Unmanned driving vehicle model recommendation method, system and device based on passenger preference fusion

Abstract

The present invention is based on the unmanned vehicle model recommendation method, system and device based on passenger preference fusion, which extracts the interaction data between the target passenger and the unmanned vehicle model and some auxiliary information from the interaction information between the passenger and the unmanned vehicle, accurately captures the real preferences and needs of the passenger, highly personalizes and accurately recommends the results, and analyzes the characteristics of the different preferences of the passengers; sets the recommendation value threshold, and when the recommendation value reaches or exceeds the recommendation value threshold, recommends the target unmanned vehicle model to the target passenger. This method realizes the accurate recommendation of unmanned vehicle models through the multi-perspective preference fusion of passengers and unmanned vehicles, improves the accuracy of unmanned vehicle model recommendation, and enhances the passenger experience.

Description

Technical Field

[0001] The present invention mainly relates to the field of traffic service technology, and specifically to a method, system and device for recommending unmanned vehicle models based on passenger preference fusion. Background Art

[0002] With the rapid development of autonomous driving technology, driverless models are becoming more and more abundant, and driverless cars are gradually being introduced into ride-hailing software. It greatly improves the convenience and safety of travel, and reduces traffic accidents caused by human errors through high-precision sensors and algorithms; driverless technology reduces operating costs, improves operating efficiency, and provides passengers with more economical and affordable travel options. How to match passengers with suitable driverless models has become an important issue.

[0003] Chinese invention CN202210985224.7 discloses a method, device and storage medium for dynamic ride-sharing scheduling of shared autonomous vehicles. According to the ride-sharing constraint conditions, multiple ride matching schemes are determined from the travel request information and vehicle information: each ride matching scheme is optimized according to the vehicle scheduling matching optimization principle; the optimal ride matching scheme is determined from the optimized multiple ride matching schemes and sent to the passenger terminal and the paired shared autonomous vehicle terminal. However, it ignores the passenger's preference for autonomous unmanned vehicles and cannot fully meet the unique preferences of different passengers. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method, system and device for recommending driverless vehicle models based on passenger preference fusion. The method accurately captures user preferences by integrating multi-perspective data of passenger travel information, travel time, carpooling records and historical evaluations, and provides users with highly personalized vehicle model recommendations. This not only improves the user experience, but also promotes the market application and popularization of driverless technology, and is of significant necessity.

[0005] The present invention proposes a method for recommending unmanned driving vehicle models based on passenger preference fusion, comprising the following steps:

[0006] S1, the passenger set and the driverless vehicle model set constitute an interaction matrix. The driverless vehicle models that have interacted with the target passenger are obtained from the interaction matrix. After preprocessing, the target passenger interaction vehicle model perspective preference matrix is ​​generated.

[0007] The target passenger's travel information, target passenger travel time period, target passenger carpooling record, and target passenger historical evaluation constitute the target passenger comprehensive information. After preprocessing, the target passenger comprehensive information is constructed by graph construction or matrix construction to generate the target passenger travel hotspot perspective preference graph, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix.

[0008] The target passenger's comprehensive perspective preference is obtained through weighted summation, and the target passenger's interactive vehicle perspective preference matrix, the target passenger's active time perspective preference matrix, the target passenger's carpooling behavior perspective preference matrix, and the target passenger's evaluation history perspective preference matrix with the largest weight are obtained. The target passenger's comprehensive perspective preference and the target passenger's initial information are then weightedly summed to obtain the target passenger's basic information;

[0009] S2. Obtain comprehensive information of the target driverless vehicle model, including the driving record and running time information of the target vehicle model, use the departure and destination as nodes, and the running time as edges to generate a travel heat location map of the target vehicle model.

[0010] The operating frequency of the target unmanned driving model is counted in different time periods, and the data with the operating frequency not less than the operating frequency threshold are sorted to form a matrix. The intelligent driving high-frequency operating time matrix is ​​taken to form the intelligent driving high-frequency operating time matrix. The intelligent driving high-frequency operating time matrix is ​​added to the initial information of the target unmanned driving model to obtain the weight of the intelligent driving optimization information. After weighted summation, the basic information of the unmanned driving model is obtained.

[0011] S3, based on the target passenger travel hotspot perspective preference graph, the target driverless vehicle travel heat location graph, taking passengers, vehicle types, departure places, and destinations as nodes, and the vehicle types used by the target passengers, ride time, and target driverless vehicle operating time as edges, to construct a hybrid graph structure.

