A method for constructing station spatial relationships based on bus card swiping data

Abstract

The present invention discloses a method for constructing the spatial relationship of stations based on a bus card swiping data set. First, a physical graph is constructed based on the bus card swiping data to represent the spatial dependency of stations. Then, a hierarchical graph is further constructed by aggregation based on the physical graph. The beneficial effect of the present invention is that the real physical graph of bus routes is fully mined from the original data of bus card swiping, reflecting the real spatial dependency of stations in the bus card swiping data. Furthermore, by converting the bus stations from fine-grained to coarse-grained through physical graph aggregation, and converting the large-scale physical graph into a hierarchical graph, the resource consumption of downstream tasks (such as traffic flow prediction tasks, etc.) can be reduced, and the computational efficiency of downstream tasks can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of spatiotemporal data mining, and in particular to a method for constructing site spatial relationships based on bus card swiping data. Background Art

[0002] Bus card swipe data is a crucial component of urban transportation management systems. It reflects the topological information and passenger flow of urban bus stops, and is a form of spatiotemporal data. Mining the spatial correlations (physical graphs, similarity graphs, functional graphs, hierarchical graphs, etc.) and temporal correlations (adjacency, periodicity, and trends) of bus card swipe data is crucial for bus passenger flow forecasting. A physical graph of a station (hereafter referred to as a "physical graph") is a graph structure that uses matrix rows and columns to represent the physical topological relationships of stations and uses matrix values ​​to store the relationships between stations. Physical graphs mathematically represent the multi-granular, multi-level spatiotemporal characteristics of stations and are widely used in bus passenger flow forecasting.

[0003] Physical maps for passenger flow forecasting are primarily constructed using existing trajectory data or crawled route data. In urban traffic management systems, bus stops can change dynamically due to external factors (such as road maintenance and large-scale events), while crawled route data is static and cannot accurately reflect the data's true spatial dependencies.

[0004] In addition, when performing graph convolution calculations on large-scale physical graphs, it may lead to low computational efficiency and high resource consumption.

[0005] (1) Physical map construction

[0006] Reference 1, "Liu L, Chen J, Wu H, et al. Physical-virtual collaboration modeling for intra- and inter-station metro ridership prediction [J]. IEEE Transactions on Intelligent Transportation Systems, 2020," directly constructs a physical graph based on the existing metro station route list. The relationships between stations are represented by an adjacency matrix, and the normalized result of each row of the matrix serves as the weight of each edge in the physical graph. In addition to constructing a physical graph, this method also constructs similarity and association graphs based on the similarity and correlation of passenger flows. However, when the bus route topology is complex or dynamically changing, it is difficult to efficiently and accurately represent the spatial dependencies between stations.

[0007] Reference 2 "Geng X, Li Y, Wang L, et al. Spatiotemporal multi-graph convolution network for ride-hailing demand forecasting [C] / / Proceedings of the AAAI conference on artificial intelligence. 2019, 33(01): 3656-3663." Based on the taxi order data in Beijing and Shanghai and the road network data for traffic connectivity assessment provided by OpenStreetMap, a region is connected to 8 adjacent regions in a 3×3 grid to construct a graph to represent the adjacency relationship between stations and build a physical map of the ride-hailing stations. This method uses taxi order data combined with road network data to construct a physical map from real data. It can be seen from this that the construction of a physical map is very important for representing the spatial dependencies of spatiotemporal data.

[0008] (2) Hierarchical graph construction

[0009] Reference 3 "Mallick T, Balaprakash P, Rask E, et al. Graph-partitioning-based diffusion convolutional recurrent neural network for large-scale traffic forecasting [J]. Transportation Research Record. 2020, 2674 (9): 473-488." proposed a graph partitioning-based extended convolutional recurrent neural network overlapping node method, using sensor locations geographically close to a given partition as nodes, and using a graph decomposition method to decompose a large highway network into smaller networks and train them independently. This method uses a graph decomposition method to achieve large-scale traffic forecasting tasks, while this patent uses a hierarchical idea and a graph aggregation method to convert a large-scale site graph into multiple subgraphs.

