A model training method, a route sorting method and a computing device

Abstract

Embodiments disclosed in the specification provide a model training method, a route sorting method and a computing device. A candidate navigation route, feature information associated with a road segment included in the candidate navigation route, and a sorting ground truth of the candidate navigation route are obtained. The feature information associated with the road segment of the candidate navigation route is input into a first feature mapping model to be trained, and a first feature vector of each road segment of the candidate navigation route is output. A standard feature vector of each road segment is determined according to at least the first feature vector of the road segment. At least the candidate navigation route and the standard feature vector of each road segment of the candidate navigation route are input into a route sorting model to be trained, and a sorting prediction value of the candidate navigation route is output. The first feature mapping model and the route sorting model are adjusted to reduce the difference between the sorting ground truth and the sorting prediction value.

Description

Technical Field

[0001] Several embodiments of this specification relate to the field of navigation technology, and in particular to a model training method, a route sorting method, and a computing device. Background Technology

[0002] In navigation scenarios, a route ranking model is typically used to rank multiple candidate navigation routes, and then recommends several (e.g., three) of the top-ranked candidate routes to the user. The user can then select one of these candidate routes to begin navigation. These candidate navigation routes consist of a series of topologically connected road segments.

[0003] When training the route ranking model, the training samples include multiple candidate navigation routes to be ranked and feature vectors of road segments within those candidate routes. The feature vectors of road segments are determined based on the attribute information of the road segment itself. The training labels corresponding to the training samples are the true ranking values ​​determined after the multiple candidate navigation routes have been actually selected or evaluated. Inputting the training samples into the route ranking model outputs the predicted ranking values ​​of the multiple candidate navigation routes. It is easy to understand that the process of training the route ranking model is essentially a process of making the predicted ranking values ​​output by the route ranking model increasingly closer to the true ranking values.

[0004] However, the inventors of this disclosure have discovered that if some road segments are not involved or are involved very little in model training, the trained route ranking model is not accurate in ranking candidate navigation routes that include these road segments. Summary of the Invention

[0005] According to a first aspect of several embodiments of this specification, a model training method is proposed, comprising:

[0006] Obtain candidate navigation routes, feature information associated with road segments included in the candidate navigation routes, and the ranking truth value of the candidate navigation routes; wherein, the feature information associated with road segments includes: the attribute information of the road segment, and the attribute information of at least one neighboring road segment in the road network that is topologically connected to the road segment;

[0007] The feature information associated with the road segments of the candidate navigation route is input into the first feature mapping model to be trained, and the first feature vector of each road segment of the candidate navigation route is output.

[0008] At least based on the first feature vector of each road segment, determine the standard feature vector of that road segment;

[0009] The candidate navigation routes and the standard feature vectors of each segment of the candidate navigation routes are input into the route ranking model to be trained, and the ranking prediction values ​​of the candidate navigation routes are output.

[0010] With the goal of reducing the difference between the true ranking value and the predicted ranking value, the first feature mapping model and the route ranking model are adjusted.

[0011] According to a second aspect of several embodiments of this specification, a route sequencing method is proposed, comprising:

[0012] Retrieve multiple candidate navigation routes to be sorted;

[0013] Obtain the standard feature vectors of the road segments included in the candidate navigation route;

[0014] The multiple candidate navigation routes, along with the standard feature vectors of each segment of each candidate navigation route, are input into a trained route ranking model, which then outputs the ranking prediction values ​​of the candidate navigation routes to be ranked.

[0015] According to a third aspect of several embodiments of this specification, a computing device is provided, including a memory and a processor; the memory is used to store computer instructions executable on the processor, and the processor is used to implement the method of the first or second aspect when executing the computer instructions.

