Taxi pick-up point recommendation method and system based on spatiotemporal directed graph convolutional network

Abstract

The method for recommending taxi pick-up points based on a spatiotemporal directed graph convolutional network includes: (1) representing the spatiotemporal graph of the trajectory. The traditional representation method is replaced by a representation based on a directed graph, and the pick-up point or drop-off point containing characteristic attributes, i.e., the point of interest, is used as a node, and a directed edge is used to represent the direction from the pick-up point to the drop-off point; (2) based on the analysis and calculation results of step (1), the computing tasks are allocated according to the representation of the spatiotemporal graph of the trajectory; (3) the roadside unit uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points. The present invention also includes a taxi pick-up point recommendation system based on a spatiotemporal directed graph convolutional network. The present invention combines a cloud-edge collaborative framework that combines edge computing with cloud computing to adapt to the processing and analysis of local, real-time, short-cycle data, and can better support real-time intelligent decision-making and execution of local business.

Description

Technical Field

[0001] The invention relates to a network-based taxi pick-up point recommendation method and system. Background Art

[0002] As a crucial component of urban public transportation, taxis play an indispensable role in urban development, serving as mobile detectors for urban population movement. Recommended taxi pick-up points play a crucial role in ensuring efficient passenger pickup for drivers and increasing revenue for taxi operators. The rapid development of connected vehicle technology has enabled vehicles to obtain accurate, real-time information about surrounding vehicles and road conditions. However, in recent years, numerous news reports have emerged about taxis disregarding traffic regulations and illegally parking in order to pick up passengers. These incidents stem from drivers arbitrarily selecting pick-up and drop-off points due to incomplete road information. This not only contributes to traffic congestion but also poses significant safety risks. Therefore, recommending taxi pick-up points is crucial. This not only reduces idle taxi runs but also shortens passenger wait times, making travel more convenient and optimizing public transportation systems.

[0003] However, accurately recommending taxi pick-up points in the context of the Internet of Vehicles (IoV) presents several challenges. This requires modeling complex urban road networks and considering the influence of surrounding environmental factors, such as points of interest (POIs). To characterize spatial correlations, some work has applied convolutional neural networks (CNNs) to spatial analysis. However, the pixels in image or video data processed by CNNs are arranged in neat matrices. CNNs are unable to process data in non-Euclidean spaces, and therefore cannot fully describe complex road network topologies. This makes modeling the relationships between taxi pick-up points difficult. Numerous studies have used only single-source data, such as data containing characteristic parameters such as traffic volume, speed, and density. However, the demand for pick-up points is significantly influenced by external factors (POIs, weather conditions, and holidays). This significantly impacts the accuracy of taxi pick-up point recommendations.

[0004] Furthermore, achieving accurate taxi pick-up point recommendations requires processing large amounts of trajectory data for model training. However, previous research has traditionally focused on analyzing data and executing decisions by uploading all data to a central server for centralized learning and computation. The data is centrally stored on this central server, and all computational tasks for the entire system are centrally deployed on this server, centrally processing all system functions. This presents two problems: 1) Under the centralized learning framework, uploading all data to the central server for preprocessing and learning places very high demands on the central server's hardware; 2) Since all data is stored on the central server, it presents security risks that can be easily attacked and can lead to data privacy leaks. These issues make effective pick-up point recommendations difficult in practice.

[0005] In summary, existing taxi trajectory-based recommendation methods have the following major problems: 1) Existing taxi trajectory-based recommendation methods do not model dynamic vehicle trajectory information well; 2) When processing and training large amounts of trajectory data, the computational pressure on the central server and the problem of privacy leakage are not taken into consideration. Summary of the Invention

[0006] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a taxi pick-up point recommendation method and system based on a spatiotemporal directed graph convolutional network under cloud-edge collaboration, which can accurately recommend pick-up points for taxi drivers.

