Method and device for multi-level evaluation of route matching degree and electronic equipment

Abstract

The application discloses a multi-level evaluation method and device for route matching degree and electronic equipment, and relates to the field of data processing. The method comprises the following steps: acquiring end point airport information of a target route, wherein the end point airport information comprises airport names and latitude and longitude coordinates of start point airports and end point airports; screening a candidate route sample set matched with the target route from an existing route library; extracting multi-dimensional features between the target route and each candidate route in the candidate route sample set according to the end point airport information of the target route, wherein the multi-dimensional features comprise basic features and cross features; inputting the multi-dimensional features into a pre-trained multi-level progressive evaluation model, and outputting matching degrees between the target route and each candidate route by the multi-level progressive evaluation model. The application solves the technical problem that the prior art cannot automatically construct quantifiable matching features between routes for new civil aviation routes without historical operation data.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to a multi-level evaluation method, apparatus, and electronic equipment for route matching. Background Technology

[0002] In the civil aviation sector, predicting the operational efficiency of newly opened routes is a crucial step for airlines in network planning and capacity allocation. Currently, the evaluation of new routes mainly relies on qualitative analysis based on expert experience, analogy methods based on simple indicator comparisons, or statistical forecasting models.

[0003] However, newly opened routes lack historical operating data, making it impossible to directly use historical operating indicators such as passenger volume and load factor to construct features. This makes it difficult to directly apply data-driven evaluation methods. Existing technologies cannot address the lack of historical operating data for newly opened civil aviation routes. The accuracy and consistency of evaluation results for these routes rely on human experience, resulting in a lack of precision and interpretability.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a multi-level evaluation method, apparatus, and electronic device for route matching, which at least solves the technical problem that the prior art cannot automatically construct quantifiable route matching characteristics for newly opened civil aviation routes without historical operating data.

[0006] According to one aspect of the embodiments of this application, a multi-level evaluation method for route matching degree is provided, comprising: obtaining endpoint airport information of a target route, the endpoint airport information including the airport names and latitude and longitude coordinates of the origin airport and the destination airport; filtering a candidate route sample set that matches the target route from an existing route database; extracting multi-dimensional features between the target route and each candidate route in the candidate route sample set based on the endpoint airport information of the target route, wherein the multi-dimensional features include basic features and cross features, the basic features including at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes respectively, and the cross features including at least the geospatial similarity features between the target route and the candidate routes; inputting the multi-dimensional features into a pre-trained multi-level progressive evaluation model, and outputting the matching degree between the target route and each candidate route by the multi-level progressive evaluation model; wherein the multi-level progressive evaluation model is a neural network model trained by supervised learning using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels.

[0007] According to another aspect of the embodiments of this application, a multi-level evaluation device for route matching degree is also provided, comprising: an acquisition unit, configured to acquire endpoint airport information of a target route, the endpoint airport information including the airport names and latitude and longitude coordinates of the origin airport and the destination airport; a filtering processing unit, configured to filter out a sample set of candidate routes that match the target route from an existing route database; an extraction processing unit, configured to extract multi-dimensional features between the target route and each candidate route in the sample set of candidate routes based on the endpoint airport information of the target route, wherein the multi-dimensional features include basic features and cross features, the basic features including at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes, and the cross features including at least the geospatial similarity features between the target route and the candidate routes; and a model processing unit, configured to input the multi-dimensional features into a pre-trained multi-level progressive evaluation model, and output the matching degree between the target route and each candidate route from the multi-level progressive evaluation model; wherein the multi-level progressive evaluation model is a neural network model trained by supervised learning using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels.

[0008] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described multi-level evaluation method for route matching.

[0009] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described multi-level evaluation method for route matching.

[0010] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described multi-level evaluation method for route matching degree.

[0011] In this embodiment, multi-dimensional features are input into a pre-trained multi-level progressive evaluation model, which outputs the matching degree between the target route and each candidate route. The multi-level progressive evaluation model is a neural network model trained through supervised learning, using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels. This pre-trained neural network model can process the input multi-dimensional features and automatically output the matching degree. Since the model has learned the mapping relationship between the multi-dimensional features of existing routes and passenger volume similarity during the training phase, for target routes without historical operational data, only the multi-dimensional features of the target route and candidate routes need to be provided for the model to automatically calculate the matching degree. Thus, this application embodiment obtains the static endpoint airport information of the target route, automatically extracts multi-dimensional features of the candidate routes based on the information, and inputs the multi-dimensional features into a neural network model pre-trained based on existing route data. The model automatically outputs the matching degree, which is beneficial to automatically constructing quantifiable inter-route matching features without relying on the historical operating data of the target route itself. This solves the technical problem that the existing technology cannot automatically construct quantifiable inter-route matching features for newly opened civil aviation routes without historical operating data. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0013] Figure 1 This is a schematic diagram of an optional multi-level evaluation method for route matching degree according to an embodiment of this application;

[0014] Figure 2 This is an architecture diagram of an optional multi-level evaluation method for route matching degree according to an embodiment of this application;

[0015] Figure 3 This is a flowchart of an optional multi-level evaluation method for route matching according to an embodiment of this application;

[0016] Figure 4 This is a schematic diagram of an optional multi-level evaluation device for route matching according to an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] According to an embodiment of this application, a method embodiment for a multi-level evaluation method of route matching degree is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0020] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0021] According to the embodiments of this application, a data processing system can be used as the execution subject of the multi-level evaluation method for route matching degree in the embodiments of this application. The system can be a software system or an embedded system combining software and hardware. Of course, the execution subject of the method in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the execution subject.

[0022] Figure 1 This is a schematic diagram of a multi-level evaluation method for route matching according to an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps:

[0023] Step S101: Obtain the endpoint airport information of the target route. The endpoint airport information includes the airport name and latitude and longitude coordinates of the origin airport and the destination airport.

[0024] For example, endpoint airport information can refer to static attribute data describing the two airports connected by the target route, including the official names of the origin airport and the destination airport, as well as their respective latitude and longitude coordinates in a global geographic coordinate system. Airport latitude and longitude coordinates can be represented as the longitude and latitude values ​​of the geometric center point of the airport runway or the center point of the terminal building on the Earth's surface. By obtaining the aforementioned endpoint airport information, the data processing system can retrieve routes with the same endpoints as the target route from an existing route database based on the airport name, or it can calculate the spatial distance and direction vectors between airports or routes based on latitude and longitude coordinates.

[0025] In some embodiments, the data processing system can pre-build an airport basic information database. This database uses the airport name as the primary key and stores each airport's unique identifier, city, latitude and longitude coordinates, and annual passenger throughput level. When the data processing system receives the origin and destination airport names of a target route, it can query the airport basic information database using string matching to retrieve the latitude and longitude coordinates corresponding to the origin and destination airport names, respectively. This embodiment allows for the use of a local database to retrieve endpoint airport information, which improves data acquisition efficiency in batch route evaluation scenarios and reduces reliance on external network services.

[0026] In other embodiments, after receiving the origin and destination airport names of the target route, the data processing system can invoke a geocoding service interface, sending the origin and destination airport names as query parameters to a geographic information service platform to obtain the latitude and longitude coordinate data corresponding to the origin and destination airport names, respectively, returned by the geographic information service platform. The data processing system can perform validity checks on the returned latitude and longitude coordinate data, including verifying whether the latitude and longitude values ​​are within a predefined reasonable range; after passing the verification, the latitude and longitude coordinate data is associated and stored with the corresponding airport name. This embodiment dynamically obtains geographic coordinate data through an external interface, which is beneficial for obtaining updated coordinate information without needing to maintain a local airport database in advance. It is suitable for route network planning scenarios where airport names are subject to updates or the evaluation scope is wide.

[0027] Step S102: Select a set of candidate routes that match the target route from the existing route database.

[0028] For example, the existing route database can refer to a structured dataset storing multiple operating routes, each of which includes at least the origin airport name, the destination airport name, and the corresponding latitude and longitude coordinates of the endpoint airports. The candidate route sample set can refer to a subset of routes selected from the existing route database that are comparable to the target route in terms of spatial or operational attributes.

[0029] In some embodiments, the data processing system can obtain the latitude and longitude coordinates of the origin airport and the destination airport of the target route, then traverse each existing route in the existing route database, calculate the first spatial distance between the origin airport of the target route and the origin airport of each existing route, and the second spatial distance between the destination airport of the target route and the destination airport of each existing route, and add the first spatial distance and the second spatial distance to obtain a comprehensive spatial distance index. The data processing system sorts the existing routes in ascending order of comprehensive spatial distance index, and selects the top preset number of existing routes with the smallest comprehensive spatial distance index to form a candidate route sample set. This embodiment, through spatial proximity filtering, is beneficial for obtaining reference routes that are geographically similar to the target route, and is suitable for route network evaluation scenarios with strong geographical clustering.

[0030] In other embodiments, the data processing system can obtain the throughput class of the origin airport and the destination airport of the target route. The throughput class can be determined based on the airport's annual passenger throughput. The data processing system filters existing routes from the existing route database that meet the following conditions: the absolute value of the difference between the throughput class of the origin airport of the existing route and the throughput class of the origin airport of the target route is less than or equal to a preset threshold, and the absolute value of the difference between the throughput class of the destination airport of the existing route and the throughput class of the destination airport of the target route is also less than or equal to the preset threshold. The data processing system constructs a candidate route sample set from all the filtered existing routes. This embodiment uses airport size class for filtering, which is beneficial for obtaining reference routes with similar positioning, thereby reducing the interference of differences in operating modes between routes of different sizes on subsequent matching evaluation.

[0031] Step S103: Based on the endpoint airport information of the target route, extract multi-dimensional features between the target route and each candidate route in the candidate route sample set. The multi-dimensional features include basic features and cross features. The basic features include at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes, and the cross features include at least the geospatial similarity features between the target route and the candidate routes.

[0032] For example, multi-dimensional features can refer to a set of multiple types of feature data used to quantify the similarity between target routes and candidate routes. Basic features can refer to the endpoint attribute features of the target or candidate routes themselves. Cross-features can refer to features obtained by calculating the difference or similarity between a certain attribute of the target route and the corresponding attribute of the candidate route. Geospatial similarity features can refer to quantitative indicators reflecting the degree of proximity of two routes in spatial geometry.

[0033] In some embodiments, the data processing system extracts the latitude and longitude coordinates of the origin airport and the destination airport of the target route from the endpoint airport information, using them as two basic features of the target route. For each candidate route in the candidate route sample set, the data processing system extracts the latitude and longitude coordinates of the origin airport and the destination airport of that candidate route, using them as two basic features of that candidate route. The data processing system can combine the four latitude and longitude coordinate values ​​of the origin airport, destination airport, and candidate route into a basic feature vector. This embodiment, by directly using the latitude and longitude coordinates of the endpoint airport as basic features, helps to completely preserve the original spatial location information of the target route and candidate routes, facilitating subsequent calculations of spatial distance or direction vectors.

[0034] In other embodiments, the data processing system calculates the flight distance of the target route and the direction vector from the origin to the destination airport based on the latitude and longitude coordinates of the origin and destination airports. For each candidate route in the candidate route sample set, the data processing system calculates the flight distance and the direction vector from the origin to the destination airport based on the latitude and longitude coordinates of the origin and destination airports. The data processing system calculates the absolute value of the ratio or difference between the flight distance of the target route and the flight distance of the candidate route to obtain the route distance similarity feature; simultaneously, it calculates the cosine of the angle between the direction vector of the target route and the direction vector of the candidate route to obtain the route direction similarity feature; the route distance similarity feature and the route direction similarity feature are used together as the geospatial similarity feature. This embodiment uses flight distance and direction vector to generate geospatial similarity features, which is beneficial for comprehensively characterizing the spatial matching degree of two routes from two dimensions: route length and direction, thereby improving the ability to represent the similarity of route operation modes.

[0035] Step S104: Input the multi-dimensional features into the pre-trained multi-level progressive evaluation model, and output the matching degree between the target route and each candidate route by the multi-level progressive evaluation model; wherein, the multi-level progressive evaluation model is a neural network model trained by supervised learning, using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels.

[0036] For example, a multi-level progressive evaluation model can refer to a machine learning model based on a neural network with multiple progressive computational levels. The matching degree can be a value ranging from 0 to 1, representing the similarity between the target route and the candidate route in terms of passenger traffic. Training samples can refer to feature vectors extracted from an existing route database, composed of multi-dimensional features of two existing routes. Training labels can refer to similarity values ​​calculated based on the actual passenger volume of the two existing routes. The data processing system inputs the extracted multi-dimensional features into the pre-trained multi-level progressive evaluation model, which automatically outputs the matching degree value according to the computational logic of each internal level.

