Passenger multi-behavior flight selection method and device for personalized sequence recommendation

Abstract

The invention discloses a flight selection method and device for passenger multi-behavior personalized sequence recommendation, and the method comprises the steps: dividing a behavior log into sessions according to the purchase behaviors of passengers, and representing the sessions as session-airport tuples; mapping the continuous timestamps of the session into a set of discrete time units; time unit embedding of three dimensions is aggregated, and embedding representing respective time modes is generated: all time modes and space modes are connected together to obtain embedding representation of passengers; performing embedding initialization on the passenger and multiple behaviors by utilizing a dependency relationship between the behaviors, and performing GCN cascade learning on the multiple behaviors of the passenger; aggregating the features learned from all behaviors by using a linear combination mode; a multi-layer perceptron is used for modeling attribute representation for each passenger to capture attribute features of the passengers, space-time representation, multi-behavior representation and passenger attribute representation of the passengers and flights are integrated together, and finally, the flight preference of the passengers is captured, and flight choices more conforming to interests are provided for the passengers.

Description

Technical Field

[0001] The present invention relates to the field of personalized recommendation, and particularly to a flight selection method and device for personalized sequential recommendation of multiple traveler behaviors. Background Art

[0002] With the popularization of mobile terminals, more and more people have changed the traditional ticket purchasing mode and are used to purchasing tickets on the direct sales websites of airlines. Personalized recommendation models travelers' preferences based on the interaction history between travelers and flights, obtains the preference probability of travelers for flights, recommends flights that travelers are interested in, and improves the travelers' experience. Currently, most of the preferences of travelers for flights ignore the sequential characteristics of the traveler's purchase behavior sequence and the dependency relationships between multiple behaviors.

[0003] Sequential recommendation attempts to model users' dynamic preferences by mining users' historical interaction behaviors and predict the next item that users may like. The earliest work used Markov chains to model the transitions of users' interaction sequences to capture users' dynamic preferences [1] . To apply to more complex recommendation scenarios, deep learning-based models (such as RNN (Recurrent Neural Network) and GNN (Graph Neural Network)) have been applied to sequential recommendation. For example: GRU4Rec [2] and SR-GNN [3] are the first models to apply RNN and GNN to sequential recommendation respectively. Transformer and attention mechanisms have also been applied to sequential recommendation [4,5] . SASRec [4] and STOSA [5] apply transformer and self-attention networks to capture users' long-term interests. While DGSR [6] captures dynamic preferences on different user sequences through a dynamic graph neural network on the interaction bipartite graph.

[0004] Users' purchase behaviors for items are a type of explicit feedback that can directly reflect users' preferences. Clicking, adding to the shopping cart, favoriting, etc. provide rich information for understanding travelers' preferences and are effective methods to alleviate the data sparsity or cold start problems in recommendation. NMTR [7] extends the neural collaborative filtering (NCF) framework to a multi-behavior setting and performs joint optimization on the cascade prediction task. Subsequently, researchers have used GNN to explore high-hop user-item interactions in multi-behavior recommendation. For example: MBGCN [8] and GHCF [9] use GCN to learn discriminative behavior representations, while MBGMN

[10] uses a graph primitive network to capture interaction diversity and behavior heterogeneity. To alleviate the data sparsity of the target behavior, MBSSL

[11] Combining self-attention and behavior-aware graph neural networks to capture the diversity and dependencies of behaviors.

[0005] In recent years, the civil aviation field has conducted systematic research on the problem of applying personalized recommendations to provide personalized services for passengers throughout the entire process of air travel. For example: flight seat recommendations

[12] ticket recommendations

[13] route selection

[14] flight recommendations

[15] and so on. They deeply mine users' behavior habits, associate passengers with "travel, transportation, dining, accommodation, and shopping", improve the travel satisfaction of passengers throughout the journey, increase user loyalty, and improve the company's revenue.

[0006] References

[0007] [1]Ruining He and Julian McAuley. 2016. Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering. In Proceedings of WWW. 507–517

[0008] [2]Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based recommendations with recurrent neural networks. In Proceedings of ICLR

[0009] [3]Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. In Proceedings of AAAI. 346–353

[0010] [4]Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recommendation. In Proceedings of ICDM. 197–206.