[0012] The target passenger basic information and the target driverless vehicle basic information are put into the hybrid graph structure, and the node representation is updated through the message passing mechanism. The target passenger basic information and the target driverless vehicle basic information are updated according to the updated nodes to obtain the target passenger comprehensive information and driverless vehicle comprehensive information.

[0013] S4. Multiply the comprehensive information of the target passenger and the comprehensive information of the driverless vehicle model to obtain the fusion information of the target passenger vehicle model; perform normalization operation on the fusion information of the target passenger vehicle model to obtain the recommendation score of the driverless vehicle model, and compare the relationship between the score and the recommendation threshold.

[0014] If the recommendation score is greater than or equal to the recommendation threshold, the driverless vehicle model is recommended to the target passenger.

[0015] The specific construction method of the target passenger travel hotspot perspective preference map in S1 is:

[0016] The starting point and destination of the target passenger are taken as nodes, the distance from the starting point to the destination is taken as an edge, and each edge is associated with the travel time information to construct a travel hotspot perspective preference graph for the target passenger.

[0017] The target passenger active time preference matrix in S1 is as follows: divide the 24 hours of a day into several time periods, obtain the target passenger travel time periods, and use the time periods when the target passengers travel most frequently as the common time periods. For target passengers who lack travel time period information, assign the default time periods as their common time periods. Summarize the common time periods of all target passengers into a matrix, which is the target passenger active time perspective preference matrix.

[0018] The target passenger carpooling behavior perspective preference matrix in S1 is specifically as follows: obtain the target passenger's ride record. If the target passenger chooses carpooling, it is marked as 1 in the corresponding position; if the target passenger does not choose carpooling, it is marked as 0. The ride records of all target passengers form a matrix, which is the target passenger carpooling behavior perspective preference matrix.

[0019] The target passenger evaluation history perspective preference matrix in S1 is: the target passenger's historical evaluation is obtained from the target passenger's comprehensive information, sorted in chronological order, and the evaluation score range is 0-5. If the target passenger lacks an evaluation record, the first evaluation record is used to fill it; if the target passenger has no evaluation record, it is assigned a full 5 points by default. All target passengers' evaluations are summarized into a matrix, which is the target passenger evaluation history perspective preference matrix.

[0020] The comprehensive information of the target passenger and the comprehensive information of the unmanned vehicle model are obtained in S3. The specific operations are as follows: the basic information of the target passenger and the basic information of the target unmanned vehicle model are put into the mixed graph structure, and the features of their neighbor nodes are collected by averaging, summing or maximum pooling. The neighbor information is combined with the characteristics of the current node to generate a new node representation vector, and the comprehensive information of the target passenger and the comprehensive information of the target unmanned vehicle model are obtained.

[0021] The weights in S1 and S2 are obtained by obtaining the target weights through a multi-head attention mechanism.

[0022] In S4, the recommended threshold is obtained. The specific method is: finding the optimal recommended threshold through the amount of information to match the target passenger and the target unmanned vehicle model.

[0023] The driverless vehicle model recommendation system based on passenger preference fusion includes:

[0024] Passenger information collection module, used to obtain the interaction information between each target passenger and the driverless vehicle;

[0025] The driverless car information collection module is used to obtain the driving records and operating time of the driverless car;

[0026] The learning module is used in the hybrid graph structure to obtain the comprehensive feature information of the target passenger and the comprehensive feature information of the driverless vehicle after learning the information of the adjacent nodes;

[0027] The driverless vehicle model recommendation module is used to determine the relationship between the recommendation score and the recommendation value threshold:

[0028] If the recommendation score is greater than or equal to the recommendation threshold, the driverless vehicle model is recommended to the target passenger.

[0029] The present invention also proposes an unmanned driving vehicle model recommendation device based on passenger preference fusion, including a processor and a memory and a computer program stored in the memory and running on the processor. The processor executes the computer program stored in the memory to implement the aforementioned unmanned driving vehicle model recommendation method based on passenger preference fusion.

[0030] The beneficial effects of the technical solution of the present invention are as follows: the present invention extracts the interaction data between the target passenger and the driverless car model and some auxiliary information from the interaction information between the passenger and the driverless car, accurately captures the real preferences and needs of the passenger, highly personalizes and accurately recommends the results, and analyzes the characteristics of the different preferences of the passengers; sets the recommendation value threshold, and when the recommendation value reaches or exceeds the recommendation value threshold, recommends the target driverless car model to the target passenger. The method achieves accurate recommendation of driverless car models by fusing the multi-perspective preferences of passengers and driverless vehicles, improves the accuracy of driverless car model recommendation, and enhances the passenger experience. DETAILED DESCRIPTION

[0031] An embodiment of the present invention provides a method for recommending an unmanned driving vehicle model based on passenger preference fusion, the method comprising the following steps:

[0032] S1, the passenger set and the driverless vehicle model set constitute an interaction matrix. The driverless vehicle models that have interacted with the target passenger are obtained from the interaction matrix. After preprocessing, the target passenger interaction vehicle model perspective preference matrix is ​​generated.

