CN108427965B - Hot spot area mining method based on road network clustering - Google Patents

Abstract

The present invention claims to protect a travel hotspot area mining method based on road network trajectory clustering. In this method, the taxi trajectories are mapped into the road network, and a clustering method combining interest points and trajectories collected from actual roads is adopted. Combined with the density peak clustering algorithm, an OPAM algorithm based on the density peak optimization of the initial center is proposed, namely DP‑OPAM. The algorithm uses the local density of data points and the shortest distance from these points to higher-density points, and uses a decision diagram to select the category of data points with higher density and the closest distance as the initial cluster center. According to the initial clustering center, the OPAM clustering algorithm with reverse learning is used to obtain the clustering result. Comparing the new algorithm with the original OPAM algorithm, the new algorithm can not only automatically determine the cluster centers, but also improve the accuracy and clustering time, and realize the analysis of user travel hotspots.

Description

A method for mining hot spots based on road network clustering

technical field

The invention belongs to a data mining method, in particular to a taxi trajectory clustering method based on a road network.

Background technique

As a hot spot in the development of transportation in the world today, intelligent transportation not only supports transportation management, but also pays more attention to meeting the needs of people's travel and public transportation. In recent years, the construction of intelligent transportation systems has developed rapidly, and many advanced technologies have been widely used in intelligent transportation systems. The wide application of GPS equipment makes the extraction of trajectory more convenient. These GPS devices can collect a large amount of mobile location sequence information and vehicle status information, and these data contain rich traffic information and user behavior information. By analyzing and mining trajectory data, we can understand traffic conditions, plan trips reasonably, discover crowd behavior characteristics, and assist in improving traffic conditions.

Taxi trajectories can cover urban road network traffic in an all-round way, reflecting not only the real-time traffic density and circulation, but also the travel patterns and regional characteristics of the crowd. Therefore, by analyzing the massive data of taxi trajectories, we can find the deep-level information hidden in the data. With the help of data mining technology, we can analyze the overall characteristics of the data and predict the development of traffic situation, and carry out traffic detection and road traffic detection for the traffic management department. Controls provide support, etc. play a major role.

As a common data mining technique, cluster analysis can be used as a tool to obtain the distribution of data, which is convenient to observe the characteristics of each cluster of data, and focus on specific clusters for further analysis. In addition, it can be used as a preprocessing step for other algorithms such as classification and qualitative induction algorithms. Trajectory clustering of moving objects, by finding similar motion trajectories, extracting motion features, etc., finds the motion laws and behavior patterns of moving objects. The trajectories of taxis are composed of discontinuous sequences of points. When measuring the similarity of trajectories, the traditional cluster analysis of trajectories mostly considers the straight-line distance between time points and points, while ignoring the actual distance reachability.

There are two main methods for cluster analysis of vehicle trajectories: one is to classify and compare the entire trajectory as an object; sort. The advantage of the former is that the method is simple, and it is convenient to intuitively evaluate the similarity between trajectories, but at the same time, this method cannot distinguish the local features of the trajectories well, and the clustering effect is often not ideal. The latter method can improve the problems caused by the former in terms of local characteristics of the trajectory, and the clustering effect is better for trajectories of different shapes. However, the disadvantage is that the method of trajectory segmentation has a great influence on the clustering results, and different segmentation methods may cause great differences in the results.

SUMMARY OF THE INVENTION

The present invention aims to solve the above problems of the prior art. A method for mining hotspot areas based on road network clustering is proposed, which can significantly improve the clustering effect and realize user travel area mining. The technical scheme of the present invention is as follows:

A method for mining hot spots based on road network clustering, comprising the following steps:

Step 1: Collect taxi trajectory data sets, perform data preprocessing including data standardization and normalization, retain valid fields, delete redundant data, and obtain preprocessed vehicle boarding and passenger trajectory points;

Step 2: Determine the latitude and longitude range of the city, and extract the city's points of interest including shopping malls and schools on the open source website;

Step 3: Obtain the road network information of the city and map the trajectory points to the road network;

Step 4: Select 80% of the pre-processed vehicle pick-up and drop-off trajectory points in step 1 as the training set, and use the improved OPAM algorithm based on the density peak to optimize the initial center to cluster the areas representing the pick-up and drop-off hot spots. The improvement points are mainly in : Use the density peak to select the initial clustering center, the selection of the initial point is more accurate and convenient, the remaining 20% is used as the test set, and the clustering effect of the model is built by 80% of the track points of the pick-up and drop-off passengers as the training set;

Step 5: Input the points of interest with road network information collected in step 2 into the model of step 4, and cluster to obtain the hotspot activity areas of residents with road network characteristics, and compare the clustering results with the collected points of interest to judge residents. Travel hotspots.

