Method and system for deducing origin-destination route and estimating travel time

Abstract

The invention relates to an origin-destination route deduction and travel time estimation method, which can deduce and generate a personalized travel route between origin and destination in combination with user preference on the premise of reducing data annotation quantity as much as possible by combining an antagonistic inverse enhancement network of an active learning mechanism. Furthermore, a plurality of weak learners fusing the graph convolutional network and the Transform network are integrated through an AdaBoost algorithm, and the graph convolutional network can obtain the high-dimensional spatial-temporal characteristics of the traffic network, thereby assisting the Transform network to better capture the long-term and short-term dependency relationship corresponding to the personalized driving route. In combination with learning results of a plurality of weak learners, a more abundant and comprehensive time sequence feature representation of a personalized driving route is obtained, finally, the travel time corresponding to the route is obtained by using the time sequence feature representation, and accurate origin-destination travel time estimation is realized.

Description

Technical Field

[0001] The present application relates to the field of traffic data prediction, and in particular to a method, system, computer device and computer-readable storage medium for inferring origin and destination routes and estimating travel time. Background Art

[0002] Travel Time Estimation (TTE) is the process of predicting the expected duration of a trip from a departure point to a destination. This technology is of great significance in many fields, such as traffic monitoring, route planning, passenger sharing, taxi / online car-hailing scheduling, etc.

[0003] In the related art, many studies focus on the travel time estimation of a given route, but few focus on the origin-destination travel time estimation (OD-TTE) of an unspecified route; however, in many practical applications, drivers only know the origin and destination before traveling, and they often choose routes based on their own preferences rather than following the recommended routes. Therefore, the existing technology ignores the problem of personalized route preferences, and at the same time, it also takes a high time cost to label the sample data for training model algorithms. Summary of the invention

[0004] The embodiments of the present application provide a method, system, computer device and computer-readable storage medium for inferring origin and destination routes and estimating travel time, so as to at least solve the problem in the related art that the travel time estimation result has poor accuracy due to ignoring personalized route preferences.

[0005] In a first aspect, an embodiment of the present application provides a method for inferring a route between a starting point and a destination and estimating travel time, the method comprising:

[0006] Acquire a road network and a corresponding trajectory data set, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point;

[0007] Iteratively training an active learning adversarial inverse reinforcement network based on the trajectory dataset and the road network to obtain an inference model for predicting a personalized driving route between the start and end points;

[0008] The inference model is used to predict the personalized driving route between the starting and ending points, and the travel time required for the personalized driving route is obtained through multiple weak learners integrated based on the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

[0009] In some embodiments, based on the trajectory dataset and the road network, iteratively training the adversarial inverse reinforcement network for active learning includes:

[0010] Using an active learning mechanism, extracting typical trajectory samples from the trajectory data set, adding normal annotations or abnormal annotations to the typical trajectory samples, and optimizing the trajectory data set by deleting the typical trajectory samples with abnormal annotations;

[0011] Based on the optimized trajectory data set, determining a reward function for constraining the model to generate a personalized driving route through iterative alternating training of the generator and the discriminator;

[0012] In each round of training, an optimized trajectory data set for each round of training is obtained through an active learning mechanism, and the generator and the discriminator are alternately trained based on the optimized trajectory data sets of multiple rounds and the reward function is continuously optimized. When the number of iterations reaches a preset number, the trained inference model is obtained.

[0013] In some embodiments, using an active learning mechanism to extract typical trajectory samples from the trajectory data set, and adding normal annotations or abnormal annotations to the typical trajectory samples includes:

[0014] Initialize a normal trajectory set, an abnormal trajectory set, and an unlabeled trajectory set, wherein the unlabeled trajectory set corresponds to different trajectory data sets;

[0015] Using the adversarial inverse reinforcement network in the current round, prediction is performed based on the trajectory samples of the unlabeled trajectory set to obtain a predicted trajectory;

[0016] It is determined whether the trajectory sample is selected as a typical trajectory according to the predicted trajectory. If so, the typical trajectory is added to the normal trajectory set or the abnormal trajectory set according to manual judgment information.

[0017] In some embodiments, based on the optimized trajectory data set, through iterative alternating training of the generator and the discriminator, determining the reward function for constraining the model to generate a personalized driving route includes:

[0018] According to the reward function, the weight of the road network is updated, and based on the road network after the weight update and the optimized trajectory data set, a learner trajectory is generated through a generator, wherein the generator is constructed based on a shortest path algorithm, and in an initial state, the reward function is a preset value;

[0019] Based on the learner trajectory, a discriminator is trained by a binary logistic regression method, wherein the discriminator outputs a similarity evaluation result for distinguishing the learner trajectory from a preset expert trajectory;

[0020] According to the similarity evaluation result output by the discriminator in each round, the reward approximator and the value approximator are iteratively updated to optimize the reward approximator and update the reward function according to the optimized reward approximator, and when the similarity evaluation result meets the expected requirements, the inference model is obtained.

[0021] In some of the embodiments, the discriminator comprises a multi-layer fully connected network and is constructed based on state parameters, action parameters, subsequent state parameters and destination parameters, and the reward function is constructed based on state parameters, action parameters, subsequent state parameters and destination parameters.

[0022] In some embodiments, obtaining the travel time required for the personalized driving route by integrating multiple weak learners based on the AdaBoost algorithm includes:

[0023] By using the plurality of weak learners, respectively performing time series prediction on the personalized driving route, and obtaining a plurality of long-term and short-term dependency sequences related to the travel time;

[0024] By global average pooling, the long-term and short-term dependency sequences are mapped from high-dimensional features to low-dimensional features, and the low-dimensional vectors are integrated by the connection operation of the AdaBoost algorithm to obtain an integrated low-dimensional vector;

[0025] The integrated low-dimensional vector is nonlinearly transformed and combined by using a connection layer to obtain a complex relationship between the personalized driving route and the travel time, and the travel time is obtained based on the complex relationship by using a multi-layer perceptron.