[0010] Reference 4 "Guo K, Hu Y, Sun Y, et al. Hierarchical Graph Convolution Network for Traffic Forecasting [C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(1): 151-159." proposed a hierarchical graph convolution network method, which mined the interactive relationship between the micro and macro layers of the graph convolution network and integrated different feature scales of road sections and regions to improve traffic forecasting performance. This method solves the problem that the graph convolution network ignores the relationship between the micro layer of the road network and the macro layer of the regional network in the traffic system, and cannot process large-scale physical graphs by constructing a hierarchical convolution network. Summary of the Invention

[0011] The purpose of the present invention is to provide a method for constructing the spatial relationship of sites based on bus card swiping data. First, a physical map is constructed based on the bus card swiping data to reflect the spatial dependency of sites, and then a hierarchical map is further constructed based on the physical map.

[0012] The technical solutions for achieving the purpose of the present invention are as follows:

[0013] A method for constructing the spatial relationship of bus stops based on bus card swiping data, comprising constructing a physical map of bus routes;

[0014] Step 1: Count bus route trajectories based on bus card swiping data: First, extract the mapping relationship of bus route trajectories’ attributes, then sort the swiping data of the trajectories by swiping time, use the stations as feature attributes, construct the stations that the trajectories pass through, and calculate the time interval between each station in each trajectories;

[0015] Step 2: Construct bus route trajectory subsets: divide the trajectory into subsets based on the route number and the boarding direction as characteristic attributes, and output the entity-attribute mapping relationship of the bus trajectory subsets;

[0016] Step 3: Construct a bus route relationship graph: Select a subset of running trajectories, use the number of stations along the running trajectory as a feature, sort the running trajectories in descending order, use the first running trajectory after sorting as the base station for the selected route, and use interpolation to complete the stations; output the bus route relationship graph, including route number, boarding direction, stations along the way, and the dependency relationship between the stations;

[0017] Step 4: Construct a physical map of bus routes: Based on the bus route relationship graph, obtain the physical map of bus routes in the form of an adjacency matrix.

[0018] Further technical solutions also include building a hierarchical graph:

[0019] Step 5: Sum the values ​​of each row in the physical map matrix to obtain the degree of the site; sort the degrees of the sites in descending order to form a dictionary dataset, where the site represents the key and the degree of the site represents the key value;

[0020] Step 6: traverse the dictionary data set and set the rows and columns adjacent to the site in the physical map matrix to 0;

[0021] Step 7: traverse the dictionary data set, and replace all the stations adjacent to the station in the bus route relationship graph with the station with the key of the dictionary as the center, so as to obtain the dependency relationship of the stations in the aggregated bus route relationship graph;

[0022] Step 8: For the case where the key value in the dictionary data set is 0, the adjacent stations are merged according to the position of the key in the bus route relationship diagram; and the aggregated hierarchical diagram is output.

[0023] The beneficial effect of this invention lies in fully mining the real physical map of bus routes from the raw data of bus card swipes, reflecting the real spatial dependencies of bus stops in the data. Furthermore, through physical graph aggregation, the bus stop structure is converted from fine-grained to coarse-grained, and the large-scale physical graph is converted into a hierarchical graph. This can reduce the resource consumption of downstream tasks (such as traffic flow prediction) and improve the computational efficiency of downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Insert a schematic diagram for the site.

[0025] Figure 2 This is an example of a physical diagram.

[0026] Figure 3 This is a schematic diagram of physical graph aggregation. DETAILED DESCRIPTION

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0028] The method for constructing the spatial relationship of stations based on the bus card swiping dataset includes:

[0029] Step 1: Preprocessing of raw card swipe data

[0030] This includes: segmenting the data into daily units; using interpolation to fill in and delete missing values; and using a rule-based approach to extract valid data in the presence of data noise.

