Personalized recommendation method for passenger carrying of taxi

Abstract

The invention discloses a personalized recommendation method for taxi passenger carrying. The method comprises the following steps: S1, generating a multi-target taxi passenger carrying hot spot area based on big data; s2, constructing a taxi travel knowledge graph; and S3, recommending a personalized cruise path of the taxi based on the knowledge graph and an interest capture recommendation algorithm. The invention provides a taxi passenger carrying personalized recommendation method which comprises the steps of multi-target taxi passenger carrying hotspot generation based on big data, taxi travel knowledge graph construction and personalized recommendation based on a knowledge graph and an interest capture recommendation algorithm. The knowledge graph is used for providing richer context information and semantic association information for the recommendation system, semantic linking is carried out on various information, so that the recommendation accuracy and interpretability are improved, an interest module is introduced, the recommendation system can better understand the requirements and interests of drivers, and the recommendation effect is improved. Fine-grained interests of taxi drivers on different attributes are considered, and more accurate personalized cruise paths are recommended for the taxi drivers according to demand preferences of the taxi drivers.

Description

Technical Field

[0001] The present invention relates to the technical field of taxi passenger transportation, and in particular to a personalized taxi passenger transportation recommendation method. Background Art

[0002] As an important supplement to urban public transportation, taxis can provide customized door-to-door services to the public 24 hours a day, compared with other public transportation modes such as buses and subways with relatively fixed service hours and routes, and have absolute advantages in comfort, convenience and flexibility. In the traditional model, taxis look for customers along the streets. This empirical cruising operation puts drivers and passengers in a state of information isolation. The supply and demand of taxis are mismatched in time and space. The phenomenon of "people looking for cars and cars looking for people" leads to the coexistence of the contradictory problems of difficulty in getting a taxi for passengers and high idle rate of taxis, resulting in reduced taxi operating efficiency and service level, and waste of vehicles and road resources. With the advent of mobile Internet and the wave of sharing economy, traditional taxis in many places have broken the single business model and connected to online car-hailing platforms to promote the reform of the cruising car industry with digitalization. Taxi drivers can make online reservations part-time while cruising to pick up passengers full-time. Online car-hailing mobile APP can It can attract traffic for taxis and help drivers get more travel orders, which has improved the operating efficiency of taxis to a certain extent; however, there are still many problems in the transformation and upgrading of the cruising car industry; although the complementary online and offline business models have enabled taxi resources to be allocated relatively reasonably, this cannot fundamentally change the traditional empty cruising mode of taxis; although a large number of users are now accustomed to using mobile APP to book a taxi, there is still a large demand for traditional telephone and hailing modes. For example, the taxi operation situation in Shanghai shows that APP taxi business only accounts for 30% of the cruising taxi business, and up to 70% of passengers still hail taxis by hailing on the roadside; a large number of taxi drivers still need to rely on experience to cruise in the city to pick up passengers, and the problems of difficulty in hailing a taxi and high empty rate still hinder the development of the cruising taxi market; at present, the issue of how to reduce the empty rate and improve the operating efficiency of taxi drivers has become the key to the development of the taxi industry.

[0003] In order to reduce the idle rate of taxis and improve the operating efficiency of drivers, a lot of research has been done on taxi hotspots and passenger route recommendations. However, most of the existing research recommends hotspot areas for passengers in different time periods through taxi trajectory data mining, without considering the travel time of taxis to hotspot areas and the changes in the attribute values ​​of various influencing factors during the journey; when recommending passenger routes, most of them only consider the shortest path between a fixed starting point and a determined target point, without considering the driver's preferences and interests for different attributes and the real-time traffic conditions of the road section during driving; at the same time, traditional recommendation algorithms are mainly divided into collaborative filtering-based recommendation algorithms, content-based recommendation algorithms and hybrid recommendation algorithms. These algorithms all have data sparsity and cold start problems. Some research on recommendation methods based on knowledge graphs has also proved the advancement of combining knowledge graphs with graph neural networks for recommendation, but these methods are relatively limited in their ability to model users' fine-grained interests in relationships and attributes in knowledge graphs.

