CN106779214B - A Multi-factor Fusion Civil Aviation Passenger Travel Prediction Method Based on Theme Model - Google Patents

Abstract

A multi-factor fusion civil aviation passenger trip prediction method based on a theme model. According to the invention, firstly, the association diagram among passengers is constructed, and subject modeling is carried out according to the preference of the passengers, so that a Passenger association diagram Travel subject Model (PGTTM) is constructed, the subject information can be enriched, and the problem of sparsity of civil aviation data can be effectively solved; secondly, a multi-factor fusion prediction framework is constructed through a Bayesian probability model, and the future travel of the passenger is accurately predicted by fusing the airline heat and PGTTM to obtain the passenger airline preference, passenger loyalty and airline market share information. The invention can effectively predict airlines and airlines of passengers going out in the future, can provide effective decision support for aviation and related industries, and provides personalized service for passengers.

Description

A Multi-factor Fusion Civil Aviation Passenger Travel Prediction Method Based on Theme Model

technical field

The invention belongs to the technical field of computer application and relates to data mining and civil aviation data analysis, in particular to a multi-factor fusion civil aviation passenger travel prediction method based on a theme model.

Background technique

With the improvement of people's living standards and the development of the Internet, a large amount of booking data has been accumulated in the passenger booking system of civil aviation, which has the characteristics of mass, sparseness and long tail, which brings challenges to the analysis of civil aviation data. Analyzing passenger travel characteristics and predicting future travel behavior based on these data is one of the most important tasks in civil aviation data analysis. Domestic and foreign research on the analysis of civil aviation passengers is in the preliminary stage, and there is not much research on the travel forecast of civil aviation passengers.

The analysis and research related to civil aviation data, such as Maalouf et al., applied cluster analysis and association rules to real airline frequent passenger data, and proposed recommendations and improvement strategies for customer relationship management ^[1] . Wang Wang et al. used a questionnaire survey combined with statistical methods to analyze the consumption motivation, airline preferences and purchasing behavior of Changchun civil aviation passenger groups ^[2] . Feng et al. constructed a heterogeneous information network on civil aviation data, and used a random walk method to find the value of low-frequency travel passengers ^[3] . Etzioni et al. explored the correlation between time and fare, and adopted a multi-strategy data mining algorithm to inform passengers of the best time to buy air tickets ^[4] .

The LDA (Latent Dirichlet Allocation) model in the topic model has better performance of text topic modeling and has good scalability ^[5] . For example, Rosen-Zvi et al. proposed ATM (Author-Topic Model) based on LDA, and performed topic modeling on authors, documents and words at the same time ^[6] . And Blei et al. proposed a supervised LDA model for the text classification problem, adding the document tags in the training corpus as observations into the LDA ^[7] . Extending the topic model or LDA model to the field of recommendation, for example, Liu et al. added the implicit feature display in the travel package data to the topic model, and proposed a personalized recommendation method of travel information ^[8] . Tan et al. expressed the passenger information in the form of feature-value pairs, used a topic model to learn the potential interest distribution of passengers, and combined with collaborative filtering to recommend travel packages ^[9] .

The social relationship between passengers is helpful for modeling. For example, Wang Kunkun et al. proposed a modeling method of passenger seat preference in civil aviation that combines individual passenger preference and relationship preference by constructing a common travel network ^[10] . Zhou Yuanwei et al. proposed a semi-supervised relationship classification algorithm based on information graphs to obtain more accurate passenger relationships and provide targeted and high-quality services ^[11] .

It is worth trying to apply the topic model to the analysis and prediction of passenger travel in civil aviation to discover the distribution of potential topics and solve the problem of massive data, and integrate the relationship between passengers into topic modeling to enrich topic information and reduce sparsity problem, so as to improve the modeling effect. In addition, by building a probabilistic model framework and integrating a variety of travel influencing factors, we will also wait and see to improve the prediction effect.

references:

[1] Maalouf L, Mansour N. Mining airline data for crm strategies. In Proceeding of the 7th WSEAS International Conference on Simulation, Modeling and Optimization, Beijing, China, pages 345-350, 2007.

