Airline ticket business management system and method and storage medium

Abstract

The invention discloses an air ticket business management system and method and a storage medium, and the system comprises the following steps: an information collection module which is used for collecting the historical travel data of a passenger, and extracting behavior variable information from the historical travel data; the behavior information analysis module is used for inputting the behavior variable information into a behavior analysis model based on a multi-layer perceptron network and outputting a first distribution mode; and the emergency degree analysis module is used for acquiring the latest travel data, acquiring a passenger priority index from the latest travel data, constructing a decision tree model based on the priority, and acquiring a second distribution mode through the decision tree model. The limitation that a traditional seat allocation algorithm allocates seats only based on a single factor or a small number of factors is avoided, the seats can be allocated to appropriate passengers more accurately, and the reasonability of seat allocation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ticket information management, and in particular to an airline ticket management system, method and storage medium. Background Art

[0002] With the continuous growth of the demand for air travel, more and more people choose to travel by air. Whether it is for business trips or leisure travel, the convenience of air transportation makes it an important means of transportation. This trend has led to an increasing number of flights by airlines, a denser route network, and increasingly fierce market competition.

[0003] In this competitive environment, airlines need to continuously improve their service quality and operational efficiency to attract more passengers and maintain market competitiveness. As one of the key links in airline operations, ticket management directly affects passengers' ticket-buying experiences, flight occupancy rates, and airlines' revenues. Therefore, how to optimize the ticket management system has become an important issue for airlines.

[0004] Currently, airline ticket management systems usually adopt relatively static seat allocation algorithms. After seats are allocated in the early ticket-booking stage, and then some seats that are refunded due to itinerary changes are reallocated. Even if the system can obtain refund information and reallocate seats, in some cases, these seats cannot be effectively utilized. For example, if there are many refunding passengers on a certain flight and the system allocates these seats to other flights, but due to insufficient consideration of factors such as passengers' itinerary connections and transfer requirements during the allocation process, some passengers cannot smoothly take the reallocated flight, resulting in the seats being idle again. In addition, when there is a large deviation between the actual number of passengers on a flight and the predicted result, the system lacks an effective dynamic adjustment mechanism and cannot optimize seat allocation in a timely manner. Therefore, an airline ticket management system, method and storage medium are proposed herein. Summary of the Invention

[0005] The purpose of the present invention is to solve the drawback in the prior art that when there are many refunded tickets on a flight and the system allocates these seats to other flights, due to insufficient consideration of key factors such as passengers' itinerary connections and transfer requirements, some passengers cannot smoothly take the reallocated flight, resulting in the seats being idle again and causing waste of resources, and to propose an airline ticket management system, method and storage medium.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An airline ticket management system, method and storage medium:

[0008] The above technical solution further includes: An airline ticket management system, comprising:

[0009] Information collection module: used to collect passengers' historical itinerary data and extract behavioral variable information from the historical itinerary data;

[0010] Behavior information analysis module: used to input the behavioral variable information into a behavior analysis model based on a multi-layer perceptron network. The number of input nodes and output nodes of the behavior analysis model is obtained through the behavioral variables, and the first allocation mode is output;

[0011] Emergency level analysis module: used to obtain the latest itinerary data, obtain the passenger priority index from the latest itinerary data, construct a decision tree model based on the priority, and obtain the second allocation mode through the decision tree model;

[0012] Allocation output module: used to use the first allocation mode and the second allocation mode as the state space of an improved Q-learning algorithm, and output the final allocation mode of the ticket through the improved Q-learning algorithm.

[0013] The process of obtaining the behavioral variable information is as follows:

[0014] Collect the historical itinerary data of passengers from the airline's ticket database, normalize the collected historical itinerary data, remove missing values, perform feature encoding on the normalized historical itinerary data to obtain itinerary variable data, and extract variable information related to passengers' behavior based on the itinerary variable data;

[0015] The variable information related to the passengers' behavior includes but is not limited to the passengers' travel frequency (calculated according to the number of historical itineraries), average transfer time, historical selection preferences for different seat types (window seat, aisle seat, etc.), and travel time preferences (such as preferring to travel in the morning, afternoon or evening);

[0016] Let the set of extracted behavioral variable information be V = {v 1 , v 2 , v 3 ,..., v m}, where m represents the number of behavioral variable information.

[0017] The steps for obtaining the first allocation mode are as follows:

[0018] Construct a behavior analysis model based on a multi-layer perceptron network;

[0019] The behavior analysis model based on the multi-layer perceptron network includes an input layer, multiple hidden layers and an output layer;

[0020] The construction process of the behavior analysis model based on the multi-layer perceptron network is as follows:

[0021] Define the number of input nodes in the input layer. The number of input nodes is determined according to the number of behavioral variable information extracted in the information collection module, that is, the number of input nodes is m, and each node corresponds to a behavioral variable;

[0022] Define the number of output nodes in the output layer. The number of output nodes is the number of types of classification type features of the behavioral variable information, and the number of output nodes is set to K;

[0023] Construct the hidden layer;

[0024] Randomly initialize the weights of each hidden layer in the multi-layer perceptron network;