[0016] In the technical solution provided in this disclosure, road segments that do not participate in or rarely participate in model training are referred to as low-frequency road segments, while road segments that participate in model training at high frequencies are referred to as high-frequency road segments. Considering that in the topology of the entire road network, the neighbor relationships formed by low-frequency road segments and their surrounding neighboring road segments may be the same or similar to those formed by high-frequency road segments and their surrounding neighboring road segments, and that such similar neighbor relationships have roughly the same reference value for navigation routes, the first idea conceived in this disclosure is to find a technical implementation method that can transfer the experience learned by the route ranking model during the training of high-frequency road segments to the analysis of low-frequency road segments, so that even when the trained route ranking model performs ranking analysis on candidate navigation routes containing low-frequency road segments, it can accurately evaluate the value of candidate navigation routes containing low-frequency road segments based on the transferred experience, thereby enabling a more accurate ranking of the priority of candidate navigation routes containing low-frequency road segments.

[0017] In terms of specific technical implementation, the first feature mapping model and the route ranking model are jointly trained, and the data composition of the training samples is changed. The training samples should include not only the candidate navigation routes and the attribute information of each segment of the candidate navigation routes, but also the attribute information of at least one neighboring segment that is topologically connected to the segment. The first feature mapping model is used to map the feature information associated with the segments into feature vectors. The feature information of a segment includes not only its own attribute information, but also the attribute information of at least one neighboring segment that is topologically connected to the segment. In this way, the feature vector of a segment input to the route ranking model contains not only the attribute information of the segment itself, but also the neighbor relationship information of the segment and its surrounding neighbors.

[0018] Thus, although the candidate navigation routes used to train the route ranking model mostly contain high-frequency road segments rather than low-frequency road segments, from the perspective of the trained route ranking model, even if a high-frequency road segment and a low-frequency road segment have different attribute information, if their corresponding neighbor relationships are the same or similar, then the route ranking model will understand the low-frequency road segment in the same way it understands the high-frequency road segment. It will more accurately evaluate the value of the low-frequency road segment to the navigation route, thereby achieving a more accurate ranking of candidate navigation routes containing the low-frequency road segment. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the road network topology provided in this publication.

[0020] Figure 2 An exemplary flowchart of a model training method is provided.

[0021] Figure 3 An example diagram illustrating the joint training of multiple models is provided.

[0022] Figure 4 An exemplary flowchart of a route sorting algorithm is provided.

[0023] Figure 5 An example diagram is provided showing how a travel navigation service sorts multiple candidate navigation routes online in real time.

[0024] Figure 6 This is a schematic diagram of the structure of a computing device provided in this disclosure.

[0025] In the accompanying drawings, identical or corresponding reference numerals denote identical or corresponding parts. Any number of elements in the drawings is for illustrative purposes only and not for limitation, and any naming is for distinction only and has no limiting meaning. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0027] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0028] The user information (including but not limited to user device information, user personal information, user location information, user travel information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0029] The navigation routes described in this disclosure refer to travel routes provided by travel navigation services and displayed on electronic maps, along which users can start and end their trips. Navigation routes typically include a single road segment, or a sequence of at least two road segments. Travel navigation services usually store global road network data, which includes a unique identifier for each road segment (e.g., road segment 1, road segment 2, etc.) and may also include a global road segment topology (which road segments each road segment connects to). The global road network data may be updated periodically or irregularly, primarily involving the addition or deletion of road segments. For example, the original road segment 1 can be split into road segment 3 and road segment 4; in this case, the original road segment 1 is deleted, and road segments 3 and 4 are added.

[0030] After determining the user's origin and destination, navigation services typically recommend one or more candidate routes. The user can then select one of these routes to begin their journey. These services usually prioritize these candidate routes using a route ranking algorithm, recommending the top few to the user.

[0031] The input to existing route ranking models consists of multiple candidate navigation routes and attribute information for each segment of each candidate route. It's easy to understand that the data for the multiple candidate navigation routes themselves can be a sequence of uniquely identified road segments. The attribute information for each road segment can be set according to actual needs, and can include travel-related characteristics of the road segment itself. The attribute information for a road segment can include at least one of the following: traffic capacity (whether it is prone to congestion), grade (main road, auxiliary road, highway, etc.), lane configuration (one-way road, two-way road, etc.), road surface type (e.g., classified according to road surface material), classification (internal road segment, external road segment), function (e.g., fire-fighting only), whether it is toll-free, intersection type (e.g., classified according to whether there are traffic lights), width, and number of lanes.