[0033] The target passenger's travel information, target passenger travel time period, target passenger carpooling record, and target passenger historical evaluation constitute the target passenger comprehensive information. After preprocessing, the target passenger comprehensive information is constructed by graph construction or matrix construction to generate the target passenger travel hotspot perspective preference graph, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix.

[0034] The target passenger's comprehensive perspective preference is obtained through weighted summation, and the target passenger's interactive vehicle model perspective preference matrix, target passenger's active time perspective preference matrix, target passenger's carpooling behavior perspective preference matrix, and target passenger's evaluation history perspective preference matrix are obtained. The preference matrix with the largest weight is then added to the target passenger's comprehensive perspective preference and the target passenger's initial information to obtain the target passenger's basic information.

[0035] It should be noted that the passenger set and the driverless vehicle model set in S1 constitute an interaction matrix, and the driverless vehicle model set that has interacted with the target passenger is obtained from the interaction matrix:

[0036] ,

[0037] in Represents the set of car models that the target passenger has interacted with, and the elements in the set represent the car models.

[0038] The interaction frequencies between the target passenger and different driverless car models are arranged from high to low to obtain the set of driverless car models that the target passenger often interacts with. The first N items equal to the threshold N are selected to form the target passenger car model preference set:

[0039] ,

[0040] The threshold N is obtained based on a dynamically adjusted threshold system, which sets different thresholds based on the matching results of the target passenger and the target driverless vehicle model in each round until the matching result reaches the expected result. Other preferences also use the dynamically adjusted threshold system to adjust the threshold N.

[0041] In addition, if the target passenger has interacted with the driverless vehicle less than N times, the vehicle with the highest interaction frequency is used to fill in; if the target passenger has no interaction data with the driverless vehicle, the default set is used Target passenger interaction models:

[0042] ,

[0043] in, The default set of target passenger interaction models. The default model.

[0044] The M target passenger vehicle model preference sets are integrated together to form a matrix with M rows and N columns, which is the target passenger interactive vehicle model perspective preference matrix. Among them, M represents the total number of passengers, and N represents the N vehicle models that the target passengers interact with most frequently.

[0045] Target passenger interaction vehicle view preference matrix It can be expressed as:

[0046] .

[0047] The target passenger's travel information, target passenger's travel time period, target passenger's carpooling record, and target passenger's historical evaluation constitute the target passenger's comprehensive information.

[0048] Preprocessing of comprehensive information of target passengers:

[0049] Define the target passenger set:

[0050] ,

[0051] in, is the target passenger set, and the elements in the set are the target passengers.

[0052] Define the target passenger origin and destination set:

[0053] ,

[0054] ,

[0055] in, S is a set of target passenger starting points, and the elements in the set are target passenger starting points; D is a set of target passenger destination points, and the elements in the set are target passenger destination points.

[0056] Define the target passenger boarding time set:

[0057] ,

[0058] in, T is a set of target passenger boarding times, and the elements in the set are the target passenger boarding times.

[0059] Target passengers , define the travel data set:

[0060] ,

[0061] in, Indicates passengers No. The travel record of the starting point, ride time and destination of each trip.

[0062] The starting point and destination of the target passenger are regarded as nodes in the graph, and the travel time is used as the edge to construct the target passenger travel hotspot perspective preference graph:

[0063] ,

[0064] in is a node set, is the edge set. Includes all nodes at the origin and destination. E i Contains the edge from the origin to the destination, for each travel record Add Edge arrive , and marked as ride time .

[0065] Preprocessing of target passenger travel time period:

[0066] Divide the 24 hours of a day into several time periods and define the time period set:

[0067] ,

[0068] in, 0-3 points, 3-6 points, 6-9 o'clock, From 9 to 12 o'clock, 12-15 o'clock, 15:00-18:00, 18-21 o'clock, It is 21-24 points.

[0069] Count the frequency of target passengers taking taxis in the target time period. According to the statistical results, sort the time periods from high to low according to the travel frequency, and take N time periods as the target passenger active time periods. If the target passenger active time periods are less than N, fill in the time period with the highest travel frequency of the target passenger. For passengers who do not travel, assign the default time period. For passengers , define the time period frequency vector: ,

[0070] in, Indicates passengers In time period Frequency of taxi rides within.