[0025] Let the weight of any node in the L-th hidden layer be initialized as w L , and let the bias of the L-th hidden layer be initialized as b L , and the number of hidden layers is M;

[0026] The construction process of the first hidden layer is as follows: Among them, represents the output result of the j-th node in the first hidden layer, represents the weight from the i-th node in the input layer to the j-th node in the first hidden layer, b 1 represents the bias of the first hidden layer, and f represents the activation function;

[0027] Construct the subsequent L-th hidden layer (L>1). The construction process of the L-th hidden layer is as follows: Among them, represents the output result of the j-th node in the L-th hidden layer, represents the output of the k-th point in the previous hidden layer of the L-th layer (reflecting the connection and information transfer between layers of the multi-layer perceptron network), represents the weight from the k-th node in the previous hidden layer of the L-th layer to the j-th node in the L-th layer, b L represents the bias of the L-th hidden layer;

[0028] Furthermore, the output of the j-th node in the first hidden layer is expressed as The output of the j-th node in the L-th hidden layer is expressed as Among them, the L-th layer refers to any hidden layer except the first layer, and L>1;

[0029] Construct the output layer. The construction process of the output layer is as follows: Among them, represents the output of the k-th node in the last hidden layer, represents the weight from the k-th node in the last hidden layer to the j-th node in the output layer, b M represents the bias of the last hidden layer;

[0030] Train the model through the cross - entropy loss function to obtain the behavior analysis model;

[0031] For the trained model, after inputting the behavior variable information, the probability distribution of the optimal output behavior variable combination is obtained through forward propagation, and this probability distribution is the first allocation mode.

[0032] The process of obtaining the passenger priority index is as follows:

[0033] Obtain the latest itinerary data of passengers from the real - time ticket data of the airline. The latest itinerary data includes but is not limited to the current booked flight information (departure place, destination, departure time, arrival time, etc.), itinerary change situations (such as whether there are ticket refunds, itinerary changes, etc.), remaining travel time (the interval from the current time to the flight departure time), membership level update situations, etc. Let the set of the obtained latest itinerary data be U = {u 1 , u 2 , u 3 ,..., u p}, where u p represents the latest itinerary data record of the p - th passenger;

[0034] Based on the obtained latest itinerary data, consider travel factors to calculate the passenger priority index. The travel factors include membership level factor: Assume that the membership level is divided into multiple levels. For the 1st passenger, its membership level is r 1 , and different weights are assigned according to the membership level. The weight of the high - level membership is higher, and the weight of the low - level membership is lower;

[0035] Itinerary urgency factor: Measure the itinerary urgency according to the remaining travel time. Let the remaining travel time be t;

[0036] Itinerary change situation factor: Itinerary change operations of passengers such as ticket refunds and itinerary changes will affect their priority. The priority of passengers with frequent itinerary changes will be reduced, and h is deducted for each itinerary change;

[0037] Considering the comprehensive travel factors, let the priority index be Z, and the calculation formula is expressed as: Z = w r + t - h.

[0038] The process of constructing the decision - tree model is as follows:

[0039] Use each feature in the latest itinerary data U and the priority index Z as the basis for constructing the decision tree. Use the membership level, remaining travel time, and itinerary change times in the latest itinerary data as the input features of the decision - tree model, and the passenger priority index as the target variable of the decision - tree model, and calculate the information gain ratio of each input feature to select the best splitting feature;

[0040] Evaluate the classification ability of the input feature (the latest trip data U) with respect to the target variable (the priority index Z);

[0041] The process of evaluating the classification ability is as follows:

[0042] Obtain one of the input features U g with the number of classes g, where 1 < g < p, and obtain the total number of classes d of the target variable Z

[0043] Based on the number of classes g and the total number of classes calculate the chi-square statistic index of the input feature U g with respect to the priority index Z. The calculation formula is: where, O gd is the observed frequency (the number of samples where the input feature U g belongs to the target variable Z d ), and E gd is the expected frequency (assuming the input feature and the target variable are independent, the number of samples where the input feature U g belongs to the target variable Z d );

[0044] Find the input feature U g and its target variable Z d that maximize the chi-square statistic index, and use this as the division basis for the current node (which means selecting the feature and grouping that can most significantly distinguish the passenger priority index to construct the decision tree). Starting from the root node, divide the data set according to the input feature and its target variable that maximize the chi-square value, and recursively repeat the above chi-square statistic calculation and node division process for each child node until the stopping condition is met;

[0045] The stopping conditions include that all samples belong to the same class (i.e., the priority indices of all passengers in the subset are in the same interval), there are no features left for division (i.e., all features have been used on the current path), and the number of samples in the node is less than the preset minimum number of samples (to prevent overfitting, the minimum number of samples included in the node is preset in advance);

[0046] When the latest trip data of a passenger is input, use this data as the input to the decision tree. The decision tree path search starts from the root node and gradually searches down the decision tree path according to the values of the features in the latest trip data. When reaching the leaf node of the decision tree, the output corresponding to the leaf node is the second allocation mode of the passenger.

