Optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost

Abstract

The invention discloses an optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost, and the method comprises the steps: building an XGBoost prediction model through employing an existing algorithm which can directly or indirectly solve multi-keyword coverage optimal path query, combining road network features and query features, and relying on a large number of historical queries; and predicting specified algorithm planning paths for different users. According to the method, an existing algorithm is divided into two types, the algorithm based on the POI candidate set is split into components, different components are combined into a reconstruction algorithm through component reconstruction, and the travel requirements of different users are met. For an algorithm based on path expansion, a path refinement and pruning optimization strategy is introduced, and an adjacent POI point refinement strategy is designed, so that the precision of the path is improved, a relation pruning strategy and a batch pruning strategy are dominated, the response time of the algorithm is effectively prolonged, and the query efficiency is improved. And in combination with historical query, feature extraction is carried out to construct a prediction model based on XGBoost, and an optimal algorithm is predicted and a path is planned for a query user.

Description

Technical Field

[0001] The present invention belongs to the technical field of spatiotemporal data management, and in particular relates to an optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost. Background Art

[0002] With the popularity of smart devices, location-based social networks, and spatial databases, a large amount of trajectory data in the road network has attracted widespread attention. In addition, users' daily travel is also inseparable from these trajectory data. For example, there are shortest distance query algorithms based on CH contraction technology and H2H tree decomposition technology. However, with the gradual refinement of user personalized needs, the shortest distance query under a single dimension obviously cannot meet the diverse needs of users. Therefore, the optimal path query under multi-keyword coverage further considers the keyword information of points of interest (POI) in the road network, and plans the shortest path for users to meet the keyword information of the user's specified travel needs.

[0003] The existing algorithms for solving the optimal path under multi-keyword coverage can be mainly divided into two categories: algorithms based on POI candidate sets and algorithms based on path expansion. The algorithm based on POI candidate sets first obtains the POI candidate set that meets the query according to the given query keywords, aiming to narrow the search space, and then constructs the optimal path according to its greedy strategy. The algorithm based on path expansion starts from the starting point and gradually constructs a large number of candidate paths that meet the query keywords until the optimal path that meets all query keywords is constructed. However, both methods have their own advantages and disadvantages: the algorithm based on candidate sets first narrows the search space by filtering POI points, and greedily constructs the path after obtaining the candidate set. Therefore, the algorithm based on POI candidate sets is often faster in response time. However, the greedy strategy to construct the path will cause the path to fall into the local optimal solution, resulting in a longer path distance. In order to plan a shorter path, the algorithm based on path expansion gradually expands the path to different POIs through path expansion, which increases the search range, so the algorithm based on path expansion takes a longer time to plan the path. Therefore, there is an obvious trade-off problem between response time and path accuracy in the existing algorithms.

[0004] From the user's experience point of view, a high-precision path should be returned to the user in a short time. However, the existing single algorithm obviously cannot meet this requirement. Specifically, if the path planning system quickly returns a low-precision route, it may increase the user's travel time. On the contrary, if the path planning system responds to a high-precision path after a long time, it will also affect the user's travel experience to a certain extent. Summary of the invention

[0005] In order to further improve the user's travel experience and return a shorter route in a short time, the present invention proposes a method for constructing an algorithm integration library based on component reconstruction and XGBoost.

[0006] The method for constructing an algorithm integrated library based on component reconstruction and XGBoost provided by the present invention comprises the following steps:

[0007] Step 1: Classify the existing algorithms for solving the optimal path under multi-keyword coverage into two categories: algorithms based on POI candidate sets and algorithms based on path expansion. The algorithms in these two categories are used as the initial algorithms in the algorithm integration library;

[0008] Step 2: According to the path construction process of the algorithm based on the POI candidate set, firstly decompose it into two types of components: point filtering component and point connection component, and optimize the point filtering component and the point connection component respectively; finally, reconstruct the algorithm based on the POI candidate set through the optimized point filtering component and point connection component to generate a reconstruction algorithm based on the POI candidate set;

[0009] Step 3: For algorithms based on path extension, the H2H index structure is first integrated to improve query efficiency; secondly, neighboring POI refinement strategies, dominant relationship pruning and batch pruning strategies are introduced to different algorithms based on path extension to improve path extension efficiency, and an optimization algorithm based on path extension is generated;

[0010] Step 4: Incorporate road network features and query features, introduce optimal algorithm evaluation criteria under different query periods based on historical query data, and construct a prediction model training set; in the prediction model training set, the optimal algorithm for each historical query belongs to one of the reconstruction algorithm based on the POI candidate set and the optimization algorithm based on path extension; then use the XGBoost algorithm to build an optimal algorithm prediction model; predict the optimal algorithm based on the characteristic attributes of the user's query, and use this optimal algorithm to plan the path for the user.