[0026] In some embodiments, the personalized driving route is respectively predicted by using a plurality of weak learners to obtain a plurality of long-term and short-term dependency sequences related to the travel time, including:

[0027] The weak learner obtains the deep topological structure information of the road network through the graph convolutional network, and uses the deep topological structure information to extract high-dimensional spatiotemporal features corresponding to the personalized driving route;

[0028] Through the Transformer module, a multi-head attention mechanism is used to perform time series prediction based on the high-dimensional spatiotemporal features and external reference information to obtain long-term and short-term dependency sequences related to the travel time.

[0029] In a second aspect, an embodiment of the present application provides a system for inferring a route between a starting point and a destination and estimating a travel time, the system comprising: an acquisition module, an inference model training module and a travel time estimation module, wherein:

[0030] The acquisition module is used to acquire a road network and a corresponding trajectory data set, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point;

[0031] The inference model training module is used to iteratively train an active learning adversarial inverse reinforcement network based on the trajectory data set and the road network to obtain an inference model for predicting a personalized driving route between the start and end point positions;

[0032] The travel time estimation module is used to predict the personalized driving route between the starting and ending points using the inference model, and obtain the travel time required for the personalized driving route through multiple weak learners integrated based on the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

[0033] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect above when executing the computer program.

[0034] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0035] Compared with the related art, the embodiment of the present application provides a method for inferring a route between origins and destinations and estimating travel time. By combining an adversarial inverse reinforcement network with an active learning mechanism, it is possible to infer and generate a personalized driving route between origins and destinations that combines user preferences while minimizing the amount of data annotation. Furthermore, the AdaBoost algorithm is used to integrate multiple weak learners of a fused graph convolutional network and a Transformer network, and the graph convolutional network can obtain high-dimensional spatiotemporal characteristics of the traffic network, thereby assisting the Transformer network to better capture the long-term and short-term dependencies corresponding to the personalized driving route. Combining the learning results of multiple weak learners, a richer and more comprehensive temporal feature representation of the personalized driving route is obtained. Finally, the temporal feature representation is used to obtain the travel time corresponding to the route, thereby achieving accurate estimation of the travel time between the origins and destinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1is a flow chart of a method for inferring a route between a starting point and a destination and estimating a travel time according to an embodiment of the present application;

[0038] Figure 2 is a schematic diagram of a training process of adversarial inverse reinforcement learning according to an embodiment of the present application;

[0039] Figure 3 is a schematic diagram of a structure of a neural network for determining a reward approximator according to an embodiment of the present application;

[0040] Figure 4 is a schematic diagram of the structure of a neural network for determining a value approximator according to an embodiment of the present application;

[0041] Figure 5 is a flow chart of another method for inferring a route between a starting point and a destination and estimating a travel time according to an embodiment of the present application;

[0042] Figure 6 It is a schematic diagram of a framework for estimating travel time between origin and destination points based on AdaBoost multi-fusion graph convolution Transformer according to an embodiment of the present application;

[0043] Figure 7 is a structural block diagram of a start-end point route inference and travel time estimation system according to an embodiment of the present application;

[0044] Figure 8 It is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0046] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as insufficient content disclosed in this application.

[0047] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0048] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantitative limitation, and may represent the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0049] In this document, it should be understood that the terms involved may be technical means for implementing a part of the present invention or other summary technical terms. For example, the terms may include:

[0050] Active learning: Active learning aims to reduce the cost of data annotation by selecting the most informative samples for annotation. In this way, the model performance can be improved while annotating as few samples as possible.

[0051] GPS trajectory: A GPS trajectory is an ordered sequence of multiple GPS records. Each record contains longitude, latitude and timestamp. The present invention maps the GPS trajectory to a road network and converts it into a route sequence containing multiple road sections.

[0052] Road network: A road network is modeled as a directed graph consisting of vertices (intersections) and edges (road segments). Although the structure of the road network is fixed, the travel cost of each edge varies with factors such as time, driver, and weather.

[0053] Personalized travel time estimation: Given a starting point for a trip o , destination d, departure time t and driver u, we first infer the most likely personalized driving route and then estimate the personalized travel time of driver u under specific start and end points and departure time.

[0054] Existing methods for estimating origin-destination travel time mainly focus on estimating travel time for specified routes. Few studies focus on origin-destination travel time estimation (OD-TTE) without specified routes, and most of these studies ignore the cost of personalized route preferences and data annotation. To fill this research gap, this application analyzes individual route preferences and proposes a personalized OD-TTE method based on AA-IRL and Transformer. By combining adversarial inverse reinforcement learning with active learning, the cost of sample annotation can be effectively reduced. After inferring possible routes, a method based on the AdaBoost multi-fusion graph convolutional Transformer module (AMGC-Transformer) is proposed to estimate travel time.

[0055] Figure 1 is a flow chart of a method for inferring a route between a starting point and a destination and estimating a travel time according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0056] S101, obtaining a trajectory data set and a road network, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point;

[0057] The road network may be a road network of any city, county, or smaller geographical area; further, in this embodiment, the starting and ending points may be any two geographical locations within the geographical area whose distance is greater than a certain preset length;

[0058] The trajectory dataset is the GPS trajectory of the user's driving process between the starting and ending points in the geographical area corresponding to the road network. It includes multiple trajectory samples, and each trajectory sample consists of multiple trajectory points. By mapping the trajectory samples to the road network, it can be converted into a route sequence containing multiple road sections.

[0059] S102, iteratively training an active learning adversarial inverse reinforcement network based on the trajectory dataset and the road network to obtain an inference model for predicting a personalized driving route between the start and end points;

[0060] It should be noted that active learning aims to reduce the cost of data annotation by selecting the most informative samples for annotation. In this way, the model performance can be improved while annotating as few samples as possible.