[0071] The M target passenger active time periods are integrated together to form a matrix with M rows and N columns, which is the target passenger active time perspective preference matrix. Among them, M represents the total number of passengers, and N represents the N target passenger active time periods with the highest travel frequency.

[0072] Target passenger active time perspective preference matrix H It can be expressed as:

[0073] .

[0074] Get the target passenger's travel record and pre-process the target passenger's travel record:

[0075] Define the carpooling set:

[0076] ,

[0077] in, Indicates passengers No. Whether the ride record is a carpool ride.

[0078] If the target passenger chooses carpooling, the corresponding position is marked as 1; if not, it is marked as 0; for passengers who carpool less than N times, the missing part needs to be filled with 0. The purpose of this is to ensure that the target passengers have N carpooling marks, and there will be no deviation in data analysis due to incomplete records. For target passengers with no riding records, by default, the carpooling marks of these passengers are all set to 0, which means there is no carpooling behavior. This default value is set based on the assumption that most new passengers have the first-time usage habits, avoiding incorrect predictions of passenger behavior without data.

[0079] The M target passenger carpooling sets are integrated together to form a matrix with M rows and N columns, which is the target passenger carpooling behavior perspective preference matrix. Among them, M represents the total number of passengers, and N represents the target passenger carpooling mark. This matrix can be used for further data analysis, such as the study of passenger carpooling habits and the optimization of carpooling services, providing important data support for subsequent passenger behavior analysis and service optimization.

[0080] The target passenger carpooling behavior perspective preference matrix is:

[0081] .

[0082] Target passenger historical evaluation preprocessing:

[0083] Define the evaluation record collection:

[0084] ,

[0085] in, Indicates passengers No. The rating of the review record.

[0086] Sort by time, select N records, and the evaluation score range is 0-5, where 0 represents very bad and 5 represents very good.

[0087] If the target passenger has fewer than N evaluation records, the first evaluation record is used to fill the number to N, ensuring that the number of evaluation records for all passengers is consistent, which is convenient for subsequent data processing and analysis. For passengers without evaluation records, the default score is 5. This default value is set to reflect that most passengers usually give higher scores when they are not particularly dissatisfied, which can also avoid analysis bias caused by lack of data.

[0088] The evaluation records of M target passengers are integrated together to form a matrix with M rows and N columns, which is the target passenger evaluation history perspective preference matrix. Among them, M represents the total number of passengers and N represents the target passenger evaluation records.

[0089] The target passenger evaluation history perspective preference matrix can be expressed as:

[0090] .

[0091] The weight of the target preference matrix is ​​calculated through the multi-head attention mechanism, taking the target passenger interaction vehicle perspective preference matrix as an example:

[0092] The target passenger interaction vehicle viewpoint preference is calculated in the first hidden layer using the following formula:

[0093] ,

[0094] in, It is the first i Quantity, yes l-1 The number of neurons in the layer, is from l-1 Tier i neurons to the current layer j The weight of the neuron, is the bias term of the neuron in the current layer, and the result of the above formula is normalized:

[0095] ,

[0096] in, The current layer j The output of a neuron, e is the base of natural logarithms,

[0097] Perform calculations for the second hidden layer:

[0098] ,

[0099] in, is the number of neurons in the penultimate layer, is from l-1 Layer by layer j neurons to the output layer k The weight of the neuron, The output layer k The bias term of a neuron, It is l-1 Tier j The output of a neuron.

[0100] Normalize the above result to get the weight of the target passenger interaction vehicle model perspective preference matrix: , the formula is as follows:

[0101] .

[0102] The calculation method of other perspective preference matrix weights is the same as the passenger interaction vehicle perspective preference weights.

[0103] The target passenger interaction vehicle perspective preference matrix, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix are multiplied by their weights and then added to obtain the target passenger comprehensive perspective preference:

[0104] ,

[0105] in, Represents the weight of the target passenger interaction vehicle model perspective preference matrix; Represents the weight of the target passenger's active time perspective preference matrix; Target passenger carpooling behavior perspective preference matrix weights; represents the weight of the target passenger's historical perspective preference matrix; Z represents the target passenger's comprehensive perspective preference.

[0106] Obtain the preference matrix with the largest weight ratio from the perspective of target passenger interaction model, the perspective of target passenger active time, the perspective of target passenger carpooling behavior, and the perspective of target passenger evaluation history. Here, the preference matrix with the largest weight ratio is set as X , passengers’ comprehensive viewing preferences Z and target passenger initial information U Use the following formula to reshape the matrix:

[0107] ,

[0108] ,

[0109] ,

[0110] in, X The preference matrix with the largest weight among the target passenger interaction vehicle model perspective preference matrix, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix. For its weight, for X The reshape matrix of is the reshape matrix of Z; for U The reshaped matrix.

