Pick-up point recommendation method, apparatus, device, and product

Abstract

This application discloses a method, apparatus, device, and product for recommending boarding points, relating to the field of intelligent transportation technology. The method includes: receiving a passenger's ride request; parsing the ride request to obtain the user's real-time location and real-time context; querying a pre-built candidate boarding anchor point library using the user's real-time location and real-time context to obtain candidate anchor point nodes and their corresponding high-frequency and low-frequency input features; extracting features from the high-frequency and low-frequency input features using a pre-built dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and fusing the high-frequency disturbance features and low-frequency regularity features to obtain fused features; and performing boarding point recommendation analysis based on the fused features to obtain a list of recommended boarding points. This improves the efficiency of boarding point recommendation.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and in particular to a method, apparatus, equipment, and product for recommending boarding points. Background Technology

[0002] With the popularization of ride-sharing, ride-hailing pick-up point recommendation has become a key link in improving user experience and system efficiency. Existing pick-up point recommendation technologies mostly focus on driver-side scheduling optimization. By analyzing historical order data, road network topology, and real-time supply and demand status, they plan pick-up routes or recommend hot spots for drivers. Mainstream methods include personalized recommendation models based on collaborative filtering (such as MPLRec), spatiotemporal prediction models based on graph neural networks, and dynamic scheduling strategies based on reinforcement learning. These technologies usually rely on a large amount of historical user trajectory data and abstract the road network into a static graph structure, using indicators such as distance or historical average speed to evaluate pick-up point availability.

[0003] However, the aforementioned technologies have significant shortcomings in practical applications. First, existing methods heavily rely on users' historical behavior data. For users with low-frequency travel or those entering unfamiliar areas in a cold start scenario, the lack of local trajectories makes it impossible to generate accurate recommendations. Furthermore, centralized models struggle to extract general patterns from massive amounts of heterogeneous data to supplement sparse data. Second, spatial data from different cities and functional zones (such as CBDs and residential areas) exhibit non-independent and identically distributed characteristics, making it difficult for a unified algorithm framework to adapt to significant regional differences. Third, existing technologies focus on learning long-term historical patterns, resulting in a slow response to high-frequency impulses caused by sudden accidents or extreme weather, leading to recommended locations being unreachable in the actual environment or significant deviations in estimated pick-up times. Finally, static road network modeling ignores the dynamic changes in real-time road capacity, severely compromising the stability and practicality of the recommendation results.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device, equipment, and product for recommending pick-up points, aiming to solve the technical problem that accurate pick-up point recommendation cannot be achieved in the ride-hailing scenario due to spatial heterogeneity and dynamic environmental interference.

[0006] To achieve the above objectives, this application proposes a boarding point recommendation method, which includes: Receive a passenger's ride request, parse the ride request to obtain the user's real-time location and real-time context; By querying the pre-built candidate boarding anchor point library using the user's real-time location and real-time context, candidate anchor point nodes and their corresponding high-frequency and low-frequency input features are obtained. The high-frequency and low-frequency input features are extracted by a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features. The high-frequency disturbance features and low-frequency regularity features are then fused to obtain fused features. Based on the fusion features, a boarding point recommendation analysis is performed to obtain a list of recommended boarding points.

[0007] In one embodiment, before the step of querying a pre-built candidate boarding anchor point database using the user's real-time location and real-time context to obtain high-frequency and low-frequency input features, the method further includes: Acquire historical order data, GPS trajectory data, Point of Interest (POI) density data, and road network topology data; The urban core area is spatially gridded using the road network topology data to obtain several candidate boarding anchor nodes, and a spatial index is constructed based on the coordinates of the several candidate boarding anchor nodes. The POI density data is used to encode the candidate boarding anchor nodes to obtain POI semantic features. By statistically analyzing the historical order data, the order demand of the candidate boarding anchor nodes within the target time window is obtained, thus acquiring the historical demand pattern information of the candidate boarding anchor nodes. By statistically analyzing the vehicle transfer records between several candidate boarding anchor points using the GPS trajectory data, the traffic transfer pattern information of the several candidate boarding anchor points is obtained. The basic time codes of the candidate boarding anchor nodes are generated according to the time period division rules; The basic time encoding, POI semantic features, and historical demand pattern information are concatenated to obtain low-frequency flow basic features. The historical speed dynamics in the GPS trajectory data and the flow transfer pattern information are concatenated to obtain high-frequency flow basic features. Based on the anchor node identifier, a candidate boarding anchor library is constructed using the spatial index, low-frequency flow basic features, and high-frequency flow basic features.

[0008] In one embodiment, the step of querying a pre-built candidate boarding anchor point library through the user's real-time location and real-time context to obtain candidate anchor point nodes and the high-frequency and low-frequency input features corresponding to the candidate anchor points includes: Based on the user's real-time location, a node search is performed using the spatial index in the candidate boarding anchor point library to obtain candidate anchor point nodes; The candidate anchor point nodes are used to extract the target low-frequency flow basic features and the target high-frequency flow basic features from the candidate boarding anchor point library. Real-time time codes are generated using real-time information from the real-time context, and the base time codes in the target low-frequency stream basic features are replaced with the real-time time codes to obtain low-frequency stream input features. The high-frequency stream input features are obtained by dynamically replacing the historical velocity dynamics in the target high-frequency stream basic features with the real-time velocity dynamics in the real-time context.

[0009] In one embodiment, before the step of extracting features from the high-frequency stream input features and low-frequency stream input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency perturbation features and low-frequency regularity features, the method further includes: A low-frequency flow extraction submodule is constructed based on Fast Fourier Transform and multi-scale convolution. The Fast Fourier Transform is used to identify the periodicity of the input low-frequency flow input features and reshape the identified time series into a two-dimensional matrix. The multi-scale convolution is used to extract the low-frequency regularity features of the two-dimensional matrix. A high-frequency flow extraction submodule is constructed based on discrete wavelet transform, which is used to decompose the input high-frequency flow input features to obtain high-frequency perturbation features; A dual-graph convolutional submodule is constructed based on the graph attention convolution of the graph attention network. The graph attention convolution is used to perform feature aggregation on the graph structure composed of low-frequency regular features and high-frequency perturbation features. A gated fusion submodule is constructed based on a feature linear modulation mechanism, which is used to dynamically adjust the fusion weights of the low-frequency and high-frequency features output by the dual-graph convolution submodule. By jointly training the low-frequency stream extraction submodule, high-frequency stream extraction submodule, dual-graph convolution submodule, and gated fusion submodule using historical sample data, a dual-stream spatiotemporal graph neural network is obtained.

[0010] In one embodiment, the step of extracting features from the high-frequency stream input features and low-frequency stream input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency perturbation features and low-frequency regularity features includes: The low-frequency stream extraction submodule performs periodic identification and multi-scale convolution on the low-frequency stream input features to obtain low-frequency regular features. The high-frequency stream input features are decomposed by the high-frequency stream extraction submodule to obtain high-frequency disturbance features; Based on the low-frequency regularity features and the geographical similarity, semantic similarity of interest points, synchronous similarity of demand, and temporal period similarity among the candidate anchor nodes, a low-frequency functional similarity map is constructed. The low-frequency functional similarity map is then aggregated using the dual-graph convolutional submodule to obtain low-frequency spatial aggregation features. Based on the high-frequency disturbance features, the flow transfer matrix between the candidate anchor nodes, and the physical reachability, a high-frequency flow graph is constructed. The high-frequency flow graph is then subjected to feature aggregation through the dual-graph convolution submodule to obtain high-frequency spatial aggregation features.