[0004] Therefore, further improvements are necessary. Summary of the invention

[0005] The purpose of the present invention is to provide a personalized recommendation method for taxi passengers, which uses knowledge graphs to provide the recommendation system with richer context information and semantic association information, and semantically links various information to improve the accuracy and explainability of the recommendation, so that the recommendation results have a basis to rely on, and enhance the driver's trust in the recommended route; introduce relationship and attribute interest modules so that the recommendation system can better understand the driver's needs and interests, consider the taxi driver's fine-grained interests in different attributes, and recommend more accurate personalized cruising routes for them according to their needs and preferences, increase the operating income of taxi drivers, reduce the waiting time for passengers to travel, and improve service quality.

[0006] A personalized taxi passenger recommendation method designed for this purpose is characterized by comprising the following steps:

[0007] S1. Generate multi-target taxi passenger hotspots based on big data;

[0008] S2, build a knowledge graph of taxi travel;

[0009] S3. Recommend personalized taxi cruising routes based on knowledge graph and interest capture recommendation algorithm.

[0010] In step S1, the steps of generating multi-target taxi passenger hotspots are as follows:

[0011] S11. Data acquisition and fusion: Acquire road network data, vehicle trajectory data, and integrate weather data, holiday data, and urban activity data; establish a real-time data update channel to ensure that the acquired road network data, trajectory data, etc. can be updated in a timely manner;

[0012] S12, Adaptive grid division: Based on factors such as population density and traffic flow in different areas, the road network is divided into several grids using adaptive geohash coding block sizes. The grids are dynamically adjusted regularly according to changes in urban development and traffic flow.

[0013] S13. Passenger hotspot statistics: Based on the order data of taxis and online ride-hailing vehicles, the number of passengers picked up in different grids at different times, the number of empty taxis and online ride-hailing vehicles, the revenue per order, the probability of picking up passengers, the average waiting time of passengers, and the distribution of passengers’ destinations are counted; the number of passengers picked up, the expected revenue, and the probability of picking up passengers for all grids on each road are calculated;

[0014] S14, graph database construction: store the number of passengers carried on each road at different time periods, expected revenue and passenger carrying probability, longitude and latitude of road intersections, geohash blocks, weather data, holiday data, urban activity data and other related information into the graph database, and establish the graph's point, edge, direction, and weight association relationship;

[0015] S15. Passenger hotspot prediction: Use artificial intelligence algorithms to predict passenger hotspot areas under different targets in the next time period.

[0016] In step S12, a finer grid division is used in areas with dense population and heavy traffic; and a relatively larger grid division is used in areas with sparse population and light traffic.

[0017] In step S14, the point corresponds to an intersection, the edge is the road between two intersections, and the weight is the total number of passengers or the probability of passengers or the expected revenue of the road in different time periods.

[0018] In step S2, the steps for constructing the taxi travel knowledge graph are as follows:

[0019] S21. Data acquisition and processing: Acquire basic data in the taxi management business system and dynamic traffic data in the map open platform;

[0020] S22, Entity and Relationship Modeling: Analyze the basic data, obtain entity data and relationship data, and perform node modeling and relationship modeling of the taxi travel knowledge graph;

[0021] S23, knowledge graph fusion: Construct a taxi travel knowledge graph based on the multi-target taxi passenger hotspots generated in step S1. Through deep learning algorithms, explore the potential relationships between different data sources to make the fused knowledge graph more coherent and logical.

[0022] In step S21, the basic data includes vehicle data, driver data, taxi company data, and passenger data; dynamic traffic data includes meter data, location data, trajectory data, and traffic conditions; data types include plain text data, structured data, and semi-structured data; among them, plain text data is in TXT format and is placed in the knowledge graph through information extraction, structured data includes linked data and databases, which are placed in the knowledge graph through graph mapping and D2R conversion respectively, and semi-structured data includes tables, lists, infoboxes and other forms, which are placed in the knowledge graph through wrappers.

[0023] In step S22, the taxi travel knowledge graph nodes include drivers, vehicles, companies, passengers, meter records, pick-up points, alighting points, passenger hot spots, and transportation facilities.

