness of the proposed hypergraph construction method, the experiment was designed as follows: Connection weights in the interaction hypergraph and directed hypergraph were removed, and different geographic distance thresholds were set. Geographic hypermaps were constructed respectively. Experimental results are as follows: Figure 4 As shown.

[0204] like Figure 4 As shown in the first and second columns, weights based on the number of visits enable the interaction hypergraph to accurately characterize individual preferences for specific locations, while weights based on location index intervals effectively quantify the directed transition relationships between locations in the directed hypergraph. The results show that both weighting methods can construct more effective hypergraphs, thereby enhancing the overall predictive performance of MAHCN. Figure 4 As shown in the third column, different The impact of the value on model performance is negligible, indicating that the proposed AHUM-G module is not sensitive to fluctuations in the geographical distance threshold and has strong robustness.

[0205] Furthermore, to further analyze the impact of hypergraph update methods on hypergraph structure, this experiment used t-SNE to visualize the feature embeddings at all locations in the NYC dataset, such as... Figure 5As shown, compared to the fixed hypergraph structure, the adaptive hypergraph structure generates a more uniform overall embedding distribution of positions, indicating that the hypergraph update method can effectively optimize the graph structure, thereby generating higher-quality node representations.

[0206] Comparative experiments on three real check-in datasets show that the proposed MAHCN exhibits higher accuracy and stronger robustness compared to existing state-of-the-art baseline methods in the next location prediction task, demonstrating superior next location prediction performance.

[0207] Example 2:

[0208] This embodiment provides a traffic next location prediction device based on multi-adaptive hypergraph collaboration, including a processor and a memory storing a number of computer instructions. When the computer instructions are executed by the processor, they implement the steps of the traffic next location prediction method based on multi-adaptive hypergraph collaboration.

[0209] For specific limitations on the next location prediction device for transportation based on multi-adaptive hypergraph collaboration, please refer to the limitations on the next location prediction method for transportation based on multi-adaptive hypergraph collaboration mentioned above, which will not be repeated here.

[0210] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, which implements the method of the present invention by running the computer program stored in the memory.

[0211] The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory stores the program, and the processor executes the program upon receiving an execution instruction.

[0212] The processor may be an integrated circuit chip with data processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0213] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0214] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for predicting next location of a traffic trip based on multi-adaptive hypergraph collaboration, characterized in that, The method comprises the following steps: Constructing an interaction hypergraph and a directed hypergraph based on individual travel trajectory sequences, and constructing a geographic hypergraph based on location-based geographic feature data; Updating the interaction hypergraph and the geographic hypergraph according to the feature similarity between locations, and introducing a graph structure enhancement strategy in the updating process to finally output the updated adaptive interaction hypergraph and the adaptive geographic hypergraph; Modeling individual travel trajectory sequences using a trajectory sequence encoder to form an individual sequence embedding matrix; Using an interaction hypergraph convolution block to process the interaction hypergraph and the adaptive interaction hypergraph respectively to obtain a location interaction embedding matrix and an adaptive location interaction embedding matrix, and extracting an individual interaction embedding matrix based on the location interaction embedding matrix; Using a directed hypergraph convolution block to process the directed hypergraph to obtain a location directed embedding matrix, and extracting an individual directed embedding matrix based on the location directed embedding matrix; Using a geographic hypergraph convolution block to process the geographic hypergraph and the adaptive geographic hypergraph respectively to obtain a location geographic embedding matrix and an adaptive location geographic embedding matrix, and extracting an individual geographic embedding matrix based on the location geographic embedding matrix; Aggregating the location interaction embedding matrix, the location directed embedding matrix and the location geographic embedding matrix to obtain a location overall embedding matrix, and aggregating the individual sequence embedding matrix, the individual interaction embedding matrix, the individual directed embedding matrix and the individual geographic embedding matrix to obtain an individual overall embedding matrix; According to the location overall embedding matrix and the individual overall embedding matrix, the highest probability of the first k positions is output as the prediction result of the next travel location of the individual through a normalization function.

2. The method of claim 1, wherein, The hyperedge set in the interaction hypergraph is composed of individuals, the node set in the interaction hypergraph is composed of locations, all locations visited by an individual are connected by taking the individual as a hyperedge, and a connection weight between the individual and the location in the interaction hypergraph is given according to the number of times the individual visits the location; The hyperedge set in the directed hypergraph is composed of source locations, the node set in the directed hypergraph is composed of target locations, all target locations corresponding to a source location are connected by taking the source location as a hyperedge, and a connection weight between the source location and the target location in the directed hypergraph is given based on the index interval of the locations; The hyperedge set and the node set in the geographic hypergraph are both composed of locations, adjacent locations within a pre-defined geographic distance threshold are identified with each location as the center, and all adjacent locations are connected by taking the center location as a hyperedge.