[0037] In some embodiments, the data processing system combines the multi-dimensional features of the target route and a candidate route into an input feature vector. The system then sequentially passes this input feature vector through a multi-level progressive evaluation model, specifically through a first, second, and third progressive network layer. The first progressive network layer outputs a basic matching degree; the second progressive network layer receives this basic matching degree and, together with operational features, calculates and outputs a revised matching degree; and the third progressive network layer receives the revised matching degree and, together with other features, calculates and outputs a final matching degree. The data processing system uses the final matching degree output by the third progressive network layer as the matching degree between the target route and the candidate route. This embodiment refines the matching degree evaluation results progressively through three progressive network layers, which helps improve the accuracy and hierarchy of the matching degree prediction.

[0038] In other embodiments, the data processing system acquires a pre-constructed training dataset containing multi-dimensional features of multiple existing route pairs and passenger volume similarity labels between these pairs. The data processing system uses the multi-dimensional features as input and the passenger volume similarity labels as supervisory signals to iteratively train an initialized neural network model. During training, the system adjusts the weight parameters of the neural network model until the prediction error on the validation set meets preset conditions. The resulting neural network model is then identified as the pre-trained multi-level progressive evaluation model. This embodiment utilizes historical operational data from existing routes to train the model, which helps the model learn the inherent mapping between multi-dimensional features and passenger volume similarity between routes, thereby obtaining relatively reliable matching prediction results when applied to new routes.

[0039] In some optional embodiments, the data processing system may not pre-set weight coefficients for each feature in the multi-dimensional features, and the importance coefficients of each feature may be automatically learned by the multi-level progressive evaluation model during training by optimizing the loss function.

[0040] In some optional embodiments, selecting a candidate route sample set matching the target route from an existing route database includes: the data processing system can perform a two-stage business rule screening on the target route to select a candidate route sample set matching the target route from the existing route database. The two-stage business rule screening includes the following first stage and second stage executed sequentially: The first stage includes: scale matching screening based on airport throughput level to obtain the endpoint airport level of the target route, and selecting m routes from the existing route database whose endpoint airport level is the same as or one level different from the endpoint airport level of the target route to form a first route candidate set, wherein the endpoint airport level includes the level of the origin airport and the level of the destination airport; The second stage includes: detecting the spatial distance between each candidate route in the first route candidate set and the target route, and selecting the k routes with the smallest spatial distance as the candidate route sample set, wherein k is a positive integer less than or equal to m.

[0041] For example, the two-stage business rule screening can refer to a sample screening process that sequentially executes two stages: scale matching screening and spatial distance screening. Airport throughput level can refer to multiple levels divided according to the annual passenger throughput of an airport, with each level corresponding to a throughput range. Endpoint airport level can refer to the throughput level of the origin airport and the destination airport of the target route. The first route candidate set can refer to the set of routes obtained after the first stage of screening that have the same or different endpoint airport levels as the target route's endpoint airport level. Spatial distance can refer to the sum of the geographical distances between the origin airport and the destination airport of the target route and the candidate routes. Through the above two-stage screening, the data processing system can first retain routes with similar airport scales in the existing route database, and then select the spatially nearest routes from them. This is beneficial for obtaining a candidate route sample set that is highly comparable to the target route in terms of flight positioning and geographical distribution, thereby improving the accuracy and computational efficiency of subsequent matching degree evaluation. A two-stage business rule screening method is adopted. First, route pairs with similar scale are retained based on airport throughput level. Then, candidate samples with the smallest spatial distance are selected based on geographical proximity. This helps to reduce noise interference between routes of different scales and provides route pair data with strong comparability for subsequent feature calculation.

[0042] In some embodiments, the data processing system pre-acquires annual passenger throughput data of civil transport airports and divides each airport into six levels according to a preset throughput range. For example, airports with annual passenger throughput greater than or equal to a first threshold are classified as Level 1, airports with annual passenger throughput greater than or equal to a second threshold but less than the first threshold are classified as Level 2, and so on. The data processing system acquires the origin airport level and destination airport level of the target route, traverses each existing route in the existing route database, and determines whether the origin airport level of the existing route is the same as or differs by one level from the origin airport level of the target route. Simultaneously, it determines whether the destination airport level of the existing route is the same as or differs by one level from the destination airport level of the target route. If both conditions are met, the existing route is included in the first route candidate set. This embodiment uses airport throughput level for scale matching and screening, which helps to exclude routes with excessively different passenger volume from the target route, thereby reducing the interference of differences in operating models between routes of different scales on subsequent evaluations.

[0043] In other embodiments, after the data processing system obtains the first candidate route set, for each candidate route in the first candidate route set, the great circle distance calculation formula is used to calculate the arc distance between the airports on the Earth's surface: first, the first spherical distance between the starting airport of the target route and the starting airport of the candidate route is calculated; then, the second spherical distance between the ending airport of the target route and the ending airport of the candidate route is calculated; finally, the first spherical distance and the second spherical distance are added together to obtain the spatial distance between the candidate route and the target route. The data processing system sorts all candidate routes in the first candidate route set in ascending order of spatial distance, and selects the top k candidate routes in the sorting result as the candidate route sample set, where k is a preset positive integer. This embodiment further uses spatial proximity filtering on the basis of scale matching, which is beneficial to obtaining reference routes that are geographically similar to the target route, thereby improving the spatial comparability and representativeness of the candidate route sample set.

[0044] In some optional embodiments, the data processing system obtains information such as the city's permanent resident population, road network density, high-speed rail network competition intensity index, geographical region, dominant industry category, main travel destinations of air passengers, air demand intensity level, and population mobility intensity level of the destination cities by calling natural language processing interfaces. The data processing system performs random sampling verification on the acquired data, comparing it with standard statistical data. When the verification pass rate is lower than a preset threshold, it triggers re-acquisition or manual review. This embodiment helps ensure the automated acquisition of basic urban information and data quality.

[0045] Optionally, the data processing system's random sampling verification of the acquired data also includes verifying the accuracy of the rate of change characteristics, which refer to the magnitude of changes in indicators of the same city at different time dimensions; if the verification finds that the data is inconsistent, the data processing system updates and corrects the erroneous items based on authoritative source data.

[0046] In some optional embodiments, the data processing system obtains basic information about the destination cities of the flight routes by calling a language model application programming interface (API). The data processing system constructs a structured instruction containing the name of the city where the airport is located and the required indicator items, and stores the data returned by the API as a structured file. The format of the basic information extracted for each airport is shown in Table 1.

[0047] Table 1

[0048]

[0049] For each airport, the extracted basic information includes: airport name, longitude and latitude of the city, city's permanent population, road network density, geographical region of the city, dominant industry category of the city, main travel destinations of air passengers, city's air travel demand intensity level, city's population mobility intensity level, and city's high-speed rail competition index. Specifically, the geographical region of the city includes Northeast China, North China, Central China, South China, East China, Northwest China, and Southwest China; the dominant industry category of the city is divided into four levels: B1, B2, B3, and B4; the main travel destinations of air passengers are divided into four levels: C1, C2, C3, and C4; the air travel demand intensity level is divided into four levels: D1, D2, D3, and D4; and the population mobility intensity level is divided into four levels: H, M, L, and N. This basic information is used to construct the endpoint attribute features of the air routes. The data processing system verifies data accuracy through random sampling. Data is considered reasonable when the sampling verification pass rate is greater than 95%. When data does not match authoritative sources or is incomplete, manual verification is performed, and the data acquisition and verification process is repeated. For example, specific level descriptions are shown in Table 2:

[0050] Table 2

[0051]

[0052]

[0053] In some optional embodiments, the data processing system calls a language model application programming interface (API) to construct a structured instruction containing the name of the city where the airport is located and the required indicators. These indicators include airport longitude, airport latitude, city resident population, city road network density, city regional affiliation, city industry type, city passenger type, city air travel demand intensity, city population mobility, and city high-speed rail competition index. The data processing system parses and stores the data returned by the API. The air travel demand intensity is divided into four levels based on urbanization rate and population density: Level 1 corresponds to an urbanization rate greater than 60% or a population density greater than 500 people per square kilometer; Level 2 corresponds to an urbanization rate of 50% to 60% or a population density of 200 to 500 people per square kilometer; Level 3 corresponds to an urbanization rate of 40% to 50% or a population density of 50 to 200 people per square kilometer; and Level 4 corresponds to an urbanization rate less than 40% or a population density less than 50 people per square kilometer. The data processing system verifies data accuracy through random sampling. Data is considered reasonable when the sampling pass rate is greater than 95%; otherwise, it undergoes manual review or is re-acquired.

[0054] For example, the data processing system performs the first stage of the two-stage business rule screening as follows: scale matching screening based on airport throughput level. The data processing system divides civil transport airports into six levels according to their annual passenger throughput: Level 1 corresponds to an annual passenger throughput of 50 million or more, Level 2 corresponds to an annual passenger throughput of 30 million or more but less than 50 million, Level 3 corresponds to an annual passenger throughput of 10 million or more but less than 30 million, Level 4 corresponds to an annual passenger throughput of 2 million or more but less than 10 million, Level 5 corresponds to an annual passenger throughput of 500,000 or more but less than 2 million, and Level 6 corresponds to an annual passenger throughput of less than 500,000.

[0055] For the target route, let the originating airport be classified as follows: The destination airport is classified as The data processing system filters candidate routes from the existing route database that meet the following criteria. Airport classification of candidate routes originating from and The absolute value of the difference is less than or equal to the preset grade difference threshold, and the destination airport grade of the candidate route is... and The absolute value of the difference is also less than or equal to the preset grade difference threshold. The data processing system will use the preset grade difference threshold. Setting it to 1 means only retaining candidate routes where the airports at both ends are of the same or different level. All candidate routes that meet the conditions are combined into the first candidate route set.

[0056] The data processing system performs the second phase as follows: spatial filtering based on geographic proximity. For each candidate route in the first route candidate set, the data processing system uses the great circle distance formula to calculate the spherical distance between the fields. The great circle distance can be calculated with reference to the following formula (1):

[0057] Formula (1);

[0058] in, Represents the Earth's average radius. Representing dimension, , and These represent the latitude values ​​of the two airports, in radians. Indicates longitude. , and These represent the longitude values ​​of the two airports, in radians.

[0059] The data processing system uses the aforementioned great circle distance formula to calculate the first distance between the origin airport of the target route and the origin airport of the candidate route, and the second distance between the destination airport of the target route and the destination airport of the candidate route. Then, the first distance and the second distance are added together to obtain the spatial distance between the candidate route and the target route. The formula for calculating the spatial distance is as follows:

[0060] Formula (2);

[0061] in, Indicates the target route. Indicates candidate routes, Indicates the origin airport of the target route. Indicates the origin airport of the candidate route. Indicates the destination airport of the target route. Indicates the destination airport of the candidate route.

[0062] The data processing system sorts all candidate routes in the first route candidate set in ascending order of spatial distance, and selects the top k candidate routes with the smallest spatial distance as the candidate route sample set, where k is a positive integer less than or equal to the total number of candidate routes in the first route candidate set.

[0063] By filtering out the interference of differences in operating modes between routes of different volumes through airport throughput levels, and then selecting spatially adjacent routes as samples through great circle distance calculation, the accuracy and calculation efficiency of subsequent matching degree assessment are improved.

[0064] In some optional embodiments, before selecting the k routes with the smallest spatial distance as the candidate route sample set, the method further includes: the data processing system can obtain the spatial distances between all candidate routes and the target route in the first route candidate set to form a spatial distance sequence; calculate the mean and standard deviation of the spatial distance sequence; determine a distance threshold based on the mean and standard deviation of the spatial distance sequence, wherein the distance threshold is equal to the sum of the mean and n times the standard deviation, and n is a preset positive integer coefficient; count the number of candidate routes in the first route candidate set whose spatial distance is less than or equal to the distance threshold, and use this number as the value of k.

[0065] For example, the spatial distance sequence can refer to an ordered or unordered set composed of the spatial distance values ​​between each candidate route and the target route in the first route candidate set. The mean of the spatial distance sequence can refer to the arithmetic mean of all spatial distance values ​​in the spatial distance sequence. The standard deviation of the spatial distance sequence can refer to the statistical measure of the degree of deviation of each spatial distance value from the mean. The distance threshold can refer to a reference boundary value used to screen candidate routes, which is calculated by adding n times the standard deviation to the mean, where n is a preset positive integer coefficient. The data processing system dynamically determines the value of k through the above statistical calculations, so that the number of candidate routes finally selected is related to the distribution characteristics of spatial distance in the first route candidate set. Using the mean and standard deviation to automatically adapt to the dispersion of spatial distance to determine the screening threshold is beneficial for dynamically adjusting the size of the candidate route sample set according to the distribution characteristics of the data itself, thereby avoiding the problem of too many or too few samples due to the use of a fixed k value.