[0032] The output of existing route ranking models is the predicted ranking of multiple candidate navigation routes. The training data used in these models consists of the ground truth rankings of these candidate routes. These ground truth rankings can be determined based on the traveler's historical choices and business evaluations. For example, the candidate navigation routes actually chosen by a traveler often have higher priority and are ranked higher. Inputting training samples into the route ranking model outputs the predicted rankings of these candidate navigation routes. It's easy to understand that training the route ranking model is essentially a process of making the predicted rankings output by the model increasingly closer to the ground truth rankings.

[0033] In practical applications, candidate navigation routes used as training samples cannot possibly cover all road segments in the road network. Some candidate navigation routes may not have participated in model training, or have participated very little in it. This means that the attribute information of these road segments has not been learned and understood by the route ranking algorithm. For ease of description, some road segments that have not participated or have participated very little in model training are called low-frequency road segments, and correspondingly, road segments that participate frequently in model training are called high-frequency road segments.

[0034] It should be noted that the reason why low-frequency road sections are not or rarely involved in model training may be that when selecting candidate navigation routes in the training samples, the selected candidate navigation routes need to be those that users have actually used, so as to facilitate the determination of the actual priority ranking of multiple candidate navigation routes.

[0035] However, after training the route ranking algorithm, it needs to be put into application. There may be candidate navigation routes containing low-frequency road segments involved in the ranking. In this case, because the trained route ranking model cannot well understand the feature vectors of low-frequency road segments (determined based on the attribute information of low-frequency road segments), the route ranking model cannot accurately evaluate the value of candidate navigation routes containing low-frequency road segments, and therefore cannot provide an accurate ranking.

[0036] Therefore, in the technical solution provided in this disclosure, considering that the neighbor relationships formed by low-frequency road segments and their surrounding neighboring road segments in the topology of the entire road network may be the same or similar to the neighbor relationships formed by high-frequency road segments and their surrounding neighboring road segments, such same or similar neighbor relationships have similar reference value for navigation routes. Figure 1 This is a schematic diagram of the road network topology provided in this disclosure. For example... Figure 1 As shown, the topology of the low-frequency road segment and its surrounding neighboring road segments, which can be highlighted in bold, is similar to the topology of the high-frequency road segment and its surrounding neighboring road segments, which is highlighted in bold. Furthermore, assuming that the attribute information of the surrounding neighboring road segments of the low-frequency road segment is also similar to that of the surrounding neighboring road segments of the high-frequency road segment, then it can be considered that the neighbor association relationship corresponding to the low-frequency road segment is similar to that corresponding to the high-frequency road segment.

[0037] Therefore, the initial idea conceived in this disclosure is to find a technical implementation method that, for low-frequency road segments and high-frequency road segments with the same or similar neighbor relationships, transfers the experience learned by the route ranking model during training for the high-frequency road segments to the analysis of the low-frequency road segments. This allows the trained route ranking model to accurately evaluate the value of candidate navigation routes containing the low-frequency road segments based on the transferred experience, even when ranking and analyzing candidate navigation routes containing the low-frequency road segments. It should be noted that the two road segments with neighbor relationships are topologically connected.

[0038] In terms of specific technical implementation, the first feature mapping model and the route ranking model are jointly trained, and the data composition of the training samples is changed. The training samples should not only include multiple candidate navigation routes and the attribute information of each segment of each candidate navigation route, but also the attribute information of at least one neighboring segment that is topologically connected to that segment. The first feature mapping model is used to map the feature information of a segment into a feature vector. The feature information of a segment includes not only its own attribute information, but also the attribute information of at least one neighboring segment. In this way, the feature vector of a segment input to the route ranking model contains not only the attribute information of the segment itself, but also the neighbor relationship information of the segment and its surrounding neighbors.

