Driver group mining method and system based on graph model

Abstract

The invention discloses a driver group mining method and system based on a graph model. The invention relates to the technical field of online car hailing. Driver node information (including ID, age, gender, driving experience and vehicle type), work history (order receiving time, place and service type) of drivers, direct relation data (relation between friends and colleagues) between the drivers and travel history geographic position data are collected from service logs of an online car-hailing platform, user feedback and interaction records of the drivers and passengers to form a data set D; according to the collected data set D between the drivers, a driver relation graph G = (V, E) is constructed, and V is a set of driver nodes; by constructing the driver relation graph and training the embedded representation of the driver nodes by using the machine learning technology, the method can capture the potential relation and similarity between the drivers, thereby realizing the accurate division of the driver group. And the online car-hailing platform can better understand the characteristics and behavior modes of the driver group.

Description

Technical Field

[0001] The present invention relates to the field of online car-hailing technology, specifically to mining online car-hailing driver groups, and more particularly to a driver group mining method and system based on a graph model. Background Art

[0002] With the growing popularity of online ride-hailing services, driver group mining aims to identify groups of drivers with similar characteristics, behavioral patterns, or experience by analyzing various driver data, such as basic information, driving behavior, and service records. This not only helps platforms better understand the characteristics of driver groups but also provides strong support for personalized services, optimized dispatch efficiency, improved service quality, and targeted marketing and customer service.

[0003] Traditional driver group mining techniques primarily rely on technologies and methods from fields such as data mining, machine learning, statistical analysis, and network science. For example, these techniques utilize driver behavior data, service records, and personal information. For example, CN201810179464.1 discloses a hotspot mining method based on road network clustering (published on August 21, 2018), which uses a clustering algorithm to divide drivers into different groups.

[0004] Alternatively, classification algorithms can be used, such as the method and device for freight insurance user classification based on behavioral data disclosed in CN202311550647.7 (published on February 20, 2024), to predict which group a driver belongs to based on their historical data. A model can be trained using the driver's historical service records to predict the likely service behavior patterns of new drivers.

[0005] However, the definition of a community using traditional technologies can be ambiguous, and different algorithms may define communities based on different criteria. This makes the results between algorithms difficult to compare and understand, reducing the accuracy and reliability of driver group mining.

[0006] To this end, the present invention proposes a driver group mining method and system based on a graph model. Summary of the Invention

[0007] In view of this, the present invention aims to provide a method and system for mining driver groups based on a graph model to solve or alleviate the technical problem existing in the prior art, namely, how to resolve the ambiguity of community definition, and to provide at least a beneficial option for this problem. The technical solution of the present invention is implemented as follows:

[0008] First, the driver group mining method based on the graph model:

[0009] (1) Overview:

[0010] The present invention aims to use graph models and machine learning techniques to mine the driver groups on the online car-hailing platform. The solution first constructs a driver relationship graph by collecting multi-dimensional data of drivers, including basic information, driving behavior, service records, direct relationship data, and geographic location information. On this basis, an improved DeepWalk algorithm is adopted to map the nodes in the graph (i.e., drivers) into a continuous vector space to capture the similarity between nodes and network structure characteristics. By generating node sequences through random walks and training node embeddings using the frequency hopping mechanism of Word2Vec, the model can learn the potential relationships between drivers. Finally, the trained model is used to calculate the similarity between driver nodes, and the drivers are divided into different groups based on the similarity. These groups reflect the similarity of drivers in terms of characteristics, behavior patterns or experience.

[0011] (2) Technical solution:

[0012] To achieve the above technical objectives, the present invention chooses to perform the following operating steps.

[0013] 2.1 Step S1, data collection:

[0014] We collect driver node information (including ID, age, gender, driving experience, and vehicle type), driver work history (order time, location, and service type), direct relationship data between drivers (friends and colleagues), and travel history geographic location data from the ride-hailing platform’s service logs, user feedback, and interaction records between drivers and passengers to form Dataset D.