[0007] The method of the present invention firstly remotely controls the GPS trajectory information to construct a spatiotemporal graph of sequential pick-up and drop-off points, and then uploads it to the roadside unit. The spatiotemporal graph uploaded to the roadside unit is then uploaded to the cloud server via the base station for integration, and the spatiotemporal data of each taxi is constructed into a total spatiotemporal graph, with the same pick-up and drop-off points merged. Finally, in order to better capture local correlations and maintain displacement invariance, and because the association between roads is defined in the form of a graph, the roadside unit uses a spatiotemporal graph directed convolutional neural network to perform convolution operations on the spatiotemporal graph, uses a neural network model for training, captures the potential features of the pick-up points, and outputs a prediction result, thereby achieving the purpose of accurately recommending taxi pick-up points. The present invention proposes a directed graph convolutional neural network (STDCN) model based on the trajectory spatiotemporal graph by combining the Internet of Vehicles. The trajectory information of each taxi is processed into trajectory spatiotemporal graph data, and the influence of the surrounding external environment (such as POI) is taken into account, and the STDCN is used to make accurate predictions. In addition, we have combined edge computing with cloud computing to form a cloud-edge collaborative framework to adapt to local, real-time, short-cycle data processing and analysis, which can better support real-time intelligent decision-making and execution of local business.

[0008] The taxi pick-up point recommendation method based on spatiotemporal directed graph convolutional network under cloud-edge collaboration in this invention has the following specific implementation steps:

[0009] (1) Representing the spatiotemporal graph of trajectories. Due to the complex nature of urban road networks, most roads in urban traffic networks generally allow for three-dimensional structures, and the characteristics of information in adjacent areas are highly correlated. Therefore, it is possible to consider using a directed graph-based representation to replace the traditional representation method. The boarding points or drop-off points containing some characteristic attributes, i.e., some points of interest, are used as nodes, and directed edges are used to represent the direction from the boarding point to the drop-off point.

[0010] In terms of form, the time series pick-up and drop-off points are connected with directed weighted edges to construct a time-space graph for each taxi. The time-space graph is composed of a weighted directed graph express:

[0011]

[0012] in v i represents a pick-up or drop-off point, and Each edge e ij Represents v i and v j The correlation between them. yes The weighted adjacency matrix of edge a ij The weight represents v i and v j The correlation between them.

[0013] s11. Construct directed edges in the spatial dimension. In the spatial dimension, edges represent the connection between upstream and downstream parts;

[0014] s12. Construct directed edges in the time dimension. Here, a directed edge is specifically defined as one that points from the previous moment to the same boarding or disembarking point at the next moment.

[0015] (2) Based on the analysis and calculation results of step (1), the calculation tasks are allocated according to the representation of the trajectory space-time graph. The specific calculation process is as follows:

[0016] After the cloud server constructs the total spatiotemporal graph, it is necessary to define a formula for calculating the edge weights between adjacent nodes to ensure that the different characteristics of nodes in both spatiotemporal and temporal dimensions can be processed using the same analysis framework. Grey correlation analysis is used to determine the correlation between nodes, which serves as the weight between edges. This includes:

[0017] s21. Determine the analysis objects in the two dimensions of time and space respectively. At the same time, divide the analysis objects into two parts: reference objects (evaluation criteria) and comparison objects (evaluation objects). In the spatial dimension, the reference object is the boarding point, and the comparison object is the drop-off point directly connected to the boarding point. Each type of object includes POI and distance information. In the temporal dimension, the reference object is the boarding and dropping-off point specified at the current moment, and the comparison object is the same boarding and dropping-off point at the previous / next moment. Each type of object includes GPS distribution at different time granularities (minutes, hours, days, weeks, etc.);

[0018] s22. Perform a non-dimensional hardening (normalization) process on the multidimensional features of the above data. Since the physical meaning of each dimension of data is different in the time and space dimensions, the dimensions of the data are also different and difficult to compare intuitively. For each comparison object in the time and space dimensions, a reference object is used to define the scaling of the multidimensional information, and the values ​​of each dimension of data fall between 0 and 1, as shown below:

[0019]

[0020] in is the mean of the comparison objects, is the mean of the reference object, and σ is the standard deviation.

[0021] Therefore, x c It can be standardized as follows:

[0022]

[0023] in

[0024] s23. Perform difference analysis. Calculate the absolute value difference △(k) on the corresponding dimension between each comparison object and the reference object one by one. The formula is as follows:

[0025]

[0026] where x c (k) is the value of the comparison object in the kth dimension, μ r (k) is the value of the reference object in the kth dimension.