[0011] In one embodiment, the step of fusing the high-frequency disturbance features and the low-frequency regularity features to obtain the fused features includes: The high-frequency energy of the candidate anchor node is obtained by calculating the sum of the absolute values ​​of the high-frequency spatial aggregation features through the gated fusion submodule. The high-frequency energy is statistically analyzed and standardized by the gated fusion submodule, and the standardized high-frequency energy is mapped to a gated value by the Sigmoid function. The gated fusion submodule multiplies the gated value element-wise with the high-frequency spatial aggregation feature to obtain the multiplication result, and then adds the multiplication result to the low-frequency spatial aggregation feature to obtain the fused feature.

[0012] In one embodiment, the step of performing pick-up point recommendation analysis based on the fusion features to obtain a list of recommended pick-up points includes: The fused features are predicted using a multi-task prediction head to obtain the demand intensity, supply density, prediction speed, and expected waiting time of the candidate anchor node. The walking time for the candidate anchor nodes is calculated based on the user's real-time location, and the trip time and dynamic trip cost are calculated based on the predicted speed and destination distance. The estimated waiting time, walking time, dynamic trip cost, and trip time are weighted by multiple objectives to obtain the score result. The candidate anchor nodes are sorted according to the scoring results to obtain a list of recommended boarding points.

[0013] Furthermore, to achieve the above objectives, this application also proposes a boarding point recommendation device, which includes: The parsing module is used to receive passengers' ride requests and parse the ride requests to obtain the user's real-time location and real-time context; The query module is used to query a pre-built candidate boarding anchor point library based on the user's real-time location and real-time context, and obtain candidate anchor point nodes and the high-frequency and low-frequency input features corresponding to the candidate anchor points. The fusion module is used to extract features from the high-frequency stream input features and low-frequency stream input features through a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and to fuse the high-frequency disturbance features and low-frequency regularity features to obtain fused features. The analysis module is used to perform boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points.

[0014] In addition, to achieve the above objectives, this application also proposes a boarding point recommendation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the boarding point recommendation method as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the boarding point recommendation method as described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the boarding point recommendation method as described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: This application proposes a method, apparatus, device, and product for recommending boarding points. The method involves receiving a passenger's ride request, parsing the request to obtain the user's real-time location and context, querying a pre-built candidate boarding point library using the user's real-time location and context to obtain candidate anchor point nodes and their corresponding high-frequency and low-frequency input features, respectively, extracting features from the high-frequency and low-frequency input features using a pre-built dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, fusing these features to obtain fused features, and performing boarding point recommendation analysis based on the fused features to obtain a recommended boarding point list. Therefore, to address spatial heterogeneity, a candidate pick-up point library is pre-built and low-frequency flow input features are extracted. The low-frequency flow extraction submodule in the dual-stream spatiotemporal graph neural network is used to learn the long-term periodic patterns of different regions, enabling the model to adapt to the differentiated needs of functional zones such as CBDs and residential areas. To address dynamic environmental interference, high-frequency flow input features are extracted and the high-frequency flow extraction submodule is used to capture instantaneous pulses caused by sudden accidents and extreme weather. Then, a fusion mechanism is used to dynamically combine low-frequency patterns with high-frequency disturbances, so that the recommendation results can reflect both the inherent characteristics of the region and respond to environmental changes in real time, thereby achieving accurate pick-up point recommendations. This solves the problem of not being able to achieve accurate pick-up point recommendations in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference, and improves the efficiency of pick-up point recommendations. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for recommending boarding points in this application, as provided in Embodiment 1. Figure 2 This is a flowchart illustrating the second embodiment of the method for recommending boarding points in this application. Figure 3 A simplified flowchart illustrating the boarding point recommendation method provided in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the module structure of the vehicle boarding point recommendation device in an embodiment of this application; Figure 5This is a schematic diagram of the device structure of the hardware operating environment involved in the boarding point recommendation method in this application embodiment.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is as follows: Acquire historical order data, GPS trajectory data, Point of Interest (POI) density data, and road network topology data; divide the urban core area into spatial grids using the road network topology data to obtain several candidate boarding anchor nodes, and construct a spatial index based on the coordinates of the candidate boarding anchor nodes; encode the candidate boarding anchor nodes using the POI density data to obtain POI semantic features; and statistically analyze the order demand of the candidate boarding anchor nodes within the target time window using the historical order data to obtain historical demand pattern information of the candidate boarding anchor nodes. By statistically analyzing vehicle transfer records between several candidate boarding anchor nodes using the GPS trajectory data, traffic transfer pattern information of the candidate boarding anchor nodes is obtained. Basic time codes for the candidate boarding anchor nodes are generated according to time period segmentation rules. The basic time codes, POI semantic features, and historical demand pattern information are concatenated to obtain low-frequency flow basic features. Historical speed dynamics from the GPS trajectory data and the traffic transfer pattern information are concatenated to obtain high-frequency flow basic features. Based on the anchor node identifiers, a candidate boarding anchor library is constructed using the spatial index, low-frequency flow basic features, and high-frequency flow basic features. Based on the user's real-time location, a node search is performed using the spatial index in the candidate boarding anchor point library to obtain candidate anchor point nodes. The target low-frequency flow basic features and target high-frequency flow basic features are extracted from the candidate boarding anchor point library using these candidate anchor points. A real-time time code is generated using real-time information from the real-time context, and the basic time code in the target low-frequency flow basic features is replaced with the real-time time code to obtain low-frequency flow input features. The historical speed dynamics in the target high-frequency flow basic features are dynamically replaced with the real-time speed dynamics from the real-time context to obtain high-frequency flow input features. A low-frequency flow extraction submodule is constructed based on Fast Fourier Transform (FFT) and multi-scale convolution. The FFT is used to identify the periodicity of the input low-frequency flow features and reshape the identified time series into a two-dimensional matrix. The multi-scale convolution is used to extract the low-frequency regularity features of the two-dimensional matrix. A high-frequency flow extraction submodule is constructed based on Discrete Wavelet Transform (DWT). The DWT is used to decompose the input high-frequency flow features to obtain high-frequency perturbation features. A dual-graph convolution submodule is constructed based on graph attention convolution of a graph attention network. The graph attention convolution is used to aggregate features of the graph structure composed of low-frequency regularity features and high-frequency perturbation features. A gated fusion submodule is constructed based on a feature linear modulation mechanism. The feature linear modulation mechanism is used to dynamically adjust the fusion weights of the low-frequency and high-frequency features output by the dual-graph convolution submodule. The low-frequency flow extraction submodule, high-frequency flow extraction submodule, dual-graph convolution submodule, and gated fusion submodule are jointly trained using historical sample data to obtain a dual-stream spatiotemporal graph neural network.The low-frequency flow extraction submodule performs periodic identification and multi-scale convolution on the low-frequency flow input features to obtain low-frequency regularity features. The high-frequency flow extraction submodule performs signal decomposition on the high-frequency flow input features to obtain high-frequency disturbance features. Based on the low-frequency regularity features and the geographical similarity, semantic similarity of interest points, demand synchronization similarity, and time period similarity between the candidate anchor nodes, a low-frequency functional similarity map is constructed. The dual-graph convolution submodule then performs feature aggregation on the low-frequency functional similarity map to obtain low-frequency spatial aggregation features. Based on the high-frequency disturbance features and the flow transfer matrix and physical reachability between the candidate anchor nodes, a high-frequency flow map is constructed. The dual-graph convolution submodule then performs feature aggregation on the high-frequency flow map to obtain high-frequency spatial aggregation features. The gating fusion submodule calculates the sum of the absolute values ​​of the high-frequency spatial aggregation features to obtain the high-frequency energy of the candidate anchor nodes. The gating fusion submodule then performs sliding window statistics and standardization on the high-frequency energy, and maps the standardized high-frequency energy to a gating value using the Sigmoid function. The gating fusion submodule then multiplies the gating value element-wise with the high-frequency spatial aggregation features to obtain the multiplication result, and adds the multiplication result to the low-frequency spatial aggregation features to obtain the fused features. A multi-task prediction head predicts the fused features to obtain the demand intensity, supply density, prediction speed, and estimated waiting time of the candidate anchor nodes. The walking time of the candidate anchor nodes is calculated based on the user's real-time location, and the travel time and dynamic travel cost are calculated based on the predicted speed and destination distance. A multi-objective weighted score is applied to the estimated waiting time, walking time, dynamic travel cost, and travel time to obtain a score result. The candidate anchor nodes are then sorted according to the score results to obtain a recommended boarding point list. This invention solves the problem of inaccurate pick-up point recommendation in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference, achieving improved efficiency in pick-up point recommendation. Based on this invention, it addresses the issues of existing methods heavily relying on historical user behavior data. For low-frequency users or cold-start scenarios entering unfamiliar areas, the lack of local trajectories hinders accurate recommendations, and centralized models struggle to extract general patterns from massive heterogeneous data to supplement sparse data, resulting in low efficiency. Therefore, this invention designs a pick-up point recommendation method and verifies its effectiveness in recommending pick-up points. Finally, the efficiency of pick-up point recommendation using this invention is significantly improved.