[0066] In some embodiments, after the data processing system calculates the mean and standard deviation of all spatial distances in the first candidate route set, it sets a preset positive integer coefficient n to 1. The data processing system calculates the mean plus one standard deviation and uses this value as the distance threshold. The data processing system iterates through the spatial distance of each candidate route in the first candidate route set, determines whether the spatial distance is less than or equal to the distance threshold, counts the number of candidate routes that meet this condition, and uses the counted number as the value of k. This embodiment sets the distance threshold to the mean plus one standard deviation, which helps to retain candidate routes whose spatial distances are distributed within one standard deviation of the mean, thereby obtaining a statistically more concentrated sample subset.

[0067] In other embodiments, after the data processing system calculates the mean and standard deviation of all spatial distances in the first candidate route set, it dynamically adjusts the value of a preset positive integer coefficient n based on the total number of candidate routes in the first candidate route set. For example, when the total number of candidate routes in the first candidate route set is greater than a first quantity threshold, the data processing system sets n to 2; when the total number of candidate routes in the first candidate route set is less than or equal to the first quantity threshold, the data processing system sets n to 1. The data processing system calculates the mean plus n times the standard deviation, uses this value as the distance threshold, counts the number of candidate routes whose spatial distance is less than or equal to the distance threshold, and uses this number as the value of k. This embodiment adaptively adjusts the value of n based on the total number of candidate routes, which is beneficial for obtaining a relatively balanced sample screening granularity in candidate route sets of different sizes, thereby improving the generalization applicability of the method.

[0068] In some optional embodiments, detecting the spatial distance between each candidate route in the first route candidate set and the target route includes: the data processing system can obtain a first distance between the origin airport of the target route and the origin airport of the candidate route; obtain a second distance between the destination airport of the target route and the destination airport of the candidate route; and add the first distance and the second distance to obtain the spatial distance between the candidate route and the target route.

[0069] For example, the first distance can refer to the geographical straight-line distance or spherical arc distance between the origin airport of the target route and the origin airport of the candidate route. The second distance can refer to the geographical straight-line distance or spherical arc distance between the destination airport of the target route and the destination airport of the candidate route. The spatial distance between the candidate route and the target route can be the sum of the first distance and the second distance, used to comprehensively measure the overall spatial proximity of the two routes in both the origin and destination dimensions. By calculating and summing the distances between the origin airports and the destination airports respectively, the data processing system can quantify the spatial distribution difference between the target route and the candidate route from the perspective of the endpoints. By decomposing the spatial proximity of routes into two independent components, origin proximity and destination proximity, it is beneficial to more precisely reflect the similarity of the two routes in the overall spatial layout, thereby providing a comprehensive spatial distance indicator for subsequent sample selection.

[0070] In some embodiments, the data processing system obtains the latitude and longitude coordinates of the originating airport of the target route and the originating airport of the candidate route, calculates the spherical arc length distance between these two coordinates using the semi-versus formula, and uses the calculation result as the first distance. The data processing system calculates the spherical arc length distance between the destination airport of the target route and the destination airport of the candidate route in the same manner, using this as the second distance. The data processing system adds the calculated first distance and second distance, and uses the sum as the spatial distance between the candidate route and the target route. This embodiment utilizes the spherical arc length distance to calculate the true geographical distance between airports, which helps improve the accuracy of spatial distance measurement, thereby obtaining spatial distance values ​​that more closely match actual geographical distribution.

[0071] In other embodiments, the data processing system acquires the latitude and longitude coordinates of the origin airport of the target route and the origin airport of the candidate route, and calculates the Euclidean plane distance between the two coordinates as the first distance; simultaneously, it calculates the Euclidean plane distance between the destination airport of the target route and the destination airport of the candidate route as the second distance. The data processing system adds the first distance and the second distance, and then divides by a preset scaling factor, using the scaled result as the spatial distance between the candidate route and the target route. This embodiment uses the less computationally intensive Euclidean distance instead of the spherical distance, which is beneficial for improving the computational efficiency in large-scale route screening scenarios and is suitable for real-time evaluation scenarios with high computational speed requirements.

[0072] In some optional embodiments, obtaining the first distance between the origin airport of the target route and the origin airport of the candidate route includes: the data processing system can obtain the latitude difference and longitude difference between the origin airports corresponding to the target route and the candidate route respectively; and determine the first distance based on the latitude difference, longitude difference and the average radius of the Earth.

[0073] For example, the latitude difference can refer to the absolute value of the difference between the latitude of the originating airport of the target route and the latitude of the originating airport of the candidate route. The longitude difference can refer to the absolute value of the difference between the longitude of the originating airport of the target route and the longitude of the originating airport of the candidate route. The average radius of the Earth can refer to the average radius of the Earth as a sphere. By acquiring the latitude and longitude differences between the two originating airports and performing spherical geometric calculations using the average radius of the Earth, the data processing system can quantify the spatial interval between the originating airports into a first distance. Calculating the first distance using the latitude difference, longitude difference, and average radius of the Earth converts the angular differences in the geographic coordinate system into actual geographic distances with units of length, thereby providing standardized metrics for subsequent spatial distance calculations.

[0074] In some embodiments, the data processing system converts the latitude values ​​of the originating airports of the target route and the candidate route to radians, respectively, and calculates the absolute value of the difference between the two radian values ​​as the latitude difference. Similarly, the data processing system converts the two longitude values ​​to radians and then calculates the absolute value of the longitude difference. The data processing system applies the semi-versus formula to calculate the first distance: first, it calculates the product of the square of half the latitude difference's sine and the product of the cosine of the originating latitude and the square of half the longitude difference's sine; then, it takes the square root of the above calculation result and calculates the arcsine; finally, it multiplies by twice the average radius of the Earth. This embodiment uses the semi-versus formula for spherical distance calculation, which is beneficial for maintaining the accuracy of distance calculation over a large geographical scale.

[0075] In other embodiments, the data processing system directly uses the latitude and longitude differences between the originating airports of the target route and the candidate routes. After converting these differences into radians, it calculates the first distance using the rectangular projection approximation formula. The rectangular projection approximation formula is: the square root of the sum of the Earth's average radius multiplied by the square of the latitude difference and the longitude difference multiplied by the square of the cosine of the originating latitude. The data processing system uses the calculated result as the first distance. This embodiment uses the rectangular approximation formula to calculate the first distance, which has lower computational complexity compared to the semi-versus formula, thus improving processing efficiency in scenarios requiring rapid calculation of spatial distances between a large number of route pairs.

[0076] In some optional embodiments, the geospatial similarity features between the target route and each candidate route in the candidate route sample set are determined based on the endpoint airport information of the target route. This includes: the data processing system can calculate the similarity between the flight distance of the target route and the flight distance of the candidate routes to obtain the route distance similarity between the target route and the candidate routes; determine the route geographical location similarity between the target route and the candidate routes based on the distance between the origin airport of the target route and the origin airport of the candidate routes, and the distance between the destination airport of the target route and the destination airport of the candidate routes; construct a target direction vector based on the latitude and longitude differences between the origin airport and the destination airport of the target route; construct a candidate direction vector based on the latitude and longitude differences between the origin airport and the destination airport of the candidate routes; calculate the dot product of the target direction vector and the candidate direction vector, and determine the route direction similarity between the target route and the candidate routes based on the dot product and the magnitudes corresponding to the target direction vector and the candidate direction vector, respectively; and use the route distance similarity, route geographical location similarity, and route direction similarity as geospatial similarity features.

[0077] For example, route distance similarity can refer to a quantified value of the closeness between the flight distance of the target route and the flight distance of the candidate route. It can be calculated as the normalized result of the ratio or difference between the two flight distances. Route geographic location similarity can refer to a spatial layout similarity index obtained by combining the distances between the origin airport and the destination airport of the target route and the candidate route. The target direction vector can refer to a vector composed of the latitude and longitude changes from the origin airport to the destination airport of the target route. The candidate direction vector can refer to a vector composed of the latitude and longitude changes from the origin airport to the destination airport of the candidate route. Route direction similarity can refer to the cosine value of the angle between the target direction vector and the candidate direction vector. By calculating the above three similarity indices and using them together as geospatial similarity features, the data processing system can comprehensively characterize the degree of matching of two routes in geospatial attributes from three dimensions: route length, endpoint spatial location correspondence, and route direction. Decomposing the spatial attributes of routes into three independent dimensions—distance, endpoint corresponding location, and direction—for quantification is beneficial for generating more discriminative and business-interpretive spatial matching features.

[0078] In some embodiments, after obtaining the flight distance values ​​of the target route and the candidate route, the data processing system calculates the absolute logarithm of the ratio of the two flight distance values, then subtracts the absolute logarithm from 1, and uses the result as the route distance similarity. The data processing system obtains a first distance between the origin airport of the target route and the origin airport of the candidate route, and a second distance between the destination airport of the target route and the destination airport of the candidate route. After normalizing the first and second distances respectively, the system sums them, then subtracts the sum from 1, and uses the result as the route geographic location similarity. The data processing system uses the latitude difference between the origin and destination airports of the target route as the first component of the target direction vector and the longitude difference as the second component; it constructs candidate direction vectors in the same way; it calculates the dot product of the target direction vector and the candidate direction vector, divides it by the product of the magnitudes of the two vectors, and uses the resulting cosine value as the route direction similarity. The data processing system combines the route distance similarity, route geographic location similarity, and route direction similarity into a three-dimensional feature vector as the geospatial similarity feature. This embodiment quantifies the distance matching degree, position matching degree, and direction matching degree through mathematical operations such as ratio, normalized summation, and cosine of the vector angle, which is beneficial to obtain three similarity components with consistent value ranges and clear physical meanings.

[0079] In other embodiments, the data processing system calculates the absolute value of the difference between the flight distance of the target route and the flight distance of the candidate route, then divides it by the larger of the two flight distances, and subtracts 1 from the result of the division to obtain the route distance similarity. The data processing system divides the first distance between the origin airport of the target route and the origin airport of the candidate route, and the second distance between the destination airport of the target route and the destination airport of the candidate route, respectively, by a preset spatial scale coefficient, calculates the square of the sum of the squares of the two divisors, and takes the reciprocal of the square root, using this reciprocal as the route geographic location similarity. The data processing system converts the target direction vector and the candidate direction vector into unit vectors and calculates their dot product, using the dot product value as the route direction similarity. The data processing system uses the weighted sum of the above three similarity values ​​to obtain the comprehensive geospatial feature as the geospatial similarity feature. This embodiment uses different normalization and aggregation methods to generate geospatial similarity features, which is beneficial for adapting to different route network distribution characteristics and evaluation needs, and improves the flexibility of the method.

[0080] Optionally, the data processing system uses the following formulas to calculate the geospatial similarity features and numerical and categorical similarities between the target route and the candidate routes.

[0081] The route distance similarity is calculated using the following formula (3):

[0082] Formula (3);

[0083] in, Indicates the flight distance of the target route. This represents the flight distance of the candidate route; the flight distance is calculated using the latitude and longitude of the origin and destination airports of the route through the semi-versus formula.

[0084] The geographical similarity of flight routes is calculated using the following formula:

[0085] Formula (4);

[0086] Formula (5);

[0087] Formula (6);

[0088] in, Indicates the latitude and longitude of the originating airport of the target route. Indicates the latitude and longitude of the originating airport of the candidate route. Indicates the latitude and longitude of the destination airport for the target route; Indicates the latitude and longitude of the destination airport for the candidate route; This represents the semi-sine distance function, used to calculate the spherical arc distance between two points on the Earth's surface; This indicates the similarity of the geographical locations of the flight routes.

[0089] The similarity of flight routes is calculated using the following formula:

[0090] Formula (7);

[0091] in, This represents the direction vector of the target route, whose components are the latitude and longitude changes between the origin and destination airports of the target route. This represents the direction vector of the candidate route, whose components are the latitude and longitude changes between the origin and destination airports of the candidate route. Represents the vector dot product. It represents the magnitude of the vector.

[0092] In some optional embodiments, the multi-level progressive evaluation model includes the following three progressive network layers: A first progressive network layer is used to determine the basic matching degree between the target route and candidate routes based on the latitude and longitude coordinates of the endpoint airports in the basic features and the route distance similarity and route geographical location similarity in the cross features; a second progressive network layer is used to differentiate and correct the basic matching degree based on the basic matching degree output by the first progressive network layer and the route direction similarity and travel demand density similarity in the cross features, and output the corrected matching degree; a third progressive network layer is used to differentiate and correct the basic matching degree based on the output of the second progressive network layer. The corrected matching degree, along with the similarity of passenger type, population mobility, and high-speed rail network competition intensity index in the cross features, are used to adjust the corrected matching degree and output the final matching degree. Among them, the similarity of air travel demand intensity, passenger type, population mobility, and high-speed rail network competition intensity index are obtained by calculating the air travel demand intensity level, passenger type level, population mobility level, and high-speed rail network competition intensity index of the endpoint cities of the target route and the candidate route, through either categorical similarity calculation or numerical similarity calculation.