[0015] Among them, data set D = (DriverNodeInfo, DriverWorkHistory, DriverRelationships, DriverGeoLocationHistory);

[0016] DriverNodeInfo is the driver node information, DriverWorkHistory is the driver's work history; DriverRelationships is a two-dimensional matrix, which is the direct relationship data between drivers. The element R[i][j] in the matrix represents the relationship between driver i and driver j (1 indicates the existence of a relationship, 0 indicates the absence of a relationship); DriverGeoLocationHistory is the travel history geographic location data.

[0017] 2.2 Step S2, build graph structure:

[0018] Based on the collected data set D between drivers, a graphical model of the driver relationship graph G = (V, E) is constructed, where V is the set of driver nodes and E is the set of edges; the weight of the edge is set according to the closeness of the number of passengers served in common.

[0019] 2.2.1 Step S200, initialize the driver node set V:

[0020] Extract all drivers’ IDs from the dataset D and initialize the driver node set V.

[0021] V=ExtractUnique(D.DriverNodeInfo,ID)

[0022] The ExtractUnique function is used to extract unique values, D.DriverNodeInfo represents the driver node information part in the dataset D, and ID represents the field name to be extracted.

[0023] 2.2.2 Step S201, initialize edge set E:

[0024] Create an empty set E to store edges between drivers.

[0025] 2.2.3 Step S202, traverse the data set D and construct edges and their weights:

[0026] Traverse the direct relationship data and travel history geographic location data between drivers in the dataset D; for each pair of drivers with a direct relationship, or a pair of drivers who have served passengers together, add an edge to the edge set E and set the weight of the edge according to the number of passengers served together.

[0027] The edge weight is calculated as:

[0028] weight ij =∑ k∈CoServicePassengers(i,j) 1;

[0029] Among them, weight ij represents the weight of the edge between driver i and driver j, and CoServicePassengers}(i,j) represents the set of passengers served by driver i and driver j.

[0030] 2.2.4 Step S203: Constructing the driver relationship graph G:

[0031] Use the driver node set V and edge set E to construct the driver relationship graph G = (V, E).

[0032] 2.3 Model training method:

[0033] P1. Set the model parameters, including the dimension of the embedding vector, the number of steps of the random walk, and the number of negative samplings.

[0034] P2. Initialize the embedding vector with random numbers as the initial weights of the model.

[0035] P3. Generate sequences by random walk: Perform a random walk on the driver relationship graph to generate a series of node sequences. Each sequence can be regarded as a "sentence", and the nodes in it can be regarded as "words".

[0036] P4. Model training: Use the skip-gram mechanism of Word2Vec to train the node embeddings. By continuously adjusting the weights of the embedding vector, minimize the loss function, that is, the similarity difference between the neighbor nodes predicted by the model and the true neighbor nodes.

[0037] P5. Model optimization: Adjust the model parameters according to the loss function value during training and the performance on the validation set. Use the Adam optimizer to accelerate the training process and reduce overfitting.

[0038] 2.4 Step S3, Driver group mining:

[0039] Use the trained model to calculate the similarity between driver nodes. Divide the drivers into different groups according to the similarity.

[0040] Divide the drivers into different groups according to the similarity, and each group has similar characteristics or behavior patterns.

[0041] 2.4.1 Step S300, Calculate the similarity between driver nodes:

[0042] Use the trained model to obtain the embedding vector of each driver node. These vectors represent the characteristics and behavior patterns of the drivers in the embedding space;

[0043] For each pair of driver nodes, calculate the cosine similarity between their embedding vectors:

[0044]

[0045] where, e u and e v represent the embedding vectors of driver nodes u and v respectively, · represents the dot product of vectors, |e u | and |e v | represent the norms of vectors e_u and e_v respectively.

[0046] Divide the drivers into different groups according to the similarity; when the similarity score between two driver nodes is higher than the preset threshold, consider them to belong to the same group.

[0047] 2.4.2 Step S301, group division:

[0048] Traverse all driver node pairs and divide the drivers into different groups based on the similarity score and the set threshold. The driver nodes in each group have similar characteristics or behavior patterns.