[0061] In this embodiment, an active learning mechanism is introduced in the personalized route inference process. In each iteration training, the most informative trajectory samples are selected for annotation, and abnormal trajectories are manually annotated. In each iteration, trajectories that are as normal as possible are used for model training, thereby gradually improving the accuracy of model training.

[0062] In the technical solution of this application, the meaning of "most informative" refers to the data that is most helpful for personalized route inference and travel time estimation, and can provide key information for model training to improve its accuracy. Specifically, the "most informative" samples can be determined by predictive probability distribution analysis and multiple prediction difference comparisons.

[0063] In an exemplary embodiment, the introduction of active learning mechanism to reduce the amount of data annotation is achieved by the following steps:

[0064] Step 1: In each iteration, initialize the normal trajectory set, abnormal trajectory set and unlabeled trajectory set.

[0065] Optionally, define the unlabeled trajectory dataset γ, the road network G, and the threshold μk in the initial state; define the normal trajectory set, abnormal trajectory set, and unlabeled trajectory set as

[0066] Step 2: Use the adversarial inverse reinforcement network in the current round to make predictions based on the unlabeled trajectory set to obtain the predicted trajectory;

[0067] In the first round of training, the A-IRL model is trained using the trajectory dataset γ. The A-IRL model uses the same starting and ending points to predict the predicted trajectory sample p′ based on the trajectory sample p in the training data.

[0068] Step 3, determine whether to select trajectory samples for labeling based on the predicted trajectory. If so, add the selected trajectory samples to the normal trajectory set or the abnormal trajectory set based on the manual operation information.

[0069] Optionally, the prediction error between the predicted trajectory sample p′ and the original trajectory sample p can be calculated by the following formula 1, and further combined with a preset threshold to determine whether the sample is selected for labeling:

[0070]

[0071] Among them, |p∩p′| represents the length of the overlapping part of the trajectory p and p′, |p| represents the length of the original trajectory p, and |p′| represents the length of the predicted trajectory p′. It can be understood that if the result of a sample trajectory meets the above discriminant, the trajectory sample is selected for labeling and divided into the unlabeled trajectory set Ω k ;

[0072] Specifically, the ratio calculated by the above formula 1 can be regarded as the similarity between the current trajectory and its corresponding predicted trajectory. If the similarity is less than or equal to the threshold, it means that the similarity between the trajectory and its predicted trajectory is low, and the trajectory can be considered to be a representative typical trajectory, which may be a trajectory with special traffic conditions, different travel times or purposes, etc.

[0073] Furthermore, normal or abnormal annotation information is manually added to it. It can be understood that the normal trajectory is divided into the normal trajectory set E k , the abnormal trajectories are divided into the abnormal trajectory set U k ; For example, if a trajectory is traveling abnormally slowly during rush hour, it may be marked as "abnormal"; a trajectory that is different from the path selected by most trajectories, including detours, deviations from recommended routes, abnormal times, etc., may also be marked as "abnormal"; and a trajectory under normal traffic conditions may be marked as "normal";.

[0074] Finally, the abnormal trajectory set U k After being deleted from the trajectory dataset, the trajectory dataset is used for model training. The purpose of this step is to ensure that the model is mainly learned based on normal trajectory data during the training process; it can ensure that in each training process, the minimum amount of annotation is used to improve the model training performance.

[0075] Figure 2 is a schematic diagram of a training process of adversarial inverse reinforcement learning according to an embodiment of the present application, such as Figure 2 As shown, the above process is executed cyclically. th In the learning process, trajectory prediction is performed and the selected set is determined as Ω in combination with the threshold μk m , and manually mark Ω m Divide into normal set E m and the exception set U m ; in the mth thIn the learning process, trajectory prediction is performed and the selected set is determined as Ω in combination with the threshold μk n , through artificial marking, Ω n Divide into normal set E n and the exception set U n ; in the mth th In the learning process, trajectory prediction is performed and the selected set is determined as Ω in combination with the threshold μk k , through artificial marking, Ω k Divide into normal set E k and the exception set U k .

[0076] The goal of using active learning is to improve model performance with minimal annotation. This means selecting trajectory samples for annotation as much as possible to select those that will improve model performance the most. In this way, the accuracy of the model can be maximized while minimizing the workload of manual annotation.

[0077] Furthermore, based on the trajectory dataset, iterative training of the active learning adversarial inverse reinforcement network includes the following steps:

[0078] Step 1: extract the most informative typical trajectory samples in the trajectory dataset, add normal or abnormal annotations to the typical trajectory samples, and optimize the trajectory dataset by deleting the typical trajectory samples with abnormal annotations;

[0079] The purpose of this step is to ensure that the model learns mainly based on normal trajectory data during training. Abnormal trajectories may interfere with the training of the model and cause the model performance to degrade. By removing abnormal trajectories, the model can focus more on learning the routes and travel time patterns under normal circumstances.

[0080] It can be understood that this step is implemented through the specific implementation step Step-Step 4 of the above-mentioned active learning; it will not be repeated here; but it should be noted that the training data set used to train the adversarial inverse reinforcement learning network is updated in real time as the training progresses.

[0081] Step 2: Based on the optimized trajectory dataset, the reward function used to assist the model in generating personalized driving routes is determined through iterative alternating training of the generator and the discriminator.

[0082] The specific steps include:

[0083] S1, initialization:

[0084] The reward function in the initial state is set according to the prior knowledge in the field. For example, if it is known that certain roads are usually congested during a specific time period, then these roads can be given lower reward values ​​in the initial reward function to guide route selection to avoid congested sections. In the absence of clear prior knowledge, reward values ​​can be randomly assigned to each section in the road network. Although this method has a certain degree of uncertainty, the reward function can be gradually adjusted through subsequent learning and optimization processes to make it closer to the actual situation;

[0085] Through the reward function r θ Update the weights of the road network; the weights of each road segment in the road network can be adjusted to reflect different traffic conditions, route preferences, etc.