3. The method of claim 1, wherein, The updating process of the interaction hypergraph is as follows: An interaction similarity matrix is constructed according to the location interaction embedding matrix; For each individual, according to the interaction similarity matrix, extract the position with the maximum similarity and not belonging to the current individual Add the extracted position to the individual traffic travel trajectory sequence of the current individual to obtain a travel trajectory expansion sequence According to the positions involved in the travel trajectory expansion sequence, an interaction similarity sub-matrix is extracted from the interaction similarity matrix, a position with the lowest similarity in the interaction similarity sub-matrix is determined, and the position with the lowest similarity in the travel trajectory expansion sequence is removed to obtain a travel trajectory correction sequence. ​​ The travel trajectory correction sequence is randomly disturbed by using a graph structure enhancement strategy, and the interaction hypergraph is reconstructed according to the travel trajectory correction sequence after random disturbance, which is denoted as an adaptive interaction hypergraph.

4. The method of claim 1, wherein, The updating process of the geographic hypergraph is as follows: A geographic similarity matrix is constructed according to the location geographic embedding matrix; A matrix with all elements being 1 is subtracted from the geographic similarity matrix as an adjustment factor; The adjustment factor is multiplied with a geographic distance matrix element by element to obtain a perceived distance matrix, the geographic distance matrix is composed of geographic distances between two locations; The perceived distance matrix is randomly disturbed by using a graph structure enhancement strategy; Based on the sensing distance matrix after random disturbance, the adjacent positions within a predefined sensing distance threshold are identified with each position as the center to obtain a position neighbor set; The position neighbor set is randomly disturbed by using a graph structure enhancement strategy, and all adjacent positions in the position neighbor set after random disturbance are connected by taking the center position as a hyperedge.

5. The method of claim 1, wherein, The trajectory sequence encoder includes a time sub-encoder and a sequence sub-encoder, and modeling individual traffic trajectory sequences by using the trajectory sequence encoder forms an individual sequence embedding matrix, which includes: The time sub-encoder encodes the time index in each individual traffic trajectory sequence by using a time2vector method to obtain a time embedding vector, and splices the time embedding vector with a learnable position embedding vector to obtain a trajectory sequence embedding matrix; The sequence sub-encoder processes each trajectory sequence embedding matrix by using a Transformer encoder, and obtains individual sequence embeddings after average pooling of the output of the Transformer encoder; All individual sequence embeddings are aggregated to form an individual sequence embedding matrix.

6. The method of claim 1, wherein, The individual interaction embedding matrix is extracted and generated based on the position interaction embedding matrix, which includes taking all position interaction embeddings belonging to the same individual in the position interaction embedding matrix, and performing average aggregation on all position interaction embeddings of the same individual to obtain an individual interaction embedding, and aggregating individual interaction embeddings of all individuals to form an individual interaction embedding matrix; The individual directed embedding matrix is extracted and generated based on the position directed embedding matrix, which includes taking all position directed embeddings belonging to the same individual in the position directed embedding matrix, and performing average aggregation on all position directed embeddings of the same individual to obtain an individual directed embedding, and aggregating individual directed embeddings of all individuals to form an individual directed embedding matrix; The individual geographic embedding matrix is extracted and generated based on the position geographic embedding matrix, which includes taking all position geographic embeddings belonging to the same individual in the position geographic embedding matrix, and performing average aggregation on all position geographic embeddings of the same individual to obtain an individual geographic embedding, and aggregating individual geographic embeddings of all individuals to form an individual geographic embedding matrix.

7. The method of claim 1, wherein, In the training and updating process of the trajectory sequence encoder, the interaction supergraph convolution block, the directed supergraph convolution block and the geographic supergraph convolution block, the total loss function includes a total contrastive learning loss, a cross-entropy loss and a KL divergence loss.

8. The method of claim 7, wherein, The total contrastive learning loss includes an interaction contrastive learning loss between the position interaction embedding matrix and the adaptive position interaction embedding matrix, a geographic contrastive learning loss between the position geographic embedding matrix and the adaptive position geographic embedding matrix, an interaction directed contrastive learning loss between the individual interaction embedding matrix and the individual directed embedding matrix, an interaction geographic contrastive learning loss between the individual interaction embedding matrix and the individual geographic embedding matrix, and a directed geographic contrastive learning loss between the individual directed embedding matrix and the individual geographic embedding matrix.

9. The method of claim 7, wherein, The KL divergence loss includes: An interaction similarity matrix is constructed according to the position interaction embedding matrix; A directed similarity matrix is constructed according to the position directed embedding matrix; A geographic similarity matrix is constructed according to the position geographic embedding matrix; An interaction similarity matrix is constructed according to the position interaction embedding matrix; A directed similarity matrix is constructed according to the position directed embedding matrix; A geographic similarity matrix is constructed according to the position geographic embedding matrix; The KL divergence between the interaction similarity matrix and the directed similarity matrix, and the KL divergence between the geographic similarity matrix and the directed similarity matrix are calculated to form a KL divergence loss.

10. A next-location prediction device for traffic travel based on multi-adaptive hypergraph collaboration, comprising a processor and a memory storing a plurality of computer instructions, characterized in that, The computer instructions, when executed by a processor, implement the steps of the method for predicting a next location of a traffic trip based on multi-adaptive hypergraph collaboration according to any one of claims 1-9.

Images