[0031] Step 2: Build the physical diagram

[0032] 2.1 Statistics of bus operation trajectories are collected. Five attributes, including bus line number, boarding direction, operation date, vehicle operation batch (daily operation schedule) and bus registration number, are used as features. All bus operation trajectories are counted to obtain the operation trajectory set.

[0033] 2.2 Construct the route stations of the running trajectory and extract the route stations of the running trajectory based on the attributes of the card swiping data such as boarding time, boarding station, disembarkation time and disembarkation station.

[0034] 2.3 Calculate the time interval between adjacent sites. Subtract the recording time of the last card swipe data at the previous site from the recording time of the first card swipe data at the next site as the time interval between the two sites. Repeat this step until the calculation of the time intervals of all adjacent sites in the running trajectory is completed.

[0035] 2.4 Based on the line number and riding direction, the running trajectory is divided to obtain the running trajectory subset.

[0036] 2.5 Select a subset of running trajectories and sort them in descending order based on the total number of stations they pass through. Take the first one in the sorted order (i.e., the one with the most stations) as the baseline for that subset. Compare the stations passed by the second-ranked trajectory with the baseline, and insert any stations not on the baseline into the baseline.

[0037] 2.6 Construct an N*N zero matrix, where N represents the total number of all stations. Map the specific numbers of the stations into numbers from 1 to N. Traverse all bus routes, set adjacent stations to 1 in the matrix, and output the physical map.

[0038] Step 3: Build a hierarchical graph

[0039] 3.1 Sum the rows of the physical graph to obtain the degree of the site, and perform descending processing on the degree of the site to form a dictionary data set, where the site represents the key and the degree of the site represents the key value.

[0040] 3.2 Traverse the dictionary dataset and set the rows and columns adjacent to the site in the physical map to 0. At the same time, replace all sites adjacent to the site in the route map with the site, centering on the dictionary key, to obtain the dependency relationship of the sites in the aggregated route map.

[0041] 3.3 Handling Isolated Values: After step 3.2, some values ​​will be aggregated around a central point and exist as a single station. These are called isolated values. Query the location of the isolated value on a specific bus route, find the nearest central point, assign the isolated value to the nearest central point, and output the aggregated hierarchical graph.

[0042] The specific embodiments are as follows:

[0043] A. Preprocess the bus card swiping dataset to obtain a valid bus card swiping dataset.

[0044] This data set contains approximately 227.6 million bus card swipe records from January 1, 2017, to February 28, 2017, in Singapore. Each record contains 13 features: card number, swipe type, ride type, bus route number, ride direction, vehicle batch number, bus registration number, boarding station number, alighting station number, boarding date, boarding time, alighting date, and alighting time. The preprocessing steps are as follows:

[0045] A1. Data segmentation. We use a time-based segmentation method to segment the data. Since the data is a single entity, processing it might require too much memory to load, so we need to segment the data. Using boarding time as a feature and daily granularity, we segment the data into 57 CSV files, storing each file for a single day.

[0046] A2. Missing Value Processing. Interpolation is used to fill in and delete missing values. First, the number of missing labels for each data item is counted. If the number of missing labels is too large, the data item is deleted. If the number of missing data items is small, the missing values ​​are filled in based on the time proximity principle, that is, the corresponding values ​​of the card swipe data at adjacent times are used to fill in the actual values.

[0047] A3. Extract valid data. Set the valid start time and valid end time, and extract data for the valid time period based on the boarding time. Only retain card swipe data with a boarding time greater than the start time and less than the end time, and delete the rest.

[0048] Through step A, we can obtain a bus card swiping dataset with high integrity and low redundancy. Its data structure is shown in Table 1-1. Each data in this dataset represents a card swiping instance.