[2] Wang Wangen, Analysis on the characteristics and behavior of Changchun civil aviation passengers [D]. Jilin University, 2010.

[3] Feng X, Xu B Y, Lu M, et al. Infrequent Passenger Value Discovery by Random Walk on Passenger-route Heterogeneous Network. Journal of Computational and Theoretical Nanoscience, 2(1):10-17, 2015.

[4] Etzioni, Oren, Tuchinda, et al. To buy or not to buy: mining airfaredata to minimize ticket purchase price[C]//ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington, USA, August. 2003: 119 -128.

[5] Blei D M, Ng A Y, Jordan M I. Latent dirichlet allocation[J]. Journal of Machine Learning Research, 2003, 3:993-1022.

[6] Rosen-Zvi M, Griffiths T, Steyvers M, et al. The author-topic model for authors and documents [C]//Proceedings of the 20th conference on Uncertainty in artificial intelligence. AUAI Press, 2004: 487-494.

[7] Blei D M, Mcauliffe J D.Supervised Topic Models[J].Advances inNeural Information Processing Systems,2010,3:327-332.

[8]Liu Q,Ge Y,Li Z,et al.Personalized Travel Package Recommendation[C]//IEEE,International Conference on Data Mining.IEEE Computer Society,2011:407-416.

[9]Tan C,Liu Q,Chen E,et al.Object-Oriented Travel PackageRecommendation[J].Acm Transactions on Intelligent Systems&Technology,2014,5(3):1-26.

[10] Wang Kunkun, Modeling and application of passenger seat preference in civil aviation [D]. Beijing Jiaotong University, 2015.

[11] Zhou Yuanwei, Design and Implementation of Civil Aviation Social Network Relationship Classification Algorithm [D]. Beijing Jiaotong University, 2013.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a multi-factor fusion based on a theme model in order to accurately predict the airlines and routes that passengers will take in the future in order to solve the problems of mass, sparseness, long tail, and diversity of factors affecting the travel of passenger booking data in civil aviation. Air passenger travel forecasting method.

The present invention adopts the theme model to carry out theme modeling for passengers and their selected airlines and routes, and introduces a passenger association graph constructed by introducing a passenger association graph travel theme model PGTTM (Passenger Graph based Travel Topic Model). Airline preference information and enrich topic information to solve the problem of civil aviation sparsity.

Next, a Bayesian probability model is introduced, which integrates airline popularity, passenger preference for airline routes obtained by PGTTM, passenger loyalty, and airline market share, and constructs a multi-factor fusion forecasting framework to more accurately predict and recommend future passengers. Airlines and routes. The above is the main content of the invention of the multi-factor fusion civil aviation passenger travel prediction method based on the theme model.

Technical scheme of the present invention

A multi-factor fusion civil aviation passenger travel prediction method based on topic model, the method includes:

Step 1): Build a travel topic model of passenger association graph. It mainly includes the construction of passenger association graph, and the topic modeling of passenger travel preference, and finally the travel topic model of passenger association graph is obtained:

Step 1.1), build a passenger association graph;

To construct a passenger association graph is to calculate the association degree between passengers, which is determined by the co-occurrence degree of passenger routes and the co-occurrence degree of attributes; the co-occurrence degree of routes is determined by the co-occurrence number of routes between passengers; the co-occurrence degree of attributes refers to Whether the age, gender, average discount, and average mileage of the passengers are the same; the information on the age, average discount, and average mileage of the passengers is obtained by the segmentation method based on variance;

Step 1.2), modeling the theme of passenger travel preference;

Based on the topic model, the topic modeling of passengers and their routes and airlines is carried out, and the potential topic distribution of passengers, routes and airlines is found and obtained, and finally the potential topic distribution of passengers is compared with the potential topic distribution of airlines and airlines. Combined, you can get the travel preference information of passengers on airlines and routes;

Step 1.3), build a travel theme model of a passenger association graph;