[0039] Thus, although the candidate navigation routes used to train the route ranking model mostly contain high-frequency road segments rather than low-frequency road segments, from the perspective of the trained route ranking model, even if a high-frequency road segment and a low-frequency road segment have different attribute information, if their corresponding neighbor relationships are the same or similar, then the route ranking model will understand the low-frequency road segment in the same way it understands the high-frequency road segment. It will more accurately evaluate the value of the low-frequency road segment to the navigation route, thereby achieving a more accurate ranking of candidate navigation routes containing the low-frequency road segment.

[0040] The technical solutions provided in this disclosure are described in detail below with reference to the accompanying drawings.

[0041] Figure 2 An exemplary flowchart of a model training method is provided, including the following steps:

[0042] S200: Obtain candidate navigation routes, feature information associated with each segment of the candidate navigation routes, and the true value of the ranking of the candidate navigation routes.

[0043] It is easy to understand that the number of candidate navigation routes in step S200 is usually multiple (two or more), thus requiring the sorting of multiple candidate navigation routes.

[0044] Those skilled in the art will understand that model training is essentially an iterative process of adjusting model parameters. Therefore, for Figure 2 The illustrated method typically requires iterative execution until the model training stopping condition is met. This stopping condition could be, for example, a set number of iterations, or it could be that the representation of the difference between the true ranking value and the predicted ranking value is less than a set threshold.

[0045] The associated feature information for each road segment includes not only the segment's attribute information but also the attribute information of at least one neighboring road segment in the road network that is topologically connected to it. The true ranking of multiple candidate navigation routes can be determined by considering actual user feedback, business experience, and other factors.

[0046] Other road segments topologically connected to a given road segment are called neighboring road segments. For example, these can be other road segments directly connected to the given road segment; these are called first-order neighboring road segments. A road segment's neighboring road segments can also be road segments that are not directly connected to the given road segment but are indirectly connected; these indirectly connected road segments are called second-order or higher-order neighboring road segments. The definition of a road segment's neighboring road segments can be flexibly defined according to the needs of the scheme, including only first-order neighboring road segments, or including first-order neighboring road segments, second-order neighboring road segments, higher-order neighboring road segments, etc. This disclosure does not impose any restrictions.

[0047] In this scheme, for each road segment, all neighboring road segments can be selected and their attribute information added to the feature information associated with this road segment. Alternatively, only some neighboring road segments can be selected and their attribute information added to the feature information associated with this road segment.

[0048] S202: Input the feature information associated with the road segments of the candidate navigation route into the first feature mapping model to be trained, and output the first feature vector of each road segment of the candidate navigation route.

[0049] For ease of description, the feature mapping model in step S202 will be referred to as the first feature mapping model. The second feature mapping model will also be mentioned later. The input of the first feature mapping model is the feature information associated with each road segment, and the output is the feature vector of each road segment (called the first feature vector).

[0050] S204: Determine the standard feature vector of each road segment based at least on the first feature vector of each road segment.

[0051] In some embodiments, the first feature vector can be directly used as the standard feature vector.

[0052] In other embodiments, the attribute information of each segment of the candidate navigation route can be input into the second feature mapping model to be trained, and the model can output a second feature vector for each segment of the candidate navigation route. In this way, a standard feature vector for each segment can be determined based on the first and second feature vectors. For example, a weighted sum of the first and second feature vectors can be calculated. This allows the route ranking model to better understand feature vectors determined solely based on the attribute information of each segment, improving the redundancy of the learning process.

[0053] S206: Input at least the candidate navigation routes and the standard feature vector of each segment of the candidate navigation routes into the route ranking model to be trained, and output the ranking prediction values ​​of the multiple candidate navigation routes.

[0054] S208: Adjust the first feature mapping model and the route ranking model with the goal of reducing the difference between the true ranking value and the predicted ranking value.

[0055] Figure 2The input to the route ranking model in the method shown can be understood as a sequence of feature vectors for each segment in the candidate navigation routes (each candidate navigation route corresponds to one sequence), and the output is the predicted ranking value of multiple candidate navigation routes. Through training, the first feature mapping model (and in some embodiments, the second feature mapping model can also be adjusted simultaneously) and the route ranking model can be adjusted simultaneously, so that the feature vector of each segment input to the route ranking model not only contains the attribute information of the segment itself, but also contains the neighbor relationship information of the segment.