Further, the step 1 is specifically: first collect the taxi trajectory data set of a certain month in the city, select the trajectory data with a relatively concentrated amount of data for one week in the city, carry out data preprocessing, retain the latitude and longitude data of the alighting track points, and the time of getting on and off the bus. Data and other valid fields, delete redundant data.

Further, the step 2 determines the range of longitude and latitude of the city, and extracts the points of interest of the city including shopping malls and schools on the open source website, specifically:

First, enter the latitude and longitude range of the target city on the open source website openstreetmap, and download the map of the entire city. In the exported OSM map data, the way represents the user's movement trajectory, and the node represents the path. Select the node label as residence, school, and shop to represent points of interest.

Further, the step 3 obtains the road network information of the city, and maps the trajectory points to the road network, specifically:

Using the TREEG network service project to obtain the electronic map data, extract the road network information of the city, and after extracting the city road network data, the GPS movement trajectory obtained above is projected onto the obtained road network map through the ST-Matching model. The trajectory points of (j-1+1) consecutive times p _i ,...,p _j on each road segment e.

Further, the step 4 is specifically as follows: first, 80% of the processed vehicle disembarkation and passenger trajectory points are selected as the training set, and an improved clustering algorithm (OPAM) based on reverse learning is used to cluster the representative getting on and off the vehicle. In the hotspot area, the improved OPAM algorithm is divided into three stages: the first stage initializes, constructs a decision diagram, selects the density peak point in the upper right corner area far away from most samples as the initial cluster center, and the number of density peak points is the cluster Number k; the second stage constructs the initial cluster center, calculates the minimum distance between each point in the data set and each cluster center, assigns the remaining sample points to the nearest initial cluster center, forms the initial division, and calculates the sum of squares of clustering errors ; The third stage is reverse learning and substituting the clustering algorithm around the center point (PAM), the k clusters obtained by the typical PAM clustering algorithm and the k reverse clusters obtained after reverse learning are arranged and combined to obtain k × k Find the cluster combination with the largest silhouette coefficient.

Further, the steps of the PAM algorithm are as follows:

(1) arbitrarily select k elements from a given data set D, and mark the selected k elements as initial representative objects or seeds o _j ;

(2) According to the Euclidean distance calculation method, calculate the distance between any non-representative object o _i in the data set D and k representative objects, and assign o _i to the cluster represented by the representative object with the closest distance;

(3) arbitrarily select a non-representative object o _random ;

(4) Calculate the total cost S:

S=dist(p,o _random )-dist(p,o _j )

(5) If the total cost S < 0, it indicates that the non-representative object o _random is a better solution, and the element o _random can replace the element o _j to form a new set of k representative objects, and continue to return to step (2) to make a new One round of object allocation;

(6) If the total cost S>0, it indicates that the representative object o _j is a better solution, go to step (3), and re-select non-representative objects to compare the total cost until the sending cost S does not change, that is, the total cost is obtained. The k clusters with the least cost.

The advantages and beneficial effects of the present invention are as follows:

The invention combines the analysis of residents' travel hotspots with the road network, and adopts the method of combining the interest area in the specific road network with the original cluster cluster. The travel hotspot solves the deficiencies of time and space in European space. This method uses the method of optimizing the initial center based on the peak density to construct a decision tree to determine the initial center, which reduces the amount of calculation and makes the clustering accuracy higher. And through the combination of special interest points and trajectories clustering, the problems of data sparsity and huge amount of calculation are solved, and user trajectory behavior analysis is realized.

Description of drawings

Fig. 1 is that the present invention provides a preferred embodiment PAM clustering algorithm flow chart;

Figure 2. Flow chart of OPAM clustering algorithm.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and in detail below with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only some of the embodiments of the invention.

