CN114969263A - Construction method, construction device and application of urban traffic knowledge map - Google Patents

Abstract

The invention discloses a method for constructing an urban traffic knowledge graph, comprising the following steps: constructing an urban traffic ontology by using an ontology building tool to form a knowledge graph mode layer; acquiring urban traffic data, extracting entities, attributes and relationships between entities, and constructing Knowledge graph data layer; combine the knowledge graph mode layer with the knowledge graph data layer to generate an urban traffic knowledge graph and store it in the database; use the knowledge representation model to infer new knowledge in the urban traffic knowledge graph and add it to the urban traffic knowledge graph , the invention also discloses a construction device and application of an urban traffic knowledge map. The invention forms a traffic knowledge system by using a knowledge graph, integrates multi-source and heterogeneous traffic big data, and mines the potential relationship between traffic entities through a knowledge inference model based on representation learning, thereby realizing the multi-source travel data in the traffic field. Effective integration and organization realize the open sharing of data in the transportation field.

Description

A construction method, construction device and application of an urban traffic knowledge graph

technical field

The invention belongs to the technical field of intelligent transportation, and in particular relates to a construction method, construction device and application of an urban traffic knowledge graph.

Background technique

Urban traffic has strong coupling, and it is necessary to coordinate "people, vehicles, roads, and environment" to complete integrated management and control, and the massive traffic travel data is related in time and space, but the integration of multi-source data is insufficient. Therefore, knowledge graphs can be used to analyze traffic travel. Data fusion, the data organization form of knowledge graph is structured, it can describe the entities existing in the real world, the attributes of entities and the relationship between two entities, but the large-scale knowledge base currently constructed, although The scale of data is large, but there is still a problem of data sparseness.

Therefore, there is an urgent need for a better construction method of urban traffic knowledge graph to solve the problem of sparse data in the knowledge graph.

SUMMARY OF THE INVENTION

In order to solve the drawbacks of the above-mentioned prior art, the present invention discloses a method for constructing an urban traffic knowledge graph that can accurately infer new knowledge. The specific technical solutions are as follows:

A method for constructing an urban traffic knowledge graph, comprising the following steps:

Use ontology construction tools to build urban traffic ontology and form a knowledge graph model layer;

Obtain urban traffic data, extract entities, attributes and relationships between entities, and build a knowledge map data layer;

Combine the knowledge graph mode layer with the knowledge graph data layer to generate an urban traffic knowledge graph and store it in the database;

The knowledge representation model is used to infer the new knowledge in the urban traffic knowledge graph, and add it to the urban traffic knowledge graph.

Further, the method of constructing the urban traffic ontology is as follows:

Adopt a seven-step method to build an urban traffic ontology, and conduct quality assessment before creating an instance;

The quality assessment specifically includes:

Verify that the transitivity of the class hierarchy is established by drawing a tree structure diagram;

Check whether the application scope and expression method of each ontology are consistent when used everywhere, and whether there is redundancy between classes and ontology;

Check whether the attribute description information is complete, whether the attribute constraints are logical, and whether the attributes are shared;

Check the scalability of the ontology;

Check the integrity, uniqueness, and logical consistency of relationships between classes.

Further, the specific method of using the knowledge representation model to infer the new knowledge in the urban traffic knowledge graph is as follows:

Divide the triplet data in the urban traffic knowledge graph into training set, validation set and test set;

constructing negative samples of the triplet data, and filtering false negatives in the negative samples;

Set the knowledge representation model hyperparameters;

Using the training set and negative samples, the knowledge representation model is trained based on the mini-batch stochastic gradient descent method, and the learning rate is adaptively adjusted during the training process through the adadelta method;

Hyperparameter tuning of the trained knowledge representation model using the validation set and negative samples;

Use the test set and negative samples to evaluate the trained knowledge representation model;

Use the trained knowledge representation model to mine the implicit relationships and missing entities of the urban traffic knowledge graph, and add it to the urban traffic knowledge graph.

Further, the method for constructing the negative sample of the triplet data is:

For the triples with a certain relationship in the training set, validation set or test set, according to Bernoulli distribution, calculate the probability of selecting the head entity or the tail entity to complete the replacement operation, and replace the entity with higher probability;

According to the relationship type constraint, the relationship determines which entities to replace, as shown in the following formula:

Among them, Δ′ is the set of constructed negative triples, d _r is the ordered index of all entities that satisfy the domain constraints of relation type r; r _r is the ordered index of all entities that satisfy the range constraints of relation type r, h is the head entity of the triplet, h' is the head entity of the negative triplet, t is the tail entity of the triplet, t' is the tail entity of the negative triplet, and r is the relation.

Further, the specific method for filtering false negative examples in the negative samples is:

Importing the triplet data and the negative triplet data in the negative sample into a relational database, using the query function in the relational database to find out the duplicate data, and removing the duplicate data in the negative sample;

Wherein, the repeated data is data that exists both in triple data and in negative triple data.

further,

The knowledge representation model is a TransD model, and the TransD knowledge representation model is as follows:

Mapping matrix:

为头实体映射向量，

为尾实体映射向量，I^mxn为单位矩阵；Among them, M _rh is the head entity mapping matrix, M _rt is the tail entity mapping matrix, r _p is the relationship vector,

map vector for head entity,

is the tail entity mapping vector, and I ^mxn is the identity matrix;

Project entity vectors into relational space:

h _⊥ =M _rh h

t _⊥ =M _rt t

h _⊥ is the head entity vector after the head entity is mapped by M _rh , t _⊥ is the tail entity vector after the tail entity is mapped by M _rt ;

Score function:

Loss function:

where γ is a hyperparameter representing the maximum separation between correct triples and negative triples. [x] ₊ =max(0,x), Δ represents the set of correct triples, and Δ′ represents the set of constructed negative triples.

Further, the specific method for evaluating the trained knowledge representation model using the test set and negative samples is as follows:

For any triplet in the test set, calculate the triplet score and the negative triplet score constructed according to the triplet and entities in the knowledge graph according to the score function in the trained knowledge inference model, and calculate the score according to the score. Rank the triplet and the negative triplet in descending order of value;

One or more evaluation indicators among the average ranking, the average penultimate ranking, the first hit rate, the top three hit rate and the top ten hit rate are used to measure the effect of link prediction task completion.

further,

The urban traffic includes public traffic and road traffic, and a public traffic knowledge map and a road traffic knowledge map are respectively constructed for public traffic and road traffic;

The method for obtaining public transportation data is to obtain the subway line and station information of the target city through the web crawler technology, and obtain the subway card swiping data of the subway line within the target time;

The method of obtaining road traffic data is to obtain the target road network data from the map database, and use the map API to obtain the target point of interest information and traffic situation data on the target road.