[0056] In some embodiments, the plurality of candidate navigation routes, the attribute information of each segment of the candidate navigation routes, and the standard feature vector of each segment of the candidate navigation routes can be specifically input into the route ranking model to be trained. This allows the route ranking model to directly understand the attribute information of each segment itself, further improving the redundancy of the learning process.

[0057] Furthermore, in some embodiments, considering that the travel navigation service periodically or irregularly re-divides the road segments in the road network to update the global road network data, the road segment type corresponding to each road segment of the candidate navigation route can be determined before inputting the feature information associated with each road segment of the candidate navigation route into the first feature mapping model to be trained. The road segment type includes road segment types known based on the global road network data (i.e., the unique identifier of this road segment can be found in the latest version of the global road network data) and road segment types unknown based on the global road network data (i.e., the unique identifier of this road segment cannot be found in the latest version of the global road network data). It is easy to understand that for road segment types with known global road network data, their neighboring road segments are known; while for road segment types with unknown global road network data, their neighboring road segments are unknown.

[0058] Therefore, candidate navigation routes may include road segments whose neighboring road segments are unknown. The feature information associated with each road segment of the candidate navigation route and the road segment type information corresponding to each road segment can be input into a first feature mapping model to be trained, outputting a first feature vector for each road segment of the candidate navigation route. Specifically, for each road segment, if the first feature mapping model determines that the road segment type corresponding to the road segment is a road segment type known based on the global road network data, then it obtains the first feature vector of the road segment based on the feature information associated with the road segment; if the first feature mapping model does not identify the neighboring road segments associated with the road segment, then it can determine that the road segment type corresponding to the road segment is a road segment type unknown based on the global road network data (i.e., a road segment with unknown neighboring road segments), and then it can obtain the first feature vector of the road segment based on the attribute information of the road segment with unknown neighboring road segments.

[0059] The effects of this embodiment are explained here. The continuous updating of global road network data may result in the deletion of some road segments. While the attribute information of these road segments can still be obtained, it is difficult to find their neighboring road segments based on the global road network data. Therefore, for road segments not in the latest version of the global road network data, the first feature mapping model cannot consider the attribute information of their neighboring road segments, but only considers the attribute information of the road segment itself to obtain its first feature vector.

[0060] Furthermore, the aforementioned road segment types may also include zero-padded road segment types generated to adapt to the model input format. Thus, for each road segment, if the first feature mapping model determines that the road segment type corresponding to that road segment is a zero-padded road segment type generated to adapt to the model input format, then the first feature vector of that road segment is determined to be 0.

[0061] It should be noted that software tools (such as TesorFlow) typically have certain requirements regarding the format of the input data when using models. For example, it must be an N*M matrix. In this scheme, N can represent the number of candidate navigation routes in a set, and M can represent the number of road segments in a single candidate navigation route. However, some candidate navigation routes have a large number of road segments, reaching M, while others have a small number, falling short of M. Therefore, a common practice is to pad candidate navigation routes with fewer road segments with meaningless empty road segments, i.e., zero-padding, to bring the total number of road segments to M. Since zero-padding is meaningless, its feature vector can be 0, thus not affecting the accuracy of the model algorithm.

[0062] Furthermore, the aforementioned software tools' requirements for the model input data format mean that if some road segments have a large number of neighboring road segments and others have a small number, then it is necessary to generate some meaningless empty road segments, i.e., padding with zero-neighboring road segments, for the road segments with fewer neighboring road segments. This implies that when the first feature mapping model analyzes the attribute information of a road segment's neighboring road segments, it needs to exclude the influence of padding with zero-neighboring road segments. Therefore, when designing the algorithm structure of the first feature mapping model, a bias correction model (whose function is to exclude the influence of padding with zero-neighboring road segments) can be added to the conventional feature mapping model algorithm structure. The specific implementation details are not an obstacle for those skilled in the art and will not be elaborated further.