[0011] [5]Ziwei Fan, Zhiwei Liu, Yu Wang, Alice Wang, Zahra Nazari, Lei Zheng, Hao Peng, and Philip S Yu. 2022. Sequential Recommendation via Stochastic Self-Attention. In Proceedings of WWW. 2036–2047

[0012] [6]Mengqi Zhang, Shu Wu, Xueli Yu, and Liang Wang. 2022. Dynamic Graph Neural Networks for Sequential Recommendation. IEEE Transactions on Knowledge and Data Engineering (2022)

[0013] [7]Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat Seng Chua, and Depeng Jin. 2019. Neural multi-task recommendation from multi-behavior data. In ICDE. IEEE, 1554-1557

[0014] [8]Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Multibehavior recommendation with graph convolutional networks. In SIGIR

[0015] [9]Chong Chen, Weizhi Ma, Min Zhang, Zhaowei Wang, Xiuqiang He, Chenyang Wang, Yiqun Liu, and Shaoping Ma. 2021. Graph Heterogeneous Multi-Relational Recommendation. In AAAI, Vol. 35. 3958-3966

[0016]

[10] Lianghao Xia, Yong Xu, Chao Huang, Peng Dai, and Liefeng Bo. 2021. Graph meta network for multi-behavior recommendation. In SIGIR. 757-766

[0017]

[11] Jingcao Xu, Chaokun Wang, Cheng Wu, Yang Song, Kai Zheng, Xiaowei Wang, Changping Wang, Guorui Zhou and Kun Gai. 2023. Multi-behavior Self-supervised Learning for Recommendation. In SIGIR. CoRR abs / 2305.18238

[0018]

[12] Youfang Lin, Kunkun Wang, Chao Zhou, Huaiyu Wan, Zhihao Wu. Modeling Civil Aviation Passengers' Preferences Based on Social Networks[J]. Journal of Beijing Jiaotong University, 2014, 38(6): 33-39

[0019]

[13] Mengxi Chen, Peng Tian, & Xiangyong Li. (2019). Research on Dynamic Pricing of Multiple Cabin Classes Considering Passengers' Choice Behavior. Chinese Journal of Management Science, 2023, 31(11): 312-320.

[0020]

[14] Xia Feng, Chen Zhang, Min Lu. Prediction of Route Selection Behavior Based on Passenger Trust Network[J]. Modern Electronics Technique, 2020, 43(4): 78-86

[0021]

[15] Siering M, Deokar A V, Janze C, et.al. Disentangling consumer recommendations: Explaining and predicting airline recommendations based on online reviews[J]. Decision Support Systems, 2018, 107: 52-63 Summary of the Invention

[0022] In view of the problem that the existing personalized recommendation method for passenger flight selection cannot well consider the time and space periodic characteristics of each passenger's preferences and the dependency relationship between multiple behaviors of passengers, the present invention provides a flight selection method and device for personalized sequential recommendation of multiple passenger behaviors. The present invention can not only accurately capture passengers' preferences for flights, but also improve passengers' travel experience, as described in detail below:

[0023] In a first aspect, a flight selection method for personalized sequential recommendation of multiple passenger behaviors, the method includes:

[0024] Dividing the behavior log into sessions according to the purchase behavior of passengers, and representing them as session-airport tuples;

[0025] Mapping the continuous timestamps of the session to a set of discrete time units; aggregating the time unit embeddings in three dimensions, and generating embeddings representing their respective time patterns: connecting all the time patterns and space patterns together to obtain the embedding representation of the passenger;

[0026] Utilizing the dependency relationship between behaviors, initializing the embeddings of passengers and multiple behaviors, and performing GCN cascade learning on multiple passenger behaviors; aggregating the features learned from all behaviors in a linear combination manner;

[0027] Using a multi-layer perceptron to model the attribute representation for each passenger to capture the attribute characteristics of the passenger, integrating the spatio-temporal representation, multi-behavior representation and passenger attribute representation of the passenger and the flight, and finally capturing the passenger's preference for the flight to provide flights that better match the interests of the passenger.

[0028] Among them, the specific operation of dividing the behavior log into sessions according to the purchase behavior of passengers and representing them as session-airport tuples is as follows:

[0029] The session s in the passenger behavior graph G u =(S, A, V, E) has an embedding vector representation as: i The embedding vector representation is:

[0030] s i =σ(W S ·AGG sess (v i,j |(s i , v i,j ) ∈ E)+b S )

[0031] Among them, σ is a non-linear activation function. The trainable weight matrix Converts the flight embedding from the K V -dimensional space to the K S -dimensional session embedding space; the aggregation function AGG sessUse a gated recurrent unit to aggregate the flight vectors of the same session s i ∈S to represent the aggregated session s i .