[0111] The input matrix is ​​linearly transformed to the new representation space by multiplying it with the weights of the corresponding matrix:

[0112] ,

[0113] ,

[0114] ,

[0115] in, for New representation space, for New representation space, for New representation space, The number of heads.

[0116] Divide the above calculation result into A head, that is , , ..., ; , , ..., ; , , ..., ,

[0117] The dimensions of the target head are:

[0118] ,

[0119] in, is the dimension of the target head, Dimensions for each head.

[0120] Then calculate the weight of the target head:

[0121] ,

[0122] in, , , Target head ，A is the weight of the target head, for The transpose of .

[0123] The function :

[0124] .

[0125] The target header can be expressed as:

[0126] ,

[0127] in, is a simplified representation of the target head weight,

[0128] Connect all the heads together to get the basic information of the target passenger :

[0129] ,

[0130] in, Basic information of target passengers. Represents the output weight matrix.

[0131] S2. Obtain comprehensive information of the target driverless vehicle model, including the driving record and operating time information of the target vehicle model, and define the vehicle model set:

[0132] ,

[0133] Target driverless vehicle operation location set:

[0134] ,

[0135] .

[0136] Target driverless vehicle model running time collection:

[0137] .

[0138] For target models , define the driving record set:

[0139] ,

[0140] in, Indicates vehicle model No. The origin, duration and destination of each trip.

[0141] Use the departure and destination as nodes and the running time as edges to build a travel heat location map for the target vehicle type ,in is a node set, is the edge set. Includes all location nodes.

[0142] Edge Set Contains edges from location to location, each edge is also associated with running time information: for each driving record Add Edge arrive , and mark the running time as .

[0143] Divide the 24 hours of a day into several fixed time periods and define the time period set:

[0144] ,

[0145] in, 0-3 points, 3-6 points, 6-9 o'clock, From 9 to 12 o'clock, 12-15 o'clock, 15:00-18:00, 18-21 o'clock, It is 21-24 points.

[0146] Then count the frequency of unmanned vehicle operation in the target time period, arrange the frequencies from high to low, take out N operating time periods of the target vehicle model, and form a matrix of M rows and N columns. This matrix is ​​called the intelligent driving high-frequency operating time.

[0147] Intelligent driving high-frequency operation time It can be expressed as:

[0148] .

[0149] This matrix has multiple advantages in the scenario of unmanned vehicle model recommendation. It optimizes the scheduling of models and recommends the most suitable models by counting the operating frequency of vehicles in different time periods. This data-driven analysis can also support customized services, thereby improving service response speed, resource utilization efficiency and passenger satisfaction.

[0150] Calculate intelligent driving optimization information:

[0151] ,

[0152] in, High-frequency operation time for intelligent driving, C is the initial information of the target autonomous driving vehicle model, Y Optimizing information for smart driving.

[0153] The high-frequency operation time of intelligent driving, the initial information of the target autonomous driving unmanned vehicle model, and the intelligent driving optimization information are processed through the function:

[0154] ,

[0155] ,

[0156] ,

[0157] in, for The reshaped matrix, for The weight of for The reshaped matrix, for The weight of for The reshaped matrix, for The weight of .

[0158] Through the matrix and the corresponding weight matrix , , Multiply to linearly transform the input matrix into the new representation space:

[0159] ,

[0160] ,

[0161] ,

[0162] in, for New representation space, for New representation space, for New representation space, The number of heads.

[0163] Then , , They are divided into heads, namely , , ..., ; , , ..., ; , , ..., ;

[0164] The dimensions of the target head are:

[0165] ,

[0166] in, is the dimension of the target head, Dimensions for each head.

[0167] Calculate the weight of the target head :

[0168] ,

[0169] Among them, the function :

[0170] .

[0171] Then the target header can be expressed as:

[0172] ,

[0173] in, Simplified representation of the target head weight.

[0174] Connect all the heads together to get the basic information of the target autonomous driving vehicle model :

[0175] ,

[0176] in, Basic information of target passengers, Represents the output weight matrix.

[0177] S3. Construct a hybrid graph structure based on the target passenger travel hotspot perspective preference graph and the target driverless vehicle travel heat location graph:

[0178] The target passengers used a specific type of vehicle for their journey, which is:

[0179] ,

[0180] A collection of trips for targeted passengers using a specific vehicle type.