[0093] For example, the first progressive network layer can refer to the network layer in the multi-level progressive evaluation model used to initially calculate the basic matching degree. Its inputs are the latitude and longitude coordinates of the endpoint airports in the basic features and the route distance similarity and route geographical location similarity in the cross features, and its output is the basic matching degree. The second progressive network layer can refer to the network layer used to differentiate and correct the basic matching degree. Its inputs are the basic matching degree output by the first progressive network layer and the route direction similarity and travel demand density similarity in the cross features, and its output is the corrected matching degree. The third progressive network layer can refer to the network layer used to fine-tune the corrected matching degree. Its inputs are the corrected matching degree output by the second progressive network layer and the passenger type similarity, population flow similarity, and high-speed rail network competition intensity index similarity in the cross features, and its output is the final matching degree. The data processing system calculates sequentially through the above three progressive network layers, which can realize multi-level matching degree evaluation from spatial basic attributes to operational attributes and then to other attributes. Decomposing the matching degree assessment process into three progressive levels—basic assessment, difference correction, and fine-tuning—is beneficial for simulating a coarse-to-fine decision-making logic, thereby improving the hierarchy and accuracy of matching degree prediction.

[0094] Optionally, the similarity of air travel demand intensity can be calculated based on spatial statistical indicators such as population density, urbanization rate, ground transportation infrastructure density, and air passenger throughput of the route endpoint cities. For example, population density can refer to the number of permanent residents per unit area of ​​a city, urbanization rate can refer to the proportion of permanent residents living in urban areas, and ground transportation infrastructure density can refer to the mileage of transportation networks such as railways and highways per unit area of ​​a city.

[0095] In some embodiments, the data processing system can construct the first progressive network layer as a random forest regressor containing fifty decision trees. The input is a feature vector composed of the latitude and longitude coordinates of the endpoint airport, route distance similarity, and route geographical location similarity. The output is a basic matching degree between 0 and 1. The data processing system constructs the second progressive network layer as another random forest regressor. The input is a feature vector formed by concatenating the basic matching degree output from the first progressive network layer with route direction similarity and travel demand density similarity. The output is a corrected matching degree. The data processing system constructs the third progressive network layer as a third random forest regressor. The input is a feature vector formed by concatenating the corrected matching degree output from the second progressive network layer with passenger type similarity, population flow similarity, and high-speed rail network competition intensity index similarity. The output is the final matching degree. This embodiment uses a random forest as the base model for each network layer, which helps to enhance the interpretability of the model while maintaining high prediction accuracy, making it easier for business personnel to understand the impact of each layer's features on the matching degree.

[0096] In other embodiments, the data processing system constructs a multilayer perceptron neural network with three hidden layers, consisting of a first, second, and third progressive network layer. The first hidden layer corresponds to the computational logic of the first progressive network layer, the second to the second, and the third to the third. The data processing system uses the latitude and longitude coordinates of the endpoint airport, route distance similarity, and route geographical location similarity as inputs to the first hidden layer. The output of the first hidden layer, along with route direction similarity and travel demand density similarity, serves as input to the second hidden layer. The output of the second hidden layer, along with passenger type similarity, population flow similarity, and high-speed rail network competition intensity index similarity, serves as input to the third hidden layer, which outputs the final matching degree. This embodiment employs an end-to-end multilayer neural network structure, which facilitates the automatic learning of the interaction relationships between features at each level through nonlinear transformations within the network, thereby improving the fitting ability of the matching degree prediction.

[0097] For example, the data processing system employs the following multi-level progressive evaluation algorithm structure. Let the original matching feature vector between the target route and a candidate route be... ,in, to This includes business scenario features among basic and cross-features, such as route distance matching degree, direction matching degree, population matching degree, regional matching degree, air travel demand intensity matching degree, passenger type matching degree, population mobility matching degree, and high-speed rail competition index matching degree.

[0098] Level 1: Basic Match Assessment. This level performs a preliminary assessment of the basic similarity between two routes based on the latitude and longitude coordinates of the endpoint airports in the basic feature subset, as well as the route distance similarity and route geographic location similarity in the cross features. The basic match score is then output. The calculation formula is as follows:

[0099] Formula (8);

[0100] in, This represents a subset of basic features, including the latitude and longitude coordinates of the endpoint airports of the target route and candidate routes, the matching degree of the city's permanent residents, the matching degree of the city's road network density, the matching degree of the region affiliation, and other basic indicators. This represents the first-level matching evaluation tool, consisting of 50 decision trees, used to analyze the matching degree of the aforementioned basic indicators.

[0101] The second level: operational feature difference correction. This level, based on the output of the first level, introduces route operational features for difference correction, and outputs the corrected matching degree. The calculation formula is as follows:

[0102] Formula (9);

[0103] in, This represents a subset of operational characteristics, including indicators reflecting the operational characteristics of a route itself, such as route distance similarity, route direction similarity, and similarity in the density of travel demand. This represents the second-level matching evaluation unit, based on the baseline matching score output from the first level. As an additional input feature, we learn how to adjust the basic matching degree based on differences in operational features.

[0104] The third level introduces other performance-related features to fine-tune the evaluation results of the first two levels, outputting the final matching degree. The calculation formula is as follows:

[0105] Formula (10);

[0106] in, This represents a subset of positioning features for aviation business scenarios, including similarity in passenger type, similarity in population mobility, and similarity in the competition intensity index of high-speed rail networks, which reflect the positioning environment and passenger behavior patterns of aviation business scenarios. This represents the third-level matching degree evaluator, which takes the outputs of the first two levels as input and learns the final impact of differences in the positioning environment of aviation business scenarios on route matching degree.

[0107] Final output This refers to the matching degree between the target route and the candidate route, with a value range of [0,1]. The larger the value, the closer the passenger traffic levels of the two routes are.

[0108] In some optional embodiments, the multi-level progressive evaluation model is pre-trained through the following training steps: obtaining a training dataset, which includes endpoint airport information for multiple existing routes and true matching degree labels between each route pair; using the training dataset as input, performing a two-stage business rule filtering operation to generate a training sample set; constructing multi-dimensional features between each route pair in the training sample set; using the multi-dimensional features of the training sample set as input and the true matching degree labels as supervision signals, updating the model parameters through a dual-objective joint optimization strategy to obtain the trained multi-level progressive evaluation model.

[0109] For example, the training dataset can refer to the original dataset used for model training, including endpoint airport information for multiple existing routes and true match labels between each route pair. The true match label can refer to the similarity value calculated based on the actual passenger volume of two existing routes, serving as the target value for model learning. The training sample set can refer to the sample set used as actual training input after the training dataset has undergone two-stage business rule filtering. The dual-objective joint optimization strategy can refer to a parameter update method that simultaneously optimizes both the accuracy of match prediction and the consistency of match ranking during model training. The data processing system first extracts the endpoint airport information of existing routes from the training dataset, performs two-stage business rule filtering to generate the training sample set, then constructs multi-dimensional features for each route pair in the training sample set, and finally updates the model parameters using the multi-dimensional features as input and the true match label as a supervision signal through the dual-objective joint optimization strategy to obtain the trained multi-level progressive evaluation model. Introducing the same two-stage filtering rules and dual-objective optimization strategy as the prediction stage in the model training stage helps ensure the consistency of the distribution of training data and prediction data, and enables the model to simultaneously possess numerical prediction accuracy and ranking reliability.

[0110] In some embodiments, the data processing system extracts endpoint airport information for multiple existing routes and monthly passenger volume data for each route pair from a historical operations database. It can generate true matching labels for each route pair and combine this data into a training dataset. The data processing system uses the training dataset as input and performs a two-stage business rule filtering process: first, it filters route pairs with endpoint airports of the same or slightly different levels based on airport throughput; then, it selects the k route pairs with the smallest spatial distance based on spatial distance. The retained route pairs are used as the training sample set. The data processing system constructs multi-dimensional features between each route pair in the training sample set, including basic features and cross-features. Using these multi-dimensional features as input and the true matching labels as supervision signals, the data processing system uses a backpropagation algorithm to iteratively update the network weights of the multi-level progressive evaluation model. In each iteration, it calculates the joint loss function value and adjusts the weights until the joint loss function value converges, resulting in a trained multi-level progressive evaluation model. This embodiment uses historical passenger volume data to generate supervisory labels and combines them with two-stage screening to construct training samples. This helps the model learn the mapping relationship between features and matching degree under the same screening logic as the evaluation stage, thereby improving the model's generalization ability in practical applications.

[0111] In other embodiments, the data processing system acquires the endpoint airport information and corresponding annual passenger volume data for multiple existing routes. It then uses time series decomposition to extract trend and seasonal components for each route, calculates true matching labels based on the similarity between trend components, and forms a training dataset. When the data processing system performs a two-stage business rule filtering, it replaces the airport throughput levels in the first stage with five scale levels generated by dynamic clustering based on annual passenger throughput, and replaces the spatial distance calculation in the second stage with Euclidean distance calculated based on the latitude and longitude of the route midpoint. After constructing multi-dimensional features, the data processing system trains a multi-level progressive evaluation model using mini-batch gradient descent. It calculates a weighted sum of mean squared error loss and sorting consistency loss on each mini-batch of data, updates the model parameters based on the gradient of the weighted sum, and outputs the trained model after completing a preset number of training rounds. This embodiment uses dynamic clustering to classify airport levels and route midpoint distance calculation as filtering methods, which is beneficial for adapting to different airport throughput distribution characteristics and data availability, improving the flexibility and adaptability of the training process.

[0112] For example, a data processing system can generate true matching labels for model training as follows: First, the total passenger volume for each route is summarized, grouped by route and month. Then, the passenger volume data is standardized according to the month dimension, using the following standardization formula:

[0113] Formula (11);

[0114] in, This represents the total number of original passengers on a given route in the current month. This represents the average number of passengers on all routes in the current month. This represents the standard deviation of passenger volume across all routes in the current month. This represents the standardized passenger volume.

[0115] For any two routes, the standardized passenger volume is denoted as follows: and The passenger volume similarity label is calculated using the following formula:

[0116] Formula (12);

[0117] in, This represents the passenger volume similarity label between routes, i.e., the passenger volume similarity or true match label. The standardized passenger volume of the two routes in the same month is denoted as follows: and For example, two routes can be a target route and a candidate route.

[0118] In some optional embodiments, the dual-objective joint optimization strategy includes: obtaining the true matching degree and predicted matching degree of each training sample in the training sample set; calculating the squared value of the difference between the true matching degree and the predicted matching degree of the training sample to obtain the first squared value corresponding to the training sample; summing the first squared values ​​corresponding to all training samples and dividing by the total number of training samples to obtain the mean squared error; obtaining the true ranking rank and predicted ranking rank of each training sample in the training sample set; calculating the squared value of the difference between the true ranking rank and the predicted ranking rank of the training sample to obtain the second squared value corresponding to the training sample; summing the second squared values ​​corresponding to all training samples, and determining the rank correlation coefficient based on the summation result and the total number of training samples, wherein the rank correlation coefficient is used to measure the consistency between the predicted ranking and the true ranking of the training samples, and the larger the value of the rank correlation coefficient, the higher the ranking consistency; constructing a joint loss function based on the mean squared error and the rank correlation coefficient; and generating a dual-objective joint optimization strategy with minimizing the joint loss function as the training objective.

[0119] For example, the true match can refer to the similarity value calculated based on actual passenger volume between two routes in the training samples, serving as a supervised target for model training. The predicted match can refer to the predicted match value output by the multi-level progressive evaluation model for the same training sample. The first square value can refer to the square of the difference between the true match and the predicted match. The mean squared error can refer to the arithmetic mean of the first square values ​​of all training samples, used to measure the numerical accuracy of the model's predictions. The true ranking rank can refer to the position of each training sample in the ranking sequence after sorting the training samples from largest to smallest according to the true match. The predicted ranking rank can refer to the position of each training sample in the ranking sequence after sorting the same group of training samples from largest to smallest according to the predicted match. The second square value can refer to the square of the difference between the true ranking rank and the predicted ranking rank. The rank correlation coefficient can refer to the Spearman rank correlation coefficient calculated based on the second square values ​​of all training samples, used to measure the consistency between the model's predicted ranking and the true ranking; a higher rank correlation coefficient indicates higher ranking consistency. The joint loss function refers to the loss function obtained by weighting the mean squared error and the rank correlation coefficient. It is used to simultaneously optimize numerical prediction accuracy and ranking consistency during training. The data processing system calculates the mean squared error and the rank correlation coefficient and constructs the joint loss function. It updates the model parameters with the goal of minimizing the joint loss function, generating a dual-objective joint optimization strategy. Considering both numerical bias and ranking bias in the matching degree during model training helps the trained model to not only predict the matching degree of new routes with numerical accuracy, but also correctly identify the relative similarity order between multiple candidate routes and the new route.