[0049] Based on the driver invitation relationship, a multi-level relationship graph of drivers is constructed. The lower the order, the closer the relationship between the drivers. We sample in the graph, select 3 drivers A, B, and C each time, and calibrate the relationship order p of AB driver pair and AC driver pair AB and p AC Based on the driver vector, calculate the correlation s between drivers A and B, and A and C AB 、s AC , and compare. If one of the following conditions is met, the estimate is correct:

[0050] p AB <p AC , and s AB >s AC ;

[0051] p AB >p AC , and s AB AC ;

[0052] (3) Mechanisms for resolving technical issues:

[0053] 3.1 Solution mechanism:

[0054] Collect multi-dimensional data of drivers from online ride-hailing platforms, including ID, age, gender, driving experience, vehicle type, order acceptance time, location, service type, relationships with friends and colleagues, and travel history location data.

[0055] Using the collected data, we constructed a driver relationship graph, where nodes represent drivers and edges represent relationships between drivers. We applied Word2Vec's frequency hopping mechanism to the driver relationship graph, generating node sequences through random walks and training embedding representations for the driver nodes. The embeddings capture the underlying relationships and similarities between drivers, ensuring that drivers with similar distances in the vector space share similar characteristics or behavioral patterns.

[0056] The trained model is used to calculate the similarity between driver nodes and classify them into different groups based on similarity. These groups have similar characteristics or behavior patterns, thus resolving the ambiguity in the definition of traditional technical communities.

[0057] 3.2 Principle Description:

[0058] ​Graphs are powerful data structures that can represent networks of relationships between nodes. By mapping drivers into a continuous vector space, embeddings capture the underlying relationships and similarities between drivers. This representation ensures that drivers who are close in distance in the vector space share similar characteristics or behavioral patterns, facilitating group segmentation and community definition.

[0059] Word2Vec is a model used to train word embeddings. In this solution, it is used to train embedding representations for driver nodes. The frequency hopping mechanism learns node embedding representations by predicting the neighboring nodes of a central node, thereby capturing the contextual relationships between nodes. In the driver relationship graph, each driver node has certain connections with its neighboring nodes. These connections are converted into similarity information in the embedding vector through the frequency hopping mechanism.

[0060] 3.3 The ambiguity of community definition:

[0061] Traditional technical community definitions are difficult to clearly define and divide due to fuzzy community boundaries and diverse member characteristics. This solution automatically divides driver groups and defines communities by constructing a driver relationship graph and using machine learning to train an embedding representation of driver nodes.

[0062] Because the embedding vector can capture the potential relationships and similarities between drivers, the divided driver groups have a high degree of consistency in characteristics or behavior patterns, thus solving the ambiguity problem in the definition of traditional technical communities.

[0063] Secondly, the driver group mining system based on the graph model:

[0064] like Figure 3 As shown, the system is used to implement the driver group mining method based on the graph model described above, which includes:

[0065] (1) A data collection module that collects multi-dimensional data on drivers from ride-hailing platforms: including basic information, driving behavior, service records, direct relationship data, and geographic location information. The data comes from the ride-hailing platform’s service logs, user feedback, and driver-passenger interaction records.

[0066] (2) Model training module for building a driver relationship graph: This module uses the frequency hopping mechanism of Word2Vec to train the embedding representation of driver nodes. This includes driver relationship graph construction, random walk, frequency hopping model (Skip-Gram), and negative sampling.

[0067] (3) A group division module that uses the trained model and embedding vector to calculate the similarity between driver nodes: drivers are divided into different groups based on similarity.

[0068] (4) An application module that provides personalized services to the ride-hailing platform based on the group segmentation results, including feedback to managers to optimize scheduling efficiency.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. Improving the Accuracy of Driver Group Segmentation: By constructing a driver relationship graph and using machine learning to train the embedding representation of driver nodes, this invention can capture the underlying relationships and similarities between drivers, thereby accurately segmenting driver groups. This helps ride-hailing platforms better understand the characteristics and behavioral patterns of driver groups.