[0011] Furthermore, in step 2, the initial algorithm based on the POI candidate set is disassembled into components, and the reconstruction method is as follows:

[0012] Step 2.1: Component disassembly of the steps of constructing a path using the initial algorithm based on the POI candidate set; Specifically, the initial algorithm based on the POI candidate set first uses the POI preprocessing method TMD or the spatial index structure IGTree index, Grid index, to filter out a large number of POI points that cannot construct the optimal path in the global road network according to the spatial positions of the query start and end points, and adds the POI points that meet the query keyword requirements to the POI candidate set C. POIThen, the initial algorithm based on the POI candidate set starts from the starting point and constructs the path according to two different greedy strategies. According to the above process of first filtering POI points and then constructing the path, an initial algorithm based on the POI candidate set is decomposed into two components. The method of filtering POI points is the point filtering component, and the method of constructing the path by the greedy strategy is the point connection component.

[0013] Step 2.2, integrating the point filtering component in step 2.1 into the H2H index to accelerate the shortest distance query that needs to be called multiple times during the POI filtering process, thereby forming an optimized point filtering component;

[0014] Step 2.3, optimizing the point connection component in step 2.1 through a parallel greedy strategy to form an optimized point connection component;

[0015] Step 2.4: Combine each optimized point filter component and point connection component to reconstruct the reconstruction algorithm based on the POI candidate set. There are 3 point filter components and 2 point connection components in the algorithm library. Finally, through combined reconstruction, 6 reconstruction algorithms based on the POI candidate set are formed.

[0016] Furthermore, the point connection components are optimized by parallel greedy strategies, including optimizing the point connection components by using the parallel "nearest neighbor greedy" strategy NN-Parallel, or optimizing the point connection components by using the parallel "maximum service quantity greedy" strategy MS-Parallel; wherein,

[0017] 1) Optimize the point connection component using the parallel nearest neighbor greedy strategy NN-Parallel. Specifically, given a query Q(s, t, {kw…}), s is the query starting point, t is the query destination, and {kw…} is the query keyword set; NN-Parallel starts from the starting point s and the end point t in parallel. POI Get the POI point closest to the starting point that can satisfy the query keyword, add it to the path, and then iterate from C POI Get POI points until all query keywords are satisfied; finally, select the path with the shortest path distance as the result;

[0018] 2) Use the greedy MS-Parallel with the maximum number of parallel services to optimize the point connection components; specifically, MS-Parallel starts from the starting point s and the end point t, first from C POI The POI point that can satisfy the most unsatisfied query keywords is obtained, and then the next POI point that can satisfy the most query keywords is obtained iteratively until all query keywords are satisfied; finally, the path with the shortest path distance is selected as the result;

[0019] Furthermore, the method of integrating H2H index, neighboring POI refinement strategy, dominant relationship pruning and batch pruning strategy in step 3 is as follows:

[0020] Step 3.1, integrate H2H index structure;

[0021] Step 3.2: Based on the integrated H2H index, for the path extension-based algorithm EPE, a path refinement strategy for neighboring POI points is introduced to shorten the path distance between a pair of adjacent POI points in the initial result path, further reducing the path length;

[0022] For the path extension-based algorithms PNE and PNE-A*, a dominance relationship pruning strategy is introduced. During the path extension process, path extension is terminated in advance for some paths that have been dominated by other paths, thereby improving the execution efficiency of the algorithm.

[0023] For the path extension-based algorithm VDE, since the VDE algorithm starts from the starting point s and follows the shortest path sp(s,t), it initiates a neighbor query for each deviation point on the shortest path to construct a candidate path.