[0027] s24. Calculate the grey relational coefficient. The formula is as follows:

[0028]

[0029] Where ρ∈[0,1] is the recognition coefficient, represents the minimum function, Represents the maximum function, i represents the reference object, and j represents the comparison object.

[0030] s25. Calculate the grey relational degree. The formula is as follows:

[0031] a ij =∑w k ξ ij (k) (6)

[0032] where w k Represents custom weight.

[0033] s26. Construct and assign subtasks. The server needs to use the aforementioned association to define two parameters, K (maximum hop count) and N (number of subgraph nodes), to construct the subgraph. If K is set, and the number of nodes is less than N, nodes with an association of 0 are added; if the number of nodes is greater than N, some nodes need to be deleted. First, the association between all other nodes in the subgraph and the target node needs to be calculated using the following formula:

[0034] a it =max{a i(i+1) ·a (i+1)(i+2) …a (t-1)t} (7)

[0035] There is at least one reachable path from node i to target node t: P = [a i(i+1) ,a (i+1)(i+2) ,…,a (t-1)t ].

[0036] Then sort the associations from all other nodes i to the target node t, and delete the points with the smallest associations one by one.

[0037] Based on the above steps, the cloud server constructs a subgraph centered on each pick-up point and assigns them as subtasks to the roadside units in the corresponding area.

[0038] (3) The roadside unit uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points. It mainly includes: input layer, STDCN layer, Dropout layer, Embedding layer, Softmax layer and output layer. The specific calculation process is as follows:

[0039] s31. In order to extract features more efficiently, the adjacency matrix is ​​mirrored into a graph structure matrix The specific formula is as follows:

[0040]

[0041] in is the adjacency matrix.

[0042] s32. Definition As a feature matrix, where N represents the number of pick-up points and M represents the attributes of POIs;

[0043] s33. Use STDCN to perform convolution operations to capture features. Each subgraph is fed into STDCN as input, and the formula is as follows:

[0044]

[0045] Where l represents the layer index, A symmetric matrix representing the adjacency matrix. W (l) Represents the weight matrix of layer l. (l) represents the input matrix of layer l, Represents the raw input of the neural network. ReLU(·) represents the activation function;

[0046] s34. In order to reduce the spatial dimension of a discrete variable while still representing it meaningfully, an embedding layer is used, as follows:

[0047] Linear(·)=W L H c (10)

[0048] Where W L represents the trainable weight matrix of the linear layer, and c represents the number of cycles of the spatiotemporal directed graph convolutional network.

[0049] s35. To classify whether a location is a pick-up point, use the Softmax function as the activation function. The formula is as follows:

[0050]

[0051] in is the output of the previous layer and the input of the Softmax function, y i is the probability that the i-th prediction object belongs to the category (pick-up point, drop-off point, both, or neither).

[0052] And using cross entropy as the loss function, the formula is as follows:

[0053]

[0054] where t i is the true value of the i-th object.

[0055] s36. Broadcast all the pick-up point information obtained in s35 to the taxis in the area to complete the recommendation of pick-up points under cloud-edge collaboration.

[0056] The system for implementing the taxi pick-up point recommendation method based on the spatiotemporal directed graph convolutional network of the present invention comprises a trajectory spatiotemporal graph identification module, a computing task allocation module, a spatiotemporal graph feature extraction module, and a pick-up point recommendation module connected in sequence;

[0057] The trajectory spatiotemporal graph identification module replaces the traditional representation method with a directed graph-based representation. The boarding or alighting points containing characteristic attributes, i.e., points of interest, are used as nodes, and directed edges are used to indicate the connection between the boarding point and the alighting point.

[0058] The computing task allocation module allocates computing tasks based on the analysis and calculation results of the trajectory space-time graph identification module and the representation of the trajectory space-time graph;

[0059] The spatiotemporal graph feature extraction and pick-up point recommendation module uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points, including: input layer, STDCN layer, Dropout layer, Embedding layer, Softmax layer and output layer.