[0063] Furthermore, it should be noted that the algorithm structure of the first feature mapping module in this paper can be a graph neural network (GAT network), convolutional neural network, or other algorithm structures that are more suitable for analyzing road network topology.

[0064] Figure 3 An example diagram illustrating the joint training of multiple models is provided. Figure 3 As shown, a first feature mapping model, a second feature mapping model, and a route ranking model are jointly trained. The input to the first feature mapping model can be the attribute information of each road segment and several neighboring road segments of multiple candidate navigation routes, and the output is a first feature vector. The input to the second feature mapping model can be the attribute information of each road segment of multiple candidate navigation routes, and the output is a second feature vector. The input to the route ranking model can be multiple candidate navigation routes (i.e., the unique identifier sequence of road segments corresponding to the candidate navigation routes) and a labeled feature vector determined based on the first and second feature vectors, and the output is the predicted ranking value of the multiple candidate navigation routes (specifically, it can be a score of the candidate navigation routes; the higher the score, the higher the ranking).

[0065] In practical applications, the first feature mapping model, the second feature mapping model, and the route ranking model mentioned above can be trained periodically or irregularly using newly acquired training samples to continuously optimize model performance.

[0066] Figure 4 An exemplary flowchart of a route sorting algorithm is provided, including the following steps:

[0067] S400: Obtain multiple candidate navigation routes to be sorted.

[0068] S402: Obtain the standard feature vector of the road segments included in the candidate navigation route.

[0069] S404: Input at least the plurality of candidate navigation routes and the standard feature vector of each segment of each candidate navigation route into the trained route ranking model, and output the ranking prediction value of the candidate navigation routes to be ranked.

[0070] For step S404, the standard feature vector of each road segment can be obtained using the feature mapping model (first feature mapping model, or first feature mapping model and second feature mapping model) described above.

[0071] In some embodiments, considering the real-time nature of travel navigation services recommending candidate navigation routes to users, the operations of obtaining the first feature vector and the second feature vector using the first feature mapping model and the second feature mapping model can be pre-completed. Specifically, the feature information associated with each road segment in the road network is pre-input into the trained first feature mapping model to construct a first feature table, which includes the first feature vector of each road segment in the road network. Similarly, the attribute information of each road segment in the road network is pre-input into the trained second feature mapping model to construct a second feature table, which includes the second feature vector of each road segment in the road network. If the global road network data is updated, the first feature mapping model and the second feature mapping model can be reused to reconstruct the first feature table and the second feature table.

[0072] Thus, when the travel navigation service recommends candidate navigation routes to users in real time, it can read the first feature vector of each segment of the candidate navigation route from the pre-stored first feature table; and read the second feature vector of each segment of the candidate navigation route from the pre-stored second feature table, and use the weighted sum of the first feature vector and the second feature vector of each segment as the standard feature vector of that segment. Figure 5 An example diagram is provided showing how a travel navigation service sorts multiple candidate navigation routes online in real time.

[0073] Furthermore, in some embodiments, if the input to the route ranking model includes not only multiple candidate navigation routes and the standard feature vector of each segment of the candidate navigation routes, but also the attribute information of each segment of each candidate route, then the input to the route ranking model during application will also include multiple candidate navigation routes, the attribute information of each segment of each candidate route, and the standard feature vector of each segment of the candidate navigation routes.

[0074] in addition, Figure 4 The route ranking model used in the method shown can be either a route ranking model trained in the prior art, or a model developed through [other methods]. Figure 3 The route ranking model trained by the method shown.

[0075] This disclosure also provides a model training apparatus, including:

[0076] The acquisition module acquires candidate navigation routes, feature information associated with road segments included in the candidate navigation routes, and the ranking truth value of the candidate navigation routes; wherein, the feature information associated with road segments includes: attribute information of the road segment, and attribute information of at least one neighboring road segment in the road network that is topologically connected to the road segment.

[0077] The mapping module inputs the feature information associated with the road segments of the candidate navigation route into the first feature mapping model to be trained, and outputs the first feature vector of each road segment of the candidate navigation route; at least based on the first feature vector of each road segment, the standard feature vector of the road segment is determined.