The technical scheme that the present invention solves the above-mentioned technical problems is:

As shown in Figure 2, the invention adopts the OPAM clustering algorithm based on the density peak optimization initial center and the road network to combine the specific steps of the hot spot mining method as follows:

Step 1: Collect the taxi trajectory data set of a city in a certain month, and select the trajectory data for a week with a relatively concentrated amount of data in the city. Carry out data preprocessing, retain valid fields such as the latitude and longitude data of the track point of getting on and off, and the time data of getting on and off, and delete redundant data.

Step 2: Enter the latitude and longitude range of the target city on the open source website openstreetmap, and download the map of the entire city. In the exported OSM map data, the way represents the user's movement trajectory, and the node represents the path. Since the way object in the OSM source data records the user's movement trajectory, the way does not only represent road information, but also building information, such as the external outline of a building. Therefore, we also need to filter the way object accordingly to remove the unnecessary way information. On the other hand, there are many node sampling points in the downloaded map source data. In the actual route planning process, you only need to know the intersection location between the way and the way. There are many tags representing different attributes in way and node. In way, only the attribute value of highway is reserved. The attribute value is residential, service, living_street, unclassified, trunk, trunk_link, secondary, secondary_link, primary, tertiary, tertiary_link, and different ways are deleted in node. But the node is the same node. Select the node label as residence, school, and shop to represent points of interest.

Step 3:

The GPS track of the driver collected above contains the instantaneous position coordinates, which hides the information on which road section the driver is on at each moment. The electronic map data is obtained by using the TREEG network service project, and the road network information of the city is extracted. After extracting the urban road network data, the above-obtained GPS movement trajectory is projected onto the obtained road network map through the ST-Matching model, and the driver passes (j-1+1) consecutive moments p _i on each road section e, ..., the trajectory points of p _j .

Step 4

Select 80% of the processed vehicle disembarkation and passenger trajectory points as the training set, and use the improved OPAM algorithm to cluster the areas representing the hotspots of getting on and off. The DP-OPAM algorithm is divided into three stages: initialization, construction of initial cluster centers, reverse learning and substitution into the PAM algorithm.

其中，

d_c为截断距离，对于大量数据而言，局部密度实质是数据点之间的相对密度，所以d_c具有鲁棒性。定义δ_i是数据点i到任何比其密度大的点的距离的最小值：δ_i＝min_j:ρj＞ρi(d_ij)对于局部密度最大的点，需要特殊处理，一般改点的值为：δ_i＝max_j(d_ij)Assuming that the data point of the sample is i, the local density is ρ _i , the calculation method of the local density ρ _i of the data point i is:

in,

_dc is the cutoff distance. For a large amount of data, the local density is essentially the relative density between data points, so _dc is robust. Definition δ _i is the minimum value of the distance from data point i to any point with greater density: δ _i =min _j:ρj>ρi (d _ij ) For the point with the largest local density, special treatment is required, and the value of the point is generally changed. is: δ _i =max _j (d _ij )

first stage initialization

(1) Initialize the distance matrix D={d _ij }i,j=1,...,n between each data point, and determine the cutoff distance.

计算样本的高密度距离δ_i。(2) Calculate the local density according to the formula S=dist(p,o _random )-dist(p,o _j ), and use the formula

Calculate the high density distance δ _i of the sample.

(3) Construct a decision diagram with ρ as the horizontal axis and δ as the vertical axis. Select the data points with high local density ρ and high density distance δ, and the density peak point in the upper right corner of most samples is obviously far away as the initial cluster. Class center, the number of density peak points is the number of clusters k.

The second stage constructs initial cluster centers

(1) Calculate the minimum distance between each point in the data set and each cluster center, assign the remaining sample points to the nearest initial cluster center, form an initial division, and calculate the sum of squares of clustering errors.

The third stage of reverse learning and substituting into the PAM algorithm

(1) Perform reverse learning on the k original clusters obtained above to obtain k corresponding reverse clusters.

(2) The k clusters obtained by the typical PAM clustering algorithm and the k reverse clusters obtained after reverse learning are arranged and combined to obtain k × k cluster combinations.

(3) Calculate the intra-cluster spacing a(o), inter-cluster spacing b(o) and silhouette coefficient s(o) of each cluster combination, compare s ₁ , s ₂ ,...,s _k×k , Find the cluster combination with the largest silhouette coefficient.