[0024] In step S3, the steps of recommending a personalized taxi cruising route are as follows:

[0025] S31. Relationship interest modeling: Multiple drivers’ potential interests are used to describe the relationship between drivers and passenger hotspots, and each interest is represented as a combination of knowledge graph relationship vectors. The independence modeling module is introduced, and the final driver-passenger hotspot representation is generated by propagating the driver’s interest representation about the relationship and the passenger hotspot representation on the graph convolutional network; a dynamic adjustment mechanism for driver interests is established to adjust the driver’s interest model in real time according to the driver’s actual operating experience and feedback;

[0026] S32, attribute interest modeling: split the initial embeddings of passenger hotspots and drivers, match the corresponding passenger hotspot attribute categories for each part, and then use the similarity prediction function to predict the driver's interest in the attribute nodes in the knowledge graph. Finally, integrate the driver's interest in passenger hotspots with the driver's interest in attributes, obtain high-order neighborhood representations through graph convolutional networks, and generate driver-passenger hotspot representations belonging to attribute interests;

[0027] S33, prediction and personalized recommendation: The driver-passenger hotspot representations obtained by the two models in step S31 and step S32 are spliced ​​separately to obtain the final driver-passenger hotspot representation, and the inner product operation is performed on the two to obtain the similarity score of the two, and then the personalized cruising route of the recommended taxi is obtained.

[0028] In step S31, the driver's potential interests include: shortest distance, shortest travel time, maximum possible profit, and highest probability of carrying passengers.

[0029] In step S32, the attribute categories of passenger hotspots include: transportation hubs, business districts, scenic spots, residential areas, office areas, hospitals, and exhibition halls.

[0030] The present invention proposes a personalized taxi passenger recommendation method based on knowledge graph and interest capture recommendation algorithm, including multi-target taxi passenger hotspot generation based on big data, taxi travel knowledge graph construction, and personalized recommendation based on knowledge graph and interest capture recommendation algorithm. The knowledge graph is used to provide the recommendation system with richer context information and semantic association information, and various information are semantically linked to improve the accuracy and interpretability of the recommendation. The interest module is introduced to enable the recommendation system to better understand the needs and interests of drivers, consider the fine-grained interests of taxi drivers in different attributes, and recommend more accurate personalized cruising routes for them according to their needs and preferences, reduce the taxi vacancy rate, improve the driver's operating efficiency, thereby increasing the operating income of taxi drivers, reducing the waiting time for passengers to travel, and improving service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention is a flowchart of generating multi-target taxi passenger hotspots in one embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of a taxi travel knowledge graph in one embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0034] See also Figure 1-Figure 2 , the personalized taxi passenger recommendation method comprises the following steps:

[0035] S1. Generate multi-target taxi passenger hotspots based on big data;

[0036] S2, build a knowledge graph of taxi travel;

[0037] S3. Recommend personalized taxi cruising routes based on knowledge graph and interest capture recommendation algorithm.

[0038] In step S1, the steps of generating multi-target taxi passenger hotspots are as follows:

[0039] S11. Data acquisition and fusion: Acquire road network data (such as OSM road data), vehicle trajectory data (taxis and online ride-hailing vehicle trajectory data), and integrate weather data, holiday data, and urban activity data (such as concerts, sports events, etc.). Weather conditions may affect passengers' travel needs. For example, the demand for taxis may increase on rainy days. During holidays and urban activities, the passenger flow in certain areas will change significantly. The integration of these data will make hotspot generation more accurate. Establish a real-time data update channel to ensure that the acquired road network data, trajectory data, etc. can be updated in a timely manner. For example, road construction information may cause congestion on some sections of the road. Timely acquisition and processing of such information can avoid recommending taxis to go to congested areas and improve the accuracy of hotspot area prediction.