[0025] In this embodiment, for ease of description, the loading point recommendation device will be used as the execution subject in the following description.

[0026] Due to the limitations of existing spatiotemporal data mining and recommendation algorithms, the accuracy of ride-hailing pick-up point recommendations needs improvement. One issue is spatial heterogeneity: different functional zones (such as CBDs and residential areas) exhibit significantly different pick-up demand patterns, making it difficult for a unified model to accommodate multiple spatial patterns, leading to decreased recommendation accuracy. Another issue is dynamic interference: users' real-time ride experience is affected by environmental factors such as sudden accidents and extreme weather. Without considering high-frequency dynamic changes, the availability and timeliness of recommended locations will significantly decrease. Furthermore, there is the issue of data sparsity. Accurate pick-up point recommendations require massive amounts of historical user trajectory data for training, but for low-frequency users or cold-start scenarios entering unfamiliar areas, insufficient data accumulation also affects the accuracy of pick-up point recommendations. Therefore, current ride-hailing recommendation systems also face challenges in environmental adaptability. Since different cities and regions have different road network structures and travel patterns, if the recommendation model does not perform targeted spatiotemporal feature modeling and dynamic fusion, recommendation efficiency will also decrease.

[0027] This application provides a solution that addresses spatial heterogeneity by pre-constructing a candidate pick-up point library and extracting low-frequency flow input features. It utilizes a low-frequency flow extraction submodule in a dual-stream spatiotemporal graph neural network to learn long-term periodic patterns in different regions, enabling the model to adapt to the differentiated needs of functional zones such as CBDs and residential areas. For dynamic environmental interference, it extracts high-frequency flow input features and captures instantaneous pulses caused by sudden accidents and extreme weather through a high-frequency flow extraction submodule. A fusion mechanism dynamically combines low-frequency patterns with high-frequency disturbances, ensuring that the recommendation results reflect both the inherent characteristics of the region and respond to environmental changes in real time. This achieves accurate pick-up point recommendations, solving the problem of inaccurate pick-up point recommendations in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference, and improving the efficiency of pick-up point recommendations.

[0028] Based on this, this application provides a method for recommending boarding points, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the vehicle boarding point recommendation method of this application.

[0029] In this embodiment, the boarding point recommendation method includes steps S01 to S04: Step S01: Receive the passenger's ride request, parse the ride request to obtain the user's real-time location and real-time context; Before the implementation of this embodiment, it should be clear that existing methods rely heavily on users' historical behavior data. For users with low-frequency travel or cold start scenarios entering unfamiliar areas, accurate recommendations cannot be generated due to the lack of local trajectories. Furthermore, centralized models struggle to extract general patterns from massive heterogeneous data to supplement sparse data. In addition, spatial data from different cities and functional zones (such as CBD and residential areas) exhibit non-independent and identically distributed characteristics, making it difficult for a unified algorithm framework to adapt to the significant differences between regions.

[0030] Therefore, in order to solve the above problems, this embodiment receives the passenger's car rental request and extracts the passenger's current latitude and longitude coordinates as the user's real-time location. At the same time, it extracts the time information of the request (including hour, day of the week, whether it is a holiday, etc.) and the current environmental status (such as weather conditions, traffic events, etc.) as real-time context, providing basic input for subsequent recommendation decisions.

[0031] Step S02: By querying the pre-built candidate boarding anchor point library through the user's real-time location and real-time context, candidate anchor point nodes and the high-frequency and low-frequency input features corresponding to the candidate anchor point nodes are obtained. Centered on the user's real-time location, the system utilizes a pre-generated spatial index (KDTree) in the candidate boarding anchor point library to quickly search for nearby candidate boarding anchor point nodes within a preset radius (e.g., 500 meters). It then extracts the low-frequency flow basic features and high-frequency flow basic features corresponding to these nodes from the library. Finally, it combines the time information and real-time speed dynamics in the real-time context to generate the final low-frequency flow input features and high-frequency flow input features.

[0032] Step S03: Extract features from the high-frequency stream input features and low-frequency stream input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and fuse the high-frequency disturbance features and low-frequency regularity features to obtain fused features; The low-frequency input features are fed into the low-frequency flow extraction submodule (based on TimesNet), which uses Fast Fourier Transform to identify periodic components and extract long-term periodic patterns to obtain low-frequency regularity features. Simultaneously, the high-frequency input features are fed into the high-frequency flow extraction submodule (based on wavelet transform), which uses Discrete Wavelet Transform to decompose the signal to capture instantaneous pulses and obtain high-frequency disturbance features. Subsequently, a dual-graph convolution submodule performs graph attention convolution on the low-frequency functional similarity graph and the high-frequency flow graph, respectively, to obtain low-frequency spatial aggregation features and high-frequency spatial aggregation features. Finally, a gated fusion submodule dynamically adjusts the fusion weights according to the fluctuation of high-frequency energy to generate fused features.

[0033] Step S04: Perform a boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points.

[0034] Finally, the fused features are input into the multi-task prediction head to predict the demand intensity, supply density, prediction speed, and expected waiting time of each candidate anchor node. Then, the walking time and dynamic travel cost are calculated by combining the user's real-time location. The candidate anchor nodes are comprehensively ranked by multi-objective weighted scoring, and the Top-K recommended boarding point list is finally output for passengers to choose from.

[0035] The above scheme utilizes the low-frequency flow extraction submodule in the dual-stream spatiotemporal graph neural network to learn the long-term periodic patterns of different regions, enabling the model to adapt to the differentiated needs of functional zones such as CBDs and residential areas. In response to dynamic environmental interference, high-frequency flow input features are extracted and instantaneous pulses caused by sudden accidents and extreme weather are captured through the high-frequency flow extraction submodule. Then, the low-frequency patterns and high-frequency disturbances are dynamically combined through a fusion mechanism, so that the recommendation results can reflect the inherent characteristics of the region and respond to environmental changes in real time, thereby achieving accurate boarding point recommendations.