The invention also discloses a device for constructing a knowledge map of urban traffic, comprising:

The ontology building module is used to use ontology building tools to build urban traffic ontology and form a knowledge graph model layer;

The data acquisition module is used to acquire urban traffic data, extract the relationship between traffic entities, attributes and entities, and build a knowledge graph data layer;

The storage module is used to combine the knowledge graph mode layer and the knowledge graph data layer to generate the urban traffic knowledge graph and store it in the database;

The reasoning module is used to use the knowledge representation model to infer new knowledge in the urban traffic knowledge graph and add it to the urban traffic knowledge graph.

The invention also discloses the application of any of the above-mentioned methods for constructing a knowledge map of urban traffic in the field of urban traffic.

By adopting the above-mentioned technical scheme, the beneficial effects of the present invention are:

The invention forms a traffic knowledge system by using a knowledge graph, integrates multi-source heterogeneous traffic big data, models the spatiotemporal relationship between traffic entities, and mines the potential relationship between traffic entities through a knowledge inference model based on representation learning , realizes the effective integration and organization of multi-source travel data in the field of transportation, forms a knowledge network in the field of transportation, and realizes the open sharing of data in the field of transportation.

Compared with the original seven-step method, the present invention adds a quality assessment step when constructing the ontology, and strictly controls the quality of the ontology added to the knowledge base through the quality assessment. The evaluation is carried out to realize the reusability of the standardized ontology, to ensure the accuracy and validity of the ontology construction, and to improve the accuracy of the subsequent use of the knowledge graph constructed by the ontology to infer new knowledge.

The present invention improves the probability of extracting the same type of entities to replace the original triples when constructing negative samples by using the relationship type constraints, which is beneficial to increase the distance between entities of the same type, that is, to increase the distance between the vector representations of entities Using the prior knowledge of the relationship to decide which entities to replace by the relationship can significantly improve the prediction accuracy of the knowledge inference model.

Description of drawings

Fig. 1 is an urban traffic knowledge map construction process according to an embodiment of the application;

Fig. 2 is an urban traffic ontology construction process according to an embodiment of the application;

3 is a class hierarchy diagram of an urban traffic ontology according to an embodiment of the application;

FIG. 4 is an inter-class relationship diagram of an urban traffic ontology according to an embodiment of the application;

FIG. 5 is a schematic diagram of visualization of road network information according to an embodiment of the present application;

6 is a schematic diagram of visualization of road traffic situation information according to an embodiment of the application;

FIG. 7 is an inter-class relationship diagram of a public transportation ontology according to an embodiment of the application;

FIG. 8 is a diagram showing the relationship between a public transportation travel user and a itinerary, and a itinerary and a site according to an embodiment of the application;

FIG. 9 is an entity and relationship diagram of a road traffic knowledge graph according to an embodiment of the application;

FIG. 10 is a relationship diagram of a road and a traffic situation according to an embodiment of the application;

11 is an explanatory diagram of a TransD model according to an embodiment of the application;

Fig. 12 is a model training process diagram according to an embodiment of the application;

13 is a flowchart of link prediction task evaluation according to an embodiment of the application;

FIG. 14 is a graph showing the change of loss value in model training according to an embodiment of the present application;

15 is a comparison diagram of the average reciprocal ranking results of multiple inference models according to an embodiment of the present application;

16 is a comparison diagram of the top ten hit rate results of multiple inference models according to an embodiment of the present application;

17 is a comparison diagram of the average speed of a street in the morning rush hour before and after school starts according to an embodiment of the application;

FIG. 18 is a schematic diagram of a road traffic knowledge graph according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

As shown in FIG. 1 , the present invention discloses a method for constructing an urban traffic knowledge graph, comprising the following steps:

Use ontology construction tools to build urban traffic ontology and form a knowledge graph model layer;

Obtain urban traffic data, extract entities, attributes and relationships between entities, and build a knowledge map data layer;

Combine the knowledge graph mode layer with the knowledge graph data layer to generate an urban traffic knowledge graph and store it in the database;

The knowledge representation model is used to infer the new knowledge in the urban traffic knowledge graph, and add it to the urban traffic knowledge graph.

The invention forms a traffic knowledge system by using a knowledge graph, integrates multi-source heterogeneous traffic big data, models the spatiotemporal relationship between traffic entities, and mines the potential relationship between traffic entities through a knowledge inference model based on representation learning , realizes the effective integration and organization of multi-source travel data in the field of transportation, forms a knowledge network in the field of transportation, and realizes the open sharing of data in the field of transportation.

In order to facilitate the understanding of the technical solution of the present invention, the construction method of the present invention is further explained below;

1. Use ontology construction tools to build urban traffic ontology (ie, urban traffic travel ontology), and form a knowledge map model layer;

The construction of the ontology in the field of urban transportation is to comprehensively consider the acquired data resources for specific business scenarios, and consider the standardization of domain terms and the wide applicability of concept categories, abstract the knowledge hierarchy in the transportation domain, and define the contents of the ontology. Classes, attributes of classes, and associations between classes.

The ontology defines the mold of the knowledge graph and describes the top-level structure of the knowledge graph. Therefore, building an ontology can analyze the system or level of domain knowledge and realize the reuse of domain knowledge.

As shown in Figure 2, this application uses the seven-step method released by Stanford University to build the ontology. The method of manually building the ontology can be selected, and the quality assessment part is added on the basis of the seven-step method. The specific ontology building process is as follows:

Step 1: Determine the domain of the ontology, clarify the purpose of the domain ontology, and the domain involved in the ontology of this application is the urban transportation domain;

Step 2: Check whether there is a reusable ontology, investigate whether there is a related ontology that has been built, and if so, it can be imported directly to save construction cost and time;

Step 3: Identify important terms, and ensure comprehensiveness when listing all terms;

The fourth step: define the class and its hierarchical structure. This application adopts the top-down method, that is, the broadest concept in the field is first defined, and then refined, as shown in Figure 3;

Step 5: Determine the attributes of the class, only relying on the class can not provide sufficient information, so it is necessary to define the attributes of the class to further describe the class. Determining the attributes of the class also includes defining the relationship between classes. There are some relationships in the urban traffic ontology. The two classes are connected by a relationship. For example, the location of a certain interest point is on a certain road, and the relationship between the interest point and the urban road is located. As shown in Figure 4, it represents the relationship between all classes in the urban traffic ontology. Relationship.