Among them, the steps of the PAM algorithm are as follows:

(1) arbitrarily select k elements from a given data set D, and mark the selected k elements as initial representative objects or seeds o _j ;

(2) According to the Euclidean distance calculation method, calculate the distance between any non-representative object o _i in the data set D and k representative objects, and assign o _i to the cluster represented by the representative object with the closest distance;

(3) arbitrarily select a non-representative object o _random ;

(4) Calculate the total cost S:

S=dist(p,o _random )-dist(p,o _j )

(5) If the total cost S < 0, it indicates that the non-representative object o _random is a better solution, and the element o _random can replace the element o _j to form a new set of k representative objects, and continue to return to step (2) to make a new One round of object allocation;

(6) If the total cost S>0, it indicates that the representative object o _j is a better solution, go to step (3), and re-select non-representative objects to compare the total cost until the sending cost S does not change, that is, the total cost is obtained. The k clusters with the least cost.

Step 5

Input the points of interest with road network information collected in step 2 into the model of step 4, and measure the similarity between the k clusters obtained from the clustering results and the representative points of interest collected, and analyze which areas the residents travel to belong to. Points of interest to analyze residential hotspots. Clustering obtains residents' hotspot activity areas with road network characteristics. Compare the clustering results with the collected points of interest to determine the hotspot areas for residents to travel.

应用Hausdorff距离计算两条轨迹中每个点到另外一条轨迹上所有点的最小值，然后从各自的最小值集合中找出最大的。当小于相似度阈值时认为和兴趣点空间上相似，将被保存到候选集合中。把候选集合中距离大于某个阈值的轨迹删除，得到离轨迹最近的路网兴趣点，即居民出行热点区域。For the trajectories _{tra and tr b to} be measured, the Hausdorff distance is used to measure the trajectory similarity. H(tr _a ,tr _b )=max{h(t _a ,tr _b ),h(tr _b ,tr _a )}, where

The Hausdorff distance is used to calculate the minimum value from each point in the two trajectories to all points on the other trajectory, and then find the maximum from the respective sets of minimum values. When it is smaller than the similarity threshold, it is considered to be spatially similar to the interest point and will be saved into the candidate set. The trajectories with distances greater than a certain threshold in the candidate set are deleted, and the road network interest points closest to the trajectories are obtained, that is, the residents' travel hotspots.

The above embodiments should be understood as only for illustrating the present invention and not for limiting the protection scope of the present invention. After reading the contents of the description of the present invention, the skilled person can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims (4)

1. A hot spot region mining method based on road network clustering is characterized by comprising the following steps:

step 1: collecting a taxi track data set, carrying out data preprocessing including data standardization and normalization, reserving effective fields, deleting redundant data, and obtaining preprocessed upper and lower taxi track points;

step 2: determining the latitude and longitude range of a city, and extracting interest points of the city including a shopping mall and a school on an open source website;

and 3, step 3: acquiring city road network information, and mapping track points into a road network;

and 4, step 4: selecting 80% of the upper and lower passenger track points of the vehicle preprocessed in the step 1 as a training set, and clustering out an area representing an upper and lower vehicle hot point by adopting an improved reverse learning-based center point partitioning clustering algorithm, wherein the improvement point is as follows: selecting an initial clustering center by using the density peak; the other 20 percent is used as a test set, and the clustering effect of a model built by taking 80 percent of upper and lower guest track points as a training set is tested;

and 5: inputting the interest points with road network information acquired in the step 3 into the model in the step 4, clustering to obtain a resident hot spot activity area with road network characteristics, comparing a clustering result with the acquired interest points, and judging a hot spot area of resident travel;

the step 4 specifically comprises the following steps: firstly, selecting 80% of processed upper and lower passenger track points of a vehicle as a training set, clustering an area representing an upper and lower hot spot by adopting an improved reverse learning-based central point partition clustering algorithm, and dividing the improved reverse learning-based central point partition clustering algorithm into three stages: initializing a first stage, constructing a decision diagram, and selecting density peak points in an upper right corner area far away from most samples as initial clustering centers, wherein the number of the density peak points is a cluster number k; constructing initial clustering centers, calculating the minimum distance between each point in the data set and each clustering center, distributing the rest sample points to the nearest initial clustering center to form initial division, and calculating the clustering error square sum; the third stage reversely learns and substitutes the algorithm for dividing and clustering around the central point, k clusters obtained by dividing and clustering around the central point and k reverse clusters obtained after reversely learning are arranged and combined to obtain k multiplied by k cluster combinations, and the cluster combination with the maximum outline coefficient is searched;

the step of dividing the clustering algorithm around the center point is as follows:

(1) arbitrarily selecting k elements from a given data set D, and marking the selected k elements as initial representative objects or seeds o _j ；

(2) Calculating any non-representative object o in the data set D according to the Euclidean distance calculation mode _i And k representative objects, and o _i Allocating to the cluster represented by the representative object closest to the cluster;

(3) arbitrarily selecting a non-representative object o _random ；

(4) Calculating the total cost S:

S＝dist(p,o _random )-dist(p,o _j )，

(5) if the total cost S < 0, indicating a non-representative object o _random Is a better solution, element o _random In place of the element o _j Forming a new k representative object set, continuously returning to the step (2), and performing a new round of object allocation;

(6) if the total cost S > 0, it indicates that the representative object o _j If the total cost is the optimal solution, turning to the step (3), reselecting the non-representative object to compare the total cost until the cost S is not changed any more, and obtaining k class clusters with the minimum total cost;

for the track tr to be measured _a And tr _b Track similarity, H (tr), is measured using the Hausdorff distance _a ,tr _b )＝max{h(tr _a ,tr _b ),h(tr _b ,tr _a ) Therein of

Calculating the distance from each point of the two tracks to the other track by using Hausdorff distanceThe minimum value with points is found out, and then the maximum value is found out from the respective minimum value set; when the similarity is smaller than the similarity threshold value, the spatial similarity with the interest point is considered, and the similarity is stored in the candidate set; deleting the tracks with the distance greater than a certain threshold value in the candidate set to obtain the road network interest points closest to the tracks, namely the resident trip hot spot areas;

let the data point of the sample be i and the local density be ρ _i Local density ρ of data point i _i The calculation method is as follows:

wherein,

d _c for the truncation distance, δ is defined _i Is the minimum of the distance of a data point i to any point greater than its local density: delta _i ＝min _j:ρj＞ρi (d _ij ) For the point with the maximum local density, special treatment is needed, and the point is changed into the following value: delta _i ＝max _j (d _ij )；

First phase initialization

(1) Initializing and solving a distance matrix D ═ D between data points _ij I, j ═ 1.. times, n, and determining a truncation distance;

(2) according to the formula S ═ dist (p, o) _random )-dist(p,o _j ) Calculating local density by formula

Calculating a high density distance δ of a sample _i ；

(3) And constructing a decision diagram with rho as a horizontal axis and delta as a vertical axis, selecting data points with higher local density rho and high density distance delta, taking density peak points of the upper right corner area far away from most samples as initial clustering centers, and taking the number of the density peak points as a clustering number k.

2. The road network clustering-based hot spot region mining method according to claim 1, wherein the step 1 specifically comprises: firstly, a taxi track data set of a certain month in a city is collected, track data of a week in the city data set is selected, data preprocessing is carried out, longitude and latitude data of track points of an upper taxi and a lower taxi are reserved, effective fields of the data of the time of the upper taxi and the lower taxi are reserved, and redundant data are deleted.

3. The road network clustering-based hotspot region mining method of claim 1, wherein the step 2 determines a city latitude and longitude range, extracts interest points of the city including shopping malls and schools from an open source website, and specifically comprises the following steps:

firstly, inputting the longitude and latitude range of a target city on an open source website opentreetmap, downloading a map of the whole city, deriving ways in OSM map data to represent the moving track of a user, representing paths by nodes, and selecting node labels as residual, school and shop to represent interest points.

4. The road network clustering-based hotspot region mining method according to claim 3, wherein the step 3 acquires road network information of a city, and maps track points into a road network, specifically:

obtaining electronic map data by adopting TAREEG network service items, extracting road network information of the city, projecting the obtained moving track on the obtained road network map through an ST-Matching model after extracting the road network data of the city, and obtaining the track points p of a driver passing through j continuous moments on each road section e _i ,…,p _j 。

Images