[0040] S12, Adaptive grid division: Based on factors such as population density and traffic flow in different areas, the adaptive geohash coding block size is used for grid division. Based on the geohash principle, the road network is divided into several grids G = {g1, g2, ..., g n}; and dynamically adjust the grid regularly according to the changes in urban development and traffic flow; for example, after a new commercial or residential area is built, the grid division of the corresponding area should be optimized to adapt to the new distribution of passenger hotspots;

[0041] S13. Passenger hotspot statistics: Based on the order data of taxis and online ride-hailing vehicles, the number of passengers picked up in different grids at different times (when the number of passengers picked up ≥ K, the grid is regarded as a passenger hotspot), the number of empty taxis and online ride-hailing vehicles, the revenue per order, the probability of picking up passengers, the average waiting time of passengers, and the distribution of passenger destinations are counted; the number of passengers picked up, the expected revenue, and the probability of picking up passengers are calculated for all grids on each road;

[0042] S14, graph database construction: store the number of passengers carried on each road at different time periods, expected revenue and passenger carrying probability, longitude and latitude of road intersections, geohash blocks, weather data, holiday data, urban activity data and other related information into the graph database, and establish the graph's point, edge, direction, and weight association relationship;

[0043] S15. Passenger hotspot prediction: Use artificial intelligence algorithms to predict passenger hotspot areas under different objectives in the next time period; obtain the latitude and longitude of the first empty taxi at the current moment, obtain the driving direction of the taxi, and obtain a certain intersection to be reached; generate hotspot areas under different objectives, such as the hotspot area closest to the taxi driver, the hotspot area with the shortest driving time, the hotspot area with the greatest possible profit, the hotspot area with the highest probability of carrying passengers, etc.

[0044] In step S12, in densely populated and busy traffic areas, a finer grid division (such as a 10-bit geohash coding block) is used to more accurately count passenger information; in sparsely populated and light traffic areas, a relatively larger grid division (such as a 6-bit geohash coding block) is used to reduce the amount of calculation.

[0045] In step S14, the point (geohash block) corresponds to a certain intersection, the edge is the road between two intersections, and the weight is the total number of passengers or the probability of passengers or the expected revenue of the road in different time periods.

[0046] In step S2, the steps for constructing the taxi travel knowledge graph are as follows:

[0047] S21. Data acquisition and processing: Acquire basic data in the taxi management business system and dynamic traffic data in the map open platform;

[0048] S22, Entity and Relationship Modeling: Analyze the basic data, obtain entity data and relationship data, and perform node modeling and relationship modeling of the taxi travel knowledge graph;

[0049] S23, knowledge graph fusion: Construct a taxi travel knowledge graph based on the multi-target taxi passenger hotspots generated in step S1. Through deep learning algorithms, explore the potential relationships between different data sources to make the fused knowledge graph more coherent and logical.

[0050] In step S21, the basic data includes vehicle data, driver data, taxi enterprise data, and passenger data; dynamic traffic data includes meter data, location data, trajectory data, and traffic conditions; data types include plain text data, structured data, and semi-structured data; among them, plain text data is in TXT format and is placed in the knowledge graph through information extraction, structured data includes link data and database, which are placed in the knowledge graph through graph mapping and D2R conversion respectively, and semi-structured data includes tables, lists, infoboxes and other forms, which are placed in the knowledge graph through wrappers; the data of the taxi travel knowledge graph mainly comes from structured data (such as person / vehicle / enterprise information, traffic event statistics, etc. stored in relational databases), and also has a small amount of unstructured data (traffic-related event descriptions), so its knowledge extraction needs to use different extraction methods for two different data formats respectively, and finally organically integrate them to form a unified knowledge structure.

[0051] In step S22, the taxi travel knowledge graph nodes include drivers, vehicles, companies, passengers, meter records, boarding points, alighting points, passenger hot spots, and transportation facilities (such as bus stops, subway stations, etc.); relationships between nodes are established, such as driver-employed-company, driver-driving-vehicle, passenger-riding-vehicle, etc., as well as the distance relationship between bus stops and nearby passenger hot spots, the association relationship between subway stations and passenger destinations, etc.