It should also be noted here that classes are inherited, and subclasses will inherit the properties of their parent classes, so properties should be placed in the broadest class, the closer to the top level, the better. Table 1 shows several classes and properties. As shown in Table 1, the property should be placed in the broadest class to ensure that subclasses can inherit the property;

Table 1

Step 6: Define the constraints of attributes, such as restricting the type of attribute values, the number range of attribute values, etc.;

Step 7: Create a concrete instance of the class after passing the quality assessment.

Among them, the quality assessment should follow the principles of clarity, consistency, scalability, simplicity and independent sharing. The specific quality assessment methods are:

(1) First, evaluate the hierarchical structure of the class. By drawing a tree structure diagram, the root node and sub-nodes are defined. Each sub-tree corresponds to independent and modular knowledge in the field. Whether transitivity is established to avoid loops in the class hierarchy;

(2) Evaluate the ontology, confirm that the expression of the ontology is clear and clear, and check whether the application scope and expression method of each ontology are consistent when used everywhere, including the consistency of the relational logic, whether there are duplicate defined classes, and whether there is a need Merged ontology to avoid redundancy;

(3) Evaluate the defined attributes and attribute constraints, check whether the description information of the attributes is complete, and whether the attribute constraints are logical; the sharedness of the evaluation attributes is widely applicable to multiple classes, not just limited to a certain class. ;

(4) Check the scalability of the ontology to ensure that the ontology can be continuously improved and updated flexibly with the continuous increase and modification of classes and attributes;

(5) Check the consistency of the logic of the relationship between classes, and evaluate the integrity of the relationship between the classes, check that the ontology is enough to include all possible relationships between classes, and check the uniqueness of the relationship between classes, that is, check whether the classes are Enough that there is only one relationship.

As shown in Figure 2, if the ontology construction fails to pass the quality assessment, it should go back to the third step to construct the ontology again until the constructed ontology can pass the quality assessment, and then create a specific instance.

Compared with the original seven-step method, the present invention adds a quality assessment step when constructing the ontology, and strictly controls the quality of the ontology added to the knowledge base through the quality assessment. The evaluation is carried out to realize the reusability of the standardized ontology, to ensure the accuracy and validity of the ontology construction, and to improve the accuracy of the subsequent use of the knowledge graph constructed by the ontology to infer new knowledge.

The ontology construction tool described in this application can be protégé 5.5.0 software, and the ontology constructed by this software can intuitively see classes, attributes and relationships.

2. Obtain urban traffic data, extract traffic entities, attributes and relationships between entities, and build a knowledge map data layer;

Urban traffic includes public traffic and urban road traffic, so the urban traffic data to be acquired in this application includes public traffic data (that is, public traffic travel data) and road traffic data (that is, road traffic travel data). It is different from the time and space of the traffic situation data that road traffic needs to obtain, so it is necessary to build the public traffic data layer and the road traffic data layer respectively, and build the public traffic knowledge map and the road traffic knowledge map respectively.

S1. Build the public transport data layer: the acquisition of public transport travel data;

The method for obtaining the travel data of public transport is to obtain the subway line and station information of the target city through the web crawler technology, obtain the subway card swiping data of the subway line within the target time, and perform preprocessing;

The specific method is as follows:

S11. Obtain subway line and station information;

This application uses the Internet crawler technology to obtain data such as subway stations and lines in the target city. The web crawler mainly uses the unique website address (URL) to find the website, and automatically grabs and downloads the target information from the website. The specific operation methods are:

Step1: First, you need to construct the incoming parameters, mainly including key value, city code, and city name. After the parameters are URL-encoded, a request is made to the target HTTP interface, that is, a request is sent;

Step2: Receive the response returned by the HTTP request, parse the returned data, and the data format is json;

Step3: Store the parsed data in the PostgreSQL database. The stored data is shown in Table 2, including the line and station name, number, station serial number, station latitude and longitude, whether it is transferable, and the route passing through this station.

Line and station information stored in Table 2

S12. Obtain subway AFC card swiping data and perform preprocessing;

The following takes the target city as Shenzhen and the target time from January 25, 2016 (Monday) to January 29, 2016 (Friday) 7:00-9:00 as an example to introduce the method of obtaining subway card swiping data and preprocessing in detail. , other cities can use this method for data acquisition and preprocessing.

Shenzhen Metro adopts the method of charging by mileage, so passengers need to swipe their cards when entering and leaving the station. As of 2020, there are 8 subway lines in Shenzhen (Shenzhen Metro Line 1-5, Line 7, 9 Line, Line 11), 182 subway stations, the data in this application is the data of Shenzhen Tong credit card, the time range is from January 25, 2016 (Monday) to January 29, 2016 (Friday), a total of 5 days The card swiping records on working days, the data format is shown in Table 3.

Table 3 Metro AFC swipe data format

In the original subway AFC card swiping data, it is necessary to distinguish the inbound and outbound type (ie transaction type) according to the field (COST_TYPE), where COST_TYPE=21 indicates that the data is inbound information, which is the transaction start state, and COST_TYPE=22 indicates that the data is outbound Inbound and outbound data needs to appear in pairs.

(1) Target data screening: Screen out the card swiping records during the morning peak (7:00-9:00), and count the number of times the same IC card is swiped during the morning peak period every day, and select the records with the number of 2 times (one-by-one). out)

(2) Elimination of abnormal data: Eliminate the card swiping records that are not paired when entering and leaving the station, that is, whether there is entry or exit, or whether there is entry or exit. Delete the missing data of the station location in the card swipe record, and delete the data with the same entry and exit location, and the data whose entry and exit time difference is greater than 5 hours.

(3) Add itinerary ID: Combine the two lines of inbound and outbound data adjacent to each other into one line of data, including the start and end time and station name, and add a column of field as the itinerary ID to distinguish the user's multiple trips within a week .

The preprocessed subway card swiping data format is shown in Table 4.

Table 4 Subway credit card data after preprocessing

S2. Build a road traffic data layer: acquisition and preprocessing of road traffic travel data;

The method of obtaining the road traffic travel data is to obtain the target road network data from the map database, and use the map API to obtain the target point of interest information and traffic situation data on the target road.

The specific steps are:

S21. Acquisition of target road network data;

Step1: Taking the target road as the area within the Fifth Ring Road of Beijing as an example, obtain the road data of the area within the Fifth Ring Road of Beijing from the map database. For example, you can download the Fifth Ring Road of Beijing on the OpenStreetMap open source map database from the BBBike website. Road data in the area; BBBike website supports a variety of data export formats (such as OSM, Shapefile, GeoJSON, etc.), and can customize the download map to get the area range, is a better target road network data download website, of course, here The download site and map database of the above are exemplary and not limiting.