[0024] Furthermore, in step 4, XGBoost is used to build the optimal algorithm prediction model. The specific process is as follows:

[0025] Step 4.1, incorporate comprehensive road network features and query features;

[0026] Step 4.2: First, obtain a large number of historical queries. Then, for each historical query, execute each algorithm based on the POI candidate set and the path expansion algorithm, and record the algorithm response time Q. time and path distance Q dis ;

[0027] Step 4.3, introduce the optimal algorithm evaluation criteria, divide the time period when users initiate queries into regular time period and peak time period; in the peak time period, define the upper limit of the algorithm response time α, then, the algorithms with response time lower than α are first excluded; finally, for each algorithm with response time lower than α, assign the algorithm that can generate the shortest path to the current query; in the regular period, the number of user-initiated queries is small, but the algorithm execution time and path distance must still be considered at the same time; therefore, first define the algorithm response time threshold β of the regular period, then, the algorithms with response time exceeding the threshold β are first excluded; for the remaining algorithms, consider the response time and path distance of each algorithm at the same time, introduce the weight parameter γ, perform weighted calculation, and obtain the score A of each algorithm. score ; Finally, A score The algorithm with the lowest score is regarded as the optimal algorithm for the current query; A score The calculation method is as follows, where D minis the shortest path distance of the current remaining algorithm, T min is the fastest response time of the current remaining algorithms;

[0028]

[0029] Step 4.4: Given a new query, first, based on the optimal algorithm prediction model, combined with the current road network characteristics and query characteristics, predict the optimal algorithm for this query, and then use the optimal algorithm to build a path and return it to the user.

[0030] In the actual large-scale road network environment: This solution introduces more road network features and query features as well as the XGBoost prediction model into the shortest path query algorithm under multi-keyword coverage for the first time. After verification in actual road networks in Asia, America and other regions, the component combination and reconstruction strategy proposed in this solution can further improve the path accuracy and response time of the existing algorithm. For the algorithm EMB based on path expansion, the path distance after adding the optimization strategy is reduced by 3-10% on average. After adding the dominant relationship pruning, the response time of the PruningKOSR algorithm is reduced by 1-2 orders of magnitude. After adding the batch pruning strategy, the response time of the DA-Prune algorithm is reduced by 75%; when processing real query requests, the path constructed by the algorithm predicted by the XGBoost optimal algorithm prediction model has obvious advantages in both response time and path distance dimensions. During the query peak period, the path accuracy constructed by the prediction model is 5% higher than that of the candidate set-based algorithm. Compared with the path expansion-based algorithm, the response time is improved by one order of magnitude. During the query regular period, the path accuracy is improved by 8% compared with the POI candidate set-based algorithm and 2% compared with the path expansion-based algorithm. In addition, the response time is also shortened by 50% compared with the path expansion-based algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0032] Figure 1 A framework diagram of the optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost provided by the present invention;

[0033] Figure 2 is a road network diagram in the embodiment;

[0034] Figure 3 Schematic diagram of TMD pre-calculation in the embodiment;

[0035] Figure 4 Schematic diagram of IGTree index in the embodiment;

[0036] Figure 5Schematic diagram of Grid index in the embodiment;

[0037] Figure 6 Schematic diagram of the neighboring POI point path refinement strategy in the embodiment;

[0038] Figure 7 Schematic diagram of dominance relationship pruning in the embodiment;

[0039] Figure 8 Schematic diagram of batch pruning in the embodiment;

[0040] Fig. 9 Schematic diagram of the optimal algorithm evaluation criteria in the embodiment. DETAILED DESCRIPTION

[0041] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0042] The present invention provides a method for constructing an optimal path algorithm integrated library under multi-keyword coverage based on component reconstruction and XGBoost.

[0043] refer to Figure 1 , the present invention utilizes two existing algorithms that can solve the optimal path under multi-keyword coverage: an algorithm based on POI candidate sets and an algorithm based on path expansion, and takes into account the road network characteristics and query characteristics that affect query efficiency. A combined reconstruction and pruning optimization strategy is proposed, and an optimal algorithm prediction model is set up with the help of historical queries and XGBoost. Finally, the optimal algorithm prediction model is used to predict the optimal algorithm for the current query, and this algorithm is selected from the algorithm integration library to plan the path. The embodiments of the method are as follows:

[0044] Example:

[0045] by Figure 2 Take the weighted undirected graph network G(V,E) as an example

[0046] Classify the existing algorithms for solving the optimal path under multi-keyword coverage;