[0052] In step S3, the steps of recommending a personalized taxi cruising route are as follows:

[0053] S31. Relationship interest modeling: Multiple potential interests of drivers are used to describe the relationship between driver-passenger hotspots, and each interest is represented as a combination of knowledge graph relationship vectors. At the same time, in order to make different interests independent of each other, an independence modeling module is introduced. In order to integrate the driver's interest in different relationships in the knowledge graph into the final driver representation, the final driver-passenger hotspot representation is generated by propagating the driver's interest representation on the relationship and the passenger hotspot representation on the graph convolutional network; the details are as follows:

[0054] Assume there is a set of m drivers represented by U = {u1,u2,…,u m}, the set of n passenger hotspots is represented by I = {i1,i2,…,i n}, then the driver-passenger hotspot interaction graph is represented as O = {(u,i)|u∈U,i∈I}; let ε be the set of entities and R be the set of relations, then the knowledge graph can be represented as g = {(h,r,t)|h∈ε,r∈R,t∈ε}, where each (h,r,t) is a specific triple;

[0055] Assuming p is a set of driver interests, the embedding vector of a driver’s interest is expressed as where e r is the embedding vector of relation r. In order to measure the importance of different relations, each relation uses a multi-layer attention mechanism to generate a corresponding attention score: f(r,p)=h T ReLU(W(r⊙p)+b), where W represents the weight matrix, b is a bias term to prevent overfitting, and h is T is the weight vector, ⊙ is the XOR logical operator; then use the softmax function to normalize f(r,p):

[0056]

[0057] Using N u ={(p,i)|(u,p,i)∈C} represents the interest distribution and the surrounding neighborhood passenger hotspots. By integrating the interest distribution and the surrounding neighborhood passenger hotspots, the driver U can be expressed as: where fu () is an aggregation function used to generate driver representations based on the driver-interest-passenger hotspot triples (u, p, i), introducing an attention mechanism for the importance of different interests to the driver, f u () can be implemented as:

[0058]

[0059] Similarly, using N i ={(r,v)|(i,r,v)∈g} represents the attributes and neighborhood information of passenger hotspot i, then passenger hotspot i can be expressed as: where f KG () is an aggregation function that extracts and aggregates information from each triple (r,i,v), and f u () can be implemented as:

[0060]

[0061] in, The (l-1)th layer represents the representation of drivers, passenger hotspots, and attributes respectively. The embedding representations of the lth layer are accumulated to obtain the final embedding representations of drivers and passenger hotspots:

[0062] A dynamic adjustment mechanism for drivers' interests should be established to adjust the drivers' interest model in real time according to their actual operating experience and feedback. For example, if a driver obtains good returns in a certain area many times, then his interest weight in that area should be increased accordingly. If a driver encounters more problems when carrying a certain type of passengers, then his interest weight in that type of passengers should be reduced accordingly.

[0063] S32. Attribute interest modeling: Design an attention mechanism to model the driver's interest in different attributes; split the initial embedding of the passenger hotspot and the driver, match the corresponding passenger hotspot attribute category for each part, and then use the similarity prediction function to predict the driver's interest in the attribute node in the knowledge graph. Finally, integrate the driver's interest in the passenger hotspot with the driver's interest in the attribute, obtain the high-order neighborhood representation through the graph convolutional network, and generate the driver-passenger hotspot representation belonging to the attribute interest; the details are as follows:

[0064] Treat each relation as an attribute space and split each entity embedding into the corresponding attribute space dimensions:

[0065]

[0066] in, Represents entity i in the corresponding attribute space r m The embedding in d m is the dimension of the embedding vector;

[0067] The passenger hotspot entity embedding at level l is recursively represented as:

[0068]

[0069] To avoid the high-order neighbor noise problem, we re-examine entity embedding from the perspective of component position. Without loss of generality, The structure is: in, yes The truncated vector of , || represents the concatenation operation;

[0070] Design an attention mechanism to model the driver’s interest in different attributes, for each pair of driver u and attribute r m Assign an interest score and generate a driver representation by combining the interest score and interaction pick-up hotspots;

[0071] Refer to the above for driver embedding u To truncate: in, yes The truncated vector of driver u for attribute r m The interest scores are:

[0072]

[0073] Among them, τ is the temperature coefficient of the hyperparameter. If driver u pays great attention to attribute r when choosing m , then the truncated representations corresponding to the passenger hotspots that the driver is interested in and the passenger hotspots that the driver is not interested in will be completely different; σ() is the sigmoid function;

[0074] The embedding representation of driver u is generated by combining the interest score and the interactive passenger hotspot:

[0075]

[0076] S33, prediction and personalized recommendation: The driver-passenger hotspot representations obtained by the two models in step S31 and step S32 are respectively spliced ​​to obtain the final driver-passenger hotspot representation, and the inner product operation is performed on the two to obtain the similarity score of the two, and then the personalized cruising route of the recommended taxi is obtained;

[0077] The inner product operation is expressed as:

[0078]

[0079] in, For the final driver, The ultimate passenger hotspot indicates that is the predicted matching score.