[0120] The data processing system can construct a multi-level progressive evaluation model, transforming the decision-making logic from coarse to fine and layer-by-layer correction into three progressive levels: basic matching degree evaluation, operational characteristic difference correction, and fine-tuning of aviation business scenario positioning performance. Through dual-objective joint optimization, it is beneficial to ensure that the model can both predict the matching degree value and identify the relative similarity ranking between candidate routes, which is conducive to realizing the complete technical process from matching degree calculation to matching degree ranking.

[0121] In some embodiments, the data processing system acquires the true matching degree and the predicted matching degree of all training samples in the training sample set. The system iterates through each training sample, calculates the difference between the true matching degree and the predicted matching degree, and uses the square of this difference as the first squared value for that training sample. The system sums the first squared values ​​for all training samples and divides the sum by the total number of training samples to obtain the mean squared error. Simultaneously, the system acquires the true and predicted rank of each training sample, calculates the square of the difference between the rank of each training sample as the second squared value, sums all the second squared values, and calculates the Spearman rank correlation coefficient based on the sum and the total number of training samples. The system adds the mean squared error to a preset weight coefficient multiplied by one minus the Spearman rank correlation coefficient to construct a joint loss function. The system optimizes by minimizing the joint loss function value, using gradient descent to iteratively update the network parameters of the multi-level progressive evaluation model until the joint loss function value converges or reaches a preset number of training epochs, thereby generating a dual-objective joint optimization strategy. This embodiment uses a weighted summation method to construct the joint loss function, which is beneficial for flexibly balancing the relative importance of numerical accuracy and ranking consistency in model optimization by adjusting the weight coefficients, and adapting to the emphasis requirements of evaluation indicators in different business scenarios.

[0122] In other embodiments, the data processing system dynamically adjusts the weights of the mean squared error (MSE) and rank correlation coefficient (MRC) terms in the joint loss function during training. The system sets initial weights such that the MSE term has a higher weight than the MRC term, primarily optimizing the numerical prediction accuracy of the matching degree in the first half of training. As the MSE on the validation set decreases gradually, the system progressively increases the weight of the MRC term while decreasing the weight of the MSE term, primarily optimizing ranking consistency in the second half of training. In each iteration, the system calculates the joint loss function value based on the current weights and updates the model parameters using an adaptive moment estimation algorithm until training is complete, generating a dual-objective joint optimization strategy. This embodiment employs a dynamic weight adjustment strategy, which is beneficial for rapidly reducing numerical bias in the early stages of training and improving ranking performance in the later stages, thereby improving the convergence efficiency and final performance of the model training.

[0123] For example, during model training, the data processing system employs a dual-objective joint optimization strategy, simultaneously optimizing the accuracy of matching degree prediction and the consistency of ranking.

[0124] Optimization objective 1: Match prediction accuracy. Mean squared error is used to measure the numerical difference between the predicted match rate and the actual match rate, as shown in the following formula:

[0125] Formula (13);

[0126] in, Indicates the first The true matching degree of each training sample, i.e., the passenger volume similarity label calculated based on the actual passenger volume. This indicates the predicted matching degree of the model output. This represents the total number of training samples.

[0127] Optimization Objective 2: Matching and Ranking Consistency. The Spearman rank correlation coefficient is used to measure the consistency between the model's predicted ranking and the actual ranking. The Spearman loss is defined as follows:

[0128] Formula (14);

[0129] in, This represents the Spearman rank correlation coefficient between the predicted match and the actual match, with a value ranging from [-1, 1]. The closer to 1, the higher the consistency of the sorting.

[0130] Joint optimization objective: The two optimization objectives mentioned above are weighted and combined to form the final loss function for model training.

[0131] Formula (15);

[0132] in, This represents the ranking consistency weight coefficient, used to balance the importance of matching prediction accuracy and ranking consistency in model optimization.

[0133] The matching degree evaluation is decomposed into three progressive levels: basic evaluation, difference correction, and fine-tuning. By using dual-objective joint optimization, both numerical prediction accuracy and relative ranking accuracy are ensured, thus realizing the process from matching degree calculation to matching degree ranking.

[0134] In some optional embodiments, after determining the matching degree between the target route and each candidate route through a multi-level progressive evaluation model, the method further includes: sorting the matching degree between the target route and each candidate route from high to low, and dividing them into at least two route levels according to the sorting results, wherein the first route level consists of the top M candidate routes with the highest matching degree, and the second route level consists of the next N candidate routes with the second highest matching degree, where M and N are positive integers; assigning a first aggregation weight to each candidate route in the first route level, and assigning a second aggregation weight to each candidate route in the second route level, wherein the first aggregation weight is greater than the second aggregation weight; multiplying the matching degree between each candidate route in the first route level and the target route by the corresponding first aggregation weight and summing them to obtain a first summation result; multiplying the matching degree between each candidate route in the second route level and the target route by the corresponding second aggregation weight and summing them to obtain a second summation result; and adding the first summation result and the second summation result to obtain a weighted aggregation value, which serves as the comprehensive matching reference value for the target route.

[0135] For example, the matching degree between the target route and each candidate route can refer to the similarity value between the target route and each candidate route output by the multi-level progressive evaluation model. Route level can refer to the different groups into which candidate routes are divided after being sorted from high to low matching degree. The first route level consists of the top M candidate routes with the highest matching degree, and the second route level consists of the next N candidate routes with the second highest matching degree, where M and N are positive integers. The first aggregation weight can refer to the weighted coefficient assigned to the matching degree of each candidate route in the first route level, and the second aggregation weight can refer to the weighted coefficient assigned to the matching degree of each candidate route in the second route level, where the first aggregation weight is greater than the second aggregation weight. The weighted aggregation value can refer to the sum of the matching degree of each candidate route multiplied by the aggregation weight corresponding to its level. The comprehensive matching reference value can refer to the weighted aggregation value, used to comprehensively characterize the overall similarity between the target route and the candidate route sample set. The data processing system sorts the target route and each candidate route from highest to lowest matching degree, divides the route into at least two levels based on the ranking, assigns decreasing aggregation weights to different levels, and then calculates the weighted sum of each level and adds them together to obtain a comprehensive matching reference value. By assigning greater weight to candidate routes with higher matching degrees, the system helps to highlight the influence of the most similar reference route in the comprehensive evaluation, while retaining the information contribution of the next most similar routes. The data processing system forms a complete technical closed loop from data acquisition, sample screening, feature construction to matching degree calculation and ranking. By integrating the operational data of multiple similar routes through the matching degree weighted aggregation mechanism, the system helps to reduce the evaluation bias caused by a single reference route and provides data support for route network planning decisions.

[0136] In some embodiments, the data processing system sorts the matching degree between the target route and each candidate route from high to low, then selects a predetermined number of candidate routes with the highest matching degree as the first route level, and the remaining candidate routes as the second route level. The data processing system assigns a first aggregation weight to each candidate route in the first route level and a second aggregation weight to each candidate route in the second route level. The data processing system calculates the product of the matching degree of each candidate route in the first route level and its first aggregation weight, and sums all products to obtain a first summation result. The data processing system calculates the product of the matching degree of each candidate route in the second route level and its second aggregation weight, and sums all products to obtain a second summation result. The data processing system adds the first summation result and the second summation result, and uses the sum as the comprehensive matching reference value for the target route. This embodiment uses a hierarchical weighted aggregation method, which is beneficial for retaining all candidate route information while ensuring that routes with higher matching degrees dominate the comprehensive reference value, thereby improving the reliability of the evaluation results.

[0137] In other embodiments, after sorting the matching degree from high to low, the data processing system divides the candidate routes into at least two route levels according to multiple preset matching degree intervals. Candidate routes with matching degrees within a first preset interval are assigned to the first route level, and those within a second preset interval are assigned to the second route level. The minimum matching degree in the first preset interval is greater than the maximum matching degree in the second preset interval. The data processing system assigns a higher aggregation weight to the first route level than to the second route level. The data processing system calculates the sum of the products of the matching degree and the corresponding aggregation weight within each route level, then adds the sum of the products for each route level, using the sum as the comprehensive matching reference value for the target route. This embodiment uses a level division method based on matching degree intervals, which allows for flexible adjustments based on the contribution of different matching degree segments to the comprehensive reference value, thereby improving the adaptability of the weighted aggregation strategy.

[0138] Figure 2This is an architecture diagram of an optional multi-level evaluation method for route matching in this application. First, the data processing system acquires basic urban data, including the latitude and longitude coordinates of the endpoint airport, the city's permanent resident population, the city's road network density, the high-speed rail network competition intensity index, the city's leading industry categories, the main travel destinations of air passengers in the city, the city's air demand intensity level, and the city's population flow intensity level. Then, the system proceeds to a data verification and completion process, sequentially executing random sampling verification, accuracy calculation, and numerical data verification. If verification fails, data updates and corrections are performed, and manual review is conducted before re-verification. After successful verification, a basic feature module (basic information of the target route's origin airport and destination airport) and a cross-feature module (geospatial similarity features and flight segment business scenario similarity) are constructed sequentially to complete multi-dimensional feature extraction. Next, a two-stage route sample screening based on business rules is performed. Finally, a multi-level progressive evaluation model outputs the matching degree and ranking.

[0139] In some optional embodiments, the basic features also include a first type of numerical urban statistical indicators and a second type of categorical urban attribute indicators; the cross features also include numerical similarity calculated based on the first type of numerical urban statistical indicators and categorical similarity calculated based on the second type of categorical urban attribute indicators; wherein, the first type of numerical urban statistical indicators includes at least one of urban resident population, urban road network density, and high-speed rail network competition intensity index; the second type of categorical urban attribute indicators includes at least one of the geographical region to which the city belongs, the city's leading industry category, the main travel destinations of urban air passengers, the city's air demand intensity level, and the city's population mobility intensity level.

[0140] For example, the first type of numerical urban statistical indicators can refer to statistics measured in continuous numerical form in the endpoint cities of flight routes, including at least one of the following: urban resident population, urban road network density, and high-speed rail network competition intensity index. The second type of categorical urban attribute indicators can refer to attribute characteristics described in discrete category form in the endpoint cities of flight routes, including at least one of the following: the geographical region to which the city belongs, the city's leading industry category, the main travel destination type of urban air passengers, the city's air demand intensity level, and the city's population flow intensity level. Numerical similarity can refer to the similarity value calculated based on the first type of numerical urban statistical indicators. Categorical similarity can refer to the similarity value calculated based on the second type of categorical urban attribute indicators. The data processing system uses the above-mentioned numerical similarity and categorical similarity as components of cross-features, and inputs them together with geospatial similarity features into a multi-level progressive evaluation model. By introducing multiple statistical indicators of endpoint cities and their similarity calculations, it is beneficial to enrich the feature dimensions of flight route matching from the perspective of urban attributes, thereby improving the model's ability to perceive the differences in urban characteristics of different flight routes and the accuracy of matching degree evaluation. By using language models to automatically extract basic information about the endpoint cities of air routes from public reports, and by designing structured instructions to ensure data quality, and by establishing a data verification and supplementation mechanism, we can effectively solve the problems of difficulty in obtaining and scaling up characteristic data for newly opened air routes.

[0141] In some embodiments, the data processing system acquires the resident population, road network density, and high-speed rail network competition intensity index of the endpoint cities of the target and candidate routes, and uses these values ​​as a first type of numerical urban statistical indicator. For each indicator in the first type of numerical urban statistical indicator, the data processing system calculates the similarity of indicator values ​​between the starting city of the target route and the starting city of the candidate route, and between the ending city of the target route and the ending city of the candidate route. The average of the two similarities is taken as the numerical similarity of that indicator. The data processing system combines the numerical similarities of each indicator into a numerical similarity feature vector. This embodiment, by calculating the similarity of the starting and ending points separately and taking the average, helps to more comprehensively reflect the degree of symmetry matching of the two routes in the numerical indicators of the endpoint cities, thereby improving the representational ability of the numerical similarity feature.

[0142] In other embodiments, the data processing system acquires the geographical region, dominant industry category, main travel destination of air passengers, air demand intensity level, and population flow intensity level of the endpoint cities of the target and candidate routes, and uses these attributes as second-category city attribute indicators. For each indicator in the second-category city attribute indicators, the data processing system pre-maps the classification levels of the indicator to corresponding numerical codes in a preset order, calculates the coded normalized distance between the starting point of the target route and the starting point of the candidate route, and the coded normalized distance between the ending point of the target route and the ending point of the candidate route, and uses the average of the two normalized distances as the categorical difference degree. Then, it subtracts the categorical difference degree from 1 to obtain the categorical similarity of the indicator. The data processing system combines the categorical similarity of each indicator into a categorical similarity feature vector. This embodiment calculates categorical similarity by using numerical codes and normalized distances, which is beneficial for converting discrete category information into continuously comparable similarity values, thereby facilitating the model to uniformly process numerical features and categorical features.