[0047] Further, the second allocation mode means continuously selecting branches according to the characteristic values in the latest itinerary data until reaching the leaf node. The leaf node corresponds to an allocation strategy. For example, the leaf node may correspond to "prioritizing a certain behavior", that is, preferentially allocating seats to customers with less remaining travel time and a high membership level. Then this is the second allocation mode for this passenger, which is also the result of the path output.

[0048] The observed frequency O gd and the expected frequency E gd are obtained as follows:

[0049] The observed frequency O gd Based on the itinerary data, determine the input features and the target variable, and count the number of samples where the input features belong to the target variable. This quantity is the observed frequency;

[0050] The expected frequency E gd is the number of samples where the input features belong to the target variable when the input features and the target variable are independent;

[0051] Obtain the number of categories of the input feature U g and the total number of categories of its target variable Z d Based on the number of categories g and the total number of categories Obtain through the formula where, represents the total number of samples of the target variable Z d in all relevant input feature categories, represents the input feature U g in the relevant target variable categories.

[0052] The steps to obtain the improved Q-learning algorithm are as follows:

[0053] Take the first allocation mode output by the behavior information analysis module and the second allocation mode output by the emergency level analysis module as the state space of the Q-learning algorithm. Let the first allocation mode be s 1 , and the second allocation mode be s 2 . A state in the state space can be represented as S=(s 1 , s 2 ). Define an action space, take all possible ticket seat allocation strategies as an action space, denoted as A, and construct an optimized reward function;

[0054] The steps to construct the optimized reward function are as follows:

[0055] Add a dynamic function to the reward function;

[0056] The operating environment of flights is dynamically changing. For example, the load factor of flights, departure time, weather conditions, etc. will all affect the optimal strategy for ticket seat allocation. The current reward function does not fully consider these dynamic factors;

[0057] Collect dynamic environment data of flights, including the load factor zk, departure time (converted to time category) qf, and weather conditions tq (coded into categories);

[0058] Design a dynamic function f zk Quantify the impact of the load factor on the reward function: f zk = a * zk, where a is the impact coefficient of the load factor;

[0059] Design a dynamic function f qf Quantify the impact of the departure time on the reward function: f qf = β * ∑ p qf, where β is the impact coefficient of the departure time, and ∑ p qf is the preference for buying window seats at all different departure times;

[0060] Design a dynamic function f tq Quantify the impact of the weather conditions on the reward function: f tq = γ * ∑ p tq, where γ is the impact coefficient of the weather conditions, and ∑ p tq is the preference for buying window seats under all different weather conditions;

[0061] The original reward function is expressed as: R = σu + (1 - σ)c, where u is the occupancy rate and c is the passenger satisfaction. The optimized reward function is

[0062] Initialize the action-value function Q(S,A) as a matrix with an initial value of 0. Select an action A (seat allocation strategy) according to the state S = (s 1 , s 2 ). Randomly select an action with probability ∈ and select the action that maximizes Q max (S,A) with probability 1 - ∈. After executing the action, observe the next state and the reward Update the action-value function: where represents the update degree of the original value by the newly obtained information, and δ is the discount factor;

[0063] Define an experience replay buffer. Each interaction obtains (Current state, executed action, obtained reward, next state) are all stored in the experience replay buffer. By iterating the above process multiple times, the Q-value is continuously updated and learning is carried out using the data in the experience replay buffer, and finally the optimized Q-learning algorithm is obtained.

[0064] The process of obtaining the final allocation mode is as follows:

[0065] According to the trained improved Q-learning algorithm, in the current state, select the action with the maximum value. This action is the optimal seat allocation action. Execute this action to allocate the passenger to the corresponding seat, and this seat allocation method is the final allocation mode of the ticket.

[0066] An airline ticket management method includes the following steps:

[0067] S1: Collect passenger historical itinerary data and extract behavioral variable information from the historical itinerary data;

[0068] S2: Input the behavioral variable information into a behavioral analysis model based on a multi-layer perceptron network. The number of input nodes and output nodes of the behavioral analysis model is obtained through the behavioral variables, and the first allocation mode is output;

[0069] S3: Obtain the latest itinerary data, obtain the passenger priority index from the latest itinerary data, construct a decision tree model based on the priority, and obtain the second allocation mode through the decision tree model;

[0070] S4: Use the first allocation mode and the second allocation mode as the state space of an improved Q-learning algorithm, and output the final allocation mode of the ticket through the improved Q-learning algorithm.

[0071] The present invention has the following beneficial effects:

[0072] 1. In the present invention, the information collection module collects passenger historical itinerary data and the latest itinerary data, comprehensively considering the passenger's behavioral habits (such as travel frequency, average transfer time, seat preference, travel time preference, etc.), real-time itinerary information (such as remaining travel time, itinerary change situation, membership level, etc.). Based on this rich information, the behavioral information analysis module and the urgency analysis module respectively output the first allocation mode and the second allocation mode, and then the final allocation mode is obtained after optimization by the improved Q-learning algorithm. This multi-factor comprehensive decision-making method avoids the limitations of traditional seat allocation algorithms that only allocate seats based on single or a small number of factors, can more accurately allocate seats to suitable passengers, and improves the rationality of seat allocation.