[0032] Among them, the embedded representation of the passenger is specifically:

[0033] Aggregate the session node embeddings into different time unit embeddings:

[0034]

[0035] Among them, the weight matrix transforms the session embedding space K S into the hour embedding space K h , the week embedding space K w and the weekday embedding space K y respectively, and AGG temp applies a time aggregation function to all sessions;

[0036]

[0037] Among them, the weight matrix converts the aggregated hour, week, and weekday embeddings into the corresponding hour, week, and weekday patterns respectively;

[0038] The calendar GNN aggregates it into airport unit embeddings according to the session node embeddings:

[0039]

[0040] Finally, connect all the time patterns and space patterns together to obtain the embedded representation of the passenger.

[0041]

[0042] Among them, the embedding initialization of the passenger and multi-behavior is:

[0043] Passenger u m and flight i n The initialization embedding can be defined as:

[0044]

[0045] Among them, · represents matrix multiplication operation. In this way, the corresponding row can be selected from the one-hot vector matrix, and then multiplied by the embedding matrices P and Q to obtain the initialization embedding vectors of passenger u m and flight i n .

[0046] Among them, the GCN cascade learning for the multi-behaviors of passengers is as follows:

[0047] By using the transformation matrix W b , transform the embedding:

[0048]

[0049] Among them, and respectively represent the transformation matrices of passenger u and flight i. and represent the initial embeddings of passenger u and flight i in the (b + 1)-th behavior.

[0050] In a second aspect, a flight selection device for personalized sequential recommendation of multi-behaviors of passengers, the device includes: a processor and a memory, and program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute any one of the methods in the first part.

[0051] In a third aspect, a computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is enabled to execute any one of the methods in the first aspect.

[0052] The beneficial effects of the technical solution provided by the present invention are:

[0053] 1. The present invention utilizes the search and interaction history of passengers, captures the temporal and spatial periodic characteristics of passenger preferences and considers the dependence relationship between passenger auxiliary behaviors (such as: purchase, browsing, query) and purchase behaviors, predicts the probability of passengers' preference for aviation products, and thus establishes a passenger flight selection model;

[0054] 2. The present invention can not only accurately capture passengers' preferences for flights, but also improve passengers' travel experience. The present invention further accurately captures passengers' preferences for flights and provides passengers with flight selections that more conform to their interests. Brief Description of the Drawings

[0055] Figure 1 is a flowchart of a flight selection method for personalized sequential recommendation of multi-behaviors of passengers;

[0056] Figure 2 is a schematic diagram of cascade learning of multi-behaviors of passengers. Detailed Embodiments

[0057] To make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below.

[0058] In the civil aviation industry, passengers' flight-taking behavior has very obvious spatio-temporal characteristics, and this time periodicity is reflected in multiple time levels. For example, from the perspective of flight-taking time, different passengers have very different preferences for flights at different times, and passengers are more willing to travel during specific holidays such as summer and winter vacations, Spring Festival, and National Day; during the day, the number of passengers traveling in the morning and afternoon is significantly higher than that at midnight; during a week, the number of passengers traveling on weekdays is significantly lower than that on weekends. This time periodicity reflects that passengers have different flight-taking preferences at different times. In addition, passengers' travel is not only sensitive to time but may also be periodic with respect to the departure and arrival cities. For example, a passenger may travel to a certain city on every Wednesday for business and return on Friday, etc. Therefore, it is crucial to accurately capture passengers' preferences by exploring the spatio-temporal characteristics of passengers' behavior, provide flight choices that better suit their interests, and improve the travel experience.

[0059] Embodiment 1

[0060] The key challenge in capturing the spatio-temporal periodicity of passengers is how to model the hierarchical time structure of passengers' behavior using the explicit time patterns of the calendar system. Therefore, the standard calendar system (e.g., days, weeks, and months) and the calendar system (e.g., hours and minutes) are used to discover the spatio-temporal preference characteristics of passengers at different time levels (hours, weeks, months), and at the same time, the dependency relationships between different behaviors of passengers are considered to obtain the dynamic preferences of passengers and better model passengers' flight choices. The process of generating the spatio-temporal feature vector of passengers is as Figure 1 shown:

[0061] I. Creation of the spatio-temporal graph of passengers' behavior and generation of session embeddings

[0062] According to the purchase behavior of passengers, their behavior logs are divided into sessions and represented as session-airport tuples where m u is the number of sessions for each passenger u. Each session s u,i contains: a flight and timestamp pair where n i represents the number of flights contained in session s u,i .