In the process of step 1.2) topic modeling, the passenger association graph in step 1.1) is added to construct the Passenger Graph based Travel Topic Model (PGTTM); PGTTM is assigned to each passenger's route and airline When the theme is used, the theme can come not only from the passenger itself, but also from other passengers associated with the passenger, which can enrich the theme information, improve the prediction performance, and alleviate the problem of the sparse travel of civil aviation passengers;

Step 2): Build a calculation model for airline popularity, passenger loyalty, and airline market share, and use these prior knowledge to help accurately predict the following:

Step 2.1), calculate the heat of the route;

For airline popularity, first count the number of times the route is taken by all passengers, and the sum of the number of times each route is taken by all passengers. On this basis, the route popularity is calculated;

Step 2.2), calculate the loyalty of the passenger to the airline;

For passenger loyalty, first count the number of times the passenger takes the airline and the sum of the number of times the passenger takes each airline. On this basis, after smoothing, the passenger's loyalty to the airline is calculated;

Step 2.3), calculate the market share of the airline on the route;

For the airline market share, first count the number of times the airline and this route are taken by all passengers as a word pair, and the number of times the airline is taken by all passengers without considering the airline. market share of the route;

Step 3): Construct a multi-factor fusion forecasting framework by integrating airline popularity, passenger preference for airline, passenger loyalty, and airline market share through a Bayesian probability model to predict the airline and airline that passengers will choose in the future:

Step 3.1), multi-factor fusion based on Bayesian probability model;

Based on the passenger's preference for the route obtained by PGTTM in step 1), the route popularity in step 2.1), the passenger loyalty in step 2.2), and the airline's market share in step 2.3), a Bayesian probability model is constructed to calculate These four factors are integrated to better model the travel behavior of passengers;

Step 3.2), multi-factor prediction based on Bayesian probability model;

For each passenger and each airline-route word pair, use the Bayesian probability model function to calculate the passenger's boarding probability; for each passenger, select several airline-route word pairs with the highest probability, Make predictions and recommendations.

Advantages and positive effects of the present invention:

·Propose the travel theme model PGTTM of passenger association graph

The invention conducts theme modeling for the travel behavior of civil aviation passengers, discovers the potential theme distribution of passengers and the airlines and routes they take, and accurately predicts the routes and other behaviors that passengers choose to travel in the future. On this basis, a passenger correlation graph is constructed and introduced, and the PGTTM is obtained, which can enrich the subject information with the help of similar passengers, improve the prediction accuracy, and solve the problem of the sparseness of civil aviation passenger travel data.

·Propose a multi-factor fusion prediction framework with the help of the Bayesian probability model function

The present invention obtains a multi-factor fusion prediction framework through a Bayesian probability model function, integrates the preference of passengers on routes obtained by PGTTM, as well as the prior knowledge of route popularity, passenger loyalty and airline market share, compared with the benchmark method, the forecasting framework can more accurately predict the airlines and routes that passengers choose to travel in the future.

Description of drawings

FIG. 1 is an overall model system diagram of the present invention.

Fig. 2 is the algorithm flow chart of the present invention.

Detailed ways

Example 1:

The method for predicting the travel of civil aviation passengers based on multi-factor fusion based on the theme model provided by the present invention will be described in detail below with reference to the accompanying drawings and specific implementations.

The present invention mainly adopts the theory and method of data mining to analyze the travel behavior of passengers in the civil aviation data. In order to ensure the normal operation of the system, in the specific implementation, it is required that the computer platform used is equipped with a memory of not less than 8G and the number of CPU cores is not low. On 4 64-bit operating systems with a main frequency of not lower than 2.6GHz, Windows7 and above, and install Oracle database, Java1.7 and above, Matlab2011b and above and other necessary software environments.

The method for predicting the travel behavior of passengers based on the theme model fusion of multiple factors provided by the present invention is as follows, and is described in conjunction with FIG. 1 and FIG. 2 .

Step 1): Build a passenger association graph travel topic model PGTTM

Step 1.1), data preprocessing and S1.1 stage of building passenger association graph;

Step 1.11), data introduction and preprocessing

Each piece of data in the passenger booking data contains the personal information and travel information of the passenger; the personal information includes the encrypted ID number that uniquely identifies the passenger, the passenger's age, gender, etc., and the travel information includes the airline taken, the departure airport, the arrival airport, and the discount. and other information.