[0078] The ranking module inputs at least the candidate navigation routes and the standard feature vectors of each segment of the candidate navigation routes into the route ranking model to be trained, and outputs the ranking prediction values ​​of the candidate navigation routes.

[0079] The adjustment module adjusts the first feature mapping model and the route ranking model with the goal of reducing the difference between the true ranking value and the predicted ranking value.

[0080] This disclosure also provides a route sorting device, including:

[0081] The acquisition module acquires multiple candidate navigation routes to be sorted and obtains the standard feature vectors of the road segments included in the candidate navigation routes.

[0082] The sorting module inputs at least the multiple candidate navigation routes and the standard feature vector of each segment of each candidate navigation route into the trained route sorting model, and outputs the sorting prediction value of the candidate navigation routes to be sorted.

[0083] This disclosure also provides a program product that executes the methods of embodiments of this disclosure during runtime.

[0084] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods of embodiments of this disclosure.

[0085] This disclosure also provides a computing device, including a memory and a processor; the memory is used to store computer instructions that can be executed on the processor, and the processor is used to implement the methods of the embodiments of this disclosure when executing the computer instructions.

[0086] Figure 6 This is a schematic diagram of the structure of a computing device provided in this disclosure. The computing device 15 may include, but is not limited to, a processor 151, a memory 152, and a bus 153 connecting different system components (including the memory 152 and the processor 151).

[0087] The memory 152 stores computer instructions that can be executed by the processor 151, enabling the processor 151 to perform the methods of any embodiment of this disclosure. The memory 152 may include a random access memory (RAM) 1521, a cache memory 1522, and / or a read-only memory (ROM) 1523. The memory 152 may also include a program tool 1525 having a set of program modules 1524, including but not limited to: an operating system, one or more application programs, other program modules, and program data. One or more combinations of these program modules may include an implementation of a network environment.

[0088] Bus 153 may include, for example, a data bus, an address bus, and a control bus. The computing device 15 can also communicate with external devices 155 via I / O interface 154, such as a keyboard or a Bluetooth device. The computing device 150 can also communicate with one or more networks via network adapter 156, such as a local area network (LAN), a wide area network (WAN), or a public network. As shown, network adapter 156 can also communicate with other modules of the computing device 15 via bus 153.

[0089] Furthermore, although the operations of the methods disclosed herein are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0090] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

[0091] The methods described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. A typical implementing device is a computer. Specifically, the computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0092] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this specification, the functions of each module can be implemented in one or more software and / or hardware.

[0093] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0095] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0096] The terminology used in the various embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of this specification. The singular forms “a,” “described,” and “the” as used in the various embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0097] It should be understood that although the terms first, second, third, etc., may be used to describe various information in various embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of various embodiments of this specification, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0098] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the method embodiments are basically similar to the method embodiments and are therefore described simply; relevant parts can be referred to the descriptions of the method embodiments. The method embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. When implementing the embodiments of this specification, the functions of each module can be implemented in one or more software and / or hardware. Alternatively, some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0099] The above description is merely a preferred embodiment of the various embodiments of this specification and is not intended to limit the various embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the various embodiments of this specification should be included within the protection scope of the various embodiments of this specification.

Claims

1. A model training method, comprising: Obtain candidate navigation routes, feature information associated with road segments included in the candidate navigation routes, and the ranking truth value of the candidate navigation routes; wherein, the feature information associated with road segments includes: the attribute information of the road segment, and the attribute information of at least one neighboring road segment in the road network that is topologically connected to the road segment; The feature information associated with the road segments of the candidate navigation route is input into the first feature mapping model to be trained, and the first feature vector of each road segment of the candidate navigation route is output. At least based on the first feature vector of each road segment, determine the standard feature vector of that road segment; The candidate navigation routes and the standard feature vectors of each segment of the candidate navigation routes are input into the route ranking model to be trained, and the ranking prediction values ​​of the candidate navigation routes are output. With the goal of reducing the difference between the true ranking value and the predicted ranking value, the first feature mapping model and the route ranking model are adjusted.