[0047] Some existing algorithms can solve the optimal path query under multi-keyword coverage, while some do not support the optimal path query under multi-keyword coverage. However, it is found that some initial algorithms can also support the optimal path query under multi-keyword coverage after integrating the latest optimization technology. Finally, two algorithms that can solve the optimal path query under multi-keyword coverage are obtained, such as Figure 1 As shown in;

[0048] For the point filtering component TMD, given a multi-keyword coverage optimal path query Q(v 1 ,v 12,{kw 1 ,kw 2}), the starting point s is v 1 , the end point t is v 12 , the query keyword set is {kw 1 ,kw 2 TMD calculates the sum of the shortest distances δ from each POI point containing the query keyword to the starting point and the end point T (s,p,t)=δ(s,p)+δ(p,t). Then get the first k shortest δ T The POI point (s, p, t) is used as a candidate point for the current query keyword. Then it is added to the POI candidate set C. POI In. Figure 3 As shown, it contains the query keyword kw 1 The POI point set is O(kw 1 )={v 2 ,v 5 ,v 8 ,v 9}, then calculate the delta of each POI point T (s,p,t), after calculation, the distances are 34,27,29,28 respectively. Set the k value to 2. Finally, kw is satisfied 1 POI point v 5 ,v 9 Added to C POI middle.

[0049] For the point filtering component IGTree, such as Figure 4 As shown, given a query Q(v 7 ,v 12 ,{kw 1 ,kw 2 ,kw 3}). First, the original road network is iteratively divided into subgraphs according to METIS, and then the subgraphs are managed using a tree structure to form an IGTree index. Then, given a query Q, IGTree first obtains its starting point v 7 and the end point v 12 The subgraph partition is located, and then the LCA of its minimum common ancestor is obtained. At this time, it is determined whether the LCA subgraph has satisfied all the query keywords. If the LCA does not satisfy all the query keywords, the parent node of the LCA is iteratively obtained until all the query keywords are satisfied. Finally, all the POI points in the LCA subgraph that meet the query keywords are added to C POI middle.

[0050] For the point filtering component Grid, such as Figure 5 As shown, given Q(v 1 ,v 12 ,{kw1 ,kw 2 ,kw 4}), Grid includes two rounds of grid expansion. The purpose of the first round of expansion is to obtain the necessary waypoints v 6 , because only v 6 Can satisfy the kw in the query 4 This keyword then specifies the second round of grid expansion, along v 1 to v 6 , v 6 to v 12 Perform bidirectional grid expansion until all query keywords are met.

[0051] For the path extension algorithm EPE, the path refinement strategy of neighboring POI points is introduced, such as Figure 6 As shown in the figure, the strategy of EPE algorithm to construct the path is to insert the keywords that satisfy the query into the shortest path sp(s, t). In order to determine the POI point to be inserted into sp(s, t), EPE algorithm uses AMD = δ(p j ,p * )+δ(p * ,p j+1 )-δ(p j ,p j+1 ) is used to calculate the detour distance of the POI point, and then insert the POI point between the point pairs with the shortest detour distance. Given a query Q(v 1 ,v 12 ,{kw 1 ,kw 2 ,kw 4}), the path P planned by EPE EPE P EPE = <v 1 ,v 3 ,v 4 ,v 7 ,v 6 ,v 9 ,v 10 ,v 12 >. Figure 6 , we can get v 3 and v 6 The shortest path between two points sp(v 3 ,v 6 ) is shorter than <v 3 ,v 4 ,v 7 ,v 6 >. Therefore, after refining the strategy of neighboring POI points, P EPE for <v 1 ,v 3 ,v 6,v 9 ,v 10 ,v 12 >.

[0052] For the path extension-based algorithm PNE, the dominant relationship pruning is introduced, such as Figure 7 As shown, there are two paths P a = <v 1 ,v 5 ,v 3 >,P b = <v 1 ,v 2 ,v 1 ,v 3 >. According to Figure 2 Keyword information, you can find P a and P b All satisfy the same keyword set {kw 1 ,kw 2}. b The length of the path is greater than P a , so P b Paths can be pruned so that no further path expansion is performed.