After preprocessing operations such as removing low-frequency passengers, removing duplicate records, and removing abnormal records, certain historical data is taken as the training set, and the rest of the data is used as the test set.

Step 1.12), segmentation method based on variance;

For example, age segmentation, extract all ages in the passenger travel records of the training set into a sorted list, traverse the minimum age to the maximum age, take each age traversed as the segmentation point, and calculate the weighted average of the variances of the two age tables after segmentation. The weight is the ratio of the age included after segmentation to the age before segmentation. Find the segmentation age value with the largest difference between the weighted average of the variance after segmentation and the variance before segmentation, which is the best segmentation point.

Step 1.13), build a passenger association graph;

The degree of association between passengers is determined by the co-occurrence degree of the airline and the co-occurrence degree of attributes; statistical calculation is performed on the training set obtained in step 1.11), and a sparse matrix expressing the number of airline co-occurrences between passengers is obtained, and each column is normalized is the route co-occurrence degree matrix; the attribute co-occurrence degree refers to whether the two passengers are the same after the passenger's age, gender, average discount, and average mileage are divided in step 1.12); Several passengers are regarded as their associated passengers, and then the degree of association between the passenger and these associated passengers is obtained by the weighted average of the co-occurrence degree of routes and the co-occurrence degree of attributes between them; thus, the association graph between passengers is constructed.

The flight route taken by the passenger is determined by the departure airport and the arrival airport, the mileage information is obtained from the distance between the two cities represented by the departure airport and the arrival airport, the price is determined by the mileage and discount information, and the average discount is determined by the passenger's total mileage and total price. Decide.

Step 1.2), use PGTTM to model passenger travel preferences

Step 1.21) obtain the S1.21 stage of the input data;

Suppose there are different U passengers (distinguished by encrypted ID numbers), C airlines, and R routes in the passenger booking records of the training set. Extract the three fields of ID number, airline and airline from the passenger booking record, and replace them with the index form respectively, that is, these three fields are represented by numbers 1~U, 1~C, 1~R respectively, and finally three fields are obtained. A vector u, c, r, all of length N (also the number of booking records in the training set), is the input data. Each row of the three vectors indicates that the passenger ui in the _ith booking record took the route _ri under the airline c _i , (1≤u _i ≤U, 1≤ci _{≤C, 1≤r i ≤} R, i=1,2,...,N).

T is the set number of topics. z represents the topic vector, with length N, and x is the passenger vector used to generate the topic, with length N. The relationship of u, c, r to z, x is that each component of them represents that the subject _zi of the airline ci and the route _ri that the passenger _ui takes is assigned by _xi , and _xi can be _{u i} , and may also be associated passengers of _ui , (1≤zi ≤T, _1≤xi ≤U, _i =1,2,...,N).

The following is the process by which passengers generate each travel behavior in PGTTM:

Dirichlet(β)，(u＝1,2,...,U，t＝1,2,...,T；θ_u是T维向量，φ_t是C维向量，

是R维向量；α，μ，β是狄利克雷分布的参数)。(1) Each passenger u corresponds to a topic distribution, and each topic t corresponds to an airline distribution and an airline distribution. The topic distribution θ _u Dirichlet(α) of passenger u, the airline distribution φ _t Dirichlet(μ) of topic t, the route distribution of topic t

Dirichlet(β), (u=1,2,...,U, t=1,2,...,T; θ _u is a T-dimensional vector, φ _t is a C-dimensional vector,

is an R-dimensional vector; α, μ, β are the parameters of the Dirichlet distribution).

航线r_i

在PGTTM中s可以是u_i本身，还可能是u_i的关联旅客，(1≤u_i≤U，1≤z_i≤T，1≤c_i≤C，1≤r_i≤R，i＝1,2,...,N)。(2) Passenger u _i first samples a passenger s, and then samples a travel theme from s, and finally selects the airline and route to take according to the travel theme. i.e. topic z _i Multinomial(θ _s ), airline c _i

route r _i

In PGTTM, s can be u _i itself, or the associated passenger of u _i , (1≤u _i ≤U, 1≤zi ≤T, 1≤ci ≤C, 1≤r _i ≤R _{, i} = 1,2,...,N).