[0086] After that, the shortest path algorithms such as Dijkstra, Bellman-Ford and Floyd-Warshall are used to collect the learner's trajectory in the road network based on the optimized trajectory data set output by the above steps. Among them, the Dijkstra algorithm is usually used to find the shortest path in the graph. In this embodiment, the Dijkstra algorithm is optionally used to determine the learner's actual driving trajectory under different weight settings; its role is equivalent to the generator.

[0087] S2, training discriminator:

[0088] In this embodiment, the discriminator D is trained by binary logistic regression. θ，φ Trained to distinguish between expert trajectories and learner trajectories, thereby identifying the route that best fits the historical data, specifically:

[0089] The purpose of this step is to train the discriminator D θ，φ To accurately distinguish the expert trajectory from the learner's trajectory; among them, the expert trajectory usually represents a more optimized and practical route selection, while the learner's trajectory is generated by the generator in the above steps, which may contain various different situations;

[0090] Specifically, some learner trajectories may be inefficient or unreasonable routes. By distinguishing these two types of trajectories, the discriminator can learn which is the better route choice, thereby providing a reference for personalized route inference. In this embodiment, the discriminator outputs a similarity evaluation result for distinguishing the learner trajectory from the preset expert trajectory.

[0091] For example, in traffic route planning, the expert trajectory may be a route frequently chosen by experienced drivers based on user preferences, which usually avoids congested roads and reduces travel time, while the learner trajectory may include the trial routes of novice drivers or the route choices in special situations.

[0092] On the other hand, by training the discriminator, we can identify the route that best matches the historical data. The historical data contains a lot of information about different routes, including traffic flow, travel time, road conditions, etc. By training the discriminator, the model can learn from these historical data which routes are more reliable and effective, so that it can better combine historical experience when making personalized route inferences and provide more accurate suggestions.

[0093] For example, historical data may show that during a certain time period, a specific combination of roads consistently leads to a shorter time to destination. By training the discriminator, the model can identify these routes and prioritize them when making route recommendations for new users.

[0094] It should also be noted that binary logistic regression is a commonly used classification algorithm. In this embodiment, a discriminator D_(θ, φ) is established and trained based on binary logistic regression, where θ represents the reward network and φ represents the value network. The binary regression logic converts the input trajectory feature into a probability value between 0 and 1, indicating the possibility that the trajectory belongs to the expert trajectory. Furthermore, by setting a threshold, the trajectory can be clearly classified as an expert trajectory or a learner trajectory.

[0095] S3. Update reward and value approximators

[0096] Specifically, according to the discriminator D θ，φ As a result, update the personalized reward approximator r θ and value approximator h φ , thereby optimizing the performance of the model.

[0097] Among them, the discriminator D θ，φ In the process of distinguishing expert trajectories from learner trajectories, feedback information is generated, which can be used to guide the update of the reward approximator and the value approximator. For example, if the discriminator finds that a trajectory is incorrectly classified as a learner trajectory, it may mean that the reward value or value assessment corresponding to the trajectory is inaccurate.

[0098] Taking traffic route planning as an example, if the discriminator determines that a route should not be recommended to the user, but in fact this route may be a reasonable choice in some cases, then the reward approximator and value approximator need to be adjusted so that the model can more accurately evaluate the value of this route.

[0099] Furthermore, the reward approximator r θ Used to evaluate the reward value under different states. According to the results of the discriminator, the parameters of the reward approximator can be adjusted so that the model can more accurately give corresponding rewards for different routes.

[0100] For example, in a traffic congested area, if the discriminator finds that the expert trajectory prefers a side road rather than a main road, then the reward approximation of this side road in the congested state can be improved, thereby guiding the model to prefer this side road when recommending routes for users.

[0101] In addition, the value approximator hφ is used to evaluate the value of different state and destination combinations. According to the results of the discriminator, the parameters of the value approximator can be updated so that the model can more accurately evaluate the value of different routes for a specific destination; if the discriminator finds that some routes perform better when reaching a specific destination, the value approximator can be adjusted to improve the value evaluation of these routes at the corresponding destination.

[0102] For example, if a user is heading to the airport, and the discriminator determines that certain routes can get to the airport faster, then the approximate value of these routes when heading to the airport destination can be increased, causing the model to give priority to these routes when recommending routes for the user.

[0103] It is understandable that updating the reward and value approximators is an ongoing process. As new data is continuously input and the discriminator is continuously optimized, the reward approximator and the value approximator also need to be constantly adjusted; an accurate reward approximator can better reflect the advantages and disadvantages of different routes, while the value approximator can help the model evaluate the value of different routes in long-term use.

[0104] Furthermore, the process of continuously optimizing the discriminator and updating the reward approximator and value approximator is the process of continuously training the model and continuously optimizing the reward function. Through multiple iterations, the model can gradually improve its performance and better adapt to different situations and user needs.

[0105] In addition, it should be noted that the discriminator proposed in this embodiment also takes into account the influence of the destination, which is defined as follows:

[0106]

[0107] Among them, s, a, s′, and g represent the state, action, next state, and destination, respectively. Specifically, combined with the application scenario of this case and the relevant knowledge of Markov decision process, it can be known that the above state s and the next state s′ represent the path, and the action a represents the selection of the next road section at the intersection. θ，φ The transformation range of (s, a, s′, β) is (0, 1), and π(a|s, g)∈{0, 1} is calculated using the Dijkstra algorithm, which represents the probability of the user choosing action a when he is in state s and has d as the destination.