[0063] The passenger-airport-flight behavior graph for each passenger u is G u =(S, A, V, E), where S, A, and V represent the sets of session nodes, airport nodes, and flight nodes of passenger u respectively. If a passenger takes a flight departing from or arriving at airport a i in session s i , there is an edge (s i , a i ) ∈ E i between session node s i ∈ S and airport node a (A) . If a passenger takes a flight in session s iDuring the period ∈S with a specific flight v i,j interacts with V, then in session s i ∈S with flight v i,j ∈V there exists (s i , v i,j ) ∈ E (V) . To capture the temporal characteristics of passenger behavior, the timestamp t of the flight i,j is used as the temporal attribute value of the edge in E (V) .

[0064] Therefore, the session s in the passenger behavior graph G u =(S, A, V, E) is embedded with the vector representation as follows: i

[0065] s i = σ(W S · AGG sess (v i,j |(s i , v i,j ) ∈ E) + b S )

[0066] where σ is a non - linear activation function. The trainable weight matrix converts the flight embedding from a K V - dimensional space to a K S - dimensional session embedding space. The aggregation function AGG sess adopts a Gated Recurrent Unit (GRU) to aggregate the flight vectors within the same session s i ∈ S into the representation of session s i .

[0067] II. Representation Learning of Passenger Temporal and Spatial Patterns

[0068] Given the session node embeddings, the idea of temporal aggregation at this layer is as follows: (1) Map the consecutive timestamps of the session to a set of discrete time units; (2) Aggregate the sessions in the same time unit into the corresponding time unit embeddings; (3) Aggregate the time unit embeddings into the embedding of the time pattern.

[0069] Map the timestamps of the session to a set of discrete time units, consider the timestamp of the first flight as the timestamp of the session, and map the timestamp to three types of time units:

[0070] · h i = hour(t i ) ∈ τ h , where τ h has 24 different values: 0 o'clock, 1 o'clock,..., 23 o'clock;

[0071] · w​i = week(t i ) ∈ τ w , where τ w is the set of weeks in a year, e.g., the 18th week;

[0072] · y i = weekday(t i ) ∈ τ y , where T y has 7 values: Sunday, Monday, …, Saturday;

[0073] Once the session nodes are mapped to the specified time unit, the Calendar-Airport GNN aggregates the session node embeddings into different time unit embeddings by applying the temporal aggregation function Agg temp to all sessions of the same time unit:

[0074]

[0075] where the weight matrix transforms the session embedding space K S into the hour embedding space K h , the week embedding space K w and the weekday embedding space K y . AGG temp is set to GRU because all flights of the same time unit can be naturally sorted according to the original timestamps.

[0076] Next, the time unit embeddings in the three dimensions (i.e., hour, week, and weekday) are further aggregated, and embeddings representing their respective time patterns are generated:

[0077]

[0078] where the weight matrix converts the aggregated hour, week, and weekday embeddings into the corresponding hour, week, and weekday patterns. Each of these time patterns captures the behavior patterns of passengers during a specific period.

[0079] Similar to the processing of the temporal aggregation layer introduced earlier, to generate the spatial (airport) pattern, the Calendar GNN first aggregates it into airport unit embeddings based on the session node embeddings:

[0080]

[0081] Finally, all the time patterns and spatial patterns are concatenated together to obtain the embedding representation of the passengers.

[0082]

[0083] III. Cascaded Learning of Multiple Passenger Behaviors

[0084] A display feedback of a passenger's purchase behavior for a flight can directly reflect the user's preference. Auxiliary behaviors such as clicks, adding to the shopping cart, and favorites provide rich information for understanding passenger preferences and are effective ways to alleviate the problem of sparse or cold-start purchase behavior data. During the ticket purchase process, passengers usually have their own behavioral sequences, and different behaviors reflect the preferences of the passengers. Therefore, the dependency relationship between behaviors can be utilized, and the information learned from one behavior can be passed to the next behavior to better capture the preferences of passengers. The specific process is as Figure 2 shown. The passenger set U and the flight set I are considered, and a multi-behavior interaction matrix {Y 1 , Y 2 , ···, Y B} sorted by the behavior sequence is used to represent the interaction situations of different behavior types. Among them, B is the number of behavior types, and Y b represents the interaction matrix of the b-th behavior type. These interaction matrices are all binary, and the value at each position is 1 or 0.

[0085] (1) Embedding Initialization of Passengers and Multiple Behaviors

[0086] The embedding vectors of passenger u ∈ U and flight i ∈ I are represented as and where d is the size of the vector. Use and to represent the embedding matrices of passengers and flights. Each passenger or flight is described by a unique ID, and this ID is represented by a one-hot vector. The one-hot vector matrix of all passengers is represented as ID U , and the one-hot vector matrix of flights is represented as ID I . The initial embeddings of passenger u m and flight i n can be defined as:

[0087]

[0088] where · represents the matrix multiplication operation. In this way, the corresponding rows can be selected from the one-hot vector matrix and then multiplied by the embedding matrices P and Q to obtain the initial embedding vectors of passenger u m and flight i n .