[0073] 2. In the present invention, the constructed optimized reward function takes into account the dynamic environmental data of flights, such as occupancy rate, departure time, weather conditions, etc. The changes in these factors will affect the ticket seat allocation strategy in real time, enabling the system to flexibly adjust according to the actual operation situation. For example, when the occupancy rate is high, the system may focus more on increasing the load factor; in bad weather, the seat allocation may be adjusted according to passenger preferences to improve passenger comfort. This dynamic adjustment ability enables the system to better handle various complex and changeable situations and adapt to different operating environments. At the same time, the system continuously learns and improves through the experience replay mechanism. Each time the interaction data for managing ticket seat allocation is stored and utilized to update the Q value. As time goes by, the system can continuously optimize the seat allocation strategy to make it more in line with the actual needs, further improving the accuracy of the ticketing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a system block diagram of an airline ticketing management system proposed by the present invention.

[0075] Figure 2 It is a method step diagram of an airline ticketing management method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0077] Embodiment 1

[0078] As Figure 1 shown, an airline ticketing management system proposed in this embodiment includes:

[0079] An information collection module: used to collect historical travel data of passengers and extract behavioral variable information from the historical travel data;

[0080] The process of obtaining behavioral variable information is as follows:

[0081] Collect historical travel data of passengers from the ticketing database of the airline, perform normalization processing on the collected historical travel data to remove missing values, perform feature encoding on the normalized historical travel data to obtain travel variable data, and extract variable information related to passenger behavior based on the travel variable data;

[0082] Specifically, in the historical travel data after normalization and missing value processing, perform feature encoding to determine those features that belong to the categorical type;

[0083] For example, classification features include the ticketing channels (official website, travel agency, third-party platform, etc.), whether there is a layover (yes, no), checked baggage situation (yes, no), etc. For each classification feature, each possible value is converted into a binary vector. For example, for the ticketing channel feature, if there are three values (official website, travel agency, third-party platform), such encoding operations are performed on all classification features to convert the original categorical data into a numerical form that can be processed by a computer for subsequent data analysis and model training. In this way, the features of the entire dataset are presented in a unified numerical variable format;

[0084] Variable information related to passenger behavior includes, but is not limited to, the passenger's travel frequency (calculated based on the number of historical trips), average layover time, historical selection preferences for different seat types (window, aisle, etc.), and travel time preferences (such as preferring to travel in the morning, afternoon, or evening);

[0085] Let the set of extracted behavioral variable information be V = {v 1 ,v 2 ,v 3 ,...,v m}, where m represents the number of behavioral variable information.

[0086] Behavior information analysis module: used to input the behavioral variable information into a behavior analysis model based on a multi-layer perceptron network. The number of input nodes and output nodes of the behavior analysis model is obtained through the behavioral variables, and the first allocation mode is output;

[0087] The steps for obtaining the first allocation mode are as follows:

[0088] Construct a behavior analysis model based on a multi-layer perceptron network;

[0089] The behavior analysis model based on a multi-layer perceptron network includes an input layer, multiple hidden layers, and an output layer;

[0090] The construction process of the behavior analysis model based on a multi-layer perceptron network is as follows:

[0091] Define the number of input nodes in the input layer. The number of input nodes is determined according to the number of behavioral variable information extracted in the information collection module, that is, the number of input nodes is m, and each node corresponds to a behavioral variable;

[0092] Define the number of output nodes in the output layer. The number of output nodes is the number of types of classification type features of the behavioral variable information, and the number of output nodes is set to K;

[0093] Construct the hidden layer;

[0094] Randomly initialize the weights of each hidden layer in the multi-layer perceptron network;

[0095] Let the weight of any node in the L-th hidden layer be initialized as w L Let the bias of the L-th hidden layer be initialized as b L The number of hidden layers is M;

[0096] The construction process of the first hidden layer is as follows: Among them, represents the output result of the j-th node in the first hidden layer, represents the weight from the i-th node in the input layer to the j-th node in the first hidden layer, b 1 represents the bias of the first hidden layer, and f represents the activation function;

[0097] Construct the subsequent L-th hidden layer (L>1). The construction process of the L-th hidden layer is as follows: Among them, represents the output result of the j-th node in the L-th hidden layer, represents the output of the k-th point in the previous hidden layer of the L-th layer (reflecting the connection and information transfer between layers in the multi-layer perceptron network), represents the weight from the k-th node in the previous hidden layer of the L-th layer to the j-th node in the L-th layer, b L represents the bias of the L-th hidden layer;

[0098] Specifically, the output of the j-th node in the first hidden layer is expressed as The output of the j-th node in the L-th hidden layer is expressed as Among them, the L-th layer refers to any hidden layer other than the first layer, and L>1;

[0099] Construct the output layer. The construction process of the output layer is as follows: Among them, represents the output of the k-th node in the last hidden layer, represents the weight from the k-th node in the last hidden layer to the j-th node in the output layer, b M represents the bias of the last hidden layer;

[0100] Train the model through the cross-entropy loss function to obtain the behavior analysis model;

[0101] For the trained model, after inputting the behavior variable information, the probability distribution of the optimal behavior variable combination is obtained through forward propagation. This probability distribution is the first allocation mode.