[0108] In the route inference method in the related art, the reward function f of the target θ，φ (s, a, s′, g) is approximated by a reward approximator r θ and a value approximator h φ limit;

[0109] Figure 3 is a schematic diagram of a structure of a neural network for determining a reward approximator according to an embodiment of the present application, Figure 4 is a schematic diagram of the structure of a neural network for determining a value approximator according to an embodiment of the present application, such as Figure 3 and Figure 4 As shown, since the application also considers the influence of the destination on the route selection, the reward function is modified as follows:

[0110] f θ，φ (s, a, s′, g) = r θ (s)+γh φ (s′, g)-h φ (s, g)

[0111] Among them, r θ (s) and h φ The variation ranges of (s, g) are (-2, 0) and (-∞, +∞), respectively;

[0112] After the last round of training, a new reward function is determined; the new reward function is used to update the weights of the road network during this round of training and to regenerate trajectories based on the updated weights.

[0113] In this embodiment, by taking the destination into consideration to build the discriminator, the applicability of different trajectories to a specific destination can be better distinguished, which helps the model to more accurately infer the best route to a specific destination, rather than just judging based on general route characteristics. For example, for different destinations to the city center and to the suburbs, traffic conditions and route choices may be very different. Considering the impact of the destination can enable the model to make more targeted route inferences.

[0114] In addition, different users may have different destination requirements. Considering the destination impact can also enable the model to provide personalized route recommendations based on the user's specific destination. This can better meet the actual needs of users and improve user satisfaction with route recommendations. At the same time, by incorporating the destination into the reward function, the model can also be more inclined to routes that can reach the destination faster and safer when selecting routes.

[0115] Through the above step S102, by continuously updating the road network weights, training the discriminator, and updating the reward and value approximator, the model's ability to infer personalized route preferences is gradually optimized, while fully considering the influence of factors such as the destination, and finally obtaining an inference model that can generate the most likely personalized driving route between the starting and ending points.

[0116] S103, using the inference model to predict the personalized driving route between the starting and ending points, and obtaining the travel time required for the personalized driving route through multiple weak learners integrated by the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

[0117] Figure 5 is a flow chart of another method for inferring a route between a starting point and a destination and estimating a travel time according to an embodiment of the present application, such as Figure 5 As shown in FIG. 1 , after obtaining the most likely personalized driving routes, predicting travel time includes the following subdivision steps:

[0118] Step 1, through multiple weak learners, the personalized driving routes are predicted in time series to obtain multiple long-term and short-term dependency sequences related to travel time;

[0119] In this embodiment, any weak learner consists of a graph convolutional network and a Transformer module;

[0120] like Figure 5 As shown in the figure, after the route sequence filling operation is performed on the most likely route, it enters the graph convolutional network (GCN); the role of the graph convolutional network is to capture the topological structure information of the road network, explore the non-Euclidean spatial associations between road sections, and make full use of the non-Euclidean structure of the road network. By analyzing the topological structure information of the road network, GCN can better understand the network environment in which these personalized driving routes are located.

[0121] Furthermore, the output of the graph convolutional network is input into the Transformer module for further processing together with external information (such as traffic conditions, driver information, atmospheric environment, etc.). This interaction enables the model to comprehensively consider the structural characteristics of the road network and the impact of external factors on travel time.

[0122] Specifically, the graph convolutional network (GCN) provides important information about spatial features to the Transformer module, helping it better capture the long-term and short-term dependencies corresponding to personalized driving routes. The spatial information provided by GCN provides a basic framework for it, enabling the Transformer to more accurately analyze the traffic relationships and travel time changes between different locations.

[0123] Furthermore, the long-term and short-term dependencies captured by Transformer mainly refer to the dependencies between travel time and different factors in time and space. In the time dimension, long-term dependencies may involve traffic flow trends in different time periods, the impact of seasonal factors on travel time, etc.; short-term dependencies may include traffic congestion at the current moment, the immediate impact of emergencies on travel time, etc.

[0124] In the spatial dimension, dependencies can include the transfer of traffic flow between different road sections, the impact of the connection relationship between nodes on the route on travel time, etc. For example, if a road section is congested, it may have a long-term impact on the travel time of subsequent sections; while the traffic conditions of the current section will have a short-term impact on the travel time of nearby sections.

[0125] By capturing these long-term and short-term dependencies, Transformer is better able to adapt to complex travel time estimation scenarios. It can accurately predict future travel times based on historical traffic data and real-time traffic information, and can quickly respond to emergencies and changes in traffic conditions.

[0126] It can be understood that each specific combination of GCN and a Transformer module can be regarded as an independent processing unit, which together constitute an MGC-Transformer (composed of GCN and Transformer) instance. These instances are each defined as a weak learner. Multiple weak learners are integrated through the AdaBoost algorithm to improve the generalization ability and prediction performance of the model.

[0127] Step 2: Perform global average pooling on the long-term and short-term dependency sequences respectively, map the long-term and short-term dependency sequences from high-dimensional features to low-dimensional features, and integrate the low-dimensional vectors through the connection operation of the AdaBoost algorithm to obtain an integrated low-dimensional vector.

[0128] It can be understood that the feature vector obtained after global average pooling represents the average features of the sequence processed by the Transformer in the spatial dimension. These feature vectors will be integrated with the outputs from different Transformer blocks to form a richer and more comprehensive feature representation.

[0129] It can be understood that by integrating the output results of multiple weak learners, the AdaBoost algorithm can integrate various information so that the subsequent processing steps can better learn the complex relationship between input data and travel time and effectively respond to the dynamic changes in traffic conditions.

[0130] Among them, global average pooling is mainly used to reduce the dimension of the output after being processed by multiple Transformer blocks and extract key features. By averaging each feature dimension, it can summarize the information of the entire sequence, reduce the number of model parameters, prevent overfitting, and provide a more concise and representative feature representation for subsequent processing steps.