[0089] (2) Cascaded Learning of Passenger Multiple Behaviors by GCN

[0090] It consists of a series of LightGCNs. Each LightGCN corresponds to a type of behavior and is used to learn the embeddings of passengers and flights under that behavior. In a cascaded manner, the embeddings learned by the previous LightGCN are used as the input features for the next LightGCN, leveraging the dependencies in the behavior chain to assist in learning the features of subsequent behaviors. Such a design can extract features from different behaviors and utilize the correlations between behaviors for recommendations.

[0091] In each LightGCN, given the input embeddings of a passenger and a flight representing the b-th type of interaction behavior, embedding propagation is performed through the user-item interaction graph to update the embeddings as follows:

[0092]

[0093]

[0094] where and respectively represent the updated embeddings of passenger u interacting with flight i under behavior b after l layers of propagation. N u represents the set of flights interacting with passenger u, and N i represents the set of passengers interacting with flight i. After L layers of propagation, the LightGCN obtains L + 1 embeddings, representing an embedding of a passenger and a flight To obtain the final passenger and flight embeddings based on the b-th type of behavior, these embeddings are simply aggregated as follows:

[0095]

[0096] In this way, the embeddings learned from each type of behavior are integrated into the final embeddings of passengers and flights. After the feature transformation operation, the embeddings learned from the previous behavior are used as the initial embeddings for the next behavior. To process the learned embeddings, a feature transformation design is introduced before passing. By using the transformation matrix W b , the embeddings are transformed as follows:

[0097]

[0098] where and respectively represent the transformation matrices of passenger u and flight i. and Denote the initial embedding of passenger \(u\) and flight \(i\) in the \((b + 1)\) - th behavior. The purpose of feature transformation is to extract useful features to facilitate the embedding learning of the next behavior. Such a cascaded GCN structure can extract features from different behaviors and utilize the dependencies in the behavior chain.

[0099] (3) Integration of Embedding Representations of Different Passenger Behaviors

[0100] To utilize the embeddings learned from different behaviors, the features learned from all behaviors are aggregated using a linear combination, i.e.,

[0101]

[0102] IV. Aggregation and Optimization of Travel Sequence Embedding and Multi - behavior Embedding

[0103] (1) Passenger Attribute Modeling

[0104] When a passenger selects a certain flight, they may be affected by various factors such as departure time, aircraft type, ticket price, etc., which are all included in the passenger's attribute feedback. A Multilayer Perceptron (MLP) is used to model the attribute representation for each passenger to capture the passenger's attribute features. It is defined as follows:

[0105]

[0106] where \(W\) a is a trainable matrix, \(B\) a is a bias matrix, and \(\sigma(.)\) is an activation function.

[0107] (2) Integration of Embedding Representations of Passengers and Flights

[0108] Finally, the spatio - temporal representation, multi - behavior representation, and passenger attribute representation of passengers and flights are integrated together:

[0109]

[0110] Finally, the prediction function is defined as:

[0111]

[0112] In terms of model optimization, the standard Bayesian Personalized Ranking (BPR) loss function is used. This loss function assumes that the score of the observed item is higher than that of the unobserved item. Specifically, for the positive sample pair \((u, i)\) and the negative sample pair \((u, j)\), the BPR loss function can be defined as follows:

[0113]

[0114] Among them, is defined as the set of positive and negative sample pairs, σ(·) represents the sigmoid function, and Θ represents all trainable parameters. At the same time, to prevent overfitting, L2 regularization is adopted, and the coefficient λ is used to adjust the strength of regularization.

[0115] In summary, the embodiment of the present invention accurately captures the passengers' preferences for flights through the above parts, and provides flight selections that are more in line with the interests of passengers.

[0116] Embodiment 2

[0117] The behavioral data of approximately 28,000 passengers on 202 flight routes of a domestic airline in China from September 1, 2020 to November 30, 2020 are used, with a total of 1,048,000 behavioral events. In order to capture the influence of the dependencies of passengers' behaviors in different time windows and multiple behaviors on passengers' preferences, all behaviors are sorted by time, and the training set and the test set are divided. The passenger interaction data from September 1, 2020 to November 15, 2020 is used as the training set data; the passenger interaction data from November 16, 2020 to November 30, 2020 is used as the test set data, and the experimental results are shown in Table 1.