[0036] Specifically, before step S02 above, which queries the pre-built candidate boarding anchor point database using the user's real-time location and real-time context to obtain high-frequency and low-frequency input features, the method further includes: Step S0201: Obtain historical order data, GPS trajectory data, point of interest (POI) density data, and road network topology data; Step S0202: The urban core area is spatially gridded using the road network topology data to obtain several candidate boarding anchor nodes, and a spatial index is constructed based on the coordinates of the several candidate boarding anchor nodes. Step S0203: Encode the candidate boarding anchor nodes with POI using the POI density data to obtain POI semantic features; Step S0204: By statistically analyzing the historical order data, the order demand of the candidate boarding anchor nodes within the target time window is obtained, thus acquiring the historical demand pattern information of the candidate boarding anchor nodes. Step S0205: By statistically analyzing the vehicle transfer records between several candidate boarding anchor nodes using the GPS trajectory data, the flow transfer pattern information of the several candidate boarding anchor nodes is obtained. Step S0206: Generate the basic time code of the candidate boarding anchor nodes according to the time period division rules; Step S0207: The basic time code, POI semantic features and historical demand pattern information are concatenated to obtain low-frequency flow basic features. The historical speed dynamics in the GPS trajectory data and the flow transfer pattern information are concatenated to obtain high-frequency flow basic features. Step S0208: Based on the anchor node identifier, a candidate boarding anchor library is constructed using the spatial index, low-frequency flow basic features, and high-frequency flow basic features.

[0037] During the construction phase of the candidate boarding anchor point library, the first step is to complete the compliant collection and de-identification of multi-source basic data. Specifically, historical order data, vehicle GPS trajectory data, urban POI density data, and full-domain road network topology data in urban ride-hailing scenarios are collected. Dirty data such as null values, abnormal drift trajectories, and invalid orders are removed from the data to complete the basic data cleaning and ensure the accuracy and effectiveness of subsequent modeling data.

[0038] After data preprocessing, the core urban travel area is uniformly divided into spatial grids based on standardized road network topology data. A 70×70 grid size is used to achieve full coverage, ultimately generating 4,900 independent candidate boarding anchor point nodes as the basic candidate units for subsequent boarding point recommendations. At the same time, a global node mapping table is constructed based on the latitude and longitude coordinates of all anchor point nodes, and the KDTree algorithm is introduced to build an efficient spatial index structure, which can achieve fast neighbor node lookup with O(logN) complexity, greatly improving the response efficiency of subsequent user location matching.

[0039] For massive POI density data, a two-layer MLP learnable encoder is built to extract semantic features. The input is a 20-dimensional original POI density vector, which is then subjected to a linear transformation to map the dimension from 20 to 32, a ReLU activation function for non-linear processing, a quadratic linear transformation to compress to 16 dimensions, and a LayerNorm layer for normalization to stabilize the values. The specific calculation formula is as follows:

[0040] The final output is a POI semantic feature with dimensions [N, 16]. It accurately depicts the differences in the regional attributes of commerce, residence, transportation, etc. around different anchor nodes, providing semantic support for subsequent spatial similarity calculation.

[0041] Based on anonymized historical order data, time-series demand statistics were conducted. Using 5-minute intervals as a single time granularity, 48 consecutive time windows were set to cover all travel periods throughout the day. The number of valid orders for each anchor node within each time window was counted, constructing a global demand matrix D with dimensions [N, T=48]. The formula for calculating the demand matrix is ​​as follows:

[0042] The demand matrix obtained above fully records the patterns of user travel demand in different regions at different times, providing core data support for calculating the demand synchronization similarity of low-frequency function graphs.

[0043] Subsequently, based on massive amounts of vehicle GPS trajectory data, the vehicle movement transfer records Count(i→j) between any two anchor nodes are statistically analyzed to quantify the frequency of traffic flow and the correlation between different nodes, constructing a full-domain node traffic transfer model that accurately reflects the dynamic transfer characteristics of real-time traffic flow on urban roads. Simultaneously, combined with the rules for dividing the day into time periods, distinguishing by hour, weekday, weekday / holiday, morning / evening peak hours, etc., standardized basic time codes are generated for all candidate anchor nodes to characterize temporal travel patterns.

[0044] After completing the multi-source feature extraction, the dual-stream basic feature concatenation operation is performed. The low-frequency stream basic feature is composed of three types of features: 16-dimensional POI static semantic features, 12-dimensional multi-dimensional time-encoding features, and 20-dimensional historical demand pattern features. After concatenation, a 48-dimensional low-frequency stream basic feature with a unified dimension is formed. The calculation formula is as follows:

[0045] The aforementioned spliced ​​low-frequency flow basic features fully preserve the inherent attributes and long-term travel patterns of the urban area. The high-frequency flow basic features are spliced ​​together from 16-dimensional GPS real-time speed dynamic features and 8-dimensional node flow transfer pattern features, forming 24-dimensional high-frequency flow basic features. The calculation formula is as follows:

[0046] Unlike the low-frequency flow fundamental characteristics mentioned above, the calculated 24-dimensional high-frequency flow fundamental focuses on capturing the instantaneous dynamic changes in road traffic.

[0047] Finally, by binding the unique identifier of each anchor node, integrating the spatial index structure, dual-stream basic features, POI semantic features, and temporal demand matrix, the construction of the full-domain candidate boarding anchor point library is completed, realizing the standardized and structured storage of urban boarding point candidate resources.

[0048] More specifically, step S02 above, which involves querying a pre-built candidate boarding anchor point library using the user's real-time location and real-time context to obtain candidate anchor point nodes and the corresponding high-frequency and low-frequency input features of the candidate anchor points, includes: Step S021: Based on the user's real-time location, perform a node search using the spatial index in the candidate boarding anchor point library to obtain candidate anchor point nodes; Step S022: Extract the target low-frequency flow basic features and the target high-frequency flow basic features from the candidate boarding anchor point library through the candidate anchor point nodes; Step S023: Generate a real-time time code using the real-time information in the real-time context, and replace the basic time code in the target low-frequency stream basic features with the real-time time code to obtain the low-frequency stream input features; Step S024: Replace the historical velocity dynamics in the target high-frequency flow basic features with the real-time velocity dynamics in the real-time context to obtain the high-frequency flow input features.

[0049] After a user initiates a ride-hailing request, the system immediately parses and obtains the user's real-time geographic location coordinates. Relying on the KDTree spatial index structure built into the candidate pick-up anchor point library, the system quickly retrieves all nearby candidate pick-up anchor point nodes within the range with the user's real-time location as the center and 500 meters as the effective recommendation radius. It then filters out a set of effective candidate nodes that meet the distance requirements, avoiding the filtering calculation of invalid pick-up points at long distances, thus improving recommendation accuracy and computational efficiency.

[0050] After identifying valid candidate anchor nodes, the pre-stored low-frequency and high-frequency basic features corresponding to each node are accurately retrieved from the pre-built candidate boarding anchor point library. This completes the batch extraction of basic features, ensuring the rapid retrieval and reuse of feature data and reducing real-time computation.

[0051] By combining the real-time context information of the user's car request, the system extracts real-time time series data such as the specific hour, day of the week, time period, and holiday status of the request, generates a brand-new real-time time code, replaces the static basic time code in the original low-frequency stream basic features, completes the real-time update of the time series features, adapts to the user's current travel time scenario, and ensures the timeliness of low-frequency regular features.

[0052] Simultaneously extract real-time GPS speed statistics and real-time traffic flow fluctuations of current urban roads, replace historical static speed data in the high-frequency flow basic features with real-time dynamic speed data, complete real-time iteration of high-frequency dynamic features, and finally generate low-frequency flow input features and high-frequency flow input features that are adapted to the user's real-time location, real-time time period, and real-time traffic conditions, providing accurate real-time input data for subsequent dual-flow spatiotemporal feature extraction.

[0053] Further, step S04 above, which involves performing a boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points, includes: Step S041: The fused features are predicted using a multi-task prediction head to obtain the demand intensity, supply density, prediction speed, and expected waiting time of the candidate anchor node. Step S042: Calculate the walking time of the candidate anchor node based on the user's real-time location, and calculate the trip time and dynamic trip cost based on the predicted speed and destination distance; Step S043: Perform a multi-objective weighted score on the estimated waiting time, walking time, dynamic trip cost, and trip time to obtain the score result; Step S044: Sort the candidate anchor nodes according to the scoring results to obtain a list of recommended boarding points.