[0143] For example, let the newly opened route be route A, the origin airport be A1, and the destination airport be A2. Let the existing route be route B, the origin airport be B1, and the destination airport be B2. Then the basic feature modules can be described as shown in Table 3:

[0144] Table 3

[0145]

[0146] In some optional embodiments, numerical similarity is calculated through the following steps: For the i-th indicator in the first type of numerical city statistical indicators, obtain the indicator values ​​of the origin and destination cities of the target route with respect to the i-th indicator, and obtain the indicator values ​​of the origin and destination cities of the candidate routes with respect to the i-th indicator, where i is an integer greater than or equal to 1; calculate the similarity between the origin cities of the target route and the candidate routes with respect to the i-th indicator, as the origin similarity for the i-th indicator; calculate the similarity between the destination cities of the target route and the candidate routes with respect to the i-th indicator, as the destination similarity for the i-th indicator; calculate the average of the origin similarity and the destination similarity for the i-th indicator to obtain the numerical similarity for the i-th indicator.

[0147] For example, the i-th indicator can refer to any indicator in the first type of numerical urban statistical indicators, where i is an integer greater than or equal to 1. Origin similarity can refer to the degree of similarity between the origin cities of the target route and the origin cities of the candidate routes regarding the i-th indicator value. Destination similarity can refer to the degree of similarity between the destination cities of the target route and the destination cities of the candidate routes regarding the i-th indicator value. Numerical similarity can refer to the average of the origin similarity and destination similarity for the i-th indicator. For each indicator in the first type of numerical urban statistical indicators, the data processing system obtains the indicator values ​​of the origin and destination cities of the target route, as well as the indicator values ​​of the origin and destination cities of the candidate routes. Then, it calculates the origin similarity and destination similarity sequentially, and finally takes the average of the two as the numerical similarity of that indicator. By calculating the similarity at the origin and destination ends separately and taking the average, it is beneficial to comprehensively reflect the overall matching degree of the two routes in terms of the numerical indicators of the endpoint cities, thereby avoiding the bias caused by considering only the matching of a single endpoint.

[0148] In some embodiments, for the i-th indicator in the first type of numerical urban statistical indicators, the data processing system substitutes the indicator values ​​of the target route's starting city and the candidate route's starting city into a preset numerical similarity function. The numerical similarity function can take the smaller value between the ratio of the two indicator values ​​and 1, or take 1 minus the absolute value of the difference between the two indicator values ​​divided by the sum of the two indicator values. The output value of this function is used as the starting point similarity. The data processing system uses the same method to substitute the indicator values ​​of the target route's ending city and the candidate route's ending city into the numerical similarity function to obtain the ending point similarity. The data processing system calculates the arithmetic mean of the starting point similarity and the ending point similarity, and uses this arithmetic mean as the numerical similarity of the i-th indicator. This embodiment uses a ratio-based or normalized difference-based similarity function, which is beneficial for uniformly mapping numerical indicators of different dimensions to a similarity range of 0 to 1, facilitating subsequent model processing.

[0149] In other embodiments, for the i-th indicator in the first type of numerical urban statistical indicators, the data processing system first performs global normalization on the four endpoint city indicator values: the target route start value, the target route end value, the candidate route start value, and the candidate route end value. Normalization can be achieved by subtracting the indicator mean and then dividing by the indicator standard deviation. The data processing system then calculates the Euclidean distance between the normalized target route start value and the candidate route start value, subtracting the Euclidean distance from 1 to obtain the start-point similarity. The same method is used to calculate the end-point similarity. The data processing system calculates a weighted average of the start-point similarity and the end-point similarity, where the weights of the start-point similarity and the end-point similarity are dynamically set according to the route direction or the throughput level of the endpoint airport. This weighted average is used as the numerical similarity of the i-th indicator. This embodiment uses global normalization and dynamic weighted averaging, which helps to eliminate the influence of differences in the dimensions of different indicators and adjusts the similarity calculation method according to the differences in the importance of endpoints, thereby improving the adaptability and accuracy of the numerical similarity.

[0150] For example, for each of the indicators in urban resident population, urban road network density, and high-speed rail network competition intensity index, the numerical similarity of the first type of numerical urban statistical indicators can be calculated using the following formula:

[0151] Formula (16);

[0152] Formula (17);

[0153] Formula (18);

[0154] in, This represents the index value of the i-th type of numerical city statistical indicator for the origin city of the target route. This represents the index value of the candidate route originating city with respect to the i-th type of numerical city statistical indicator. This represents the index value of the i-th type of numerical city statistical indicator for the destination city of the target route. This represents the index value of the candidate route's destination city with respect to the i-th type of numerical city statistical indicator. This represents the starting point similarity for the i-th indicator. This represents the endpoint similarity for the i-th indicator. This represents the numerical similarity of the i-th indicator.

[0155] In some optional embodiments, the categorical similarity is calculated through the following steps: For each indicator in the second category of categorical city attribute indicators, the classification levels of the indicator are pre-mapped to corresponding numerical codes in a preset order, so that each level corresponds to a unique code value; based on the indicator level to which the starting city of the target route belongs, the numerical codes that have a mapping relationship with that indicator level are queried to obtain the starting code value and the ending code value of the target route; based on the indicator level to which the starting city of the candidate route belongs, the numerical codes that have a mapping relationship with that indicator level are queried to obtain the starting code value and the ending code value of the candidate route; Calculate the absolute value of the difference between the starting point code value of the target route and the starting point code value of the candidate route, as the starting point code absolute difference; divide the starting point code absolute difference by the code value range to obtain the starting point code distance, where the code value range is equal to the difference between the maximum and minimum code values; calculate the absolute value of the difference between the ending point code value of the target route and the ending point code value of the candidate route, as the ending point code absolute difference; divide the ending point code absolute difference by the code value range to obtain the ending point code distance; calculate the average of the starting point code distance and the ending point code distance to obtain the average code distance, and then calculate the class similarity, where the class similarity is equal to 1 minus the average code distance.

[0156] For example, the second category of categorical city attribute indicators can refer to city attribute features described in discrete levels, with each indicator's classification level corresponding to a unique numerical code in a preset order. The numerical code can refer to the mapping result of converting the classification level into an integer value. The code value range can refer to the difference between the maximum and minimum code values. The average code distance can refer to the average of the starting point code distance and the ending point code distance. The categorical similarity can refer to the result obtained by subtracting the average code distance from 1. For each indicator in the second category of categorical city attribute indicators, the data processing system first establishes a mapping relationship between classification levels and numerical codes, then obtains the code values ​​corresponding to the starting and ending cities of the target and candidate routes, calculates the starting point code distance and the ending point code distance respectively, takes the average of the two to obtain the average code distance, and finally subtracts this average code distance from 1 to obtain the categorical similarity. Converting discrete classification levels into continuous categorical similarity values ​​through coding and normalized distance calculation helps to ensure that categorical features and numerical features participate in model calculations on the same numerical scale, thereby improving the uniformity and comparability of the feature system.

[0157] The data processing system can construct a two-layer feature system that includes basic features and cross features. The cross features are further divided into numerical similarity, geospatial similarity features, and categorical similarity. Specialized features for civil aviation business scenarios, such as route direction matching degree, population flow degree similarity, and travel demand density similarity, can be designed to make the feature system have good interpretability.

[0158] In some embodiments, for each indicator in the second category of city attribute indicators, the data processing system maps each level sequentially to a continuous integer code starting from 1, according to the order of magnitude or degree represented by the classification level from low to high. The data processing system obtains the code value corresponding to the indicator level of the target route's starting city as the target route's starting point code value, and obtains the code value corresponding to the indicator level of the target route's ending city as the target route's ending point code value; it also obtains the code values ​​of candidate route starting cities as candidate route starting points code values, and candidate route ending cities as candidate route ending points code values. The data processing system calculates the absolute value of the difference between the target route's starting point code value and the candidate route's starting point code value, divides the absolute value by the code value range, and obtains the starting point code distance; the ending point code distance is obtained in the same way. The data processing system calculates the arithmetic mean of the starting point code distance and the ending point code distance to obtain the average code distance, and then subtracts the average code distance from 1 to obtain the classification similarity. This embodiment uses a continuous coding of classification levels from low to high, which helps to ensure that the difference between code values ​​truly reflects the degree difference between levels, thereby improving the rationality of the classification similarity calculation.

[0159] In other embodiments, for city attribute indicators in the second category that do not have a natural order relationship, such as the geographical region to which a city belongs, the data processing system uses one-hot encoding before calculating similarity. Specifically, the data processing system expands each classification level into a binary vector with a length equal to the total number of classification levels. Each level corresponds to a dimension with a value of 1, while other dimensions have values ​​of 0. The data processing system obtains the one-hot encoded vectors of the starting city of the target route and the starting cities of the candidate routes, respectively. It calculates the cosine similarity between the two vectors as the complement of the starting point encoding distance, i.e., the starting point similarity. The same method is used to calculate the ending point similarity. Then, the average of the starting point similarity and the ending point similarity is calculated, and this average is directly used as the classification similarity. This embodiment uses one-hot encoding and cosine similarity calculation for unordered classification indicators, which helps to avoid introducing non-existent magnitude relationships due to artificially assigned sequential encoding, thereby improving the accuracy of classification similarity calculation.

[0160] For example, for each indicator in the categories of a city's geographical region, dominant industry category, main travel destination of air passengers, air demand intensity level, and population mobility intensity level, the following steps are used to calculate the categorical similarity.

[0161] First, the various classification levels of the indicator are mapped to continuous integer codes in order of increasing severity, with the code value range being the difference between the maximum and minimum code values.

[0162] Then, the categorical similarity is calculated using the following formula:

[0163] Formula (19);

[0164] in, This represents the numerical code of the originating city of the target route with respect to a certain second-category city attribute indicator. This represents the numerical code of the candidate route originating city for a certain second-category city attribute indicator. This represents the numerical code of the destination city of the target route with respect to a certain second-category city attribute indicator. This represents the numerical code of the candidate route's destination city for a certain second-category city attribute indicator. This indicates the range of values ​​that can be represented by the code, specifically the difference between the maximum and minimum values. This indicates the similarity between different categories.

[0165] Data processing systems can employ domain-specific knowledge graphs to extract endpoint city attribute information, offering high controllability of the extraction results. Alternatively, network embedding techniques can be used to quantify the similarity of flight routes within the network topology, capturing potential similar features through graph structures. However, knowledge graph construction and maintenance are costly, and network embedding methods have relatively weak interpretability. In contrast, this technical solution utilizes a language model-based approach to acquire basic data, achieving a better balance between automation, generalization, and interpretability.

[0166] For example, six cross-categorized characteristics—route revenue level matching, regional matching, industry type matching, passenger type matching, air travel demand intensity matching, and population mobility matching—combined to characterize the degree of type matching between new and existing routes in their endpoint cities. In civil aviation, revenue level reflects a city's consumption level and travel preferences; regional affiliation reflects the common characteristics of geographical location and business scenario positioning; industry type determines the main purpose of passenger travel (business travel, tourism, or working / visiting relatives); passenger type reveals the core composition of air passenger sources (high-end business travelers, cultural tourism travelers, business travelers, or working / visiting relatives); air travel demand intensity comprehensively characterizes the intensity of air travel demand determined by a city's population density, industrial agglomeration, and transportation infrastructure level; and population mobility reflects a city's population attractiveness and vitality, directly affecting the stability and seasonal fluctuations of air passenger sources. These six characteristics collectively point to a core business question: do the endpoint cities of new and existing routes belong to the same "type category" in terms of social structure and passenger behavior patterns? The calculation methods for route revenue level similarity, route regional similarity, route industry type similarity, route passenger type similarity, route travel demand intensity similarity, and route population mobility similarity are the same. Taking route population mobility similarity as an example: First, the population mobility characteristics are numerically encoded, with H, M, L, and N encoded as 1, 2, 3, and 4 respectively, and then calculated as follows:

[0167] 1 - (|Population flow code of originating city of route 1 - Population flow code of originating city of route 2| / (MAX-Min) + |Population flow code of destination city of route 1 - Population flow code of destination city of route 2| / (MAX-Min)) / 2. Where MAX is the maximum value of the code and Min is the minimum value of the code.

[0168] By calculating basic features and cross features, a multi-dimensional feature system representing the similarity between the new route A and the existing route B can be obtained, namely the feature system for evaluating the value of the new route, as shown in Table 4:

[0169] Table 4

[0170]

[0171] Based on the aforementioned features, a value evaluation index is constructed by calculating passenger volume or revenue between routes. The similarity labels for these features and passenger volume or revenue between routes are trained on an existing route dataset. No weights are assigned to features during the feature design phase; the model automatically learns the importance coefficient of each feature. When evaluating the value of a new route, these features can be input to find the route with the highest similarity in passenger volume or revenue to the newly opened route, thereby guiding airlines' business decisions regarding new route openings.