[0071] Optimizing Personalized Services: Based on the driver group breakdown, ride-hailing platforms can provide personalized services to different driver groups. For example, they can offer better order matching services to those with high efficiency in accepting orders, and more precise dispatch strategies to those serving specific time periods or regions. This helps improve driver satisfaction and loyalty.

[0072] 3. Improving operational efficiency: Through in-depth analysis of the driver base, ride-hailing platforms can identify operational bottlenecks and problems and implement corresponding optimization measures. For example, to address driver shortages in certain areas or service periods, they can adjust dispatch strategies or add incentive mechanisms to attract more drivers to the service. This helps improve the platform's operational efficiency and service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0074] Figure 1 Schematic diagram of the method flow of the present invention;

[0075] Figure 2 Schematic diagram of the model execution method of the present invention;

[0076] Figure 3 Schematic diagram of the system composition of the present invention. DETAILED DESCRIPTION

[0077] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0078] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0079] Explanation of relevant terms:

[0080] (1) ID: The driver’s unique identifier, used to distinguish different drivers.

[0081] (2) Age: The driver’s age data, reflecting the driver’s age characteristics.

[0082] (3) Gender: The driver’s gender data, indicating the driver’s gender characteristics.

[0083] (4) Driving experience: The length of time a driver has been engaged in driving work, which reflects the driver's driving proficiency.

[0084] (5) Vehicle type: The type of vehicle driven by the driver, reflecting the driver's vehicle preference or work needs.

[0085] (6) Order acceptance time, location and service type: records the driver's order acceptance activities, including the specific time and location of the order acceptance and the type of service provided (such as express, private car, etc.), reflecting the driver's working hours and service scope.

[0086] (7) Friends and colleagues: Social or work relationships between drivers, indicating close connections between drivers.

[0087] (8) Travel history geographic location data: records the geographic location information of the driver's acceptance of orders and travel, reflecting the driver's service scope and geographical distribution.

[0088] (9) Graph structure: A data structure used to represent the relationship network between drivers. Nodes represent drivers and edges represent the relationships between drivers.

[0089] (10) Word2Vec’s frequency hopping mechanism: A model for training word vectors (here, driver vectors) that learns the embedded representation of nodes by predicting the neighboring nodes of the central node.

[0090] (11) Similarity: A measure of the similarity between two drivers, usually based on the distance or similarity of the driver vectors.

[0091] (12) Group: A collection of drivers with similar characteristics or behavior patterns. By dividing the groups, we can better understand the characteristics of the driver group.

[0092] (13) Characteristics or behavioral patterns: indicators or patterns that describe the characteristics of individual or group drivers, such as driving habits, service preferences, etc.

[0093] Example 1: Figures 1-2 As shown, this embodiment discloses a driver group mining method based on a graph model. To improve service quality and optimize scheduling strategies, online ride-hailing platforms need to conduct group mining on drivers to identify groups of drivers with similar characteristics, behavior patterns, or experience. This solution collects multidimensional data on drivers, constructs a driver relationship graph, and uses an improved DeepWalk algorithm to train a model, ultimately achieving driver group mining. The method includes:

[0094] In this embodiment, regarding step S1, data collection:

[0095] We collect driver node information (including ID, age, gender, driving experience, and vehicle type), driver work history (order time, location, and service type), direct relationship data between drivers (friends and colleagues), and travel history geographic location data from the ride-hailing platform’s service logs, user feedback, and interaction records between drivers and passengers to form Dataset D.

[0096] (1) Extract the driver’s ID, age, gender, driving experience, and vehicle model information from the service log of the online car-hailing platform to form DriverNodeInfo.

[0097] Example: DriverNodeInfo=[{ID:1,Age:30,Gender:'Male',DrivingExperience:5,VehicleType:'SUV'},...]