[0080] In step S31, the driver's potential interests include: shortest distance, shortest travel time, maximum possible profit, and highest probability of carrying passengers.

[0081] In step S32, the attribute categories of passenger hotspots include: transportation hubs, business districts, scenic spots, residential areas, office areas, hospitals, and exhibition halls.

[0082] This taxi passenger personalized recommendation method is different from the existing technology that mostly recommends passenger hot spots in different time periods through mining taxi historical trajectory data, without considering the travel time of taxis to hot spots and the changes in the attribute values ​​of various influencing factors during the journey, and does not fully explore the potential needs and interest preferences of passengers; therefore, this taxi passenger personalized recommendation method generates hot spots under different goals, and constructs attributes of interest to drivers considering various changes in the process of traveling to hot spots, such as the hot spots closest to taxi drivers, the hot spots with the shortest travel time, the hot spots with the greatest possible benefits, the hot spots with the highest probability of carrying passengers, etc.

[0083] The personalized taxi passenger recommendation method is different from the existing technology in that most of them only consider the shortest path between a fixed starting point and a determined target point when recommending passenger routes, and do not consider the driver's preferences and interests for different attributes and the real-time traffic conditions of the road section during driving; at the same time, traditional recommendation algorithms are mainly divided into collaborative filtering-based recommendation algorithms, content-based recommendation algorithms and hybrid recommendation algorithms, all of which have data sparsity and cold start problems. Some recommendation methods based on knowledge graphs have also proven the advancement of combining knowledge graphs with graph neural networks for recommendation, but these methods are relatively limited in their ability to model users' fine-grained interests in relationships and attributes in knowledge graphs; for this reason, a personalized taxi passenger recommendation method based on knowledge graphs and interest capture recommendation algorithms is proposed, which uses knowledge graphs to provide the recommendation system with richer contextual information and semantic association information, and semantically links various information to improve the accuracy and interpretability of recommendations, introduces relationship and attribute interest modules, enables the recommendation system to better understand the needs and interests of drivers, considers the fine-grained interests of taxi drivers in different attributes, and recommends more accurate personalized cruising routes for them according to their needs and preferences.

[0084] The above is a preferred embodiment of the present invention, which shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for illustrating the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A personalized taxi passenger recommendation method, characterized by: The following steps are involved: S1. Generate multi-target taxi passenger hotspots based on big data; S2, build a knowledge graph of taxi travel; S3. Recommend personalized taxi cruising routes based on knowledge graph and interest capture recommendation algorithm.

2. The method for personalized taxi passenger recommendation according to claim 1, characterized in that: In step S1, the steps of generating multi-target taxi passenger hotspots are as follows: S11. Data acquisition and fusion: Acquire road network data, vehicle trajectory data, and integrate weather data, holiday data, and urban activity data; establish a real-time data update channel to ensure that the acquired road network data, trajectory data, etc. can be updated in a timely manner; S12, Adaptive grid division: Based on factors such as population density and traffic flow in different areas, the road network is divided into several grids using adaptive geohash coding block sizes. The grids are dynamically adjusted regularly according to changes in urban development and traffic flow. S13. Passenger hotspot statistics: Based on the order data of taxis and online ride-hailing vehicles, the number of passengers picked up in different grids at different times, the number of empty taxis and online ride-hailing vehicles, the revenue per order, the probability of picking up passengers, the average waiting time of passengers, and the distribution of passengers’ destinations are counted; the number of passengers picked up, the expected revenue, and the probability of picking up passengers for all grids on each road are calculated; S14, graph database construction: store the number of passengers carried on each road at different time periods, expected revenue and passenger carrying probability, longitude and latitude of road intersections, geohash blocks, weather data, holiday data, urban activity data and other related information into the graph database, and establish the graph's point, edge, direction, and weight association relationship; S15. Passenger hotspot prediction: Use artificial intelligence algorithms to predict passenger hotspot areas under different targets in the next time period.