[0102] Among them, after inputting the information of the behavior variable, the output probability distribution is obtained through forward propagation. Here, the forward propagation process refers to the calculation process from the input layer through the hidden layer to the output layer. The input layer inputs the information of the behavior variable, the hidden layer performs weighted summation on the input and processes it through an activation function, and the output layer converts the output of the hidden layer into a probability distribution to obtain the probability values of each category, thereby determining the first allocation mode;

[0103] Specifically, through the output layer, the output of the last hidden layer can be converted into a probability distribution corresponding to each category of the first allocation mode, so as to realize the prediction of the first allocation mode. For example, if the output layer has 4 nodes (corresponding to 4 seat types), after being processed by the hidden layer and the output layer, the 4 probability values obtained respectively represent the possibility that the passenger is assigned to these 4 seat types.

[0104] Emergency level analysis module: used to obtain the latest itinerary data, obtain the passenger priority index from the latest itinerary data, build a decision tree model based on the priority, and obtain the second allocation mode through the decision tree model;

[0105] The process of obtaining the passenger priority index is as follows:

[0106] Obtain the latest itinerary data of the passenger from the real-time ticketing data of the airline. The latest itinerary data includes but is not limited to the currently booked flight information (departure place, destination, departure time, arrival time, etc.), itinerary change situation (such as whether there are operations such as ticket refund and flight change), remaining travel time (the interval from the current time to the flight departure time), membership level update situation, etc. Let the set of the obtained latest itinerary data be U = {u 1 ,u 2 ,u 3 ,...,u p}, where u p represents the latest itinerary data record of the p-th passenger;

[0107] Based on the obtained latest itinerary data, consider the travel factors to calculate the passenger priority index. The travel factors include membership level factor: assume that the membership level is divided into multiple levels. For the first passenger, his membership level is r 1 ,and different weights are assigned according to the membership level. The weight of the senior membership level is higher, and the weight of the junior membership level is lower;

[0108] Itinerary emergency level factor: measure the itinerary emergency level according to the remaining travel time. Let the remaining travel time be t;

[0109] Itinerary change situation factor: The itinerary change operations such as passenger ticket refund and flight change will affect their priority. The priority of passengers with frequent flight changes will be reduced, and h will be deducted for each flight change;

[0110] Considering comprehensive travel factors, let the priority index be Z, and its calculation formula is expressed as: Z = w r + t - h.

[0111] The construction process of the decision tree model is as follows:

[0112] Take each feature in the latest trip data U and the priority index Z as the basis for constructing the decision tree. Use the membership level, remaining travel time, and number of itinerary changes in the latest trip data as the input features of the decision tree model, and the passenger priority index as the target variable of the decision tree model. Calculate the information gain ratio of each input feature to select the best splitting feature;

[0113] Evaluate the classification ability of the input features (latest trip data U) for the target variable (priority index Z);

[0114] The process of evaluating the classification ability is as follows:

[0115] Obtain the number of categories g of one of the input features U g , where 1 < g < p, and obtain the total number of categories of the target variable Z d ;

[0116] Based on the number of categories g and the total number of categories calculate the chi-square statistic index of the input feature U g for the priority index Z. The calculation formula is: where O gd is the observed frequency (the number of samples where the input feature U g belongs to the target variable Z d ), and E gd is the expected frequency (assuming the input feature and the target variable are independent, the number of samples where the input feature U g belongs to the target variable Z d );

[0117] Find the input feature U g and its target variable Z d that maximize the chi-square statistic index. Use this as the splitting basis for the current node (which means selecting the feature and grouping that can most significantly distinguish the passenger priority index to construct the decision tree). Starting from the root node, divide the data set according to the input feature and its target variable that maximize the chi-square value, and recursively repeat the above chi-square statistic calculation and node splitting process for each child node until the stopping condition is met;

[0118] The stopping conditions include that all samples belong to the same class (i.e., the priority indices of all passengers in the subset are in the same interval), there are no more features to divide (i.e., all features have been used on the current path), and the number of samples in the node is less than the preset minimum number of samples (to prevent overfitting, the minimum number of samples contained in the node is preset in advance).

[0119] When the latest trip data of a passenger is input, these data are used as the input of the decision tree. The decision tree path search starts from the root node according to the branch situation of the decision tree, and gradually searches downward along the values of the features in the latest trip data. When reaching the leaf node of the decision tree, the path output corresponding to the leaf node is the second allocation mode of the passenger.

[0120] Specifically, the second allocation mode means continuously selecting branches according to the feature values in the latest trip data until reaching the leaf node. The leaf node corresponds to an allocation strategy. For example, the leaf node may correspond to "giving priority to a certain behavior", that is, giving priority to allocating seats to customers with less remaining travel time and a high membership level. Then this is the second allocation mode of the passenger, which is also the result of the path output.