Step2: Use OSM2GMNS to output road network data that conforms to the GMNS standard from the downloaded target road network. The full name of GMNS is General Modeling Network Specification, which defines a flexible and unified representation format for multi-mode transportation networks. OSM2GMNS also provides a simplified intersection function. , the output file includes road nodes (node.csv) and road connection arcs (link.csv). The main field descriptions are shown in Table 5 and Table 6.

Table 5 Field descriptions of road nodes

Table 6 Field descriptions for road connecting arcs

Among them, the node type (OSM_HIGHWAY) is divided into intersections with expressways, intersections with traffic signal control, and intersections without traffic signal control; the link class (LINK_TYPE_NAME) is divided into expressways, arterial roads, secondary arterial roads, Branch road, community road.

Step3: Import the data into the relational database and visualize it in the QGIS software.

As shown in Figure 5, taking the intersection of Yuquan Road and Shijingshan Road as an example, the dotted line in the figure is the road connecting arc, the dots are the road nodes, and there is a connecting arc between every two nodes.

S22, the acquisition of the interest point of the target road;

POI (Point of Interest) generally refers to points with practical significance in the real world, such as facilities or buildings related to people's lives, such as parking lots, schools, hospitals, etc. POI data generally includes name, type, address and Basic attributes such as latitude and longitude.

Points of interest can be obtained through the search POI interface in the map API. This application uses the AutoNavi map API to obtain the information of primary and secondary schools within the Fifth Ring Road of Beijing as an example to describe the process of obtaining points of interest.

Step1: First determine the query area, use QGIS to draw the boundary of Beijing's Fifth Ring Road, export it to GeoJSON format, and process it as a csv file (boundary coordinate pair) with one column of longitude and the other of latitude.

Step2: Because each request returns a maximum of 1000 POI information, it is necessary to divide the large area into multiple small grids. Set the grid size to 10km*10km, map the boundary coordinates to the grid, and finally get the coordinates of the lower left corner vertex of each small grid in the area.

Step3: Determine the code of the queried POI type (141202 for secondary schools and 141203 for primary schools). After that, the POI information in each grid is obtained in turn. The request parameters are the key value, the vertex coordinate pair, and the POI type code of the query. After the parameters are URL-encoded, a request is sent to the target HTTP interface.

Step4: Parse the data returned in json format and store it in the database. In the end, the number of primary schools and 442 middle schools in the fifth ring is 598 (there is a situation where a school has multiple campuses, that is, a school name corresponds to multiple POI data). The POI information field descriptions are shown in Table 7.

Table 7 POI information field description

S23. Acquisition of target road traffic situation data;

This application takes the study of the relationship between students going to and from school and the road traffic situation as an example to describe the acquisition steps of the target road traffic data:

According to Beijing's 2020 autumn semester start schedule, from August 29, elementary, junior high, and high school classes will start, and September 7, all grades of elementary, junior high and high schools will start. 50, school ends earlier than 16:30; middle school students arrive later than 7:30 and dismiss earlier than 17:30.

In order to study the relationship between students going to and from school and the road traffic situation, the time range for collecting the traffic situation in this application is one week before the start of school for students in 2020: August 24 (Monday) to August 28 (Friday), when students start school The next week: September 21st (Monday) to September 25th (Friday), the daily collection period is from 6:30 to 10:30 in the morning (4 hours), and from 16:00 to 21:00 in the afternoon (5 hours) ), the spatial scope of collecting traffic situation is the traffic situation data of roads in the area within the Fifth Ring Road of Beijing.

Obtain traffic situation data through the traffic situation interface of AutoNavi Maps API. The method of obtaining traffic situation data is similar to the method of obtaining POI. The data is also queried according to the method of rectangular area, but the length of the diagonal of the rectangle is required to be less than 10 kilometers. The large area needs to be divided into multiple small grids to break through the limits of this area.

Step1: Determine the query area, use QGIS to draw the boundary of the Fifth Ring Road in Beijing, export the data and process it as a csv file with one column for longitude and the other for latitude, that is, the pair of boundary coordinates

Step2: Divide the area into multiple small grids of 7km*7km, map the boundary coordinates to the grids, and finally obtain the coordinates of the lower left corner of each small grid in the area.

Step3: Obtain the traffic situation data of the roads in each grid in turn. The request parameters are the key value, the road level, the vertex coordinate pair of the rectangular area, and the return data format type. After the parameters are URL encoded, a request is made to the target HTTP interface. Because the traffic situation service of AutoNavi Maps API limits the daily call volume of individual developers to 2,000 times/day, the excess will be blocked, so it is necessary to set a single KEY request more than 2,000 times, switch to the next KEY, and continue to send requests.

Step4: The road condition information is updated every 2 minutes, and the returned json data is parsed and stored in the database. As shown in Table 8, the fields included in the returned result include road name, direction description, driving angle, road condition, speed (km/h), etc.

Table 8 Description of fields of traffic situation information

Among them, the driving angle (ANGLE) reflects the driving direction of the vehicle on the road, in which the due east direction is set to zero degrees, and it is a positive number when it rotates in the counterclockwise direction, and the value range is [0, 360]. The traffic status (STATUS) reflects the traffic status of the road, where 0 represents unknown state, 1 represents smooth traffic state, 2 represents vehicle slowing state, 3 represents traffic congestion, and 4 represents serious congestion state.

Figure 6 shows the result of visualization of traffic situation data in QGIS at a certain time. It can be found that the roads with traffic situation return in the area within the Fifth Ring Road of Beijing are mainly urban expressways and trunk roads.

S24. Preprocessing of road traffic travel data:

(1) Unified data sampling interval: The real-time traffic situation data of the road is obtained by using web crawler technology. The data on the AutoNavi map platform is updated every 2 minutes, but due to the instability of the network connection, the return time interval of the traffic situation data is different. Between 4 minutes and 7 minutes. In order to facilitate subsequent analysis, the data sampling interval needs to be fixed at 5 minutes.

(2) Data matching and screening: Because most roads with traffic situation return are highways and main roads, the road network data with traffic situation data (including network nodes and road network connection arcs) to eliminate invalid and redundant data, thereby improving data quality.

3. Combine the knowledge graph mode layer and the knowledge graph data layer to generate an urban traffic knowledge graph and store it in the database;

Due to the difference in time and space based on card swiping data and traffic situation data, the public transportation data layer and the urban road traffic data layer are constructed respectively. It is also necessary to generate the public transportation knowledge map and the road traffic knowledge map respectively, and store them in the database. The applied urban traffic knowledge map includes public traffic knowledge map and road traffic knowledge map.