[0172] Figure 3 A flowchart illustrating an optional multi-level evaluation method for route matching degree according to this application is shown. First, the data processing system acquires the target route endpoint airport information and urban statistical indicators, specifically, the target route endpoint airport information, first-type numerical urban statistical indicators, and second-type categorical urban attribute indicators. Then, a data verification process is initiated. If verification fails, the data is updated and corrected, and the system returns for re-verification. Upon successful verification, basic features are constructed sequentially, including at least the latitude and longitude coordinates of the endpoint airports, and cross-features are constructed, including at least geospatial similarity features, completing multi-dimensional feature extraction. Next, a two-stage business rule screening is performed. Finally, a multi-level progressive evaluation model outputs the matching degree and ranking, i.e., the matching degree, progressive network layer, and ranking weighted aggregation are output. The entire process forms a technical closed loop from data acquisition, verification, feature construction, sample screening to matching degree output and ranking.

[0173] See Figure 4 According to another aspect of the embodiments of this application, a multi-level evaluation device for route matching degree is also provided, comprising: an acquisition unit, configured to acquire endpoint airport information of a target route, the endpoint airport information including the airport names and latitude and longitude coordinates of the origin airport and the destination airport; a filtering processing unit, configured to filter out a set of candidate route samples that match the target route from an existing route database; an extraction processing unit, configured to extract multi-dimensional features between the target route and each candidate route in the candidate route sample set based on the endpoint airport information of the target route, wherein the multi-dimensional features include basic features and cross features, the basic features including at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes respectively, and the cross features including at least the geospatial similarity features between the target route and the candidate routes; and a model processing unit, configured to input the multi-dimensional features into a pre-trained multi-level progressive evaluation model, and output the matching degree between the target route and each candidate route by the multi-level progressive evaluation model; wherein the multi-level progressive evaluation model is a neural network model trained by supervised learning using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels.

[0174] Optionally, the filtering processing unit includes: a two-stage filtering subunit, used to perform two-stage business rule filtering on the target route, filtering out a candidate route sample set that matches the target route from the existing route database, wherein the two-stage business rule filtering includes the following first stage and second stage executed sequentially; the two-stage filtering subunit includes: a first filtering submodule, used to execute the first stage, including: scale matching filtering based on airport throughput level, obtaining the endpoint airport level of the target route, filtering out m routes from the existing route database whose endpoint airport level is the same as or one level different from the endpoint airport level of the target route, forming a first route candidate set, wherein the endpoint airport level includes the level of the origin airport and the level of the destination airport; a second filtering submodule, used to execute the second stage, including: detecting the spatial distance between each candidate route in the first route candidate set and the target route, selecting the k routes with the smallest spatial distance as the candidate route sample set, wherein k is a positive integer less than or equal to m.

[0175] Optionally, the second screening submodule further includes: a spatial distance sequence generation component, used to obtain the spatial distances between all candidate routes in the first route candidate set and the target route before selecting the k routes with the smallest spatial distances as the candidate route sample set, forming a spatial distance sequence; a statistical calculation component, used to calculate the mean and standard deviation of the spatial distance sequence; a threshold determination component, used to determine the distance threshold based on the mean and standard deviation of the spatial distance sequence, wherein the distance threshold is equal to the sum of the mean and n times the standard deviation, where n is a preset positive integer coefficient; and a quantity statistics component, used to count the number of candidate routes in the first route candidate set whose spatial distance is less than or equal to the distance threshold, and use this number as the value of k.

[0176] Optionally, the second filtering submodule includes: a first distance acquisition component for acquiring a first distance between the origin airport of the target route and the origin airport of the candidate route; a second distance acquisition component for acquiring a second distance between the destination airport of the target route and the destination airport of the candidate route; and a spatial distance calculation component for adding the first distance and the second distance to obtain the spatial distance between the candidate route and the target route.

[0177] Optionally, the first distance acquisition component includes: a latitude and longitude difference calculation subcomponent, used to acquire the latitude and longitude differences between the origin airports corresponding to the target route and the candidate route, respectively; and a spherical distance determination subcomponent, used to determine the first distance based on the latitude difference, longitude difference, and the average radius of the Earth.

[0178] Optionally, the extraction and processing unit includes: a geospatial similarity feature extraction subunit, used to determine the geospatial similarity features between the target route and each candidate route in the candidate route sample set based on the endpoint airport information of the target route; the geospatial similarity feature extraction subunit includes: a route distance similarity calculation module, used to calculate the similarity between the flight distance of the target route and the flight distance of the candidate routes, to obtain the route distance similarity between the target route and the candidate routes; and a route geographic location similarity calculation module, used to determine the target route based on the distance between the origin airport of the target route and the origin airport of the candidate routes, and the distance between the destination airport of the target route and the destination airport of the candidate routes. The system includes: a route geolocation similarity module for the target route and candidate routes; a direction vector construction module for constructing a target direction vector based on the latitude and longitude differences between the origin and destination airports of the target route; and a candidate direction vector for constructing a candidate direction vector based on the latitude and longitude differences between the origin and destination airports of the candidate routes; a route direction similarity calculation module for calculating the dot product of the target and candidate direction vectors and determining the route direction similarity between the target and candidate routes based on the dot product and the magnitudes of the target and candidate direction vectors; and a feature combination module for using route distance similarity, route geolocation similarity, and route direction similarity as geospatial similarity features.

[0179] Optionally, the multi-level progressive evaluation model includes the following three progressive network layers: The first progressive network layer determines the basic matching degree between the target route and candidate routes based on the latitude and longitude coordinates of the endpoint airports in the basic features and the route distance similarity and route geographical location similarity in the cross features; the second progressive network layer differentiates and corrects the basic matching degree based on the basic matching degree output by the first progressive network layer and the route direction similarity and travel demand intensity similarity in the cross features, outputting the corrected matching degree; the third progressive network layer adjusts the corrected matching degree based on the corrected matching degree output by the second progressive network layer and the passenger type similarity, population mobility similarity, and high-speed rail network competition intensity index similarity in the cross features, outputting the final matching degree; wherein, the travel demand intensity similarity, passenger type similarity, population mobility similarity, and high-speed rail network competition intensity index similarity are data obtained by categorical similarity calculation or numerical similarity calculation based on the travel demand intensity level, passenger type level, population mobility level, and high-speed rail network competition intensity index of the endpoint cities of the target route and candidate routes, respectively.

[0180] Optionally, the multi-level progressive evaluation model is pre-trained through the following training units: a training data acquisition subunit, used to acquire the training dataset, which includes endpoint airport information for multiple existing routes and true matching degree labels between each route pair; a sample screening subunit, used to take the training dataset as input, perform a two-stage business rule screening operation, and generate a training sample set; a feature construction subunit, used to construct multi-dimensional features between each route pair in the training sample set; and a model training subunit, used to take the multi-dimensional features of the training sample set as input, take the true matching degree labels as supervision signals, and update the model parameters through a dual-objective joint optimization strategy to obtain the trained multi-level progressive evaluation model.

[0181] Optionally, the model training subunit includes: a matching degree acquisition module, used to acquire the true matching degree and predicted matching degree of each training sample in the training sample set; a first square value calculation module, used to calculate the square of the difference between the true matching degree and the predicted matching degree of the training sample to obtain the first square value corresponding to the training sample; a mean square error calculation module, used to sum the first square values ​​corresponding to all training samples and divide by the total number of training samples to obtain the mean square error; a ranking rank acquisition module, used to acquire the true ranking rank and predicted ranking rank of each training sample in the training sample set; and a second square value calculation module, used to calculate the true ranking of the training samples. The squared difference between the rank and the predicted rank is used to obtain the second squared value for the training sample. The rank correlation coefficient determination module sums the second squared values ​​for all training samples and determines the rank correlation coefficient based on the summation and the total number of training samples. The rank correlation coefficient measures the consistency between the predicted and actual rankings of the training samples; a larger rank correlation coefficient indicates higher ranking consistency. The joint loss function construction module constructs a joint loss function based on the mean squared error and the rank correlation coefficient. The optimization strategy generation module generates a dual-objective joint optimization strategy with minimizing the joint loss function as the training objective.

[0182] Optionally, the device further includes: a sorting and hierarchical unit, used to sort the matching degree between the target route and each candidate route from high to low after determining the matching degree between the target route and each candidate route through a multi-level progressive evaluation model, and divide the route into at least two route levels according to the sorting results, wherein the first route level consists of the top M candidate routes with the highest matching degree, and the second route level consists of the next N candidate routes with the second highest matching degree, where M and N are positive integers; and a weight allocation unit, used to assign a first aggregation weight to each candidate route in the first route level, and a weight to each candidate route in the second route level. The system selects routes and assigns a second aggregation weight, where the first aggregation weight is greater than the second aggregation weight. A first summation unit multiplies the matching degree of each candidate route at the first route level with the target route by its corresponding first aggregation weight and then sums the results to obtain a first summation result. A second summation unit multiplies the matching degree of each candidate route at the second route level with the target route by its corresponding second aggregation weight and then sums the results to obtain a second summation result. A weighted aggregation unit adds the first and second summation results to obtain a weighted aggregation value, which serves as a comprehensive matching reference value for the target route.

[0183] Optionally, the basic features also include the first type of numerical urban statistical indicators and the second type of categorical urban attribute indicators; the cross features also include numerical similarity calculated based on the first type of numerical urban statistical indicators and categorical similarity calculated based on the second type of categorical urban attribute indicators; wherein, the first type of numerical urban statistical indicators includes at least one of urban resident population, urban road network density, and high-speed rail network competition intensity index; the second type of categorical urban attribute indicators includes at least one of the following: the geographical region to which the city belongs, the city's leading industry category, the main travel destinations of urban air passengers, the intensity level of urban air travel demand, and the intensity level of urban population flow.

[0184] Optionally, the extraction and processing unit further includes a numerical similarity calculation subunit, used to calculate numerical similarity through the following steps: for the i-th indicator in the first type of numerical city statistical indicators, obtain the indicator values ​​of the starting city and ending city of the target route with respect to the i-th indicator, and the indicator values ​​of the starting city and ending city of the candidate route with respect to the i-th indicator, where i is an integer greater than or equal to 1; calculate the similarity between the target route and the candidate route with respect to the indicator values ​​of the starting cities with respect to the i-th indicator, as the starting similarity for the i-th indicator; calculate the similarity between the target route and the candidate route with respect to the indicator values ​​of the ending cities with respect to the i-th indicator, as the ending similarity for the i-th indicator; calculate the average of the starting similarity and the ending similarity for the i-th indicator to obtain the numerical similarity for the i-th indicator.

[0185] Optionally, the extraction and processing unit further includes: a categorical similarity calculation subunit, used to calculate categorical similarity through the following steps: for each indicator in the second category of categorical city attribute indicators, pre-map each classification level of the indicator to a corresponding numerical code in a preset order, so that each level corresponds to a unique code value; based on the indicator level to which the starting city of the target route belongs, query the numerical codes that have a mapping relationship with the indicator level to obtain the starting code value and the ending code value of the target route; based on the indicator level to which the starting city of the candidate route belongs, query the numerical codes that have a mapping relationship with the indicator level to obtain the starting code value and the candidate route. The process involves calculating the absolute difference between the starting point code value of the target route and the starting point code value of the candidate route, as the absolute difference of the starting point code. This absolute difference is then divided by the range of code values ​​to obtain the starting point code distance, where the range of code values ​​is equal to the difference between the maximum and minimum code values. The process also involves calculating the absolute difference between the ending point code value of the target route and the ending point code value of the candidate route, as the absolute difference of the ending point code. This absolute difference is then divided by the range of code values ​​to obtain the ending point code distance. Finally, the average of the starting point code distance and the ending point code distance is calculated to obtain the average code distance. Finally, the classification similarity is calculated, where the classification similarity is equal to 1 minus the average code distance.

[0186] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described multi-level evaluation method for route matching.

[0187] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described multi-level evaluation method for route matching.

[0188] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described multi-level evaluation method for route matching degree.

[0189] The sequence numbers of the embodiments described above are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. In the above embodiments of this application, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. It should be understood that the disclosed technical content in the several embodiments provided in this application can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units can be a logical functional division, and in actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces; the indirect coupling or communication connection between units or modules can be electrical or other forms. The units described as separate components may or may not be physically separated; the components shown as units may or may not be physical units, that is, they can be located in one place or distributed across multiple units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution in this embodiment.