[0049] Table 1-1 Description of the original card swiping data structure

[0050] Serial number Field Name illustrate Serial number Field Name illustrate 1 Card_Number Card number 8 Boarding_stop_stn Boarding station number 2 Card_Type Card swipe type 9 Alighting_stop_stn Get-off station number 3 Travel_Mode Ride Type 10 Ride_start_date Boarding date 4 Bus_Service_Number Bus route number 11 Ride_start_time Boarding time 5 Direction Directions 12 Ride_end_date Drop-off date 6 Bus_Trip_Num Vehicle operation batch number 13 Ride_end_time Get-off time 7 Bus_Reg_Num Bus registration number

[0051] B. Statistics of bus route operation trajectories: First, extract the mapping relationship of the attributes of the bus route operation trajectory, then sort the card swiping data of the operation trajectory by the card swiping time, use the station as the characteristic attribute, construct the stations that the operation trajectory passes through in sequence, and calculate the time interval between each station passed by the operation trajectory.

[0052] B1, bus line running trajectory TR = {tr1,tr2,tr3,…,tr i}, i represents the total number of bus trajectories, which include five attributes: bus route number L, boarding direction D, travel date RD, vehicle batch T, and bus registration number R. The mapping relationship between the five attributes of bus route trajectories is shown in Table 1-2. For example, as shown in Table 1-3, trajectory tr1 is the trajectory of bus route 3, boarding direction 0, travel date January 1, 2017, batch 1, and vehicle registration number 1013. The card swipe record for trajectory tr1 is shown in Table 1-3.

[0053] Table 1-2 Attribute mapping relationship of bus route operation trajectory

[0054] Running track TR Line No. L Travel direction D Operation Date RD Run batch T Vehicle registration number R <![CDATA[tr1,…,tr i ]]> <![CDATA[l1,…,l j ]]> <![CDATA[d0,d1]]> <![CDATA[rd1,…,rd k ]]> <![CDATA[t1,…,t n ]]> <![CDATA[R1,…,R m ]]>

[0055] Table 1-3 Card swiping record table of running track tr1

[0056]

[0057]

[0058] Table 1-4 Card swiping stations and card swiping time C′ after sorting of running trajectory tr1 tr1

[0059] Card swiping site Card swiping time Card swiping site Card swiping time Card swiping site Card swiping time Card swiping site Card swiping time 65009 5:50:46 65179 6:03:15 65199 6:08:58 76179 6:51:24 65009 5:50:48 65169 6:04:42 65199 6:09:00 76149 6:55:22 65271 5:56:43 65079 6:07:09 77339 6:15:45 76149 6:56:05 65271 5:56:46 65079 6:07:09 77339 6:15:49 76051 6:58:45 65239 6:00:43 65199 6:08:40 77339 6:16:09 75009 6:59:58 65239 6:00:43 65199 6:08:40 77339 6:16:11 75009 7:00:03 65179 6:03:13 65199 6:08:43 77289 6:17:51 65179 6:03:13 65199 6:08:46 77329 6:27:37

[0060] Table 1-5 Sites S passed by running trajectory tr1 tr1

[0061] Site 1 Site 2 Site 3 Site 4 Site 5 Site 6 Site 7 65009 65271 65239 65179 65169 65079 65199 Site 8 Site 9 Site 10 Site 11 Site 12 Site 13 Site 14 77339 77289 77329 76179 76149 76051 75009

[0062] B2. Construct bus route operation trajectory tr i Based on the idea that there are already existing stations with card swiping data, and the actual situation that Singapore buses require swiping cards to get on and off, we first merge the boarding station, boarding time, and disembarkation station, disembarkation time of the passenger card swiping data to obtain the set C of card swiping time and card swiping stations. tri ={(st1,s1),(st2,s1),(st3,s1),…,(st n ,s m )}, where st represents the card swiping time, s represents the card swiping station, n represents the total number of card swiping times, and m represents the running trajectory tr i As shown in Table 1-3: The total number of card swipe times n for track tr1 is 30. n >st n-1 The method for the set C tri Sort and get the new set C′ of card swiping time and card swiping site tri , output bus line running trajectory tri The set of pathway sites S tri ={s1,s2,s3,…,s m For example: the set C′ of card swiping sites and card swiping times after tr1 is sorted tr1 As shown in Table 1-4, S tr1 As shown in Table 1-5, the running trajectory tr1 passes through 14 stations in total, that is, m is 14.