(1) Selection of database;

The database used to store the knowledge graph can be a graph database or a relational database and other storage databases. In this application, a graph database is preferred as the storage database. The graph database can be selected from Neo4j, TitanDB, etc. Edges form graphs and express information intuitively through a graphical structure. The unique data structure of graph databases can effectively store and express the knowledge in the knowledge graph and the relationship between entities. For example, nodes in a graph database represent entities, and edges represent entities. Therefore, when storing knowledge graphs, the storage effect of graph databases is better than other storage databases such as relational databases.

(2) Storage of public transportation knowledge graph;

Combine the knowledge graph mode layer with the public transportation knowledge graph data layer to generate the public transportation knowledge graph and store it in the database;

As shown in Figure 7, Figure 7 shows the inter-class relationship of the urban traffic ontology about public transportation, the classes in the ontology correspond to the labels of the nodes in the graph database, and the entities are stored in the form of nodes in the knowledge graph. The inter-class relationship, that is, the relationship between entities, is divided into the following categories: the relationship between the subway station and the line (subway_station-belong-subway_line), the adjacent relationship between the subway station entities (subway_station-belong-subway_line), public transportation users and The attribution relationship of the trip (user-has-trip), the start and end relationship between the trip and the station (trip-start_from/end_at-subway_station).

The same public transportation user will have multiple trips in a week, so there is a one-to-many relationship between the user and the trip. As shown in Figure 8, there is a start or end relationship between the trip and the station.

(3) Storage of road traffic knowledge graph;

Combine the knowledge graph mode layer with the road traffic knowledge graph data layer to generate the road traffic knowledge graph and store it in the database;

The entities in the road traffic knowledge graph include urban roads, points of interest, and traffic situations in the domain ontology, and entities representing spatiotemporal relational data are added. As shown in Figure 9, 9 is the entity and the entity contained in the road traffic knowledge graph. relationship between.

Among them, the relationship between the road and the traffic situation is represented by a multi-step relationship path. Date and time are the time attributes of the traffic situation. The road entity is first associated with the date, then the date is associated with the time entity, and the last time is associated with the traffic situation entity. The three-step relationship path of the road entity is associated with the traffic situation, as shown in Figure 10.

Fourth, use the knowledge representation model to infer new knowledge and add it to the knowledge graph.

Although the public transportation knowledge graph and road traffic knowledge graph constructed in this application integrate multi-source data, there is a problem of sparse data, so it is necessary to excavate the missing entities and potential relationships in the knowledge graph and add it to the knowledge graph .

The present application mines the implicit relationship in the urban traffic knowledge graph through the knowledge reasoning model to complement the knowledge graph, and verifies the validity of the model through the link prediction task on the urban traffic knowledge graph. The specific steps are as follows:

S1. Construction of knowledge reasoning model

Conventional knowledge reasoning models can be used in this application to carry out knowledge reasoning mining, such as TransE and TransH models, but according to the characteristics of knowledge reasoning in this application, the preferred TransD model is used as a knowledge reasoning model. The main idea of the model is to use the projection vector to construct The dynamic mapping matrix of , encodes entities as low-dimensional embedding vectors in the relation space, taking into account that entities and relations have different types and properties, as shown in Figure 11.

In the TransD model, the first vector (h, r, t) represents the actual meaning of the entity or relationship, and the second vector (h _p , r _p , t _p ) is used for the construction of the mapping matrix, which consists of entities and The projection vector of the relationship is jointly determined, and the entity can be mapped from the entity space to the vector space, and each mapping matrix is initialized with the identity matrix I, and the vector operation is used to replace the matrix multiplication operation, which effectively reduces the amount of calculation.

为头实体映射向量，

为尾实体映射向量，I^mxn为单位矩阵；Among them, M _rh is the head entity mapping matrix, M _rt is the tail entity mapping matrix, r _p is the relationship vector,

map vector for head entity,

is the tail entity mapping vector, and I ^mxn is the identity matrix;

Project the entity vector into the relational space and embed as:

h _⊥ =M _rh h

t _⊥ =M _rt t

h _⊥ is the head entity vector after the head entity is mapped by M _rh , t _⊥ is the tail entity vector after the tail entity is mapped by M _rt ;

The TransD model regards the relation vector r in the triplet (h, r, t) as obtained by the translation operation of the projection vector of the head entity vector h and the projection vector of the tail entity vector t in the relation space, that is, in the relation space The sum of the projection vector of the head entity and the projection vector of the relation is approximately equal to the projection vector of the tail entity. So define a score function based on L2 Euclidean distance to measure the distance between these two vectors:

The model adds L2 norm constraints to the vector, which can make the parameters related to the model smaller, avoid the problem of over-fitting of the model, and make it have a strong generalization ability.

||h|| ₂ ≤1,||t|| ₂ ≤1,||r|| ₂ ≤1,||h _⊥ || ₂ ≤1,||t _⊥ || ₂ ≤1

From the above score function, it can be seen that for a correct triplet, the larger the score value is, the better, while the wrong triplet is expected to have a smaller score value, the better. Therefore, this application defines a distance interval-based ranking loss function. , with minimizing the loss function value as the training objective of the model.

where L is the loss function and γ is a hyperparameter representing the maximum separation between correct triples and negative triples. [x] ₊ =max(0,x), Δ represents the set of correct triples, and Δ′ represents the set of constructed negative triples.

S2. Data set division, model parameter setting and negative sample construction;

S21. Data set division;

Export the entity and relational data in the urban traffic knowledge graph stored in the relational database. For example, when the graph database Neo4j is used to store the knowledge graph, the APOC (A Package of Components) plug-in of Neo4j can be used to export the entity and relational data, and use the database The search and filtering function preprocesses the data into data in csv format, entity data is stored in the form of entity name plus the id corresponding to the entity, relational data is stored in the form of the relationship name plus the id corresponding to the relationship, and triples are stored in the form of the header entity id plus The tail entity id is added to the relationship id, and the triplet data is divided into training set, test set and validation set according to a predetermined ratio. The predetermined ratio is selected according to actual needs, for example, 85% of the training set: 10% of the test set: 5 % of the validation set.

S22. Model parameter settings

There are many hyperparameters in the TransD model, that is, the parameter needs to be manually set during training. The hyperparameters included in the model include the learning rate α, the embedding dimension k, the number of samples in each batch batch_size, and the interval γ. The size of each parameter is set to The influence and setting range of the model are:

① Learning rate α: When the learning rate is set too large, the loss value may fail to converge; if the learning rate is set too small, the training time required for the model to converge becomes longer. The learning rate setting range of this application is {0.001, 0.01, 0.1 };

②Embedding dimension k; if the embedding dimension is set too low, the performance is not enough, and if the embedding dimension is set too high, it is easy to over-fit, and the setting range of the embedding dimension is {20, 50, 100, 150};

③The number of samples in each batch batch size: If the batch size is set too large, the program may crash due to limited memory space. If the batch size is set too small, it will make the model convergence more difficult, because the parameters obtained if the sample size is too small are not representative. , which affects the generalization performance of the model. The number of samples in each batch is set in the range of {64, 128, 256, ;

④Interval γ: The setting range is {0.25, 0.5, 1, 2, 3}.