[0053] For the path extension-based algorithm VDE, such as Figure 8 As shown in Figure 2. The VDE algorithm first obtains sp(s, t), and then iteratively executes the deviation strategy along each point on the shortest path. In order to further improve the efficiency of deviation, a batch pruning strategy is introduced. Assume that the current deviation point is v d , by performing a neighbor query, we obtain v d The nearest neighbor v NN ,At this point, it can be found that there is a shortest path sp(v d ,v NN )= <v d ,v a ,v b ,v NN > Arrive at v NN , and v a ,v b These two points are exactly on sp(v d ,t) as the subsequent deviation point. Therefore, v a ,v b These two points are batch pruned without deviation.

[0054] The above component reconstruction algorithm and the pruning optimization algorithm together constitute Figure 1 Then, with the help of a large number of historical queries, the time periods when users initiate queries are divided into peak periods and regular periods, and a historical query dataset is constructed. Fig. 9 As shown, during peak hours, the upper limit of the algorithm response time is defined as α = 0.15, and then the algorithms with response time lower than α are excluded first. Finally, for each algorithm with response time lower than α, the algorithm ⑥Grid-MS that can generate the shortest path is assigned to the current query. In regular cycles, the number of user-initiated queries is small, but the execution time and path distance of the algorithm must still be taken into account. Therefore, the algorithm response time threshold β = 1.0 for the regular cycle is first defined, and then the algorithms with response time exceeding the threshold β are excluded first. For the remaining algorithms, the response time and path distance of each algorithm are considered at the same time, and the weight parameter γ is introduced to perform weighted calculations to obtain the score A of each algorithm. score Finally, A score The algorithm with the lowest score ⑦EPE is selected as the optimal algorithm for the current query. score The calculation method is as follows, where D min is the shortest path distance of the current remaining algorithm, T min is the fastest response time of the currently remaining algorithms.

[0055]

[0056] Subsequently, the historical queries of each assigned optimal algorithm are used as data, and the optimal algorithm prediction model is constructed using the XGBoost algorithm. Finally, given a new query, first, the optimal algorithm is predicted for this query based on the optimal algorithm prediction model, combined with the current road network characteristics and query characteristics, and then the optimal algorithm is used to construct the path and returned to the user.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost is characterized by: The steps include: Step 1: Classify the existing algorithms for solving the optimal path under multi-keyword coverage into two categories: algorithms based on POI candidate sets and algorithms based on path expansion, and use the algorithms in these two categories as the initial algorithms in the algorithm integration library; Step 2: According to the path construction process of the algorithm based on the POI candidate set, firstly decompose it into two types of components: point filtering component and point connection component, and optimize the point filtering component and the point connection component respectively; finally, reconstruct the algorithm based on the POI candidate set through the optimized point filtering component and point connection component to generate a reconstruction algorithm based on the POI candidate set; Step 3: For algorithms based on path extension, the H2H index structure is first integrated to improve query efficiency; secondly, neighboring POI refinement strategies, dominant relationship pruning and batch pruning strategies are introduced to different algorithms based on path extension to improve path extension efficiency, and an optimization algorithm based on path extension is generated; Step 4: Incorporate road network features and query features, introduce optimal algorithm evaluation criteria under different query periods based on historical query data, and construct a prediction model training set; in the prediction model training set, the optimal algorithm for each historical query belongs to one of the reconstruction algorithm based on the POI candidate set and the optimization algorithm based on path extension; then use the XGBoost algorithm to build an optimal algorithm prediction model; predict the optimal algorithm based on the characteristic attributes of the user's query, and use this optimal algorithm to plan the path for the user.

2. The optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost according to claim 1, characterized in that: Step 2: Decompose the initial algorithm based on the POI candidate set into components. The reconstruction method is as follows: Step 2.1: Component disassembly of the steps of constructing a path using the initial algorithm based on the POI candidate set; Specifically, the initial algorithm based on the POI candidate set first uses the POI preprocessing method TMD or the spatial index structure IGTree index, Grid index, to filter out a large number of POI points that cannot construct the optimal path in the global road network according to the spatial positions of the query start and end points, and adds the POI points that meet the query keyword requirements to the POI candidate set C. POI Then, the initial algorithm based on the POI candidate set constructs the path from the starting point according to two different greedy strategies; According to the above process of first filtering POI points and then constructing the path, an initial algorithm based on the POI candidate set is decomposed into two components; The method of filtering POI points is the point filtering component, and the method of constructing paths with the greedy strategy is the point connection component; Step 2.2, integrating the point filtering component in step 2.1 into the H2H index to accelerate the shortest distance query that needs to be called multiple times during the POI filtering process, thereby forming an optimized point filtering component; Step 2.3, optimizing the point connection component in step 2.1 through a parallel greedy strategy to form an optimized point connection component; Step 2.4, combine each optimized point filtering component and point connection component to reconstruct the reconstruction algorithm based on the POI candidate set; there are 3 point filtering components and 2 point connection components in the algorithm library. Finally, through combined reconstruction, 6 reconstruction algorithms based on the POI candidate set are formed.