[0060] The taxi pick-up point recommendation method based on the spatiotemporal directed graph convolutional network of the present invention first constructs the GPS trajectory information into a spatiotemporal graph of the time-series pick-up and drop-off points, and then uploads it to the roadside unit. The spatiotemporal graph uploaded to the roadside unit is then uploaded to the cloud server through the base station for integration, and the spatiotemporal data of each taxi is constructed into a total spatiotemporal graph, and the same pick-up and drop-off points are merged. Finally, in order to better capture local correlations and maintain displacement invariance, and because the association between roads is defined in the form of a graph, the roadside unit uses a spatiotemporal directed graph convolutional neural network to perform convolution operations on the spatiotemporal graph, uses a neural network model for training, captures the potential features of the pick-up points, and outputs a prediction result to achieve the purpose of accurately recommending taxi pick-up points. The present invention proposes a spatiotemporal directed graph convolutional neural network (STDCN) model by combining the Internet of Vehicles, processes the trajectory information of each taxi into trajectory spatiotemporal graph data, and takes into account the influence of the surrounding external environment (such as POI), and uses STDCN to make accurate predictions. In addition, the present invention combines edge computing with cloud computing in a cloud-edge collaborative framework to adapt to local, real-time, short-cycle data processing and analysis, which can better support real-time intelligent decision-making and execution of local business.

[0061] The beneficial effects of the present invention are: (1) a spatiotemporal graph directed graph convolutional neural network model is proposed, which connects the time series pick-up points with directed weighted edges to construct a trajectory spatiotemporal graph. In addition, the directed graph convolutional network is used to recommend taxi pick-up points, so that taxis can pick up passengers efficiently; (2) the present invention adopts a cloud-edge collaborative framework to combine cloud computing and edge computing. Most of the computing load is integrated into the roadside unit, and the cloud computing center only distributes tasks. This reduces the pressure on the cloud server and can also protect the privacy of the data; BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is an overall flow chart of the method of the present invention.

[0063] Figure 2 This is the overall structure diagram of the spatiotemporal directed graph convolutional neural network (STDCN) model of the method of the present invention. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0065] The taxi pick-up point recommendation method based on spatiotemporal directed graph convolutional network under cloud-edge collaboration in this invention has the following specific implementation steps:

[0066] (1) Representing the spatiotemporal graph of trajectories. Due to the complex nature of urban road networks, most roads in urban traffic networks generally allow for three-dimensional structures, and the characteristics of information in adjacent areas are highly correlated. Therefore, it is possible to consider using a directed graph-based representation to replace the traditional representation method. The boarding points or drop-off points containing some characteristic attributes, i.e., some points of interest, are used as nodes, and directed edges are used to represent the direction from the boarding point to the drop-off point.

[0067] In terms of form, the time series pick-up and drop-off points are connected with directed weighted edges to construct a time-space graph for each taxi. The time-space graph is composed of a weighted directed graph express:

[0068]

[0069] in v i represents a pick-up or drop-off point, and Each edge e ij Represents v i and v j The correlation between them. yes The weighted adjacency matrix of edge a ij The weight represents v i and v j The correlation between them.

[0070] s11. Construct directed edges in the spatial dimension. In the spatial dimension, edges represent the connection between upstream and downstream parts;

[0071] s12. Construct directed edges in the time dimension. Here, a directed edge is specifically defined as one that points from the previous moment to the same boarding or disembarking point at the next moment.

[0072] (2) Based on the analysis and calculation results of step (1), the calculation tasks are allocated according to the representation of the trajectory space-time graph. The specific calculation process is as follows:

[0073] After the cloud server constructs the total spatiotemporal graph, it is necessary to define a formula for calculating the edge weights between adjacent nodes to ensure that the different characteristics of nodes in both spatiotemporal and temporal dimensions can be processed using the same analysis framework. Grey correlation analysis is used to determine the correlation between nodes, which serves as the weight between edges. This includes:

[0074] s21. Determine the analysis objects in the two dimensions of time and space respectively. At the same time, divide the analysis objects into two parts: reference objects (evaluation criteria) and comparison objects (evaluation objects). In the spatial dimension, the reference object is the boarding point, and the comparison object is the drop-off point directly connected to the boarding point. Each type of object includes POI and distance information. In the temporal dimension, the reference object is the boarding and dropping-off point specified at the current moment, and the comparison object is the same boarding and dropping-off point at the previous / next moment. Each type of object includes GPS distribution at different time granularities (minutes, hours, days, weeks, etc.);

[0075] s22. Perform a non-dimensional hardening (normalization) process on the multidimensional features of the above data. Since the physical meaning of each dimension of data is different in the time and space dimensions, the dimensions of the data are also different and difficult to compare intuitively. For each comparison object in the time and space dimensions, a reference object is used to define the scaling of the multidimensional information, and the values ​​of each dimension of data fall between 0 and 1, as shown below:

[0076]

[0077] in is the mean of the comparison objects, is the mean of the reference object, and σ is the standard deviation.