[0054] Before inputting the multi-task prediction head, the fused features are first projected along the time dimension, compressing and mapping the features of the historical 48 time steps into the predicted features of the next 12 time steps (corresponding to the next 60 minutes). This is accomplished through dimensional permutation and linear transformation, and the specific calculation formula is as follows:

[0055] Thus output It serves as the common input for the four prediction branches.

[0056] The prediction head contains four sets of prediction branches with consistent structure. Each branch employs a hierarchical structure combining Linear (48→64) dimensional transformation, ReLU nonlinear activation, and Linear (64→1) dimensional compression with a dedicated activation function. Combined with output scaling parameters of s_d=10.0, s_s=10.0, s_v=100.0, and s_e=30.0, it outputs four types of prediction results: λ_d (demand intensity), λ_s (supply density), V (prediction velocity), and T_wait (estimated waiting time). Demand intensity:

[0057] Supply density:

[0058] Predicted speed:

[0059] Estimated waiting time:

[0060] All four prediction results have dimensions of [B, 12, N, 1]. The value of the last prediction time step in the future is selected as the effective scoring indicator to accurately predict the travel service status of each boarding point in the next 60 minutes.

[0061] Based on the user's real-time location and the coordinates of each candidate anchor point node, walking time is accurately calculated. A path correction factor is introduced to compensate for the deviation between the straight-line distance and the actual walking path. The specific calculation formula is as follows:

[0062] The actual time it takes for the user to walk to the pick-up point is obtained.

[0063] Simultaneously, the travel time is estimated by combining the predicted road speed V with the distance to the user's destination. With mileage And calculate the dynamic trip cost, the billing formula is:

[0064] Where BaseFee=10 yuan, PerMinFee=0.5 yuan / minute, PerKmFee=2.5 yuan / km, and T_trip and D_trip are calculated from the predicted speed V and the distance to the destination.

[0065] The system integrates four core metrics: estimated waiting time, user walking time, dynamic trip cost, and vehicle trip time. Weights are set at w=[0.5, 0.2, 0.2, 0.1]. Normalized baselines of s_wait=10min, s_walk=10min, s_price=20 yuan, and s_trip=20min are used to calculate the comprehensive score. The calculation formula is as follows:

[0066] The lower the score obtained above, the better the overall travel experience, service efficiency, and cost-effectiveness of the pick-up point.

[0067] Based on the comprehensive score ranking rule from smallest to largest, all candidate pick-up anchor points within a 500-meter range are prioritized and the top five optimal nodes are selected to generate a Top-5 recommended pick-up point list. The list also includes complete information such as the precise coordinates of each recommended anchor point, predicted waiting time, walking distance, and estimated trip cost, providing users with diversified and accurate pick-up point selection options.

[0068] This embodiment, through the above-described scheme, specifically receives a passenger's ride request, parses the request to obtain the user's real-time location and real-time context; queries a pre-built candidate boarding anchor point library using the user's real-time location and real-time context to obtain candidate anchor point nodes and their corresponding high-frequency and low-frequency input features; extracts features from the high-frequency and low-frequency input features using a pre-built dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and fuses these features to obtain fused features; performs boarding point recommendation analysis based on the fused features to obtain a recommended boarding point list.

[0069] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Before step S03, which involves extracting features from the high-frequency and low-frequency input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, the boarding point recommendation method further includes steps S0301 to S0305: Step S0301: A low-frequency flow extraction submodule is constructed based on Fast Fourier Transform and multi-scale convolution. The Fast Fourier Transform is used to perform periodic identification on the input low-frequency flow input features and reshape the identified time series into a two-dimensional matrix. The multi-scale convolution is used to extract the low-frequency regularity features of the two-dimensional matrix. Step S0302: A high-frequency flow extraction submodule is constructed based on discrete wavelet transform. The discrete wavelet transform is used to decompose the input high-frequency flow input features to obtain high-frequency perturbation features. Step S0303: A dual-graph convolution submodule is constructed based on the graph attention convolution of the graph attention network. The graph attention convolution is used to perform feature aggregation on the graph structure composed of low-frequency regular features and high-frequency perturbation features. Step S0304: A gated fusion submodule is constructed based on the feature linear modulation mechanism, wherein the feature linear modulation mechanism is used to dynamically adjust the fusion weights of the low-frequency features and high-frequency features output by the dual-graph convolution submodule. Step S0305: The low-frequency stream extraction submodule, high-frequency stream extraction submodule, dual-graph convolution submodule, and gated fusion submodule are jointly trained using historical sample data to obtain a dual-stream spatiotemporal graph neural network.

[0070] The low-frequency flow extraction submodule is built around the core concept of mining time-series periodic patterns. It utilizes Fast Fourier Transform (FFT) to perform frequency domain analysis on the input low-frequency time-series feature sequences, automatically identifying periodic components in the time-series data and selecting the top k frequency parameters with the highest intensity. This is achieved through the following formula: The core period length of the time series data is calculated, and then the one-dimensional time series data is reshaped according to the calculated period:

[0071] The two-dimensional feature matrix of the structure enables the structured reconstruction of temporal features.

[0072] Subsequently, 3×3, 5×5, and 7×7 multi-scale Inception convolutional kernels were introduced to extract features and capture long-term urban travel patterns at different periodic scales. The multi-scale convolutional output features were then concatenated and fused, and uniformly mapped to 48-dimensional low-frequency features via linear projection. The specific calculation formula is as follows:

[0073] The above calculation formula is used to complete the functional construction of the low-frequency flow extraction submodule, accurately mining the stable cycle characteristics of urban travel.

[0074] The high-frequency stream extraction submodule is built on Discrete Wavelet Transform (DWT) and focuses on capturing the instantaneous abrupt changes in road traffic characteristics. For the input high-frequency dynamic feature signal, it performs multi-level signal decomposition using learnable CNN operators P and U, extracting the detail and approximate components of the signal respectively. The decomposition process is as follows:

[0075]

[0076] P and U are learnable CNN operators. The multi-level decomposition operation is repeated three times to fully explore the short-term traffic flow pulse disturbance features caused by extreme weather, traffic accidents, temporary traffic control and other scenarios. Finally, 48-dimensional multi-scale high-frequency disturbance features are output to achieve high sensitivity perception of instantaneous changes in road conditions and make up for the shortcomings of traditional models that only learn long-term patterns and ignore dynamic disturbances.

[0077] The dual-graph convolutional submodule uses the dynamic graph attention network GATv2 as its core architecture, adapting to the dual-graph modeling requirements of this embodiment. It supports independent feature clustering of static functional similarity graphs and dynamic flow graphs, and the module has a built-in learnable feature transformation matrix. The system employs a 96-dimensional learnable attention vector 'a', with a fixed configuration of 2 heads, 48 ​​input / output channels, 1 edge dimension, and shared parameters. Residual connections and a LayerNorm normalization structure ensure gradient stability during deep network propagation. This allows for accurate modeling of nonlinear spatial relationships between anchor nodes, adapting to the feature learning needs of complex urban road networks. Building upon dual-stream feature extraction and spatial graph convolution, the model incorporates a prototype embedding layer to address data sparsity and cold-start issues. The system pre-maintains a set of feature prototypes representing typical urban functional patterns, each corresponding to a standardized regional travel feature (e.g., CBD business mode, residential commuting mode, transportation hub mode). When facing sparse data in unfamiliar areas or new user cold-start scenarios, the system performs similarity matching between the current anchor point's POI semantic features and demand pattern features and the prototype library. The prototype feature with the highest matching degree is embedded as a completion feature into the node representation, achieving cross-regional knowledge transfer. Even with extremely scarce historical order data in a specific region, feature completion can be achieved by associating with general prototype features, ensuring recommendation accuracy.