3. The optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost according to claim 2, characterized in that: The point connection components are optimized by parallel greedy strategies, including using the parallel "nearest neighbor greedy" strategy NN-Parallel to optimize the point connection components, or using the parallel "maximum service number greedy" MS-Parallel to optimize the point connection components; wherein, 1) Optimize the point connection component using the parallel nearest neighbor greedy strategy NN-Parallel. Specifically, given a query Q(s, t, {kw...}), s is the query starting point, t is the query destination, and {kw...} is the query keyword set; NN-Parallel starts from the starting point s and the end point t in parallel. POI Get the POI point closest to the starting point that can satisfy the query keyword, add it to the path, and then iterate from C POI Get POI points until all query keywords are satisfied; finally, select the path with the shortest path distance as the result; 2) Use the greedy MS-Parallel with the maximum number of parallel services to optimize the point connection components; specifically, MS-Parallel starts from the starting point s and the end point t, first from C POI The POI point that can satisfy the most unsatisfied query keywords is obtained, and then the next POI point that can satisfy the most query keywords is obtained iteratively until all query keywords are satisfied; finally, the path with the shortest path distance is selected as the result.

4. The optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost according to claim 1, characterized in that: Step 3 integrates H2H index, neighboring POI refinement strategy, dominant relationship pruning and batch pruning strategy as follows: Step 3.1, integrate H2H index structure; Step 3.2: Based on the integrated H2H index, for the path extension-based algorithm EPE, a path refinement strategy for neighboring POI points is introduced to shorten the path distance between a pair of adjacent POI points in the initial result path, further reducing the path length; For the path extension-based algorithms PNE and PNE-A*, a dominance relationship pruning strategy is introduced. During the path extension process, path extension is terminated in advance for some paths that are already dominated by other paths, thereby improving the execution efficiency of the algorithm. For the path extension-based algorithm VDE, since the VDE algorithm starts from the starting point s and follows the shortest path sp(s, t), it initiates a neighbor query for each deviation point on the shortest path to construct a candidate path.

5. The optimal path query method under multi-keyword coverage based on component reconstruction and XGBoost according to claim 1, characterized in that: In step 4, XGBoost is used to build the optimal algorithm prediction model. The specific process is as follows: Step 4.1, incorporate comprehensive road network features and query features; Step 4.2: First, obtain a large number of historical queries. Then, for each historical query, execute each algorithm based on the POI candidate set and the path expansion algorithm, and record the algorithm response time Q. time and path distance Q dis ; Step 4.3, introduce the optimal algorithm evaluation criteria, and divide the time period when users initiate queries into regular time period and peak time period; During peak hours, an upper bound α of the algorithm response time is defined, and then algorithms with response time lower than α are excluded first; finally, for each algorithm with response time lower than α, the algorithm that can generate the shortest path is assigned to the current query; during regular periods, the number of queries initiated by users is small, but the execution time and path distance of the algorithm must still be considered at the same time; therefore, a threshold β of the algorithm response time during regular periods is first defined, and then algorithms with response time exceeding the threshold β are excluded first; for the remaining algorithms, the response time and path distance of each algorithm are considered at the same time, and a weight parameter γ is introduced to perform weighted calculation to obtain the score A of each algorithm. score ; Finally, A score The algorithm with the lowest score is regarded as the optimal algorithm for the current query; A score The calculation method is as follows, where D min is the shortest path distance of the current remaining algorithm, T min is the fastest response time of the current remaining algorithms; Step 4.4: Given a new query, first, based on the optimal algorithm prediction model, combined with the current road network characteristics and query characteristics, predict the optimal algorithm for this query, and then use the optimal algorithm to build a path and return it to the user.

Images