[0181] A specific type of vehicle is used to perform the trip, the set is:

[0182] ,

[0183] A collection of vehicles of a particular type that are used to perform a journey.

[0184] The passenger travel time set is:

[0185] ,

[0186] Gather passengers at boarding time.

[0187] The set of driverless car running time is:

[0188] ,

[0189] in, Gather passengers at boarding time. and are the set of target customer’s ride time and target vehicle model’s operating time respectively.

[0190] The passengers, vehicle types, departure places and destinations are taken as nodes, and the vehicle type used by the target passenger, the time of the ride, and the operating time of the target driverless car are taken as edges to construct a hybrid graph structure. The basic information of the target passenger and the basic information of the target driverless car model are used to update the representation of the node through the message passing mechanism. This process is achieved by considering the information of the node's adjacent nodes. The specific steps are as follows:

[0191] The basic information of the target passenger and the basic information of the target driverless vehicle are put into the hybrid graph structure. , collect neighbor nodes Features , features are aggregated by averaging, summing or max pooling, taking averaging as an example:

[0192] ,

[0193] in, is the average feature of the target node, is the neighbor node, are the features of the collected neighbor nodes, .

[0194] Combining the collected neighbor information and the current node’s own characteristics, a new node representation vector is generated:

[0195] ,

[0196] in Is a node In the The layer representation, Is a node u In the The layer representation, is a nonlinear transformation function.

[0197] By stacking multiple layers of features, the model can gradually expand the contextual information of the nodes, thereby better capturing the complex relationships between the nodes in the graph. After the message transmission of the layer, the comprehensive information of the target passenger and the comprehensive information of the target autonomous driving unmanned vehicle are obtained:

[0198] ,

[0199] in, It is in Neighbor information collected by the layer, Is a node u In the L -1 layer representation, Is a node u In the t The layer representation, It is L Layer nonlinear transformation function.

[0200] S4. Multiply the comprehensive information of the target passenger and the comprehensive information of the driverless vehicle model to obtain the fusion information of the target passenger and vehicle model :

[0201] .

[0202] Normalize the target passenger vehicle model fusion information to obtain the recommended value of the target autonomous driving unmanned vehicle model :

[0203] .

[0204] Where e is the base of natural logarithms, approximately equal to 2.71828, It is the passenger vehicle type fusion information.

[0205] If the target autonomous driving vehicle model has the recommended value If the value is greater than or equal to 0.5, the target autonomous driving vehicle model will be recommended to the target passenger. And according to the recommendation result, the dynamic threshold system will adjust the target preference threshold to obtain better results in the next round of recommendation. Intelligent data filling and filtering will also adjust the data accordingly according to the results until the optimal result is achieved.

[0206] The present invention also proposes an unmanned driving vehicle model recommendation system based on passenger preference fusion, comprising:

[0207] Target passenger basic information acquisition module: extracts the interaction data and some related information between the target passenger and the driverless vehicle model from the target passenger and driverless vehicle model interaction matrix and the passenger comprehensive information table, generates the passenger interaction vehicle model perspective preference, passenger travel hotspot perspective preference map, passenger active time perspective preference, passenger carpooling behavior perspective preference, passenger evaluation history perspective preference, and auxiliary information that can enhance the passenger experience. The target passenger basic information is obtained by integrating the passenger initial information and part of the auxiliary information;

[0208] Target driverless vehicle basic information acquisition module: obtain the driverless vehicle travel popularity location map and intelligent driving high-frequency operation time through the comprehensive information table of the driverless vehicle, and integrate the intelligent driving high-frequency operation time and the initial information of the driverless vehicle to obtain the basic information of the driverless vehicle;

[0209] Hybrid graph construction and update module: Based on the target passenger travel hotspot perspective preference graph and the target driverless vehicle model travel heat location graph, a hybrid graph structure is constructed, and the nodes of the target passenger basic information and the target driverless vehicle model basic information are updated to obtain the target passenger comprehensive information and the target driverless vehicle model comprehensive information;

[0210] Unmanned vehicle model recommendation module: used to process the comprehensive information of target passengers and the comprehensive information of target unmanned vehicle models, obtain the recommended score of the target unmanned vehicle model, and determine the relationship between the recommended score and the recommended value threshold.

[0211] The present invention also proposes an unmanned driving vehicle model recommendation device based on passenger preference fusion, including a processor and a memory and a computer program stored in the memory and running on the processor. The processor executes the computer program stored in the memory to implement the aforementioned unmanned driving vehicle model recommendation method based on passenger preference fusion.