[0078] Therefore, x c It can be standardized as follows:

[0079]

[0080] in

[0081] s23. Perform difference analysis. Calculate the absolute value difference △(k) on the corresponding dimension between each comparison object and the reference object one by one. The formula is as follows:

[0082]

[0083] where x c (k) is the value of the comparison object in the kth dimension, μ r (k) is the value of the reference object in the kth dimension.

[0084] s24. Calculate the grey relational coefficient. The formula is as follows:

[0085]

[0086] Where ρ∈[0,1] is the recognition coefficient, represents the minimum function, Represents the maximum function, i represents the reference object, and j represents the comparison object.

[0087] s25. Calculate the grey relational degree. The formula is as follows:

[0088] a ij =∑w k ξ ij (k) (6)

[0089] where w k Represents custom weight.

[0090] s26. Construct and assign subtasks. The server needs to use the aforementioned association to define two parameters, K (maximum hop count) and N (number of subgraph nodes), to construct the subgraph. If K is set, and the number of nodes is less than N, nodes with an association of 0 are added; if the number of nodes is greater than N, some nodes need to be deleted. First, the association between all other nodes in the subgraph and the target node needs to be calculated using the following formula:

[0091] a it =max{a i(i+1) ·a (i+1)(i+2) …a (t-1)t} (7)

[0092] There is at least one reachable path from node i to target node t: P = [a i(i+1) ,a (i+1)(i+2) ,…,a (t-1)t ].

[0093] Then sort the associations from all other nodes i to the target node t, and delete the points with the smallest associations one by one.

[0094] Based on the above steps, the cloud server constructs a subgraph centered on each pick-up point and assigns them as subtasks to the roadside units in the corresponding area.

[0095] (3) The roadside unit uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points. It mainly includes: input layer, STDCN layer, Dropout layer, Embedding layer, Softmax layer and output layer. The specific calculation process is as follows:

[0096] s31. In order to extract features more efficiently, the adjacency matrix is ​​mirrored into a graph structure matrix The specific formula is as follows:

[0097]

[0098] in is the adjacency matrix.

[0099] s32. Definition As a feature matrix, where N represents the number of pick-up points and M represents the attributes of POIs;

[0100] s33. Use STDCN to perform convolution operations to capture features. Each subgraph is fed into STDCN as input, and the formula is as follows:

[0101]

[0102] Where l represents the layer index, A symmetric matrix representing the adjacency matrix. W (l) Represents the weight matrix of layer l. (l) represents the input matrix of layer l, Represents the raw input of the neural network. ReLU(·) represents the activation function;

[0103] s34. In order to reduce the spatial dimension of a discrete variable while still representing it meaningfully, an embedding layer is used, as follows:

[0104] Linear(·)=W L H c (10)

[0105] Where W L represents the trainable weight matrix of the linear layer, and c represents the number of cycles of the spatiotemporal directed graph convolutional network.

[0106] s35. To classify whether a location is a pick-up point, use the Softmax function as the activation function. The formula is as follows:

[0107]

[0108] in is the output of the previous layer and the input of the Softmax function, y i is the probability that the i-th prediction object belongs to the category (pick-up point, drop-off point, both, or neither).

[0109] And using cross entropy as the loss function, the formula is as follows:

[0110]

[0111] where t i is the true value of the i-th object.

[0112] s36. Broadcast all the pick-up point information obtained in s35 to the taxis in the area to complete the recommendation of pick-up points under cloud-edge collaboration.