[0098] (2) Collect the driver's work history:

[0099] The driver's order acceptance time, location, and service type are extracted from the service log to form DriverWorkHistory.

[0100] Example: DriverWorkHistory=[{ID:1,Orders:[{Time:'2023-01-01',Location:'CityA',ServiceType:'Express'},...]},...]

[0101] (3) Collecting direct relationship data between drivers:

[0102] Information such as friends and colleagues between drivers is extracted from user feedback and driver interaction records to form a two-dimensional matrix of DriverRelationships.

[0103] Example: DriverRelationships = [[0,1,0],[1,0,1],[0,1,0]] (indicates that Driver 1 has a direct relationship with Driver 2, and Driver 2 has a direct relationship with Driver 3)

[0104] (4) Collect travel history location data:

[0105] Extract the driver's travel history geographic location data from the service log to form

[0106] DriverGeoLocationHistory.

[0107] Example: DriverGeoLocationHistory = [{ID:1,Locations:[{Latitude:34.05,Longitude:-118.25},...]},...]

[0108] Among them, data set D = (DriverNodeInfo, DriverWorkHistory, DriverRelationships, DriverGeoLocationHistory);

[0109] DriverNodeInfo is the driver node information, DriverWorkHistory is the driver's work history; DriverRelationships is a two-dimensional matrix, which is the direct relationship data between drivers. The element R[i][j] in the matrix represents the relationship between driver i and driver j (1 indicates the existence of a relationship, 0 indicates the absence of a relationship); DriverGeoLocationHistory is the travel history geographic location data.

[0110] In this embodiment, regarding step S2, a graph structure is constructed:

[0111] Based on the collected data set D between drivers, a graphical model of the driver relationship graph G = (V, E) is constructed, where V is the set of driver nodes and E is the set of edges; the weight of the edge is set according to the closeness of the number of passengers served in common.

[0112] Step S200, initialize the driver node set V:

[0113] Extract all driver IDs from DriverNodeInfo and initialize the driver node set V.

[0114] V=ExtractUnique(D.DriverNodeInfo,ID);

[0115] V = [1, 2, 3, ...];

[0116] The ExtractUnique function is used to extract unique values, D.DriverNodeInfo represents the driver node information in dataset D, and ID represents the field name to be extracted.

[0117] Step S201, initialize the edge set E:

[0118] Create an empty set E to store edges between drivers.

[0119] Step S202: traverse the data set D and construct edges and their weights:

[0120] Traverse DriverRelationships and for each pair of drivers that have a direct relationship, add an edge to E with a weight of 1. The weight of the edge is calculated as:

[0121] weight ij =∑ k∈CoServicePassengers(i,j) 1;

[0122] Among them, weight ij represents the weight of the edge between driver i and driver j, and CoServicePassengers}(i,j) represents the set of passengers served by driver i and driver j.

[0123] For example: If driver 1 and driver 2 serve 5 passengers together, then weight_{1,2}=5

[0124] Step S203 , constructing a driver relationship graph G: using the driver node set V and the edge set E to construct a driver relationship graph G=(V, E).

[0125] In this embodiment, regarding step S3, driver group mining: Use the trained model to calculate the similarity between driver nodes. Divide the drivers into different groups according to the similarity. Divide the drivers into different groups according to the similarity, and each group has similar characteristics or behavior patterns.

[0126] Specifically, in step S300, calculate the similarity between driver nodes: Use the trained model to obtain the embedding vectors of each driver node. These vectors represent the characteristics and behavior patterns of the drivers in the embedding space;

[0127] For each pair of driver nodes, calculate the cosine similarity between their embedding vectors:

[0128]

[0129] where, e u and e v respectively represent the embedding vectors of driver nodes u and v, · represents the dot product of vectors, |e u | and |e v | respectively represent the norms of vectors e_u and e_v. 1]

[0130] Divide the drivers into different groups according to the similarity; when the similarity score between two driver nodes is higher than the preset threshold, it is considered that they belong to the same group.