S23. Construction of negative samples

Since there are only correct triples in the urban traffic knowledge graph, it is necessary to construct wrong triples as negative samples to participate in the training, verification and evaluation of the model. The construction method of the negative samples in this application is as follows:

S231, Bernoulli negative sampling

Considering the different types of relationships, when constructing negative triples for triples with a certain relationship, the party with a smaller number of entities should have a greater probability to be selected for the replacement operation, and statistics with a certain relationship All triples of r, the amount of data about:

① The average number of tail entities associated with the head entity, denoted as tpt

②The average number of head entities associated with tail entities, denoted as hpt

The random variable X takes only two values, 0 and 1, and the corresponding probability is:

The construction of the final negative sample obeys the Bernoulli distribution with parameter p, and the distribution law of the random variable X is:

P(X=x)=p ^x (1-p) ^(1-x) ,x=0,1

For a correct triple with relation r, the probability of selecting the head entity is p, the probability of selecting the tail entity is 1-p, and the entity with higher probability is replaced to construct a negative triple.

S232. Relation type constraint

The type constraints of the relationship are represented by defining the entity type that the relationship should be associated with, and the prior knowledge of the relationship type is used to determine which entities to replace by the relationship, and the following variables are defined:

①Ordered index domain _{r of all entities within the domain constraints that satisfy the relation type r}

② The ordered index range _r of all entities within the range constraint of the relation type r is satisfied

For all triples with a certain relation r, when constructing a negative triple, calculate the probability of selecting the head or tail entity to complete the replacement operation according to the Bernoulli distribution. Select from the entity subset within the relationship type, if the tail entity is selected, select from the entity subset within the scope of the relationship type, as shown in the following formula:

Among them, Δ′ is the set of constructed negative triples, d _r is the ordered index of all entities that satisfy the domain constraints of relation type r; r _r is the ordered index of all entities that satisfy the range constraints of relation type r, h is the head entity of the triplet, h' is the head entity of the negative triplet, t is the tail entity of the triplet, t' is the tail entity of the negative triplet, and r is the relation.

The construction method of the negative triplet in the prior art is: for any triplet (h, r, t), randomly extract an entity from the set E containing all entities and use the head entity or the head entity in the original triplet. If the tail entity is replaced, an erroneous triple will be obtained. However, due to the existence of one-to-many, many-to-one and many-to-many relationship types, this method of randomly sampling and constructing negative samples will introduce many wrong negative samples. That is, False Negative (that is, data in both triples and negative triples) This application uses Bernoulli negative sampling to perform negative triples on triples with a certain relationship. During construction, the party with a smaller number of entities is selected for the replacement operation, which greatly reduces the amount of false negatives in negative samples.

At the same time, by using the relationship type constraint, the probability of extracting the same type of entities to replace the original triples when constructing negative samples is improved, which is beneficial to increase the distance between the same type of entities, that is, to increase the vector representation of entities. The difference, using prior knowledge of the relationship, determines which entities to replace by the relationship, can significantly improve the accuracy of the model prediction.

At the same time, even if the negative samples are constructed using the relation type constraint method, it is still unavoidable that there are a small number of false negatives in the negative samples. The present application also discloses a method for filtering out the false negatives in the negative samples. The tuple data (that is, the triplet data in the training set, the validation set and the test set) and the negative triplet data in the negative sample are imported into the relational database, for example, the PostgreSQL database can be selected, and the duplicate data can be duplicated using the query function in the relational database. Find out, and remove the duplicate data in the negative sample; wherein, the duplicate data is the data that exists both in the triplet data and in the negative triplet data.

Through the filtering operation, the interference of negative samples during model retraining, validation and prediction is avoided, and the accuracy of the model is further improved.

S3. Training of knowledge inference model

Bring the training set divided in S21 and the model parameters set in S22 into the inference model constructed by S1 for model training, and use the negative sample construction method described in S23 to construct a negative sample of the training set for the triples in the training set. The negative samples of the training set and the training set are brought into the knowledge inference model for model training,

At the same time, the Mini-batch Gradient Descent method can be used to update the parameters during model training to obtain the minimum value of the loss function. The model updates the vector representation through continuous iteration until the loss function value of the model converges. , or the model has been trained to the maximum number of times, and the embedding representation of entities and relationships is obtained after the training is completed. The model training process is shown in Figure 12.

During the training process, an inappropriate setting of the learning rate will adversely affect the learning effect. If the learning rate is set too large, the model may not converge, and the loss value will fluctuate continuously. If the learning rate is set too small, the model will converge faster. It is slow and requires a long training time. In this application, when using the small batch stochastic gradient descent method to update and learn the parameters, the adadelta method is used to adjust the learning rate adaptively during the training process, which prevents the manually set learning rate from being inconsistent. The impact of appropriateness on learning outcomes.

S4. Verification of knowledge reasoning model

The verification set divided in S2 is brought into the knowledge inference model trained in S3 for preliminary verification and evaluation, and the hyperparameters are adjusted according to the verification results.

S5. Evaluation of Knowledge Reasoning Models

This application verifies the validity of the model by linking the prediction task, and selects multiple evaluation indicators to evaluate the comprehensive ability of the inference model;

S51. Link prediction

Link prediction refers to the task of predicting another entity that has a specific relationship with a given entity, i.e., for a triple (h r t), predict the head entity h given the known relationship r and tail entity t, denoted as ( ?r t), or make predictions on tail entity t given head entity h and relation r, denoted (h r ?), or make predictions on relation r given head entity h and tail entity t , denoted as (h?t).

S52, construct negative samples of the test set and rank them

Using the negative sample construction method and filtering operation described in S23, a negative sample of the test set is constructed for the triples in the test set. Taking the predicted tail entity as an example, the relationship type determines which entities in the entity set are selected to replace the tail entity. And after the filtering operation is performed, the triplet score and the negative triplet score constructed according to the triplet and the entities in the knowledge graph are calculated according to the score function in the trained knowledge inference model, and the scores are calculated from the largest to the largest. Rank the triplet and the negative triplet constructed according to the triplet and entities in the knowledge graph from the smallest to the smallest;

At the same time, for the predicted head entity or the relationship between predicted entities, the above method can be used to count the ranking situation.