[0131] For the output sequence after multiple Transformer modules, assume that its structure is [B, L, D], where B represents the batch size, L represents the sequence length, and D represents the feature dimension. Global average pooling averages the entire sequence length L on each feature dimension of each batch.

[0132] Specifically, the calculation method is to calculate the average value of all positions for each feature dimension d (d∈{1, 2, ..., D}) in each batch, that is, sum up the values ​​of all [i, j, d] (where i∈{1, 2, ..., B}, j∈{1, 2, ..., L}), and then divide it by the sequence length L to obtain an output with a shape of [B, D].

[0133] Step 3: Use multiple fully connected layers to perform nonlinear transformation and combined connection on the integrated low-dimensional limits to obtain the complex relationship between personalized driving routes and travel time, and obtain the travel time based on the complex relationship.

[0134] Among them, the multi-layer perceptron is usually composed of multiple fully connected layers. The result of global average pooling first enters the first fully connected layer, which performs linear transformation and nonlinear activation function processing (such as ReLU, etc.) on the input to map the input features to a new high-dimensional space. Then, the subsequent fully connected layers continue to further transform and process the features, gradually extracting features that are highly correlated with travel time.

[0135] The concatenated results are processed again by a multi-layer perceptron to further learn and integrate these comprehensive features. This multi-layer perceptron can adjust the weights according to the concatenated features, capture the complex relationship between different features, and finally output the estimated travel time between the origin and destination. This estimated value can be a scalar, representing the predicted travel time.

[0136] Through the above step S103, different from the traditional technology of estimating travel time according to a predetermined route, the present application scheme combines the preferences and driving habits of each driver to perform route inference, and then obtains a personalized route. After the personalized driving route is input into the inference model, the model can make full use of the sequence of long-term and short-term dependency information processed by the Transformer, and adopts the AdaBoost algorithm. By integrating the output results of multiple weak learners, it can integrate various information, and then perform dimensionality reduction and feature extraction through global average pooling, and then gradually learn the complex relationship between the input data and the travel time through the nonlinear transformation and result connection of the multi-layer perceptron, so as to achieve accurate travel time estimation of the starting and ending points.

[0137] In addition, the training process of AMGC-Transformer is an iterative process, which achieves accurate prediction of travel time through the gradual optimization and integration of multiple weak learners (MGC-Transformer model). Figure 6 is a schematic diagram of a framework for estimating travel time between origin and destination points based on AdaBoost multi-fusion graph convolution Transformer according to an embodiment of the present application, such as Figure 6 As shown, the specific training process includes the following steps:

[0138] 1. Train the i-th MGC-Transformer model using the i-th training subset and predict the travel time y of the training set T i (γ):

[0139] The data is divided into multiple training subsets, and the i-th training subset is used to train the i-th MGC-Transformer model. Each MGC-Transformer consists of a graph convolutional network (GCN) and a Transformer. After training, the model predicts the travel time for the entire training set T and obtains the prediction result y i (γ).

[0140] 2. Calculate the relative error of the i-th weak learner on each sample using the following formula:

[0141]

[0142] For each sample q in the training set, the relative error of the i-th weak learner is calculated. i (q) is the predicted travel time of sample q by the i-th weak learner, and y(q)) is the actual travel time of sample q. By calculating the relative error, the prediction accuracy of each weak learner on a single sample can be measured.

[0143] 3. Calculate the total relative error of the i-th weak learner using the following formula:

[0144]

[0145] The relative errors of all samples are summed to obtain the total relative error of the i-th weak learner. This total relative error reflects the overall performance of the weak learner on the entire training set.

[0146] 4. Update the weight of the i-th weak learner using the following formula:

[0147]

[0148] Among them, according to the total relative error of the weak learner, its weight is calculated, and this weight will be used for the subsequent update of the sample weight and the integration of the strong learner. If the total relative error of the weak learner is small, it means that it performs well on the training set, and the weight will increase accordingly; otherwise, the weight will decrease.

[0149] 5. Update the weight of each sample in the training set using the following formula:

[0150]

[0151] For each sample q in the training set, update its weight, D i (q) is the weight of sample q in the previous iteration, D i+1 (q) is the updated weight, a i is the weight coefficient of the ith weak learner. The process of updating the weight takes into account the weight of the weak learner, the predicted travel time of the sample by the weak learner, and the actual travel time of the sample. The purpose of this is to make the model pay more attention to those samples that are wrongly predicted and improve the model's learning ability for difficult samples.

[0152] By repeating the above steps, each weak learner is continuously optimized and the weight of the samples is continuously adjusted. Finally, multiple weak learners are linearly integrated to form a strong learner (AMGC-Transformer) to achieve accurate estimation of travel time. This ensemble learning method can improve prediction performance and enhance the robustness of the model to uncertainty.

[0153] Through the above steps S101 to S103, combined with the adversarial inverse reinforcement network of the active learning mechanism, it is possible to infer and generate a personalized driving route between the starting and ending points that combines user preferences while minimizing the amount of data annotation; further, the AdaBoost algorithm is used to integrate multiple weak learners of the fused graph convolutional network and the Transformer network, and the graph convolutional network can obtain the high-dimensional spatiotemporal characteristics of the traffic network, thereby assisting the Transformer network to better capture the long-term and short-term dependencies corresponding to the personalized driving route; combining the learning results of multiple weak learners, a richer and more comprehensive temporal feature representation of the personalized driving route is obtained, and finally, the travel time corresponding to the route is obtained using the temporal feature representation, which can achieve accurate estimation of the travel time between the starting and ending points.