[0063] B3. Calculate S tri The time difference between adjacent sites is calculated as shown in formula (1):

[0064] td m，m-1 =head(s m )-end(s m-1 ), m>1 (1)

[0065] Among them, m represents the running trajectory tr i The total number of stations passed, td represents the time difference between stations, head represents the time of the first card swipe record at a station, and end represents the time of the last card swipe record at a station. For example, in Table 1-4, the time difference between stations 65271 and 65009 is 5 minutes and 55 seconds, which is 5:56:43 minus 5:50:48.

[0066] B4. Construct a subset of bus running trajectories. Divide the running trajectories by the route number and the boarding direction, and obtain the running trajectory subset SUBTR = {subtr1, subtr2, subtr3, ..., subtr k}, since the riding direction is 0 and 1, k represents the total number of bus line numbers L multiplied by 2. k Including multiple running tracks tr i , and the site dependency S of the trajectory tri The time interval between the stations is td. The running trajectory subset subtr k The entity-attribute relationship table is shown in Table 1-6.

[0067] Through step B, output the bus trajectory subset subtr extracted from the bus passenger flow dataset k and its properties.

[0068] Table 1-6 Bus trajectory subset entity-attribute mapping table

[0069]

[0070] C. Based on the bus trajectory subset subtr kThe route map is constructed based on the data and its attributes. A subset of running trajectories is selected, characterized by the number of stations along the route. The trajectories are sorted in descending order, and one trajectory is used as the base station for selecting the route. Interpolation is used to complete the stations. The bus route relationship diagram is output, including the route number, boarding direction, stations along the route, and the dependency relationship between the stations.

[0071] C1. Select the subset of bus trajectories subtr k , with the set of sites along the trajectory S tri The total number of sites is characterized by S tri Total number of sites tri-1 The total number of sites is the rule, for the running trajectory tr i Sort and get a new subset of bus trajectory n represents a subset of bus trajectories The running track of tr i The total number of .

[0072] C2, take the first one in the arrangement as the line reference, that is, take tr1 as the line reference, and use formula (2) to complete the line station, that is, put the i The sites that are not in tr1 are completed into tr1.

[0073] tr1←tr1+(tr i -tr1) (2)

[0074] C3. The "+" operation in formula 2 is an insertion and completion operation. The location of the site insertion needs to be determined based on the time difference between adjacent sites. The specific algorithm is shown in Algorithm 1-1. Figure 1 As shown. First get tr i The difference between tr1 and D and F sites, and store their site indexes 2 and 4 in array d[0,1]. Then insert sites D and F into tr1 respectively, first get the difference between site D in tr1 and tr i In the same recent site A, and site A in tr i The index in tr1 is stored in index0, and the index in tr1 is stored in index1; calculate and compare the time interval tr between sites i td (A，D) and tr1td (A，B) , if tr i td (A，D) Less than tr1td (A，B) , then insert D between AB, by Figure 1 It can be seen that tr i td (A，D) Greater than tr1td (A，B) , so the loop compares tr i ​td (A，D) and tr1td (A，C) , tr1td (A，E) , found tr i td (A，D) Less than tr1td (A，E) , so insert site D before site E in tr1. Similarly, insert site F into tr1.

[0075] Through step C, the station relationship graph of all routes in the card swiping dataset is output, which includes the route number L, the riding direction D, the passing stations S, and the dependency relationship of the stations.

[0076] Algorithm 1-1 Site Insertion Algorithm

[0077]

[0078]

[0079] D. Build a physical adjacency matrix based on the site relationship graph of all lines N represents the total number of sites in the card swiping dataset, and the calculation formula for P(i,j) is:

[0080]

[0081] Through step D, the construction of the adjacency matrix of the physical graph is completed. The schematic diagram of the adjacency matrix of the physical graph is as follows Figure 2 shown.