(T×R维)是PGTTM要推断的参数。就是根据已有的旅客u和其搭乘行为c、r，反向推断它们的主题分布。Passenger-subject distribution θ (U×T dimension), subject-airline distribution φ (T×C dimension), subject-airline distribution

(T×R dimension) is the parameter to be inferred by PGTTM. It is to infer their topic distributions in reverse based on the existing passengers u and their boarding behaviors c and r.

Step 1.22) S1.22 stage of initialization operation;

The initial state of the passenger x for assigning the theme is set equal to the passenger u boarding. Then randomly initialize the topic vector z with T topics. (ie 1≤zi≤T, _i =1,2,...,N).

Let C ^UT be a U×T-dimensional matrix, which represents the number of times passengers are assigned to each topic, which is obtained by the vectors x and z; C ^TC is a T×C-dimensional matrix, which represents the number of topics assigned to each airline, which is calculated by the vectors z and c. Statistically obtained; C ^TR is a T×R-dimensional matrix, which represents the number of times the subject is assigned to each route, and is obtained from the statistics of the vectors z and r. These three matrices are the subject count matrices of passengers, airlines, and routes, respectively.

Set the maximum number of iterations NN; construct a vector order of length N whose values are spread from 1 to N, but the order is randomly shuffled.

Step 1.23) Update the S1.23 stage of the topic count matrix without considering the topic assignments of the current passenger, current airline and airline;

都减1。Update the three topic count matrices without considering the component of topic z subscripted by order _i , that is

both minus 1.

Step 1.24) S1.24 stage of sampling a passenger used to generate a new theme for the current airline and airline;

和航线

重新采样的主题由当前旅客

产生，还是由

关联图中的关联旅客产生。而由

的哪一个关联旅客产生，则由一个多项分布来决定，该多项分布的参数是该旅客与其关联旅客的关联度。假设采样旅客为s，

是采样概率，取决于两个分布的参数。Determined by a Bernoulli distribution with parameter τ as the current airline

and route

Resampled topics by current travelers

produced by

The associated passengers in the association graph are generated. and by

Which of the associated passengers is generated is determined by a multinomial distribution whose parameter is the degree of association between the passenger and its associated passengers. Suppose the sample passenger is s,

is the sampling probability, which depends on the parameters of the two distributions.

Step 1.25) The S1.25 stage of reassigning a new topic for the current airline and airline using the Gibbs sampling formula;

和当前航线

重新分配的新主题是t(t＝1,2,...,T)的概率。公式如下：According to the Gibbs sampling formula, it is calculated that the passenger s is the current airline

and current route

The reassigned new topic is the probability of t(t=1,2,...,T). The formula is as follows:

是旅客s分配主题t的次数，

是主题t分配给航空公司

的次数，

是主题t分配给航线

的次数，

是步骤1.23)得到的、根据

采样旅客s的概率。The meaning of the formula is the probability of sampling passenger s for the current passenger and sampling the new topic t for the current airline and airline. Among them, the vector with the subscript -order _i means that the component with the subscript order _i is not considered,

is the number of times passenger s assigns topic t,

is the topic t assigned to the airline

number of times,

is the subject t assigned to the route

number of times,

is obtained in step 1.23), according to

Probability of sampling passenger s.

Finally, a multinomial distribution is formed with these T probability values as parameters, and a new topic is sampled as topic.

Step 1.26) update the S1.26 stage for generating the subject vector and subject vector;

更新为s，根据步骤1.25)在z中将

更新为topic。According to step 1.24) in x the

Update to s, in z according to step 1.25)

Update to topic.

Step 1.27) Update the S1.27 stage of the three topic count matrices;

都加1。After the passenger vector and topic vector of the generated topic are updated in step 1.26), let

Both add 1.