[0154] In a second aspect, this embodiment also provides a system for inferring a route between a starting point and a destination and estimating travel time. Figure 7 is a structural block diagram of a system for inferring origin and destination routes and estimating travel time according to an embodiment of the present application, such as Figure 7 As shown, the system includes: an acquisition module 70, an inference model training module 71 and a travel time estimation module 72, wherein:

[0155] The acquisition module 70 is used to acquire a road network and a corresponding trajectory data set, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point;

[0156] The inference model training module 71 is used to iteratively train the active learning adversarial inverse reinforcement network based on the trajectory data set and the road network to obtain an inference model for predicting a personalized driving route between the start and end point locations;

[0157] The travel time estimation module 72 is used to predict the personalized driving route between the starting and ending points using the inference model, and obtain the travel time required for the personalized driving route through multiple weak learners integrated based on the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

[0158] Through the above system, combined with the adversarial inverse reinforcement network of the active learning mechanism, it is possible to infer and generate personalized driving routes between the starting and ending points that combine user preferences while minimizing the amount of data annotation; further, through the AdaBoost algorithm, multiple weak learners of the fusion graph convolutional network and the Transformer network are integrated together, and the learning results of multiple weak learners are combined to obtain a richer and more comprehensive temporal feature representation of the personalized driving route. Finally, the travel time corresponding to the route is obtained using the temporal feature representation, which can achieve accurate estimation of the travel time between the starting and ending points.

[0159] Finally, in order to verify the effectiveness of the origin and destination route inference and travel time estimation of this embodiment, a detailed comparison will be conducted through experiments as follows:

[0160] Numerical experiments:

[0161] The data was collected using online car-hailing trajectory data from a certain online car-hailing platform's travel plan. The data collection area is located in the downtown area of ​​a certain city in China, and contains a simplified road network with a total of 5044 road sections. Its latitude range is 30.65°N to 30.73°N, and its longitude range is 104.04°E to 104.13°E, covering an area of ​​approximately 64 square kilometers. 300 online car-hailing drivers were selected based on the number of trajectories, and the online car-hailing trajectory data was divided into training set, validation set, and test set.

[0162] In order to evaluate the data mining capabilities of multiple models on unlabeled trajectory datasets, 12,000 trajectories were randomly generated and inserted into the training set. The start-end pairs of these artificially generated trajectories were randomly selected from real historical trajectory datasets. After data processing, the training set, validation set, and test set contained 95,141 trajectories (plus 12,000 artificially generated trajectories), 21,047 trajectories, and 21,876 trajectories, respectively. During the training process, the AA-IRL model proposed in the present application scheme selects informative trajectories from the training set and annotates them using active learning methods. In addition, the map matching method in the prior art is used to match the GPS trajectory to the road network.

[0163] The application scheme mainly includes two steps: inferring possible driving routes and estimating the travel time of the inferred routes. Therefore, multiple evaluation indicators are used to evaluate the effects of these two steps. In order to evaluate the performance of different methods in origin-destination travel time estimation (OD-TTE), mean absolute error (MAE), mean absolute percentage error (MAPE) and satisfaction (SR) are used as evaluation indicators, which are widely used in travel time prediction.

[0164] Experimental results:

[0165] In order to show the experimental results more comprehensively, the trajectories in the test set are divided into three categories: short distance (≤2.5 km), medium distance (2.5-5.0 km) and long distance (≥5.0 km). The test set contains 5986 short distance trajectories, 8563 medium distance trajectories and 7327 long distance trajectories respectively. Compared with A-IRL, the AA-IRL model proposed in this application scheme not only modifies the reward function, but also combines active learning with A-IRL to annotate the training trajectories at a lower time cost. The model proposed in this application scheme surpasses the existing IRL model in the performance of personalized route inference. Existing IRL studies show relatively good performance in route inference, but they are susceptible to noise and rely on data annotation. When the training set is unlabeled and contains a certain proportion of abnormal trajectories, the performance of the existing inverse reinforcement learning model will deteriorate. The F1 scores of MED-IRL, A-IRL, MEDD-IRL and AMED-IRL are 0.6923, 0.7073, 0.7034 and 0.7054 respectively. Compared with the existing IRL technical solution, the model proposed in the present application solution improves the F1 score by 0.7408-0.7073=3.35%.

[0166] At the same time, as the route distance increases, the improvement of the accuracy of route inference in the present application scheme is more obvious. The longer the travel distance, the more difficult it is to predict which route the user will choose. For short routes, both the model proposed in the present application scheme and the existing IRL model can achieve satisfactory performance, and the F1 scores of all models are higher than 0.84. Although the F1 score of the model proposed in the present application scheme decreases as the route becomes longer, its advantage over existing research is expanding. For short, medium and long routes, the F1 scores of the model proposed in the present application scheme are 2.54%, 3.43% and 3.91% higher than A-IRL, respectively. Compared with the task of estimating route travel time, the main challenge of OD-TTE lies in inferring the route that the user will choose. Accurately inferring the user's route choice is very important for estimating the driver's personalized start-end travel time.

[0167] This embodiment further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0168] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0169] S1, obtain the road network and its corresponding trajectory dataset;

[0170] S2, based on the trajectory dataset and road network, iteratively trains the active learning adversarial inverse reinforcement network to obtain an inference model for predicting personalized driving routes between the origin and destination locations;

[0171] S3 uses the inference model to predict the personalized driving route between the starting and ending points, and obtains the travel time required for the personalized driving route through multiple weak learners integrated based on the AdaBoost algorithm.

[0172] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be described in detail here.

[0173] In one embodiment, Figure 8 is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application, such as Figure 8 As shown, an electronic device is provided, which may be a server, and its internal structure diagram may be as shown in Figure 8 As shown. The electronic device includes a processor, a network interface, an internal memory and a non-volatile memory connected through an internal bus, wherein the non-volatile memory stores an operating system, a computer program and a database. The processor is used to provide computing and control capabilities, the network interface is used to communicate with an external terminal through a network connection, the internal memory is used to provide an environment for the operation of the operating system, the computer program is executed by the processor to implement a method for inferring a route between a starting point and a destination and estimating a travel time, and the database is used to store data.