2. The method as described in claim 1, wherein, The neighboring road segments that are topologically connected to each road segment include: Other road sections that this section connects to.

3. The method as described in claim 1, wherein, The method further includes: The attribute information of each segment of the candidate navigation route is input into the second feature mapping model to be trained, and the second feature vector of each segment of the candidate navigation route is output. At least based on the first feature vector of each road segment, determine the standard feature vector of that road segment, including: The weighted sum of the first and second feature vectors of the road segment is used as the standard feature vector of the road segment; The adjustment of the first feature mapping model and the route ranking model includes: Adjust the first feature mapping model, the second feature mapping model, and the route sorting model.

4. The method of claim 3, wherein inputting at least the candidate navigation route and the standard feature vector of each segment of the candidate navigation route into the route ranking model to be trained includes: The candidate navigation routes, the attribute information of each segment of the candidate navigation routes, and the standard feature vector of each segment of the candidate navigation routes are input into the route ranking model to be trained.

5. The method according to any one of claims 1-4, wherein, If the candidate navigation route includes road segments with unknown neighboring road segments, and the feature information associated with the unknown neighboring road segments is input into the first feature mapping model to be trained, and the first feature mapping model to be trained fails to identify the neighboring road segments associated with the road segment, the method further includes: Based on the attribute information of the unknown neighboring road segment, the first feature vector of the road segment is obtained.

6. The method of claim 5, wherein, For each road segment, if the first feature mapping model determines that the road segment is a zero-padding road segment generated to adapt to the model input format, then the first feature vector of the road segment is determined to be 0.

7. A route sorting method, comprising: Retrieve multiple candidate navigation routes to be sorted; Obtain the standard feature vectors of the road segments included in the candidate navigation route; The The standard feature vector is determined at least based on the first feature vector of each road segment. The first feature vector is obtained by inputting the feature information associated with the road segments of the candidate navigation route into a trained first feature mapping model. The feature information associated with the road segment includes: the attribute information of the road segment, and the attribute information of at least one neighboring road segment in the road network that is topologically connected to the road segment. The multiple candidate navigation routes to be sorted, and the standard feature vector of each segment of each candidate navigation route, are input into the trained route sorting model, and the sorting prediction value of the candidate navigation routes to be sorted is output.

8. The method of claim 7, wherein obtaining the standard feature vector of each segment of each candidate navigation route includes: Read the first feature vector of each segment of each candidate navigation route from the pre-stored first feature table; Read the second feature vector of each segment of each candidate navigation route from the pre-stored second feature table; The weighted sum of the first and second feature vectors of the road segment is used as the standard feature vector of the road segment.

9. The method of claim 8, further comprising: The feature information associated with each road segment in the road network is input into the trained first feature mapping model to construct and store the first feature table, which includes the first feature vector of each road segment in the road network. The feature information associated with each road segment includes: the attribute information of the road segment, and the attribute information of at least one neighboring road segment in the road network that is topologically connected to the road segment. The attribute information of each road segment in the road network is input into the trained second feature mapping model to construct and store a second feature table, which includes the second feature vector of each road segment in the road network.

10. The method of claim 7, wherein at least the plurality of candidate navigation routes to be ranked, and the standard feature vector of each segment of each candidate navigation route, are input into a trained route ranking model, and the ranking prediction value of the candidate navigation routes to be ranked is output, comprising: The multiple candidate navigation routes to be sorted, the attribute information of each segment of each candidate navigation route, and the standard feature vector of each segment of each candidate navigation route are input into the trained route sorting model, and the sorting prediction value of the multiple candidate navigation routes to be sorted is output.

11. The method according to any one of claims 7-10, wherein, The trained route ranking model is obtained by pre-training using any one of the methods described in weights 1-6.

12. A computing device, comprising a memory and a processor; the memory being configured to store computer instructions executable on the processor, the processor being configured to implement the method of any one of claims 1-11 when executing the computer instructions.

Images