S53. Evaluation indicators

This application can choose the mean rank (MR), the mean reciprocal rank (Mean Reciprocal Rank, MRR), the first hit rate (hits@1), the top three hit rate (hits@3) and the top ten hit rate (hits@10 ) to measure the effect of link prediction task completion.

①Mean Rank (MR): The average rank represents the rank order of the correct test set triples in the set of test set negative triples obtained by all replaced head or tail entities. The higher the ranking, the better the effect of the model. it is good.

②Mean Reciprocal Rank (MRR): The average reciprocal ranking can reflect the overall ranking of the triples in all test sets in the constructed list of negative triples in the test set, as shown in the following formula.

Among them, T represents the set of triples in all test sets, |T| represents the amount of data in the set T, and rank(h, r _i , t) represents the rank order corresponding to the correct triples.

③The first hit rate (hits@1): The first hit rate represents the ratio of the number of triples ranked first in the test set to the total number of triples in the test set, as shown in the following formula.

Among them, Ind(x) represents the indicator function, which is used to judge whether the correct triple (h, r, t) is in the first place, if it is satisfied, output 1, otherwise output 0.

Similarly, the top three hit rate and the top ten hit rate can be defined. For example, the top ten hit rate (hits@10) indicates that the number of triples in the top ten in the test set accounts for all the triples in the test set. ratio of numbers. as shown in the following formula.

The higher the value of the top ten hit rate, the better the effect of the model.

In summary, the process of evaluating the model using the link prediction task is shown in Figure 13.

In order to verify the improvement effect of the selection of knowledge reasoning model, filtering operation and relation type constraints on the prediction results, this application compares the influence of the presence or absence of filtering operations, relation type constraints and multiple reasoning models on the prediction task results, and the results are as follows Show:

(1) The influence of the presence or absence of filtering operations on the prediction results

Table 9 Comparison of prediction results with and without filtering operations

As shown in Table 9, on the Mean Reciprocal Rank (MRR) indicator, the filtering operation is improved by 30% compared with the non-filtering operation. The filtering operation is increased by 10.2%, and the top three hit rate indicators and the first hit rate indicators are also improved, indicating that the filtering operation can effectively improve the accuracy of the model prediction.

(2) The influence of relationship type constraints on the prediction results

On the basis of the filtering operation, compare the prediction results obtained by using the relation type constraints when constructing negative triples. Relational type constraints refer to the use of prior knowledge of relational types to determine which entities are replaced by the relation when constructing a triple. Unused relationship type constraint refers to randomly extracting entities from all entities for replacement when constructing negative triples. The prediction results are shown in Table 10.

Table 10 Comparison of prediction results with and without relation type constraints

As shown in Table 10, the mean rank (MR) of the head or tail entities of the correct triplet with relation type constraint operation is significantly better than that without relation type constraint, and the same is true for the rest of the evaluation metrics, indicating that there is a relation type Constraints can improve the performance of the model on the link prediction task.

(3) The influence of different knowledge inference models on the prediction results

This application compares the TransD model with the classic reasoning models TransE and TransH, and also performs filtering operations and relationship type constraints. The results of the three models in the entity link prediction task are shown in Tables 11-12.

The TransE model represents the relationship as the translation of the head and tail entity vectors, the TransH model represents the relationship as the translation of the entity vector on a specific relationship hyperplane, and the TransD model projects the vector representation of the entity into the relationship vector space through a dynamic mapping matrix. Think of the relation vector as a translation of the entity's projection vector. The optimal hyperparameter settings for different models are shown in Table 13.

Table 13 Values of hyperparameters for different models

This application draws a process in which the loss function value of the TransD model continues to decrease until it converges during the training process, as shown in Figure 14, where the horizontal axis in the figure is the number of iterations of the model, and the vertical axis is the loss function value of the model.

On the data set of the public transportation travel knowledge graph constructed in this application, the performance and accuracy of different reasoning models are compared according to the evaluation indicators by linking the prediction task. The result of the tail entity.

Table 11 Results of predicting head entities

Table 12 Results for predicting tail entities

Select the evaluation indicators MRR and hists@10 as representatives, and draw a bar graph, as shown in Figures 15-16, which can intuitively compare the prediction results of different models. outperforms the other two inference models. For example, when predicting head entities, the mean reciprocal ranking (MRR) of the TransD model is improved by 6.6% compared to that of the TransH model, and when predicting tail entities, the metric is improved by 22.4%.

To sum up, in this application, the TransD model is selected as the knowledge inference model, the negative samples are constructed through the relationship type constraints, and the false negative examples in the negative samples are eliminated by the filtering operation, which effectively improves the accuracy of the prediction results. The prediction effect of the knowledge inference model trained by the method described in this application is better, and the inferred entity and the relationship between the entities are more accurate when the urban traffic knowledge map is completed.

The present application also discloses an apparatus for constructing an urban traffic knowledge graph, including an ontology construction module, a data acquisition module, a storage module, and a reasoning module.

Among them, the ontology building module is used to construct the urban traffic ontology by using ontology building tools to form the knowledge graph model layer; the data acquisition module is used to obtain the urban traffic data, extract the relationship between traffic entities, attributes and entities, and construct the knowledge graph data layer; the storage module is used to combine the knowledge graph mode layer and the knowledge graph data layer to generate the urban traffic knowledge graph and store it in the database; the inference module is used to use the knowledge representation model to infer new knowledge in the urban traffic knowledge graph. Add the knowledge map of urban traffic.

The present application also discloses the application of any of the above-mentioned methods for constructing an urban traffic knowledge graph in the field of urban traffic, for example, applying the method for constructing an urban traffic knowledge graph of the present application to predict the traffic state so as to reasonably arrange travel time; The applied knowledge graph can query the shortest distance path between subway stations; the applied knowledge graph can be used to query commuter companions based on the similarity of travel chains.

Example 1

The method described in this application will be used below to predict the speed value of a road near a primary and secondary school at 7:15 after the school starts, and verify whether the knowledge inference model of this application can complement the speed value of the road at a certain moment. ;

By comparing the actual average speed of the road before and after the school starts, it is found that the time when the average speed of the road decreases during the morning rush hour after school starts will be about half an hour earlier, as shown in Figure 17. After school started, the average road speed dropped from 45km\h to 40km\h from 7:00 in the morning. At 7:15, the average road speed actually dropped to 35km\h.