[0131] Specifically, in step S301, divide groups: Traverse all pairs of driver nodes, and divide the drivers into different groups according to the similarity score and the set threshold. The driver nodes within each group have similar characteristics or behavior patterns.

[0132] Construct a multi-order relationship graph of drivers based on the driver invitation relationship. The lower the order, the closer the relationship between drivers. We sample in the graph. Each time we sample, we select 3 drivers A, B, and C, and calibrate the relationship orders p AB and p2] AC of the driver pairs AB and AC. Based on the driver vectors, calculate the correlations s AB and s AC between drivers A and B, and A and C, and compare. If one of the following conditions is met, it means the prediction is correct:

[0133] p AB < p AC and s AB > s AC ;

[0134] p<0000)35]]> p AC and s AB < s AC ;

[0135] ​Example 2:

[0136] 2.3 Model training method:

[0137] P1. Set the model parameters, including the dimension of the embedding vector, the number of random walk steps, and the number of negative samples.

[0138] P1.1: Set the dimension of the embedding vector:

[0139] The dimension of the embedding vector is a hyperparameter that determines the size of the embedding space. Generally, the higher the dimension, the richer the information the model can capture, but the higher the computational cost.

[0140] P1.2: Set the number of steps of random walk:

[0141] The number of random walk steps determines the length of the generated node sequence. The longer the number of steps, the wider the range of nodes that the sequence can cover, but it may also introduce more noise.

[0142] P1.3: Set the number of negative samples:

[0143] The number of negative samples is used to control the number of negative examples during training. Negative samples are used to improve the generalization ability of the model, but too many negative samples may also cause instability in the training process.

[0144] P2. Use random numbers to initialize the embedding vector as the initial weight of the model.

[0145] P3. Random Walk Sequence Generation: Perform a random walk on the driver relationship graph to generate a series of node sequences. Each sequence can be regarded as a "sentence" and the nodes in it can be regarded as "words".

[0146] P3.1: Perform a random walk on the driver relationship graph:

[0147] for each node in V:

[0148] sequences.append(RandomWalk(node,walk_length,G))

[0149] Among them, the RandomWalk function is used to perform random walk, and G represents the driver relationship graph.

[0150] P4. Model training: Use Word2Vec's skip-gram mechanism to train node embeddings. By continuously adjusting the weights of the embedding vectors, we minimize the loss function, which is the difference in similarity between the neighbor nodes predicted by the model and the actual neighbor nodes:

[0151] model=SkipGramModel(embedding_dim,num_negative_samples)

[0152] model.train(sequences,num_epochs=num_epochs) %num_epochs represents the number of training rounds

[0153] Among them, SkipGramModel represents the frequency hopping model class, sequences represents the generated node sequence set, and num_epochs represents the number of training rounds.

[0154] The loss function is:

[0155] L=-∑ (u,v)∈PositiveSamples logσ(e u ·e v ′)-∑ (u,v′)∈NegativeSamples logσ(-e u ·e v′ ′)

[0156] Among them, e u and e v 'represents the embedding vectors of node u and node v (v is the neighbor of u), σ represents the sigmoid function, PositiveSamples represents the positive sample set, and NegativeSamples represents the negative sample set.

[0157] P5. Model optimization: Adjust model parameters based on the loss function value during training and the performance on the validation set. Use the Adam optimizer to speed up the training process and reduce overfitting:

[0158] optimizer=AdamOptimizer(learning_rate=learning_rate)

[0159] model.optimize(optimizer, validation_data) % validation_data represents the validation set data

[0160] Among them, AdamOptimizer represents the Adam optimizer class, learning_rate represents the learning rate, and validation_data represents the validation set data.

[0161] All of the above embodiments merely represent implementation methods of the present invention in practical applications. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the scope of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the appended claims.