[0118] Table 1 Comparison of passenger flight selection results between this method and other baselines

[0119]

[0120]

[0121] The experimental results show that the multi-behavior model can achieve better performance than the single-behavior model, which proves the benefits of using auxiliary behaviors (i.e., browsing and collecting) to predict the target behavior (i.e., purchase). The MBDP model achieves the best performance and is significantly better than all baselines in terms of each metric of the dataset. For all passengers,

[0122] MBDP achieves an average improvement of MAR 22.9%, HR@k 15.7%, and Precision@k 19.1% respectively based on the strongest baseline. This confirms that the proposed passenger-flight multi-behavior model can extract passenger embeddings with stronger predictive power in flight recommendations.

[0123] I. Description of the data and format used in the implementation case

[0124] (1) Data used: Use the behavior data of approximately 28,000 passengers on 202 flight routes of a domestic airline in China from September 1, 2020 to November 30, 2020, with a total of 1,048,000 behavior events. The passenger behaviors are divided into several behaviors such as query, browse, and purchase, among which query and browse are auxiliary behaviors and purchase is the target behavior. Use the passenger interaction data from September 1, 2020 to November 15, 2020 as the training set data; use the passenger interaction data from November 16, 2020 to November 30, 2020 as the test set data.

[0125] (2) Data format

[0126] Each row in the training data table represents an interaction record between a passenger and a flight, including passenger ID, flight ID, spatial information, interaction time, and interaction event type.

[0127] Each row of the test data represents an interaction record between a passenger and a flight, including passenger ID, flight ID, spatial information, interaction time, and interaction event type.

[0128] Each row of passenger attributes represents the attribute information of a passenger, including column name, passenger ID, airport ID, and 6 passenger attributes.

[0129] II. Loading training data

[0130] Input the path of the parameter data file, load the data, and group the data by timestamp (t). This step divides the data into multiple groups according to the timestamp, and each group represents a time window.

[0131] By traversing the time window groups, extract the passenger ID, flight ID, airport ID, and label within each time window and store them as a nested list.

[0132] Finally, return a list of lists of sub-lists, which store the passenger ID, flight ID, airport ID, and label respectively. Among them, each sub-list represents the passengers, flights, and labels within a time window.

[0133] III. Data preprocessing

[0134] Read two files containing passenger records and passenger attribute information. Convert the time field in the records to a timestamp, and divide the data into a training set and a test set according to the timestamp.

[0135] Traverse the training set passenger records, extract the passenger attribute information, and organize the passenger interaction events into conversations in chronological order to generate a list of passenger conversations and a list of conversation times. Also, take the last purchase event as the candidate item for each passenger, along with the interaction time, and convert it into the form of month, day of the week, and hour.

[0136] Similarly, traverse the passenger records in the test set, extract the attribute information of the passengers, find the candidate items for each passenger in the test set, record the time of the candidate flights, and convert the timestamps into a list form of month, day of the week, and hour.

[0137] IV. Generation of Passenger Spatiotemporal Infiltration

[0138] (1) Generate passenger and flight representations and store them separately;

[0139] (2) Generate session representations: Obtain the infiltration of flights within each session and convert it into a session embedding;

[0140] (3) Extract time features: Obtain three time levels (hour, week, month) to extract time features and generate time feature embeddings;

[0141] (4) Extract space features: Extract airport infiltration and generate space feature embeddings;

[0142] (5) Update passenger representations:

[0143] V. Generation of Passenger Multi-behavior Infiltration Representations

[0144] Input the interaction records of each passenger in the training data, construct a multi-behavior graph for the passengers, learn the dependencies between multi-behaviors in different behavior graphs in a set order, and at the same time perform information propagation and integration within each behavior graph to update the infiltration representations of passengers, flights, and various behaviors, and further obtain the multi-behavior infiltration representations of passengers and flights.

[0145] VI. Connection of Infiltration Representations of Passengers and Flights

[0146] Connect the infiltration representations of passengers and flights obtained from the spatiotemporal model and the multi-behavior model to obtain the final representations of passengers and flights.

[0147]

[0148] VII. Calculate Prediction Scores through a Fully Connected Layer

[0149]

[0150] VIII. Calculate LOSS Value

[0151] In terms of model optimization, the standard BPR (Bayesian Personalized Ranking) loss function is used.

[0152] Negative Sampling: During the training phase, perform negative sampling and calculate the scores of negative samples.

[0153] Output: Return the calculated loss and the predicted probabilities or scores for the given passengers and candidate flights.