3. The method for personalized taxi passenger recommendation according to claim 2, characterized in that: In step S12, a finer grid division is used in areas with dense population and heavy traffic; and a relatively larger grid division is used in areas with sparse population and light traffic.

4. The method for personalized taxi passenger recommendation according to claim 2, characterized in that: In step S14, the point corresponds to an intersection, the edge is the road between two intersections, and the weight is the total number of passengers or the probability of passengers or the expected revenue of the road in different time periods.

5. The method for personalized taxi passenger recommendation according to claim 1, characterized in that: In step S2, the steps for constructing the taxi travel knowledge graph are as follows: S21. Data acquisition and processing: Acquire basic data in the taxi management business system and dynamic traffic data in the map open platform; S22, Entity and Relationship Modeling: Analyze the basic data, obtain entity data and relationship data, and perform node modeling and relationship modeling of the taxi travel knowledge graph; S23, knowledge graph fusion: Construct a taxi travel knowledge graph based on the multi-target taxi passenger hotspots generated in step S1. Through deep learning algorithms, explore the potential relationships between different data sources to make the fused knowledge graph more coherent and logical.

6. The method for personalized taxi passenger recommendation according to claim 5, characterized in that: In step S21, the basic data includes vehicle data, driver data, taxi company data, and passenger data; dynamic traffic data includes meter data, location data, trajectory data, and traffic conditions; data types include plain text data, structured data, and semi-structured data; among them, plain text data is in TXT format and is placed in the knowledge graph through information extraction, structured data includes linked data and databases, which are placed in the knowledge graph through graph mapping and D2R conversion respectively, and semi-structured data includes tables, lists, infoboxes and other forms, which are placed in the knowledge graph through wrappers.

7. The method for personalized taxi passenger recommendation according to claim 5, characterized in that: In step S22, the taxi travel knowledge graph nodes include drivers, vehicles, companies, passengers, meter records, pick-up points, alighting points, passenger hot spots, and transportation facilities.

8. The method for personalized taxi passenger recommendation according to claim 1, characterized in that: In step S3, the steps of recommending a personalized taxi cruising route are as follows: S31. Relationship interest modeling: Multiple drivers’ potential interests are used to describe the relationship between drivers and passenger hotspots, and each interest is represented as a combination of knowledge graph relationship vectors. The independence modeling module is introduced, and the final driver-passenger hotspot representation is generated by propagating the driver’s interest representation about the relationship and the passenger hotspot representation on the graph convolutional network; a dynamic adjustment mechanism for driver interests is established to adjust the driver’s interest model in real time according to the driver’s actual operating experience and feedback; S32, attribute interest modeling: split the initial embeddings of passenger hotspots and drivers, match the corresponding passenger hotspot attribute categories for each part, and then use the similarity prediction function to predict the driver's interest in the attribute nodes in the knowledge graph. Finally, integrate the driver's interest in passenger hotspots with the driver's interest in attributes, obtain high-order neighborhood representations through graph convolutional networks, and generate driver-passenger hotspot representations belonging to attribute interests; S33, prediction and personalized recommendation: The driver-passenger hotspot representations obtained by the two models in step S31 and step S32 are spliced ​​separately to obtain the final driver-passenger hotspot representation, and the inner product operation is performed on the two to obtain the similarity score of the two, and then the personalized cruising route of the recommended taxi is obtained.

9. The method for personalized taxi passenger recommendation according to claim 8, characterized in that: In step S31, the driver's potential interests include: shortest distance, shortest travel time, maximum possible profit, and highest probability of carrying passengers.

10. The method for personalized taxi passenger recommendation according to claim 8, characterized in that: In step S32, the attribute categories of passenger hotspots include: transportation hubs, business districts, scenic spots, residential areas, office areas, hospitals, and exhibition halls.

Images