[0082] E. Hierarchical graph construction. Aggregating the route physical graph is divided into multiple rounds of center point aggregation operations and isolated point merging operations. The former step is to first calculate the sum of the values ​​in each row of the physical graph matrix, obtain the data of the center point and the sum of adjacent points in descending order, form dictionary data with rows as the center point and the related adjacent points, and set the rows or columns in the matrix with the same values ​​to 0. This is performed in descending order until there are no adjacent stations. The latter step is to extract the isolated points and the aggregated center points of the aggregated matrix, replace the station data in the bus route with the center points, and then merge the remaining isolated points into the adjacent center points.

[0083] E1. Sum the values ​​of each row of the physical map to get the degree of the site (Num1, Num2, Num3, ..., Num N ), N represents the total number of sites in the physical graph, and the degree of the site is expressed as Num i -1>Num i The rules are processed in descending order to form a dictionary data set DN = {(s1, Num1), (s2, Num2), (s3, Num3), ..., (s N , Num N)}, where the site represents the key and the degree of the site represents the key value. Figure 3 As shown: the degree of site 4 is 5, the degree of site 7 is 4, and the degree of site 9 is 3.

[0084] E2. Traverse the dictionary dataset DN and set the rows and columns adjacent to the site in the physical map to 0. Figure 3 As shown: N 4j Set to 0, N j4 Set to 0.

[0085] E3. Traverse the dictionary dataset DN, and replace all the sites adjacent to the site in the route map with the site with the key of the dictionary as the center, and obtain the dependency relationship of the sites in the aggregated route map. Figure 3 As shown in (b), sites 1, 2, 3, 5, and 7 are all replaced by site 4; site 12 is replaced by site 6; sites 10 and 11 are replaced by site 9, and site 13 is replaced by site 8.

[0086] E4. Handle isolated points. If the key value in the dictionary data set is 0, merge it into the adjacent station according to the position of the key in the route map. Figure 3 In (a), station 14 is an isolated point. According to the position of the line, station 14 is merged into station 8, and station 14 is replaced by station 8. The final output is as follows: Figure 3 The hierarchical diagram shown in (c).

[0087] After step E, the 4850 sites in the card swiping data are aggregated into 285 hierarchical graphs.

Claims

1. A method for constructing the spatial relationship of bus stops based on bus card swiping data, characterized in that: This includes constructing a physical map of bus routes; Step 1: Count bus route trajectories based on bus card swiping data: First, extract the mapping relationship of bus route trajectories’ attributes, then sort the swiping data of the trajectories by swiping time, use the stations as feature attributes, construct the stations that the trajectories pass through, and calculate the time interval between each station in each trajectories; Step 2: Construct bus route trajectory subsets: divide the trajectory into subsets based on the route number and the boarding direction as characteristic attributes, and output the entity-attribute mapping relationship of the bus trajectory subsets; Step 3: Construct a bus route relationship diagram: select a subset of running trajectories, use the number of stations along the running trajectories as a feature, sort the running trajectories in descending order, use the first running trajectory after sorting as the base station for the selected route, and use interpolation to complete the stations; Output bus route relationship diagram, including route number, riding direction, passing stations, and station dependencies; Step 4: Construct a physical map of bus routes: Based on the bus route relationship graph, obtain the physical map of bus routes in the form of an adjacency matrix; Also includes building a hierarchical graph: Step 5: Sum the values ​​of each row in the physical map matrix to obtain the degree of the site; sort the degrees of the sites in descending order to form a dictionary dataset, where the site represents the key and the degree of the site represents the key value; Step 6: traverse the dictionary data set and set the rows and columns adjacent to the site in the physical map matrix to 0; Step 7: traverse the dictionary data set, and replace all the stations adjacent to the station in the bus route relationship graph with the station with the key of the dictionary as the center, so as to obtain the dependency relationship of the stations in the aggregated bus route relationship graph; Step 8: For the case where the key value in the dictionary data set is 0, the adjacent station is merged according to the position of the key in the bus route relationship diagram; Output the aggregated hierarchical graph.

Images