[0154] IX. Loading test data:

[0155] Input the test data file of the input parameters, and perform the same data preprocessing as the training data. Organize the data according to the time window.

[0156] X. Test results:

[0157] Use the behavior data of approximately 28,000 passengers on 202 flight routes of a domestic airline in China from September 1, 2020 to November 30, 2020, with a total of 1,048,000 behavior events. In order to capture the influence of the dependencies of passengers' preferences on different time windows and various behaviors, sort all behaviors by time, and divide the training set and the test set. Use the passenger interaction data from September 1, 2020 to November 15, 2020 as the training set data; use the passenger interaction data from November 16, 2020 to November 30, 2020 as the test set data. The experimental results are shown in Table 1.

[0158] Table 1 Comparison of passengers' willingness to pay between this method and other baselines

[0159]

[0160] The experimental results show that the multi-behavior model can achieve better performance than the single-behavior model, which proves the benefits of using auxiliary behaviors (i.e., browsing and collecting) to predict the target behavior (i.e., purchase). The MBDP model achieves the best performance and is significantly better than all baselines in terms of each metric of the dataset. For all passengers, MBDP achieves an average improvement of MAR 22.9%, HR@k 15.7%, and Precision@k 19.1% based on the strongest baseline. This confirms that the proposed passenger-flight multi-behavior model can extract passenger embeddings with stronger predictive ability in flight recommendations.

[0161] Example 3

[0162] A flight selection device for personalized sequential recommendation of passengers' multi-behaviors, the device includes: a processor and a memory, and program instructions are stored in the memory. The processor calls the program instructions stored in the memory to enable the device to execute the following method steps in Example 1:

[0163] A method for personalized sequential recommendation of passengers' multi-behaviors for flight selection, the method includes:

[0164] Divide the behavior log into sessions according to the purchase behavior of passengers, and represent them as session-airport tuples;

[0165] Map the continuous timestamps of the session into a set of discrete time units; aggregate the time unit embeddings in three dimensions and generate embeddings representing their respective time patterns: concatenate all the time patterns and spatial patterns to obtain the embedding representation of the passenger;

[0166] Utilize the dependencies between behaviors to initialize the embeddings of the passenger and multi-behaviors, and perform GCN cascade learning on the passenger's multi-behaviors; aggregate the features learned from all behaviors in a linear combination manner;

[0167] Use a multi-layer perceptron to model the attribute representation for each passenger to capture the passenger's attribute features, integrate the spatio-temporal representation, multi-behavior representation, and passenger attribute representation of the passenger and the flight, and finally capture the passenger's preference for the flight to provide flight selections that better match the interests of the passenger.

[0168] Among them, according to the purchase behavior of the passenger, the behavior log is divided into sessions and represented as session-airport tuples, specifically:

[0169] The passenger behavior graph G u =(S, A, V, E) in the session s i The embedding vector is represented as:

[0170] s i =σ(W S ·AGG sess (v i,j |(s i , v i,j )∈E)+b s )

[0171] Among them, σ is a non-linear activation function. The trainable weight matrix Converts the flight embedding from the K V -dimensional space to the K S -dimensional session embedding space; the aggregation function AGG sess Adopts a gated recurrent unit to aggregate the flight vectors within the same session si∈S to represent the session s i .

[0172] Among them, the embedding representation of the passenger is specifically:

[0173] Aggregate the session node embeddings into different time unit embeddings:

[0174]

[0175] Among them, the weight matrix Respectively transforms the session embedding space K S Into the hour embedding space K h , the week embedding space K w And the weekday embedding space Ky , AGG temp is the time aggregation function for all session applications;

[0176]

[0177] Among them, the weight matrix converts the aggregated hour, week, and weekday embeddings into corresponding hour, week, and weekday patterns respectively;

[0178] The calendar GNN aggregates it into airport unit embeddings according to the session node embeddings:

[0179]

[0180] Finally, all the time patterns and space patterns are concatenated together to obtain the embedding representation of the passenger.

[0181]

[0182] Among them, the initial embeddings of the passenger and multi-behavior are:

[0183] Passenger u m and flight i n The initial embedding can be defined as:

[0184]

[0185] Among them, · represents the matrix multiplication operation. In this way, the corresponding row can be selected from the one-hot vector matrix, and then multiplied by the embedding matrices P and Q to obtain the initial embedding vector of passenger u m and flight i n .

[0186] Among them, the GCN cascade learning for the passenger multi-behavior is:

[0187] By using the transformation matrix W b , the embedding is transformed:

[0188]

[0189] Among them, and represent the transformation matrices of passenger u and flight i respectively. and represent the initial embeddings of passenger u and flight i in the (b + 1)-th behavior.