The method described in this application is used to predict the speed of the above-mentioned road at 7:15 after the school starts. The relationship between entities and entities included in the road traffic knowledge graph is shown in Figure 18. The school is located on the road, and the road and the There are relational paths between velocity values. A certain road and its associated entities are selected as the experimental data set, the number of entities is 53, the relationship is 523 pairs, the training set is 417, the verification set is 53, and the test set is 53. For the relationship between time and speed (time_speed), take time as the head entity, and predict the speed value of the tail entity. Table 18 below shows the top five entities when predicting tail entities, and the bolded ones in the table indicate the correct tail entities.

Table 14 Top five entity names when predicting tail entities

As shown in Table 14, the tail entity that ranks first among the tail entities predicted by the method described in this application is the correct tail entity, indicating that the prediction result of the knowledge inference model of this application is relatively accurate, and the knowledge of this application can be used. The inference model completes the speed value of the road at a certain moment.

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art can still Modification of the technical solutions, or equivalent replacement of some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of the technical solutions of the present invention. within the scope of the claims and description.

Claims (10)

1. A construction method of an urban traffic knowledge graph is characterized by comprising the following steps: the method comprises the following steps:

constructing an urban traffic body by using a body construction tool to form a knowledge map mode layer;

acquiring urban traffic data, extracting the relation among entities, attributes and entities, and constructing a knowledge map data layer;

combining the knowledge map mode layer with the knowledge map data layer to generate an urban traffic knowledge map, and storing the urban traffic knowledge map in a database;

and deducing new knowledge in the urban traffic knowledge map by using the knowledge representation model, and supplementing the urban traffic knowledge map.

2. The construction method according to claim 1, characterized in that: the method for constructing the urban traffic ontology comprises the following steps:

constructing an urban traffic body by adopting a seven-step method, and performing quality evaluation before creating an example;

the quality assessment specifically comprises:

verifying whether the transitivity of the hierarchical structure of the class is established or not by drawing a tree structure diagram;

checking whether the application range and the expression mode of each ontology are consistent when used everywhere, and whether redundancy of classes and ontologies appears;

checking whether the description information of the attribute is complete, whether the attribute constraint accords with the logic and whether the attribute has the sharing property;

checking the expandability of the body;

the integrity, uniqueness and logical consistency of the relationships between the classes are checked.

3. The construction method according to claim 1, characterized in that: the specific method for reasoning out new knowledge in the urban traffic knowledge map by using the knowledge representation model comprises the following steps:

dividing triple data in the urban traffic knowledge map into a training set, a verification set and a test set;

constructing a negative sample of the ternary data, and filtering false negative examples in the negative sample;

setting a knowledge representation model hyper-parameter;

training a knowledge representation model based on a small batch random gradient descent method by using a training set and a negative sample, and adaptively adjusting the learning rate in the training process by using an adapelta method;

carrying out hyper-parameter adjustment on the trained knowledge representation model by using the verification set and the negative samples;

evaluating the trained knowledge representation model by using the test set and the negative sample;

and mining implicit relations and missing entities of the urban traffic knowledge graph by using the trained knowledge representation model, and supplementing the urban traffic knowledge graph.

4. The construction method according to claim 3, wherein:

the method for constructing the negative sample of the triple data comprises the following steps:

calculating the probability of selecting a head entity or a tail entity to complete replacement operation according to Bernoulli distribution for triples with a certain relation in a training set, a verification set or a test set, and replacing the entities with higher probability;

according to the relationship type constraint, the relationship decides which entities to replace, as shown in the following formula:

where Δ' is the set of constructed negative triplets, d _r An ordered index for all entities within the domain constraint that satisfy the relationship type r; r is _r In order to satisfy the ordered indexes of all entities within the range constraint of the relationship type r, h is a head entity of the triple, h 'is a head entity of the negative triple, t is a tail entity of the triple, t' is a tail entity of the negative triple, and r is a relationship.

5. The construction method according to claim 3, wherein: the specific method for filtering the false negative examples in the negative sample comprises the following steps:

importing the ternary group data and the negative ternary group data in the negative sample into a relational database, finding out repeated data by using a query function in the relational database, and removing the repeated data in the negative sample;

wherein the repeated data is data existing in both the ternary data and the negative ternary data.

6. The construction method according to claim 3, wherein:

the knowledge representation model is a TransD model, and the TransD knowledge representation model is as follows:

mapping a matrix:

wherein M is _rh Mapping matrices, M, for the head entity _rt Mapping matrix, r, for tail entities _p In the form of a relationship vector, the relationship vector,

a vector is mapped for the head entity,

mapping vectors for tail entities, I ^mxn Is an identity matrix;

projecting the entity vector into a relationship space:

h _⊥ ＝M _rh h

t _⊥ ＝M _rt t

h _⊥ is head entity composed of M _rh Mapped head entity vector, t _⊥ Is a tail entity composed of M _rt Mapping the tail entity vector;

the score function:

loss function:

where γ is a hyperparameter, indicating the maximum separation between the correct and negative triplets. [ x ] of] ₊ Max (0, x), Δ represents the set of correct triples, and Δ' represents the set of constructed negative triples.

7. The construction method according to claim 6, wherein:

the specific method for evaluating the trained knowledge representation model by using the test set and the negative samples comprises the following steps:

for any triple in the test set, calculating a triple score and a negative triple score constructed according to the triple and entities in the knowledge graph according to a score function in a trained knowledge inference model, and ranking the triple and the negative triple from large to small according to the score value;

and measuring the completion effect of the link prediction task by adopting one or more evaluation indexes of average ranking, average reciprocal ranking, top hit rate, top three hit rate and top ten hit rate.

8. The construction method according to claim 1, characterized in that:

the urban traffic comprises public traffic and road traffic, and a public traffic knowledge map and a road traffic knowledge map are respectively constructed for the public traffic and the road traffic;

the method for acquiring the public traffic data comprises the steps of acquiring subway line and station information of a target city through a web crawler technology, and acquiring subway card swiping data of the subway line within a target time;

the method for acquiring the road traffic data comprises the steps of acquiring target road network data from a map database, and acquiring target interest point information and traffic situation data on a target road by utilizing a map API.

9. A construction device of an urban traffic knowledge map is characterized in that: the method comprises the following steps:

the body construction module is used for constructing the urban traffic body by using the body construction tool to form a knowledge map mode layer;

the data acquisition module is used for acquiring urban traffic data, extracting traffic entities, attributes and relationships among the entities and constructing a knowledge map data layer;

the storage module is used for combining the knowledge map mode layer with the knowledge map data layer to generate an urban traffic knowledge map and storing the urban traffic knowledge map into a database;

and the reasoning module is used for reasoning new knowledge in the urban traffic knowledge map by using the knowledge representation model and supplementing the urban traffic knowledge map.

10. The application of the urban traffic knowledge base map construction method of any one of claims 1 to 8 in the field of urban traffic.

Images