[0174] Those skilled in the art will understand that Figure 8 The structure shown in the figure is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the electronic device to which the scheme of the present application is applied. Specifically, the electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0175] In addition, in combination with the adaptive evolution warning method for abnormal traffic behavior in the above embodiment, the embodiment of the present application can provide a storage medium for implementation. The storage medium stores a computer program; when the computer program is executed by the processor, any of the start-end point route inference and travel time estimation methods in the above embodiment is implemented.

[0176] In one embodiment, a computer program product is provided, including a computer program, characterized in that when the computer program is executed by a processor, any one of the methods for inferring a route between a starting point and a destination and estimating a travel time in the above-mentioned embodiments is implemented.

[0177] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0178] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for inferring a route between a starting point and a destination and estimating travel time, characterized in that: The method comprises: Acquire a road network and a corresponding trajectory data set, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point; Iteratively training an active learning adversarial inverse reinforcement network based on the trajectory dataset and the road network to obtain an inference model for predicting a personalized driving route between the start and end points; The inference model is used to predict the personalized driving route between the starting and ending points, and the travel time required for the personalized driving route is obtained through multiple weak learners integrated based on the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

2. The method according to claim 1, characterized in that Based on the trajectory dataset and the road network, iteratively training an adversarial inverse reinforcement network for active learning includes: Using an active learning mechanism, extracting typical trajectory samples from the trajectory data set, adding normal annotations or abnormal annotations to the typical trajectory samples, and optimizing the trajectory data set by deleting the typical trajectory samples with abnormal annotations; Based on the optimized trajectory data set, determining a reward function for constraining the model to generate a personalized driving route through iterative alternating training of the generator and the discriminator; In each round of training, an optimized trajectory data set for each round of training is obtained through an active learning mechanism, and the generator and the discriminator are alternately trained based on the optimized trajectory data sets of multiple rounds and the reward function is continuously optimized. When the number of iterations reaches a preset number, the trained inference model is obtained.

3. The method according to claim 2, characterized in that Using an active learning mechanism, extracting typical trajectory samples from the trajectory data set, and adding normal annotations or abnormal annotations to the typical trajectory samples include: Initialize a normal trajectory set, an abnormal trajectory set, and an unlabeled trajectory set, wherein the unlabeled trajectory set corresponds to different trajectory data sets; Using the adversarial inverse reinforcement network in the current round, prediction is performed based on the trajectory samples of the unlabeled trajectory set to obtain a predicted trajectory; It is determined whether the trajectory sample is selected as a typical trajectory according to the predicted trajectory. If so, the typical trajectory is added to the normal trajectory set or the abnormal trajectory set according to manual judgment information.

4. The method according to claim 2, characterized in that: Based on the optimized trajectory data set, through iterative alternating training of the generator and the discriminator, the reward function used to constrain the model to generate a personalized driving route includes: According to the reward function, the weight of the road network is updated, and based on the road network after the weight update and the optimized trajectory data set, a learner trajectory is generated through a generator, wherein the generator is constructed based on a shortest path algorithm, and in an initial state, the reward function is a preset value; Based on the learner trajectory, a discriminator is trained by a binary logistic regression method, wherein the discriminator outputs a similarity evaluation result for distinguishing the learner trajectory from a preset expert trajectory; According to the similarity evaluation result output by the discriminator in each round, the reward approximator and the value approximator are iteratively updated to optimize the reward approximator and update the reward function according to the optimized reward approximator, and when the similarity evaluation result meets the expected requirements, the inference model is obtained.

5. The method according to claim 4, characterized in that The discriminator includes a multi-layer fully connected network and is constructed based on state parameters, action parameters, subsequent state parameters and destination parameters. The reward function is constructed based on state parameters, action parameters, subsequent state parameters and destination parameters.

6. The method according to claim 1, characterized in that The travel time required for obtaining the personalized driving route by integrating multiple weak learners based on the AdaBoost algorithm includes: By using the plurality of weak learners, respectively performing time series prediction on the personalized driving route, and obtaining a plurality of long-term and short-term dependency sequences related to the travel time; By global average pooling, the long-term and short-term dependency sequences are mapped from high-dimensional features to low-dimensional features, and the low-dimensional vectors are integrated by the connection operation of the AdaBoost algorithm to obtain an integrated low-dimensional vector; The integrated low-dimensional vector is nonlinearly transformed and combined by using a connection layer to obtain a complex relationship between the personalized driving route and the travel time, and the travel time is obtained based on the complex relationship by using a multi-layer perceptron.

7. The method according to claim 6, characterized in that By using the multiple weak learners, the personalized driving route is respectively predicted in time series, and multiple long-term and short-term dependency sequences related to the travel time are obtained, including: The weak learner obtains the deep topological structure information of the road network through the graph convolutional network, and uses the deep topological structure information to extract high-dimensional spatiotemporal features corresponding to the personalized driving route; Through the Transformer module, a multi-head attention mechanism is used to perform time series prediction based on the high-dimensional spatiotemporal features and external reference information to obtain long-term and short-term dependency sequences related to the travel time.

8. A system for inferring a route between a starting point and a destination and estimating travel time, characterized in that: The system comprises: an acquisition module, an inference model training module and a travel time estimation module, wherein: The acquisition module is used to acquire a road network and a corresponding trajectory data set, wherein the trajectory data set includes a plurality of trajectory samples, and each trajectory sample includes positioning trajectory information between a starting point and an end point; The inference model training module is used to iteratively train an active learning adversarial inverse reinforcement network based on the trajectory data set and the road network to obtain an inference model for predicting a personalized driving route between the start and end point positions; The travel time estimation module is used to predict the personalized driving route between the starting and ending points using the inference model, and obtain the travel time required for the personalized driving route through multiple weak learners integrated based on the AdaBoost algorithm, wherein the weak learners include a graph convolutional network and a Transformer module.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images