[0190] It should be noted here that the device description in the above embodiments corresponds to the method description in the embodiments, and the embodiments of the present invention will not be elaborated here.

[0191] The execution entities of the above-mentioned processor and memory can be devices with computing functions such as a computer, a single-chip microcomputer, a microcontroller, etc. In specific implementation, the embodiments of the present invention do not limit the execution entity, and it is selected according to the needs in actual applications.

[0192] Data signals are transmitted between the memory and the processor through a bus, and the embodiments of the present invention will not elaborate on this.

[0193] Based on the same inventive concept, the embodiments of the present invention also provide a computer-readable storage medium. The storage medium includes a stored program, and when the program runs, it controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0194] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0195] It should be noted here that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not elaborate on this.

[0196] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0197] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium. The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium or a semiconductor medium, etc.

[0198] Except for special instructions, the embodiments of the present invention do not limit the models of each device, and any device that can complete the above functions can be used.

[0199] Those skilled in the art can understand that the drawings are only schematic diagrams of a preferred embodiment, and the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0200] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A flight selection method for personalized sequential recommendation of multiple passenger behaviors, characterized in that, The method includes: Dividing the behavior log into sessions according to the purchase behavior of passengers and representing them as session-airport tuples; Mapping the continuous timestamps of the sessions into a set of discrete time units; aggregating the time unit embeddings in three dimensions and generating embeddings representing their respective time patterns: concatenating all the time patterns and spatial patterns together to obtain the embedding representation of the passenger; Utilizing the dependency relationships between behaviors to initialize the embeddings of passengers and multi-behaviors, and performing GCN cascaded learning on the multi-behaviors of passengers; aggregating the features learned from all behaviors in a linear combination manner; Using a multi-layer perceptron to model the attribute representation for each passenger to capture the attribute features of the passenger, integrating the spatio-temporal representation, multi-behavior representation, and passenger attribute representation of the passenger and the flight, and finally capturing the flight preferences of the passenger to provide flight selections more in line with the interests of the passenger.

2. The flight selection method for personalized sequence recommendation of multiple passenger behaviors according to claim 1, characterized in that, The specific operation of dividing the behavior log into sessions according to the purchase behavior of passengers and representing them as session-airport tuples is as follows: Passenger Behavior Diagram G u = session s in (S, A, V, E) i The embedding vector representation is as follows: s i = σ(W S · AGG sess (v i,j |(s i , v i,j ) ∈ E) + b S ) Among them, σ is a non-linear activation function, and the training weight matrix Embed the flight from K V Convert the dimensional space to K S Dimensional session embedding space; Aggregation function AGG sess Use a gated recurrent unit to aggregate the flight vectors within the same session s i ∈S to represent the session s i Representation of.

3. The flight selection method for personalized sequence recommendation of multiple passenger behaviors according to claim 1, characterized in that, The specific embedding representation of the passenger is as follows: Aggregating the session node embeddings into different time unit embeddings: Among them, the weight matrix will transform the session embedding space K S into the hour embedding space K h , the week embedding space K w and the weekday embedding space K y , respectively; AGG temp applies a time aggregation function to all sessions; Among them, the weight matrix respectively converts the aggregated hourly, weekly, and weekday embeddings into the corresponding hourly, weekly, and weekday patterns; The Calendar GNN aggregates them into airport unit embeddings according to the session node embeddings: Finally, concatenating all the time patterns and spatial patterns together to obtain the embedding representation of the passenger:

4. The flight selection method for personalized sequence recommendation of multiple passenger behaviors according to claim 1, characterized in that The embedding initialization of the passenger and multiple behaviors is as follows: passenger u m and flight i n The initialized embedding is defined as: Among them, · represents matrix multiplication operation, select the corresponding rows from the one-hot vector matrix, and then multiply them with the embedding matrices P and Q to obtain the initial embedding vectors of passenger u m and flight i n respectively.

5. The flight selection method for personalized sequence recommendation of multiple passenger behaviors according to claim 1, characterized in that The specific operation of performing GCN cascaded learning on the multi-behaviors of passengers is as follows: By using the transformation matrix W b , transform the embedding: Among them, and respectively represent the transition matrices of passenger u and flight i, and represent the initial embeddings of passenger u and flight i in the (b + 1)-th behavior.

6. A flight selection device for personalized sequence recommendation of multiple passenger behaviors, characterized in that, The device includes: a processor and a memory. Program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by the processor, the processor is enabled to execute the method described in any one of claims 1-5.

Images