[0113] As attached Figure 1 In the method of the present invention, the remote control unit first constructs the GPS trajectory information into a time-series spatiotemporal graph of pick-up and drop-off points, and then uploads it to the roadside unit. The spatiotemporal graph uploaded to the roadside unit is then uploaded to the cloud server via the base station for integration. The spatiotemporal data of each taxi is constructed into an overall spatiotemporal graph, and the same pick-up and drop-off points are merged. Finally, in order to better capture local correlations and maintain displacement invariance, and because the associations between roads are defined in the form of a graph, the roadside unit uses a directed convolutional neural network of the spatiotemporal graph to perform convolution operations on the spatiotemporal graph. The neural network model is used for training, capturing the potential features of the pick-up points, and outputting prediction results to achieve the purpose of accurately recommending taxi pick-up points.

[0114] As attached Figure 2 This method uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points. It mainly consists of an input layer, a STDCN layer, a dropout layer, an embedding layer, a softmax layer, and an output layer. The input layer contains an adjacency matrix and a feature matrix. The deepening STDCN layer performs convolution operations to capture features, and each subgraph is used as input to feed the STDCN layer. The dropout layer prevents gradient vanishing as the neural network deepens. Through encoding in the embedding layer and classification in the softmax layer, the output layer determines whether a point is a pick-up point and recommends it.

[0115] The system for implementing the taxi pick-up point recommendation method based on a spatiotemporal directed graph convolutional network of the present invention comprises a trajectory spatiotemporal graph identification module, a computational task allocation module, and a spatiotemporal graph feature extraction and pick-up point recommendation module, which are connected in sequence. The trajectory spatiotemporal graph identification module includes the contents of step (1) of the method of the present invention, the computational task allocation module includes the contents of step (2), and the spatiotemporal graph feature extraction and pick-up point recommendation module includes the contents of step (3).

[0116] Implementation application cases show that the taxi pick-up point recommendation method based on spatiotemporal directed graph convolutional network under cloud-edge collaboration proposed in this invention is effective. Compared with other design methods, the method of this invention effectively models dynamic vehicle trajectory information, and considering the computational pressure and privacy leakage issues of the central server, the design scheme is more in line with actual data processing conditions.