[0162] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0163] At the same time, those skilled in the art will understand that all or part of the processes in all the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

Claims

1. A driver group mining method based on a graph model, characterized by: The following steps are included: S1, collects driver node information, driver work history, direct relationship data between drivers, and travel history geographic location data from the service logs, user feedback, and interaction records between drivers and passengers of the online ride-hailing platform to form dataset D; S2, based on the collected dataset D between drivers, build a graph model of the driver relationship graph G; S3, using the trained model, calculates the similarity between driver nodes; Drivers are divided into different groups based on similarities.

2. The driver group mining method according to claim 1, characterized in that: In S1, the collected driver node information includes ID, age, gender, driving experience and vehicle type; the driver's work history includes order acceptance time, location and service type; and the direct relationship data includes friend and colleague relationships.

3. The driver group mining method according to claim 2, characterized in that: In the S1, the data set D = (DriverNodeInfo, DriverWorkHistory, DriverRelationships, DriverGeoLocationHistory); DriverNodeInfo is the driver node information, DriverWorkHistory is the driver's work history; DriverRelationships is a two-dimensional matrix, which is the direct relationship data between drivers. The element R[i][j] in the matrix represents the relationship between driver i and driver j, where 1 indicates the existence of a relationship and 0 indicates the absence of a relationship; DriverGeoLocationHistory is the travel history geographic location data.

4. The driver group mining method according to claim 1, characterized in that: In S2, the weight of the edge is set according to the closeness of the number of passengers served together.

5. The driver group mining method according to claim 4, characterized in that: The execution process of S2 includes: S200, extract all drivers’ IDs from dataset D and initialize driver node set V: V=ExtractUnique(D.DriverNodeInfo,ID) The ExtractUnique function is used to extract unique values, D.DriverNodeInfo represents the driver node information in dataset D, and ID represents the field name to be extracted. S201, create an empty set E to store the edges between drivers; S202, traverse the direct relationship data and travel history geographic location data between drivers in data set D; for each pair of drivers with a direct relationship, or a pair of drivers who have served passengers together, add an edge to the edge set E, and set the weight of the edge according to the number of passengers served together.

6. The driver group mining method according to claim 5, characterized in that: In S2, the edge weight is calculated as: weight ij =∑ k∈CoServicePassengers(i,j) 1; Among them, weight ij represents the weight of the edge between driver i and driver j, CoServicePassengers}(i,j) represents the set of passengers served by driver i and driver j together; Use the driver node set V and edge set E to construct the driver relationship graph G = (V, E).

7. The driver group mining method according to claim 1, characterized in that: The execution process of S3 includes: S300, for each pair of driver nodes, calculate the cosine similarity between their embedding vectors: where, e u and e v represent the embedding vectors of driver nodes u and v respectively, · represents the dot product of vectors, |e u | and |e v | represent the norms of vectors e_u and e_v respectively; When the similarity scores of two driver nodes are higher than the preset threshold, they are considered to belong to the same group; Traverse all driver node pairs and divide drivers into different groups based on similarity scores and set thresholds; driver nodes in each group have similar characteristics or behavior patterns.

8. The driver group mining method according to claim 7, characterized in that: In S3, a multi-order relationship graph of drivers is constructed based on the driver invitation relationship. The lower the order, the closer the relationship between the drivers. Sampling is performed in the graph. Three drivers A, B, and C are selected for each sampling. The order p of the relationship between the AB driver pair and the AC driver pair is calibrated. AB and p AC ; Based on the driver vector, calculate the correlation s between drivers A and B, and A and C AB 、s AC , and compare; if one of the following conditions is met, the estimate is correct: p AB <p AC , and s AB >s AC ; p AB >p AC , and s AB AC .​ 9. A system for implementing the driver group mining method according to any one of claims 1 to 8, characterized in that: The system comprises: A data collection module that collects multi-dimensional data of drivers from ride-hailing platforms; A model training module for building a driver relationship graph; A group segmentation module that uses the trained model and embedding vectors to calculate the similarity between driver nodes; An application module that provides personalized services for online ride-hailing platforms based on group segmentation results.

10. The system according to claim 9, characterized in that: The model training module uses the frequency hopping mechanism of Word2Vec to train the embedding representation of the driver node.

Images