[0078] The gated fusion submodule is built on the FiLM feature linear modulation mechanism. It can autonomously sense the ambient wave intensity and dynamically adjust the dual-stream feature fusion weights. The module supports fully automated operation of high-frequency energy calculation, sliding window statistics, Z-score normalization, Sigmoid gated mapping, and feature weighted fusion. A standard deviation lower bound constraint of 0.15 is set to effectively avoid the fusion instability caused by numerical fluctuations, and realize adaptive feature fusion under stable and sudden road conditions, taking into account both model stability and real-time performance.

[0079] After completing the construction of all sub-modules, a training dataset is constructed by collecting massive amounts of real urban ride-hailing travel sample data. End-to-end joint training is performed on the low-frequency stream extraction sub-module, high-frequency stream extraction sub-module, dual-graph convolution sub-module, and gated fusion sub-module. The learnable parameters of each module are iteratively optimized through backpropagation, and the correlation between urban spatial heterogeneous features, temporal periodic patterns, and dynamic road condition disturbances is continuously fitted. Finally, a high-precision and highly robust dual-stream spatiotemporal graph neural network model is trained to adapt to the pick-up point recommendation needs of various urban travel scenarios.

[0080] Specifically, step S03 above, which involves extracting features from the high-frequency and low-frequency input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, includes: Step S031: The low-frequency stream input features are periodically identified and multi-scale convolutionally processed by the low-frequency stream extraction submodule to obtain low-frequency regular features; Step S032: The high-frequency stream input features are decomposed by the high-frequency stream extraction submodule to obtain high-frequency disturbance features; Step S033: Based on the low-frequency regularity features and the geographical similarity, semantic similarity of interest points, synchronous similarity of demand, and temporal period similarity between the candidate anchor nodes, a low-frequency functional similarity map is constructed, and the low-frequency functional similarity map is aggregated through the dual-graph convolution submodule to obtain low-frequency spatial aggregation features. Step S034: Based on the high-frequency disturbance features, the flow transfer matrix between the candidate anchor nodes, and the physical reachability, a high-frequency flow graph is constructed, and the high-frequency flow graph is subjected to feature aggregation through the dual-graph convolution submodule to obtain high-frequency spatial aggregation features.

[0081] After completing the feature extraction of the dual-flow submodule, a low-frequency functional similarity map is constructed based on the output of the trained low-frequency flow extraction submodule and the multi-dimensional similarity features of urban anchor nodes. This map uses the inherent functional attributes of nodes as the core association basis and integrates four types of core similarity indicators to complete the edge weight calculation, including: Geographical similarity:

[0082] in, Let be the geographical distance between node i and node j.

[0083] POI semantic similarity is calculated based on the cosine similarity of the node POI encoding vectors. The calculation formula is as follows:

[0084] Demand synchronization similarity is calculated based on the correlation of historical demand sequences between nodes. The calculation formula is as follows:

[0085] Temporal similarity is calculated based on the overlap of peak periods of nodes, using the following formula:

[0086] The formula for weighted fusion of the four types of similarity is:

[0087] Where α_geo=0.3, α_poi=0.4, α_dem=0.2, α_tt=0.1, the sum of the four terms is 1, and W_LF(i,j) is a scalar representing the functional association strength between i and j.

[0088] The constructed low-frequency functional similarity map is input into the dual-graph convolutional submodule, and node feature aggregation is completed through a parameter-sharing GATv2 attention network. First, the attention coefficients between nodes are calculated using the following formula:

[0089] in , For node feature vectors (taken during low-frequency propagation) High-frequency propagation ), For the edge weights of the corresponding graph, This is a vector concatenation operation. The learnable feature transformation matrix, This is a learnable attention vector.

[0090] Normalizing the attention coefficients yields the neighborhood aggregation weights:

[0091] in, Let i be the set of neighboring nodes of node i.

[0092] Finally, the output features are obtained through weighted aggregation and residual connection, and the calculation formula is as follows:

[0093] Perform the above operation on the low-frequency features to obtain the low-frequency spatial aggregated features h_lf_gnn with dimensions [N,48]. Perform the same operation on the high-frequency features to obtain the high-frequency spatial aggregated features h_hf_gnn with dimensions [N,48].

[0094] The system synchronously extracts instantaneous disturbance features from the high-frequency flow extraction submodule and constructs a high-frequency flow map by combining it with the dynamic traffic characteristics of the road network. The core of this construction relies on two types of dynamic indicators: The traffic transfer matrix counts the number of times a vehicle travels from node i to node j in the GPS trajectory. The calculation formula is as follows:

[0095] Physical reachability, defined as the average travel time from node i to node j, is calculated using the following formula:

[0096] Weights are assigned to the two types of indicators respectively. =0.7、 =0.3, weighted fusion yields high-frequency graph edge weights. The calculation formula is:

[0097] By performing attention convolution operations on the high-frequency flow graph through the dual-graph convolution submodule, the dynamic spatial correlation of nodes under sudden road conditions is captured, and dynamic feature aggregation is completed. The output high-frequency spatial aggregation features with dimensions [N, 48] are accurately adapted to the feature modeling requirements of instantaneous road condition fluctuation scenarios. The dual-graph independent modeling mode completely realizes the decoupling learning of the static functional rules of urban road networks and dynamic traffic flow changes, and greatly improves the recommendation robustness of the model under different time periods and different road conditions.

[0098] More specifically, step S03 above, which involves fusing the high-frequency disturbance features and the low-frequency regularity features to obtain the fused features, includes: Step S035: Calculate the sum of the absolute values ​​of the high-frequency spatial aggregation features using the gated fusion submodule to obtain the high-frequency energy of the candidate anchor node; Step S036: The high-frequency energy is statistically analyzed and standardized by the gated fusion submodule, and the standardized high-frequency energy is mapped to a gated value by the Sigmoid function. Step S037: The gated value is multiplied element-wise with the high-frequency spatial aggregation feature by the gated fusion submodule to obtain the multiplication result, and the multiplication result is added to the low-frequency spatial aggregation feature to obtain the fused feature.

[0099] After completing the aggregation of low-frequency and high-frequency spatial features, the FiLM gated fusion submodule is started to complete adaptive feature fusion. First, L1 energy calculation is performed on the high-frequency spatial aggregation features of each anchor node and each time dimension to extract the high-frequency feature fluctuation energy. :

[0100] The intensity of disturbances in the current road traffic environment is quantified; the higher the energy value, the more significant the sudden fluctuations in road conditions.

[0101] Subsequently, a sliding window statistical analysis with a window size of 2 was performed on the high-frequency energy sequence to calculate the local mean in real time. with standard deviation :

[0102] Simultaneously set a lower bound for the standard deviation. This will prevent numerical instability caused by extreme data and ensure the stability of statistical results.

[0103] Z-score standardization is performed based on the statistically obtained mean and standard deviation. Learnable parameters a=1.0 and b=-1.0 are then mapped to gate values ​​using the following formula:

[0104] in, For the gate value, expand to Align with the feature dimension to achieve precise control per feature dimension.

[0105] Under normal, stable road conditions, environmental fluctuation values ​​are low, and the threshold value is close to 0. The model focuses on low-frequency spatial aggregation features and relies on long-term urban travel patterns to make recommendations. In sudden scenarios such as rainstorms, accidents, and traffic jams, environmental fluctuation values ​​surge, and the threshold value approaches 1. The model significantly increases the weight of high-frequency disturbance features and corrects the recommendation results in real time.