Step 1.28) Calculate the S1.28 stage of passenger-topic, topic-airline, and topic-airline distribution after the iteration is over;

The number of iterations is from 1 to NN, and i is from 1 to N, as the outer loop and the inner loop, respectively, to continuously resample the passengers that generate the topic and the topics assigned to airlines and routes, that is, repeat steps 1.23) to 1.27). After the iteration is completed, the passenger-topic distribution θ, the topic-airline distribution φ, and the topic-airline distribution can be obtained according to the following formulas

Wherein, u=1,2,...,U, c=1,2,...,C, r=1,2,...,R, t=1,2,...,T.

Step 1.29) The S1.29 stage of calculating the passenger's preference for the route;

PGTTM is used to model passengers' preferences for airlines and routes. For example, P(u|r) is used to represent the passenger's preference for the airline, which is also the attraction of the airline to passengers. The calculation formula is as follows:

where u=1,2,...,U, r=1,2,...,R.

Step 2): Calculate airline popularity, passenger loyalty, and airline market share:

Step 2.1), the S2.1 stage of calculating the route heat;

Route popularity is represented by P(r), which indicates the probability that passengers choose route r when they travel. The formula is as follows:

Among them, count(r) represents the number of times the route r appears in the passenger booking records in 2010, r=1,2,...,R.

Step 2.2), the S2.2 stage of calculating passenger loyalty;

Passenger loyalty is represented by P(c|u), which indicates the probability that passenger u chooses airline c when he travels. The formula is as follows:

Among them, count(u,c) represents the number of times that passenger u chooses airline c in the passenger booking record in 2010, c=1,2,...,C, u=1,2,...,U.

Step 2.3), the S2.3 stage of calculating the airline market share;

The airline market share is represented by P(c|r), which indicates the probability that route r belongs to the route of airline c. The formula is as follows:

Among them, count(c,r) represents the number of records co-occurred by airline c and route r in the passenger booking records in 2010, c=1,2,...,C, r=1,2,... ., R.

Step 3): Introduce a Bayesian probability model, build a multi-factor fusion prediction framework, calculate the probability of passengers taking airlines and routes, and carry out the S3 stage of prediction and recommendation:

Step 3.1), use the Bayesian probability model to construct a multi-factor fusion prediction framework;

The airline preference of passengers obtained by PGTTM in step 1), and the airline popularity, passenger loyalty, and airline market share in step 2) are fused together using a Bayesian probability model to construct a multi-factor fusion prediction framework. The Bayesian probability model used in the present invention is derived as follows:

First, for a fixed passenger u, P(u) is a constant, we can get

according to

P(r,c,u)=P(r)*P(u|r)*P(c|u,r)≈P(r)*P(u|r)*[αP(c|u)+ (1-α)P(c|r)], so the required Bayesian probability function can be obtained as follows:

logP(r,c|u)∝log{P(r)*P(u|r)*[αP(c|u)+(1-α)P(c|r)]}

Among them, P(r,c|u) represents the probability that the passenger u chooses the route r under the airline c, α is a parameter that can be set, and the log on both sides of the formula is to avoid the obtained probability value being too small.

The last formula is the required Bayesian probability model, and it is also a multi-factor fusion prediction framework, which integrates airline popularity P(r), passenger preference for airline P(u|r), passenger loyalty P(c|u), Airline market share P(c|r). (c=1,2,...,C, r=1,2,...,R, u=1,2,...,U).

Step 3.2), predict the airline and route that the passenger will choose in the future;

According to the multi-factor prediction framework in step 3.1), assuming that there are a total of W airline-airline word pairs in the training set, for each passenger u, the probability of taking each airline-airline word pair can be calculated. Sort from large to small, and then find the top K (TopK) airline-airline word pairs with the largest values as the prediction objects, and recommend them. By comparing the prediction results with the test set, the prediction accuracy is obtained.