[0121] Observed frequency O gd and expected frequency E gd are obtained as follows:

[0122] Observed frequency O gd Based on the trip data, the input features and the target variable are determined, and the number of samples in which the input features belong to the target variable is counted. This number is the observed frequency;

[0123] Expected frequency E gd is the number of samples in which the input features belong to the target variable when the input features and the target variable are independent;

[0124] The number of categories of the input feature U g and the total number of categories of its target variable Z d are obtained. Based on the number of categories g and the total number of categories through the formula where, represents the total number of samples of the target variable Z d in all relevant input feature categories, represents the total number of samples of the input feature U g in the relevant target variable categories;

[0125] Specifically, through the observed frequency and the expected frequency, the chi-square statistical index can be further calculated to evaluate the classification ability of the input features for the target variable (priority index).

[0126] Allocation Output Module: It is used to take the first allocation mode and the second allocation mode as the state space of an improved Q-learning algorithm, and output the final allocation mode of tickets through the improved Q-learning algorithm.

[0127] The obtaining steps of the improved Q-learning algorithm are as follows:

[0128] Take the first allocation mode output by the behavior information analysis module and the second allocation mode output by the emergency degree analysis module as the state space of the Q-learning algorithm. Let the first allocation mode be s 1 , and the second allocation mode be s 2 . A state in the state space can be represented as S=(s 1 , s 2 ). Define an action space, take all possible ticket seat allocation strategies as an action space, denoted as A, and construct an optimized reward function;

[0129] The construction steps of the optimized reward function are as follows:

[0130] Add a dynamic function to the reward function;

[0131] The operating environment of the flight is dynamically changing. For example, the occupancy rate of the flight, departure time, weather conditions, etc. will all affect the optimal strategy for ticket seat allocation. The current reward function does not fully consider these dynamic factors;

[0132] Collect the dynamic environment data of the flight, including the occupancy rate zk, departure time (converted to time category) qf, and weather conditions tq (coded into categories);

[0133] Design a dynamic function f zk Quantify the impact of the occupancy rate on the reward function: f zk =a*zk, where a is the impact coefficient of the occupancy rate;

[0134] Design a dynamic function f qf Quantify the impact of the departure time on the reward function: f qf =β*∑ p qf, where β is the impact coefficient of the departure time, and ∑ p qf is the preference for buying window seats at all different departure times;

[0135] Design a dynamic function f tq Quantify the impact of the weather conditions on the reward function: f tq =γ*∑ p tq, where γ is the impact coefficient of the weather conditions, and ∑ p tq is the preference for buying window seats under all different weather conditions;

[0136] The original reward function is expressed as: R = σu + (1 - σ)c, where u is the occupancy rate and c is the passenger satisfaction. After optimization, the reward function is

[0137] Initialize the action-value function Q(S,A) as a matrix with an initial value of 0. According to the state S=(s 1 , s 2 ), select the action A (seat allocation strategy). Randomly select an action with probability ∈ and select the action that maximizes Q max (S,A) with probability 1 - ∈. After executing the action, observe the next state and the reward Update the action-value function: where represents the degree of update of the new acquired information to the original value, and δ is the discount factor;

[0138] Define an experience replay buffer. Each interaction's (current state, executed action, obtained reward, next state) is stored in the experience replay buffer. By iterating the above process multiple times, continuously update the Q value and learn using the data in the experience replay buffer, and finally obtain the optimized Q-learning algorithm.

[0139] The process of obtaining the final allocation mode is as follows:

[0140] According to the trained improved Q-learning algorithm, in the current state, select the action with the maximum value. This action is the optimal seat allocation action. Execute this action to allocate passengers to the corresponding seats, and this seat allocation method is the final allocation mode of the ticket.

[0141] In this embodiment, by obtaining the first allocation mode and the second allocation mode, the state space S is constructed. Here, the allocation mode is for seat allocation in the ticket system. In the ticket management system, seat allocation is the core function. The system needs to consider various factors to determine how to allocate seats for passengers, including passenger priority, flight occupancy rate, passenger preferences, etc. According to the trained improved Q-learning algorithm, in the current state space S, select the action A with the maximum Q value. The application of the improved Q-learning algorithm in the ticket management system is to optimize the seat allocation strategy, which can find the optimal seat allocation plan according to various factors (such as passenger behavior, flight dynamics, etc.). This is the embodiment of the intelligent management of seat allocation in the ticket management system. Execute this action A to allocate passengers to the corresponding seats, and this seat allocation method is the final allocation mode of the ticket. The output of the entire process is the seat allocation result of the ticket, which is the ultimate goal of the ticket management system. According to the input and algorithm optimization, provide reasonable seat allocation for passengers.