[0106] Finally, the FiLM feature fusion operation is performed, using low-frequency spatial aggregation features as the base and high-frequency features as the superposition after gated modulation. The calculation formula is as follows:

[0107] Generate final fusion features that take into account both long-term patterns and real-time dynamics. This enables adaptive feature modulation in all weather and all scenarios, completely solving the technical shortcomings of traditional fixed-weight fusion that cannot adapt to dynamic road conditions.

[0108] This embodiment, through the above scheme, specifically constructs a low-frequency flow extraction submodule based on Fast Fourier Transform (FFT) and multi-scale convolution. The FFT is used to perform periodic identification on the input low-frequency flow features and reshape the identified time series into a two-dimensional matrix. The multi-scale convolution is used to extract the low-frequency regularity features of the two-dimensional matrix. A high-frequency flow extraction submodule is constructed based on Discrete Wavelet Transform (DWT). The DWT is used to decompose the input high-frequency flow features to obtain high-frequency perturbation features. A dual-graph convolution submodule is constructed based on graph attention convolution of a graph attention network. The graph attention convolution is used to perform feature aggregation on the graph structure composed of low-frequency regularity features and high-frequency perturbation features. A gated fusion submodule is constructed based on a feature linear modulation mechanism. The feature linear modulation mechanism is used to dynamically adjust the fusion weights of the low-frequency and high-frequency features output by the dual-graph convolution submodule. The low-frequency flow extraction submodule, high-frequency flow extraction submodule, dual-graph convolution submodule, and gated fusion submodule are jointly trained using historical sample data to obtain a dual-stream spatiotemporal graph neural network.

[0109] For example, to help understand the implementation process of the boarding point recommendation method obtained in this embodiment combined with the above embodiment one, please refer to... Figure 3 , Figure 3 A simplified flowchart of a boarding point recommendation method is provided, specifically: This embodiment relies on five core modules working in a step-by-step, closed-loop iteration to complete the process. It covers the entire chain of data preprocessing, dual-stream feature decoupling and encoding, dual-graph space modeling, adaptive feature fusion, and intelligent decision recommendation, and can fully cover the needs for accurate pick-up point recommendation in various urban ride-hailing travel scenarios.

[0110] First, the spatiotemporal data acquisition and preprocessing module is activated. Corresponding to the data input layer, it batch collects multi-source data such as anonymized orders, GPS trajectories, POI density, and road network topology across the entire city. This process includes data cleaning, 70×70 grid anchor point construction, KDTree spatial index building, POI semantic encoding, and time-series demand matrix statistics. Standardized 48-dimensional low-frequency input features and 24-dimensional high-frequency input features are generated respectively, and a structured candidate vehicle-mounting anchor point library is constructed to provide a standardized data foundation for subsequent model operations. At the same time, multi-dimensional data coupling modeling solves the problem of traditional models having a single data dimension and being unable to adapt to spatial heterogeneity.

[0111] After data preprocessing, the data enters the dual-stream spatiotemporal feature encoding module. The module completes feature decoupling learning through a parallel dual-path structure. The low-frequency stream relies on the TimesNet architecture and mines long-term stable urban travel patterns through FFT periodicity recognition, two-dimensional temporal reshaping, and multi-scale Inception convolution. The high-frequency stream relies on discrete wavelet multi-level decomposition to capture instantaneous traffic pulse disturbances. This achieves the separation and extraction of long-term low-frequency patterns and short-term high-frequency disturbances, completely breaking the technical limitations of traditional single-stream feature extraction and improving the model's dynamic environment perception capabilities.

[0112] After feature encoding is completed, the spatial dependency modeling module is connected to execute the innovative dual-graph GATv2 modeling logic. Low-frequency functional similarity graphs are constructed based on four-dimensional similarity, and high-frequency flow graphs are constructed based on traffic flow transfer and road accessibility. The dual-graph features are independently aggregated through a parameter-sharing dynamic graph attention network. At the same time, typical urban functional patterns are aggregated and matched through a prototype embedding layer. Cross-regional knowledge transfer is used to make up for the lack of features in local data sparsity and cold start scenarios. Residual connectivity and layer normalization are combined to ensure model stability, accurately characterize the static functional associations and dynamic traffic associations of the urban road network, and improve the model's generalization ability.

[0113] Then, the module enters the dynamic feature fusion module. Through high-frequency energy sensing, sliding window statistics, and adaptive gating generation, it dynamically adjusts the dual-stream feature fusion weights. Based on the real-time road condition fluctuation state, it autonomously balances the proportion of long-term patterns and real-time disturbances, outputs the globally optimal fused features, realizes full-scene adaptive feature modulation, and ensures the accuracy of feature expression in complex environments.

[0114] Finally, the results are output through the boarding point evaluation and recommendation decision module. The fused features are mapped to the future prediction period through time projection. Based on the multi-task prediction head, core business indicators such as supply and demand, speed, and waiting time are output. Multi-objective weighted scoring is completed by combining walking distance and dynamic billing rules. Finally, the top-5 optimal boarding points within 500 meters are selected and output as a list.

[0115] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the boarding point recommendation method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0116] This application also provides a boarding point recommendation device, please refer to... Figure 4 The boarding point recommendation device includes: The parsing module 10 is used to receive a passenger's car rental request and parse the car rental request to obtain the user's real-time location and real-time context; The query module 20 is used to query a pre-built candidate boarding anchor point library through the user's real-time location and real-time context to obtain candidate anchor point nodes and the high-frequency stream input features and low-frequency stream input features corresponding to the candidate anchor point nodes. The fusion module 30 is used to extract features from the high-frequency stream input features and the low-frequency stream input features through a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and to fuse the high-frequency disturbance features and low-frequency regularity features to obtain fused features. Analysis module 40 is used to perform boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points.

[0117] The pick-up point recommendation device provided in this application, employing the pick-up point recommendation method in the above embodiments, can solve the technical problem that accurate pick-up point recommendation cannot be achieved in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference. Compared with the prior art, the beneficial effects of the pick-up point recommendation device provided in this application are the same as those of the pick-up point recommendation method provided in the above embodiments, and other technical features in the pick-up point recommendation device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0118] This application provides a boarding point recommendation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the boarding point recommendation method in Embodiment 1 above.

[0119] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing a pick-up point recommendation device according to embodiments of this application. The pick-up point recommendation device in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The boarding point recommendation device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0120] like Figure 5As shown, the boarding point recommendation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the boarding point recommendation device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the boarding point recommendation device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows boarding point recommendation devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0121] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0122] The pick-up point recommendation device provided in this application, employing the pick-up point recommendation method described in the above embodiments, can solve the technical problem of inaccurate pick-up point recommendation in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference. Compared with the prior art, the beneficial effects of the pick-up point recommendation device provided in this application are the same as those of the pick-up point recommendation method provided in the above embodiments, and other technical features of this pick-up point recommendation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0123] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0125] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the boarding point recommendation method in the above embodiments.

[0126] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0127] The aforementioned computer-readable storage medium may be included in the boarding point recommendation device; or it may exist independently and not be assembled into the boarding point recommendation device.

[0128] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the pick-up point recommendation device, the pick-up point recommendation device performs the following actions: receives a passenger's ride request; parses the ride request to obtain the user's real-time location and real-time context; queries a pre-built candidate pick-up anchor point library using the user's real-time location and real-time context to obtain candidate anchor point nodes and their corresponding high-frequency and low-frequency input features; extracts features from the high-frequency and low-frequency input features using a pre-built dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and fuses these features to obtain fused features; and performs pick-up point recommendation analysis based on the fused features to obtain a list of recommended pick-up points.