For example, for a passenger 17464755. (encrypted ID number), the airline 290 (the real name code) and the route CTU-CAN (airport three-character code, Chengdu Shuangliu Airport-Guangzhou Baiyun Airport) in the booking data The representative (c, r) is substituted into the multi-factor fusion prediction framework function of step 3.1) for calculation. Assuming that the calculated value is the largest compared to other W-1 word pairs, it is a matter of course that this word pair is used as the prediction object. If the passenger actually took the route under the airline in the test set, then for Top1, the prediction accuracy is 1. (c=1,2,...,C, r=1,2,...,R, u=1,2,...,U).

It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive, so the present invention is not limited to the embodiments described in the specific implementation manner. The other embodiments obtained also belong to the protection scope of the present invention.

Claims (1)

1. A multi-factor fusion civil aviation passenger travel prediction method based on the theme model, using data mining theory and methods to analyze passenger travel behavior in civil aviation data, the operating environment requires the computer platform used to be equipped with a memory of no less than 8G, CPU A 64-bit operating system with at least 4 cores and a main frequency of at least 2.6GHz, Windows 7 and above, Oracle database, Java 1.7 and above, and Matlab 2011b and above must be installed; , the method includes: Step 1): Build a travel topic model of passenger association graph; including building a passenger association graph, and subjecting the travel selection probability distribution of passengers to topic modeling, and finally obtaining a travel topic model of passenger association graph: Step 1.1), build a passenger association graph; To construct a passenger association graph is to calculate the association degree between passengers, which is determined by the co-occurrence degree of passenger routes and the co-occurrence degree of attributes; the co-occurrence degree of routes is determined by the co-occurrence number of routes between passengers; the co-occurrence degree of attributes refers to Whether the age, gender, average discount, and average mileage of the passengers are the same; the information on the age, average discount, and average mileage of the passengers is obtained by the segmentation method based on variance; Step 1.2), modeling the subject of passenger travel selection probability distribution; Based on the topic model, the topic modeling of passengers and their routes and airlines is carried out, and the potential topic distribution of passengers, routes and airlines is found and obtained, and finally the potential topic distribution of passengers is compared with the potential topic distribution of airlines and airlines. Combined, the probability distribution information of passengers' travel choices to airlines and routes is obtained; Step 1.3), build a travel theme model of a passenger association graph; In the process of step 1.2) topic modeling, the passenger association graph in step 1.1) is added to construct the Passenger Graph based Travel Topic Model (PGTTM); PGTTM is assigned to each passenger's route and airline When the theme is used, the theme can not only come from the passenger itself, but also from other passengers associated with the passenger; this can enrich the theme information, improve the prediction performance, and alleviate the problem of the sparse travel of civil aviation passengers; Step 2): Build a calculation model for airline popularity, passenger loyalty, and airline market share, and use these prior knowledge to help accurately predict the following: Step 2.1), calculate the heat of the route; For airline popularity, first count the number of times the route is taken by all passengers, and the sum of the number of times each route is taken by all passengers. On this basis, the route popularity is calculated; Step 2.2), calculate the loyalty of the passenger to the airline; For passenger loyalty, first count the number of times the passenger takes the airline and the sum of the number of times the passenger takes each airline. On this basis, after smoothing, the passenger's loyalty to the airline is calculated; Step 2.3), calculate the market share of the airline on the route; For the airline market share, first count the number of times the airline and this route are taken by all passengers as a word pair, and the number of times the airline is taken by all passengers without considering the airline. market share of the route; Step 3): Build a multi-factor fusion prediction framework through the Bayesian probability model; integrate the popularity of routes, the probability distribution of passengers' choice of routes, passenger loyalty, and airline market share through the Bayesian probability model. Routes and airlines make predictions: Step 3.1), multi-factor fusion based on Bayesian probability model; A Bayesian probability model is constructed based on the probability distribution of passengers' choice of routes obtained by PGTTM in step 1.3), the popularity of routes in step 2.1), the passenger loyalty in step 2.2), and the airline market share in step 2.3). , Integrate these four factors to better model the travel behavior of passengers; Step 3.2), multi-factor prediction based on Bayesian probability model; For each passenger and each airline-route word pair, use the Bayesian probability model function to calculate the passenger's boarding probability; for each passenger, select several airline-route word pairs with the highest probability, Make predictions and recommendations.

Images