[0212] The beneficial effects of the technical solution of the present invention are: The beneficial effects of the technical solution of the present invention are: The present invention extracts the interaction data between the target passenger and the driverless car model and some auxiliary information from the interaction information between the passenger and the driverless car, accurately captures the real preferences and needs of the passenger, highly personalizes and accurately recommends the results, and analyzes the characteristics of the different preferences of the passengers; sets the recommendation value threshold, and when the recommendation value reaches or exceeds the recommendation value threshold, recommends the target driverless car model to the target passenger. This method achieves accurate recommendation of driverless car models by fusing the multi-perspective preferences of passengers and driverless vehicles, improves the accuracy of driverless car model recommendations, and enhances the passenger experience.

Claims

1. A method for recommending unmanned vehicle models based on passenger preference fusion, characterized in that: The steps include: S1, the passenger set and the driverless vehicle model set constitute an interaction matrix. The driverless vehicle models that have interacted with the target passenger are obtained from the interaction matrix. After preprocessing, the target passenger interaction vehicle model perspective preference matrix is ​​generated. The target passenger's travel information, target passenger travel time period, target passenger carpooling record, and target passenger historical evaluation constitute the target passenger comprehensive information. After preprocessing, the target passenger comprehensive information is constructed by graph construction or matrix construction to generate the target passenger travel hotspot perspective preference graph, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix. The target passenger interaction vehicle model perspective preference matrix, the target passenger active time perspective preference matrix, the target passenger carpooling behavior perspective preference matrix, and the target passenger evaluation history perspective preference matrix are respectively multiplied by their weights and then added to obtain the target passenger comprehensive perspective preference, and the preference matrix with the largest weight among the target passenger interaction vehicle model perspective preference matrix, the target passenger active time perspective preference matrix, the target passenger carpooling behavior perspective preference matrix, and the target passenger evaluation history perspective preference matrix is ​​obtained, and the target passenger comprehensive perspective preference and the target passenger initial information are weighted and summed to obtain the target passenger basic information; S2. Obtain comprehensive information of the target driverless vehicle model, including the driving record and running time information of the target vehicle model, use the departure and destination as nodes, and the running time as edges to generate a travel heat location map of the target vehicle model. The operating frequency of the target unmanned driving model is counted in different time periods, and the data with the operating frequency not less than the operating frequency threshold are sorted to form a matrix. The intelligent driving high-frequency operating time matrix is ​​taken to form the intelligent driving high-frequency operating time matrix. The intelligent driving high-frequency operating time matrix is ​​added to the initial information of the target unmanned driving model to obtain the weight of the intelligent driving optimization information. After weighted summation, the basic information of the target unmanned driving model is obtained. S3. Based on the target passenger travel hotspot perspective preference graph and the target driverless vehicle travel heat location graph, the passengers, vehicle types, departure places, and destinations are used as nodes, and the vehicle types used by the target passengers, ride time, and target driverless vehicle operation time are used as edges to construct a hybrid graph structure. The node representation is updated through a message passing mechanism, and the basic information of the target passenger and the basic information of the target driverless vehicle model are updated according to the updated node to obtain the comprehensive information of the target passenger and the comprehensive information of the target driverless vehicle model; S4, multiplying the comprehensive information of the target passenger and the comprehensive information of the target driverless vehicle model to obtain the fusion information of the target passenger and vehicle model; The target passenger vehicle model fusion information is normalized and the target driverless vehicle model recommendation score is obtained. The relationship between the score and the recommendation threshold is compared. If the recommendation score is greater than or equal to the recommendation threshold, the target driverless vehicle model is recommended to the target passenger.

2. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The specific construction method of the target passenger travel hotspot perspective preference map in S1 is: The starting point and destination of the target passenger are taken as nodes, the distance from the starting point to the destination is taken as an edge, and each edge is associated with the travel time information to construct a travel hotspot perspective preference graph for the target passenger.

3. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The target passenger active time preference matrix in S1 is specifically: The 24 hours of a day are divided into several time periods, and the travel time periods of the target passengers are obtained. The time periods when the target passengers travel most frequently are taken as the common time periods. For the target passengers who lack travel time period information, the default time periods are assigned as the passenger active time periods. The common time periods of all target passengers are summarized into a matrix, which is the target passenger active time perspective preference matrix.

4. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The target passenger carpooling behavior perspective preference matrix in S1 is specifically: Get the target passenger's riding record. If the target passenger chooses carpooling, it is marked as 1 in the corresponding position; if not, it is marked as 0. The riding records of all target passengers form a matrix, which is the target passenger's carpooling behavior perspective preference matrix.

5. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The target passenger evaluation history perspective preference matrix in S1 is: The historical evaluation of the target passenger is obtained from the comprehensive information of the target passenger and sorted in chronological order. The evaluation score range is 0-5. If the target passenger lacks an evaluation record, the first evaluation record is used to fill it; if the target passenger has no evaluation record, it is assigned a full 5 points by default. The evaluations of all target passengers are summarized into a matrix, which is the target passenger evaluation history perspective preference matrix.

6. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The specific operation of updating the node representation in S3 is: The features of its neighbor nodes are collected by averaging, summing or maximum pooling, and the neighbor information is combined with the current node's own features to generate a new node representation vector.

7. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The weight acquisition method in S1 and S2 is: obtaining the target weight through a multi-head attention mechanism.

8. The method for recommending unmanned vehicle models based on passenger preference fusion according to claim 1, characterized in that: The specific method of obtaining the travel heat location map of the target vehicle type in S2 is: Locate the stop points of the target driverless vehicle model, count the locations of the stop points, determine the locations where the driverless target vehicle model runs most frequently, use the departure and destination as nodes, and the running time as edges to obtain the travel heat location map of the target vehicle model.

9. An unmanned vehicle model recommendation system based on passenger preference fusion, characterized in that: include: Target passenger basic information acquisition module: The passenger set and the driverless vehicle model set constitute an interaction matrix, and the driverless vehicle models that have interacted with the target passenger are obtained from the interaction matrix. After preprocessing, the target passenger interaction vehicle model perspective preference matrix is ​​generated. The target passenger's travel information, target passenger travel time period, target passenger carpooling record, and target passenger historical evaluation constitute the target passenger comprehensive information. After preprocessing, the target passenger comprehensive information is then constructed through graph construction or matrix construction to generate the target passenger travel hotspot perspective preference graph, target passenger active time perspective preference matrix, target passenger carpooling behavior perspective preference matrix, and target passenger evaluation history perspective preference matrix; The target passenger interaction vehicle model perspective preference matrix, the target passenger active time perspective preference matrix, the target passenger carpooling behavior perspective preference matrix, and the target passenger evaluation history perspective preference matrix are respectively multiplied by their weights and then added to obtain the target passenger comprehensive perspective preference, and the preference matrix with the largest weight among the target passenger interaction vehicle model perspective preference matrix, the target passenger active time perspective preference matrix, the target passenger carpooling behavior perspective preference matrix, and the target passenger evaluation history perspective preference matrix is ​​obtained, and the target passenger comprehensive perspective preference and the target passenger initial information are weighted and summed to obtain the target passenger basic information; Target unmanned driving model basic information acquisition module: obtain the comprehensive information of the target unmanned driving model, including the driving record and running time information of the target model, take the departure place and destination as nodes, and the running time as the edge, generate the target model travel heat location map, count the running frequency of the target unmanned driving model in different time periods, sort them from high to low to form a matrix, take the data in the matrix with the running frequency not less than the running frequency threshold to form the intelligent driving high-frequency running time matrix, add the intelligent driving high-frequency running time matrix and the initial information of the target unmanned driving model to obtain the weight of the intelligent driving optimization information, and after weighted summation, obtain the basic information of the target unmanned driving model; Hybrid graph construction and update module: Based on the target passenger travel hotspot perspective preference graph and the target driverless vehicle model travel heat location graph, the passenger, vehicle model, departure place, and destination are used as nodes, and the vehicle model used by the target passenger, the ride time, and the target driverless vehicle running time are used as edges to construct a hybrid graph structure. The node representation is updated through the message passing mechanism. The basic information of the target passenger and the basic information of the target driverless vehicle are updated according to the updated nodes to obtain the comprehensive information of the target passenger and the comprehensive information of the target driverless vehicle. Unmanned driving vehicle model recommendation module: multiply the comprehensive information of the target passenger and the comprehensive information of the target unmanned driving vehicle model to obtain the fusion information of the target passenger model; The target passenger vehicle model fusion information is normalized and calculated to obtain the target driverless vehicle model recommendation score, which is compared with the recommendation threshold. If the recommendation score is greater than or equal to the recommendation threshold, the target driverless vehicle model is recommended to the target passenger.

10. An unmanned driving vehicle model recommendation device based on passenger preference fusion, characterized in that: The method comprises a processor and a memory, and when the processor executes a stored computer program, the method for recommending an unmanned driving vehicle model based on passenger preference fusion as described in any one of claims 1 to 8 is implemented.