[0142] Example 2

[0143] As Figure 2 shown, based on Example 1, this example proposes an air ticket management method, including the following steps:

[0144] S1: Collect passengers' historical itinerary data and extract behavioral variable information from the historical itinerary data;

[0145] S2: Input the behavioral variable information into a behavioral analysis model based on a multi-layer perceptron network. The number of input nodes and output nodes of the behavioral analysis model is obtained through the behavioral variables, and the first allocation mode is output;

[0146] S3: Obtain the latest itinerary data, obtain the passenger priority index from the latest itinerary data, construct a decision tree model based on the priority, and obtain the second allocation mode through the decision tree model;

[0147] S4: Use the first allocation mode and the second allocation mode as the state space of an improved Q-learning algorithm, and output the final allocation mode of the tickets through the improved Q-learning algorithm.

[0148] Example 3

[0149] In this example, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program runs, it implements the operation mode of the above-provided air ticket management system.

[0150] Since the storage medium introduced in this example is the storage medium used in implementing the air ticket management system in the embodiments of the present application, based on the air ticket management system introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manner and various variations of the storage medium in this example. Therefore, the specific implementation of how this storage medium implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the storage medium used in the air ticket management system in the embodiments of the present application, it falls within the scope of protection of the present application.

[0151] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0152] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An airline ticket management system, characterized in that: The steps include: Information collection module: used to collect passengers' historical travel data and extract behavioral variable information from the historical travel data; Behavior information analysis module: used to input behavior variable information into a behavior analysis model based on a multi-layer perceptron network, wherein the number of input nodes and output nodes of the behavior analysis model is obtained through the behavior variables, and a first allocation mode is output; Urgency analysis module: used to obtain the latest travel data, obtain the passenger priority index from the latest travel data, build a decision tree model based on the priority index, and obtain the second allocation mode through the decision tree model; Allocation output module: used to use the first allocation mode and the second allocation mode as a state space of an improved Q learning algorithm, and output the final allocation mode of the ticket through the improved Q learning algorithm.

2. An airline ticket management system according to claim 1, characterized in that: The process of obtaining the behavior variable information is as follows: Collect historical itinerary data of passengers from the airline's ticketing database, normalize the collected historical itinerary data, perform feature encoding on the processed historical itinerary data, obtain itinerary variable data, and extract variable information related to passenger behavior based on the itinerary variable data; The variable information related to the passenger behavior includes the passenger's travel frequency, average transfer time, historical selection preference for different seat types, and travel time preference; Suppose the extracted behavior variable information set is V = {v1, v2, v3, ..., v m }, where m represents the number of behavioral variable information.

3. An airline ticket management system according to claim 1, characterized in that: The steps of obtaining the first allocation mode are: Build a behavior analysis model based on a multi-layer perceptron network; The behavior analysis model based on the multi-layer perceptron network includes an input layer, multiple hidden layers and an output layer; The construction process of the behavior analysis model based on the multi-layer perceptron network is as follows: Define the number of input nodes in the input layer. The number of input nodes is determined according to the number of behavioral variable information extracted in the information acquisition module. That is, the number of input nodes is m, and each node corresponds to a behavioral variable. Define the number of output nodes in the output layer. The number of output nodes is the number of classification type features of the behavior variable information. The number of output nodes is set to K; The construction process of the multiple hidden layers is as follows: Randomly initialize the weights of each hidden layer in the multi-layer perceptron network; Assume that the weight of any node in the Lth hidden layer is initialized to w L , let the bias of the Lth hidden layer be initialized to b L , the number of hidden layers is M; The construction process of the first hidden layer is: in, represents the output result of the jth node in the first hidden layer, represents the weight from the i-th node in the input layer to the j-th node in the first hidden layer, b1 represents the bias of the first hidden layer, and f represents the activation function; Construct the subsequent Lth hidden layer (L>1), the construction process of the Lth hidden layer is: in, represents the output result of the jth node of the Lth hidden layer, represents the output of the kth point of the hidden layer before the Lth layer, represents the weight from the kth node in the previous hidden layer of the Lth layer to the jth node in the Lth layer, b L represents the bias of the Lth hidden layer; The process of constructing the output layer is: The construction process of the output layer is: in, represents the output of the kth node in the last hidden layer, represents the weight from the kth node in the last hidden layer to the jth node in the output layer, b M represents the bias of the last hidden layer; The model is trained through the cross entropy loss function to obtain the behavior analysis model. After the behavior variable information is input, the probability distribution of the output optimal behavior variable combination is obtained through the forward propagation of the behavior analysis model. This probability distribution is the first allocation mode.

4. The airline ticket management system according to claim 1, characterized in that: The process of obtaining the passenger priority index is as follows: The latest itinerary data of the passenger is obtained from the real-time ticketing data of the airline. The latest itinerary data includes the currently booked flight data, itinerary change data, remaining travel time, and membership level update data. Suppose the latest itinerary data set obtained is U = {u1, u2, u3, ..., u p }, where u p Represents the latest travel data record of the pth passenger; Based on the latest acquired itinerary data, the passenger's priority index Z is calculated by taking into account travel factors, including membership level factors, itinerary urgency factors, and comprehensive travel factors.