[0117] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A taxi pick-up point recommendation method based on a spatiotemporal directed graph convolutional network includes the following steps: (1) Representing the spatiotemporal graph of trajectories; using a directed graph-based representation to replace the traditional representation method, taking the boarding or disembarking points containing characteristic attributes, i.e., points of interest, as nodes, and using directed edges to represent the direction from the boarding point to the disembarking point; Formally, the time-series pick-up and drop-off points are connected with directed weighted edges to construct a space-time graph for each taxi. The space-time graph is represented by a weighted directed graph D: D={V,E,A} (1) Where V={v p }, v p represents a boarding point or alighting point, and E={e pd }, each edge e pd Represents v p and v d the correlation between is the weighted adjacency matrix of D, edge a pd The weight represents v p and v d The degree of correlation between s11. Construct directed edges in the spatial dimension; in the spatial dimension, edges represent connections between upstream and downstream parts; s12. Construct directed edges in the time dimension. Here, a directed edge is defined as one that points from the previous moment to the same boarding or disembarking point at the next moment. (2) Based on the analysis and calculation results of step (1), the calculation tasks are allocated according to the representation of the trajectory space-time graph. The specific calculation process is as follows: After the cloud server builds the total spatiotemporal graph, it is necessary to define the calculation formula for the edge weights between adjacent nodes to ensure that the different characteristics of the nodes in both spatiotemporal and temporal dimensions are processed using the same analysis framework; Use grey correlation analysis to determine the correlation between nodes, which is used as the weight between edges. Specifically, it includes: s21. Determine the analysis objects in both the temporal and spatial dimensions. The analysis objects are divided into two parts: reference objects and comparison objects. The reference object serves as the evaluation standard, and the comparison object serves as the evaluation object. In the spatial dimension, the reference object is the boarding point, and the comparison object is the directly connected drop-off point. Each object type includes POI and distance information. In the temporal dimension, the reference object is the designated boarding point at the current moment, and the comparison object is the same boarding point at the previous or next moment. Each object type includes GPS distribution at different time granularities. s22. Quantitatively normalize the multidimensional features of POI, distance information, and GPS distribution. Because the physical meanings of the data in each dimension differ across time and space, the dimensions of the data also differ, making intuitive comparison difficult. For each comparison object in time and space, a reference object is used to define the scaling of the multidimensional information, keeping the values ​​of each dimension between 0 and 1, as shown below: in is the mean of the comparison objects, is the mean of the reference object, σ is the standard deviation; Therefore, x c Normalization, the formula is as follows: in s23. Perform difference analysis; calculate the absolute value difference Δ(k) on the corresponding dimension between each comparison object and the reference object one by one, using the following formula: in is the value of the comparison object in the kth dimension, μ r (k) is the value of the reference object in the kth dimension; s24. Calculate the grey relational coefficient. The formula is as follows: Where ρ∈[0,1] is the recognition coefficient, represents the minimum function, represents the maximum function, i represents the reference object, and j represents the comparison object; s25. Calculate the grey relational degree. The formula is as follows: a ij =∑w k x ij (k) (6) where w k Represents custom weight; s26. Construct and assign subtasks. The server needs to use the aforementioned association to define two parameters: the maximum number of hops K and the number of subgraph nodes N to construct the subgraph. If K is set and the number of nodes is less than N, nodes with an association of 0 are added. If the number of nodes is greater than N, some nodes need to be deleted. First, the association between all other nodes in the subgraph and the target node needs to be calculated. The formula is as follows: a xt =max{a x(x+1) ·a (x+1)(x+2) …a (t-1)t } (7) There is at least one reachable path from node x to target node t: P = [a x(x+1) ,a (x+1)(x+2) ,…,a (t-1)t ]; Then sort the associations from all other nodes x to the target node t, and delete the points with the smallest associations in turn; Based on the above steps, the cloud server constructs a subgraph centered on each pick-up point and assigns them as subtasks to the roadside units in the corresponding area; (3) The roadside unit uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points. The network includes an input layer, a STDCN layer, a dropout layer, an embedding layer, a softmax layer, and an output layer. The STDCN layer is a directed graph convolutional neural network model based on the spatiotemporal graph of trajectories. The specific calculation process is as follows: s31. In order to extract features more effectively, the adjacency matrix is ​​mirrored into the graph structure matrix A′. The specific formula is as follows: A′=A+A T (8) in is the adjacency matrix; s32. Definition As a feature matrix; where N represents the number of subgraph nodes and M represents the attributes of POI; s33. Use STDCN to perform convolution operations to capture features; each subgraph is fed into STDCN as input, and the formula is as follows: A (l+1) =ReLU(A′H (l) W (l) ) (9) Where l represents the layer index, A′ represents the symmetric matrix of the adjacency matrix; W (l) Represents the weight matrix of the lth layer; H (l) represents the input matrix of layer l, Represents the original input of the neural network; ReLU(·) represents the activation function; s34. In order to reduce the spatial dimension of a discrete variable while still representing it meaningfully, an embedding layer is used, as follows: Linear(·)=W L H c (10) Where W L represents the trainable weight matrix of the linear layer, and c represents the number of cycles of the spatiotemporal directed graph convolutional network; s35. To classify whether a location is a pick-up point, use the Softmax function as the activation function. The formula is as follows: in is the output of the previous layer and the input of the Softmax function, y λ is the probability that the λth predicted object belongs to the pick-up point, the drop-off point, both, or neither; And using cross entropy as the loss function, the formula is as follows: where t λ is the true value of the λth predicted object; s36. Broadcast all the pick-up point information obtained in s35 to the taxis in the area to complete the recommendation of pick-up points under cloud-edge collaboration.

2. A system for implementing the taxi pick-up point recommendation method based on a spatiotemporal directed graph convolutional network according to claim 1, characterized in that: It includes a trajectory space-time graph identification module, a calculation task allocation module, a space-time graph feature extraction module and a passenger point recommendation module connected in sequence; The trajectory spatiotemporal graph identification module replaces the traditional representation method with a directed graph-based representation. The boarding or alighting points containing characteristic attributes, i.e., points of interest, are used as nodes, and directed edges are used to indicate the connection between the boarding point and the alighting point. The computing task allocation module allocates computing tasks based on the analysis and calculation results of the trajectory space-time graph identification module and the representation of the trajectory space-time graph; The spatiotemporal graph feature extraction and pick-up point recommendation module uses a spatiotemporal directed graph convolutional network to extract spatiotemporal graph features and recommend pick-up points, including: input layer, STDCN layer, Dropout layer, Embedding layer, Softmax layer and output layer.

Images