[0190] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes: USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, optical disks, and other media capable of storing program code. The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-level evaluation method for route matching degree, characterized in that, include: Obtain the endpoint airport information of the target route, which includes the airport names and latitude and longitude coordinates of the origin airport and the destination airport; Select a set of candidate routes that match the target route from the existing route database; Based on the endpoint airport information of the target route, multi-dimensional features are extracted between the target route and each candidate route in the candidate route sample set. The multi-dimensional features include basic features and cross features. The basic features include at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes, and the cross features include at least the geospatial similarity features between the target route and the candidate routes. The multi-dimensional features are input into a pre-trained multi-level progressive evaluation model, which outputs the matching degree between the target route and each candidate route. The multi-level progressive evaluation model is a neural network model trained by supervised learning, using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels. This includes selecting a candidate route sample set that matches the target route from an existing route database, including: A two-stage business rule screening is performed on the target route to select a candidate route sample set that matches the target route from the existing route database. The two-stage business rule screening includes the following first stage and second stage executed in sequence. The first stage includes: scale matching and filtering based on airport throughput level, obtaining the endpoint airport level of the target route, and filtering out m routes from the existing route database whose endpoint airport level is the same as or one level different from the endpoint airport level of the target route to form a first route candidate set, wherein the endpoint airport level includes the level of the origin airport and the level of the destination airport. The second stage includes: detecting the spatial distance between each candidate route in the first route candidate set and the target route, and selecting the k routes with the smallest spatial distance as the candidate route sample set, where k is a positive integer less than or equal to m; The multi-level progressive evaluation model includes the following three progressive network layers, wherein: The first progressive network layer is used to determine the basic matching degree between the target route and the candidate route based on the latitude and longitude coordinates of the endpoint airports in the basic features and the route distance similarity and route geographical location similarity in the cross features. The second progressive network layer is used to differentiate and correct the basic matching degree based on the basic matching degree output by the first progressive network layer and the similarity of route direction and similarity of travel demand density in the cross features, and output the corrected matching degree. The third progressive network layer is used to adjust the corrected matching degree based on the corrected matching degree output by the second progressive network layer and the similarity of passenger type, similarity of population mobility, and similarity of high-speed railway network competition intensity index in the cross features, and output the final matching degree. The similarity of air travel demand intensity, passenger type, population mobility, and high-speed rail network competition intensity index is obtained by using either a categorical similarity calculation method or a numerical similarity calculation method based on the air travel demand intensity level, passenger type level, population mobility level, and high-speed rail network competition intensity index of the endpoint cities of the target route and the candidate routes, respectively.

2. The method according to claim 1, characterized in that, Before selecting the k routes with the smallest spatial distances as the candidate route sample set, the method further includes: Obtain the spatial distances between all candidate routes in the first route candidate set and the target route to form a spatial distance sequence; Calculate the mean and standard deviation of the spatial distance sequence; A distance threshold is determined based on the mean and standard deviation of the spatial distance sequence, wherein the distance threshold is equal to the sum of the mean and n times the standard deviation, and n is a preset positive integer coefficient; The number of candidate routes in the first route candidate set whose spatial distance is less than or equal to the distance threshold is counted, and this number is used as the value of k.

3. The method according to claim 1, characterized in that, Detecting the spatial distance between each candidate route in the first route candidate set and the target route includes: Obtain the first distance between the origin airport of the target route and the origin airport of the candidate route; Obtain the second distance between the destination airport of the target route and the destination airport of the candidate route; The first distance and the second distance are added together to obtain the spatial distance between the candidate route and the target route.

4. The method according to claim 3, characterized in that, Obtaining the first distance between the origin airport of the target route and the origin airport of the candidate route includes: Obtain the latitude and longitude differences between the origin airports corresponding to the target route and the candidate routes, respectively; The first distance is determined based on the latitude difference, the longitude difference, and the average radius of the Earth.

5. The method according to claim 1, characterized in that, Based on the endpoint airport information of the target route, determine the geospatial similarity features between the target route and each candidate route in the candidate route sample set, including: Calculate the similarity between the flight distance of the target route and the flight distance of the candidate route to obtain the route distance similarity between the target route and the candidate route; The geographical similarity between the target route and the candidate route is determined based on the distance between the origin airport of the target route and the origin airport of the candidate route, and the distance between the destination airport of the target route and the destination airport of the candidate route. A target direction vector is constructed based on the latitude and longitude differences between the origin and destination airports of the target route; a candidate direction vector is constructed based on the latitude and longitude differences between the origin and destination airports of the candidate routes. Calculate the dot product of the target direction vector and the candidate direction vector, and determine the route direction similarity between the target route and the candidate route based on the dot product and the magnitudes corresponding to the target direction vector and the candidate direction vector, respectively. The route distance similarity, the route geographical location similarity, and the route direction similarity are used as the geospatial similarity features.

6. The method according to claim 1, characterized in that, The multi-level progressive evaluation model is pre-trained through the following training steps: Obtain a training dataset, which includes endpoint airport information for multiple existing routes and true matching degree labels between each route pair; Using the training dataset as input, the two-stage business rule filtering operation is performed to generate a training sample set; Construct multi-dimensional features between each flight path pair in the training sample set; Using the multi-dimensional features of the training sample set as input and the true matching degree label as supervision signal, the model parameters are updated through a dual-objective joint optimization strategy to obtain a trained multi-level progressive evaluation model.

7. The method according to claim 6, characterized in that, The dual-objective joint optimization strategy includes: Obtain the true matching degree and predicted matching degree of each training sample in the training sample set; Calculate the square of the difference between the true matching degree and the predicted matching degree of the training sample to obtain the first square value corresponding to the training sample; The mean squared error is obtained by summing the first squared values ​​corresponding to all training samples and dividing by the total number of training samples. Obtain the true rank and predicted rank of each training sample in the training sample set; Calculate the square of the difference between the true ranking rank and the predicted ranking rank of the training sample to obtain the second square value corresponding to the training sample; The second squared values ​​corresponding to all training samples are summed, and the rank correlation coefficient is determined based on the summation result and the total number of training samples. The rank correlation coefficient is used to measure the consistency between the predicted ranking and the actual ranking of the training samples. The larger the value of the rank correlation coefficient, the higher the ranking consistency. A joint loss function is constructed based on the mean squared error and the rank correlation coefficient; The dual-objective joint optimization strategy is generated with minimizing the joint loss function as the training objective.

8. The method according to claim 1, characterized in that, After determining the matching degree between the target route and each candidate route using the multi-level progressive evaluation model, the method further includes: The matching degree between the target route and each candidate route is sorted from high to low, and the route is divided into at least two route levels according to the sorting results. The first route level is the top M candidate routes with the highest matching degree, and the second route level is the next N candidate routes with the second highest matching degree, where M and N are positive integers. A first aggregation weight is assigned to each candidate route at the first route level, and a second aggregation weight is assigned to each candidate route at the second route level, wherein the first aggregation weight is greater than the second aggregation weight. The matching degree between each candidate route at the first route level and the target route is multiplied by the corresponding first aggregation weight and then summed to obtain the first summation result; The matching degree between each candidate route at the second route level and the target route is multiplied by the corresponding second aggregation weight and then summed to obtain the second summation result; The first summation result and the second summation result are added together to obtain a weighted aggregate value, which is used as the comprehensive matching reference value for the target route.

9. The method according to claim 1, characterized in that, The basic features also include a first type of numerical urban statistical indicators and a second type of categorical urban attribute indicators; the cross features also include numerical similarity calculated based on the first type of numerical urban statistical indicators and categorical similarity calculated based on the second type of categorical urban attribute indicators. The first category of numerical urban statistical indicators includes at least one of the following: urban resident population, urban road network density, and high-speed railway network competition intensity index; the second category of categorized urban attribute indicators includes at least one of the following: the geographical region to which the city belongs, the city's leading industry category, the main travel destinations of urban air passengers, the intensity level of urban air travel demand, and the intensity level of urban population flow.

10. The method according to claim 9, characterized in that, The numerical similarity is calculated through the following steps: For the i-th indicator in the first type of numerical city statistical indicators, obtain the indicator values ​​of the starting city and the ending city of the target route with respect to the i-th indicator, and obtain the indicator values ​​of the starting city and the ending city of the candidate route with respect to the i-th indicator, where i is an integer greater than or equal to 1. Calculate the similarity of the index values ​​of the i-th indicator between the origin cities of the target route and the candidate routes, and use it as the origin similarity for the i-th indicator; Calculate the similarity of the index values ​​of the target route and the candidate routes with respect to the destination cities for the i-th index, and use it as the destination similarity for the i-th index; Calculate the average of the starting point similarity and the ending point similarity for the i-th indicator to obtain the numerical similarity of the i-th indicator.

11. The method according to claim 10, characterized in that, The categorical similarity is calculated through the following steps: For each indicator in the second category of city attribute indicators, each classification level of the indicator is pre-mapped to a corresponding numerical code in a preset order, so that each level corresponds to a unique code value. Based on the indicator level of the city of origin of the target route, query the numerical code that has a mapping relationship with the indicator level to obtain the origin code value and destination code value of the target route. Based on the indicator level of the candidate route origin city, query the numerical code that has a mapping relationship with the indicator level to obtain the candidate route origin code value and candidate route destination code value. Calculate the absolute value of the difference between the origin code value of the target route and the origin code value of the candidate route, and use it as the origin code absolute difference; Divide the absolute difference of the starting code by the range of code values ​​to obtain the starting code distance, wherein the range of code values ​​is equal to the difference between the maximum and minimum code values; Calculate the absolute value of the difference between the destination code value of the target route and the destination code value of the candidate route, and use it as the destination code absolute difference; divide the destination code absolute difference by the range of code values ​​to obtain the destination code distance; The average coding distance is obtained by calculating the average coding distance between the starting point and the ending point. Then, the class similarity is calculated, where the class similarity is equal to 1 minus the average coding distance.

12. A multi-level evaluation device for route matching degree, characterized in that, include: The acquisition unit is used to acquire the endpoint airport information of the target route, which includes the airport names and latitude and longitude coordinates of the origin airport and the destination airport. The filtering processing unit is used to filter out a set of candidate route samples that match the target route from the existing route database; The extraction and processing unit is used to extract multi-dimensional features between the target route and each candidate route in the candidate route sample set based on the endpoint airport information of the target route. The multi-dimensional features include basic features and cross features. The basic features include at least the latitude and longitude coordinates of the endpoint airports corresponding to the target route and the candidate routes, and the cross features include at least the geospatial similarity features between the target route and the candidate routes. The model processing unit is used to input the multi-dimensional features into a pre-trained multi-level progressive evaluation model, and the multi-level progressive evaluation model outputs the matching degree between the target route and each candidate route; wherein, the multi-level progressive evaluation model is a neural network model trained by supervised learning, using the multi-dimensional features of existing routes as training samples and the passenger volume similarity between existing routes as training labels. The filtering processing unit includes a two-stage filtering subunit, used to perform two-stage business rule filtering on the target route, filtering out a candidate route sample set that matches the target route from the existing route database. The two-stage business rule filtering includes the following first stage and second stage executed sequentially. The two-stage filtering subunit includes a first filtering submodule, used to execute the first stage, including: scale matching filtering based on airport throughput level, obtaining the endpoint airport level of the target route, and filtering out m routes from the existing route database whose endpoint airport level is the same as or one level different from the endpoint airport level of the target route to form a first route candidate set, wherein the endpoint airport level includes the level of the origin airport and the level of the destination airport; and a second filtering submodule, used to execute the second stage, including: detecting the spatial distance between each candidate route in the first route candidate set and the target route, and selecting the k routes with the smallest spatial distance as the candidate route sample set, wherein k is a positive integer less than or equal to m. The multi-level progressive evaluation model includes the following three progressive network layers, wherein: The first progressive network layer is used to determine the basic matching degree between the target route and the candidate route based on the latitude and longitude coordinates of the endpoint airports in the basic features and the route distance similarity and route geographical location similarity in the cross features. The second progressive network layer is used to differentiate and correct the basic matching degree based on the basic matching degree output by the first progressive network layer and the similarity of route direction and similarity of travel demand density in the cross features, and output the corrected matching degree. The third progressive network layer is used to adjust the corrected matching degree based on the corrected matching degree output by the second progressive network layer and the similarity of passenger type, similarity of population mobility, and similarity of high-speed railway network competition intensity index in the cross features, and output the final matching degree. The similarity of air travel demand intensity, passenger type, population mobility, and high-speed rail network competition intensity index is obtained by using either a categorical similarity calculation method or a numerical similarity calculation method based on the air travel demand intensity level, passenger type level, population mobility level, and high-speed rail network competition intensity index of the endpoint cities of the target route and the candidate routes, respectively.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device on which the computer-readable storage medium is located performs the multi-level evaluation method for route matching degree as described in any one of claims 1 to 11.

14. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the multi-level evaluation method for route matching degree as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the multi-level evaluation method for route matching degree as described in any one of claims 1 to 11.

Images