5. The airline ticket management system according to claim 1, characterized in that: The construction process of the decision tree model is: The features and priority index in the latest itinerary data are used as the basis for building a decision tree. The membership level, remaining travel time, and number of changes in the latest itinerary data are used as input features of the decision tree model, and the passenger priority index is used as the target variable of the decision tree model. The information gain ratio of each input feature is calculated to select the best partitioning feature. Evaluate the classification ability of input features on the target variable; The process of evaluating classification ability is: Obtain one of the input features U g The number of categories g, where 1 < g < p, and obtain the target variable Z d The total number of categories Based on the number of categories g and the total number of categories Calculate the input feature U g The chi-square statistical index of the priority index Z is calculated as follows: Among them, O gd is the observed frequency, E gd is the expected frequency; Find the input feature U that maximizes the chi-squared statistic g and its target variable Z d , using this as the basis for dividing the current node, starting from the root node, divide the data set according to the input features and target variables that maximize the chi-square value, recursively repeat the above chi-square statistic calculation and node division process for each child node until the stopping condition is met, and finally obtain a decision tree model; The acquisition process of the second allocation mode is: When the latest travel data of a passenger is input, the latest travel data is used as the input of the decision tree. The decision tree path search starts from the root node according to the branching of the decision tree, and gradually searches down the path of the decision tree according to the value of the feature in the latest travel data. When the leaf node of the decision tree is reached, the path output corresponding to the leaf node is the second allocation mode of the passenger.

6. An airline ticket management system according to claim 5, characterized in that: The stopping conditions include that all samples belong to the same class, there is no feature to continue dividing, and the number of samples in the node is less than a preset minimum number of samples.

7. An airline ticket management system according to claim 5, characterized in that: The observation frequency O gd and the expected frequency E gd The acquisition process is: The observation frequency O gd Determine the input features and target variables based on the trip data, and count the number of samples where the input features belong to the target variable, i.e. the observation frequency; The expected frequency E gd is the number of samples where the input feature belongs to the target variable when the input feature is independent of the target variable; Get input feature U g The number of categories and its target variable Z d The total number of categories, based on the number of categories g and the total number of categories By formula Get, where represents the target variable Z d The total number of samples in all relevant input feature categories, Represents the input feature U g The total number of samples in the associated target variable category.

8. An airline ticket management system according to claim 1, characterized in that: The steps for obtaining the improved Q learning algorithm are: The first allocation mode output by the behavior information analysis module and the second allocation mode output by the urgency analysis module are used as the state space of the Q learning algorithm. Let the first allocation mode be s1 and the second allocation mode be s2. A state in the state space can be expressed as S = (s1, s2). Define an action space, take all possible ticket seat allocation strategies as an action space, express it as A, and construct an optimization reward function. The steps for constructing the optimization reward function are: Add a dynamic function to the reward function; Collect dynamic environmental data of the flight, including passenger load factor zk, take-off time qf, and weather conditions tq; Design a dynamic function f zk Quantify the impact of load factor on the reward function: f zk =a*zk, where a is the influence coefficient of passenger load factor; Design a dynamic function f qf Quantify the impact of takeoff time on the reward function: f qf =β*∑ p qf, where β is the influence coefficient of take-off time, ∑ p qf is the preference for buying window tickets at all different departure times; Design a dynamic function f tq Quantify the impact of weather conditions on the reward function: f tq =γ*∑ p tq, where γ is the influence coefficient of weather conditions, ∑ p tq is the preference for buying window tickets under all different weather conditions; The original reward function is expressed as: R = σu + (1-σ) c, where u is the occupancy rate and c is the passenger satisfaction. The optimized reward function is Initialize the action-value function Q(S,A) as a matrix with the initial value set to 0. Select action A (seat allocation strategy) according to the state S = (s1, s2). Select actions randomly with probability ∈, and select the action that maximizes Q with probability 1-∈ max (S,A), after executing the action, observe the next state and rewards Update the action-value function: in, It indicates the update degree of the new information to the original value, and δ is the discount factor; Define an experience replay buffer, each interaction gets All are stored in the experience replay buffer, the Q value is updated and the data in the experience replay buffer is used for learning, and finally an optimized Q learning algorithm is obtained; The process of obtaining the final allocation mode is as follows: According to the trained improved Q-learning algorithm, in the current state, the action with the maximum value is selected, which is the optimal seat allocation action. The action is executed to allocate passengers to the corresponding seats. This seat allocation method is the final ticket allocation mode.

9. An airline ticket management method, which is implemented based on an airline ticket management system according to any one of claims 1 to 8, characterized in that: include: S1: Collect passengers’ historical travel data and extract behavioral variable information from the historical travel data; S2: inputting the behavior variable information into a behavior analysis model based on a multi-layer perceptron network, wherein the number of input nodes and output nodes of the behavior analysis model is obtained through the behavior variables, and outputting a first allocation mode; S3: Obtain the latest travel data, obtain the passenger priority index from the latest travel data, build a decision tree model based on the priority, and obtain the second allocation mode through the decision tree model; S4: The first allocation mode and the second allocation mode are used as a state space of an improved Q-learning algorithm, and the final ticket allocation mode is output through the improved Q-learning algorithm.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, an airline ticket management system according to any one of claims 1 to 7 is implemented.

Images