[0129] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0130] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0131] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0132] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described pick-up point recommendation method. This solves the technical problem that, in ride-hailing scenarios, accurate pick-up point recommendations cannot be achieved due to spatial heterogeneity and dynamic environmental interference. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the pick-up point recommendation method provided in the above embodiments, and will not be repeated here.

[0133] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the boarding point recommendation method as described above.

[0134] The computer program product provided in this application can solve the technical problem that accurate pick-up point recommendation cannot be achieved in ride-hailing scenarios due to spatial heterogeneity and dynamic environmental interference. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the pick-up point recommendation method provided in the above embodiments, and will not be repeated here.

[0135] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for recommending boarding points, characterized in that, The method for recommending boarding points includes: Receive a passenger's ride request, parse the ride request to obtain the user's real-time location and real-time context; By querying the pre-built candidate boarding anchor point library using the user's real-time location and real-time context, candidate anchor point nodes and their corresponding high-frequency and low-frequency input features are obtained. The high-frequency and low-frequency input features are extracted by a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features. The high-frequency disturbance features and low-frequency regularity features are then fused to obtain fused features. Based on the fusion features, a boarding point recommendation analysis is performed to obtain a list of recommended boarding points.

2. The boarding point recommendation method as described in claim 1, characterized in that, Before the step of querying a pre-built candidate boarding anchor point database using the user's real-time location and real-time context to obtain high-frequency and low-frequency input features, the method further includes: Acquire historical order data, GPS trajectory data, Point of Interest (POI) density data, and road network topology data; The urban core area is spatially gridded using the road network topology data to obtain several candidate boarding anchor nodes, and a spatial index is constructed based on the coordinates of the several candidate boarding anchor nodes. The POI density data is used to encode the candidate boarding anchor nodes to obtain POI semantic features. By statistically analyzing the historical order data, the order demand of the candidate boarding anchor nodes within the target time window is obtained, thus acquiring the historical demand pattern information of the candidate boarding anchor nodes. By statistically analyzing the vehicle transfer records between several candidate boarding anchor points using the GPS trajectory data, the traffic transfer pattern information of the several candidate boarding anchor points is obtained. The basic time codes of the candidate boarding anchor nodes are generated according to the time period division rules; The basic time encoding, POI semantic features, and historical demand pattern information are concatenated to obtain low-frequency flow basic features. The historical speed dynamics in the GPS trajectory data and the flow transfer pattern information are concatenated to obtain high-frequency flow basic features. Based on the anchor node identifier, a candidate boarding anchor library is constructed using the spatial index, low-frequency flow basic features, and high-frequency flow basic features.

3. The boarding point recommendation method as described in claim 1, characterized in that, The step of querying a pre-built candidate boarding anchor point library using the user's real-time location and real-time context to obtain candidate anchor point nodes and the high-frequency and low-frequency input features corresponding to the candidate anchor points includes: Based on the user's real-time location, a node search is performed using the spatial index in the candidate boarding anchor point library to obtain candidate anchor point nodes; The candidate anchor point nodes are used to extract the target low-frequency flow basic features and the target high-frequency flow basic features from the candidate boarding anchor point library. Real-time time codes are generated using real-time information from the real-time context, and the base time codes in the target low-frequency stream basic features are replaced with the real-time time codes to obtain low-frequency stream input features. The high-frequency stream input features are obtained by dynamically replacing the historical velocity dynamics in the target high-frequency stream basic features with the real-time velocity dynamics in the real-time context.

4. The boarding point recommendation method as described in claim 1, characterized in that, Before the step of extracting features from the high-frequency and low-frequency input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, the method further includes: A low-frequency flow extraction submodule is constructed based on Fast Fourier Transform and multi-scale convolution. The Fast Fourier Transform is used to identify the periodicity of the input low-frequency flow input features and reshape the identified time series into a two-dimensional matrix. The multi-scale convolution is used to extract the low-frequency regularity features of the two-dimensional matrix. A high-frequency flow extraction submodule is constructed based on discrete wavelet transform, which is used to decompose the input high-frequency flow input features to obtain high-frequency perturbation features; A dual-graph convolutional submodule is constructed based on the graph attention convolution of the graph attention network. The graph attention convolution is used to perform feature aggregation on the graph structure composed of low-frequency regular features and high-frequency perturbation features. A gated fusion submodule is constructed based on a feature linear modulation mechanism, which is used to dynamically adjust the fusion weights of the low-frequency and high-frequency features output by the dual-graph convolution submodule. By jointly training the low-frequency stream extraction submodule, high-frequency stream extraction submodule, dual-graph convolution submodule, and gated fusion submodule using historical sample data, a dual-stream spatiotemporal graph neural network is obtained.

5. The boarding point recommendation method as described in claim 4, characterized in that, The step of extracting features from the high-frequency and low-frequency input features using a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features includes: The low-frequency stream extraction submodule performs periodic identification and multi-scale convolution on the low-frequency stream input features to obtain low-frequency regular features. The high-frequency stream input features are decomposed by the high-frequency stream extraction submodule to obtain high-frequency disturbance features; Based on the low-frequency regularity features and the geographical similarity, semantic similarity of interest points, synchronous similarity of demand, and temporal period similarity among the candidate anchor nodes, a low-frequency functional similarity map is constructed. The low-frequency functional similarity map is then aggregated using the dual-graph convolutional submodule to obtain low-frequency spatial aggregation features. Based on the high-frequency disturbance features, the flow transfer matrix between the candidate anchor nodes, and the physical reachability, a high-frequency flow graph is constructed. The high-frequency flow graph is then subjected to feature aggregation through the dual-graph convolution submodule to obtain high-frequency spatial aggregation features.

6. The boarding point recommendation method as described in claim 5, characterized in that, The step of fusing the high-frequency disturbance features and the low-frequency regularity features to obtain the fused features includes: The high-frequency energy of the candidate anchor node is obtained by calculating the sum of the absolute values ​​of the high-frequency spatial aggregation features through the gated fusion submodule. The high-frequency energy is statistically analyzed and standardized by the gated fusion submodule, and the standardized high-frequency energy is mapped to a gated value by the Sigmoid function. The gated fusion submodule multiplies the gated value element-wise with the high-frequency spatial aggregation feature to obtain the multiplication result, and then adds the multiplication result to the low-frequency spatial aggregation feature to obtain the fused feature.

7. The boarding point recommendation method as described in claim 1, characterized in that, The step of performing boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points includes: The fused features are predicted using a multi-task prediction head to obtain the demand intensity, supply density, prediction speed, and expected waiting time of the candidate anchor node. The walking time for the candidate anchor nodes is calculated based on the user's real-time location, and the trip time and dynamic trip cost are calculated based on the predicted speed and destination distance. The estimated waiting time, walking time, dynamic trip cost, and trip time are weighted by multiple objectives to obtain the score result. The candidate anchor nodes are sorted according to the scoring results to obtain a list of recommended boarding points.

8. A boarding point recommendation device, characterized in that, The boarding point recommendation device includes: The parsing module is used to receive passengers' ride requests and parse the ride requests to obtain the user's real-time location and real-time context; The query module is used to query a pre-built candidate boarding anchor point library based on the user's real-time location and real-time context, and obtain candidate anchor point nodes and the high-frequency and low-frequency input features corresponding to the candidate anchor points. The fusion module is used to extract features from the high-frequency stream input features and low-frequency stream input features through a pre-constructed dual-stream spatiotemporal graph neural network to obtain high-frequency disturbance features and low-frequency regularity features, and to fuse the high-frequency disturbance features and low-frequency regularity features to obtain fused features. The analysis module is used to perform boarding point recommendation analysis based on the fusion features to obtain a list of recommended boarding points.

9. A boarding point recommendation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the boarding point recommendation method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the boarding point recommendation method as described in any one of claims 1 to 7.

Images